fs/btrfs/inode.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Copyright (C) 2007 Oracle.  All rights reserved.
   4  */
   5
   6 #include <linux/kernel.h>
   7 #include <linux/bio.h>
   8 #include <linux/buffer_head.h>
   9 #include <linux/file.h>
  10 #include <linux/fs.h>
  11 #include <linux/pagemap.h>
  12 #include <linux/highmem.h>
  13 #include <linux/time.h>
  14 #include <linux/init.h>
  15 #include <linux/string.h>
  16 #include <linux/backing-dev.h>
  17 #include <linux/writeback.h>
  18 #include <linux/compat.h>
  19 #include <linux/xattr.h>
  20 #include <linux/posix_acl.h>
  21 #include <linux/falloc.h>
  22 #include <linux/slab.h>
  23 #include <linux/ratelimit.h>
  24 #include <linux/btrfs.h>
  25 #include <linux/blkdev.h>
  26 #include <linux/posix_acl_xattr.h>
  27 #include <linux/uio.h>
  28 #include <linux/magic.h>
  29 #include <linux/iversion.h>
  30 #include <linux/swap.h>
  31 #include <linux/sched/mm.h>
  32 #include <asm/unaligned.h>
  33 #include "misc.h"
  34 #include "ctree.h"
  35 #include "disk-io.h"
  36 #include "transaction.h"
  37 #include "btrfs_inode.h"
  38 #include "print-tree.h"
  39 #include "ordered-data.h"
  40 #include "xattr.h"
  41 #include "tree-log.h"
  42 #include "volumes.h"
  43 #include "compression.h"
  44 #include "locking.h"
  45 #include "free-space-cache.h"
  46 #include "inode-map.h"
  47 #include "backref.h"
  48 #include "props.h"
  49 #include "qgroup.h"
  50 #include "delalloc-space.h"
  51 #include "block-group.h"
  52
  53 struct btrfs_iget_args {
  54         struct btrfs_key *location;
  55         struct btrfs_root *root;
  56 };
  57
  58 struct btrfs_dio_data {
  59         u64 reserve;
  60         u64 unsubmitted_oe_range_start;
  61         u64 unsubmitted_oe_range_end;
  62         int overwrite;
  63 };
  64
  65 static const struct inode_operations btrfs_dir_inode_operations;
  66 static const struct inode_operations btrfs_symlink_inode_operations;
  67 static const struct inode_operations btrfs_dir_ro_inode_operations;
  68 static const struct inode_operations btrfs_special_inode_operations;
  69 static const struct inode_operations btrfs_file_inode_operations;
  70 static const struct address_space_operations btrfs_aops;
  71 static const struct file_operations btrfs_dir_file_operations;
  72 static const struct extent_io_ops btrfs_extent_io_ops;
  73
  74 static struct kmem_cache *btrfs_inode_cachep;
  75 struct kmem_cache *btrfs_trans_handle_cachep;
  76 struct kmem_cache *btrfs_path_cachep;
  77 struct kmem_cache *btrfs_free_space_cachep;
  78 struct kmem_cache *btrfs_free_space_bitmap_cachep;
  79
  80 static int btrfs_setsize(struct inode *inode, struct iattr *attr);
  81 static int btrfs_truncate(struct inode *inode, bool skip_writeback);
  82 static int btrfs_finish_ordered_io(struct btrfs_ordered_extent *ordered_extent);
  83 static noinline int cow_file_range(struct inode *inode,
  84                                    struct page *locked_page,
  85                                    u64 start, u64 end, int *page_started,
  86                                    unsigned long *nr_written, int unlock);
  87 static struct extent_map *create_io_em(struct inode *inode, u64 start, u64 len,
  88                                        u64 orig_start, u64 block_start,
  89                                        u64 block_len, u64 orig_block_len,
  90                                        u64 ram_bytes, int compress_type,
  91                                        int type);
  92
  93 static void __endio_write_update_ordered(struct inode *inode,
  94                                          const u64 offset, const u64 bytes,
  95                                          const bool uptodate);
  96
  97 /*
  98  * Cleanup all submitted ordered extents in specified range to handle errors
  99  * from the btrfs_run_delalloc_range() callback.
 100  *
 101  * NOTE: caller must ensure that when an error happens, it can not call
 102  * extent_clear_unlock_delalloc() to clear both the bits EXTENT_DO_ACCOUNTING
 103  * and EXTENT_DELALLOC simultaneously, because that causes the reserved metadata
 104  * to be released, which we want to happen only when finishing the ordered
 105  * extent (btrfs_finish_ordered_io()).
 106  */
 107 static inline void btrfs_cleanup_ordered_extents(struct inode *inode,
 108                                                  struct page *locked_page,
 109                                                  u64 offset, u64 bytes)
 110 {
 111         unsigned long index = offset >> PAGE_SHIFT;
 112         unsigned long end_index = (offset + bytes - 1) >> PAGE_SHIFT;
 113         u64 page_start = page_offset(locked_page);
 114         u64 page_end = page_start + PAGE_SIZE - 1;
 115
 116         struct page *page;
 117
 118         while (index <= end_index) {
 119                 page = find_get_page(inode->i_mapping, index);
 120                 index++;
 121                 if (!page)
 122                         continue;
 123                 ClearPagePrivate2(page);
 124                 put_page(page);
 125         }
 126
 127         /*
 128          * In case this page belongs to the delalloc range being instantiated
 129          * then skip it, since the first page of a range is going to be
 130          * properly cleaned up by the caller of run_delalloc_range
 131          */
 132         if (page_start >= offset && page_end <= (offset + bytes - 1)) {
 133                 offset += PAGE_SIZE;
 134                 bytes -= PAGE_SIZE;
 135         }
 136
 137         return __endio_write_update_ordered(inode, offset, bytes, false);
 138 }
 139
 140 static int btrfs_dirty_inode(struct inode *inode);
 141
 142 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
 143 void btrfs_test_inode_set_ops(struct inode *inode)
 144 {
 145         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
 146 }
 147 #endif
 148
 149 static int btrfs_init_inode_security(struct btrfs_trans_handle *trans,
 150                                      struct inode *inode,  struct inode *dir,
 151                                      const struct qstr *qstr)
 152 {
 153         int err;
 154
 155         err = btrfs_init_acl(trans, inode, dir);
 156         if (!err)
 157                 err = btrfs_xattr_security_init(trans, inode, dir, qstr);
 158         return err;
 159 }
 160
 161 /*
 162  * this does all the hard work for inserting an inline extent into
 163  * the btree.  The caller should have done a btrfs_drop_extents so that
 164  * no overlapping inline items exist in the btree
 165  */
 166 static int insert_inline_extent(struct btrfs_trans_handle *trans,
 167                                 struct btrfs_path *path, int extent_inserted,
 168                                 struct btrfs_root *root, struct inode *inode,
 169                                 u64 start, size_t size, size_t compressed_size,
 170                                 int compress_type,
 171                                 struct page **compressed_pages)
 172 {
 173         struct extent_buffer *leaf;
 174         struct page *page = NULL;
 175         char *kaddr;
 176         unsigned long ptr;
 177         struct btrfs_file_extent_item *ei;
 178         int ret;
 179         size_t cur_size = size;
 180         unsigned long offset;
 181
 182         ASSERT((compressed_size > 0 && compressed_pages) ||
 183                (compressed_size == 0 && !compressed_pages));
 184
 185         if (compressed_size && compressed_pages)
 186                 cur_size = compressed_size;
 187
 188         inode_add_bytes(inode, size);
 189
 190         if (!extent_inserted) {
 191                 struct btrfs_key key;
 192                 size_t datasize;
 193
 194                 key.objectid = btrfs_ino(BTRFS_I(inode));
 195                 key.offset = start;
 196                 key.type = BTRFS_EXTENT_DATA_KEY;
 197
 198                 datasize = btrfs_file_extent_calc_inline_size(cur_size);
 199                 path->leave_spinning = 1;
 200                 ret = btrfs_insert_empty_item(trans, root, path, &key,
 201                                               datasize);
 202                 if (ret)
 203                         goto fail;
 204         }
 205         leaf = path->nodes[0];
 206         ei = btrfs_item_ptr(leaf, path->slots[0],
 207                             struct btrfs_file_extent_item);
 208         btrfs_set_file_extent_generation(leaf, ei, trans->transid);
 209         btrfs_set_file_extent_type(leaf, ei, BTRFS_FILE_EXTENT_INLINE);
 210         btrfs_set_file_extent_encryption(leaf, ei, 0);
 211         btrfs_set_file_extent_other_encoding(leaf, ei, 0);
 212         btrfs_set_file_extent_ram_bytes(leaf, ei, size);
 213         ptr = btrfs_file_extent_inline_start(ei);
 214
 215         if (compress_type != BTRFS_COMPRESS_NONE) {
 216                 struct page *cpage;
 217                 int i = 0;
 218                 while (compressed_size > 0) {
 219                         cpage = compressed_pages[i];
 220                         cur_size = min_t(unsigned long, compressed_size,
 221                                        PAGE_SIZE);
 222
 223                         kaddr = kmap_atomic(cpage);
 224                         write_extent_buffer(leaf, kaddr, ptr, cur_size);
 225                         kunmap_atomic(kaddr);
 226
 227                         i++;
 228                         ptr += cur_size;
 229                         compressed_size -= cur_size;
 230                 }
 231                 btrfs_set_file_extent_compression(leaf, ei,
 232                                                   compress_type);
 233         } else {
 234                 page = find_get_page(inode->i_mapping,
 235                                      start >> PAGE_SHIFT);
 236                 btrfs_set_file_extent_compression(leaf, ei, 0);
 237                 kaddr = kmap_atomic(page);
 238                 offset = offset_in_page(start);
 239                 write_extent_buffer(leaf, kaddr + offset, ptr, size);
 240                 kunmap_atomic(kaddr);
 241                 put_page(page);
 242         }
 243         btrfs_mark_buffer_dirty(leaf);
 244         btrfs_release_path(path);
 245
 246         /*
 247          * we're an inline extent, so nobody can
 248          * extend the file past i_size without locking
 249          * a page we already have locked.
 250          *
 251          * We must do any isize and inode updates
 252          * before we unlock the pages.  Otherwise we
 253          * could end up racing with unlink.
 254          */
 255         BTRFS_I(inode)->disk_i_size = inode->i_size;
 256         ret = btrfs_update_inode(trans, root, inode);
 257
 258 fail:
 259         return ret;
 260 }
 261
 262
 263 /*
 264  * conditionally insert an inline extent into the file.  This
 265  * does the checks required to make sure the data is small enough
 266  * to fit as an inline extent.
 267  */
 268 static noinline int cow_file_range_inline(struct inode *inode, u64 start,
 269                                           u64 end, size_t compressed_size,
 270                                           int compress_type,
 271                                           struct page **compressed_pages)
 272 {
 273         struct btrfs_root *root = BTRFS_I(inode)->root;
 274         struct btrfs_fs_info *fs_info = root->fs_info;
 275         struct btrfs_trans_handle *trans;
 276         u64 isize = i_size_read(inode);
 277         u64 actual_end = min(end + 1, isize);
 278         u64 inline_len = actual_end - start;
 279         u64 aligned_end = ALIGN(end, fs_info->sectorsize);
 280         u64 data_len = inline_len;
 281         int ret;
 282         struct btrfs_path *path;
 283         int extent_inserted = 0;
 284         u32 extent_item_size;
 285
 286         if (compressed_size)
 287                 data_len = compressed_size;
 288
 289         if (start > 0 ||
 290             actual_end > fs_info->sectorsize ||
 291             data_len > BTRFS_MAX_INLINE_DATA_SIZE(fs_info) ||
 292             (!compressed_size &&
 293             (actual_end & (fs_info->sectorsize - 1)) == 0) ||
 294             end + 1 < isize ||
 295             data_len > fs_info->max_inline) {
 296                 return 1;
 297         }
 298
 299         path = btrfs_alloc_path();
 300         if (!path)
 301                 return -ENOMEM;
 302
 303         trans = btrfs_join_transaction(root);
 304         if (IS_ERR(trans)) {
 305                 btrfs_free_path(path);
 306                 return PTR_ERR(trans);
 307         }
 308         trans->block_rsv = &BTRFS_I(inode)->block_rsv;
 309
 310         if (compressed_size && compressed_pages)
 311                 extent_item_size = btrfs_file_extent_calc_inline_size(
 312                    compressed_size);
 313         else
 314                 extent_item_size = btrfs_file_extent_calc_inline_size(
 315                     inline_len);
 316
 317         ret = __btrfs_drop_extents(trans, root, inode, path,
 318                                    start, aligned_end, NULL,
 319                                    1, 1, extent_item_size, &extent_inserted);
 320         if (ret) {
 321                 btrfs_abort_transaction(trans, ret);
 322                 goto out;
 323         }
 324
 325         if (isize > actual_end)
 326                 inline_len = min_t(u64, isize, actual_end);
 327         ret = insert_inline_extent(trans, path, extent_inserted,
 328                                    root, inode, start,
 329                                    inline_len, compressed_size,
 330                                    compress_type, compressed_pages);
 331         if (ret && ret != -ENOSPC) {
 332                 btrfs_abort_transaction(trans, ret);
 333                 goto out;
 334         } else if (ret == -ENOSPC) {
 335                 ret = 1;
 336                 goto out;
 337         }
 338
 339         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &BTRFS_I(inode)->runtime_flags);
 340         btrfs_drop_extent_cache(BTRFS_I(inode), start, aligned_end - 1, 0);
 341 out:
 342         /*
 343          * Don't forget to free the reserved space, as for inlined extent
 344          * it won't count as data extent, free them directly here.
 345          * And at reserve time, it's always aligned to page size, so
 346          * just free one page here.
 347          */
 348         btrfs_qgroup_free_data(inode, NULL, 0, PAGE_SIZE);
 349         btrfs_free_path(path);
 350         btrfs_end_transaction(trans);
 351         return ret;
 352 }
 353
 354 struct async_extent {
 355         u64 start;
 356         u64 ram_size;
 357         u64 compressed_size;
 358         struct page **pages;
 359         unsigned long nr_pages;
 360         int compress_type;
 361         struct list_head list;
 362 };
 363
 364 struct async_chunk {
 365         struct inode *inode;
 366         struct page *locked_page;
 367         u64 start;
 368         u64 end;
 369         unsigned int write_flags;
 370         struct list_head extents;
 371         struct cgroup_subsys_state *blkcg_css;
 372         struct btrfs_work work;
 373         atomic_t *pending;
 374 };
 375
 376 struct async_cow {
 377         /* Number of chunks in flight; must be first in the structure */
 378         atomic_t num_chunks;
 379         struct async_chunk chunks[];
 380 };
 381
 382 static noinline int add_async_extent(struct async_chunk *cow,
 383                                      u64 start, u64 ram_size,
 384                                      u64 compressed_size,
 385                                      struct page **pages,
 386                                      unsigned long nr_pages,
 387                                      int compress_type)
 388 {
 389         struct async_extent *async_extent;
 390
 391         async_extent = kmalloc(sizeof(*async_extent), GFP_NOFS);
 392         BUG_ON(!async_extent); /* -ENOMEM */
 393         async_extent->start = start;
 394         async_extent->ram_size = ram_size;
 395         async_extent->compressed_size = compressed_size;
 396         async_extent->pages = pages;
 397         async_extent->nr_pages = nr_pages;
 398         async_extent->compress_type = compress_type;
 399         list_add_tail(&async_extent->list, &cow->extents);
 400         return 0;
 401 }
 402
 403 /*
 404  * Check if the inode has flags compatible with compression
 405  */
 406 static inline bool inode_can_compress(struct inode *inode)
 407 {
 408         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW ||
 409             BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)
 410                 return false;
 411         return true;
 412 }
 413
 414 /*
 415  * Check if the inode needs to be submitted to compression, based on mount
 416  * options, defragmentation, properties or heuristics.
 417  */
 418 static inline int inode_need_compress(struct inode *inode, u64 start, u64 end)
 419 {
 420         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
 421
 422         if (!inode_can_compress(inode)) {
 423                 WARN(IS_ENABLED(CONFIG_BTRFS_DEBUG),
 424                         KERN_ERR "BTRFS: unexpected compression for ino %llu\n",
 425                         btrfs_ino(BTRFS_I(inode)));
 426                 return 0;
 427         }
 428         /* force compress */
 429         if (btrfs_test_opt(fs_info, FORCE_COMPRESS))
 430                 return 1;
 431         /* defrag ioctl */
 432         if (BTRFS_I(inode)->defrag_compress)
 433                 return 1;
 434         /* bad compression ratios */
 435         if (BTRFS_I(inode)->flags & BTRFS_INODE_NOCOMPRESS)
 436                 return 0;
 437         if (btrfs_test_opt(fs_info, COMPRESS) ||
 438             BTRFS_I(inode)->flags & BTRFS_INODE_COMPRESS ||
 439             BTRFS_I(inode)->prop_compress)
 440                 return btrfs_compress_heuristic(inode, start, end);
 441         return 0;
 442 }
 443
 444 static inline void inode_should_defrag(struct btrfs_inode *inode,
 445                 u64 start, u64 end, u64 num_bytes, u64 small_write)
 446 {
 447         /* If this is a small write inside eof, kick off a defrag */
 448         if (num_bytes < small_write &&
 449             (start > 0 || end + 1 < inode->disk_i_size))
 450                 btrfs_add_inode_defrag(NULL, inode);
 451 }
 452
 453 /*
 454  * we create compressed extents in two phases.  The first
 455  * phase compresses a range of pages that have already been
 456  * locked (both pages and state bits are locked).
 457  *
 458  * This is done inside an ordered work queue, and the compression
 459  * is spread across many cpus.  The actual IO submission is step
 460  * two, and the ordered work queue takes care of making sure that
 461  * happens in the same order things were put onto the queue by
 462  * writepages and friends.
 463  *
 464  * If this code finds it can't get good compression, it puts an
 465  * entry onto the work queue to write the uncompressed bytes.  This
 466  * makes sure that both compressed inodes and uncompressed inodes
 467  * are written in the same order that the flusher thread sent them
 468  * down.
 469  */
 470 static noinline int compress_file_range(struct async_chunk *async_chunk)
 471 {
 472         struct inode *inode = async_chunk->inode;
 473         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
 474         u64 blocksize = fs_info->sectorsize;
 475         u64 start = async_chunk->start;
 476         u64 end = async_chunk->end;
 477         u64 actual_end;
 478         u64 i_size;
 479         int ret = 0;
 480         struct page **pages = NULL;
 481         unsigned long nr_pages;
 482         unsigned long total_compressed = 0;
 483         unsigned long total_in = 0;
 484         int i;
 485         int will_compress;
 486         int compress_type = fs_info->compress_type;
 487         int compressed_extents = 0;
 488         int redirty = 0;
 489
 490         inode_should_defrag(BTRFS_I(inode), start, end, end - start + 1,
 491                         SZ_16K);
 492
 493         /*
 494          * We need to save i_size before now because it could change in between
 495          * us evaluating the size and assigning it.  This is because we lock and
 496          * unlock the page in truncate and fallocate, and then modify the i_size
 497          * later on.
 498          *
 499          * The barriers are to emulate READ_ONCE, remove that once i_size_read
 500          * does that for us.
 501          */
 502         barrier();
 503         i_size = i_size_read(inode);
 504         barrier();
 505         actual_end = min_t(u64, i_size, end + 1);
 506 again:
 507         will_compress = 0;
 508         nr_pages = (end >> PAGE_SHIFT) - (start >> PAGE_SHIFT) + 1;
 509         BUILD_BUG_ON((BTRFS_MAX_COMPRESSED % PAGE_SIZE) != 0);
 510         nr_pages = min_t(unsigned long, nr_pages,
 511                         BTRFS_MAX_COMPRESSED / PAGE_SIZE);
 512
 513         /*
 514          * we don't want to send crud past the end of i_size through
 515          * compression, that's just a waste of CPU time.  So, if the
 516          * end of the file is before the start of our current
 517          * requested range of bytes, we bail out to the uncompressed
 518          * cleanup code that can deal with all of this.
 519          *
 520          * It isn't really the fastest way to fix things, but this is a
 521          * very uncommon corner.
 522          */
 523         if (actual_end <= start)
 524                 goto cleanup_and_bail_uncompressed;
 525
 526         total_compressed = actual_end - start;
 527
 528         /*
 529          * skip compression for a small file range(<=blocksize) that
 530          * isn't an inline extent, since it doesn't save disk space at all.
 531          */
 532         if (total_compressed <= blocksize &&
 533            (start > 0 || end + 1 < BTRFS_I(inode)->disk_i_size))
 534                 goto cleanup_and_bail_uncompressed;
 535
 536         total_compressed = min_t(unsigned long, total_compressed,
 537                         BTRFS_MAX_UNCOMPRESSED);
 538         total_in = 0;
 539         ret = 0;
 540
 541         /*
 542          * we do compression for mount -o compress and when the
 543          * inode has not been flagged as nocompress.  This flag can
 544          * change at any time if we discover bad compression ratios.
 545          */
 546         if (inode_need_compress(inode, start, end)) {
 547                 WARN_ON(pages);
 548                 pages = kcalloc(nr_pages, sizeof(struct page *), GFP_NOFS);
 549                 if (!pages) {
 550                         /* just bail out to the uncompressed code */
 551                         nr_pages = 0;
 552                         goto cont;
 553                 }
 554
 555                 if (BTRFS_I(inode)->defrag_compress)
 556                         compress_type = BTRFS_I(inode)->defrag_compress;
 557                 else if (BTRFS_I(inode)->prop_compress)
 558                         compress_type = BTRFS_I(inode)->prop_compress;
 559
 560                 /*
 561                  * we need to call clear_page_dirty_for_io on each
 562                  * page in the range.  Otherwise applications with the file
 563                  * mmap'd can wander in and change the page contents while
 564                  * we are compressing them.
 565                  *
 566                  * If the compression fails for any reason, we set the pages
 567                  * dirty again later on.
 568                  *
 569                  * Note that the remaining part is redirtied, the start pointer
 570                  * has moved, the end is the original one.
 571                  */
 572                 if (!redirty) {
 573                         extent_range_clear_dirty_for_io(inode, start, end);
 574                         redirty = 1;
 575                 }
 576
 577                 /* Compression level is applied here and only here */
 578                 ret = btrfs_compress_pages(
 579                         compress_type | (fs_info->compress_level << 4),
 580                                            inode->i_mapping, start,
 581                                            pages,
 582                                            &nr_pages,
 583                                            &total_in,
 584                                            &total_compressed);
 585
 586                 if (!ret) {
 587                         unsigned long offset = offset_in_page(total_compressed);
 588                         struct page *page = pages[nr_pages - 1];
 589                         char *kaddr;
 590
 591                         /* zero the tail end of the last page, we might be
 592                          * sending it down to disk
 593                          */
 594                         if (offset) {
 595                                 kaddr = kmap_atomic(page);
 596                                 memset(kaddr + offset, 0,
 597                                        PAGE_SIZE - offset);
 598                                 kunmap_atomic(kaddr);
 599                         }
 600                         will_compress = 1;
 601                 }
 602         }
 603 cont:
 604         if (start == 0) {
 605                 /* lets try to make an inline extent */
 606                 if (ret || total_in < actual_end) {
 607                         /* we didn't compress the entire range, try
 608                          * to make an uncompressed inline extent.
 609                          */
 610                         ret = cow_file_range_inline(inode, start, end, 0,
 611                                                     BTRFS_COMPRESS_NONE, NULL);
 612                 } else {
 613                         /* try making a compressed inline extent */
 614                         ret = cow_file_range_inline(inode, start, end,
 615                                                     total_compressed,
 616                                                     compress_type, pages);
 617                 }
 618                 if (ret <= 0) {
 619                         unsigned long clear_flags = EXTENT_DELALLOC |
 620                                 EXTENT_DELALLOC_NEW | EXTENT_DEFRAG |
 621                                 EXTENT_DO_ACCOUNTING;
 622                         unsigned long page_error_op;
 623
 624                         page_error_op = ret < 0 ? PAGE_SET_ERROR : 0;
 625
 626                         /*
 627                          * inline extent creation worked or returned error,
 628                          * we don't need to create any more async work items.
 629                          * Unlock and free up our temp pages.
 630                          *
 631                          * We use DO_ACCOUNTING here because we need the
 632                          * delalloc_release_metadata to be done _after_ we drop
 633                          * our outstanding extent for clearing delalloc for this
 634                          * range.
 635                          */
 636                         extent_clear_unlock_delalloc(inode, start, end, NULL,
 637                                                      clear_flags,
 638                                                      PAGE_UNLOCK |
 639                                                      PAGE_CLEAR_DIRTY |
 640                                                      PAGE_SET_WRITEBACK |
 641                                                      page_error_op |
 642                                                      PAGE_END_WRITEBACK);
 643
 644                         for (i = 0; i < nr_pages; i++) {
 645                                 WARN_ON(pages[i]->mapping);
 646                                 put_page(pages[i]);
 647                         }
 648                         kfree(pages);
 649
 650                         return 0;
 651                 }
 652         }
 653
 654         if (will_compress) {
 655                 /*
 656                  * we aren't doing an inline extent round the compressed size
 657                  * up to a block size boundary so the allocator does sane
 658                  * things
 659                  */
 660                 total_compressed = ALIGN(total_compressed, blocksize);
 661
 662                 /*
 663                  * one last check to make sure the compression is really a
 664                  * win, compare the page count read with the blocks on disk,
 665                  * compression must free at least one sector size
 666                  */
 667                 total_in = ALIGN(total_in, PAGE_SIZE);
 668                 if (total_compressed + blocksize <= total_in) {
 669                         compressed_extents++;
 670
 671                         /*
 672                          * The async work queues will take care of doing actual
 673                          * allocation on disk for these compressed pages, and
 674                          * will submit them to the elevator.
 675                          */
 676                         add_async_extent(async_chunk, start, total_in,
 677                                         total_compressed, pages, nr_pages,
 678                                         compress_type);
 679
 680                         if (start + total_in < end) {
 681                                 start += total_in;
 682                                 pages = NULL;
 683                                 cond_resched();
 684                                 goto again;
 685                         }
 686                         return compressed_extents;
 687                 }
 688         }
 689         if (pages) {
 690                 /*
 691                  * the compression code ran but failed to make things smaller,
 692                  * free any pages it allocated and our page pointer array
 693                  */
 694                 for (i = 0; i < nr_pages; i++) {
 695                         WARN_ON(pages[i]->mapping);
 696                         put_page(pages[i]);
 697                 }
 698                 kfree(pages);
 699                 pages = NULL;
 700                 total_compressed = 0;
 701                 nr_pages = 0;
 702
 703                 /* flag the file so we don't compress in the future */
 704                 if (!btrfs_test_opt(fs_info, FORCE_COMPRESS) &&
 705                     !(BTRFS_I(inode)->prop_compress)) {
 706                         BTRFS_I(inode)->flags |= BTRFS_INODE_NOCOMPRESS;
 707                 }
 708         }
 709 cleanup_and_bail_uncompressed:
 710         /*
 711          * No compression, but we still need to write the pages in the file
 712          * we've been given so far.  redirty the locked page if it corresponds
 713          * to our extent and set things up for the async work queue to run
 714          * cow_file_range to do the normal delalloc dance.
 715          */
 716         if (async_chunk->locked_page &&
 717             (page_offset(async_chunk->locked_page) >= start &&
 718              page_offset(async_chunk->locked_page)) <= end) {
 719                 __set_page_dirty_nobuffers(async_chunk->locked_page);
 720                 /* unlocked later on in the async handlers */
 721         }
 722
 723         if (redirty)
 724                 extent_range_redirty_for_io(inode, start, end);
 725         add_async_extent(async_chunk, start, end - start + 1, 0, NULL, 0,
 726                          BTRFS_COMPRESS_NONE);
 727         compressed_extents++;
 728
 729         return compressed_extents;
 730 }
 731
 732 static void free_async_extent_pages(struct async_extent *async_extent)
 733 {
 734         int i;
 735
 736         if (!async_extent->pages)
 737                 return;
 738
 739         for (i = 0; i < async_extent->nr_pages; i++) {
 740                 WARN_ON(async_extent->pages[i]->mapping);
 741                 put_page(async_extent->pages[i]);
 742         }
 743         kfree(async_extent->pages);
 744         async_extent->nr_pages = 0;
 745         async_extent->pages = NULL;
 746 }
 747
 748 /*
 749  * phase two of compressed writeback.  This is the ordered portion
 750  * of the code, which only gets called in the order the work was
 751  * queued.  We walk all the async extents created by compress_file_range
 752  * and send them down to the disk.
 753  */
 754 static noinline void submit_compressed_extents(struct async_chunk *async_chunk)
 755 {
 756         struct inode *inode = async_chunk->inode;
 757         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
 758         struct async_extent *async_extent;
 759         u64 alloc_hint = 0;
 760         struct btrfs_key ins;
 761         struct extent_map *em;
 762         struct btrfs_root *root = BTRFS_I(inode)->root;
 763         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
 764         int ret = 0;
 765
 766 again:
 767         while (!list_empty(&async_chunk->extents)) {
 768                 async_extent = list_entry(async_chunk->extents.next,
 769                                           struct async_extent, list);
 770                 list_del(&async_extent->list);
 771
 772 retry:
 773                 lock_extent(io_tree, async_extent->start,
 774                             async_extent->start + async_extent->ram_size - 1);
 775                 /* did the compression code fall back to uncompressed IO? */
 776                 if (!async_extent->pages) {
 777                         int page_started = 0;
 778                         unsigned long nr_written = 0;
 779
 780                         /* allocate blocks */
 781                         ret = cow_file_range(inode, async_chunk->locked_page,
 782                                              async_extent->start,
 783                                              async_extent->start +
 784                                              async_extent->ram_size - 1,
 785                                              &page_started, &nr_written, 0);
 786
 787                         /* JDM XXX */
 788
 789                         /*
 790                          * if page_started, cow_file_range inserted an
 791                          * inline extent and took care of all the unlocking
 792                          * and IO for us.  Otherwise, we need to submit
 793                          * all those pages down to the drive.
 794                          */
 795                         if (!page_started && !ret)
 796                                 extent_write_locked_range(inode,
 797                                                   async_extent->start,
 798                                                   async_extent->start +
 799                                                   async_extent->ram_size - 1,
 800                                                   WB_SYNC_ALL);
 801                         else if (ret && async_chunk->locked_page)
 802                                 unlock_page(async_chunk->locked_page);
 803                         kfree(async_extent);
 804                         cond_resched();
 805                         continue;
 806                 }
 807
 808                 ret = btrfs_reserve_extent(root, async_extent->ram_size,
 809                                            async_extent->compressed_size,
 810                                            async_extent->compressed_size,
 811                                            0, alloc_hint, &ins, 1, 1);
 812                 if (ret) {
 813                         free_async_extent_pages(async_extent);
 814
 815                         if (ret == -ENOSPC) {
 816                                 unlock_extent(io_tree, async_extent->start,
 817                                               async_extent->start +
 818                                               async_extent->ram_size - 1);
 819
 820                                 /*
 821                                  * we need to redirty the pages if we decide to
 822                                  * fallback to uncompressed IO, otherwise we
 823                                  * will not submit these pages down to lower
 824                                  * layers.
 825                                  */
 826                                 extent_range_redirty_for_io(inode,
 827                                                 async_extent->start,
 828                                                 async_extent->start +
 829                                                 async_extent->ram_size - 1);
 830
 831                                 goto retry;
 832                         }
 833                         goto out_free;
 834                 }
 835                 /*
 836                  * here we're doing allocation and writeback of the
 837                  * compressed pages
 838                  */
 839                 em = create_io_em(inode, async_extent->start,
 840                                   async_extent->ram_size, /* len */
 841                                   async_extent->start, /* orig_start */
 842                                   ins.objectid, /* block_start */
 843                                   ins.offset, /* block_len */
 844                                   ins.offset, /* orig_block_len */
 845                                   async_extent->ram_size, /* ram_bytes */
 846                                   async_extent->compress_type,
 847                                   BTRFS_ORDERED_COMPRESSED);
 848                 if (IS_ERR(em))
 849                         /* ret value is not necessary due to void function */
 850                         goto out_free_reserve;
 851                 free_extent_map(em);
 852
 853                 ret = btrfs_add_ordered_extent_compress(inode,
 854                                                 async_extent->start,
 855                                                 ins.objectid,
 856                                                 async_extent->ram_size,
 857                                                 ins.offset,
 858                                                 BTRFS_ORDERED_COMPRESSED,
 859                                                 async_extent->compress_type);
 860                 if (ret) {
 861                         btrfs_drop_extent_cache(BTRFS_I(inode),
 862                                                 async_extent->start,
 863                                                 async_extent->start +
 864                                                 async_extent->ram_size - 1, 0);
 865                         goto out_free_reserve;
 866                 }
 867                 btrfs_dec_block_group_reservations(fs_info, ins.objectid);
 868
 869                 /*
 870                  * clear dirty, set writeback and unlock the pages.
 871                  */
 872                 extent_clear_unlock_delalloc(inode, async_extent->start,
 873                                 async_extent->start +
 874                                 async_extent->ram_size - 1,
 875                                 NULL, EXTENT_LOCKED | EXTENT_DELALLOC,
 876                                 PAGE_UNLOCK | PAGE_CLEAR_DIRTY |
 877                                 PAGE_SET_WRITEBACK);
 878                 if (btrfs_submit_compressed_write(inode,
 879                                     async_extent->start,
 880                                     async_extent->ram_size,
 881                                     ins.objectid,
 882                                     ins.offset, async_extent->pages,
 883                                     async_extent->nr_pages,
 884                                     async_chunk->write_flags,
 885                                     async_chunk->blkcg_css)) {
 886                         struct page *p = async_extent->pages[0];
 887                         const u64 start = async_extent->start;
 888                         const u64 end = start + async_extent->ram_size - 1;
 889
 890                         p->mapping = inode->i_mapping;
 891                         btrfs_writepage_endio_finish_ordered(p, start, end, 0);
 892
 893                         p->mapping = NULL;
 894                         extent_clear_unlock_delalloc(inode, start, end,
 895                                                      NULL, 0,
 896                                                      PAGE_END_WRITEBACK |
 897                                                      PAGE_SET_ERROR);
 898                         free_async_extent_pages(async_extent);
 899                 }
 900                 alloc_hint = ins.objectid + ins.offset;
 901                 kfree(async_extent);
 902                 cond_resched();
 903         }
 904         return;
 905 out_free_reserve:
 906         btrfs_dec_block_group_reservations(fs_info, ins.objectid);
 907         btrfs_free_reserved_extent(fs_info, ins.objectid, ins.offset, 1);
 908 out_free:
 909         extent_clear_unlock_delalloc(inode, async_extent->start,
 910                                      async_extent->start +
 911                                      async_extent->ram_size - 1,
 912                                      NULL, EXTENT_LOCKED | EXTENT_DELALLOC |
 913                                      EXTENT_DELALLOC_NEW |
 914                                      EXTENT_DEFRAG | EXTENT_DO_ACCOUNTING,
 915                                      PAGE_UNLOCK | PAGE_CLEAR_DIRTY |
 916                                      PAGE_SET_WRITEBACK | PAGE_END_WRITEBACK |
 917                                      PAGE_SET_ERROR);
 918         free_async_extent_pages(async_extent);
 919         kfree(async_extent);
 920         goto again;
 921 }
 922
 923 static u64 get_extent_allocation_hint(struct inode *inode, u64 start,
 924                                       u64 num_bytes)
 925 {
 926         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
 927         struct extent_map *em;
 928         u64 alloc_hint = 0;
 929
 930         read_lock(&em_tree->lock);
 931         em = search_extent_mapping(em_tree, start, num_bytes);
 932         if (em) {
 933                 /*
 934                  * if block start isn't an actual block number then find the
 935                  * first block in this inode and use that as a hint.  If that
 936                  * block is also bogus then just don't worry about it.
 937                  */
 938                 if (em->block_start >= EXTENT_MAP_LAST_BYTE) {
 939                         free_extent_map(em);
 940                         em = search_extent_mapping(em_tree, 0, 0);
 941                         if (em && em->block_start < EXTENT_MAP_LAST_BYTE)
 942                                 alloc_hint = em->block_start;
 943                         if (em)
 944                                 free_extent_map(em);
 945                 } else {
 946                         alloc_hint = em->block_start;
 947                         free_extent_map(em);
 948                 }
 949         }
 950         read_unlock(&em_tree->lock);
 951
 952         return alloc_hint;
 953 }
 954
 955 /*
 956  * when extent_io.c finds a delayed allocation range in the file,
 957  * the call backs end up in this code.  The basic idea is to
 958  * allocate extents on disk for the range, and create ordered data structs
 959  * in ram to track those extents.
 960  *
 961  * locked_page is the page that writepage had locked already.  We use
 962  * it to make sure we don't do extra locks or unlocks.
 963  *
 964  * *page_started is set to one if we unlock locked_page and do everything
 965  * required to start IO on it.  It may be clean and already done with
 966  * IO when we return.
 967  */
 968 static noinline int cow_file_range(struct inode *inode,
 969                                    struct page *locked_page,
 970                                    u64 start, u64 end, int *page_started,
 971                                    unsigned long *nr_written, int unlock)
 972 {
 973         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
 974         struct btrfs_root *root = BTRFS_I(inode)->root;
 975         u64 alloc_hint = 0;
 976         u64 num_bytes;
 977         unsigned long ram_size;
 978         u64 cur_alloc_size = 0;
 979         u64 blocksize = fs_info->sectorsize;
 980         struct btrfs_key ins;
 981         struct extent_map *em;
 982         unsigned clear_bits;
 983         unsigned long page_ops;
 984         bool extent_reserved = false;
 985         int ret = 0;
 986
 987         if (btrfs_is_free_space_inode(BTRFS_I(inode))) {
 988                 WARN_ON_ONCE(1);
 989                 ret = -EINVAL;
 990                 goto out_unlock;
 991         }
 992
 993         num_bytes = ALIGN(end - start + 1, blocksize);
 994         num_bytes = max(blocksize,  num_bytes);
 995         ASSERT(num_bytes <= btrfs_super_total_bytes(fs_info->super_copy));
 996
 997         inode_should_defrag(BTRFS_I(inode), start, end, num_bytes, SZ_64K);
 998
 999         if (start == 0) {
1000                 /* lets try to make an inline extent */
1001                 ret = cow_file_range_inline(inode, start, end, 0,
1002                                             BTRFS_COMPRESS_NONE, NULL);
1003                 if (ret == 0) {
1004                         /*
1005                          * We use DO_ACCOUNTING here because we need the
1006                          * delalloc_release_metadata to be run _after_ we drop
1007                          * our outstanding extent for clearing delalloc for this
1008                          * range.
1009                          */
1010                         extent_clear_unlock_delalloc(inode, start, end, NULL,
1011                                      EXTENT_LOCKED | EXTENT_DELALLOC |
1012                                      EXTENT_DELALLOC_NEW | EXTENT_DEFRAG |
1013                                      EXTENT_DO_ACCOUNTING, PAGE_UNLOCK |
1014                                      PAGE_CLEAR_DIRTY | PAGE_SET_WRITEBACK |
1015                                      PAGE_END_WRITEBACK);
1016                         *nr_written = *nr_written +
1017                              (end - start + PAGE_SIZE) / PAGE_SIZE;
1018                         *page_started = 1;
1019                         goto out;
1020                 } else if (ret < 0) {
1021                         goto out_unlock;
1022                 }
1023         }
1024
1025         alloc_hint = get_extent_allocation_hint(inode, start, num_bytes);
1026         btrfs_drop_extent_cache(BTRFS_I(inode), start,
1027                         start + num_bytes - 1, 0);
1028
1029         while (num_bytes > 0) {
1030                 cur_alloc_size = num_bytes;
1031                 ret = btrfs_reserve_extent(root, cur_alloc_size, cur_alloc_size,
1032                                            fs_info->sectorsize, 0, alloc_hint,
1033                                            &ins, 1, 1);
1034                 if (ret < 0)
1035                         goto out_unlock;
1036                 cur_alloc_size = ins.offset;
1037                 extent_reserved = true;
1038
1039                 ram_size = ins.offset;
1040                 em = create_io_em(inode, start, ins.offset, /* len */
1041                                   start, /* orig_start */
1042                                   ins.objectid, /* block_start */
1043                                   ins.offset, /* block_len */
1044                                   ins.offset, /* orig_block_len */
1045                                   ram_size, /* ram_bytes */
1046                                   BTRFS_COMPRESS_NONE, /* compress_type */
1047                                   BTRFS_ORDERED_REGULAR /* type */);
1048                 if (IS_ERR(em)) {
1049                         ret = PTR_ERR(em);
1050                         goto out_reserve;
1051                 }
1052                 free_extent_map(em);
1053
1054                 ret = btrfs_add_ordered_extent(inode, start, ins.objectid,
1055                                                ram_size, cur_alloc_size, 0);
1056                 if (ret)
1057                         goto out_drop_extent_cache;
1058
1059                 if (root->root_key.objectid ==
1060                     BTRFS_DATA_RELOC_TREE_OBJECTID) {
1061                         ret = btrfs_reloc_clone_csums(inode, start,
1062                                                       cur_alloc_size);
1063                         /*
1064                          * Only drop cache here, and process as normal.
1065                          *
1066                          * We must not allow extent_clear_unlock_delalloc()
1067                          * at out_unlock label to free meta of this ordered
1068                          * extent, as its meta should be freed by
1069                          * btrfs_finish_ordered_io().
1070                          *
1071                          * So we must continue until @start is increased to
1072                          * skip current ordered extent.
1073                          */
1074                         if (ret)
1075                                 btrfs_drop_extent_cache(BTRFS_I(inode), start,
1076                                                 start + ram_size - 1, 0);
1077                 }
1078
1079                 btrfs_dec_block_group_reservations(fs_info, ins.objectid);
1080
1081                 /* we're not doing compressed IO, don't unlock the first
1082                  * page (which the caller expects to stay locked), don't
1083                  * clear any dirty bits and don't set any writeback bits
1084                  *
1085                  * Do set the Private2 bit so we know this page was properly
1086                  * setup for writepage
1087                  */
1088                 page_ops = unlock ? PAGE_UNLOCK : 0;
1089                 page_ops |= PAGE_SET_PRIVATE2;
1090
1091                 extent_clear_unlock_delalloc(inode, start,
1092                                              start + ram_size - 1,
1093                                              locked_page,
1094                                              EXTENT_LOCKED | EXTENT_DELALLOC,
1095                                              page_ops);
1096                 if (num_bytes < cur_alloc_size)
1097                         num_bytes = 0;
1098                 else
1099                         num_bytes -= cur_alloc_size;
1100                 alloc_hint = ins.objectid + ins.offset;
1101                 start += cur_alloc_size;
1102                 extent_reserved = false;
1103
1104                 /*
1105                  * btrfs_reloc_clone_csums() error, since start is increased
1106                  * extent_clear_unlock_delalloc() at out_unlock label won't
1107                  * free metadata of current ordered extent, we're OK to exit.
1108                  */
1109                 if (ret)
1110                         goto out_unlock;
1111         }
1112 out:
1113         return ret;
1114
1115 out_drop_extent_cache:
1116         btrfs_drop_extent_cache(BTRFS_I(inode), start, start + ram_size - 1, 0);
1117 out_reserve:
1118         btrfs_dec_block_group_reservations(fs_info, ins.objectid);
1119         btrfs_free_reserved_extent(fs_info, ins.objectid, ins.offset, 1);
1120 out_unlock:
1121         clear_bits = EXTENT_LOCKED | EXTENT_DELALLOC | EXTENT_DELALLOC_NEW |
1122                 EXTENT_DEFRAG | EXTENT_CLEAR_META_RESV;
1123         page_ops = PAGE_UNLOCK | PAGE_CLEAR_DIRTY | PAGE_SET_WRITEBACK |
1124                 PAGE_END_WRITEBACK;
1125         /*
1126          * If we reserved an extent for our delalloc range (or a subrange) and
1127          * failed to create the respective ordered extent, then it means that
1128          * when we reserved the extent we decremented the extent's size from
1129          * the data space_info's bytes_may_use counter and incremented the
1130          * space_info's bytes_reserved counter by the same amount. We must make
1131          * sure extent_clear_unlock_delalloc() does not try to decrement again
1132          * the data space_info's bytes_may_use counter, therefore we do not pass
1133          * it the flag EXTENT_CLEAR_DATA_RESV.
1134          */
1135         if (extent_reserved) {
1136                 extent_clear_unlock_delalloc(inode, start,
1137                                              start + cur_alloc_size,
1138                                              locked_page,
1139                                              clear_bits,
1140                                              page_ops);
1141                 start += cur_alloc_size;
1142                 if (start >= end)
1143                         goto out;
1144         }
1145         extent_clear_unlock_delalloc(inode, start, end, locked_page,
1146                                      clear_bits | EXTENT_CLEAR_DATA_RESV,
1147                                      page_ops);
1148         goto out;
1149 }
1150
1151 /*
1152  * work queue call back to started compression on a file and pages
1153  */
1154 static noinline void async_cow_start(struct btrfs_work *work)
1155 {
1156         struct async_chunk *async_chunk;
1157         int compressed_extents;
1158
1159         async_chunk = container_of(work, struct async_chunk, work);
1160
1161         compressed_extents = compress_file_range(async_chunk);
1162         if (compressed_extents == 0) {
1163                 btrfs_add_delayed_iput(async_chunk->inode);
1164                 async_chunk->inode = NULL;
1165         }
1166 }
1167
1168 /*
1169  * work queue call back to submit previously compressed pages
1170  */
1171 static noinline void async_cow_submit(struct btrfs_work *work)
1172 {
1173         struct async_chunk *async_chunk = container_of(work, struct async_chunk,
1174                                                      work);
1175         struct btrfs_fs_info *fs_info = btrfs_work_owner(work);
1176         unsigned long nr_pages;
1177
1178         nr_pages = (async_chunk->end - async_chunk->start + PAGE_SIZE) >>
1179                 PAGE_SHIFT;
1180
1181         /* atomic_sub_return implies a barrier */
1182         if (atomic_sub_return(nr_pages, &fs_info->async_delalloc_pages) <
1183             5 * SZ_1M)
1184                 cond_wake_up_nomb(&fs_info->async_submit_wait);
1185
1186         /*
1187          * ->inode could be NULL if async_chunk_start has failed to compress,
1188          * in which case we don't have anything to submit, yet we need to
1189          * always adjust ->async_delalloc_pages as its paired with the init
1190          * happening in cow_file_range_async
1191          */
1192         if (async_chunk->inode)
1193                 submit_compressed_extents(async_chunk);
1194 }
1195
1196 static noinline void async_cow_free(struct btrfs_work *work)
1197 {
1198         struct async_chunk *async_chunk;
1199
1200         async_chunk = container_of(work, struct async_chunk, work);
1201         if (async_chunk->inode)
1202                 btrfs_add_delayed_iput(async_chunk->inode);
1203         if (async_chunk->blkcg_css)
1204                 css_put(async_chunk->blkcg_css);
1205         /*
1206          * Since the pointer to 'pending' is at the beginning of the array of
1207          * async_chunk's, freeing it ensures the whole array has been freed.
1208          */
1209         if (atomic_dec_and_test(async_chunk->pending))
1210                 kvfree(async_chunk->pending);
1211 }
1212
1213 static int cow_file_range_async(struct inode *inode,
1214                                 struct writeback_control *wbc,
1215                                 struct page *locked_page,
1216                                 u64 start, u64 end, int *page_started,
1217                                 unsigned long *nr_written)
1218 {
1219         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1220         struct cgroup_subsys_state *blkcg_css = wbc_blkcg_css(wbc);
1221         struct async_cow *ctx;
1222         struct async_chunk *async_chunk;
1223         unsigned long nr_pages;
1224         u64 cur_end;
1225         u64 num_chunks = DIV_ROUND_UP(end - start, SZ_512K);
1226         int i;
1227         bool should_compress;
1228         unsigned nofs_flag;
1229         const unsigned int write_flags = wbc_to_write_flags(wbc);
1230
1231         unlock_extent(&BTRFS_I(inode)->io_tree, start, end);
1232
1233         if (BTRFS_I(inode)->flags & BTRFS_INODE_NOCOMPRESS &&
1234             !btrfs_test_opt(fs_info, FORCE_COMPRESS)) {
1235                 num_chunks = 1;
1236                 should_compress = false;
1237         } else {
1238                 should_compress = true;
1239         }
1240
1241         nofs_flag = memalloc_nofs_save();
1242         ctx = kvmalloc(struct_size(ctx, chunks, num_chunks), GFP_KERNEL);
1243         memalloc_nofs_restore(nofs_flag);
1244
1245         if (!ctx) {
1246                 unsigned clear_bits = EXTENT_LOCKED | EXTENT_DELALLOC |
1247                         EXTENT_DELALLOC_NEW | EXTENT_DEFRAG |
1248                         EXTENT_DO_ACCOUNTING;
1249                 unsigned long page_ops = PAGE_UNLOCK | PAGE_CLEAR_DIRTY |
1250                         PAGE_SET_WRITEBACK | PAGE_END_WRITEBACK |
1251                         PAGE_SET_ERROR;
1252
1253                 extent_clear_unlock_delalloc(inode, start, end, locked_page,
1254                                              clear_bits, page_ops);
1255                 return -ENOMEM;
1256         }
1257
1258         async_chunk = ctx->chunks;
1259         atomic_set(&ctx->num_chunks, num_chunks);
1260
1261         for (i = 0; i < num_chunks; i++) {
1262                 if (should_compress)
1263                         cur_end = min(end, start + SZ_512K - 1);
1264                 else
1265                         cur_end = end;
1266
1267                 /*
1268                  * igrab is called higher up in the call chain, take only the
1269                  * lightweight reference for the callback lifetime
1270                  */
1271                 ihold(inode);
1272                 async_chunk[i].pending = &ctx->num_chunks;
1273                 async_chunk[i].inode = inode;
1274                 async_chunk[i].start = start;
1275                 async_chunk[i].end = cur_end;
1276                 async_chunk[i].write_flags = write_flags;
1277                 INIT_LIST_HEAD(&async_chunk[i].extents);
1278
1279                 /*
1280                  * The locked_page comes all the way from writepage and its
1281                  * the original page we were actually given.  As we spread
1282                  * this large delalloc region across multiple async_chunk
1283                  * structs, only the first struct needs a pointer to locked_page
1284                  *
1285                  * This way we don't need racey decisions about who is supposed
1286                  * to unlock it.
1287                  */
1288                 if (locked_page) {
1289                         /*
1290                          * Depending on the compressibility, the pages might or
1291                          * might not go through async.  We want all of them to
1292                          * be accounted against wbc once.  Let's do it here
1293                          * before the paths diverge.  wbc accounting is used
1294                          * only for foreign writeback detection and doesn't
1295                          * need full accuracy.  Just account the whole thing
1296                          * against the first page.
1297                          */
1298                         wbc_account_cgroup_owner(wbc, locked_page,
1299                                                  cur_end - start);
1300                         async_chunk[i].locked_page = locked_page;
1301                         locked_page = NULL;
1302                 } else {
1303                         async_chunk[i].locked_page = NULL;
1304                 }
1305
1306                 if (blkcg_css != blkcg_root_css) {
1307                         css_get(blkcg_css);
1308                         async_chunk[i].blkcg_css = blkcg_css;
1309                 } else {
1310                         async_chunk[i].blkcg_css = NULL;
1311                 }
1312
1313                 btrfs_init_work(&async_chunk[i].work, async_cow_start,
1314                                 async_cow_submit, async_cow_free);
1315
1316                 nr_pages = DIV_ROUND_UP(cur_end - start, PAGE_SIZE);
1317                 atomic_add(nr_pages, &fs_info->async_delalloc_pages);
1318
1319                 btrfs_queue_work(fs_info->delalloc_workers, &async_chunk[i].work);
1320
1321                 *nr_written += nr_pages;
1322                 start = cur_end + 1;
1323         }
1324         *page_started = 1;
1325         return 0;
1326 }
1327
1328 static noinline int csum_exist_in_range(struct btrfs_fs_info *fs_info,
1329                                         u64 bytenr, u64 num_bytes)
1330 {
1331         int ret;
1332         struct btrfs_ordered_sum *sums;
1333         LIST_HEAD(list);
1334
1335         ret = btrfs_lookup_csums_range(fs_info->csum_root, bytenr,
1336                                        bytenr + num_bytes - 1, &list, 0);
1337         if (ret == 0 && list_empty(&list))
1338                 return 0;
1339
1340         while (!list_empty(&list)) {
1341                 sums = list_entry(list.next, struct btrfs_ordered_sum, list);
1342                 list_del(&sums->list);
1343                 kfree(sums);
1344         }
1345         if (ret < 0)
1346                 return ret;
1347         return 1;
1348 }
1349
1350 /*
1351  * when nowcow writeback call back.  This checks for snapshots or COW copies
1352  * of the extents that exist in the file, and COWs the file as required.
1353  *
1354  * If no cow copies or snapshots exist, we write directly to the existing
1355  * blocks on disk
1356  */
1357 static noinline int run_delalloc_nocow(struct inode *inode,
1358                                        struct page *locked_page,
1359                                        const u64 start, const u64 end,
1360                                        int *page_started, int force,
1361                                        unsigned long *nr_written)
1362 {
1363         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1364         struct btrfs_root *root = BTRFS_I(inode)->root;
1365         struct btrfs_path *path;
1366         u64 cow_start = (u64)-1;
1367         u64 cur_offset = start;
1368         int ret;
1369         bool check_prev = true;
1370         const bool freespace_inode = btrfs_is_free_space_inode(BTRFS_I(inode));
1371         u64 ino = btrfs_ino(BTRFS_I(inode));
1372         bool nocow = false;
1373         u64 disk_bytenr = 0;
1374
1375         path = btrfs_alloc_path();
1376         if (!path) {
1377                 extent_clear_unlock_delalloc(inode, start, end, locked_page,
1378                                              EXTENT_LOCKED | EXTENT_DELALLOC |
1379                                              EXTENT_DO_ACCOUNTING |
1380                                              EXTENT_DEFRAG, PAGE_UNLOCK |
1381                                              PAGE_CLEAR_DIRTY |
1382                                              PAGE_SET_WRITEBACK |
1383                                              PAGE_END_WRITEBACK);
1384                 return -ENOMEM;
1385         }
1386
1387         while (1) {
1388                 struct btrfs_key found_key;
1389                 struct btrfs_file_extent_item *fi;
1390                 struct extent_buffer *leaf;
1391                 u64 extent_end;
1392                 u64 extent_offset;
1393                 u64 num_bytes = 0;
1394                 u64 disk_num_bytes;
1395                 u64 ram_bytes;
1396                 int extent_type;
1397
1398                 nocow = false;
1399
1400                 ret = btrfs_lookup_file_extent(NULL, root, path, ino,
1401                                                cur_offset, 0);
1402                 if (ret < 0)
1403                         goto error;
1404
1405                 /*
1406                  * If there is no extent for our range when doing the initial
1407                  * search, then go back to the previous slot as it will be the
1408                  * one containing the search offset
1409                  */
1410                 if (ret > 0 && path->slots[0] > 0 && check_prev) {
1411                         leaf = path->nodes[0];
1412                         btrfs_item_key_to_cpu(leaf, &found_key,
1413                                               path->slots[0] - 1);
1414                         if (found_key.objectid == ino &&
1415                             found_key.type == BTRFS_EXTENT_DATA_KEY)
1416                                 path->slots[0]--;
1417                 }
1418                 check_prev = false;
1419 next_slot:
1420                 /* Go to next leaf if we have exhausted the current one */
1421                 leaf = path->nodes[0];
1422                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
1423                         ret = btrfs_next_leaf(root, path);
1424                         if (ret < 0) {
1425                                 if (cow_start != (u64)-1)
1426                                         cur_offset = cow_start;
1427                                 goto error;
1428                         }
1429                         if (ret > 0)
1430                                 break;
1431                         leaf = path->nodes[0];
1432                 }
1433
1434                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
1435
1436                 /* Didn't find anything for our INO */
1437                 if (found_key.objectid > ino)
1438                         break;
1439                 /*
1440                  * Keep searching until we find an EXTENT_ITEM or there are no
1441                  * more extents for this inode
1442                  */
1443                 if (WARN_ON_ONCE(found_key.objectid < ino) ||
1444                     found_key.type < BTRFS_EXTENT_DATA_KEY) {
1445                         path->slots[0]++;
1446                         goto next_slot;
1447                 }
1448
1449                 /* Found key is not EXTENT_DATA_KEY or starts after req range */
1450                 if (found_key.type > BTRFS_EXTENT_DATA_KEY ||
1451                     found_key.offset > end)
1452                         break;
1453
1454                 /*
1455                  * If the found extent starts after requested offset, then
1456                  * adjust extent_end to be right before this extent begins
1457                  */
1458                 if (found_key.offset > cur_offset) {
1459                         extent_end = found_key.offset;
1460                         extent_type = 0;
1461                         goto out_check;
1462                 }
1463
1464                 /*
1465                  * Found extent which begins before our range and potentially
1466                  * intersect it
1467                  */
1468                 fi = btrfs_item_ptr(leaf, path->slots[0],
1469                                     struct btrfs_file_extent_item);
1470                 extent_type = btrfs_file_extent_type(leaf, fi);
1471
1472                 ram_bytes = btrfs_file_extent_ram_bytes(leaf, fi);
1473                 if (extent_type == BTRFS_FILE_EXTENT_REG ||
1474                     extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
1475                         disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1476                         extent_offset = btrfs_file_extent_offset(leaf, fi);
1477                         extent_end = found_key.offset +
1478                                 btrfs_file_extent_num_bytes(leaf, fi);
1479                         disk_num_bytes =
1480                                 btrfs_file_extent_disk_num_bytes(leaf, fi);
1481                         /*
1482                          * If the extent we got ends before our current offset,
1483                          * skip to the next extent.
1484                          */
1485                         if (extent_end <= cur_offset) {
1486                                 path->slots[0]++;
1487                                 goto next_slot;
1488                         }
1489                         /* Skip holes */
1490                         if (disk_bytenr == 0)
1491                                 goto out_check;
1492                         /* Skip compressed/encrypted/encoded extents */
1493                         if (btrfs_file_extent_compression(leaf, fi) ||
1494                             btrfs_file_extent_encryption(leaf, fi) ||
1495                             btrfs_file_extent_other_encoding(leaf, fi))
1496                                 goto out_check;
1497                         /*
1498                          * If extent is created before the last volume's snapshot
1499                          * this implies the extent is shared, hence we can't do
1500                          * nocow. This is the same check as in
1501                          * btrfs_cross_ref_exist but without calling
1502                          * btrfs_search_slot.
1503                          */
1504                         if (!freespace_inode &&
1505                             btrfs_file_extent_generation(leaf, fi) <=
1506                             btrfs_root_last_snapshot(&root->root_item))
1507                                 goto out_check;
1508                         if (extent_type == BTRFS_FILE_EXTENT_REG && !force)
1509                                 goto out_check;
1510                         /* If extent is RO, we must COW it */
1511                         if (btrfs_extent_readonly(fs_info, disk_bytenr))
1512                                 goto out_check;
1513                         ret = btrfs_cross_ref_exist(root, ino,
1514                                                     found_key.offset -
1515                                                     extent_offset, disk_bytenr);
1516                         if (ret) {
1517                                 /*
1518                                  * ret could be -EIO if the above fails to read
1519                                  * metadata.
1520                                  */
1521                                 if (ret < 0) {
1522                                         if (cow_start != (u64)-1)
1523                                                 cur_offset = cow_start;
1524                                         goto error;
1525                                 }
1526
1527                                 WARN_ON_ONCE(freespace_inode);
1528                                 goto out_check;
1529                         }
1530                         disk_bytenr += extent_offset;
1531                         disk_bytenr += cur_offset - found_key.offset;
1532                         num_bytes = min(end + 1, extent_end) - cur_offset;
1533                         /*
1534                          * If there are pending snapshots for this root, we
1535                          * fall into common COW way
1536                          */
1537                         if (!freespace_inode && atomic_read(&root->snapshot_force_cow))
1538                                 goto out_check;
1539                         /*
1540                          * force cow if csum exists in the range.
1541                          * this ensure that csum for a given extent are
1542                          * either valid or do not exist.
1543                          */
1544                         ret = csum_exist_in_range(fs_info, disk_bytenr,
1545                                                   num_bytes);
1546                         if (ret) {
1547                                 /*
1548                                  * ret could be -EIO if the above fails to read
1549                                  * metadata.
1550                                  */
1551                                 if (ret < 0) {
1552                                         if (cow_start != (u64)-1)
1553                                                 cur_offset = cow_start;
1554                                         goto error;
1555                                 }
1556                                 WARN_ON_ONCE(freespace_inode);
1557                                 goto out_check;
1558                         }
1559                         if (!btrfs_inc_nocow_writers(fs_info, disk_bytenr))
1560                                 goto out_check;
1561                         nocow = true;
1562                 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
1563                         extent_end = found_key.offset + ram_bytes;
1564                         extent_end = ALIGN(extent_end, fs_info->sectorsize);
1565                         /* Skip extents outside of our requested range */
1566                         if (extent_end <= start) {
1567                                 path->slots[0]++;
1568                                 goto next_slot;
1569                         }
1570                 } else {
1571                         /* If this triggers then we have a memory corruption */
1572                         BUG();
1573                 }
1574 out_check:
1575                 /*
1576                  * If nocow is false then record the beginning of the range
1577                  * that needs to be COWed
1578                  */
1579                 if (!nocow) {
1580                         if (cow_start == (u64)-1)
1581                                 cow_start = cur_offset;
1582                         cur_offset = extent_end;
1583                         if (cur_offset > end)
1584                                 break;
1585                         path->slots[0]++;
1586                         goto next_slot;
1587                 }
1588
1589                 btrfs_release_path(path);
1590
1591                 /*
1592                  * COW range from cow_start to found_key.offset - 1. As the key
1593                  * will contain the beginning of the first extent that can be
1594                  * NOCOW, following one which needs to be COW'ed
1595                  */
1596                 if (cow_start != (u64)-1) {
1597                         ret = cow_file_range(inode, locked_page,
1598                                              cow_start, found_key.offset - 1,
1599                                              page_started, nr_written, 1);
1600                         if (ret) {
1601                                 if (nocow)
1602                                         btrfs_dec_nocow_writers(fs_info,
1603                                                                 disk_bytenr);
1604                                 goto error;
1605                         }
1606                         cow_start = (u64)-1;
1607                 }
1608
1609                 if (extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
1610                         u64 orig_start = found_key.offset - extent_offset;
1611                         struct extent_map *em;
1612
1613                         em = create_io_em(inode, cur_offset, num_bytes,
1614                                           orig_start,
1615                                           disk_bytenr, /* block_start */
1616                                           num_bytes, /* block_len */
1617                                           disk_num_bytes, /* orig_block_len */
1618                                           ram_bytes, BTRFS_COMPRESS_NONE,
1619                                           BTRFS_ORDERED_PREALLOC);
1620                         if (IS_ERR(em)) {
1621                                 if (nocow)
1622                                         btrfs_dec_nocow_writers(fs_info,
1623                                                                 disk_bytenr);
1624                                 ret = PTR_ERR(em);
1625                                 goto error;
1626                         }
1627                         free_extent_map(em);
1628                         ret = btrfs_add_ordered_extent(inode, cur_offset,
1629                                                        disk_bytenr, num_bytes,
1630                                                        num_bytes,
1631                                                        BTRFS_ORDERED_PREALLOC);
1632                         if (ret) {
1633                                 btrfs_drop_extent_cache(BTRFS_I(inode),
1634                                                         cur_offset,
1635                                                         cur_offset + num_bytes - 1,
1636                                                         0);
1637                                 goto error;
1638                         }
1639                 } else {
1640                         ret = btrfs_add_ordered_extent(inode, cur_offset,
1641                                                        disk_bytenr, num_bytes,
1642                                                        num_bytes,
1643                                                        BTRFS_ORDERED_NOCOW);
1644                         if (ret)
1645                                 goto error;
1646                 }
1647
1648                 if (nocow)
1649                         btrfs_dec_nocow_writers(fs_info, disk_bytenr);
1650                 nocow = false;
1651
1652                 if (root->root_key.objectid ==
1653                     BTRFS_DATA_RELOC_TREE_OBJECTID)
1654                         /*
1655                          * Error handled later, as we must prevent
1656                          * extent_clear_unlock_delalloc() in error handler
1657                          * from freeing metadata of created ordered extent.
1658                          */
1659                         ret = btrfs_reloc_clone_csums(inode, cur_offset,
1660                                                       num_bytes);
1661
1662                 extent_clear_unlock_delalloc(inode, cur_offset,
1663                                              cur_offset + num_bytes - 1,
1664                                              locked_page, EXTENT_LOCKED |
1665                                              EXTENT_DELALLOC |
1666                                              EXTENT_CLEAR_DATA_RESV,
1667                                              PAGE_UNLOCK | PAGE_SET_PRIVATE2);
1668
1669                 cur_offset = extent_end;
1670
1671                 /*
1672                  * btrfs_reloc_clone_csums() error, now we're OK to call error
1673                  * handler, as metadata for created ordered extent will only
1674                  * be freed by btrfs_finish_ordered_io().
1675                  */
1676                 if (ret)
1677                         goto error;
1678                 if (cur_offset > end)
1679                         break;
1680         }
1681         btrfs_release_path(path);
1682
1683         if (cur_offset <= end && cow_start == (u64)-1)
1684                 cow_start = cur_offset;
1685
1686         if (cow_start != (u64)-1) {
1687                 cur_offset = end;
1688                 ret = cow_file_range(inode, locked_page, cow_start, end,
1689                                      page_started, nr_written, 1);
1690                 if (ret)
1691                         goto error;
1692         }
1693
1694 error:
1695         if (nocow)
1696                 btrfs_dec_nocow_writers(fs_info, disk_bytenr);
1697
1698         if (ret && cur_offset < end)
1699                 extent_clear_unlock_delalloc(inode, cur_offset, end,
1700                                              locked_page, EXTENT_LOCKED |
1701                                              EXTENT_DELALLOC | EXTENT_DEFRAG |
1702                                              EXTENT_DO_ACCOUNTING, PAGE_UNLOCK |
1703                                              PAGE_CLEAR_DIRTY |
1704                                              PAGE_SET_WRITEBACK |
1705                                              PAGE_END_WRITEBACK);
1706         btrfs_free_path(path);
1707         return ret;
1708 }
1709
1710 static inline int need_force_cow(struct inode *inode, u64 start, u64 end)
1711 {
1712
1713         if (!(BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW) &&
1714             !(BTRFS_I(inode)->flags & BTRFS_INODE_PREALLOC))
1715                 return 0;
1716
1717         /*
1718          * @defrag_bytes is a hint value, no spinlock held here,
1719          * if is not zero, it means the file is defragging.
1720          * Force cow if given extent needs to be defragged.
1721          */
1722         if (BTRFS_I(inode)->defrag_bytes &&
1723             test_range_bit(&BTRFS_I(inode)->io_tree, start, end,
1724                            EXTENT_DEFRAG, 0, NULL))
1725                 return 1;
1726
1727         return 0;
1728 }
1729
1730 /*
1731  * Function to process delayed allocation (create CoW) for ranges which are
1732  * being touched for the first time.
1733  */
1734 int btrfs_run_delalloc_range(struct inode *inode, struct page *locked_page,
1735                 u64 start, u64 end, int *page_started, unsigned long *nr_written,
1736                 struct writeback_control *wbc)
1737 {
1738         int ret;
1739         int force_cow = need_force_cow(inode, start, end);
1740
1741         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW && !force_cow) {
1742                 ret = run_delalloc_nocow(inode, locked_page, start, end,
1743                                          page_started, 1, nr_written);
1744         } else if (BTRFS_I(inode)->flags & BTRFS_INODE_PREALLOC && !force_cow) {
1745                 ret = run_delalloc_nocow(inode, locked_page, start, end,
1746                                          page_started, 0, nr_written);
1747         } else if (!inode_can_compress(inode) ||
1748                    !inode_need_compress(inode, start, end)) {
1749                 ret = cow_file_range(inode, locked_page, start, end,
1750                                       page_started, nr_written, 1);
1751         } else {
1752                 set_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
1753                         &BTRFS_I(inode)->runtime_flags);
1754                 ret = cow_file_range_async(inode, wbc, locked_page, start, end,
1755                                            page_started, nr_written);
1756         }
1757         if (ret)
1758                 btrfs_cleanup_ordered_extents(inode, locked_page, start,
1759                                               end - start + 1);
1760         return ret;
1761 }
1762
1763 void btrfs_split_delalloc_extent(struct inode *inode,
1764                                  struct extent_state *orig, u64 split)
1765 {
1766         u64 size;
1767
1768         /* not delalloc, ignore it */
1769         if (!(orig->state & EXTENT_DELALLOC))
1770                 return;
1771
1772         size = orig->end - orig->start + 1;
1773         if (size > BTRFS_MAX_EXTENT_SIZE) {
1774                 u32 num_extents;
1775                 u64 new_size;
1776
1777                 /*
1778                  * See the explanation in btrfs_merge_delalloc_extent, the same
1779                  * applies here, just in reverse.
1780                  */
1781                 new_size = orig->end - split + 1;
1782                 num_extents = count_max_extents(new_size);
1783                 new_size = split - orig->start;
1784                 num_extents += count_max_extents(new_size);
1785                 if (count_max_extents(size) >= num_extents)
1786                         return;
1787         }
1788
1789         spin_lock(&BTRFS_I(inode)->lock);
1790         btrfs_mod_outstanding_extents(BTRFS_I(inode), 1);
1791         spin_unlock(&BTRFS_I(inode)->lock);
1792 }
1793
1794 /*
1795  * Handle merged delayed allocation extents so we can keep track of new extents
1796  * that are just merged onto old extents, such as when we are doing sequential
1797  * writes, so we can properly account for the metadata space we'll need.
1798  */
1799 void btrfs_merge_delalloc_extent(struct inode *inode, struct extent_state *new,
1800                                  struct extent_state *other)
1801 {
1802         u64 new_size, old_size;
1803         u32 num_extents;
1804
1805         /* not delalloc, ignore it */
1806         if (!(other->state & EXTENT_DELALLOC))
1807                 return;
1808
1809         if (new->start > other->start)
1810                 new_size = new->end - other->start + 1;
1811         else
1812                 new_size = other->end - new->start + 1;
1813
1814         /* we're not bigger than the max, unreserve the space and go */
1815         if (new_size <= BTRFS_MAX_EXTENT_SIZE) {
1816                 spin_lock(&BTRFS_I(inode)->lock);
1817                 btrfs_mod_outstanding_extents(BTRFS_I(inode), -1);
1818                 spin_unlock(&BTRFS_I(inode)->lock);
1819                 return;
1820         }
1821
1822         /*
1823          * We have to add up either side to figure out how many extents were
1824          * accounted for before we merged into one big extent.  If the number of
1825          * extents we accounted for is <= the amount we need for the new range
1826          * then we can return, otherwise drop.  Think of it like this
1827          *
1828          * [ 4k][MAX_SIZE]
1829          *
1830          * So we've grown the extent by a MAX_SIZE extent, this would mean we
1831          * need 2 outstanding extents, on one side we have 1 and the other side
1832          * we have 1 so they are == and we can return.  But in this case
1833          *
1834          * [MAX_SIZE+4k][MAX_SIZE+4k]
1835          *
1836          * Each range on their own accounts for 2 extents, but merged together
1837          * they are only 3 extents worth of accounting, so we need to drop in
1838          * this case.
1839          */
1840         old_size = other->end - other->start + 1;
1841         num_extents = count_max_extents(old_size);
1842         old_size = new->end - new->start + 1;
1843         num_extents += count_max_extents(old_size);
1844         if (count_max_extents(new_size) >= num_extents)
1845                 return;
1846
1847         spin_lock(&BTRFS_I(inode)->lock);
1848         btrfs_mod_outstanding_extents(BTRFS_I(inode), -1);
1849         spin_unlock(&BTRFS_I(inode)->lock);
1850 }
1851
1852 static void btrfs_add_delalloc_inodes(struct btrfs_root *root,
1853                                       struct inode *inode)
1854 {
1855         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1856
1857         spin_lock(&root->delalloc_lock);
1858         if (list_empty(&BTRFS_I(inode)->delalloc_inodes)) {
1859                 list_add_tail(&BTRFS_I(inode)->delalloc_inodes,
1860                               &root->delalloc_inodes);
1861                 set_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1862                         &BTRFS_I(inode)->runtime_flags);
1863                 root->nr_delalloc_inodes++;
1864                 if (root->nr_delalloc_inodes == 1) {
1865                         spin_lock(&fs_info->delalloc_root_lock);
1866                         BUG_ON(!list_empty(&root->delalloc_root));
1867                         list_add_tail(&root->delalloc_root,
1868                                       &fs_info->delalloc_roots);
1869                         spin_unlock(&fs_info->delalloc_root_lock);
1870                 }
1871         }
1872         spin_unlock(&root->delalloc_lock);
1873 }
1874
1875
1876 void __btrfs_del_delalloc_inode(struct btrfs_root *root,
1877                                 struct btrfs_inode *inode)
1878 {
1879         struct btrfs_fs_info *fs_info = root->fs_info;
1880
1881         if (!list_empty(&inode->delalloc_inodes)) {
1882                 list_del_init(&inode->delalloc_inodes);
1883                 clear_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1884                           &inode->runtime_flags);
1885                 root->nr_delalloc_inodes--;
1886                 if (!root->nr_delalloc_inodes) {
1887                         ASSERT(list_empty(&root->delalloc_inodes));
1888                         spin_lock(&fs_info->delalloc_root_lock);
1889                         BUG_ON(list_empty(&root->delalloc_root));
1890                         list_del_init(&root->delalloc_root);
1891                         spin_unlock(&fs_info->delalloc_root_lock);
1892                 }
1893         }
1894 }
1895
1896 static void btrfs_del_delalloc_inode(struct btrfs_root *root,
1897                                      struct btrfs_inode *inode)
1898 {
1899         spin_lock(&root->delalloc_lock);
1900         __btrfs_del_delalloc_inode(root, inode);
1901         spin_unlock(&root->delalloc_lock);
1902 }
1903
1904 /*
1905  * Properly track delayed allocation bytes in the inode and to maintain the
1906  * list of inodes that have pending delalloc work to be done.
1907  */
1908 void btrfs_set_delalloc_extent(struct inode *inode, struct extent_state *state,
1909                                unsigned *bits)
1910 {
1911         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1912
1913         if ((*bits & EXTENT_DEFRAG) && !(*bits & EXTENT_DELALLOC))
1914                 WARN_ON(1);
1915         /*
1916          * set_bit and clear bit hooks normally require _irqsave/restore
1917          * but in this case, we are only testing for the DELALLOC
1918          * bit, which is only set or cleared with irqs on
1919          */
1920         if (!(state->state & EXTENT_DELALLOC) && (*bits & EXTENT_DELALLOC)) {
1921                 struct btrfs_root *root = BTRFS_I(inode)->root;
1922                 u64 len = state->end + 1 - state->start;
1923                 u32 num_extents = count_max_extents(len);
1924                 bool do_list = !btrfs_is_free_space_inode(BTRFS_I(inode));
1925
1926                 spin_lock(&BTRFS_I(inode)->lock);
1927                 btrfs_mod_outstanding_extents(BTRFS_I(inode), num_extents);
1928                 spin_unlock(&BTRFS_I(inode)->lock);
1929
1930                 /* For sanity tests */
1931                 if (btrfs_is_testing(fs_info))
1932                         return;
1933
1934                 percpu_counter_add_batch(&fs_info->delalloc_bytes, len,
1935                                          fs_info->delalloc_batch);
1936                 spin_lock(&BTRFS_I(inode)->lock);
1937                 BTRFS_I(inode)->delalloc_bytes += len;
1938                 if (*bits & EXTENT_DEFRAG)
1939                         BTRFS_I(inode)->defrag_bytes += len;
1940                 if (do_list && !test_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1941                                          &BTRFS_I(inode)->runtime_flags))
1942                         btrfs_add_delalloc_inodes(root, inode);
1943                 spin_unlock(&BTRFS_I(inode)->lock);
1944         }
1945
1946         if (!(state->state & EXTENT_DELALLOC_NEW) &&
1947             (*bits & EXTENT_DELALLOC_NEW)) {
1948                 spin_lock(&BTRFS_I(inode)->lock);
1949                 BTRFS_I(inode)->new_delalloc_bytes += state->end + 1 -
1950                         state->start;
1951                 spin_unlock(&BTRFS_I(inode)->lock);
1952         }
1953 }
1954
1955 /*
1956  * Once a range is no longer delalloc this function ensures that proper
1957  * accounting happens.
1958  */
1959 void btrfs_clear_delalloc_extent(struct inode *vfs_inode,
1960                                  struct extent_state *state, unsigned *bits)
1961 {
1962         struct btrfs_inode *inode = BTRFS_I(vfs_inode);
1963         struct btrfs_fs_info *fs_info = btrfs_sb(vfs_inode->i_sb);
1964         u64 len = state->end + 1 - state->start;
1965         u32 num_extents = count_max_extents(len);
1966
1967         if ((state->state & EXTENT_DEFRAG) && (*bits & EXTENT_DEFRAG)) {
1968                 spin_lock(&inode->lock);
1969                 inode->defrag_bytes -= len;
1970                 spin_unlock(&inode->lock);
1971         }
1972
1973         /*
1974          * set_bit and clear bit hooks normally require _irqsave/restore
1975          * but in this case, we are only testing for the DELALLOC
1976          * bit, which is only set or cleared with irqs on
1977          */
1978         if ((state->state & EXTENT_DELALLOC) && (*bits & EXTENT_DELALLOC)) {
1979                 struct btrfs_root *root = inode->root;
1980                 bool do_list = !btrfs_is_free_space_inode(inode);
1981
1982                 spin_lock(&inode->lock);
1983                 btrfs_mod_outstanding_extents(inode, -num_extents);
1984                 spin_unlock(&inode->lock);
1985
1986                 /*
1987                  * We don't reserve metadata space for space cache inodes so we
1988                  * don't need to call delalloc_release_metadata if there is an
1989                  * error.
1990                  */
1991                 if (*bits & EXTENT_CLEAR_META_RESV &&
1992                     root != fs_info->tree_root)
1993                         btrfs_delalloc_release_metadata(inode, len, false);
1994
1995                 /* For sanity tests. */
1996                 if (btrfs_is_testing(fs_info))
1997                         return;
1998
1999                 if (root->root_key.objectid != BTRFS_DATA_RELOC_TREE_OBJECTID &&
2000                     do_list && !(state->state & EXTENT_NORESERVE) &&
2001                     (*bits & EXTENT_CLEAR_DATA_RESV))
2002                         btrfs_free_reserved_data_space_noquota(
2003                                         &inode->vfs_inode,
2004                                         state->start, len);
2005
2006                 percpu_counter_add_batch(&fs_info->delalloc_bytes, -len,
2007                                          fs_info->delalloc_batch);
2008                 spin_lock(&inode->lock);
2009                 inode->delalloc_bytes -= len;
2010                 if (do_list && inode->delalloc_bytes == 0 &&
2011                     test_bit(BTRFS_INODE_IN_DELALLOC_LIST,
2012                                         &inode->runtime_flags))
2013                         btrfs_del_delalloc_inode(root, inode);
2014                 spin_unlock(&inode->lock);
2015         }
2016
2017         if ((state->state & EXTENT_DELALLOC_NEW) &&
2018             (*bits & EXTENT_DELALLOC_NEW)) {
2019                 spin_lock(&inode->lock);
2020                 ASSERT(inode->new_delalloc_bytes >= len);
2021                 inode->new_delalloc_bytes -= len;
2022                 spin_unlock(&inode->lock);
2023         }
2024 }
2025
2026 /*
2027  * btrfs_bio_fits_in_stripe - Checks whether the size of the given bio will fit
2028  * in a chunk's stripe. This function ensures that bios do not span a
2029  * stripe/chunk
2030  *
2031  * @page - The page we are about to add to the bio
2032  * @size - size we want to add to the bio
2033  * @bio - bio we want to ensure is smaller than a stripe
2034  * @bio_flags - flags of the bio
2035  *
2036  * return 1 if page cannot be added to the bio
2037  * return 0 if page can be added to the bio
2038  * return error otherwise
2039  */
2040 int btrfs_bio_fits_in_stripe(struct page *page, size_t size, struct bio *bio,
2041                              unsigned long bio_flags)
2042 {
2043         struct inode *inode = page->mapping->host;
2044         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2045         u64 logical = (u64)bio->bi_iter.bi_sector << 9;
2046         u64 length = 0;
2047         u64 map_length;
2048         int ret;
2049         struct btrfs_io_geometry geom;
2050
2051         if (bio_flags & EXTENT_BIO_COMPRESSED)
2052                 return 0;
2053
2054         length = bio->bi_iter.bi_size;
2055         map_length = length;
2056         ret = btrfs_get_io_geometry(fs_info, btrfs_op(bio), logical, map_length,
2057                                     &geom);
2058         if (ret < 0)
2059                 return ret;
2060
2061         if (geom.len < length + size)
2062                 return 1;
2063         return 0;
2064 }
2065
2066 /*
2067  * in order to insert checksums into the metadata in large chunks,
2068  * we wait until bio submission time.   All the pages in the bio are
2069  * checksummed and sums are attached onto the ordered extent record.
2070  *
2071  * At IO completion time the cums attached on the ordered extent record
2072  * are inserted into the btree
2073  */
2074 static blk_status_t btrfs_submit_bio_start(void *private_data, struct bio *bio,
2075                                     u64 bio_offset)
2076 {
2077         struct inode *inode = private_data;
2078         blk_status_t ret = 0;
2079
2080         ret = btrfs_csum_one_bio(inode, bio, 0, 0);
2081         BUG_ON(ret); /* -ENOMEM */
2082         return 0;
2083 }
2084
2085 /*
2086  * extent_io.c submission hook. This does the right thing for csum calculation
2087  * on write, or reading the csums from the tree before a read.
2088  *
2089  * Rules about async/sync submit,
2090  * a) read:                             sync submit
2091  *
2092  * b) write without checksum:           sync submit
2093  *
2094  * c) write with checksum:
2095  *    c-1) if bio is issued by fsync:   sync submit
2096  *         (sync_writers != 0)
2097  *
2098  *    c-2) if root is reloc root:       sync submit
2099  *         (only in case of buffered IO)
2100  *
2101  *    c-3) otherwise:                   async submit
2102  */
2103 static blk_status_t btrfs_submit_bio_hook(struct inode *inode, struct bio *bio,
2104                                           int mirror_num,
2105                                           unsigned long bio_flags)
2106
2107 {
2108         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2109         struct btrfs_root *root = BTRFS_I(inode)->root;
2110         enum btrfs_wq_endio_type metadata = BTRFS_WQ_ENDIO_DATA;
2111         blk_status_t ret = 0;
2112         int skip_sum;
2113         int async = !atomic_read(&BTRFS_I(inode)->sync_writers);
2114
2115         skip_sum = BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM;
2116
2117         if (btrfs_is_free_space_inode(BTRFS_I(inode)))
2118                 metadata = BTRFS_WQ_ENDIO_FREE_SPACE;
2119
2120         if (bio_op(bio) != REQ_OP_WRITE) {
2121                 ret = btrfs_bio_wq_end_io(fs_info, bio, metadata);
2122                 if (ret)
2123                         goto out;
2124
2125                 if (bio_flags & EXTENT_BIO_COMPRESSED) {
2126                         ret = btrfs_submit_compressed_read(inode, bio,
2127                                                            mirror_num,
2128                                                            bio_flags);
2129                         goto out;
2130                 } else if (!skip_sum) {
2131                         ret = btrfs_lookup_bio_sums(inode, bio, NULL);
2132                         if (ret)
2133                                 goto out;
2134                 }
2135                 goto mapit;
2136         } else if (async && !skip_sum) {
2137                 /* csum items have already been cloned */
2138                 if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
2139                         goto mapit;
2140                 /* we're doing a write, do the async checksumming */
2141                 ret = btrfs_wq_submit_bio(fs_info, bio, mirror_num, bio_flags,
2142                                           0, inode, btrfs_submit_bio_start);
2143                 goto out;
2144         } else if (!skip_sum) {
2145                 ret = btrfs_csum_one_bio(inode, bio, 0, 0);
2146                 if (ret)
2147                         goto out;
2148         }
2149
2150 mapit:
2151         ret = btrfs_map_bio(fs_info, bio, mirror_num);
2152
2153 out:
2154         if (ret) {
2155                 bio->bi_status = ret;
2156                 bio_endio(bio);
2157         }
2158         return ret;
2159 }
2160
2161 /*
2162  * given a list of ordered sums record them in the inode.  This happens
2163  * at IO completion time based on sums calculated at bio submission time.
2164  */
2165 static noinline int add_pending_csums(struct btrfs_trans_handle *trans,
2166                              struct inode *inode, struct list_head *list)
2167 {
2168         struct btrfs_ordered_sum *sum;
2169         int ret;
2170
2171         list_for_each_entry(sum, list, list) {
2172                 trans->adding_csums = true;
2173                 ret = btrfs_csum_file_blocks(trans,
2174                        BTRFS_I(inode)->root->fs_info->csum_root, sum);
2175                 trans->adding_csums = false;
2176                 if (ret)
2177                         return ret;
2178         }
2179         return 0;
2180 }
2181
2182 int btrfs_set_extent_delalloc(struct inode *inode, u64 start, u64 end,
2183                               unsigned int extra_bits,
2184                               struct extent_state **cached_state)
2185 {
2186         WARN_ON(PAGE_ALIGNED(end));
2187         return set_extent_delalloc(&BTRFS_I(inode)->io_tree, start, end,
2188                                    extra_bits, cached_state);
2189 }
2190
2191 /* see btrfs_writepage_start_hook for details on why this is required */
2192 struct btrfs_writepage_fixup {
2193         struct page *page;
2194         struct btrfs_work work;
2195 };
2196
2197 static void btrfs_writepage_fixup_worker(struct btrfs_work *work)
2198 {
2199         struct btrfs_writepage_fixup *fixup;
2200         struct btrfs_ordered_extent *ordered;
2201         struct extent_state *cached_state = NULL;
2202         struct extent_changeset *data_reserved = NULL;
2203         struct page *page;
2204         struct inode *inode;
2205         u64 page_start;
2206         u64 page_end;
2207         int ret;
2208
2209         fixup = container_of(work, struct btrfs_writepage_fixup, work);
2210         page = fixup->page;
2211 again:
2212         lock_page(page);
2213         if (!page->mapping || !PageDirty(page) || !PageChecked(page)) {
2214                 ClearPageChecked(page);
2215                 goto out_page;
2216         }
2217
2218         inode = page->mapping->host;
2219         page_start = page_offset(page);
2220         page_end = page_offset(page) + PAGE_SIZE - 1;
2221
2222         lock_extent_bits(&BTRFS_I(inode)->io_tree, page_start, page_end,
2223                          &cached_state);
2224
2225         /* already ordered? We're done */
2226         if (PagePrivate2(page))
2227                 goto out;
2228
2229         ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), page_start,
2230                                         PAGE_SIZE);
2231         if (ordered) {
2232                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, page_start,
2233                                      page_end, &cached_state);
2234                 unlock_page(page);
2235                 btrfs_start_ordered_extent(inode, ordered, 1);
2236                 btrfs_put_ordered_extent(ordered);
2237                 goto again;
2238         }
2239
2240         ret = btrfs_delalloc_reserve_space(inode, &data_reserved, page_start,
2241                                            PAGE_SIZE);
2242         if (ret) {
2243                 mapping_set_error(page->mapping, ret);
2244                 end_extent_writepage(page, ret, page_start, page_end);
2245                 ClearPageChecked(page);
2246                 goto out;
2247          }
2248
2249         ret = btrfs_set_extent_delalloc(inode, page_start, page_end, 0,
2250                                         &cached_state);
2251         if (ret) {
2252                 mapping_set_error(page->mapping, ret);
2253                 end_extent_writepage(page, ret, page_start, page_end);
2254                 ClearPageChecked(page);
2255                 goto out_reserved;
2256         }
2257
2258         ClearPageChecked(page);
2259         set_page_dirty(page);
2260 out_reserved:
2261         btrfs_delalloc_release_extents(BTRFS_I(inode), PAGE_SIZE);
2262         if (ret)
2263                 btrfs_delalloc_release_space(inode, data_reserved, page_start,
2264                                              PAGE_SIZE, true);
2265 out:
2266         unlock_extent_cached(&BTRFS_I(inode)->io_tree, page_start, page_end,
2267                              &cached_state);
2268 out_page:
2269         unlock_page(page);
2270         put_page(page);
2271         kfree(fixup);
2272         extent_changeset_free(data_reserved);
2273 }
2274
2275 /*
2276  * There are a few paths in the higher layers of the kernel that directly
2277  * set the page dirty bit without asking the filesystem if it is a
2278  * good idea.  This causes problems because we want to make sure COW
2279  * properly happens and the data=ordered rules are followed.
2280  *
2281  * In our case any range that doesn't have the ORDERED bit set
2282  * hasn't been properly setup for IO.  We kick off an async process
2283  * to fix it up.  The async helper will wait for ordered extents, set
2284  * the delalloc bit and make it safe to write the page.
2285  */
2286 int btrfs_writepage_cow_fixup(struct page *page, u64 start, u64 end)
2287 {
2288         struct inode *inode = page->mapping->host;
2289         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2290         struct btrfs_writepage_fixup *fixup;
2291
2292         /* this page is properly in the ordered list */
2293         if (TestClearPagePrivate2(page))
2294                 return 0;
2295
2296         if (PageChecked(page))
2297                 return -EAGAIN;
2298
2299         fixup = kzalloc(sizeof(*fixup), GFP_NOFS);
2300         if (!fixup)
2301                 return -EAGAIN;
2302
2303         SetPageChecked(page);
2304         get_page(page);
2305         btrfs_init_work(&fixup->work, btrfs_writepage_fixup_worker, NULL, NULL);
2306         fixup->page = page;
2307         btrfs_queue_work(fs_info->fixup_workers, &fixup->work);
2308         return -EBUSY;
2309 }
2310
2311 static int insert_reserved_file_extent(struct btrfs_trans_handle *trans,
2312                                        struct inode *inode, u64 file_pos,
2313                                        u64 disk_bytenr, u64 disk_num_bytes,
2314                                        u64 num_bytes, u64 ram_bytes,
2315                                        u8 compression, u8 encryption,
2316                                        u16 other_encoding, int extent_type)
2317 {
2318         struct btrfs_root *root = BTRFS_I(inode)->root;
2319         struct btrfs_file_extent_item *fi;
2320         struct btrfs_path *path;
2321         struct extent_buffer *leaf;
2322         struct btrfs_key ins;
2323         u64 qg_released;
2324         int extent_inserted = 0;
2325         int ret;
2326
2327         path = btrfs_alloc_path();
2328         if (!path)
2329                 return -ENOMEM;
2330
2331         /*
2332          * we may be replacing one extent in the tree with another.
2333          * The new extent is pinned in the extent map, and we don't want
2334          * to drop it from the cache until it is completely in the btree.
2335          *
2336          * So, tell btrfs_drop_extents to leave this extent in the cache.
2337          * the caller is expected to unpin it and allow it to be merged
2338          * with the others.
2339          */
2340         ret = __btrfs_drop_extents(trans, root, inode, path, file_pos,
2341                                    file_pos + num_bytes, NULL, 0,
2342                                    1, sizeof(*fi), &extent_inserted);
2343         if (ret)
2344                 goto out;
2345
2346         if (!extent_inserted) {
2347                 ins.objectid = btrfs_ino(BTRFS_I(inode));
2348                 ins.offset = file_pos;
2349                 ins.type = BTRFS_EXTENT_DATA_KEY;
2350
2351                 path->leave_spinning = 1;
2352                 ret = btrfs_insert_empty_item(trans, root, path, &ins,
2353                                               sizeof(*fi));
2354                 if (ret)
2355                         goto out;
2356         }
2357         leaf = path->nodes[0];
2358         fi = btrfs_item_ptr(leaf, path->slots[0],
2359                             struct btrfs_file_extent_item);
2360         btrfs_set_file_extent_generation(leaf, fi, trans->transid);
2361         btrfs_set_file_extent_type(leaf, fi, extent_type);
2362         btrfs_set_file_extent_disk_bytenr(leaf, fi, disk_bytenr);
2363         btrfs_set_file_extent_disk_num_bytes(leaf, fi, disk_num_bytes);
2364         btrfs_set_file_extent_offset(leaf, fi, 0);
2365         btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2366         btrfs_set_file_extent_ram_bytes(leaf, fi, ram_bytes);
2367         btrfs_set_file_extent_compression(leaf, fi, compression);
2368         btrfs_set_file_extent_encryption(leaf, fi, encryption);
2369         btrfs_set_file_extent_other_encoding(leaf, fi, other_encoding);
2370
2371         btrfs_mark_buffer_dirty(leaf);
2372         btrfs_release_path(path);
2373
2374         inode_add_bytes(inode, num_bytes);
2375
2376         ins.objectid = disk_bytenr;
2377         ins.offset = disk_num_bytes;
2378         ins.type = BTRFS_EXTENT_ITEM_KEY;
2379
2380         /*
2381          * Release the reserved range from inode dirty range map, as it is
2382          * already moved into delayed_ref_head
2383          */
2384         ret = btrfs_qgroup_release_data(inode, file_pos, ram_bytes);
2385         if (ret < 0)
2386                 goto out;
2387         qg_released = ret;
2388         ret = btrfs_alloc_reserved_file_extent(trans, root,
2389                                                btrfs_ino(BTRFS_I(inode)),
2390                                                file_pos, qg_released, &ins);
2391 out:
2392         btrfs_free_path(path);
2393
2394         return ret;
2395 }
2396
2397 /* snapshot-aware defrag */
2398 struct sa_defrag_extent_backref {
2399         struct rb_node node;
2400         struct old_sa_defrag_extent *old;
2401         u64 root_id;
2402         u64 inum;
2403         u64 file_pos;
2404         u64 extent_offset;
2405         u64 num_bytes;
2406         u64 generation;
2407 };
2408
2409 struct old_sa_defrag_extent {
2410         struct list_head list;
2411         struct new_sa_defrag_extent *new;
2412
2413         u64 extent_offset;
2414         u64 bytenr;
2415         u64 offset;
2416         u64 len;
2417         int count;
2418 };
2419
2420 struct new_sa_defrag_extent {
2421         struct rb_root root;
2422         struct list_head head;
2423         struct btrfs_path *path;
2424         struct inode *inode;
2425         u64 file_pos;
2426         u64 len;
2427         u64 bytenr;
2428         u64 disk_len;
2429         u8 compress_type;
2430 };
2431
2432 static int backref_comp(struct sa_defrag_extent_backref *b1,
2433                         struct sa_defrag_extent_backref *b2)
2434 {
2435         if (b1->root_id < b2->root_id)
2436                 return -1;
2437         else if (b1->root_id > b2->root_id)
2438                 return 1;
2439
2440         if (b1->inum < b2->inum)
2441                 return -1;
2442         else if (b1->inum > b2->inum)
2443                 return 1;
2444
2445         if (b1->file_pos < b2->file_pos)
2446                 return -1;
2447         else if (b1->file_pos > b2->file_pos)
2448                 return 1;
2449
2450         /*
2451          * [------------------------------] ===> (a range of space)
2452          *     |<--->|   |<---->| =============> (fs/file tree A)
2453          * |<---------------------------->| ===> (fs/file tree B)
2454          *
2455          * A range of space can refer to two file extents in one tree while
2456          * refer to only one file extent in another tree.
2457          *
2458          * So we may process a disk offset more than one time(two extents in A)
2459          * and locate at the same extent(one extent in B), then insert two same
2460          * backrefs(both refer to the extent in B).
2461          */
2462         return 0;
2463 }
2464
2465 static void backref_insert(struct rb_root *root,
2466                            struct sa_defrag_extent_backref *backref)
2467 {
2468         struct rb_node **p = &root->rb_node;
2469         struct rb_node *parent = NULL;
2470         struct sa_defrag_extent_backref *entry;
2471         int ret;
2472
2473         while (*p) {
2474                 parent = *p;
2475                 entry = rb_entry(parent, struct sa_defrag_extent_backref, node);
2476
2477                 ret = backref_comp(backref, entry);
2478                 if (ret < 0)
2479                         p = &(*p)->rb_left;
2480                 else
2481                         p = &(*p)->rb_right;
2482         }
2483
2484         rb_link_node(&backref->node, parent, p);
2485         rb_insert_color(&backref->node, root);
2486 }
2487
2488 /*
2489  * Note the backref might has changed, and in this case we just return 0.
2490  */
2491 static noinline int record_one_backref(u64 inum, u64 offset, u64 root_id,
2492                                        void *ctx)
2493 {
2494         struct btrfs_file_extent_item *extent;
2495         struct old_sa_defrag_extent *old = ctx;
2496         struct new_sa_defrag_extent *new = old->new;
2497         struct btrfs_path *path = new->path;
2498         struct btrfs_key key;
2499         struct btrfs_root *root;
2500         struct sa_defrag_extent_backref *backref;
2501         struct extent_buffer *leaf;
2502         struct inode *inode = new->inode;
2503         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2504         int slot;
2505         int ret;
2506         u64 extent_offset;
2507         u64 num_bytes;
2508
2509         if (BTRFS_I(inode)->root->root_key.objectid == root_id &&
2510             inum == btrfs_ino(BTRFS_I(inode)))
2511                 return 0;
2512
2513         key.objectid = root_id;
2514         key.type = BTRFS_ROOT_ITEM_KEY;
2515         key.offset = (u64)-1;
2516
2517         root = btrfs_read_fs_root_no_name(fs_info, &key);
2518         if (IS_ERR(root)) {
2519                 if (PTR_ERR(root) == -ENOENT)
2520                         return 0;
2521                 WARN_ON(1);
2522                 btrfs_debug(fs_info, "inum=%llu, offset=%llu, root_id=%llu",
2523                          inum, offset, root_id);
2524                 return PTR_ERR(root);
2525         }
2526
2527         key.objectid = inum;
2528         key.type = BTRFS_EXTENT_DATA_KEY;
2529         if (offset > (u64)-1 << 32)
2530                 key.offset = 0;
2531         else
2532                 key.offset = offset;
2533
2534         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2535         if (WARN_ON(ret < 0))
2536                 return ret;
2537         ret = 0;
2538
2539         while (1) {
2540                 cond_resched();
2541
2542                 leaf = path->nodes[0];
2543                 slot = path->slots[0];
2544
2545                 if (slot >= btrfs_header_nritems(leaf)) {
2546                         ret = btrfs_next_leaf(root, path);
2547                         if (ret < 0) {
2548                                 goto out;
2549                         } else if (ret > 0) {
2550                                 ret = 0;
2551                                 goto out;
2552                         }
2553                         continue;
2554                 }
2555
2556                 path->slots[0]++;
2557
2558                 btrfs_item_key_to_cpu(leaf, &key, slot);
2559
2560                 if (key.objectid > inum)
2561                         goto out;
2562
2563                 if (key.objectid < inum || key.type != BTRFS_EXTENT_DATA_KEY)
2564                         continue;
2565
2566                 extent = btrfs_item_ptr(leaf, slot,
2567                                         struct btrfs_file_extent_item);
2568
2569                 if (btrfs_file_extent_disk_bytenr(leaf, extent) != old->bytenr)
2570                         continue;
2571
2572                 /*
2573                  * 'offset' refers to the exact key.offset,
2574                  * NOT the 'offset' field in btrfs_extent_data_ref, ie.
2575                  * (key.offset - extent_offset).
2576                  */
2577                 if (key.offset != offset)
2578                         continue;
2579
2580                 extent_offset = btrfs_file_extent_offset(leaf, extent);
2581                 num_bytes = btrfs_file_extent_num_bytes(leaf, extent);
2582
2583                 if (extent_offset >= old->extent_offset + old->offset +
2584                     old->len || extent_offset + num_bytes <=
2585                     old->extent_offset + old->offset)
2586                         continue;
2587                 break;
2588         }
2589
2590         backref = kmalloc(sizeof(*backref), GFP_NOFS);
2591         if (!backref) {
2592                 ret = -ENOENT;
2593                 goto out;
2594         }
2595
2596         backref->root_id = root_id;
2597         backref->inum = inum;
2598         backref->file_pos = offset;
2599         backref->num_bytes = num_bytes;
2600         backref->extent_offset = extent_offset;
2601         backref->generation = btrfs_file_extent_generation(leaf, extent);
2602         backref->old = old;
2603         backref_insert(&new->root, backref);
2604         old->count++;
2605 out:
2606         btrfs_release_path(path);
2607         WARN_ON(ret);
2608         return ret;
2609 }
2610
2611 static noinline bool record_extent_backrefs(struct btrfs_path *path,
2612                                    struct new_sa_defrag_extent *new)
2613 {
2614         struct btrfs_fs_info *fs_info = btrfs_sb(new->inode->i_sb);
2615         struct old_sa_defrag_extent *old, *tmp;
2616         int ret;
2617
2618         new->path = path;
2619
2620         list_for_each_entry_safe(old, tmp, &new->head, list) {
2621                 ret = iterate_inodes_from_logical(old->bytenr +
2622                                                   old->extent_offset, fs_info,
2623                                                   path, record_one_backref,
2624                                                   old, false);
2625                 if (ret < 0 && ret != -ENOENT)
2626                         return false;
2627
2628                 /* no backref to be processed for this extent */
2629                 if (!old->count) {
2630                         list_del(&old->list);
2631                         kfree(old);
2632                 }
2633         }
2634
2635         if (list_empty(&new->head))
2636                 return false;
2637
2638         return true;
2639 }
2640
2641 static int relink_is_mergable(struct extent_buffer *leaf,
2642                               struct btrfs_file_extent_item *fi,
2643                               struct new_sa_defrag_extent *new)
2644 {
2645         if (btrfs_file_extent_disk_bytenr(leaf, fi) != new->bytenr)
2646                 return 0;
2647
2648         if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG)
2649                 return 0;
2650
2651         if (btrfs_file_extent_compression(leaf, fi) != new->compress_type)
2652                 return 0;
2653
2654         if (btrfs_file_extent_encryption(leaf, fi) ||
2655             btrfs_file_extent_other_encoding(leaf, fi))
2656                 return 0;
2657
2658         return 1;
2659 }
2660
2661 /*
2662  * Note the backref might has changed, and in this case we just return 0.
2663  */
2664 static noinline int relink_extent_backref(struct btrfs_path *path,
2665                                  struct sa_defrag_extent_backref *prev,
2666                                  struct sa_defrag_extent_backref *backref)
2667 {
2668         struct btrfs_file_extent_item *extent;
2669         struct btrfs_file_extent_item *item;
2670         struct btrfs_ordered_extent *ordered;
2671         struct btrfs_trans_handle *trans;
2672         struct btrfs_ref ref = { 0 };
2673         struct btrfs_root *root;
2674         struct btrfs_key key;
2675         struct extent_buffer *leaf;
2676         struct old_sa_defrag_extent *old = backref->old;
2677         struct new_sa_defrag_extent *new = old->new;
2678         struct btrfs_fs_info *fs_info = btrfs_sb(new->inode->i_sb);
2679         struct inode *inode;
2680         struct extent_state *cached = NULL;
2681         int ret = 0;
2682         u64 start;
2683         u64 len;
2684         u64 lock_start;
2685         u64 lock_end;
2686         bool merge = false;
2687         int index;
2688
2689         if (prev && prev->root_id == backref->root_id &&
2690             prev->inum == backref->inum &&
2691             prev->file_pos + prev->num_bytes == backref->file_pos)
2692                 merge = true;
2693
2694         /* step 1: get root */
2695         key.objectid = backref->root_id;
2696         key.type = BTRFS_ROOT_ITEM_KEY;
2697         key.offset = (u64)-1;
2698
2699         index = srcu_read_lock(&fs_info->subvol_srcu);
2700
2701         root = btrfs_read_fs_root_no_name(fs_info, &key);
2702         if (IS_ERR(root)) {
2703                 srcu_read_unlock(&fs_info->subvol_srcu, index);
2704                 if (PTR_ERR(root) == -ENOENT)
2705                         return 0;
2706                 return PTR_ERR(root);
2707         }
2708
2709         if (btrfs_root_readonly(root)) {
2710                 srcu_read_unlock(&fs_info->subvol_srcu, index);
2711                 return 0;
2712         }
2713
2714         /* step 2: get inode */
2715         key.objectid = backref->inum;
2716         key.type = BTRFS_INODE_ITEM_KEY;
2717         key.offset = 0;
2718
2719         inode = btrfs_iget(fs_info->sb, &key, root);
2720         if (IS_ERR(inode)) {
2721                 srcu_read_unlock(&fs_info->subvol_srcu, index);
2722                 return 0;
2723         }
2724
2725         srcu_read_unlock(&fs_info->subvol_srcu, index);
2726
2727         /* step 3: relink backref */
2728         lock_start = backref->file_pos;
2729         lock_end = backref->file_pos + backref->num_bytes - 1;
2730         lock_extent_bits(&BTRFS_I(inode)->io_tree, lock_start, lock_end,
2731                          &cached);
2732
2733         ordered = btrfs_lookup_first_ordered_extent(inode, lock_end);
2734         if (ordered) {
2735                 btrfs_put_ordered_extent(ordered);
2736                 goto out_unlock;
2737         }
2738
2739         trans = btrfs_join_transaction(root);
2740         if (IS_ERR(trans)) {
2741                 ret = PTR_ERR(trans);
2742                 goto out_unlock;
2743         }
2744
2745         key.objectid = backref->inum;
2746         key.type = BTRFS_EXTENT_DATA_KEY;
2747         key.offset = backref->file_pos;
2748
2749         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2750         if (ret < 0) {
2751                 goto out_free_path;
2752         } else if (ret > 0) {
2753                 ret = 0;
2754                 goto out_free_path;
2755         }
2756
2757         extent = btrfs_item_ptr(path->nodes[0], path->slots[0],
2758                                 struct btrfs_file_extent_item);
2759
2760         if (btrfs_file_extent_generation(path->nodes[0], extent) !=
2761             backref->generation)
2762                 goto out_free_path;
2763
2764         btrfs_release_path(path);
2765
2766         start = backref->file_pos;
2767         if (backref->extent_offset < old->extent_offset + old->offset)
2768                 start += old->extent_offset + old->offset -
2769                          backref->extent_offset;
2770
2771         len = min(backref->extent_offset + backref->num_bytes,
2772                   old->extent_offset + old->offset + old->len);
2773         len -= max(backref->extent_offset, old->extent_offset + old->offset);
2774
2775         ret = btrfs_drop_extents(trans, root, inode, start,
2776                                  start + len, 1);
2777         if (ret)
2778                 goto out_free_path;
2779 again:
2780         key.objectid = btrfs_ino(BTRFS_I(inode));
2781         key.type = BTRFS_EXTENT_DATA_KEY;
2782         key.offset = start;
2783
2784         path->leave_spinning = 1;
2785         if (merge) {
2786                 struct btrfs_file_extent_item *fi;
2787                 u64 extent_len;
2788                 struct btrfs_key found_key;
2789
2790                 ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
2791                 if (ret < 0)
2792                         goto out_free_path;
2793
2794                 path->slots[0]--;
2795                 leaf = path->nodes[0];
2796                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
2797
2798                 fi = btrfs_item_ptr(leaf, path->slots[0],
2799                                     struct btrfs_file_extent_item);
2800                 extent_len = btrfs_file_extent_num_bytes(leaf, fi);
2801
2802                 if (extent_len + found_key.offset == start &&
2803                     relink_is_mergable(leaf, fi, new)) {
2804                         btrfs_set_file_extent_num_bytes(leaf, fi,
2805                                                         extent_len + len);
2806                         btrfs_mark_buffer_dirty(leaf);
2807                         inode_add_bytes(inode, len);
2808
2809                         ret = 1;
2810                         goto out_free_path;
2811                 } else {
2812                         merge = false;
2813                         btrfs_release_path(path);
2814                         goto again;
2815                 }
2816         }
2817
2818         ret = btrfs_insert_empty_item(trans, root, path, &key,
2819                                         sizeof(*extent));
2820         if (ret) {
2821                 btrfs_abort_transaction(trans, ret);
2822                 goto out_free_path;
2823         }
2824
2825         leaf = path->nodes[0];
2826         item = btrfs_item_ptr(leaf, path->slots[0],
2827                                 struct btrfs_file_extent_item);
2828         btrfs_set_file_extent_disk_bytenr(leaf, item, new->bytenr);
2829         btrfs_set_file_extent_disk_num_bytes(leaf, item, new->disk_len);
2830         btrfs_set_file_extent_offset(leaf, item, start - new->file_pos);
2831         btrfs_set_file_extent_num_bytes(leaf, item, len);
2832         btrfs_set_file_extent_ram_bytes(leaf, item, new->len);
2833         btrfs_set_file_extent_generation(leaf, item, trans->transid);
2834         btrfs_set_file_extent_type(leaf, item, BTRFS_FILE_EXTENT_REG);
2835         btrfs_set_file_extent_compression(leaf, item, new->compress_type);
2836         btrfs_set_file_extent_encryption(leaf, item, 0);
2837         btrfs_set_file_extent_other_encoding(leaf, item, 0);
2838
2839         btrfs_mark_buffer_dirty(leaf);
2840         inode_add_bytes(inode, len);
2841         btrfs_release_path(path);
2842
2843         btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF, new->bytenr,
2844                                new->disk_len, 0);
2845         btrfs_init_data_ref(&ref, backref->root_id, backref->inum,
2846                             new->file_pos);  /* start - extent_offset */
2847         ret = btrfs_inc_extent_ref(trans, &ref);
2848         if (ret) {
2849                 btrfs_abort_transaction(trans, ret);
2850                 goto out_free_path;
2851         }
2852
2853         ret = 1;
2854 out_free_path:
2855         btrfs_release_path(path);
2856         path->leave_spinning = 0;
2857         btrfs_end_transaction(trans);
2858 out_unlock:
2859         unlock_extent_cached(&BTRFS_I(inode)->io_tree, lock_start, lock_end,
2860                              &cached);
2861         iput(inode);
2862         return ret;
2863 }
2864
2865 static void free_sa_defrag_extent(struct new_sa_defrag_extent *new)
2866 {
2867         struct old_sa_defrag_extent *old, *tmp;
2868
2869         if (!new)
2870                 return;
2871
2872         list_for_each_entry_safe(old, tmp, &new->head, list) {
2873                 kfree(old);
2874         }
2875         kfree(new);
2876 }
2877
2878 static void relink_file_extents(struct new_sa_defrag_extent *new)
2879 {
2880         struct btrfs_fs_info *fs_info = btrfs_sb(new->inode->i_sb);
2881         struct btrfs_path *path;
2882         struct sa_defrag_extent_backref *backref;
2883         struct sa_defrag_extent_backref *prev = NULL;
2884         struct rb_node *node;
2885         int ret;
2886
2887         path = btrfs_alloc_path();
2888         if (!path)
2889                 return;
2890
2891         if (!record_extent_backrefs(path, new)) {
2892                 btrfs_free_path(path);
2893                 goto out;
2894         }
2895         btrfs_release_path(path);
2896
2897         while (1) {
2898                 node = rb_first(&new->root);
2899                 if (!node)
2900                         break;
2901                 rb_erase(node, &new->root);
2902
2903                 backref = rb_entry(node, struct sa_defrag_extent_backref, node);
2904
2905                 ret = relink_extent_backref(path, prev, backref);
2906                 WARN_ON(ret < 0);
2907
2908                 kfree(prev);
2909
2910                 if (ret == 1)
2911                         prev = backref;
2912                 else
2913                         prev = NULL;
2914                 cond_resched();
2915         }
2916         kfree(prev);
2917
2918         btrfs_free_path(path);
2919 out:
2920         free_sa_defrag_extent(new);
2921
2922         atomic_dec(&fs_info->defrag_running);
2923         wake_up(&fs_info->transaction_wait);
2924 }
2925
2926 static struct new_sa_defrag_extent *
2927 record_old_file_extents(struct inode *inode,
2928                         struct btrfs_ordered_extent *ordered)
2929 {
2930         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2931         struct btrfs_root *root = BTRFS_I(inode)->root;
2932         struct btrfs_path *path;
2933         struct btrfs_key key;
2934         struct old_sa_defrag_extent *old;
2935         struct new_sa_defrag_extent *new;
2936         int ret;
2937
2938         new = kmalloc(sizeof(*new), GFP_NOFS);
2939         if (!new)
2940                 return NULL;
2941
2942         new->inode = inode;
2943         new->file_pos = ordered->file_offset;
2944         new->len = ordered->len;
2945         new->bytenr = ordered->start;
2946         new->disk_len = ordered->disk_len;
2947         new->compress_type = ordered->compress_type;
2948         new->root = RB_ROOT;
2949         INIT_LIST_HEAD(&new->head);
2950
2951         path = btrfs_alloc_path();
2952         if (!path)
2953                 goto out_kfree;
2954
2955         key.objectid = btrfs_ino(BTRFS_I(inode));
2956         key.type = BTRFS_EXTENT_DATA_KEY;
2957         key.offset = new->file_pos;
2958
2959         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2960         if (ret < 0)
2961                 goto out_free_path;
2962         if (ret > 0 && path->slots[0] > 0)
2963                 path->slots[0]--;
2964
2965         /* find out all the old extents for the file range */
2966         while (1) {
2967                 struct btrfs_file_extent_item *extent;
2968                 struct extent_buffer *l;
2969                 int slot;
2970                 u64 num_bytes;
2971                 u64 offset;
2972                 u64 end;
2973                 u64 disk_bytenr;
2974                 u64 extent_offset;
2975
2976                 l = path->nodes[0];
2977                 slot = path->slots[0];
2978
2979                 if (slot >= btrfs_header_nritems(l)) {
2980                         ret = btrfs_next_leaf(root, path);
2981                         if (ret < 0)
2982                                 goto out_free_path;
2983                         else if (ret > 0)
2984                                 break;
2985                         continue;
2986                 }
2987
2988                 btrfs_item_key_to_cpu(l, &key, slot);
2989
2990                 if (key.objectid != btrfs_ino(BTRFS_I(inode)))
2991                         break;
2992                 if (key.type != BTRFS_EXTENT_DATA_KEY)
2993                         break;
2994                 if (key.offset >= new->file_pos + new->len)
2995                         break;
2996
2997                 extent = btrfs_item_ptr(l, slot, struct btrfs_file_extent_item);
2998
2999                 num_bytes = btrfs_file_extent_num_bytes(l, extent);
3000                 if (key.offset + num_bytes < new->file_pos)
3001                         goto next;
3002
3003                 disk_bytenr = btrfs_file_extent_disk_bytenr(l, extent);
3004                 if (!disk_bytenr)
3005                         goto next;
3006
3007                 extent_offset = btrfs_file_extent_offset(l, extent);
3008
3009                 old = kmalloc(sizeof(*old), GFP_NOFS);
3010                 if (!old)
3011                         goto out_free_path;
3012
3013                 offset = max(new->file_pos, key.offset);
3014                 end = min(new->file_pos + new->len, key.offset + num_bytes);
3015
3016                 old->bytenr = disk_bytenr;
3017                 old->extent_offset = extent_offset;
3018                 old->offset = offset - key.offset;
3019                 old->len = end - offset;
3020                 old->new = new;
3021                 old->count = 0;
3022                 list_add_tail(&old->list, &new->head);
3023 next:
3024                 path->slots[0]++;
3025                 cond_resched();
3026         }
3027
3028         btrfs_free_path(path);
3029         atomic_inc(&fs_info->defrag_running);
3030
3031         return new;
3032
3033 out_free_path:
3034         btrfs_free_path(path);
3035 out_kfree:
3036         free_sa_defrag_extent(new);
3037         return NULL;
3038 }
3039
3040 static void btrfs_release_delalloc_bytes(struct btrfs_fs_info *fs_info,
3041                                          u64 start, u64 len)
3042 {
3043         struct btrfs_block_group *cache;
3044
3045         cache = btrfs_lookup_block_group(fs_info, start);
3046         ASSERT(cache);
3047
3048         spin_lock(&cache->lock);
3049         cache->delalloc_bytes -= len;
3050         spin_unlock(&cache->lock);
3051
3052         btrfs_put_block_group(cache);
3053 }
3054
3055 /* as ordered data IO finishes, this gets called so we can finish
3056  * an ordered extent if the range of bytes in the file it covers are
3057  * fully written.
3058  */
3059 static int btrfs_finish_ordered_io(struct btrfs_ordered_extent *ordered_extent)
3060 {
3061         struct inode *inode = ordered_extent->inode;
3062         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
3063         struct btrfs_root *root = BTRFS_I(inode)->root;
3064         struct btrfs_trans_handle *trans = NULL;
3065         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
3066         struct extent_state *cached_state = NULL;
3067         struct new_sa_defrag_extent *new = NULL;
3068         int compress_type = 0;
3069         int ret = 0;
3070         u64 logical_len = ordered_extent->len;
3071         bool freespace_inode;
3072         bool truncated = false;
3073         bool range_locked = false;
3074         bool clear_new_delalloc_bytes = false;
3075         bool clear_reserved_extent = true;
3076
3077         if (!test_bit(BTRFS_ORDERED_NOCOW, &ordered_extent->flags) &&
3078             !test_bit(BTRFS_ORDERED_PREALLOC, &ordered_extent->flags) &&
3079             !test_bit(BTRFS_ORDERED_DIRECT, &ordered_extent->flags))
3080                 clear_new_delalloc_bytes = true;
3081
3082         freespace_inode = btrfs_is_free_space_inode(BTRFS_I(inode));
3083
3084         if (test_bit(BTRFS_ORDERED_IOERR, &ordered_extent->flags)) {
3085                 ret = -EIO;
3086                 goto out;
3087         }
3088
3089         btrfs_free_io_failure_record(BTRFS_I(inode),
3090                         ordered_extent->file_offset,
3091                         ordered_extent->file_offset +
3092                         ordered_extent->len - 1);
3093
3094         if (test_bit(BTRFS_ORDERED_TRUNCATED, &ordered_extent->flags)) {
3095                 truncated = true;
3096                 logical_len = ordered_extent->truncated_len;
3097                 /* Truncated the entire extent, don't bother adding */
3098                 if (!logical_len)
3099                         goto out;
3100         }
3101
3102         if (test_bit(BTRFS_ORDERED_NOCOW, &ordered_extent->flags)) {
3103                 BUG_ON(!list_empty(&ordered_extent->list)); /* Logic error */
3104
3105                 /*
3106                  * For mwrite(mmap + memset to write) case, we still reserve
3107                  * space for NOCOW range.
3108                  * As NOCOW won't cause a new delayed ref, just free the space
3109                  */
3110                 btrfs_qgroup_free_data(inode, NULL, ordered_extent->file_offset,
3111                                        ordered_extent->len);
3112                 btrfs_ordered_update_i_size(inode, 0, ordered_extent);
3113                 if (freespace_inode)
3114                         trans = btrfs_join_transaction_spacecache(root);
3115                 else
3116                         trans = btrfs_join_transaction(root);
3117                 if (IS_ERR(trans)) {
3118                         ret = PTR_ERR(trans);
3119                         trans = NULL;
3120                         goto out;
3121                 }
3122                 trans->block_rsv = &BTRFS_I(inode)->block_rsv;
3123                 ret = btrfs_update_inode_fallback(trans, root, inode);
3124                 if (ret) /* -ENOMEM or corruption */
3125                         btrfs_abort_transaction(trans, ret);
3126                 goto out;
3127         }
3128
3129         range_locked = true;
3130         lock_extent_bits(io_tree, ordered_extent->file_offset,
3131                          ordered_extent->file_offset + ordered_extent->len - 1,
3132                          &cached_state);
3133
3134         ret = test_range_bit(io_tree, ordered_extent->file_offset,
3135                         ordered_extent->file_offset + ordered_extent->len - 1,
3136                         EXTENT_DEFRAG, 0, cached_state);
3137         if (ret) {
3138                 u64 last_snapshot = btrfs_root_last_snapshot(&root->root_item);
3139                 if (0 && last_snapshot >= BTRFS_I(inode)->generation)
3140                         /* the inode is shared */
3141                         new = record_old_file_extents(inode, ordered_extent);
3142
3143                 clear_extent_bit(io_tree, ordered_extent->file_offset,
3144                         ordered_extent->file_offset + ordered_extent->len - 1,
3145                         EXTENT_DEFRAG, 0, 0, &cached_state);
3146         }
3147
3148         if (freespace_inode)
3149                 trans = btrfs_join_transaction_spacecache(root);
3150         else
3151                 trans = btrfs_join_transaction(root);
3152         if (IS_ERR(trans)) {
3153                 ret = PTR_ERR(trans);
3154                 trans = NULL;
3155                 goto out;
3156         }
3157
3158         trans->block_rsv = &BTRFS_I(inode)->block_rsv;
3159
3160         if (test_bit(BTRFS_ORDERED_COMPRESSED, &ordered_extent->flags))
3161                 compress_type = ordered_extent->compress_type;
3162         if (test_bit(BTRFS_ORDERED_PREALLOC, &ordered_extent->flags)) {
3163                 BUG_ON(compress_type);
3164                 btrfs_qgroup_free_data(inode, NULL, ordered_extent->file_offset,
3165                                        ordered_extent->len);
3166                 ret = btrfs_mark_extent_written(trans, BTRFS_I(inode),
3167                                                 ordered_extent->file_offset,
3168                                                 ordered_extent->file_offset +
3169                                                 logical_len);
3170         } else {
3171                 BUG_ON(root == fs_info->tree_root);
3172                 ret = insert_reserved_file_extent(trans, inode,
3173                                                 ordered_extent->file_offset,
3174                                                 ordered_extent->start,
3175                                                 ordered_extent->disk_len,
3176                                                 logical_len, logical_len,
3177                                                 compress_type, 0, 0,
3178                                                 BTRFS_FILE_EXTENT_REG);
3179                 if (!ret) {
3180                         clear_reserved_extent = false;
3181                         btrfs_release_delalloc_bytes(fs_info,
3182                                                      ordered_extent->start,
3183                                                      ordered_extent->disk_len);
3184                 }
3185         }
3186         unpin_extent_cache(&BTRFS_I(inode)->extent_tree,
3187                            ordered_extent->file_offset, ordered_extent->len,
3188                            trans->transid);
3189         if (ret < 0) {
3190                 btrfs_abort_transaction(trans, ret);
3191                 goto out;
3192         }
3193
3194         ret = add_pending_csums(trans, inode, &ordered_extent->list);
3195         if (ret) {
3196                 btrfs_abort_transaction(trans, ret);
3197                 goto out;
3198         }
3199
3200         btrfs_ordered_update_i_size(inode, 0, ordered_extent);
3201         ret = btrfs_update_inode_fallback(trans, root, inode);
3202         if (ret) { /* -ENOMEM or corruption */
3203                 btrfs_abort_transaction(trans, ret);
3204                 goto out;
3205         }
3206         ret = 0;
3207 out:
3208         if (range_locked || clear_new_delalloc_bytes) {
3209                 unsigned int clear_bits = 0;
3210
3211                 if (range_locked)
3212                         clear_bits |= EXTENT_LOCKED;
3213                 if (clear_new_delalloc_bytes)
3214                         clear_bits |= EXTENT_DELALLOC_NEW;
3215                 clear_extent_bit(&BTRFS_I(inode)->io_tree,
3216                                  ordered_extent->file_offset,
3217                                  ordered_extent->file_offset +
3218                                  ordered_extent->len - 1,
3219                                  clear_bits,
3220                                  (clear_bits & EXTENT_LOCKED) ? 1 : 0,
3221                                  0, &cached_state);
3222         }
3223
3224         if (trans)
3225                 btrfs_end_transaction(trans);
3226
3227         if (ret || truncated) {
3228                 u64 start, end;
3229
3230                 if (truncated)
3231                         start = ordered_extent->file_offset + logical_len;
3232                 else
3233                         start = ordered_extent->file_offset;
3234                 end = ordered_extent->file_offset + ordered_extent->len - 1;
3235                 clear_extent_uptodate(io_tree, start, end, NULL);
3236
3237                 /* Drop the cache for the part of the extent we didn't write. */
3238                 btrfs_drop_extent_cache(BTRFS_I(inode), start, end, 0);
3239
3240                 /*
3241                  * If the ordered extent had an IOERR or something else went
3242                  * wrong we need to return the space for this ordered extent
3243                  * back to the allocator.  We only free the extent in the
3244                  * truncated case if we didn't write out the extent at all.
3245                  *
3246                  * If we made it past insert_reserved_file_extent before we
3247                  * errored out then we don't need to do this as the accounting
3248                  * has already been done.
3249                  */
3250                 if ((ret || !logical_len) &&
3251                     clear_reserved_extent &&
3252                     !test_bit(BTRFS_ORDERED_NOCOW, &ordered_extent->flags) &&
3253                     !test_bit(BTRFS_ORDERED_PREALLOC, &ordered_extent->flags))
3254                         btrfs_free_reserved_extent(fs_info,
3255                                                    ordered_extent->start,
3256                                                    ordered_extent->disk_len, 1);
3257         }
3258
3259
3260         /*
3261          * This needs to be done to make sure anybody waiting knows we are done
3262          * updating everything for this ordered extent.
3263          */
3264         btrfs_remove_ordered_extent(inode, ordered_extent);
3265
3266         /* for snapshot-aware defrag */
3267         if (new) {
3268                 if (ret) {
3269                         free_sa_defrag_extent(new);
3270                         atomic_dec(&fs_info->defrag_running);
3271                 } else {
3272                         relink_file_extents(new);
3273                 }
3274         }
3275
3276         /* once for us */
3277         btrfs_put_ordered_extent(ordered_extent);
3278         /* once for the tree */
3279         btrfs_put_ordered_extent(ordered_extent);
3280
3281         return ret;
3282 }
3283
3284 static void finish_ordered_fn(struct btrfs_work *work)
3285 {
3286         struct btrfs_ordered_extent *ordered_extent;
3287         ordered_extent = container_of(work, struct btrfs_ordered_extent, work);
3288         btrfs_finish_ordered_io(ordered_extent);
3289 }
3290
3291 void btrfs_writepage_endio_finish_ordered(struct page *page, u64 start,
3292                                           u64 end, int uptodate)
3293 {
3294         struct inode *inode = page->mapping->host;
3295         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
3296         struct btrfs_ordered_extent *ordered_extent = NULL;
3297         struct btrfs_workqueue *wq;
3298
3299         trace_btrfs_writepage_end_io_hook(page, start, end, uptodate);
3300
3301         ClearPagePrivate2(page);
3302         if (!btrfs_dec_test_ordered_pending(inode, &ordered_extent, start,
3303                                             end - start + 1, uptodate))
3304                 return;
3305
3306         if (btrfs_is_free_space_inode(BTRFS_I(inode)))
3307                 wq = fs_info->endio_freespace_worker;
3308         else
3309                 wq = fs_info->endio_write_workers;
3310
3311         btrfs_init_work(&ordered_extent->work, finish_ordered_fn, NULL, NULL);
3312         btrfs_queue_work(wq, &ordered_extent->work);
3313 }
3314
3315 static int __readpage_endio_check(struct inode *inode,
3316                                   struct btrfs_io_bio *io_bio,
3317                                   int icsum, struct page *page,
3318                                   int pgoff, u64 start, size_t len)
3319 {
3320         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
3321         SHASH_DESC_ON_STACK(shash, fs_info->csum_shash);
3322         char *kaddr;
3323         u16 csum_size = btrfs_super_csum_size(fs_info->super_copy);
3324         u8 *csum_expected;
3325         u8 csum[BTRFS_CSUM_SIZE];
3326
3327         csum_expected = ((u8 *)io_bio->csum) + icsum * csum_size;
3328
3329         kaddr = kmap_atomic(page);
3330         shash->tfm = fs_info->csum_shash;
3331
3332         crypto_shash_init(shash);
3333         crypto_shash_update(shash, kaddr + pgoff, len);
3334         crypto_shash_final(shash, csum);
3335
3336         if (memcmp(csum, csum_expected, csum_size))
3337                 goto zeroit;
3338
3339         kunmap_atomic(kaddr);
3340         return 0;
3341 zeroit:
3342         btrfs_print_data_csum_error(BTRFS_I(inode), start, csum, csum_expected,
3343                                     io_bio->mirror_num);
3344         memset(kaddr + pgoff, 1, len);
3345         flush_dcache_page(page);
3346         kunmap_atomic(kaddr);
3347         return -EIO;
3348 }
3349
3350 /*
3351  * when reads are done, we need to check csums to verify the data is correct
3352  * if there's a match, we allow the bio to finish.  If not, the code in
3353  * extent_io.c will try to find good copies for us.
3354  */
3355 static int btrfs_readpage_end_io_hook(struct btrfs_io_bio *io_bio,
3356                                       u64 phy_offset, struct page *page,
3357                                       u64 start, u64 end, int mirror)
3358 {
3359         size_t offset = start - page_offset(page);
3360         struct inode *inode = page->mapping->host;
3361         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
3362         struct btrfs_root *root = BTRFS_I(inode)->root;
3363
3364         if (PageChecked(page)) {
3365                 ClearPageChecked(page);
3366                 return 0;
3367         }
3368
3369         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)
3370                 return 0;
3371
3372         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID &&
3373             test_range_bit(io_tree, start, end, EXTENT_NODATASUM, 1, NULL)) {
3374                 clear_extent_bits(io_tree, start, end, EXTENT_NODATASUM);
3375                 return 0;
3376         }
3377
3378         phy_offset >>= inode->i_sb->s_blocksize_bits;
3379         return __readpage_endio_check(inode, io_bio, phy_offset, page, offset,
3380                                       start, (size_t)(end - start + 1));
3381 }
3382
3383 /*
3384  * btrfs_add_delayed_iput - perform a delayed iput on @inode
3385  *
3386  * @inode: The inode we want to perform iput on
3387  *
3388  * This function uses the generic vfs_inode::i_count to track whether we should
3389  * just decrement it (in case it's > 1) or if this is the last iput then link
3390  * the inode to the delayed iput machinery. Delayed iputs are processed at
3391  * transaction commit time/superblock commit/cleaner kthread.
3392  */
3393 void btrfs_add_delayed_iput(struct inode *inode)
3394 {
3395         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
3396         struct btrfs_inode *binode = BTRFS_I(inode);
3397
3398         if (atomic_add_unless(&inode->i_count, -1, 1))
3399                 return;
3400
3401         atomic_inc(&fs_info->nr_delayed_iputs);
3402         spin_lock(&fs_info->delayed_iput_lock);
3403         ASSERT(list_empty(&binode->delayed_iput));
3404         list_add_tail(&binode->delayed_iput, &fs_info->delayed_iputs);
3405         spin_unlock(&fs_info->delayed_iput_lock);
3406         if (!test_bit(BTRFS_FS_CLEANER_RUNNING, &fs_info->flags))
3407                 wake_up_process(fs_info->cleaner_kthread);
3408 }
3409
3410 static void run_delayed_iput_locked(struct btrfs_fs_info *fs_info,
3411                                     struct btrfs_inode *inode)
3412 {
3413         list_del_init(&inode->delayed_iput);
3414         spin_unlock(&fs_info->delayed_iput_lock);
3415         iput(&inode->vfs_inode);
3416         if (atomic_dec_and_test(&fs_info->nr_delayed_iputs))
3417                 wake_up(&fs_info->delayed_iputs_wait);
3418         spin_lock(&fs_info->delayed_iput_lock);
3419 }
3420
3421 static void btrfs_run_delayed_iput(struct btrfs_fs_info *fs_info,
3422                                    struct btrfs_inode *inode)
3423 {
3424         if (!list_empty(&inode->delayed_iput)) {
3425                 spin_lock(&fs_info->delayed_iput_lock);
3426                 if (!list_empty(&inode->delayed_iput))
3427                         run_delayed_iput_locked(fs_info, inode);
3428                 spin_unlock(&fs_info->delayed_iput_lock);
3429         }
3430 }
3431
3432 void btrfs_run_delayed_iputs(struct btrfs_fs_info *fs_info)
3433 {
3434
3435         spin_lock(&fs_info->delayed_iput_lock);
3436         while (!list_empty(&fs_info->delayed_iputs)) {
3437                 struct btrfs_inode *inode;
3438
3439                 inode = list_first_entry(&fs_info->delayed_iputs,
3440                                 struct btrfs_inode, delayed_iput);
3441                 run_delayed_iput_locked(fs_info, inode);
3442         }
3443         spin_unlock(&fs_info->delayed_iput_lock);
3444 }
3445
3446 /**
3447  * btrfs_wait_on_delayed_iputs - wait on the delayed iputs to be done running
3448  * @fs_info - the fs_info for this fs
3449  * @return - EINTR if we were killed, 0 if nothing's pending
3450  *
3451  * This will wait on any delayed iputs that are currently running with KILLABLE
3452  * set.  Once they are all done running we will return, unless we are killed in
3453  * which case we return EINTR. This helps in user operations like fallocate etc
3454  * that might get blocked on the iputs.
3455  */
3456 int btrfs_wait_on_delayed_iputs(struct btrfs_fs_info *fs_info)
3457 {
3458         int ret = wait_event_killable(fs_info->delayed_iputs_wait,
3459                         atomic_read(&fs_info->nr_delayed_iputs) == 0);
3460         if (ret)
3461                 return -EINTR;
3462         return 0;
3463 }
3464
3465 /*
3466  * This creates an orphan entry for the given inode in case something goes wrong
3467  * in the middle of an unlink.
3468  */
3469 int btrfs_orphan_add(struct btrfs_trans_handle *trans,
3470                      struct btrfs_inode *inode)
3471 {
3472         int ret;
3473
3474         ret = btrfs_insert_orphan_item(trans, inode->root, btrfs_ino(inode));
3475         if (ret && ret != -EEXIST) {
3476                 btrfs_abort_transaction(trans, ret);
3477                 return ret;
3478         }
3479
3480         return 0;
3481 }
3482
3483 /*
3484  * We have done the delete so we can go ahead and remove the orphan item for
3485  * this particular inode.
3486  */
3487 static int btrfs_orphan_del(struct btrfs_trans_handle *trans,
3488                             struct btrfs_inode *inode)
3489 {
3490         return btrfs_del_orphan_item(trans, inode->root, btrfs_ino(inode));
3491 }
3492
3493 /*
3494  * this cleans up any orphans that may be left on the list from the last use
3495  * of this root.
3496  */
3497 int btrfs_orphan_cleanup(struct btrfs_root *root)
3498 {
3499         struct btrfs_fs_info *fs_info = root->fs_info;
3500         struct btrfs_path *path;
3501         struct extent_buffer *leaf;
3502         struct btrfs_key key, found_key;
3503         struct btrfs_trans_handle *trans;
3504         struct inode *inode;
3505         u64 last_objectid = 0;
3506         int ret = 0, nr_unlink = 0;
3507
3508         if (cmpxchg(&root->orphan_cleanup_state, 0, ORPHAN_CLEANUP_STARTED))
3509                 return 0;
3510
3511         path = btrfs_alloc_path();
3512         if (!path) {
3513                 ret = -ENOMEM;
3514                 goto out;
3515         }
3516         path->reada = READA_BACK;
3517
3518         key.objectid = BTRFS_ORPHAN_OBJECTID;
3519         key.type = BTRFS_ORPHAN_ITEM_KEY;
3520         key.offset = (u64)-1;
3521
3522         while (1) {
3523                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3524                 if (ret < 0)
3525                         goto out;
3526
3527                 /*
3528                  * if ret == 0 means we found what we were searching for, which
3529                  * is weird, but possible, so only screw with path if we didn't
3530                  * find the key and see if we have stuff that matches
3531                  */
3532                 if (ret > 0) {
3533                         ret = 0;
3534                         if (path->slots[0] == 0)
3535                                 break;
3536                         path->slots[0]--;
3537                 }
3538
3539                 /* pull out the item */
3540                 leaf = path->nodes[0];
3541                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
3542
3543                 /* make sure the item matches what we want */
3544                 if (found_key.objectid != BTRFS_ORPHAN_OBJECTID)
3545                         break;
3546                 if (found_key.type != BTRFS_ORPHAN_ITEM_KEY)
3547                         break;
3548
3549                 /* release the path since we're done with it */
3550                 btrfs_release_path(path);
3551
3552                 /*
3553                  * this is where we are basically btrfs_lookup, without the
3554                  * crossing root thing.  we store the inode number in the
3555                  * offset of the orphan item.
3556                  */
3557
3558                 if (found_key.offset == last_objectid) {
3559                         btrfs_err(fs_info,
3560                                   "Error removing orphan entry, stopping orphan cleanup");
3561                         ret = -EINVAL;
3562                         goto out;
3563                 }
3564
3565                 last_objectid = found_key.offset;
3566
3567                 found_key.objectid = found_key.offset;
3568                 found_key.type = BTRFS_INODE_ITEM_KEY;
3569                 found_key.offset = 0;
3570                 inode = btrfs_iget(fs_info->sb, &found_key, root);
3571                 ret = PTR_ERR_OR_ZERO(inode);
3572                 if (ret && ret != -ENOENT)
3573                         goto out;
3574
3575                 if (ret == -ENOENT && root == fs_info->tree_root) {
3576                         struct btrfs_root *dead_root;
3577                         struct btrfs_fs_info *fs_info = root->fs_info;
3578                         int is_dead_root = 0;
3579
3580                         /*
3581                          * this is an orphan in the tree root. Currently these
3582                          * could come from 2 sources:
3583                          *  a) a snapshot deletion in progress
3584                          *  b) a free space cache inode
3585                          * We need to distinguish those two, as the snapshot
3586                          * orphan must not get deleted.
3587                          * find_dead_roots already ran before us, so if this
3588                          * is a snapshot deletion, we should find the root
3589                          * in the dead_roots list
3590                          */
3591                         spin_lock(&fs_info->trans_lock);
3592                         list_for_each_entry(dead_root, &fs_info->dead_roots,
3593                                             root_list) {
3594                                 if (dead_root->root_key.objectid ==
3595                                     found_key.objectid) {
3596                                         is_dead_root = 1;
3597                                         break;
3598                                 }
3599                         }
3600                         spin_unlock(&fs_info->trans_lock);
3601                         if (is_dead_root) {
3602                                 /* prevent this orphan from being found again */
3603                                 key.offset = found_key.objectid - 1;
3604                                 continue;
3605                         }
3606
3607                 }
3608
3609                 /*
3610                  * If we have an inode with links, there are a couple of
3611                  * possibilities. Old kernels (before v3.12) used to create an
3612                  * orphan item for truncate indicating that there were possibly
3613                  * extent items past i_size that needed to be deleted. In v3.12,
3614                  * truncate was changed to update i_size in sync with the extent
3615                  * items, but the (useless) orphan item was still created. Since
3616                  * v4.18, we don't create the orphan item for truncate at all.
3617                  *
3618                  * So, this item could mean that we need to do a truncate, but
3619                  * only if this filesystem was last used on a pre-v3.12 kernel
3620                  * and was not cleanly unmounted. The odds of that are quite
3621                  * slim, and it's a pain to do the truncate now, so just delete
3622                  * the orphan item.
3623                  *
3624                  * It's also possible that this orphan item was supposed to be
3625                  * deleted but wasn't. The inode number may have been reused,
3626                  * but either way, we can delete the orphan item.
3627                  */
3628                 if (ret == -ENOENT || inode->i_nlink) {
3629                         if (!ret)
3630                                 iput(inode);
3631                         trans = btrfs_start_transaction(root, 1);
3632                         if (IS_ERR(trans)) {
3633                                 ret = PTR_ERR(trans);
3634                                 goto out;
3635                         }
3636                         btrfs_debug(fs_info, "auto deleting %Lu",
3637                                     found_key.objectid);
3638                         ret = btrfs_del_orphan_item(trans, root,
3639                                                     found_key.objectid);
3640                         btrfs_end_transaction(trans);
3641                         if (ret)
3642                                 goto out;
3643                         continue;
3644                 }
3645
3646                 nr_unlink++;
3647
3648                 /* this will do delete_inode and everything for us */
3649                 iput(inode);
3650         }
3651         /* release the path since we're done with it */
3652         btrfs_release_path(path);
3653
3654         root->orphan_cleanup_state = ORPHAN_CLEANUP_DONE;
3655
3656         if (test_bit(BTRFS_ROOT_ORPHAN_ITEM_INSERTED, &root->state)) {
3657                 trans = btrfs_join_transaction(root);
3658                 if (!IS_ERR(trans))
3659                         btrfs_end_transaction(trans);
3660         }
3661
3662         if (nr_unlink)
3663                 btrfs_debug(fs_info, "unlinked %d orphans", nr_unlink);
3664
3665 out:
3666         if (ret)
3667                 btrfs_err(fs_info, "could not do orphan cleanup %d", ret);
3668         btrfs_free_path(path);
3669         return ret;
3670 }
3671
3672 /*
3673  * very simple check to peek ahead in the leaf looking for xattrs.  If we
3674  * don't find any xattrs, we know there can't be any acls.
3675  *
3676  * slot is the slot the inode is in, objectid is the objectid of the inode
3677  */
3678 static noinline int acls_after_inode_item(struct extent_buffer *leaf,
3679                                           int slot, u64 objectid,
3680                                           int *first_xattr_slot)
3681 {
3682         u32 nritems = btrfs_header_nritems(leaf);
3683         struct btrfs_key found_key;
3684         static u64 xattr_access = 0;
3685         static u64 xattr_default = 0;
3686         int scanned = 0;
3687
3688         if (!xattr_access) {
3689                 xattr_access = btrfs_name_hash(XATTR_NAME_POSIX_ACL_ACCESS,
3690                                         strlen(XATTR_NAME_POSIX_ACL_ACCESS));
3691                 xattr_default = btrfs_name_hash(XATTR_NAME_POSIX_ACL_DEFAULT,
3692                                         strlen(XATTR_NAME_POSIX_ACL_DEFAULT));
3693         }
3694
3695         slot++;
3696         *first_xattr_slot = -1;
3697         while (slot < nritems) {
3698                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
3699
3700                 /* we found a different objectid, there must not be acls */
3701                 if (found_key.objectid != objectid)
3702                         return 0;
3703
3704                 /* we found an xattr, assume we've got an acl */
3705                 if (found_key.type == BTRFS_XATTR_ITEM_KEY) {
3706                         if (*first_xattr_slot == -1)
3707                                 *first_xattr_slot = slot;
3708                         if (found_key.offset == xattr_access ||
3709                             found_key.offset == xattr_default)
3710                                 return 1;
3711                 }
3712
3713                 /*
3714                  * we found a key greater than an xattr key, there can't
3715                  * be any acls later on
3716                  */
3717                 if (found_key.type > BTRFS_XATTR_ITEM_KEY)
3718                         return 0;
3719
3720                 slot++;
3721                 scanned++;
3722
3723                 /*
3724                  * it goes inode, inode backrefs, xattrs, extents,
3725                  * so if there are a ton of hard links to an inode there can
3726                  * be a lot of backrefs.  Don't waste time searching too hard,
3727                  * this is just an optimization
3728                  */
3729                 if (scanned >= 8)
3730                         break;
3731         }
3732         /* we hit the end of the leaf before we found an xattr or
3733          * something larger than an xattr.  We have to assume the inode
3734          * has acls
3735          */
3736         if (*first_xattr_slot == -1)
3737                 *first_xattr_slot = slot;
3738         return 1;
3739 }
3740
3741 /*
3742  * read an inode from the btree into the in-memory inode
3743  */
3744 static int btrfs_read_locked_inode(struct inode *inode,
3745                                    struct btrfs_path *in_path)
3746 {
3747         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
3748         struct btrfs_path *path = in_path;
3749         struct extent_buffer *leaf;
3750         struct btrfs_inode_item *inode_item;
3751         struct btrfs_root *root = BTRFS_I(inode)->root;
3752         struct btrfs_key location;
3753         unsigned long ptr;
3754         int maybe_acls;
3755         u32 rdev;
3756         int ret;
3757         bool filled = false;
3758         int first_xattr_slot;
3759
3760         ret = btrfs_fill_inode(inode, &rdev);
3761         if (!ret)
3762                 filled = true;
3763
3764         if (!path) {
3765                 path = btrfs_alloc_path();
3766                 if (!path)
3767                         return -ENOMEM;
3768         }
3769
3770         memcpy(&location, &BTRFS_I(inode)->location, sizeof(location));
3771
3772         ret = btrfs_lookup_inode(NULL, root, path, &location, 0);
3773         if (ret) {
3774                 if (path != in_path)
3775                         btrfs_free_path(path);
3776                 return ret;
3777         }
3778
3779         leaf = path->nodes[0];
3780
3781         if (filled)
3782                 goto cache_index;
3783
3784         inode_item = btrfs_item_ptr(leaf, path->slots[0],
3785                                     struct btrfs_inode_item);
3786         inode->i_mode = btrfs_inode_mode(leaf, inode_item);
3787         set_nlink(inode, btrfs_inode_nlink(leaf, inode_item));
3788         i_uid_write(inode, btrfs_inode_uid(leaf, inode_item));
3789         i_gid_write(inode, btrfs_inode_gid(leaf, inode_item));
3790         btrfs_i_size_write(BTRFS_I(inode), btrfs_inode_size(leaf, inode_item));
3791
3792         inode->i_atime.tv_sec = btrfs_timespec_sec(leaf, &inode_item->atime);
3793         inode->i_atime.tv_nsec = btrfs_timespec_nsec(leaf, &inode_item->atime);
3794
3795         inode->i_mtime.tv_sec = btrfs_timespec_sec(leaf, &inode_item->mtime);
3796         inode->i_mtime.tv_nsec = btrfs_timespec_nsec(leaf, &inode_item->mtime);
3797
3798         inode->i_ctime.tv_sec = btrfs_timespec_sec(leaf, &inode_item->ctime);
3799         inode->i_ctime.tv_nsec = btrfs_timespec_nsec(leaf, &inode_item->ctime);
3800
3801         BTRFS_I(inode)->i_otime.tv_sec =
3802                 btrfs_timespec_sec(leaf, &inode_item->otime);
3803         BTRFS_I(inode)->i_otime.tv_nsec =
3804                 btrfs_timespec_nsec(leaf, &inode_item->otime);
3805
3806         inode_set_bytes(inode, btrfs_inode_nbytes(leaf, inode_item));
3807         BTRFS_I(inode)->generation = btrfs_inode_generation(leaf, inode_item);
3808         BTRFS_I(inode)->last_trans = btrfs_inode_transid(leaf, inode_item);
3809
3810         inode_set_iversion_queried(inode,
3811                                    btrfs_inode_sequence(leaf, inode_item));
3812         inode->i_generation = BTRFS_I(inode)->generation;
3813         inode->i_rdev = 0;
3814         rdev = btrfs_inode_rdev(leaf, inode_item);
3815
3816         BTRFS_I(inode)->index_cnt = (u64)-1;
3817         BTRFS_I(inode)->flags = btrfs_inode_flags(leaf, inode_item);
3818
3819 cache_index:
3820         /*
3821          * If we were modified in the current generation and evicted from memory
3822          * and then re-read we need to do a full sync since we don't have any
3823          * idea about which extents were modified before we were evicted from
3824          * cache.
3825          *
3826          * This is required for both inode re-read from disk and delayed inode
3827          * in delayed_nodes_tree.
3828          */
3829         if (BTRFS_I(inode)->last_trans == fs_info->generation)
3830                 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
3831                         &BTRFS_I(inode)->runtime_flags);
3832
3833         /*
3834          * We don't persist the id of the transaction where an unlink operation
3835          * against the inode was last made. So here we assume the inode might
3836          * have been evicted, and therefore the exact value of last_unlink_trans
3837          * lost, and set it to last_trans to avoid metadata inconsistencies
3838          * between the inode and its parent if the inode is fsync'ed and the log
3839          * replayed. For example, in the scenario:
3840          *
3841          * touch mydir/foo
3842          * ln mydir/foo mydir/bar
3843          * sync
3844          * unlink mydir/bar
3845          * echo 2 > /proc/sys/vm/drop_caches   # evicts inode
3846          * xfs_io -c fsync mydir/foo
3847          * <power failure>
3848          * mount fs, triggers fsync log replay
3849          *
3850          * We must make sure that when we fsync our inode foo we also log its
3851          * parent inode, otherwise after log replay the parent still has the
3852          * dentry with the "bar" name but our inode foo has a link count of 1
3853          * and doesn't have an inode ref with the name "bar" anymore.
3854          *
3855          * Setting last_unlink_trans to last_trans is a pessimistic approach,
3856          * but it guarantees correctness at the expense of occasional full
3857          * transaction commits on fsync if our inode is a directory, or if our
3858          * inode is not a directory, logging its parent unnecessarily.
3859          */
3860         BTRFS_I(inode)->last_unlink_trans = BTRFS_I(inode)->last_trans;
3861
3862         path->slots[0]++;
3863         if (inode->i_nlink != 1 ||
3864             path->slots[0] >= btrfs_header_nritems(leaf))
3865                 goto cache_acl;
3866
3867         btrfs_item_key_to_cpu(leaf, &location, path->slots[0]);
3868         if (location.objectid != btrfs_ino(BTRFS_I(inode)))
3869                 goto cache_acl;
3870
3871         ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
3872         if (location.type == BTRFS_INODE_REF_KEY) {
3873                 struct btrfs_inode_ref *ref;
3874
3875                 ref = (struct btrfs_inode_ref *)ptr;
3876                 BTRFS_I(inode)->dir_index = btrfs_inode_ref_index(leaf, ref);
3877         } else if (location.type == BTRFS_INODE_EXTREF_KEY) {
3878                 struct btrfs_inode_extref *extref;
3879
3880                 extref = (struct btrfs_inode_extref *)ptr;
3881                 BTRFS_I(inode)->dir_index = btrfs_inode_extref_index(leaf,
3882                                                                      extref);
3883         }
3884 cache_acl:
3885         /*
3886          * try to precache a NULL acl entry for files that don't have
3887          * any xattrs or acls
3888          */
3889         maybe_acls = acls_after_inode_item(leaf, path->slots[0],
3890                         btrfs_ino(BTRFS_I(inode)), &first_xattr_slot);
3891         if (first_xattr_slot != -1) {
3892                 path->slots[0] = first_xattr_slot;
3893                 ret = btrfs_load_inode_props(inode, path);
3894                 if (ret)
3895                         btrfs_err(fs_info,
3896                                   "error loading props for ino %llu (root %llu): %d",
3897                                   btrfs_ino(BTRFS_I(inode)),
3898                                   root->root_key.objectid, ret);
3899         }
3900         if (path != in_path)
3901                 btrfs_free_path(path);
3902
3903         if (!maybe_acls)
3904                 cache_no_acl(inode);
3905
3906         switch (inode->i_mode & S_IFMT) {
3907         case S_IFREG:
3908                 inode->i_mapping->a_ops = &btrfs_aops;
3909                 BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
3910                 inode->i_fop = &btrfs_file_operations;
3911                 inode->i_op = &btrfs_file_inode_operations;
3912                 break;
3913         case S_IFDIR:
3914                 inode->i_fop = &btrfs_dir_file_operations;
3915                 inode->i_op = &btrfs_dir_inode_operations;
3916                 break;
3917         case S_IFLNK:
3918                 inode->i_op = &btrfs_symlink_inode_operations;
3919                 inode_nohighmem(inode);
3920                 inode->i_mapping->a_ops = &btrfs_aops;
3921                 break;
3922         default:
3923                 inode->i_op = &btrfs_special_inode_operations;
3924                 init_special_inode(inode, inode->i_mode, rdev);
3925                 break;
3926         }
3927
3928         btrfs_sync_inode_flags_to_i_flags(inode);
3929         return 0;
3930 }
3931
3932 /*
3933  * given a leaf and an inode, copy the inode fields into the leaf
3934  */
3935 static void fill_inode_item(struct btrfs_trans_handle *trans,
3936                             struct extent_buffer *leaf,
3937                             struct btrfs_inode_item *item,
3938                             struct inode *inode)
3939 {
3940         struct btrfs_map_token token;
3941
3942         btrfs_init_map_token(&token, leaf);
3943
3944         btrfs_set_token_inode_uid(leaf, item, i_uid_read(inode), &token);
3945         btrfs_set_token_inode_gid(leaf, item, i_gid_read(inode), &token);
3946         btrfs_set_token_inode_size(leaf, item, BTRFS_I(inode)->disk_i_size,
3947                                    &token);
3948         btrfs_set_token_inode_mode(leaf, item, inode->i_mode, &token);
3949         btrfs_set_token_inode_nlink(leaf, item, inode->i_nlink, &token);
3950
3951         btrfs_set_token_timespec_sec(leaf, &item->atime,
3952                                      inode->i_atime.tv_sec, &token);
3953         btrfs_set_token_timespec_nsec(leaf, &item->atime,
3954                                       inode->i_atime.tv_nsec, &token);
3955
3956         btrfs_set_token_timespec_sec(leaf, &item->mtime,
3957                                      inode->i_mtime.tv_sec, &token);
3958         btrfs_set_token_timespec_nsec(leaf, &item->mtime,
3959                                       inode->i_mtime.tv_nsec, &token);
3960
3961         btrfs_set_token_timespec_sec(leaf, &item->ctime,
3962                                      inode->i_ctime.tv_sec, &token);
3963         btrfs_set_token_timespec_nsec(leaf, &item->ctime,
3964                                       inode->i_ctime.tv_nsec, &token);
3965
3966         btrfs_set_token_timespec_sec(leaf, &item->otime,
3967                                      BTRFS_I(inode)->i_otime.tv_sec, &token);
3968         btrfs_set_token_timespec_nsec(leaf, &item->otime,
3969                                       BTRFS_I(inode)->i_otime.tv_nsec, &token);
3970
3971         btrfs_set_token_inode_nbytes(leaf, item, inode_get_bytes(inode),
3972                                      &token);
3973         btrfs_set_token_inode_generation(leaf, item, BTRFS_I(inode)->generation,
3974                                          &token);
3975         btrfs_set_token_inode_sequence(leaf, item, inode_peek_iversion(inode),
3976                                        &token);
3977         btrfs_set_token_inode_transid(leaf, item, trans->transid, &token);
3978         btrfs_set_token_inode_rdev(leaf, item, inode->i_rdev, &token);
3979         btrfs_set_token_inode_flags(leaf, item, BTRFS_I(inode)->flags, &token);
3980         btrfs_set_token_inode_block_group(leaf, item, 0, &token);
3981 }
3982
3983 /*
3984  * copy everything in the in-memory inode into the btree.
3985  */
3986 static noinline int btrfs_update_inode_item(struct btrfs_trans_handle *trans,
3987                                 struct btrfs_root *root, struct inode *inode)
3988 {
3989         struct btrfs_inode_item *inode_item;
3990         struct btrfs_path *path;
3991         struct extent_buffer *leaf;
3992         int ret;
3993
3994         path = btrfs_alloc_path();
3995         if (!path)
3996                 return -ENOMEM;
3997
3998         path->leave_spinning = 1;
3999         ret = btrfs_lookup_inode(trans, root, path, &BTRFS_I(inode)->location,
4000                                  1);
4001         if (ret) {
4002                 if (ret > 0)
4003                         ret = -ENOENT;
4004                 goto failed;
4005         }
4006
4007         leaf = path->nodes[0];
4008         inode_item = btrfs_item_ptr(leaf, path->slots[0],
4009                                     struct btrfs_inode_item);
4010
4011         fill_inode_item(trans, leaf, inode_item, inode);
4012         btrfs_mark_buffer_dirty(leaf);
4013         btrfs_set_inode_last_trans(trans, inode);
4014         ret = 0;
4015 failed:
4016         btrfs_free_path(path);
4017         return ret;
4018 }
4019
4020 /*
4021  * copy everything in the in-memory inode into the btree.
4022  */
4023 noinline int btrfs_update_inode(struct btrfs_trans_handle *trans,
4024                                 struct btrfs_root *root, struct inode *inode)
4025 {
4026         struct btrfs_fs_info *fs_info = root->fs_info;
4027         int ret;
4028
4029         /*
4030          * If the inode is a free space inode, we can deadlock during commit
4031          * if we put it into the delayed code.
4032          *
4033          * The data relocation inode should also be directly updated
4034          * without delay
4035          */
4036         if (!btrfs_is_free_space_inode(BTRFS_I(inode))
4037             && root->root_key.objectid != BTRFS_DATA_RELOC_TREE_OBJECTID
4038             && !test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags)) {
4039                 btrfs_update_root_times(trans, root);
4040
4041                 ret = btrfs_delayed_update_inode(trans, root, inode);
4042                 if (!ret)
4043                         btrfs_set_inode_last_trans(trans, inode);
4044                 return ret;
4045         }
4046
4047         return btrfs_update_inode_item(trans, root, inode);
4048 }
4049
4050 noinline int btrfs_update_inode_fallback(struct btrfs_trans_handle *trans,
4051                                          struct btrfs_root *root,
4052                                          struct inode *inode)
4053 {
4054         int ret;
4055
4056         ret = btrfs_update_inode(trans, root, inode);
4057         if (ret == -ENOSPC)
4058                 return btrfs_update_inode_item(trans, root, inode);
4059         return ret;
4060 }
4061
4062 /*
4063  * unlink helper that gets used here in inode.c and in the tree logging
4064  * recovery code.  It remove a link in a directory with a given name, and
4065  * also drops the back refs in the inode to the directory
4066  */
4067 static int __btrfs_unlink_inode(struct btrfs_trans_handle *trans,
4068                                 struct btrfs_root *root,
4069                                 struct btrfs_inode *dir,
4070                                 struct btrfs_inode *inode,
4071                                 const char *name, int name_len)
4072 {
4073         struct btrfs_fs_info *fs_info = root->fs_info;
4074         struct btrfs_path *path;
4075         int ret = 0;
4076         struct btrfs_dir_item *di;
4077         u64 index;
4078         u64 ino = btrfs_ino(inode);
4079         u64 dir_ino = btrfs_ino(dir);
4080
4081         path = btrfs_alloc_path();
4082         if (!path) {
4083                 ret = -ENOMEM;
4084                 goto out;
4085         }
4086
4087         path->leave_spinning = 1;
4088         di = btrfs_lookup_dir_item(trans, root, path, dir_ino,
4089                                     name, name_len, -1);
4090         if (IS_ERR_OR_NULL(di)) {
4091                 ret = di ? PTR_ERR(di) : -ENOENT;
4092                 goto err;
4093         }
4094         ret = btrfs_delete_one_dir_name(trans, root, path, di);
4095         if (ret)
4096                 goto err;
4097         btrfs_release_path(path);
4098
4099         /*
4100          * If we don't have dir index, we have to get it by looking up
4101          * the inode ref, since we get the inode ref, remove it directly,
4102          * it is unnecessary to do delayed deletion.
4103          *
4104          * But if we have dir index, needn't search inode ref to get it.
4105          * Since the inode ref is close to the inode item, it is better
4106          * that we delay to delete it, and just do this deletion when
4107          * we update the inode item.
4108          */
4109         if (inode->dir_index) {
4110                 ret = btrfs_delayed_delete_inode_ref(inode);
4111                 if (!ret) {
4112                         index = inode->dir_index;
4113                         goto skip_backref;
4114                 }
4115         }
4116
4117         ret = btrfs_del_inode_ref(trans, root, name, name_len, ino,
4118                                   dir_ino, &index);
4119         if (ret) {
4120                 btrfs_info(fs_info,
4121                         "failed to delete reference to %.*s, inode %llu parent %llu",
4122                         name_len, name, ino, dir_ino);
4123                 btrfs_abort_transaction(trans, ret);
4124                 goto err;
4125         }
4126 skip_backref:
4127         ret = btrfs_delete_delayed_dir_index(trans, dir, index);
4128         if (ret) {
4129                 btrfs_abort_transaction(trans, ret);
4130                 goto err;
4131         }
4132
4133         ret = btrfs_del_inode_ref_in_log(trans, root, name, name_len, inode,
4134                         dir_ino);
4135         if (ret != 0 && ret != -ENOENT) {
4136                 btrfs_abort_transaction(trans, ret);
4137                 goto err;
4138         }
4139
4140         ret = btrfs_del_dir_entries_in_log(trans, root, name, name_len, dir,
4141                         index);
4142         if (ret == -ENOENT)
4143                 ret = 0;
4144         else if (ret)
4145                 btrfs_abort_transaction(trans, ret);
4146
4147         /*
4148          * If we have a pending delayed iput we could end up with the final iput
4149          * being run in btrfs-cleaner context.  If we have enough of these built
4150          * up we can end up burning a lot of time in btrfs-cleaner without any
4151          * way to throttle the unlinks.  Since we're currently holding a ref on
4152          * the inode we can run the delayed iput here without any issues as the
4153          * final iput won't be done until after we drop the ref we're currently
4154          * holding.
4155          */
4156         btrfs_run_delayed_iput(fs_info, inode);
4157 err:
4158         btrfs_free_path(path);
4159         if (ret)
4160                 goto out;
4161
4162         btrfs_i_size_write(dir, dir->vfs_inode.i_size - name_len * 2);
4163         inode_inc_iversion(&inode->vfs_inode);
4164         inode_inc_iversion(&dir->vfs_inode);
4165         inode->vfs_inode.i_ctime = dir->vfs_inode.i_mtime =
4166                 dir->vfs_inode.i_ctime = current_time(&inode->vfs_inode);
4167         ret = btrfs_update_inode(trans, root, &dir->vfs_inode);
4168 out:
4169         return ret;
4170 }
4171
4172 int btrfs_unlink_inode(struct btrfs_trans_handle *trans,
4173                        struct btrfs_root *root,
4174                        struct btrfs_inode *dir, struct btrfs_inode *inode,
4175                        const char *name, int name_len)
4176 {
4177         int ret;
4178         ret = __btrfs_unlink_inode(trans, root, dir, inode, name, name_len);
4179         if (!ret) {
4180                 drop_nlink(&inode->vfs_inode);
4181                 ret = btrfs_update_inode(trans, root, &inode->vfs_inode);
4182         }
4183         return ret;
4184 }
4185
4186 /*
4187  * helper to start transaction for unlink and rmdir.
4188  *
4189  * unlink and rmdir are special in btrfs, they do not always free space, so
4190  * if we cannot make our reservations the normal way try and see if there is
4191  * plenty of slack room in the global reserve to migrate, otherwise we cannot
4192  * allow the unlink to occur.
4193  */
4194 static struct btrfs_trans_handle *__unlink_start_trans(struct inode *dir)
4195 {
4196         struct btrfs_root *root = BTRFS_I(dir)->root;
4197
4198         /*
4199          * 1 for the possible orphan item
4200          * 1 for the dir item
4201          * 1 for the dir index
4202          * 1 for the inode ref
4203          * 1 for the inode
4204          */
4205         return btrfs_start_transaction_fallback_global_rsv(root, 5, 5);
4206 }
4207
4208 static int btrfs_unlink(struct inode *dir, struct dentry *dentry)
4209 {
4210         struct btrfs_root *root = BTRFS_I(dir)->root;
4211         struct btrfs_trans_handle *trans;
4212         struct inode *inode = d_inode(dentry);
4213         int ret;
4214
4215         trans = __unlink_start_trans(dir);
4216         if (IS_ERR(trans))
4217                 return PTR_ERR(trans);
4218
4219         btrfs_record_unlink_dir(trans, BTRFS_I(dir), BTRFS_I(d_inode(dentry)),
4220                         0);
4221
4222         ret = btrfs_unlink_inode(trans, root, BTRFS_I(dir),
4223                         BTRFS_I(d_inode(dentry)), dentry->d_name.name,
4224                         dentry->d_name.len);
4225         if (ret)
4226                 goto out;
4227
4228         if (inode->i_nlink == 0) {
4229                 ret = btrfs_orphan_add(trans, BTRFS_I(inode));
4230                 if (ret)
4231                         goto out;
4232         }
4233
4234 out:
4235         btrfs_end_transaction(trans);
4236         btrfs_btree_balance_dirty(root->fs_info);
4237         return ret;
4238 }
4239
4240 static int btrfs_unlink_subvol(struct btrfs_trans_handle *trans,
4241                                struct inode *dir, struct dentry *dentry)
4242 {
4243         struct btrfs_root *root = BTRFS_I(dir)->root;
4244         struct btrfs_inode *inode = BTRFS_I(d_inode(dentry));
4245         struct btrfs_path *path;
4246         struct extent_buffer *leaf;
4247         struct btrfs_dir_item *di;
4248         struct btrfs_key key;
4249         const char *name = dentry->d_name.name;
4250         int name_len = dentry->d_name.len;
4251         u64 index;
4252         int ret;
4253         u64 objectid;
4254         u64 dir_ino = btrfs_ino(BTRFS_I(dir));
4255
4256         if (btrfs_ino(inode) == BTRFS_FIRST_FREE_OBJECTID) {
4257                 objectid = inode->root->root_key.objectid;
4258         } else if (btrfs_ino(inode) == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID) {
4259                 objectid = inode->location.objectid;
4260         } else {
4261                 WARN_ON(1);
4262                 return -EINVAL;
4263         }
4264
4265         path = btrfs_alloc_path();
4266         if (!path)
4267                 return -ENOMEM;
4268
4269         di = btrfs_lookup_dir_item(trans, root, path, dir_ino,
4270                                    name, name_len, -1);
4271         if (IS_ERR_OR_NULL(di)) {
4272                 ret = di ? PTR_ERR(di) : -ENOENT;
4273                 goto out;
4274         }
4275
4276         leaf = path->nodes[0];
4277         btrfs_dir_item_key_to_cpu(leaf, di, &key);
4278         WARN_ON(key.type != BTRFS_ROOT_ITEM_KEY || key.objectid != objectid);
4279         ret = btrfs_delete_one_dir_name(trans, root, path, di);
4280         if (ret) {
4281                 btrfs_abort_transaction(trans, ret);
4282                 goto out;
4283         }
4284         btrfs_release_path(path);
4285
4286         /*
4287          * This is a placeholder inode for a subvolume we didn't have a
4288          * reference to at the time of the snapshot creation.  In the meantime
4289          * we could have renamed the real subvol link into our snapshot, so
4290          * depending on btrfs_del_root_ref to return -ENOENT here is incorret.
4291          * Instead simply lookup the dir_index_item for this entry so we can
4292          * remove it.  Otherwise we know we have a ref to the root and we can
4293          * call btrfs_del_root_ref, and it _shouldn't_ fail.
4294          */
4295         if (btrfs_ino(inode) == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID) {
4296                 di = btrfs_search_dir_index_item(root, path, dir_ino,
4297                                                  name, name_len);
4298                 if (IS_ERR_OR_NULL(di)) {
4299                         if (!di)
4300                                 ret = -ENOENT;
4301                         else
4302                                 ret = PTR_ERR(di);
4303                         btrfs_abort_transaction(trans, ret);
4304                         goto out;
4305                 }
4306
4307                 leaf = path->nodes[0];
4308                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
4309                 index = key.offset;
4310                 btrfs_release_path(path);
4311         } else {
4312                 ret = btrfs_del_root_ref(trans, objectid,
4313                                          root->root_key.objectid, dir_ino,
4314                                          &index, name, name_len);
4315                 if (ret) {
4316                         btrfs_abort_transaction(trans, ret);
4317                         goto out;
4318                 }
4319         }
4320
4321         ret = btrfs_delete_delayed_dir_index(trans, BTRFS_I(dir), index);
4322         if (ret) {
4323                 btrfs_abort_transaction(trans, ret);
4324                 goto out;
4325         }
4326
4327         btrfs_i_size_write(BTRFS_I(dir), dir->i_size - name_len * 2);
4328         inode_inc_iversion(dir);
4329         dir->i_mtime = dir->i_ctime = current_time(dir);
4330         ret = btrfs_update_inode_fallback(trans, root, dir);
4331         if (ret)
4332                 btrfs_abort_transaction(trans, ret);
4333 out:
4334         btrfs_free_path(path);
4335         return ret;
4336 }
4337
4338 /*
4339  * Helper to check if the subvolume references other subvolumes or if it's
4340  * default.
4341  */
4342 static noinline int may_destroy_subvol(struct btrfs_root *root)
4343 {
4344         struct btrfs_fs_info *fs_info = root->fs_info;
4345         struct btrfs_path *path;
4346         struct btrfs_dir_item *di;
4347         struct btrfs_key key;
4348         u64 dir_id;
4349         int ret;
4350
4351         path = btrfs_alloc_path();
4352         if (!path)
4353                 return -ENOMEM;
4354
4355         /* Make sure this root isn't set as the default subvol */
4356         dir_id = btrfs_super_root_dir(fs_info->super_copy);
4357         di = btrfs_lookup_dir_item(NULL, fs_info->tree_root, path,
4358                                    dir_id, "default", 7, 0);
4359         if (di && !IS_ERR(di)) {
4360                 btrfs_dir_item_key_to_cpu(path->nodes[0], di, &key);
4361                 if (key.objectid == root->root_key.objectid) {
4362                         ret = -EPERM;
4363                         btrfs_err(fs_info,
4364                                   "deleting default subvolume %llu is not allowed",
4365                                   key.objectid);
4366                         goto out;
4367                 }
4368                 btrfs_release_path(path);
4369         }
4370
4371         key.objectid = root->root_key.objectid;
4372         key.type = BTRFS_ROOT_REF_KEY;
4373         key.offset = (u64)-1;
4374
4375         ret = btrfs_search_slot(NULL, fs_info->tree_root, &key, path, 0, 0);
4376         if (ret < 0)
4377                 goto out;
4378         BUG_ON(ret == 0);
4379
4380         ret = 0;
4381         if (path->slots[0] > 0) {
4382                 path->slots[0]--;
4383                 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
4384                 if (key.objectid == root->root_key.objectid &&
4385                     key.type == BTRFS_ROOT_REF_KEY)
4386                         ret = -ENOTEMPTY;
4387         }
4388 out:
4389         btrfs_free_path(path);
4390         return ret;
4391 }
4392
4393 /* Delete all dentries for inodes belonging to the root */
4394 static void btrfs_prune_dentries(struct btrfs_root *root)
4395 {
4396         struct btrfs_fs_info *fs_info = root->fs_info;
4397         struct rb_node *node;
4398         struct rb_node *prev;
4399         struct btrfs_inode *entry;
4400         struct inode *inode;
4401         u64 objectid = 0;
4402
4403         if (!test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state))
4404                 WARN_ON(btrfs_root_refs(&root->root_item) != 0);
4405
4406         spin_lock(&root->inode_lock);
4407 again:
4408         node = root->inode_tree.rb_node;
4409         prev = NULL;
4410         while (node) {
4411                 prev = node;
4412                 entry = rb_entry(node, struct btrfs_inode, rb_node);
4413
4414                 if (objectid < btrfs_ino(entry))
4415                         node = node->rb_left;
4416                 else if (objectid > btrfs_ino(entry))
4417                         node = node->rb_right;
4418                 else
4419                         break;
4420         }
4421         if (!node) {
4422                 while (prev) {
4423                         entry = rb_entry(prev, struct btrfs_inode, rb_node);
4424                         if (objectid <= btrfs_ino(entry)) {
4425                                 node = prev;
4426                                 break;
4427                         }
4428                         prev = rb_next(prev);
4429                 }
4430         }
4431         while (node) {
4432                 entry = rb_entry(node, struct btrfs_inode, rb_node);
4433                 objectid = btrfs_ino(entry) + 1;
4434                 inode = igrab(&entry->vfs_inode);
4435                 if (inode) {
4436                         spin_unlock(&root->inode_lock);
4437                         if (atomic_read(&inode->i_count) > 1)
4438                                 d_prune_aliases(inode);
4439                         /*
4440                          * btrfs_drop_inode will have it removed from the inode
4441                          * cache when its usage count hits zero.
4442                          */
4443                         iput(inode);
4444                         cond_resched();
4445                         spin_lock(&root->inode_lock);
4446                         goto again;
4447                 }
4448
4449                 if (cond_resched_lock(&root->inode_lock))
4450                         goto again;
4451
4452                 node = rb_next(node);
4453         }
4454         spin_unlock(&root->inode_lock);
4455 }
4456
4457 int btrfs_delete_subvolume(struct inode *dir, struct dentry *dentry)
4458 {
4459         struct btrfs_fs_info *fs_info = btrfs_sb(dentry->d_sb);
4460         struct btrfs_root *root = BTRFS_I(dir)->root;
4461         struct inode *inode = d_inode(dentry);
4462         struct btrfs_root *dest = BTRFS_I(inode)->root;
4463         struct btrfs_trans_handle *trans;
4464         struct btrfs_block_rsv block_rsv;
4465         u64 root_flags;
4466         int ret;
4467         int err;
4468
4469         /*
4470          * Don't allow to delete a subvolume with send in progress. This is
4471          * inside the inode lock so the error handling that has to drop the bit
4472          * again is not run concurrently.
4473          */
4474         spin_lock(&dest->root_item_lock);
4475         if (dest->send_in_progress) {
4476                 spin_unlock(&dest->root_item_lock);
4477                 btrfs_warn(fs_info,
4478                            "attempt to delete subvolume %llu during send",
4479                            dest->root_key.objectid);
4480                 return -EPERM;
4481         }
4482         root_flags = btrfs_root_flags(&dest->root_item);
4483         btrfs_set_root_flags(&dest->root_item,
4484                              root_flags | BTRFS_ROOT_SUBVOL_DEAD);
4485         spin_unlock(&dest->root_item_lock);
4486
4487         down_write(&fs_info->subvol_sem);
4488
4489         err = may_destroy_subvol(dest);
4490         if (err)
4491                 goto out_up_write;
4492
4493         btrfs_init_block_rsv(&block_rsv, BTRFS_BLOCK_RSV_TEMP);
4494         /*
4495          * One for dir inode,
4496          * two for dir entries,
4497          * two for root ref/backref.
4498          */
4499         err = btrfs_subvolume_reserve_metadata(root, &block_rsv, 5, true);
4500         if (err)
4501                 goto out_up_write;
4502
4503         trans = btrfs_start_transaction(root, 0);
4504         if (IS_ERR(trans)) {
4505                 err = PTR_ERR(trans);
4506                 goto out_release;
4507         }
4508         trans->block_rsv = &block_rsv;
4509         trans->bytes_reserved = block_rsv.size;
4510
4511         btrfs_record_snapshot_destroy(trans, BTRFS_I(dir));
4512
4513         ret = btrfs_unlink_subvol(trans, dir, dentry);
4514         if (ret) {
4515                 err = ret;
4516                 btrfs_abort_transaction(trans, ret);
4517                 goto out_end_trans;
4518         }
4519
4520         btrfs_record_root_in_trans(trans, dest);
4521
4522         memset(&dest->root_item.drop_progress, 0,
4523                 sizeof(dest->root_item.drop_progress));
4524         dest->root_item.drop_level = 0;
4525         btrfs_set_root_refs(&dest->root_item, 0);
4526
4527         if (!test_and_set_bit(BTRFS_ROOT_ORPHAN_ITEM_INSERTED, &dest->state)) {
4528                 ret = btrfs_insert_orphan_item(trans,
4529                                         fs_info->tree_root,
4530                                         dest->root_key.objectid);
4531                 if (ret) {
4532                         btrfs_abort_transaction(trans, ret);
4533                         err = ret;
4534                         goto out_end_trans;
4535                 }
4536         }
4537
4538         ret = btrfs_uuid_tree_remove(trans, dest->root_item.uuid,
4539                                   BTRFS_UUID_KEY_SUBVOL,
4540                                   dest->root_key.objectid);
4541         if (ret && ret != -ENOENT) {
4542                 btrfs_abort_transaction(trans, ret);
4543                 err = ret;
4544                 goto out_end_trans;
4545         }
4546         if (!btrfs_is_empty_uuid(dest->root_item.received_uuid)) {
4547                 ret = btrfs_uuid_tree_remove(trans,
4548                                           dest->root_item.received_uuid,
4549                                           BTRFS_UUID_KEY_RECEIVED_SUBVOL,
4550                                           dest->root_key.objectid);
4551                 if (ret && ret != -ENOENT) {
4552                         btrfs_abort_transaction(trans, ret);
4553                         err = ret;
4554                         goto out_end_trans;
4555                 }
4556         }
4557
4558 out_end_trans:
4559         trans->block_rsv = NULL;
4560         trans->bytes_reserved = 0;
4561         ret = btrfs_end_transaction(trans);
4562         if (ret && !err)
4563                 err = ret;
4564         inode->i_flags |= S_DEAD;
4565 out_release:
4566         btrfs_subvolume_release_metadata(fs_info, &block_rsv);
4567 out_up_write:
4568         up_write(&fs_info->subvol_sem);
4569         if (err) {
4570                 spin_lock(&dest->root_item_lock);
4571                 root_flags = btrfs_root_flags(&dest->root_item);
4572                 btrfs_set_root_flags(&dest->root_item,
4573                                 root_flags & ~BTRFS_ROOT_SUBVOL_DEAD);
4574                 spin_unlock(&dest->root_item_lock);
4575         } else {
4576                 d_invalidate(dentry);
4577                 btrfs_prune_dentries(dest);
4578                 ASSERT(dest->send_in_progress == 0);
4579
4580                 /* the last ref */
4581                 if (dest->ino_cache_inode) {
4582                         iput(dest->ino_cache_inode);
4583                         dest->ino_cache_inode = NULL;
4584                 }
4585         }
4586
4587         return err;
4588 }
4589
4590 static int btrfs_rmdir(struct inode *dir, struct dentry *dentry)
4591 {
4592         struct inode *inode = d_inode(dentry);
4593         int err = 0;
4594         struct btrfs_root *root = BTRFS_I(dir)->root;
4595         struct btrfs_trans_handle *trans;
4596         u64 last_unlink_trans;
4597
4598         if (inode->i_size > BTRFS_EMPTY_DIR_SIZE)
4599                 return -ENOTEMPTY;
4600         if (btrfs_ino(BTRFS_I(inode)) == BTRFS_FIRST_FREE_OBJECTID)
4601                 return btrfs_delete_subvolume(dir, dentry);
4602
4603         trans = __unlink_start_trans(dir);
4604         if (IS_ERR(trans))
4605                 return PTR_ERR(trans);
4606
4607         if (unlikely(btrfs_ino(BTRFS_I(inode)) == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)) {
4608                 err = btrfs_unlink_subvol(trans, dir, dentry);
4609                 goto out;
4610         }
4611
4612         err = btrfs_orphan_add(trans, BTRFS_I(inode));
4613         if (err)
4614                 goto out;
4615
4616         last_unlink_trans = BTRFS_I(inode)->last_unlink_trans;
4617
4618         /* now the directory is empty */
4619         err = btrfs_unlink_inode(trans, root, BTRFS_I(dir),
4620                         BTRFS_I(d_inode(dentry)), dentry->d_name.name,
4621                         dentry->d_name.len);
4622         if (!err) {
4623                 btrfs_i_size_write(BTRFS_I(inode), 0);
4624                 /*
4625                  * Propagate the last_unlink_trans value of the deleted dir to
4626                  * its parent directory. This is to prevent an unrecoverable
4627                  * log tree in the case we do something like this:
4628                  * 1) create dir foo
4629                  * 2) create snapshot under dir foo
4630                  * 3) delete the snapshot
4631                  * 4) rmdir foo
4632                  * 5) mkdir foo
4633                  * 6) fsync foo or some file inside foo
4634                  */
4635                 if (last_unlink_trans >= trans->transid)
4636                         BTRFS_I(dir)->last_unlink_trans = last_unlink_trans;
4637         }
4638 out:
4639         btrfs_end_transaction(trans);
4640         btrfs_btree_balance_dirty(root->fs_info);
4641
4642         return err;
4643 }
4644
4645 /*
4646  * Return this if we need to call truncate_block for the last bit of the
4647  * truncate.
4648  */
4649 #define NEED_TRUNCATE_BLOCK 1
4650
4651 /*
4652  * this can truncate away extent items, csum items and directory items.
4653  * It starts at a high offset and removes keys until it can't find
4654  * any higher than new_size
4655  *
4656  * csum items that cross the new i_size are truncated to the new size
4657  * as well.
4658  *
4659  * min_type is the minimum key type to truncate down to.  If set to 0, this
4660  * will kill all the items on this inode, including the INODE_ITEM_KEY.
4661  */
4662 int btrfs_truncate_inode_items(struct btrfs_trans_handle *trans,
4663                                struct btrfs_root *root,
4664                                struct inode *inode,
4665                                u64 new_size, u32 min_type)
4666 {
4667         struct btrfs_fs_info *fs_info = root->fs_info;
4668         struct btrfs_path *path;
4669         struct extent_buffer *leaf;
4670         struct btrfs_file_extent_item *fi;
4671         struct btrfs_key key;
4672         struct btrfs_key found_key;
4673         u64 extent_start = 0;
4674         u64 extent_num_bytes = 0;
4675         u64 extent_offset = 0;
4676         u64 item_end = 0;
4677         u64 last_size = new_size;
4678         u32 found_type = (u8)-1;
4679         int found_extent;
4680         int del_item;
4681         int pending_del_nr = 0;
4682         int pending_del_slot = 0;
4683         int extent_type = -1;
4684         int ret;
4685         u64 ino = btrfs_ino(BTRFS_I(inode));
4686         u64 bytes_deleted = 0;
4687         bool be_nice = false;
4688         bool should_throttle = false;
4689
4690         BUG_ON(new_size > 0 && min_type != BTRFS_EXTENT_DATA_KEY);
4691
4692         /*
4693          * for non-free space inodes and ref cows, we want to back off from
4694          * time to time
4695          */
4696         if (!btrfs_is_free_space_inode(BTRFS_I(inode)) &&
4697             test_bit(BTRFS_ROOT_REF_COWS, &root->state))
4698                 be_nice = true;
4699
4700         path = btrfs_alloc_path();
4701         if (!path)
4702                 return -ENOMEM;
4703         path->reada = READA_BACK;
4704
4705         /*
4706          * We want to drop from the next block forward in case this new size is
4707          * not block aligned since we will be keeping the last block of the
4708          * extent just the way it is.
4709          */
4710         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
4711             root == fs_info->tree_root)
4712                 btrfs_drop_extent_cache(BTRFS_I(inode), ALIGN(new_size,
4713                                         fs_info->sectorsize),
4714                                         (u64)-1, 0);
4715
4716         /*
4717          * This function is also used to drop the items in the log tree before
4718          * we relog the inode, so if root != BTRFS_I(inode)->root, it means
4719          * it is used to drop the logged items. So we shouldn't kill the delayed
4720          * items.
4721          */
4722         if (min_type == 0 && root == BTRFS_I(inode)->root)
4723                 btrfs_kill_delayed_inode_items(BTRFS_I(inode));
4724
4725         key.objectid = ino;
4726         key.offset = (u64)-1;
4727         key.type = (u8)-1;
4728
4729 search_again:
4730         /*
4731          * with a 16K leaf size and 128MB extents, you can actually queue
4732          * up a huge file in a single leaf.  Most of the time that
4733          * bytes_deleted is > 0, it will be huge by the time we get here
4734          */
4735         if (be_nice && bytes_deleted > SZ_32M &&
4736             btrfs_should_end_transaction(trans)) {
4737                 ret = -EAGAIN;
4738                 goto out;
4739         }
4740
4741         path->leave_spinning = 1;
4742         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
4743         if (ret < 0)
4744                 goto out;
4745
4746         if (ret > 0) {
4747                 ret = 0;
4748                 /* there are no items in the tree for us to truncate, we're
4749                  * done
4750                  */
4751                 if (path->slots[0] == 0)
4752                         goto out;
4753                 path->slots[0]--;
4754         }
4755
4756         while (1) {
4757                 fi = NULL;
4758                 leaf = path->nodes[0];
4759                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
4760                 found_type = found_key.type;
4761
4762                 if (found_key.objectid != ino)
4763                         break;
4764
4765                 if (found_type < min_type)
4766                         break;
4767
4768                 item_end = found_key.offset;
4769                 if (found_type == BTRFS_EXTENT_DATA_KEY) {
4770                         fi = btrfs_item_ptr(leaf, path->slots[0],
4771                                             struct btrfs_file_extent_item);
4772                         extent_type = btrfs_file_extent_type(leaf, fi);
4773                         if (extent_type != BTRFS_FILE_EXTENT_INLINE) {
4774                                 item_end +=
4775                                     btrfs_file_extent_num_bytes(leaf, fi);
4776
4777                                 trace_btrfs_truncate_show_fi_regular(
4778                                         BTRFS_I(inode), leaf, fi,
4779                                         found_key.offset);
4780                         } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
4781                                 item_end += btrfs_file_extent_ram_bytes(leaf,
4782                                                                         fi);
4783
4784                                 trace_btrfs_truncate_show_fi_inline(
4785                                         BTRFS_I(inode), leaf, fi, path->slots[0],
4786                                         found_key.offset);
4787                         }
4788                         item_end--;
4789                 }
4790                 if (found_type > min_type) {
4791                         del_item = 1;
4792                 } else {
4793                         if (item_end < new_size)
4794                                 break;
4795                         if (found_key.offset >= new_size)
4796                                 del_item = 1;
4797                         else
4798                                 del_item = 0;
4799                 }
4800                 found_extent = 0;
4801                 /* FIXME, shrink the extent if the ref count is only 1 */
4802                 if (found_type != BTRFS_EXTENT_DATA_KEY)
4803                         goto delete;
4804
4805                 if (extent_type != BTRFS_FILE_EXTENT_INLINE) {
4806                         u64 num_dec;
4807                         extent_start = btrfs_file_extent_disk_bytenr(leaf, fi);
4808                         if (!del_item) {
4809                                 u64 orig_num_bytes =
4810                                         btrfs_file_extent_num_bytes(leaf, fi);
4811                                 extent_num_bytes = ALIGN(new_size -
4812                                                 found_key.offset,
4813                                                 fs_info->sectorsize);
4814                                 btrfs_set_file_extent_num_bytes(leaf, fi,
4815                                                          extent_num_bytes);
4816                                 num_dec = (orig_num_bytes -
4817                                            extent_num_bytes);
4818                                 if (test_bit(BTRFS_ROOT_REF_COWS,
4819                                              &root->state) &&
4820                                     extent_start != 0)
4821                                         inode_sub_bytes(inode, num_dec);
4822                                 btrfs_mark_buffer_dirty(leaf);
4823                         } else {
4824                                 extent_num_bytes =
4825                                         btrfs_file_extent_disk_num_bytes(leaf,
4826                                                                          fi);
4827                                 extent_offset = found_key.offset -
4828                                         btrfs_file_extent_offset(leaf, fi);
4829
4830                                 /* FIXME blocksize != 4096 */
4831                                 num_dec = btrfs_file_extent_num_bytes(leaf, fi);
4832                                 if (extent_start != 0) {
4833                                         found_extent = 1;
4834                                         if (test_bit(BTRFS_ROOT_REF_COWS,
4835                                                      &root->state))
4836                                                 inode_sub_bytes(inode, num_dec);
4837                                 }
4838                         }
4839                 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
4840                         /*
4841                          * we can't truncate inline items that have had
4842                          * special encodings
4843                          */
4844                         if (!del_item &&
4845                             btrfs_file_extent_encryption(leaf, fi) == 0 &&
4846                             btrfs_file_extent_other_encoding(leaf, fi) == 0 &&
4847                             btrfs_file_extent_compression(leaf, fi) == 0) {
4848                                 u32 size = (u32)(new_size - found_key.offset);
4849
4850                                 btrfs_set_file_extent_ram_bytes(leaf, fi, size);
4851                                 size = btrfs_file_extent_calc_inline_size(size);
4852                                 btrfs_truncate_item(path, size, 1);
4853                         } else if (!del_item) {
4854                                 /*
4855                                  * We have to bail so the last_size is set to
4856                                  * just before this extent.
4857                                  */
4858                                 ret = NEED_TRUNCATE_BLOCK;
4859                                 break;
4860                         }
4861
4862                         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state))
4863                                 inode_sub_bytes(inode, item_end + 1 - new_size);
4864                 }
4865 delete:
4866                 if (del_item)
4867                         last_size = found_key.offset;
4868                 else
4869                         last_size = new_size;
4870                 if (del_item) {
4871                         if (!pending_del_nr) {
4872                                 /* no pending yet, add ourselves */
4873                                 pending_del_slot = path->slots[0];
4874                                 pending_del_nr = 1;
4875                         } else if (pending_del_nr &&
4876                                    path->slots[0] + 1 == pending_del_slot) {
4877                                 /* hop on the pending chunk */
4878                                 pending_del_nr++;
4879                                 pending_del_slot = path->slots[0];
4880                         } else {
4881                                 BUG();
4882                         }
4883                 } else {
4884                         break;
4885                 }
4886                 should_throttle = false;
4887
4888                 if (found_extent &&
4889                     (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
4890                      root == fs_info->tree_root)) {
4891                         struct btrfs_ref ref = { 0 };
4892
4893                         btrfs_set_path_blocking(path);
4894                         bytes_deleted += extent_num_bytes;
4895
4896                         btrfs_init_generic_ref(&ref, BTRFS_DROP_DELAYED_REF,
4897                                         extent_start, extent_num_bytes, 0);
4898                         ref.real_root = root->root_key.objectid;
4899                         btrfs_init_data_ref(&ref, btrfs_header_owner(leaf),
4900                                         ino, extent_offset);
4901                         ret = btrfs_free_extent(trans, &ref);
4902                         if (ret) {
4903                                 btrfs_abort_transaction(trans, ret);
4904                                 break;
4905                         }
4906                         if (be_nice) {
4907                                 if (btrfs_should_throttle_delayed_refs(trans))
4908                                         should_throttle = true;
4909                         }
4910                 }
4911
4912                 if (found_type == BTRFS_INODE_ITEM_KEY)
4913                         break;
4914
4915                 if (path->slots[0] == 0 ||
4916                     path->slots[0] != pending_del_slot ||
4917                     should_throttle) {
4918                         if (pending_del_nr) {
4919                                 ret = btrfs_del_items(trans, root, path,
4920                                                 pending_del_slot,
4921                                                 pending_del_nr);
4922                                 if (ret) {
4923                                         btrfs_abort_transaction(trans, ret);
4924                                         break;
4925                                 }
4926                                 pending_del_nr = 0;
4927                         }
4928                         btrfs_release_path(path);
4929
4930                         /*
4931                          * We can generate a lot of delayed refs, so we need to
4932                          * throttle every once and a while and make sure we're
4933                          * adding enough space to keep up with the work we are
4934                          * generating.  Since we hold a transaction here we
4935                          * can't flush, and we don't want to FLUSH_LIMIT because
4936                          * we could have generated too many delayed refs to
4937                          * actually allocate, so just bail if we're short and
4938                          * let the normal reservation dance happen higher up.
4939                          */
4940                         if (should_throttle) {
4941                                 ret = btrfs_delayed_refs_rsv_refill(fs_info,
4942                                                         BTRFS_RESERVE_NO_FLUSH);
4943                                 if (ret) {
4944                                         ret = -EAGAIN;
4945                                         break;
4946                                 }
4947                         }
4948                         goto search_again;
4949                 } else {
4950                         path->slots[0]--;
4951                 }
4952         }
4953 out:
4954         if (ret >= 0 && pending_del_nr) {
4955                 int err;
4956
4957                 err = btrfs_del_items(trans, root, path, pending_del_slot,
4958                                       pending_del_nr);
4959                 if (err) {
4960                         btrfs_abort_transaction(trans, err);
4961                         ret = err;
4962                 }
4963         }
4964         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
4965                 ASSERT(last_size >= new_size);
4966                 if (!ret && last_size > new_size)
4967                         last_size = new_size;
4968                 btrfs_ordered_update_i_size(inode, last_size, NULL);
4969         }
4970
4971         btrfs_free_path(path);
4972         return ret;
4973 }
4974
4975 /*
4976  * btrfs_truncate_block - read, zero a chunk and write a block
4977  * @inode - inode that we're zeroing
4978  * @from - the offset to start zeroing
4979  * @len - the length to zero, 0 to zero the entire range respective to the
4980  *      offset
4981  * @front - zero up to the offset instead of from the offset on
4982  *
4983  * This will find the block for the "from" offset and cow the block and zero the
4984  * part we want to zero.  This is used with truncate and hole punching.
4985  */
4986 int btrfs_truncate_block(struct inode *inode, loff_t from, loff_t len,
4987                         int front)
4988 {
4989         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
4990         struct address_space *mapping = inode->i_mapping;
4991         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
4992         struct btrfs_ordered_extent *ordered;
4993         struct extent_state *cached_state = NULL;
4994         struct extent_changeset *data_reserved = NULL;
4995         char *kaddr;
4996         u32 blocksize = fs_info->sectorsize;
4997         pgoff_t index = from >> PAGE_SHIFT;
4998         unsigned offset = from & (blocksize - 1);
4999         struct page *page;
5000         gfp_t mask = btrfs_alloc_write_mask(mapping);
5001         int ret = 0;
5002         u64 block_start;
5003         u64 block_end;
5004
5005         if (IS_ALIGNED(offset, blocksize) &&
5006             (!len || IS_ALIGNED(len, blocksize)))
5007                 goto out;
5008
5009         block_start = round_down(from, blocksize);
5010         block_end = block_start + blocksize - 1;
5011
5012         ret = btrfs_delalloc_reserve_space(inode, &data_reserved,
5013                                            block_start, blocksize);
5014         if (ret)
5015                 goto out;
5016
5017 again:
5018         page = find_or_create_page(mapping, index, mask);
5019         if (!page) {
5020                 btrfs_delalloc_release_space(inode, data_reserved,
5021                                              block_start, blocksize, true);
5022                 btrfs_delalloc_release_extents(BTRFS_I(inode), blocksize);
5023                 ret = -ENOMEM;
5024                 goto out;
5025         }
5026
5027         if (!PageUptodate(page)) {
5028                 ret = btrfs_readpage(NULL, page);
5029                 lock_page(page);
5030                 if (page->mapping != mapping) {
5031                         unlock_page(page);
5032                         put_page(page);
5033                         goto again;
5034                 }
5035                 if (!PageUptodate(page)) {
5036                         ret = -EIO;
5037                         goto out_unlock;
5038                 }
5039         }
5040         wait_on_page_writeback(page);
5041
5042         lock_extent_bits(io_tree, block_start, block_end, &cached_state);
5043         set_page_extent_mapped(page);
5044
5045         ordered = btrfs_lookup_ordered_extent(inode, block_start);
5046         if (ordered) {
5047                 unlock_extent_cached(io_tree, block_start, block_end,
5048                                      &cached_state);
5049                 unlock_page(page);
5050                 put_page(page);
5051                 btrfs_start_ordered_extent(inode, ordered, 1);
5052                 btrfs_put_ordered_extent(ordered);
5053                 goto again;
5054         }
5055
5056         clear_extent_bit(&BTRFS_I(inode)->io_tree, block_start, block_end,
5057                          EXTENT_DELALLOC | EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG,
5058                          0, 0, &cached_state);
5059
5060         ret = btrfs_set_extent_delalloc(inode, block_start, block_end, 0,
5061                                         &cached_state);
5062         if (ret) {
5063                 unlock_extent_cached(io_tree, block_start, block_end,
5064                                      &cached_state);
5065                 goto out_unlock;
5066         }
5067
5068         if (offset != blocksize) {
5069                 if (!len)
5070                         len = blocksize - offset;
5071                 kaddr = kmap(page);
5072                 if (front)
5073                         memset(kaddr + (block_start - page_offset(page)),
5074                                 0, offset);
5075                 else
5076                         memset(kaddr + (block_start - page_offset(page)) +  offset,
5077                                 0, len);
5078                 flush_dcache_page(page);
5079                 kunmap(page);
5080         }
5081         ClearPageChecked(page);
5082         set_page_dirty(page);
5083         unlock_extent_cached(io_tree, block_start, block_end, &cached_state);
5084
5085 out_unlock:
5086         if (ret)
5087                 btrfs_delalloc_release_space(inode, data_reserved, block_start,
5088                                              blocksize, true);
5089         btrfs_delalloc_release_extents(BTRFS_I(inode), blocksize);
5090         unlock_page(page);
5091         put_page(page);
5092 out:
5093         extent_changeset_free(data_reserved);
5094         return ret;
5095 }
5096
5097 static int maybe_insert_hole(struct btrfs_root *root, struct inode *inode,
5098                              u64 offset, u64 len)
5099 {
5100         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
5101         struct btrfs_trans_handle *trans;
5102         int ret;
5103
5104         /*
5105          * Still need to make sure the inode looks like it's been updated so
5106          * that any holes get logged if we fsync.
5107          */
5108         if (btrfs_fs_incompat(fs_info, NO_HOLES)) {
5109                 BTRFS_I(inode)->last_trans = fs_info->generation;
5110                 BTRFS_I(inode)->last_sub_trans = root->log_transid;
5111                 BTRFS_I(inode)->last_log_commit = root->last_log_commit;
5112                 return 0;
5113         }
5114
5115         /*
5116          * 1 - for the one we're dropping
5117          * 1 - for the one we're adding
5118          * 1 - for updating the inode.
5119          */
5120         trans = btrfs_start_transaction(root, 3);
5121         if (IS_ERR(trans))
5122                 return PTR_ERR(trans);
5123
5124         ret = btrfs_drop_extents(trans, root, inode, offset, offset + len, 1);
5125         if (ret) {
5126                 btrfs_abort_transaction(trans, ret);
5127                 btrfs_end_transaction(trans);
5128                 return ret;
5129         }
5130
5131         ret = btrfs_insert_file_extent(trans, root, btrfs_ino(BTRFS_I(inode)),
5132                         offset, 0, 0, len, 0, len, 0, 0, 0);
5133         if (ret)
5134                 btrfs_abort_transaction(trans, ret);
5135         else
5136                 btrfs_update_inode(trans, root, inode);
5137         btrfs_end_transaction(trans);
5138         return ret;
5139 }
5140
5141 /*
5142  * This function puts in dummy file extents for the area we're creating a hole
5143  * for.  So if we are truncating this file to a larger size we need to insert
5144  * these file extents so that btrfs_get_extent will return a EXTENT_MAP_HOLE for
5145  * the range between oldsize and size
5146  */
5147 int btrfs_cont_expand(struct inode *inode, loff_t oldsize, loff_t size)
5148 {
5149         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
5150         struct btrfs_root *root = BTRFS_I(inode)->root;
5151         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
5152         struct extent_map *em = NULL;
5153         struct extent_state *cached_state = NULL;
5154         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
5155         u64 hole_start = ALIGN(oldsize, fs_info->sectorsize);
5156         u64 block_end = ALIGN(size, fs_info->sectorsize);
5157         u64 last_byte;
5158         u64 cur_offset;
5159         u64 hole_size;
5160         int err = 0;
5161
5162         /*
5163          * If our size started in the middle of a block we need to zero out the
5164          * rest of the block before we expand the i_size, otherwise we could
5165          * expose stale data.
5166          */
5167         err = btrfs_truncate_block(inode, oldsize, 0, 0);
5168         if (err)
5169                 return err;
5170
5171         if (size <= hole_start)
5172                 return 0;
5173
5174         btrfs_lock_and_flush_ordered_range(io_tree, BTRFS_I(inode), hole_start,
5175                                            block_end - 1, &cached_state);
5176         cur_offset = hole_start;
5177         while (1) {
5178                 em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, cur_offset,
5179                                 block_end - cur_offset, 0);
5180                 if (IS_ERR(em)) {
5181                         err = PTR_ERR(em);
5182                         em = NULL;
5183                         break;
5184                 }
5185                 last_byte = min(extent_map_end(em), block_end);
5186                 last_byte = ALIGN(last_byte, fs_info->sectorsize);
5187                 if (!test_bit(EXTENT_FLAG_PREALLOC, &em->flags)) {
5188                         struct extent_map *hole_em;
5189                         hole_size = last_byte - cur_offset;
5190
5191                         err = maybe_insert_hole(root, inode, cur_offset,
5192                                                 hole_size);
5193                         if (err)
5194                                 break;
5195                         btrfs_drop_extent_cache(BTRFS_I(inode), cur_offset,
5196                                                 cur_offset + hole_size - 1, 0);
5197                         hole_em = alloc_extent_map();
5198                         if (!hole_em) {
5199                                 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
5200                                         &BTRFS_I(inode)->runtime_flags);
5201                                 goto next;
5202                         }
5203                         hole_em->start = cur_offset;
5204                         hole_em->len = hole_size;
5205                         hole_em->orig_start = cur_offset;
5206
5207                         hole_em->block_start = EXTENT_MAP_HOLE;
5208                         hole_em->block_len = 0;
5209                         hole_em->orig_block_len = 0;
5210                         hole_em->ram_bytes = hole_size;
5211                         hole_em->compress_type = BTRFS_COMPRESS_NONE;
5212                         hole_em->generation = fs_info->generation;
5213
5214                         while (1) {
5215                                 write_lock(&em_tree->lock);
5216                                 err = add_extent_mapping(em_tree, hole_em, 1);
5217                                 write_unlock(&em_tree->lock);
5218                                 if (err != -EEXIST)
5219                                         break;
5220                                 btrfs_drop_extent_cache(BTRFS_I(inode),
5221                                                         cur_offset,
5222                                                         cur_offset +
5223                                                         hole_size - 1, 0);
5224                         }
5225                         free_extent_map(hole_em);
5226                 }
5227 next:
5228                 free_extent_map(em);
5229                 em = NULL;
5230                 cur_offset = last_byte;
5231                 if (cur_offset >= block_end)
5232                         break;
5233         }
5234         free_extent_map(em);
5235         unlock_extent_cached(io_tree, hole_start, block_end - 1, &cached_state);
5236         return err;
5237 }
5238
5239 static int btrfs_setsize(struct inode *inode, struct iattr *attr)
5240 {
5241         struct btrfs_root *root = BTRFS_I(inode)->root;
5242         struct btrfs_trans_handle *trans;
5243         loff_t oldsize = i_size_read(inode);
5244         loff_t newsize = attr->ia_size;
5245         int mask = attr->ia_valid;
5246         int ret;
5247
5248         /*
5249          * The regular truncate() case without ATTR_CTIME and ATTR_MTIME is a
5250          * special case where we need to update the times despite not having
5251          * these flags set.  For all other operations the VFS set these flags
5252          * explicitly if it wants a timestamp update.
5253          */
5254         if (newsize != oldsize) {
5255                 inode_inc_iversion(inode);
5256                 if (!(mask & (ATTR_CTIME | ATTR_MTIME)))
5257                         inode->i_ctime = inode->i_mtime =
5258                                 current_time(inode);
5259         }
5260
5261         if (newsize > oldsize) {
5262                 /*
5263                  * Don't do an expanding truncate while snapshotting is ongoing.
5264                  * This is to ensure the snapshot captures a fully consistent
5265                  * state of this file - if the snapshot captures this expanding
5266                  * truncation, it must capture all writes that happened before
5267                  * this truncation.
5268                  */
5269                 btrfs_wait_for_snapshot_creation(root);
5270                 ret = btrfs_cont_expand(inode, oldsize, newsize);
5271                 if (ret) {
5272                         btrfs_end_write_no_snapshotting(root);
5273                         return ret;
5274                 }
5275
5276                 trans = btrfs_start_transaction(root, 1);
5277                 if (IS_ERR(trans)) {
5278                         btrfs_end_write_no_snapshotting(root);
5279                         return PTR_ERR(trans);
5280                 }
5281
5282                 i_size_write(inode, newsize);
5283                 btrfs_ordered_update_i_size(inode, i_size_read(inode), NULL);
5284                 pagecache_isize_extended(inode, oldsize, newsize);
5285                 ret = btrfs_update_inode(trans, root, inode);
5286                 btrfs_end_write_no_snapshotting(root);
5287                 btrfs_end_transaction(trans);
5288         } else {
5289
5290                 /*
5291                  * We're truncating a file that used to have good data down to
5292                  * zero. Make sure it gets into the ordered flush list so that
5293                  * any new writes get down to disk quickly.
5294                  */
5295                 if (newsize == 0)
5296                         set_bit(BTRFS_INODE_ORDERED_DATA_CLOSE,
5297                                 &BTRFS_I(inode)->runtime_flags);
5298
5299                 truncate_setsize(inode, newsize);
5300
5301                 /* Disable nonlocked read DIO to avoid the endless truncate */
5302                 btrfs_inode_block_unlocked_dio(BTRFS_I(inode));
5303                 inode_dio_wait(inode);
5304                 btrfs_inode_resume_unlocked_dio(BTRFS_I(inode));
5305
5306                 ret = btrfs_truncate(inode, newsize == oldsize);
5307                 if (ret && inode->i_nlink) {
5308                         int err;
5309
5310                         /*
5311                          * Truncate failed, so fix up the in-memory size. We
5312                          * adjusted disk_i_size down as we removed extents, so
5313                          * wait for disk_i_size to be stable and then update the
5314                          * in-memory size to match.
5315                          */
5316                         err = btrfs_wait_ordered_range(inode, 0, (u64)-1);
5317                         if (err)
5318                                 return err;
5319                         i_size_write(inode, BTRFS_I(inode)->disk_i_size);
5320                 }
5321         }
5322
5323         return ret;
5324 }
5325
5326 static int btrfs_setattr(struct dentry *dentry, struct iattr *attr)
5327 {
5328         struct inode *inode = d_inode(dentry);
5329         struct btrfs_root *root = BTRFS_I(inode)->root;
5330         int err;
5331
5332         if (btrfs_root_readonly(root))
5333                 return -EROFS;
5334
5335         err = setattr_prepare(dentry, attr);
5336         if (err)
5337                 return err;
5338
5339         if (S_ISREG(inode->i_mode) && (attr->ia_valid & ATTR_SIZE)) {
5340                 err = btrfs_setsize(inode, attr);
5341                 if (err)
5342                         return err;
5343         }
5344
5345         if (attr->ia_valid) {
5346                 setattr_copy(inode, attr);
5347                 inode_inc_iversion(inode);
5348                 err = btrfs_dirty_inode(inode);
5349
5350                 if (!err && attr->ia_valid & ATTR_MODE)
5351                         err = posix_acl_chmod(inode, inode->i_mode);
5352         }
5353
5354         return err;
5355 }
5356
5357 /*
5358  * While truncating the inode pages during eviction, we get the VFS calling
5359  * btrfs_invalidatepage() against each page of the inode. This is slow because
5360  * the calls to btrfs_invalidatepage() result in a huge amount of calls to
5361  * lock_extent_bits() and clear_extent_bit(), which keep merging and splitting
5362  * extent_state structures over and over, wasting lots of time.
5363  *
5364  * Therefore if the inode is being evicted, let btrfs_invalidatepage() skip all
5365  * those expensive operations on a per page basis and do only the ordered io
5366  * finishing, while we release here the extent_map and extent_state structures,
5367  * without the excessive merging and splitting.
5368  */
5369 static void evict_inode_truncate_pages(struct inode *inode)
5370 {
5371         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
5372         struct extent_map_tree *map_tree = &BTRFS_I(inode)->extent_tree;
5373         struct rb_node *node;
5374
5375         ASSERT(inode->i_state & I_FREEING);
5376         truncate_inode_pages_final(&inode->i_data);
5377
5378         write_lock(&map_tree->lock);
5379         while (!RB_EMPTY_ROOT(&map_tree->map.rb_root)) {
5380                 struct extent_map *em;
5381
5382                 node = rb_first_cached(&map_tree->map);
5383                 em = rb_entry(node, struct extent_map, rb_node);
5384                 clear_bit(EXTENT_FLAG_PINNED, &em->flags);
5385                 clear_bit(EXTENT_FLAG_LOGGING, &em->flags);
5386                 remove_extent_mapping(map_tree, em);
5387                 free_extent_map(em);
5388                 if (need_resched()) {
5389                         write_unlock(&map_tree->lock);
5390                         cond_resched();
5391                         write_lock(&map_tree->lock);
5392                 }
5393         }
5394         write_unlock(&map_tree->lock);
5395
5396         /*
5397          * Keep looping until we have no more ranges in the io tree.
5398          * We can have ongoing bios started by readpages (called from readahead)
5399          * that have their endio callback (extent_io.c:end_bio_extent_readpage)
5400          * still in progress (unlocked the pages in the bio but did not yet
5401          * unlocked the ranges in the io tree). Therefore this means some
5402          * ranges can still be locked and eviction started because before
5403          * submitting those bios, which are executed by a separate task (work
5404          * queue kthread), inode references (inode->i_count) were not taken
5405          * (which would be dropped in the end io callback of each bio).
5406          * Therefore here we effectively end up waiting for those bios and
5407          * anyone else holding locked ranges without having bumped the inode's
5408          * reference count - if we don't do it, when they access the inode's
5409          * io_tree to unlock a range it may be too late, leading to an
5410          * use-after-free issue.
5411          */
5412         spin_lock(&io_tree->lock);
5413         while (!RB_EMPTY_ROOT(&io_tree->state)) {
5414                 struct extent_state *state;
5415                 struct extent_state *cached_state = NULL;
5416                 u64 start;
5417                 u64 end;
5418                 unsigned state_flags;
5419
5420                 node = rb_first(&io_tree->state);
5421                 state = rb_entry(node, struct extent_state, rb_node);
5422                 start = state->start;
5423                 end = state->end;
5424                 state_flags = state->state;
5425                 spin_unlock(&io_tree->lock);
5426
5427                 lock_extent_bits(io_tree, start, end, &cached_state);
5428
5429                 /*
5430                  * If still has DELALLOC flag, the extent didn't reach disk,
5431                  * and its reserved space won't be freed by delayed_ref.
5432                  * So we need to free its reserved space here.
5433                  * (Refer to comment in btrfs_invalidatepage, case 2)
5434                  *
5435                  * Note, end is the bytenr of last byte, so we need + 1 here.
5436                  */
5437                 if (state_flags & EXTENT_DELALLOC)
5438                         btrfs_qgroup_free_data(inode, NULL, start, end - start + 1);
5439
5440                 clear_extent_bit(io_tree, start, end,
5441                                  EXTENT_LOCKED | EXTENT_DELALLOC |
5442                                  EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG, 1, 1,
5443                                  &cached_state);
5444
5445                 cond_resched();
5446                 spin_lock(&io_tree->lock);
5447         }
5448         spin_unlock(&io_tree->lock);
5449 }
5450
5451 static struct btrfs_trans_handle *evict_refill_and_join(struct btrfs_root *root,
5452                                                         struct btrfs_block_rsv *rsv)
5453 {
5454         struct btrfs_fs_info *fs_info = root->fs_info;
5455         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5456         struct btrfs_trans_handle *trans;
5457         u64 delayed_refs_extra = btrfs_calc_insert_metadata_size(fs_info, 1);
5458         int ret;
5459
5460         /*
5461          * Eviction should be taking place at some place safe because of our
5462          * delayed iputs.  However the normal flushing code will run delayed
5463          * iputs, so we cannot use FLUSH_ALL otherwise we'll deadlock.
5464          *
5465          * We reserve the delayed_refs_extra here again because we can't use
5466          * btrfs_start_transaction(root, 0) for the same deadlocky reason as
5467          * above.  We reserve our extra bit here because we generate a ton of
5468          * delayed refs activity by truncating.
5469          *
5470          * If we cannot make our reservation we'll attempt to steal from the
5471          * global reserve, because we really want to be able to free up space.
5472          */
5473         ret = btrfs_block_rsv_refill(root, rsv, rsv->size + delayed_refs_extra,
5474                                      BTRFS_RESERVE_FLUSH_EVICT);
5475         if (ret) {
5476                 /*
5477                  * Try to steal from the global reserve if there is space for
5478                  * it.
5479                  */
5480                 if (btrfs_check_space_for_delayed_refs(fs_info) ||
5481                     btrfs_block_rsv_migrate(global_rsv, rsv, rsv->size, 0)) {
5482                         btrfs_warn(fs_info,
5483                                    "could not allocate space for delete; will truncate on mount");
5484                         return ERR_PTR(-ENOSPC);
5485                 }
5486                 delayed_refs_extra = 0;
5487         }
5488
5489         trans = btrfs_join_transaction(root);
5490         if (IS_ERR(trans))
5491                 return trans;
5492
5493         if (delayed_refs_extra) {
5494                 trans->block_rsv = &fs_info->trans_block_rsv;
5495                 trans->bytes_reserved = delayed_refs_extra;
5496                 btrfs_block_rsv_migrate(rsv, trans->block_rsv,
5497                                         delayed_refs_extra, 1);
5498         }
5499         return trans;
5500 }
5501
5502 void btrfs_evict_inode(struct inode *inode)
5503 {
5504         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
5505         struct btrfs_trans_handle *trans;
5506         struct btrfs_root *root = BTRFS_I(inode)->root;
5507         struct btrfs_block_rsv *rsv;
5508         int ret;
5509
5510         trace_btrfs_inode_evict(inode);
5511
5512         if (!root) {
5513                 clear_inode(inode);
5514                 return;
5515         }
5516
5517         evict_inode_truncate_pages(inode);
5518
5519         if (inode->i_nlink &&
5520             ((btrfs_root_refs(&root->root_item) != 0 &&
5521               root->root_key.objectid != BTRFS_ROOT_TREE_OBJECTID) ||
5522              btrfs_is_free_space_inode(BTRFS_I(inode))))
5523                 goto no_delete;
5524
5525         if (is_bad_inode(inode))
5526                 goto no_delete;
5527
5528         btrfs_free_io_failure_record(BTRFS_I(inode), 0, (u64)-1);
5529
5530         if (test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags))
5531                 goto no_delete;
5532
5533         if (inode->i_nlink > 0) {
5534                 BUG_ON(btrfs_root_refs(&root->root_item) != 0 &&
5535                        root->root_key.objectid != BTRFS_ROOT_TREE_OBJECTID);
5536                 goto no_delete;
5537         }
5538
5539         ret = btrfs_commit_inode_delayed_inode(BTRFS_I(inode));
5540         if (ret)
5541                 goto no_delete;
5542
5543         rsv = btrfs_alloc_block_rsv(fs_info, BTRFS_BLOCK_RSV_TEMP);
5544         if (!rsv)
5545                 goto no_delete;
5546         rsv->size = btrfs_calc_metadata_size(fs_info, 1);
5547         rsv->failfast = 1;
5548
5549         btrfs_i_size_write(BTRFS_I(inode), 0);
5550
5551         while (1) {
5552                 trans = evict_refill_and_join(root, rsv);
5553                 if (IS_ERR(trans))
5554                         goto free_rsv;
5555
5556                 trans->block_rsv = rsv;
5557
5558                 ret = btrfs_truncate_inode_items(trans, root, inode, 0, 0);
5559                 trans->block_rsv = &fs_info->trans_block_rsv;
5560                 btrfs_end_transaction(trans);
5561                 btrfs_btree_balance_dirty(fs_info);
5562                 if (ret && ret != -ENOSPC && ret != -EAGAIN)
5563                         goto free_rsv;
5564                 else if (!ret)
5565                         break;
5566         }
5567
5568         /*
5569          * Errors here aren't a big deal, it just means we leave orphan items in
5570          * the tree. They will be cleaned up on the next mount. If the inode
5571          * number gets reused, cleanup deletes the orphan item without doing
5572          * anything, and unlink reuses the existing orphan item.
5573          *
5574          * If it turns out that we are dropping too many of these, we might want
5575          * to add a mechanism for retrying these after a commit.
5576          */
5577         trans = evict_refill_and_join(root, rsv);
5578         if (!IS_ERR(trans)) {
5579                 trans->block_rsv = rsv;
5580                 btrfs_orphan_del(trans, BTRFS_I(inode));
5581                 trans->block_rsv = &fs_info->trans_block_rsv;
5582                 btrfs_end_transaction(trans);
5583         }
5584
5585         if (!(root == fs_info->tree_root ||
5586               root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID))
5587                 btrfs_return_ino(root, btrfs_ino(BTRFS_I(inode)));
5588
5589 free_rsv:
5590         btrfs_free_block_rsv(fs_info, rsv);
5591 no_delete:
5592         /*
5593          * If we didn't successfully delete, the orphan item will still be in
5594          * the tree and we'll retry on the next mount. Again, we might also want
5595          * to retry these periodically in the future.
5596          */
5597         btrfs_remove_delayed_node(BTRFS_I(inode));
5598         clear_inode(inode);
5599 }
5600
5601 /*
5602  * Return the key found in the dir entry in the location pointer, fill @type
5603  * with BTRFS_FT_*, and return 0.
5604  *
5605  * If no dir entries were found, returns -ENOENT.
5606  * If found a corrupted location in dir entry, returns -EUCLEAN.
5607  */
5608 static int btrfs_inode_by_name(struct inode *dir, struct dentry *dentry,
5609                                struct btrfs_key *location, u8 *type)
5610 {
5611         const char *name = dentry->d_name.name;
5612         int namelen = dentry->d_name.len;
5613         struct btrfs_dir_item *di;
5614         struct btrfs_path *path;
5615         struct btrfs_root *root = BTRFS_I(dir)->root;
5616         int ret = 0;
5617
5618         path = btrfs_alloc_path();
5619         if (!path)
5620                 return -ENOMEM;
5621
5622         di = btrfs_lookup_dir_item(NULL, root, path, btrfs_ino(BTRFS_I(dir)),
5623                         name, namelen, 0);
5624         if (IS_ERR_OR_NULL(di)) {
5625                 ret = di ? PTR_ERR(di) : -ENOENT;
5626                 goto out;
5627         }
5628
5629         btrfs_dir_item_key_to_cpu(path->nodes[0], di, location);
5630         if (location->type != BTRFS_INODE_ITEM_KEY &&
5631             location->type != BTRFS_ROOT_ITEM_KEY) {
5632                 ret = -EUCLEAN;
5633                 btrfs_warn(root->fs_info,
5634 "%s gets something invalid in DIR_ITEM (name %s, directory ino %llu, location(%llu %u %llu))",
5635                            __func__, name, btrfs_ino(BTRFS_I(dir)),
5636                            location->objectid, location->type, location->offset);
5637         }
5638         if (!ret)
5639                 *type = btrfs_dir_type(path->nodes[0], di);
5640 out:
5641         btrfs_free_path(path);
5642         return ret;
5643 }
5644
5645 /*
5646  * when we hit a tree root in a directory, the btrfs part of the inode
5647  * needs to be changed to reflect the root directory of the tree root.  This
5648  * is kind of like crossing a mount point.
5649  */
5650 static int fixup_tree_root_location(struct btrfs_fs_info *fs_info,
5651                                     struct inode *dir,
5652                                     struct dentry *dentry,
5653                                     struct btrfs_key *location,
5654                                     struct btrfs_root **sub_root)
5655 {
5656         struct btrfs_path *path;
5657         struct btrfs_root *new_root;
5658         struct btrfs_root_ref *ref;
5659         struct extent_buffer *leaf;
5660         struct btrfs_key key;
5661         int ret;
5662         int err = 0;
5663
5664         path = btrfs_alloc_path();
5665         if (!path) {
5666                 err = -ENOMEM;
5667                 goto out;
5668         }
5669
5670         err = -ENOENT;
5671         key.objectid = BTRFS_I(dir)->root->root_key.objectid;
5672         key.type = BTRFS_ROOT_REF_KEY;
5673         key.offset = location->objectid;
5674
5675         ret = btrfs_search_slot(NULL, fs_info->tree_root, &key, path, 0, 0);
5676         if (ret) {
5677                 if (ret < 0)
5678                         err = ret;
5679                 goto out;
5680         }
5681
5682         leaf = path->nodes[0];
5683         ref = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_root_ref);
5684         if (btrfs_root_ref_dirid(leaf, ref) != btrfs_ino(BTRFS_I(dir)) ||
5685             btrfs_root_ref_name_len(leaf, ref) != dentry->d_name.len)
5686                 goto out;
5687
5688         ret = memcmp_extent_buffer(leaf, dentry->d_name.name,
5689                                    (unsigned long)(ref + 1),
5690                                    dentry->d_name.len);
5691         if (ret)
5692                 goto out;
5693
5694         btrfs_release_path(path);
5695
5696         new_root = btrfs_read_fs_root_no_name(fs_info, location);
5697         if (IS_ERR(new_root)) {
5698                 err = PTR_ERR(new_root);
5699                 goto out;
5700         }
5701
5702         *sub_root = new_root;
5703         location->objectid = btrfs_root_dirid(&new_root->root_item);
5704         location->type = BTRFS_INODE_ITEM_KEY;
5705         location->offset = 0;
5706         err = 0;
5707 out:
5708         btrfs_free_path(path);
5709         return err;
5710 }
5711
5712 static void inode_tree_add(struct inode *inode)
5713 {
5714         struct btrfs_root *root = BTRFS_I(inode)->root;
5715         struct btrfs_inode *entry;
5716         struct rb_node **p;
5717         struct rb_node *parent;
5718         struct rb_node *new = &BTRFS_I(inode)->rb_node;
5719         u64 ino = btrfs_ino(BTRFS_I(inode));
5720
5721         if (inode_unhashed(inode))
5722                 return;
5723         parent = NULL;
5724         spin_lock(&root->inode_lock);
5725         p = &root->inode_tree.rb_node;
5726         while (*p) {
5727                 parent = *p;
5728                 entry = rb_entry(parent, struct btrfs_inode, rb_node);
5729
5730                 if (ino < btrfs_ino(entry))
5731                         p = &parent->rb_left;
5732                 else if (ino > btrfs_ino(entry))
5733                         p = &parent->rb_right;
5734                 else {
5735                         WARN_ON(!(entry->vfs_inode.i_state &
5736                                   (I_WILL_FREE | I_FREEING)));
5737                         rb_replace_node(parent, new, &root->inode_tree);
5738                         RB_CLEAR_NODE(parent);
5739                         spin_unlock(&root->inode_lock);
5740                         return;
5741                 }
5742         }
5743         rb_link_node(new, parent, p);
5744         rb_insert_color(new, &root->inode_tree);
5745         spin_unlock(&root->inode_lock);
5746 }
5747
5748 static void inode_tree_del(struct inode *inode)
5749 {
5750         struct btrfs_root *root = BTRFS_I(inode)->root;
5751         int empty = 0;
5752
5753         spin_lock(&root->inode_lock);
5754         if (!RB_EMPTY_NODE(&BTRFS_I(inode)->rb_node)) {
5755                 rb_erase(&BTRFS_I(inode)->rb_node, &root->inode_tree);
5756                 RB_CLEAR_NODE(&BTRFS_I(inode)->rb_node);
5757                 empty = RB_EMPTY_ROOT(&root->inode_tree);
5758         }
5759         spin_unlock(&root->inode_lock);
5760
5761         if (empty && btrfs_root_refs(&root->root_item) == 0) {
5762                 spin_lock(&root->inode_lock);
5763                 empty = RB_EMPTY_ROOT(&root->inode_tree);
5764                 spin_unlock(&root->inode_lock);
5765                 if (empty)
5766                         btrfs_add_dead_root(root);
5767         }
5768 }
5769
5770
5771 static int btrfs_init_locked_inode(struct inode *inode, void *p)
5772 {
5773         struct btrfs_iget_args *args = p;
5774         inode->i_ino = args->location->objectid;
5775         memcpy(&BTRFS_I(inode)->location, args->location,
5776                sizeof(*args->location));
5777         BTRFS_I(inode)->root = args->root;
5778         return 0;
5779 }
5780
5781 static int btrfs_find_actor(struct inode *inode, void *opaque)
5782 {
5783         struct btrfs_iget_args *args = opaque;
5784         return args->location->objectid == BTRFS_I(inode)->location.objectid &&
5785                 args->root == BTRFS_I(inode)->root;
5786 }
5787
5788 static struct inode *btrfs_iget_locked(struct super_block *s,
5789                                        struct btrfs_key *location,
5790                                        struct btrfs_root *root)
5791 {
5792         struct inode *inode;
5793         struct btrfs_iget_args args;
5794         unsigned long hashval = btrfs_inode_hash(location->objectid, root);
5795
5796         args.location = location;
5797         args.root = root;
5798
5799         inode = iget5_locked(s, hashval, btrfs_find_actor,
5800                              btrfs_init_locked_inode,
5801                              (void *)&args);
5802         return inode;
5803 }
5804
5805 /*
5806  * Get an inode object given its location and corresponding root.
5807  * Path can be preallocated to prevent recursing back to iget through
5808  * allocator. NULL is also valid but may require an additional allocation
5809  * later.
5810  */
5811 struct inode *btrfs_iget_path(struct super_block *s, struct btrfs_key *location,
5812                               struct btrfs_root *root, struct btrfs_path *path)
5813 {
5814         struct inode *inode;
5815
5816         inode = btrfs_iget_locked(s, location, root);
5817         if (!inode)
5818                 return ERR_PTR(-ENOMEM);
5819
5820         if (inode->i_state & I_NEW) {
5821                 int ret;
5822
5823                 ret = btrfs_read_locked_inode(inode, path);
5824                 if (!ret) {
5825                         inode_tree_add(inode);
5826                         unlock_new_inode(inode);
5827                 } else {
5828                         iget_failed(inode);
5829                         /*
5830                          * ret > 0 can come from btrfs_search_slot called by
5831                          * btrfs_read_locked_inode, this means the inode item
5832                          * was not found.
5833                          */
5834                         if (ret > 0)
5835                                 ret = -ENOENT;
5836                         inode = ERR_PTR(ret);
5837                 }
5838         }
5839
5840         return inode;
5841 }
5842
5843 struct inode *btrfs_iget(struct super_block *s, struct btrfs_key *location,
5844                          struct btrfs_root *root)
5845 {
5846         return btrfs_iget_path(s, location, root, NULL);
5847 }
5848
5849 static struct inode *new_simple_dir(struct super_block *s,
5850                                     struct btrfs_key *key,
5851                                     struct btrfs_root *root)
5852 {
5853         struct inode *inode = new_inode(s);
5854
5855         if (!inode)
5856                 return ERR_PTR(-ENOMEM);
5857
5858         BTRFS_I(inode)->root = root;
5859         memcpy(&BTRFS_I(inode)->location, key, sizeof(*key));
5860         set_bit(BTRFS_INODE_DUMMY, &BTRFS_I(inode)->runtime_flags);
5861
5862         inode->i_ino = BTRFS_EMPTY_SUBVOL_DIR_OBJECTID;
5863         inode->i_op = &btrfs_dir_ro_inode_operations;
5864         inode->i_opflags &= ~IOP_XATTR;
5865         inode->i_fop = &simple_dir_operations;
5866         inode->i_mode = S_IFDIR | S_IRUGO | S_IWUSR | S_IXUGO;
5867         inode->i_mtime = current_time(inode);
5868         inode->i_atime = inode->i_mtime;
5869         inode->i_ctime = inode->i_mtime;
5870         BTRFS_I(inode)->i_otime = inode->i_mtime;
5871
5872         return inode;
5873 }
5874
5875 static inline u8 btrfs_inode_type(struct inode *inode)
5876 {
5877         /*
5878          * Compile-time asserts that generic FT_* types still match
5879          * BTRFS_FT_* types
5880          */
5881         BUILD_BUG_ON(BTRFS_FT_UNKNOWN != FT_UNKNOWN);
5882         BUILD_BUG_ON(BTRFS_FT_REG_FILE != FT_REG_FILE);
5883         BUILD_BUG_ON(BTRFS_FT_DIR != FT_DIR);
5884         BUILD_BUG_ON(BTRFS_FT_CHRDEV != FT_CHRDEV);
5885         BUILD_BUG_ON(BTRFS_FT_BLKDEV != FT_BLKDEV);
5886         BUILD_BUG_ON(BTRFS_FT_FIFO != FT_FIFO);
5887         BUILD_BUG_ON(BTRFS_FT_SOCK != FT_SOCK);
5888         BUILD_BUG_ON(BTRFS_FT_SYMLINK != FT_SYMLINK);
5889
5890         return fs_umode_to_ftype(inode->i_mode);
5891 }
5892
5893 struct inode *btrfs_lookup_dentry(struct inode *dir, struct dentry *dentry)
5894 {
5895         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
5896         struct inode *inode;
5897         struct btrfs_root *root = BTRFS_I(dir)->root;
5898         struct btrfs_root *sub_root = root;
5899         struct btrfs_key location;
5900         u8 di_type = 0;
5901         int index;
5902         int ret = 0;
5903
5904         if (dentry->d_name.len > BTRFS_NAME_LEN)
5905                 return ERR_PTR(-ENAMETOOLONG);
5906
5907         ret = btrfs_inode_by_name(dir, dentry, &location, &di_type);
5908         if (ret < 0)
5909                 return ERR_PTR(ret);
5910
5911         if (location.type == BTRFS_INODE_ITEM_KEY) {
5912                 inode = btrfs_iget(dir->i_sb, &location, root);
5913                 if (IS_ERR(inode))
5914                         return inode;
5915
5916                 /* Do extra check against inode mode with di_type */
5917                 if (btrfs_inode_type(inode) != di_type) {
5918                         btrfs_crit(fs_info,
5919 "inode mode mismatch with dir: inode mode=0%o btrfs type=%u dir type=%u",
5920                                   inode->i_mode, btrfs_inode_type(inode),
5921                                   di_type);
5922                         iput(inode);
5923                         return ERR_PTR(-EUCLEAN);
5924                 }
5925                 return inode;
5926         }
5927
5928         index = srcu_read_lock(&fs_info->subvol_srcu);
5929         ret = fixup_tree_root_location(fs_info, dir, dentry,
5930                                        &location, &sub_root);
5931         if (ret < 0) {
5932                 if (ret != -ENOENT)
5933                         inode = ERR_PTR(ret);
5934                 else
5935                         inode = new_simple_dir(dir->i_sb, &location, sub_root);
5936         } else {
5937                 inode = btrfs_iget(dir->i_sb, &location, sub_root);
5938         }
5939         srcu_read_unlock(&fs_info->subvol_srcu, index);
5940
5941         if (!IS_ERR(inode) && root != sub_root) {
5942                 down_read(&fs_info->cleanup_work_sem);
5943                 if (!sb_rdonly(inode->i_sb))
5944                         ret = btrfs_orphan_cleanup(sub_root);
5945                 up_read(&fs_info->cleanup_work_sem);
5946                 if (ret) {
5947                         iput(inode);
5948                         inode = ERR_PTR(ret);
5949                 }
5950         }
5951
5952         return inode;
5953 }
5954
5955 static int btrfs_dentry_delete(const struct dentry *dentry)
5956 {
5957         struct btrfs_root *root;
5958         struct inode *inode = d_inode(dentry);
5959
5960         if (!inode && !IS_ROOT(dentry))
5961                 inode = d_inode(dentry->d_parent);
5962
5963         if (inode) {
5964                 root = BTRFS_I(inode)->root;
5965                 if (btrfs_root_refs(&root->root_item) == 0)
5966                         return 1;
5967
5968                 if (btrfs_ino(BTRFS_I(inode)) == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)
5969                         return 1;
5970         }
5971         return 0;
5972 }
5973
5974 static struct dentry *btrfs_lookup(struct inode *dir, struct dentry *dentry,
5975                                    unsigned int flags)
5976 {
5977         struct inode *inode = btrfs_lookup_dentry(dir, dentry);
5978
5979         if (inode == ERR_PTR(-ENOENT))
5980                 inode = NULL;
5981         return d_splice_alias(inode, dentry);
5982 }
5983
5984 /*
5985  * All this infrastructure exists because dir_emit can fault, and we are holding
5986  * the tree lock when doing readdir.  For now just allocate a buffer and copy
5987  * our information into that, and then dir_emit from the buffer.  This is
5988  * similar to what NFS does, only we don't keep the buffer around in pagecache
5989  * because I'm afraid I'll mess that up.  Long term we need to make filldir do
5990  * copy_to_user_inatomic so we don't have to worry about page faulting under the
5991  * tree lock.
5992  */
5993 static int btrfs_opendir(struct inode *inode, struct file *file)
5994 {
5995         struct btrfs_file_private *private;
5996
5997         private = kzalloc(sizeof(struct btrfs_file_private), GFP_KERNEL);
5998         if (!private)
5999                 return -ENOMEM;
6000         private->filldir_buf = kzalloc(PAGE_SIZE, GFP_KERNEL);
6001         if (!private->filldir_buf) {
6002                 kfree(private);
6003                 return -ENOMEM;
6004         }
6005         file->private_data = private;
6006         return 0;
6007 }
6008
6009 struct dir_entry {
6010         u64 ino;
6011         u64 offset;
6012         unsigned type;
6013         int name_len;
6014 };
6015
6016 static int btrfs_filldir(void *addr, int entries, struct dir_context *ctx)
6017 {
6018         while (entries--) {
6019                 struct dir_entry *entry = addr;
6020                 char *name = (char *)(entry + 1);
6021
6022                 ctx->pos = get_unaligned(&entry->offset);
6023                 if (!dir_emit(ctx, name, get_unaligned(&entry->name_len),
6024                                          get_unaligned(&entry->ino),
6025                                          get_unaligned(&entry->type)))
6026                         return 1;
6027                 addr += sizeof(struct dir_entry) +
6028                         get_unaligned(&entry->name_len);
6029                 ctx->pos++;
6030         }
6031         return 0;
6032 }
6033
6034 static int btrfs_real_readdir(struct file *file, struct dir_context *ctx)
6035 {
6036         struct inode *inode = file_inode(file);
6037         struct btrfs_root *root = BTRFS_I(inode)->root;
6038         struct btrfs_file_private *private = file->private_data;
6039         struct btrfs_dir_item *di;
6040         struct btrfs_key key;
6041         struct btrfs_key found_key;
6042         struct btrfs_path *path;
6043         void *addr;
6044         struct list_head ins_list;
6045         struct list_head del_list;
6046         int ret;
6047         struct extent_buffer *leaf;
6048         int slot;
6049         char *name_ptr;
6050         int name_len;
6051         int entries = 0;
6052         int total_len = 0;
6053         bool put = false;
6054         struct btrfs_key location;
6055
6056         if (!dir_emit_dots(file, ctx))
6057                 return 0;
6058
6059         path = btrfs_alloc_path();
6060         if (!path)
6061                 return -ENOMEM;
6062
6063         addr = private->filldir_buf;
6064         path->reada = READA_FORWARD;
6065
6066         INIT_LIST_HEAD(&ins_list);
6067         INIT_LIST_HEAD(&del_list);
6068         put = btrfs_readdir_get_delayed_items(inode, &ins_list, &del_list);
6069
6070 again:
6071         key.type = BTRFS_DIR_INDEX_KEY;
6072         key.offset = ctx->pos;
6073         key.objectid = btrfs_ino(BTRFS_I(inode));
6074
6075         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
6076         if (ret < 0)
6077                 goto err;
6078
6079         while (1) {
6080                 struct dir_entry *entry;
6081
6082                 leaf = path->nodes[0];
6083                 slot = path->slots[0];
6084                 if (slot >= btrfs_header_nritems(leaf)) {
6085                         ret = btrfs_next_leaf(root, path);
6086                         if (ret < 0)
6087                                 goto err;
6088                         else if (ret > 0)
6089                                 break;
6090                         continue;
6091                 }
6092
6093                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
6094
6095                 if (found_key.objectid != key.objectid)
6096                         break;
6097                 if (found_key.type != BTRFS_DIR_INDEX_KEY)
6098                         break;
6099                 if (found_key.offset < ctx->pos)
6100                         goto next;
6101                 if (btrfs_should_delete_dir_index(&del_list, found_key.offset))
6102                         goto next;
6103                 di = btrfs_item_ptr(leaf, slot, struct btrfs_dir_item);
6104                 name_len = btrfs_dir_name_len(leaf, di);
6105                 if ((total_len + sizeof(struct dir_entry) + name_len) >=
6106                     PAGE_SIZE) {
6107                         btrfs_release_path(path);
6108                         ret = btrfs_filldir(private->filldir_buf, entries, ctx);
6109                         if (ret)
6110                                 goto nopos;
6111                         addr = private->filldir_buf;
6112                         entries = 0;
6113                         total_len = 0;
6114                         goto again;
6115                 }
6116
6117                 entry = addr;
6118                 put_unaligned(name_len, &entry->name_len);
6119                 name_ptr = (char *)(entry + 1);
6120                 read_extent_buffer(leaf, name_ptr, (unsigned long)(di + 1),
6121                                    name_len);
6122                 put_unaligned(fs_ftype_to_dtype(btrfs_dir_type(leaf, di)),
6123                                 &entry->type);
6124                 btrfs_dir_item_key_to_cpu(leaf, di, &location);
6125                 put_unaligned(location.objectid, &entry->ino);
6126                 put_unaligned(found_key.offset, &entry->offset);
6127                 entries++;
6128                 addr += sizeof(struct dir_entry) + name_len;
6129                 total_len += sizeof(struct dir_entry) + name_len;
6130 next:
6131                 path->slots[0]++;
6132         }
6133         btrfs_release_path(path);
6134
6135         ret = btrfs_filldir(private->filldir_buf, entries, ctx);
6136         if (ret)
6137                 goto nopos;
6138
6139         ret = btrfs_readdir_delayed_dir_index(ctx, &ins_list);
6140         if (ret)
6141                 goto nopos;
6142
6143         /*
6144          * Stop new entries from being returned after we return the last
6145          * entry.
6146          *
6147          * New directory entries are assigned a strictly increasing
6148          * offset.  This means that new entries created during readdir
6149          * are *guaranteed* to be seen in the future by that readdir.
6150          * This has broken buggy programs which operate on names as
6151          * they're returned by readdir.  Until we re-use freed offsets
6152          * we have this hack to stop new entries from being returned
6153          * under the assumption that they'll never reach this huge
6154          * offset.
6155          *
6156          * This is being careful not to overflow 32bit loff_t unless the
6157          * last entry requires it because doing so has broken 32bit apps
6158          * in the past.
6159          */
6160         if (ctx->pos >= INT_MAX)
6161                 ctx->pos = LLONG_MAX;
6162         else
6163                 ctx->pos = INT_MAX;
6164 nopos:
6165         ret = 0;
6166 err:
6167         if (put)
6168                 btrfs_readdir_put_delayed_items(inode, &ins_list, &del_list);
6169         btrfs_free_path(path);
6170         return ret;
6171 }
6172
6173 /*
6174  * This is somewhat expensive, updating the tree every time the
6175  * inode changes.  But, it is most likely to find the inode in cache.
6176  * FIXME, needs more benchmarking...there are no reasons other than performance
6177  * to keep or drop this code.
6178  */
6179 static int btrfs_dirty_inode(struct inode *inode)
6180 {
6181         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
6182         struct btrfs_root *root = BTRFS_I(inode)->root;
6183         struct btrfs_trans_handle *trans;
6184         int ret;
6185
6186         if (test_bit(BTRFS_INODE_DUMMY, &BTRFS_I(inode)->runtime_flags))
6187                 return 0;
6188
6189         trans = btrfs_join_transaction(root);
6190         if (IS_ERR(trans))
6191                 return PTR_ERR(trans);
6192
6193         ret = btrfs_update_inode(trans, root, inode);
6194         if (ret && ret == -ENOSPC) {
6195                 /* whoops, lets try again with the full transaction */
6196                 btrfs_end_transaction(trans);
6197                 trans = btrfs_start_transaction(root, 1);
6198                 if (IS_ERR(trans))
6199                         return PTR_ERR(trans);
6200
6201                 ret = btrfs_update_inode(trans, root, inode);
6202         }
6203         btrfs_end_transaction(trans);
6204         if (BTRFS_I(inode)->delayed_node)
6205                 btrfs_balance_delayed_items(fs_info);
6206
6207         return ret;
6208 }
6209
6210 /*
6211  * This is a copy of file_update_time.  We need this so we can return error on
6212  * ENOSPC for updating the inode in the case of file write and mmap writes.
6213  */
6214 static int btrfs_update_time(struct inode *inode, struct timespec64 *now,
6215                              int flags)
6216 {
6217         struct btrfs_root *root = BTRFS_I(inode)->root;
6218         bool dirty = flags & ~S_VERSION;
6219
6220         if (btrfs_root_readonly(root))
6221                 return -EROFS;
6222
6223         if (flags & S_VERSION)
6224                 dirty |= inode_maybe_inc_iversion(inode, dirty);
6225         if (flags & S_CTIME)
6226                 inode->i_ctime = *now;
6227         if (flags & S_MTIME)
6228                 inode->i_mtime = *now;
6229         if (flags & S_ATIME)
6230                 inode->i_atime = *now;
6231         return dirty ? btrfs_dirty_inode(inode) : 0;
6232 }
6233
6234 /*
6235  * find the highest existing sequence number in a directory
6236  * and then set the in-memory index_cnt variable to reflect
6237  * free sequence numbers
6238  */
6239 static int btrfs_set_inode_index_count(struct btrfs_inode *inode)
6240 {
6241         struct btrfs_root *root = inode->root;
6242         struct btrfs_key key, found_key;
6243         struct btrfs_path *path;
6244         struct extent_buffer *leaf;
6245         int ret;
6246
6247         key.objectid = btrfs_ino(inode);
6248         key.type = BTRFS_DIR_INDEX_KEY;
6249         key.offset = (u64)-1;
6250
6251         path = btrfs_alloc_path();
6252         if (!path)
6253                 return -ENOMEM;
6254
6255         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
6256         if (ret < 0)
6257                 goto out;
6258         /* FIXME: we should be able to handle this */
6259         if (ret == 0)
6260                 goto out;
6261         ret = 0;
6262
6263         /*
6264          * MAGIC NUMBER EXPLANATION:
6265          * since we search a directory based on f_pos we have to start at 2
6266          * since '.' and '..' have f_pos of 0 and 1 respectively, so everybody
6267          * else has to start at 2
6268          */
6269         if (path->slots[0] == 0) {
6270                 inode->index_cnt = 2;
6271                 goto out;
6272         }
6273
6274         path->slots[0]--;
6275
6276         leaf = path->nodes[0];
6277         btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
6278
6279         if (found_key.objectid != btrfs_ino(inode) ||
6280             found_key.type != BTRFS_DIR_INDEX_KEY) {
6281                 inode->index_cnt = 2;
6282                 goto out;
6283         }
6284
6285         inode->index_cnt = found_key.offset + 1;
6286 out:
6287         btrfs_free_path(path);
6288         return ret;
6289 }
6290
6291 /*
6292  * helper to find a free sequence number in a given directory.  This current
6293  * code is very simple, later versions will do smarter things in the btree
6294  */
6295 int btrfs_set_inode_index(struct btrfs_inode *dir, u64 *index)
6296 {
6297         int ret = 0;
6298
6299         if (dir->index_cnt == (u64)-1) {
6300                 ret = btrfs_inode_delayed_dir_index_count(dir);
6301                 if (ret) {
6302                         ret = btrfs_set_inode_index_count(dir);
6303                         if (ret)
6304                                 return ret;
6305                 }
6306         }
6307
6308         *index = dir->index_cnt;
6309         dir->index_cnt++;
6310
6311         return ret;
6312 }
6313
6314 static int btrfs_insert_inode_locked(struct inode *inode)
6315 {
6316         struct btrfs_iget_args args;
6317         args.location = &BTRFS_I(inode)->location;
6318         args.root = BTRFS_I(inode)->root;
6319
6320         return insert_inode_locked4(inode,
6321                    btrfs_inode_hash(inode->i_ino, BTRFS_I(inode)->root),
6322                    btrfs_find_actor, &args);
6323 }
6324
6325 /*
6326  * Inherit flags from the parent inode.
6327  *
6328  * Currently only the compression flags and the cow flags are inherited.
6329  */
6330 static void btrfs_inherit_iflags(struct inode *inode, struct inode *dir)
6331 {
6332         unsigned int flags;
6333
6334         if (!dir)
6335                 return;
6336
6337         flags = BTRFS_I(dir)->flags;
6338
6339         if (flags & BTRFS_INODE_NOCOMPRESS) {
6340                 BTRFS_I(inode)->flags &= ~BTRFS_INODE_COMPRESS;
6341                 BTRFS_I(inode)->flags |= BTRFS_INODE_NOCOMPRESS;
6342         } else if (flags & BTRFS_INODE_COMPRESS) {
6343                 BTRFS_I(inode)->flags &= ~BTRFS_INODE_NOCOMPRESS;
6344                 BTRFS_I(inode)->flags |= BTRFS_INODE_COMPRESS;
6345         }
6346
6347         if (flags & BTRFS_INODE_NODATACOW) {
6348                 BTRFS_I(inode)->flags |= BTRFS_INODE_NODATACOW;
6349                 if (S_ISREG(inode->i_mode))
6350                         BTRFS_I(inode)->flags |= BTRFS_INODE_NODATASUM;
6351         }
6352
6353         btrfs_sync_inode_flags_to_i_flags(inode);
6354 }
6355
6356 static struct inode *btrfs_new_inode(struct btrfs_trans_handle *trans,
6357                                      struct btrfs_root *root,
6358                                      struct inode *dir,
6359                                      const char *name, int name_len,
6360                                      u64 ref_objectid, u64 objectid,
6361                                      umode_t mode, u64 *index)
6362 {
6363         struct btrfs_fs_info *fs_info = root->fs_info;
6364         struct inode *inode;
6365         struct btrfs_inode_item *inode_item;
6366         struct btrfs_key *location;
6367         struct btrfs_path *path;
6368         struct btrfs_inode_ref *ref;
6369         struct btrfs_key key[2];
6370         u32 sizes[2];
6371         int nitems = name ? 2 : 1;
6372         unsigned long ptr;
6373         unsigned int nofs_flag;
6374         int ret;
6375
6376         path = btrfs_alloc_path();
6377         if (!path)
6378                 return ERR_PTR(-ENOMEM);
6379
6380         nofs_flag = memalloc_nofs_save();
6381         inode = new_inode(fs_info->sb);
6382         memalloc_nofs_restore(nofs_flag);
6383         if (!inode) {
6384                 btrfs_free_path(path);
6385                 return ERR_PTR(-ENOMEM);
6386         }
6387
6388         /*
6389          * O_TMPFILE, set link count to 0, so that after this point,
6390          * we fill in an inode item with the correct link count.
6391          */
6392         if (!name)
6393                 set_nlink(inode, 0);
6394
6395         /*
6396          * we have to initialize this early, so we can reclaim the inode
6397          * number if we fail afterwards in this function.
6398          */
6399         inode->i_ino = objectid;
6400
6401         if (dir && name) {
6402                 trace_btrfs_inode_request(dir);
6403
6404                 ret = btrfs_set_inode_index(BTRFS_I(dir), index);
6405                 if (ret) {
6406                         btrfs_free_path(path);
6407                         iput(inode);
6408                         return ERR_PTR(ret);
6409                 }
6410         } else if (dir) {
6411                 *index = 0;
6412         }
6413         /*
6414          * index_cnt is ignored for everything but a dir,
6415          * btrfs_set_inode_index_count has an explanation for the magic
6416          * number
6417          */
6418         BTRFS_I(inode)->index_cnt = 2;
6419         BTRFS_I(inode)->dir_index = *index;
6420         BTRFS_I(inode)->root = root;
6421         BTRFS_I(inode)->generation = trans->transid;
6422         inode->i_generation = BTRFS_I(inode)->generation;
6423
6424         /*
6425          * We could have gotten an inode number from somebody who was fsynced
6426          * and then removed in this same transaction, so let's just set full
6427          * sync since it will be a full sync anyway and this will blow away the
6428          * old info in the log.
6429          */
6430         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &BTRFS_I(inode)->runtime_flags);
6431
6432         key[0].objectid = objectid;
6433         key[0].type = BTRFS_INODE_ITEM_KEY;
6434         key[0].offset = 0;
6435
6436         sizes[0] = sizeof(struct btrfs_inode_item);
6437
6438         if (name) {
6439                 /*
6440                  * Start new inodes with an inode_ref. This is slightly more
6441                  * efficient for small numbers of hard links since they will
6442                  * be packed into one item. Extended refs will kick in if we
6443                  * add more hard links than can fit in the ref item.
6444                  */
6445                 key[1].objectid = objectid;
6446                 key[1].type = BTRFS_INODE_REF_KEY;
6447                 key[1].offset = ref_objectid;
6448
6449                 sizes[1] = name_len + sizeof(*ref);
6450         }
6451
6452         location = &BTRFS_I(inode)->location;
6453         location->objectid = objectid;
6454         location->offset = 0;
6455         location->type = BTRFS_INODE_ITEM_KEY;
6456
6457         ret = btrfs_insert_inode_locked(inode);
6458         if (ret < 0) {
6459                 iput(inode);
6460                 goto fail;
6461         }
6462
6463         path->leave_spinning = 1;
6464         ret = btrfs_insert_empty_items(trans, root, path, key, sizes, nitems);
6465         if (ret != 0)
6466                 goto fail_unlock;
6467
6468         inode_init_owner(inode, dir, mode);
6469         inode_set_bytes(inode, 0);
6470
6471         inode->i_mtime = current_time(inode);
6472         inode->i_atime = inode->i_mtime;
6473         inode->i_ctime = inode->i_mtime;
6474         BTRFS_I(inode)->i_otime = inode->i_mtime;
6475
6476         inode_item = btrfs_item_ptr(path->nodes[0], path->slots[0],
6477                                   struct btrfs_inode_item);
6478         memzero_extent_buffer(path->nodes[0], (unsigned long)inode_item,
6479                              sizeof(*inode_item));
6480         fill_inode_item(trans, path->nodes[0], inode_item, inode);
6481
6482         if (name) {
6483                 ref = btrfs_item_ptr(path->nodes[0], path->slots[0] + 1,
6484                                      struct btrfs_inode_ref);
6485                 btrfs_set_inode_ref_name_len(path->nodes[0], ref, name_len);
6486                 btrfs_set_inode_ref_index(path->nodes[0], ref, *index);
6487                 ptr = (unsigned long)(ref + 1);
6488                 write_extent_buffer(path->nodes[0], name, ptr, name_len);
6489         }
6490
6491         btrfs_mark_buffer_dirty(path->nodes[0]);
6492         btrfs_free_path(path);
6493
6494         btrfs_inherit_iflags(inode, dir);
6495
6496         if (S_ISREG(mode)) {
6497                 if (btrfs_test_opt(fs_info, NODATASUM))
6498                         BTRFS_I(inode)->flags |= BTRFS_INODE_NODATASUM;
6499                 if (btrfs_test_opt(fs_info, NODATACOW))
6500                         BTRFS_I(inode)->flags |= BTRFS_INODE_NODATACOW |
6501                                 BTRFS_INODE_NODATASUM;
6502         }
6503
6504         inode_tree_add(inode);
6505
6506         trace_btrfs_inode_new(inode);
6507         btrfs_set_inode_last_trans(trans, inode);
6508
6509         btrfs_update_root_times(trans, root);
6510
6511         ret = btrfs_inode_inherit_props(trans, inode, dir);
6512         if (ret)
6513                 btrfs_err(fs_info,
6514                           "error inheriting props for ino %llu (root %llu): %d",
6515                         btrfs_ino(BTRFS_I(inode)), root->root_key.objectid, ret);
6516
6517         return inode;
6518
6519 fail_unlock:
6520         discard_new_inode(inode);
6521 fail:
6522         if (dir && name)
6523                 BTRFS_I(dir)->index_cnt--;
6524         btrfs_free_path(path);
6525         return ERR_PTR(ret);
6526 }
6527
6528 /*
6529  * utility function to add 'inode' into 'parent_inode' with
6530  * a give name and a given sequence number.
6531  * if 'add_backref' is true, also insert a backref from the
6532  * inode to the parent directory.
6533  */
6534 int btrfs_add_link(struct btrfs_trans_handle *trans,
6535                    struct btrfs_inode *parent_inode, struct btrfs_inode *inode,
6536                    const char *name, int name_len, int add_backref, u64 index)
6537 {
6538         int ret = 0;
6539         struct btrfs_key key;
6540         struct btrfs_root *root = parent_inode->root;
6541         u64 ino = btrfs_ino(inode);
6542         u64 parent_ino = btrfs_ino(parent_inode);
6543
6544         if (unlikely(ino == BTRFS_FIRST_FREE_OBJECTID)) {
6545                 memcpy(&key, &inode->root->root_key, sizeof(key));
6546         } else {
6547                 key.objectid = ino;
6548                 key.type = BTRFS_INODE_ITEM_KEY;
6549                 key.offset = 0;
6550         }
6551
6552         if (unlikely(ino == BTRFS_FIRST_FREE_OBJECTID)) {
6553                 ret = btrfs_add_root_ref(trans, key.objectid,
6554                                          root->root_key.objectid, parent_ino,
6555                                          index, name, name_len);
6556         } else if (add_backref) {
6557                 ret = btrfs_insert_inode_ref(trans, root, name, name_len, ino,
6558                                              parent_ino, index);
6559         }
6560
6561         /* Nothing to clean up yet */
6562         if (ret)
6563                 return ret;
6564
6565         ret = btrfs_insert_dir_item(trans, name, name_len, parent_inode, &key,
6566                                     btrfs_inode_type(&inode->vfs_inode), index);
6567         if (ret == -EEXIST || ret == -EOVERFLOW)
6568                 goto fail_dir_item;
6569         else if (ret) {
6570                 btrfs_abort_transaction(trans, ret);
6571                 return ret;
6572         }
6573
6574         btrfs_i_size_write(parent_inode, parent_inode->vfs_inode.i_size +
6575                            name_len * 2);
6576         inode_inc_iversion(&parent_inode->vfs_inode);
6577         /*
6578          * If we are replaying a log tree, we do not want to update the mtime
6579          * and ctime of the parent directory with the current time, since the
6580          * log replay procedure is responsible for setting them to their correct
6581          * values (the ones it had when the fsync was done).
6582          */
6583         if (!test_bit(BTRFS_FS_LOG_RECOVERING, &root->fs_info->flags)) {
6584                 struct timespec64 now = current_time(&parent_inode->vfs_inode);
6585
6586                 parent_inode->vfs_inode.i_mtime = now;
6587                 parent_inode->vfs_inode.i_ctime = now;
6588         }
6589         ret = btrfs_update_inode(trans, root, &parent_inode->vfs_inode);
6590         if (ret)
6591                 btrfs_abort_transaction(trans, ret);
6592         return ret;
6593
6594 fail_dir_item:
6595         if (unlikely(ino == BTRFS_FIRST_FREE_OBJECTID)) {
6596                 u64 local_index;
6597                 int err;
6598                 err = btrfs_del_root_ref(trans, key.objectid,
6599                                          root->root_key.objectid, parent_ino,
6600                                          &local_index, name, name_len);
6601                 if (err)
6602                         btrfs_abort_transaction(trans, err);
6603         } else if (add_backref) {
6604                 u64 local_index;
6605                 int err;
6606
6607                 err = btrfs_del_inode_ref(trans, root, name, name_len,
6608                                           ino, parent_ino, &local_index);
6609                 if (err)
6610                         btrfs_abort_transaction(trans, err);
6611         }
6612
6613         /* Return the original error code */
6614         return ret;
6615 }
6616
6617 static int btrfs_add_nondir(struct btrfs_trans_handle *trans,
6618                             struct btrfs_inode *dir, struct dentry *dentry,
6619                             struct btrfs_inode *inode, int backref, u64 index)
6620 {
6621         int err = btrfs_add_link(trans, dir, inode,
6622                                  dentry->d_name.name, dentry->d_name.len,
6623                                  backref, index);
6624         if (err > 0)
6625                 err = -EEXIST;
6626         return err;
6627 }
6628
6629 static int btrfs_mknod(struct inode *dir, struct dentry *dentry,
6630                         umode_t mode, dev_t rdev)
6631 {
6632         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
6633         struct btrfs_trans_handle *trans;
6634         struct btrfs_root *root = BTRFS_I(dir)->root;
6635         struct inode *inode = NULL;
6636         int err;
6637         u64 objectid;
6638         u64 index = 0;
6639
6640         /*
6641          * 2 for inode item and ref
6642          * 2 for dir items
6643          * 1 for xattr if selinux is on
6644          */
6645         trans = btrfs_start_transaction(root, 5);
6646         if (IS_ERR(trans))
6647                 return PTR_ERR(trans);
6648
6649         err = btrfs_find_free_ino(root, &objectid);
6650         if (err)
6651                 goto out_unlock;
6652
6653         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
6654                         dentry->d_name.len, btrfs_ino(BTRFS_I(dir)), objectid,
6655                         mode, &index);
6656         if (IS_ERR(inode)) {
6657                 err = PTR_ERR(inode);
6658                 inode = NULL;
6659                 goto out_unlock;
6660         }
6661
6662         /*
6663         * If the active LSM wants to access the inode during
6664         * d_instantiate it needs these. Smack checks to see
6665         * if the filesystem supports xattrs by looking at the
6666         * ops vector.
6667         */
6668         inode->i_op = &btrfs_special_inode_operations;
6669         init_special_inode(inode, inode->i_mode, rdev);
6670
6671         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
6672         if (err)
6673                 goto out_unlock;
6674
6675         err = btrfs_add_nondir(trans, BTRFS_I(dir), dentry, BTRFS_I(inode),
6676                         0, index);
6677         if (err)
6678                 goto out_unlock;
6679
6680         btrfs_update_inode(trans, root, inode);
6681         d_instantiate_new(dentry, inode);
6682
6683 out_unlock:
6684         btrfs_end_transaction(trans);
6685         btrfs_btree_balance_dirty(fs_info);
6686         if (err && inode) {
6687                 inode_dec_link_count(inode);
6688                 discard_new_inode(inode);
6689         }
6690         return err;
6691 }
6692
6693 static int btrfs_create(struct inode *dir, struct dentry *dentry,
6694                         umode_t mode, bool excl)
6695 {
6696         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
6697         struct btrfs_trans_handle *trans;
6698         struct btrfs_root *root = BTRFS_I(dir)->root;
6699         struct inode *inode = NULL;
6700         int err;
6701         u64 objectid;
6702         u64 index = 0;
6703
6704         /*
6705          * 2 for inode item and ref
6706          * 2 for dir items
6707          * 1 for xattr if selinux is on
6708          */
6709         trans = btrfs_start_transaction(root, 5);
6710         if (IS_ERR(trans))
6711                 return PTR_ERR(trans);
6712
6713         err = btrfs_find_free_ino(root, &objectid);
6714         if (err)
6715                 goto out_unlock;
6716
6717         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
6718                         dentry->d_name.len, btrfs_ino(BTRFS_I(dir)), objectid,
6719                         mode, &index);
6720         if (IS_ERR(inode)) {
6721                 err = PTR_ERR(inode);
6722                 inode = NULL;
6723                 goto out_unlock;
6724         }
6725         /*
6726         * If the active LSM wants to access the inode during
6727         * d_instantiate it needs these. Smack checks to see
6728         * if the filesystem supports xattrs by looking at the
6729         * ops vector.
6730         */
6731         inode->i_fop = &btrfs_file_operations;
6732         inode->i_op = &btrfs_file_inode_operations;
6733         inode->i_mapping->a_ops = &btrfs_aops;
6734
6735         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
6736         if (err)
6737                 goto out_unlock;
6738
6739         err = btrfs_update_inode(trans, root, inode);
6740         if (err)
6741                 goto out_unlock;
6742
6743         err = btrfs_add_nondir(trans, BTRFS_I(dir), dentry, BTRFS_I(inode),
6744                         0, index);
6745         if (err)
6746                 goto out_unlock;
6747
6748         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
6749         d_instantiate_new(dentry, inode);
6750
6751 out_unlock:
6752         btrfs_end_transaction(trans);
6753         if (err && inode) {
6754                 inode_dec_link_count(inode);
6755                 discard_new_inode(inode);
6756         }
6757         btrfs_btree_balance_dirty(fs_info);
6758         return err;
6759 }
6760
6761 static int btrfs_link(struct dentry *old_dentry, struct inode *dir,
6762                       struct dentry *dentry)
6763 {
6764         struct btrfs_trans_handle *trans = NULL;
6765         struct btrfs_root *root = BTRFS_I(dir)->root;
6766         struct inode *inode = d_inode(old_dentry);
6767         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
6768         u64 index;
6769         int err;
6770         int drop_inode = 0;
6771
6772         /* do not allow sys_link's with other subvols of the same device */
6773         if (root->root_key.objectid != BTRFS_I(inode)->root->root_key.objectid)
6774                 return -EXDEV;
6775
6776         if (inode->i_nlink >= BTRFS_LINK_MAX)
6777                 return -EMLINK;
6778
6779         err = btrfs_set_inode_index(BTRFS_I(dir), &index);
6780         if (err)
6781                 goto fail;
6782
6783         /*
6784          * 2 items for inode and inode ref
6785          * 2 items for dir items
6786          * 1 item for parent inode
6787          * 1 item for orphan item deletion if O_TMPFILE
6788          */
6789         trans = btrfs_start_transaction(root, inode->i_nlink ? 5 : 6);
6790         if (IS_ERR(trans)) {
6791                 err = PTR_ERR(trans);
6792                 trans = NULL;
6793                 goto fail;
6794         }
6795
6796         /* There are several dir indexes for this inode, clear the cache. */
6797         BTRFS_I(inode)->dir_index = 0ULL;
6798         inc_nlink(inode);
6799         inode_inc_iversion(inode);
6800         inode->i_ctime = current_time(inode);
6801         ihold(inode);
6802         set_bit(BTRFS_INODE_COPY_EVERYTHING, &BTRFS_I(inode)->runtime_flags);
6803
6804         err = btrfs_add_nondir(trans, BTRFS_I(dir), dentry, BTRFS_I(inode),
6805                         1, index);
6806
6807         if (err) {
6808                 drop_inode = 1;
6809         } else {
6810                 struct dentry *parent = dentry->d_parent;
6811                 int ret;
6812
6813                 err = btrfs_update_inode(trans, root, inode);
6814                 if (err)
6815                         goto fail;
6816                 if (inode->i_nlink == 1) {
6817                         /*
6818                          * If new hard link count is 1, it's a file created
6819                          * with open(2) O_TMPFILE flag.
6820                          */
6821                         err = btrfs_orphan_del(trans, BTRFS_I(inode));
6822                         if (err)
6823                                 goto fail;
6824                 }
6825                 d_instantiate(dentry, inode);
6826                 ret = btrfs_log_new_name(trans, BTRFS_I(inode), NULL, parent,
6827                                          true, NULL);
6828                 if (ret == BTRFS_NEED_TRANS_COMMIT) {
6829                         err = btrfs_commit_transaction(trans);
6830                         trans = NULL;
6831                 }
6832         }
6833
6834 fail:
6835         if (trans)
6836                 btrfs_end_transaction(trans);
6837         if (drop_inode) {
6838                 inode_dec_link_count(inode);
6839                 iput(inode);
6840         }
6841         btrfs_btree_balance_dirty(fs_info);
6842         return err;
6843 }
6844
6845 static int btrfs_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode)
6846 {
6847         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
6848         struct inode *inode = NULL;
6849         struct btrfs_trans_handle *trans;
6850         struct btrfs_root *root = BTRFS_I(dir)->root;
6851         int err = 0;
6852         u64 objectid = 0;
6853         u64 index = 0;
6854
6855         /*
6856          * 2 items for inode and ref
6857          * 2 items for dir items
6858          * 1 for xattr if selinux is on
6859          */
6860         trans = btrfs_start_transaction(root, 5);
6861         if (IS_ERR(trans))
6862                 return PTR_ERR(trans);
6863
6864         err = btrfs_find_free_ino(root, &objectid);
6865         if (err)
6866                 goto out_fail;
6867
6868         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
6869                         dentry->d_name.len, btrfs_ino(BTRFS_I(dir)), objectid,
6870                         S_IFDIR | mode, &index);
6871         if (IS_ERR(inode)) {
6872                 err = PTR_ERR(inode);
6873                 inode = NULL;
6874                 goto out_fail;
6875         }
6876
6877         /* these must be set before we unlock the inode */
6878         inode->i_op = &btrfs_dir_inode_operations;
6879         inode->i_fop = &btrfs_dir_file_operations;
6880
6881         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
6882         if (err)
6883                 goto out_fail;
6884
6885         btrfs_i_size_write(BTRFS_I(inode), 0);
6886         err = btrfs_update_inode(trans, root, inode);
6887         if (err)
6888                 goto out_fail;
6889
6890         err = btrfs_add_link(trans, BTRFS_I(dir), BTRFS_I(inode),
6891                         dentry->d_name.name,
6892                         dentry->d_name.len, 0, index);
6893         if (err)
6894                 goto out_fail;
6895
6896         d_instantiate_new(dentry, inode);
6897
6898 out_fail:
6899         btrfs_end_transaction(trans);
6900         if (err && inode) {
6901                 inode_dec_link_count(inode);
6902                 discard_new_inode(inode);
6903         }
6904         btrfs_btree_balance_dirty(fs_info);
6905         return err;
6906 }
6907
6908 static noinline int uncompress_inline(struct btrfs_path *path,
6909                                       struct page *page,
6910                                       size_t pg_offset, u64 extent_offset,
6911                                       struct btrfs_file_extent_item *item)
6912 {
6913         int ret;
6914         struct extent_buffer *leaf = path->nodes[0];
6915         char *tmp;
6916         size_t max_size;
6917         unsigned long inline_size;
6918         unsigned long ptr;
6919         int compress_type;
6920
6921         WARN_ON(pg_offset != 0);
6922         compress_type = btrfs_file_extent_compression(leaf, item);
6923         max_size = btrfs_file_extent_ram_bytes(leaf, item);
6924         inline_size = btrfs_file_extent_inline_item_len(leaf,
6925                                         btrfs_item_nr(path->slots[0]));
6926         tmp = kmalloc(inline_size, GFP_NOFS);
6927         if (!tmp)
6928                 return -ENOMEM;
6929         ptr = btrfs_file_extent_inline_start(item);
6930
6931         read_extent_buffer(leaf, tmp, ptr, inline_size);
6932
6933         max_size = min_t(unsigned long, PAGE_SIZE, max_size);
6934         ret = btrfs_decompress(compress_type, tmp, page,
6935                                extent_offset, inline_size, max_size);
6936
6937         /*
6938          * decompression code contains a memset to fill in any space between the end
6939          * of the uncompressed data and the end of max_size in case the decompressed
6940          * data ends up shorter than ram_bytes.  That doesn't cover the hole between
6941          * the end of an inline extent and the beginning of the next block, so we
6942          * cover that region here.
6943          */
6944
6945         if (max_size + pg_offset < PAGE_SIZE) {
6946                 char *map = kmap(page);
6947                 memset(map + pg_offset + max_size, 0, PAGE_SIZE - max_size - pg_offset);
6948                 kunmap(page);
6949         }
6950         kfree(tmp);
6951         return ret;
6952 }
6953
6954 /*
6955  * a bit scary, this does extent mapping from logical file offset to the disk.
6956  * the ugly parts come from merging extents from the disk with the in-ram
6957  * representation.  This gets more complex because of the data=ordered code,
6958  * where the in-ram extents might be locked pending data=ordered completion.
6959  *
6960  * This also copies inline extents directly into the page.
6961  */
6962 struct extent_map *btrfs_get_extent(struct btrfs_inode *inode,
6963                                     struct page *page,
6964                                     size_t pg_offset, u64 start, u64 len,
6965                                     int create)
6966 {
6967         struct btrfs_fs_info *fs_info = inode->root->fs_info;
6968         int ret;
6969         int err = 0;
6970         u64 extent_start = 0;
6971         u64 extent_end = 0;
6972         u64 objectid = btrfs_ino(inode);
6973         int extent_type = -1;
6974         struct btrfs_path *path = NULL;
6975         struct btrfs_root *root = inode->root;
6976         struct btrfs_file_extent_item *item;
6977         struct extent_buffer *leaf;
6978         struct btrfs_key found_key;
6979         struct extent_map *em = NULL;
6980         struct extent_map_tree *em_tree = &inode->extent_tree;
6981         struct extent_io_tree *io_tree = &inode->io_tree;
6982         const bool new_inline = !page || create;
6983
6984         read_lock(&em_tree->lock);
6985         em = lookup_extent_mapping(em_tree, start, len);
6986         read_unlock(&em_tree->lock);
6987
6988         if (em) {
6989                 if (em->start > start || em->start + em->len <= start)
6990                         free_extent_map(em);
6991                 else if (em->block_start == EXTENT_MAP_INLINE && page)
6992                         free_extent_map(em);
6993                 else
6994                         goto out;
6995         }
6996         em = alloc_extent_map();
6997         if (!em) {
6998                 err = -ENOMEM;
6999                 goto out;
7000         }
7001         em->start = EXTENT_MAP_HOLE;
7002         em->orig_start = EXTENT_MAP_HOLE;
7003         em->len = (u64)-1;
7004         em->block_len = (u64)-1;
7005
7006         path = btrfs_alloc_path();
7007         if (!path) {
7008                 err = -ENOMEM;
7009                 goto out;
7010         }
7011
7012         /* Chances are we'll be called again, so go ahead and do readahead */
7013         path->reada = READA_FORWARD;
7014
7015         /*
7016          * Unless we're going to uncompress the inline extent, no sleep would
7017          * happen.
7018          */
7019         path->leave_spinning = 1;
7020
7021         ret = btrfs_lookup_file_extent(NULL, root, path, objectid, start, 0);
7022         if (ret < 0) {
7023                 err = ret;
7024                 goto out;
7025         } else if (ret > 0) {
7026                 if (path->slots[0] == 0)
7027                         goto not_found;
7028                 path->slots[0]--;
7029         }
7030
7031         leaf = path->nodes[0];
7032         item = btrfs_item_ptr(leaf, path->slots[0],
7033                               struct btrfs_file_extent_item);
7034         btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
7035         if (found_key.objectid != objectid ||
7036             found_key.type != BTRFS_EXTENT_DATA_KEY) {
7037                 /*
7038                  * If we backup past the first extent we want to move forward
7039                  * and see if there is an extent in front of us, otherwise we'll
7040                  * say there is a hole for our whole search range which can
7041                  * cause problems.
7042                  */
7043                 extent_end = start;
7044                 goto next;
7045         }
7046
7047         extent_type = btrfs_file_extent_type(leaf, item);
7048         extent_start = found_key.offset;
7049         if (extent_type == BTRFS_FILE_EXTENT_REG ||
7050             extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
7051                 /* Only regular file could have regular/prealloc extent */
7052                 if (!S_ISREG(inode->vfs_inode.i_mode)) {
7053                         ret = -EUCLEAN;
7054                         btrfs_crit(fs_info,
7055                 "regular/prealloc extent found for non-regular inode %llu",
7056                                    btrfs_ino(inode));
7057                         goto out;
7058                 }
7059                 extent_end = extent_start +
7060                        btrfs_file_extent_num_bytes(leaf, item);
7061
7062                 trace_btrfs_get_extent_show_fi_regular(inode, leaf, item,
7063                                                        extent_start);
7064         } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
7065                 size_t size;
7066
7067                 size = btrfs_file_extent_ram_bytes(leaf, item);
7068                 extent_end = ALIGN(extent_start + size,
7069                                    fs_info->sectorsize);
7070
7071                 trace_btrfs_get_extent_show_fi_inline(inode, leaf, item,
7072                                                       path->slots[0],
7073                                                       extent_start);
7074         }
7075 next:
7076         if (start >= extent_end) {
7077                 path->slots[0]++;
7078                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
7079                         ret = btrfs_next_leaf(root, path);
7080                         if (ret < 0) {
7081                                 err = ret;
7082                                 goto out;
7083                         } else if (ret > 0) {
7084                                 goto not_found;
7085                         }
7086                         leaf = path->nodes[0];
7087                 }
7088                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
7089                 if (found_key.objectid != objectid ||
7090                     found_key.type != BTRFS_EXTENT_DATA_KEY)
7091                         goto not_found;
7092                 if (start + len <= found_key.offset)
7093                         goto not_found;
7094                 if (start > found_key.offset)
7095                         goto next;
7096
7097                 /* New extent overlaps with existing one */
7098                 em->start = start;
7099                 em->orig_start = start;
7100                 em->len = found_key.offset - start;
7101                 em->block_start = EXTENT_MAP_HOLE;
7102                 goto insert;
7103         }
7104
7105         btrfs_extent_item_to_extent_map(inode, path, item,
7106                         new_inline, em);
7107
7108         if (extent_type == BTRFS_FILE_EXTENT_REG ||
7109             extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
7110                 goto insert;
7111         } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
7112                 unsigned long ptr;
7113                 char *map;
7114                 size_t size;
7115                 size_t extent_offset;
7116                 size_t copy_size;
7117
7118                 if (new_inline)
7119                         goto out;
7120
7121                 size = btrfs_file_extent_ram_bytes(leaf, item);
7122                 extent_offset = page_offset(page) + pg_offset - extent_start;
7123                 copy_size = min_t(u64, PAGE_SIZE - pg_offset,
7124                                   size - extent_offset);
7125                 em->start = extent_start + extent_offset;
7126                 em->len = ALIGN(copy_size, fs_info->sectorsize);
7127                 em->orig_block_len = em->len;
7128                 em->orig_start = em->start;
7129                 ptr = btrfs_file_extent_inline_start(item) + extent_offset;
7130
7131                 btrfs_set_path_blocking(path);
7132                 if (!PageUptodate(page)) {
7133                         if (btrfs_file_extent_compression(leaf, item) !=
7134                             BTRFS_COMPRESS_NONE) {
7135                                 ret = uncompress_inline(path, page, pg_offset,
7136                                                         extent_offset, item);
7137                                 if (ret) {
7138                                         err = ret;
7139                                         goto out;
7140                                 }
7141                         } else {
7142                                 map = kmap(page);
7143                                 read_extent_buffer(leaf, map + pg_offset, ptr,
7144                                                    copy_size);
7145                                 if (pg_offset + copy_size < PAGE_SIZE) {
7146                                         memset(map + pg_offset + copy_size, 0,
7147                                                PAGE_SIZE - pg_offset -
7148                                                copy_size);
7149                                 }
7150                                 kunmap(page);
7151                         }
7152                         flush_dcache_page(page);
7153                 }
7154                 set_extent_uptodate(io_tree, em->start,
7155                                     extent_map_end(em) - 1, NULL, GFP_NOFS);
7156                 goto insert;
7157         }
7158 not_found:
7159         em->start = start;
7160         em->orig_start = start;
7161         em->len = len;
7162         em->block_start = EXTENT_MAP_HOLE;
7163 insert:
7164         btrfs_release_path(path);
7165         if (em->start > start || extent_map_end(em) <= start) {
7166                 btrfs_err(fs_info,
7167                           "bad extent! em: [%llu %llu] passed [%llu %llu]",
7168                           em->start, em->len, start, len);
7169                 err = -EIO;
7170                 goto out;
7171         }
7172
7173         err = 0;
7174         write_lock(&em_tree->lock);
7175         err = btrfs_add_extent_mapping(fs_info, em_tree, &em, start, len);
7176         write_unlock(&em_tree->lock);
7177 out:
7178         btrfs_free_path(path);
7179
7180         trace_btrfs_get_extent(root, inode, em);
7181
7182         if (err) {
7183                 free_extent_map(em);
7184                 return ERR_PTR(err);
7185         }
7186         BUG_ON(!em); /* Error is always set */
7187         return em;
7188 }
7189
7190 struct extent_map *btrfs_get_extent_fiemap(struct btrfs_inode *inode,
7191                                            u64 start, u64 len)
7192 {
7193         struct extent_map *em;
7194         struct extent_map *hole_em = NULL;
7195         u64 delalloc_start = start;
7196         u64 end;
7197         u64 delalloc_len;
7198         u64 delalloc_end;
7199         int err = 0;
7200
7201         em = btrfs_get_extent(inode, NULL, 0, start, len, 0);
7202         if (IS_ERR(em))
7203                 return em;
7204         /*
7205          * If our em maps to:
7206          * - a hole or
7207          * - a pre-alloc extent,
7208          * there might actually be delalloc bytes behind it.
7209          */
7210         if (em->block_start != EXTENT_MAP_HOLE &&
7211             !test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
7212                 return em;
7213         else
7214                 hole_em = em;
7215
7216         /* check to see if we've wrapped (len == -1 or similar) */
7217         end = start + len;
7218         if (end < start)
7219                 end = (u64)-1;
7220         else
7221                 end -= 1;
7222
7223         em = NULL;
7224
7225         /* ok, we didn't find anything, lets look for delalloc */
7226         delalloc_len = count_range_bits(&inode->io_tree, &delalloc_start,
7227                                  end, len, EXTENT_DELALLOC, 1);
7228         delalloc_end = delalloc_start + delalloc_len;
7229         if (delalloc_end < delalloc_start)
7230                 delalloc_end = (u64)-1;
7231
7232         /*
7233          * We didn't find anything useful, return the original results from
7234          * get_extent()
7235          */
7236         if (delalloc_start > end || delalloc_end <= start) {
7237                 em = hole_em;
7238                 hole_em = NULL;
7239                 goto out;
7240         }
7241
7242         /*
7243          * Adjust the delalloc_start to make sure it doesn't go backwards from
7244          * the start they passed in
7245          */
7246         delalloc_start = max(start, delalloc_start);
7247         delalloc_len = delalloc_end - delalloc_start;
7248
7249         if (delalloc_len > 0) {
7250                 u64 hole_start;
7251                 u64 hole_len;
7252                 const u64 hole_end = extent_map_end(hole_em);
7253
7254                 em = alloc_extent_map();
7255                 if (!em) {
7256                         err = -ENOMEM;
7257                         goto out;
7258                 }
7259
7260                 ASSERT(hole_em);
7261                 /*
7262                  * When btrfs_get_extent can't find anything it returns one
7263                  * huge hole
7264                  *
7265                  * Make sure what it found really fits our range, and adjust to
7266                  * make sure it is based on the start from the caller
7267                  */
7268                 if (hole_end <= start || hole_em->start > end) {
7269                        free_extent_map(hole_em);
7270                        hole_em = NULL;
7271                 } else {
7272                        hole_start = max(hole_em->start, start);
7273                        hole_len = hole_end - hole_start;
7274                 }
7275
7276                 if (hole_em && delalloc_start > hole_start) {
7277                         /*
7278                          * Our hole starts before our delalloc, so we have to
7279                          * return just the parts of the hole that go until the
7280                          * delalloc starts
7281                          */
7282                         em->len = min(hole_len, delalloc_start - hole_start);
7283                         em->start = hole_start;
7284                         em->orig_start = hole_start;
7285                         /*
7286                          * Don't adjust block start at all, it is fixed at
7287                          * EXTENT_MAP_HOLE
7288                          */
7289                         em->block_start = hole_em->block_start;
7290                         em->block_len = hole_len;
7291                         if (test_bit(EXTENT_FLAG_PREALLOC, &hole_em->flags))
7292                                 set_bit(EXTENT_FLAG_PREALLOC, &em->flags);
7293                 } else {
7294                         /*
7295                          * Hole is out of passed range or it starts after
7296                          * delalloc range
7297                          */
7298                         em->start = delalloc_start;
7299                         em->len = delalloc_len;
7300                         em->orig_start = delalloc_start;
7301                         em->block_start = EXTENT_MAP_DELALLOC;
7302                         em->block_len = delalloc_len;
7303                 }
7304         } else {
7305                 return hole_em;
7306         }
7307 out:
7308
7309         free_extent_map(hole_em);
7310         if (err) {
7311                 free_extent_map(em);
7312                 return ERR_PTR(err);
7313         }
7314         return em;
7315 }
7316
7317 static struct extent_map *btrfs_create_dio_extent(struct inode *inode,
7318                                                   const u64 start,
7319                                                   const u64 len,
7320                                                   const u64 orig_start,
7321                                                   const u64 block_start,
7322                                                   const u64 block_len,
7323                                                   const u64 orig_block_len,
7324                                                   const u64 ram_bytes,
7325                                                   const int type)
7326 {
7327         struct extent_map *em = NULL;
7328         int ret;
7329
7330         if (type != BTRFS_ORDERED_NOCOW) {
7331                 em = create_io_em(inode, start, len, orig_start,
7332                                   block_start, block_len, orig_block_len,
7333                                   ram_bytes,
7334                                   BTRFS_COMPRESS_NONE, /* compress_type */
7335                                   type);
7336                 if (IS_ERR(em))
7337                         goto out;
7338         }
7339         ret = btrfs_add_ordered_extent_dio(inode, start, block_start,
7340                                            len, block_len, type);
7341         if (ret) {
7342                 if (em) {
7343                         free_extent_map(em);
7344                         btrfs_drop_extent_cache(BTRFS_I(inode), start,
7345                                                 start + len - 1, 0);
7346                 }
7347                 em = ERR_PTR(ret);
7348         }
7349  out:
7350
7351         return em;
7352 }
7353
7354 static struct extent_map *btrfs_new_extent_direct(struct inode *inode,
7355                                                   u64 start, u64 len)
7356 {
7357         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7358         struct btrfs_root *root = BTRFS_I(inode)->root;
7359         struct extent_map *em;
7360         struct btrfs_key ins;
7361         u64 alloc_hint;
7362         int ret;
7363
7364         alloc_hint = get_extent_allocation_hint(inode, start, len);
7365         ret = btrfs_reserve_extent(root, len, len, fs_info->sectorsize,
7366                                    0, alloc_hint, &ins, 1, 1);
7367         if (ret)
7368                 return ERR_PTR(ret);
7369
7370         em = btrfs_create_dio_extent(inode, start, ins.offset, start,
7371                                      ins.objectid, ins.offset, ins.offset,
7372                                      ins.offset, BTRFS_ORDERED_REGULAR);
7373         btrfs_dec_block_group_reservations(fs_info, ins.objectid);
7374         if (IS_ERR(em))
7375                 btrfs_free_reserved_extent(fs_info, ins.objectid,
7376                                            ins.offset, 1);
7377
7378         return em;
7379 }
7380
7381 /*
7382  * returns 1 when the nocow is safe, < 1 on error, 0 if the
7383  * block must be cow'd
7384  */
7385 noinline int can_nocow_extent(struct inode *inode, u64 offset, u64 *len,
7386                               u64 *orig_start, u64 *orig_block_len,
7387                               u64 *ram_bytes)
7388 {
7389         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7390         struct btrfs_path *path;
7391         int ret;
7392         struct extent_buffer *leaf;
7393         struct btrfs_root *root = BTRFS_I(inode)->root;
7394         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
7395         struct btrfs_file_extent_item *fi;
7396         struct btrfs_key key;
7397         u64 disk_bytenr;
7398         u64 backref_offset;
7399         u64 extent_end;
7400         u64 num_bytes;
7401         int slot;
7402         int found_type;
7403         bool nocow = (BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW);
7404
7405         path = btrfs_alloc_path();
7406         if (!path)
7407                 return -ENOMEM;
7408
7409         ret = btrfs_lookup_file_extent(NULL, root, path,
7410                         btrfs_ino(BTRFS_I(inode)), offset, 0);
7411         if (ret < 0)
7412                 goto out;
7413
7414         slot = path->slots[0];
7415         if (ret == 1) {
7416                 if (slot == 0) {
7417                         /* can't find the item, must cow */
7418                         ret = 0;
7419                         goto out;
7420                 }
7421                 slot--;
7422         }
7423         ret = 0;
7424         leaf = path->nodes[0];
7425         btrfs_item_key_to_cpu(leaf, &key, slot);
7426         if (key.objectid != btrfs_ino(BTRFS_I(inode)) ||
7427             key.type != BTRFS_EXTENT_DATA_KEY) {
7428                 /* not our file or wrong item type, must cow */
7429                 goto out;
7430         }
7431
7432         if (key.offset > offset) {
7433                 /* Wrong offset, must cow */
7434                 goto out;
7435         }
7436
7437         fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
7438         found_type = btrfs_file_extent_type(leaf, fi);
7439         if (found_type != BTRFS_FILE_EXTENT_REG &&
7440             found_type != BTRFS_FILE_EXTENT_PREALLOC) {
7441                 /* not a regular extent, must cow */
7442                 goto out;
7443         }
7444
7445         if (!nocow && found_type == BTRFS_FILE_EXTENT_REG)
7446                 goto out;
7447
7448         extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
7449         if (extent_end <= offset)
7450                 goto out;
7451
7452         disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
7453         if (disk_bytenr == 0)
7454                 goto out;
7455
7456         if (btrfs_file_extent_compression(leaf, fi) ||
7457             btrfs_file_extent_encryption(leaf, fi) ||
7458             btrfs_file_extent_other_encoding(leaf, fi))
7459                 goto out;
7460
7461         /*
7462          * Do the same check as in btrfs_cross_ref_exist but without the
7463          * unnecessary search.
7464          */
7465         if (btrfs_file_extent_generation(leaf, fi) <=
7466             btrfs_root_last_snapshot(&root->root_item))
7467                 goto out;
7468
7469         backref_offset = btrfs_file_extent_offset(leaf, fi);
7470
7471         if (orig_start) {
7472                 *orig_start = key.offset - backref_offset;
7473                 *orig_block_len = btrfs_file_extent_disk_num_bytes(leaf, fi);
7474                 *ram_bytes = btrfs_file_extent_ram_bytes(leaf, fi);
7475         }
7476
7477         if (btrfs_extent_readonly(fs_info, disk_bytenr))
7478                 goto out;
7479
7480         num_bytes = min(offset + *len, extent_end) - offset;
7481         if (!nocow && found_type == BTRFS_FILE_EXTENT_PREALLOC) {
7482                 u64 range_end;
7483
7484                 range_end = round_up(offset + num_bytes,
7485                                      root->fs_info->sectorsize) - 1;
7486                 ret = test_range_bit(io_tree, offset, range_end,
7487                                      EXTENT_DELALLOC, 0, NULL);
7488                 if (ret) {
7489                         ret = -EAGAIN;
7490                         goto out;
7491                 }
7492         }
7493
7494         btrfs_release_path(path);
7495
7496         /*
7497          * look for other files referencing this extent, if we
7498          * find any we must cow
7499          */
7500
7501         ret = btrfs_cross_ref_exist(root, btrfs_ino(BTRFS_I(inode)),
7502                                     key.offset - backref_offset, disk_bytenr);
7503         if (ret) {
7504                 ret = 0;
7505                 goto out;
7506         }
7507
7508         /*
7509          * adjust disk_bytenr and num_bytes to cover just the bytes
7510          * in this extent we are about to write.  If there
7511          * are any csums in that range we have to cow in order
7512          * to keep the csums correct
7513          */
7514         disk_bytenr += backref_offset;
7515         disk_bytenr += offset - key.offset;
7516         if (csum_exist_in_range(fs_info, disk_bytenr, num_bytes))
7517                 goto out;
7518         /*
7519          * all of the above have passed, it is safe to overwrite this extent
7520          * without cow
7521          */
7522         *len = num_bytes;
7523         ret = 1;
7524 out:
7525         btrfs_free_path(path);
7526         return ret;
7527 }
7528
7529 static int lock_extent_direct(struct inode *inode, u64 lockstart, u64 lockend,
7530                               struct extent_state **cached_state, int writing)
7531 {
7532         struct btrfs_ordered_extent *ordered;
7533         int ret = 0;
7534
7535         while (1) {
7536                 lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend,
7537                                  cached_state);
7538                 /*
7539                  * We're concerned with the entire range that we're going to be
7540                  * doing DIO to, so we need to make sure there's no ordered
7541                  * extents in this range.
7542                  */
7543                 ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), lockstart,
7544                                                      lockend - lockstart + 1);
7545
7546                 /*
7547                  * We need to make sure there are no buffered pages in this
7548                  * range either, we could have raced between the invalidate in
7549                  * generic_file_direct_write and locking the extent.  The
7550                  * invalidate needs to happen so that reads after a write do not
7551                  * get stale data.
7552                  */
7553                 if (!ordered &&
7554                     (!writing || !filemap_range_has_page(inode->i_mapping,
7555                                                          lockstart, lockend)))
7556                         break;
7557
7558                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
7559                                      cached_state);
7560
7561                 if (ordered) {
7562                         /*
7563                          * If we are doing a DIO read and the ordered extent we
7564                          * found is for a buffered write, we can not wait for it
7565                          * to complete and retry, because if we do so we can
7566                          * deadlock with concurrent buffered writes on page
7567                          * locks. This happens only if our DIO read covers more
7568                          * than one extent map, if at this point has already
7569                          * created an ordered extent for a previous extent map
7570                          * and locked its range in the inode's io tree, and a
7571                          * concurrent write against that previous extent map's
7572                          * range and this range started (we unlock the ranges
7573                          * in the io tree only when the bios complete and
7574                          * buffered writes always lock pages before attempting
7575                          * to lock range in the io tree).
7576                          */
7577                         if (writing ||
7578                             test_bit(BTRFS_ORDERED_DIRECT, &ordered->flags))
7579                                 btrfs_start_ordered_extent(inode, ordered, 1);
7580                         else
7581                                 ret = -ENOTBLK;
7582                         btrfs_put_ordered_extent(ordered);
7583                 } else {
7584                         /*
7585                          * We could trigger writeback for this range (and wait
7586                          * for it to complete) and then invalidate the pages for
7587                          * this range (through invalidate_inode_pages2_range()),
7588                          * but that can lead us to a deadlock with a concurrent
7589                          * call to readpages() (a buffered read or a defrag call
7590                          * triggered a readahead) on a page lock due to an
7591                          * ordered dio extent we created before but did not have
7592                          * yet a corresponding bio submitted (whence it can not
7593                          * complete), which makes readpages() wait for that
7594                          * ordered extent to complete while holding a lock on
7595                          * that page.
7596                          */
7597                         ret = -ENOTBLK;
7598                 }
7599
7600                 if (ret)
7601                         break;
7602
7603                 cond_resched();
7604         }
7605
7606         return ret;
7607 }
7608
7609 /* The callers of this must take lock_extent() */
7610 static struct extent_map *create_io_em(struct inode *inode, u64 start, u64 len,
7611                                        u64 orig_start, u64 block_start,
7612                                        u64 block_len, u64 orig_block_len,
7613                                        u64 ram_bytes, int compress_type,
7614                                        int type)
7615 {
7616         struct extent_map_tree *em_tree;
7617         struct extent_map *em;
7618         int ret;
7619
7620         ASSERT(type == BTRFS_ORDERED_PREALLOC ||
7621                type == BTRFS_ORDERED_COMPRESSED ||
7622                type == BTRFS_ORDERED_NOCOW ||
7623                type == BTRFS_ORDERED_REGULAR);
7624
7625         em_tree = &BTRFS_I(inode)->extent_tree;
7626         em = alloc_extent_map();
7627         if (!em)
7628                 return ERR_PTR(-ENOMEM);
7629
7630         em->start = start;
7631         em->orig_start = orig_start;
7632         em->len = len;
7633         em->block_len = block_len;
7634         em->block_start = block_start;
7635         em->orig_block_len = orig_block_len;
7636         em->ram_bytes = ram_bytes;
7637         em->generation = -1;
7638         set_bit(EXTENT_FLAG_PINNED, &em->flags);
7639         if (type == BTRFS_ORDERED_PREALLOC) {
7640                 set_bit(EXTENT_FLAG_FILLING, &em->flags);
7641         } else if (type == BTRFS_ORDERED_COMPRESSED) {
7642                 set_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
7643                 em->compress_type = compress_type;
7644         }
7645
7646         do {
7647                 btrfs_drop_extent_cache(BTRFS_I(inode), em->start,
7648                                 em->start + em->len - 1, 0);
7649                 write_lock(&em_tree->lock);
7650                 ret = add_extent_mapping(em_tree, em, 1);
7651                 write_unlock(&em_tree->lock);
7652                 /*
7653                  * The caller has taken lock_extent(), who could race with us
7654                  * to add em?
7655                  */
7656         } while (ret == -EEXIST);
7657
7658         if (ret) {
7659                 free_extent_map(em);
7660                 return ERR_PTR(ret);
7661         }
7662
7663         /* em got 2 refs now, callers needs to do free_extent_map once. */
7664         return em;
7665 }
7666
7667
7668 static int btrfs_get_blocks_direct_read(struct extent_map *em,
7669                                         struct buffer_head *bh_result,
7670                                         struct inode *inode,
7671                                         u64 start, u64 len)
7672 {
7673         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7674
7675         if (em->block_start == EXTENT_MAP_HOLE ||
7676                         test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
7677                 return -ENOENT;
7678
7679         len = min(len, em->len - (start - em->start));
7680
7681         bh_result->b_blocknr = (em->block_start + (start - em->start)) >>
7682                 inode->i_blkbits;
7683         bh_result->b_size = len;
7684         bh_result->b_bdev = fs_info->fs_devices->latest_bdev;
7685         set_buffer_mapped(bh_result);
7686
7687         return 0;
7688 }
7689
7690 static int btrfs_get_blocks_direct_write(struct extent_map **map,
7691                                          struct buffer_head *bh_result,
7692                                          struct inode *inode,
7693                                          struct btrfs_dio_data *dio_data,
7694                                          u64 start, u64 len)
7695 {
7696         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7697         struct extent_map *em = *map;
7698         int ret = 0;
7699
7700         /*
7701          * We don't allocate a new extent in the following cases
7702          *
7703          * 1) The inode is marked as NODATACOW. In this case we'll just use the
7704          * existing extent.
7705          * 2) The extent is marked as PREALLOC. We're good to go here and can
7706          * just use the extent.
7707          *
7708          */
7709         if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags) ||
7710             ((BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW) &&
7711              em->block_start != EXTENT_MAP_HOLE)) {
7712                 int type;
7713                 u64 block_start, orig_start, orig_block_len, ram_bytes;
7714
7715                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
7716                         type = BTRFS_ORDERED_PREALLOC;
7717                 else
7718                         type = BTRFS_ORDERED_NOCOW;
7719                 len = min(len, em->len - (start - em->start));
7720                 block_start = em->block_start + (start - em->start);
7721
7722                 if (can_nocow_extent(inode, start, &len, &orig_start,
7723                                      &orig_block_len, &ram_bytes) == 1 &&
7724                     btrfs_inc_nocow_writers(fs_info, block_start)) {
7725                         struct extent_map *em2;
7726
7727                         em2 = btrfs_create_dio_extent(inode, start, len,
7728                                                       orig_start, block_start,
7729                                                       len, orig_block_len,
7730                                                       ram_bytes, type);
7731                         btrfs_dec_nocow_writers(fs_info, block_start);
7732                         if (type == BTRFS_ORDERED_PREALLOC) {
7733                                 free_extent_map(em);
7734                                 *map = em = em2;
7735                         }
7736
7737                         if (em2 && IS_ERR(em2)) {
7738                                 ret = PTR_ERR(em2);
7739                                 goto out;
7740                         }
7741                         /*
7742                          * For inode marked NODATACOW or extent marked PREALLOC,
7743                          * use the existing or preallocated extent, so does not
7744                          * need to adjust btrfs_space_info's bytes_may_use.
7745                          */
7746                         btrfs_free_reserved_data_space_noquota(inode, start,
7747                                                                len);
7748                         goto skip_cow;
7749                 }
7750         }
7751
7752         /* this will cow the extent */
7753         len = bh_result->b_size;
7754         free_extent_map(em);
7755         *map = em = btrfs_new_extent_direct(inode, start, len);
7756         if (IS_ERR(em)) {
7757                 ret = PTR_ERR(em);
7758                 goto out;
7759         }
7760
7761         len = min(len, em->len - (start - em->start));
7762
7763 skip_cow:
7764         bh_result->b_blocknr = (em->block_start + (start - em->start)) >>
7765                 inode->i_blkbits;
7766         bh_result->b_size = len;
7767         bh_result->b_bdev = fs_info->fs_devices->latest_bdev;
7768         set_buffer_mapped(bh_result);
7769
7770         if (!test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
7771                 set_buffer_new(bh_result);
7772
7773         /*
7774          * Need to update the i_size under the extent lock so buffered
7775          * readers will get the updated i_size when we unlock.
7776          */
7777         if (!dio_data->overwrite && start + len > i_size_read(inode))
7778                 i_size_write(inode, start + len);
7779
7780         WARN_ON(dio_data->reserve < len);
7781         dio_data->reserve -= len;
7782         dio_data->unsubmitted_oe_range_end = start + len;
7783         current->journal_info = dio_data;
7784 out:
7785         return ret;
7786 }
7787
7788 static int btrfs_get_blocks_direct(struct inode *inode, sector_t iblock,
7789                                    struct buffer_head *bh_result, int create)
7790 {
7791         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7792         struct extent_map *em;
7793         struct extent_state *cached_state = NULL;
7794         struct btrfs_dio_data *dio_data = NULL;
7795         u64 start = iblock << inode->i_blkbits;
7796         u64 lockstart, lockend;
7797         u64 len = bh_result->b_size;
7798         int ret = 0;
7799
7800         if (!create)
7801                 len = min_t(u64, len, fs_info->sectorsize);
7802
7803         lockstart = start;
7804         lockend = start + len - 1;
7805
7806         if (current->journal_info) {
7807                 /*
7808                  * Need to pull our outstanding extents and set journal_info to NULL so
7809                  * that anything that needs to check if there's a transaction doesn't get
7810                  * confused.
7811                  */
7812                 dio_data = current->journal_info;
7813                 current->journal_info = NULL;
7814         }
7815
7816         /*
7817          * If this errors out it's because we couldn't invalidate pagecache for
7818          * this range and we need to fallback to buffered.
7819          */
7820         if (lock_extent_direct(inode, lockstart, lockend, &cached_state,
7821                                create)) {
7822                 ret = -ENOTBLK;
7823                 goto err;
7824         }
7825
7826         em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, start, len, 0);
7827         if (IS_ERR(em)) {
7828                 ret = PTR_ERR(em);
7829                 goto unlock_err;
7830         }
7831
7832         /*
7833          * Ok for INLINE and COMPRESSED extents we need to fallback on buffered
7834          * io.  INLINE is special, and we could probably kludge it in here, but
7835          * it's still buffered so for safety lets just fall back to the generic
7836          * buffered path.
7837          *
7838          * For COMPRESSED we _have_ to read the entire extent in so we can
7839          * decompress it, so there will be buffering required no matter what we
7840          * do, so go ahead and fallback to buffered.
7841          *
7842          * We return -ENOTBLK because that's what makes DIO go ahead and go back
7843          * to buffered IO.  Don't blame me, this is the price we pay for using
7844          * the generic code.
7845          */
7846         if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags) ||
7847             em->block_start == EXTENT_MAP_INLINE) {
7848                 free_extent_map(em);
7849                 ret = -ENOTBLK;
7850                 goto unlock_err;
7851         }
7852
7853         if (create) {
7854                 ret = btrfs_get_blocks_direct_write(&em, bh_result, inode,
7855                                                     dio_data, start, len);
7856                 if (ret < 0)
7857                         goto unlock_err;
7858
7859                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart,
7860                                      lockend, &cached_state);
7861         } else {
7862                 ret = btrfs_get_blocks_direct_read(em, bh_result, inode,
7863                                                    start, len);
7864                 /* Can be negative only if we read from a hole */
7865                 if (ret < 0) {
7866                         ret = 0;
7867                         free_extent_map(em);
7868                         goto unlock_err;
7869                 }
7870                 /*
7871                  * We need to unlock only the end area that we aren't using.
7872                  * The rest is going to be unlocked by the endio routine.
7873                  */
7874                 lockstart = start + bh_result->b_size;
7875                 if (lockstart < lockend) {
7876                         unlock_extent_cached(&BTRFS_I(inode)->io_tree,
7877                                              lockstart, lockend, &cached_state);
7878                 } else {
7879                         free_extent_state(cached_state);
7880                 }
7881         }
7882
7883         free_extent_map(em);
7884
7885         return 0;
7886
7887 unlock_err:
7888         unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
7889                              &cached_state);
7890 err:
7891         if (dio_data)
7892                 current->journal_info = dio_data;
7893         return ret;
7894 }
7895
7896 static inline blk_status_t submit_dio_repair_bio(struct inode *inode,
7897                                                  struct bio *bio,
7898                                                  int mirror_num)
7899 {
7900         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7901         blk_status_t ret;
7902
7903         BUG_ON(bio_op(bio) == REQ_OP_WRITE);
7904
7905         ret = btrfs_bio_wq_end_io(fs_info, bio, BTRFS_WQ_ENDIO_DIO_REPAIR);
7906         if (ret)
7907                 return ret;
7908
7909         ret = btrfs_map_bio(fs_info, bio, mirror_num);
7910
7911         return ret;
7912 }
7913
7914 static int btrfs_check_dio_repairable(struct inode *inode,
7915                                       struct bio *failed_bio,
7916                                       struct io_failure_record *failrec,
7917                                       int failed_mirror)
7918 {
7919         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7920         int num_copies;
7921
7922         num_copies = btrfs_num_copies(fs_info, failrec->logical, failrec->len);
7923         if (num_copies == 1) {
7924                 /*
7925                  * we only have a single copy of the data, so don't bother with
7926                  * all the retry and error correction code that follows. no
7927                  * matter what the error is, it is very likely to persist.
7928                  */
7929                 btrfs_debug(fs_info,
7930                         "Check DIO Repairable: cannot repair, num_copies=%d, next_mirror %d, failed_mirror %d",
7931                         num_copies, failrec->this_mirror, failed_mirror);
7932                 return 0;
7933         }
7934
7935         failrec->failed_mirror = failed_mirror;
7936         failrec->this_mirror++;
7937         if (failrec->this_mirror == failed_mirror)
7938                 failrec->this_mirror++;
7939
7940         if (failrec->this_mirror > num_copies) {
7941                 btrfs_debug(fs_info,
7942                         "Check DIO Repairable: (fail) num_copies=%d, next_mirror %d, failed_mirror %d",
7943                         num_copies, failrec->this_mirror, failed_mirror);
7944                 return 0;
7945         }
7946
7947         return 1;
7948 }
7949
7950 static blk_status_t dio_read_error(struct inode *inode, struct bio *failed_bio,
7951                                    struct page *page, unsigned int pgoff,
7952                                    u64 start, u64 end, int failed_mirror,
7953                                    bio_end_io_t *repair_endio, void *repair_arg)
7954 {
7955         struct io_failure_record *failrec;
7956         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
7957         struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
7958         struct bio *bio;
7959         int isector;
7960         unsigned int read_mode = 0;
7961         int segs;
7962         int ret;
7963         blk_status_t status;
7964         struct bio_vec bvec;
7965
7966         BUG_ON(bio_op(failed_bio) == REQ_OP_WRITE);
7967
7968         ret = btrfs_get_io_failure_record(inode, start, end, &failrec);
7969         if (ret)
7970                 return errno_to_blk_status(ret);
7971
7972         ret = btrfs_check_dio_repairable(inode, failed_bio, failrec,
7973                                          failed_mirror);
7974         if (!ret) {
7975                 free_io_failure(failure_tree, io_tree, failrec);
7976                 return BLK_STS_IOERR;
7977         }
7978
7979         segs = bio_segments(failed_bio);
7980         bio_get_first_bvec(failed_bio, &bvec);
7981         if (segs > 1 ||
7982             (bvec.bv_len > btrfs_inode_sectorsize(inode)))
7983                 read_mode |= REQ_FAILFAST_DEV;
7984
7985         isector = start - btrfs_io_bio(failed_bio)->logical;
7986         isector >>= inode->i_sb->s_blocksize_bits;
7987         bio = btrfs_create_repair_bio(inode, failed_bio, failrec, page,
7988                                 pgoff, isector, repair_endio, repair_arg);
7989         bio->bi_opf = REQ_OP_READ | read_mode;
7990
7991         btrfs_debug(BTRFS_I(inode)->root->fs_info,
7992                     "repair DIO read error: submitting new dio read[%#x] to this_mirror=%d, in_validation=%d",
7993                     read_mode, failrec->this_mirror, failrec->in_validation);
7994
7995         status = submit_dio_repair_bio(inode, bio, failrec->this_mirror);
7996         if (status) {
7997                 free_io_failure(failure_tree, io_tree, failrec);
7998                 bio_put(bio);
7999         }
8000
8001         return status;
8002 }
8003
8004 struct btrfs_retry_complete {
8005         struct completion done;
8006         struct inode *inode;
8007         u64 start;
8008         int uptodate;
8009 };
8010
8011 static void btrfs_retry_endio_nocsum(struct bio *bio)
8012 {
8013         struct btrfs_retry_complete *done = bio->bi_private;
8014         struct inode *inode = done->inode;
8015         struct bio_vec *bvec;
8016         struct extent_io_tree *io_tree, *failure_tree;
8017         struct bvec_iter_all iter_all;
8018
8019         if (bio->bi_status)
8020                 goto end;
8021
8022         ASSERT(bio->bi_vcnt == 1);
8023         io_tree = &BTRFS_I(inode)->io_tree;
8024         failure_tree = &BTRFS_I(inode)->io_failure_tree;
8025         ASSERT(bio_first_bvec_all(bio)->bv_len == btrfs_inode_sectorsize(inode));
8026
8027         done->uptodate = 1;
8028         ASSERT(!bio_flagged(bio, BIO_CLONED));
8029         bio_for_each_segment_all(bvec, bio, iter_all)
8030                 clean_io_failure(BTRFS_I(inode)->root->fs_info, failure_tree,
8031                                  io_tree, done->start, bvec->bv_page,
8032                                  btrfs_ino(BTRFS_I(inode)), 0);
8033 end:
8034         complete(&done->done);
8035         bio_put(bio);
8036 }
8037
8038 static blk_status_t __btrfs_correct_data_nocsum(struct inode *inode,
8039                                                 struct btrfs_io_bio *io_bio)
8040 {
8041         struct btrfs_fs_info *fs_info;
8042         struct bio_vec bvec;
8043         struct bvec_iter iter;
8044         struct btrfs_retry_complete done;
8045         u64 start;
8046         unsigned int pgoff;
8047         u32 sectorsize;
8048         int nr_sectors;
8049         blk_status_t ret;
8050         blk_status_t err = BLK_STS_OK;
8051
8052         fs_info = BTRFS_I(inode)->root->fs_info;
8053         sectorsize = fs_info->sectorsize;
8054
8055         start = io_bio->logical;
8056         done.inode = inode;
8057         io_bio->bio.bi_iter = io_bio->iter;
8058
8059         bio_for_each_segment(bvec, &io_bio->bio, iter) {
8060                 nr_sectors = BTRFS_BYTES_TO_BLKS(fs_info, bvec.bv_len);
8061                 pgoff = bvec.bv_offset;
8062
8063 next_block_or_try_again:
8064                 done.uptodate = 0;
8065                 done.start = start;
8066                 init_completion(&done.done);
8067
8068                 ret = dio_read_error(inode, &io_bio->bio, bvec.bv_page,
8069                                 pgoff, start, start + sectorsize - 1,
8070                                 io_bio->mirror_num,
8071                                 btrfs_retry_endio_nocsum, &done);
8072                 if (ret) {
8073                         err = ret;
8074                         goto next;
8075                 }
8076
8077                 wait_for_completion_io(&done.done);
8078
8079                 if (!done.uptodate) {
8080                         /* We might have another mirror, so try again */
8081                         goto next_block_or_try_again;
8082                 }
8083
8084 next:
8085                 start += sectorsize;
8086
8087                 nr_sectors--;
8088                 if (nr_sectors) {
8089                         pgoff += sectorsize;
8090                         ASSERT(pgoff < PAGE_SIZE);
8091                         goto next_block_or_try_again;
8092                 }
8093         }
8094
8095         return err;
8096 }
8097
8098 static void btrfs_retry_endio(struct bio *bio)
8099 {
8100         struct btrfs_retry_complete *done = bio->bi_private;
8101         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
8102         struct extent_io_tree *io_tree, *failure_tree;
8103         struct inode *inode = done->inode;
8104         struct bio_vec *bvec;
8105         int uptodate;
8106         int ret;
8107         int i = 0;
8108         struct bvec_iter_all iter_all;
8109
8110         if (bio->bi_status)
8111                 goto end;
8112
8113         uptodate = 1;
8114
8115         ASSERT(bio->bi_vcnt == 1);
8116         ASSERT(bio_first_bvec_all(bio)->bv_len == btrfs_inode_sectorsize(done->inode));
8117
8118         io_tree = &BTRFS_I(inode)->io_tree;
8119         failure_tree = &BTRFS_I(inode)->io_failure_tree;
8120
8121         ASSERT(!bio_flagged(bio, BIO_CLONED));
8122         bio_for_each_segment_all(bvec, bio, iter_all) {
8123                 ret = __readpage_endio_check(inode, io_bio, i, bvec->bv_page,
8124                                              bvec->bv_offset, done->start,
8125                                              bvec->bv_len);
8126                 if (!ret)
8127                         clean_io_failure(BTRFS_I(inode)->root->fs_info,
8128                                          failure_tree, io_tree, done->start,
8129                                          bvec->bv_page,
8130                                          btrfs_ino(BTRFS_I(inode)),
8131                                          bvec->bv_offset);
8132                 else
8133                         uptodate = 0;
8134                 i++;
8135         }
8136
8137         done->uptodate = uptodate;
8138 end:
8139         complete(&done->done);
8140         bio_put(bio);
8141 }
8142
8143 static blk_status_t __btrfs_subio_endio_read(struct inode *inode,
8144                 struct btrfs_io_bio *io_bio, blk_status_t err)
8145 {
8146         struct btrfs_fs_info *fs_info;
8147         struct bio_vec bvec;
8148         struct bvec_iter iter;
8149         struct btrfs_retry_complete done;
8150         u64 start;
8151         u64 offset = 0;
8152         u32 sectorsize;
8153         int nr_sectors;
8154         unsigned int pgoff;
8155         int csum_pos;
8156         bool uptodate = (err == 0);
8157         int ret;
8158         blk_status_t status;
8159
8160         fs_info = BTRFS_I(inode)->root->fs_info;
8161         sectorsize = fs_info->sectorsize;
8162
8163         err = BLK_STS_OK;
8164         start = io_bio->logical;
8165         done.inode = inode;
8166         io_bio->bio.bi_iter = io_bio->iter;
8167
8168         bio_for_each_segment(bvec, &io_bio->bio, iter) {
8169                 nr_sectors = BTRFS_BYTES_TO_BLKS(fs_info, bvec.bv_len);
8170
8171                 pgoff = bvec.bv_offset;
8172 next_block:
8173                 if (uptodate) {
8174                         csum_pos = BTRFS_BYTES_TO_BLKS(fs_info, offset);
8175                         ret = __readpage_endio_check(inode, io_bio, csum_pos,
8176                                         bvec.bv_page, pgoff, start, sectorsize);
8177                         if (likely(!ret))
8178                                 goto next;
8179                 }
8180 try_again:
8181                 done.uptodate = 0;
8182                 done.start = start;
8183                 init_completion(&done.done);
8184
8185                 status = dio_read_error(inode, &io_bio->bio, bvec.bv_page,
8186                                         pgoff, start, start + sectorsize - 1,
8187                                         io_bio->mirror_num, btrfs_retry_endio,
8188                                         &done);
8189                 if (status) {
8190                         err = status;
8191                         goto next;
8192                 }
8193
8194                 wait_for_completion_io(&done.done);
8195
8196                 if (!done.uptodate) {
8197                         /* We might have another mirror, so try again */
8198                         goto try_again;
8199                 }
8200 next:
8201                 offset += sectorsize;
8202                 start += sectorsize;
8203
8204                 ASSERT(nr_sectors);
8205
8206                 nr_sectors--;
8207                 if (nr_sectors) {
8208                         pgoff += sectorsize;
8209                         ASSERT(pgoff < PAGE_SIZE);
8210                         goto next_block;
8211                 }
8212         }
8213
8214         return err;
8215 }
8216
8217 static blk_status_t btrfs_subio_endio_read(struct inode *inode,
8218                 struct btrfs_io_bio *io_bio, blk_status_t err)
8219 {
8220         bool skip_csum = BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM;
8221
8222         if (skip_csum) {
8223                 if (unlikely(err))
8224                         return __btrfs_correct_data_nocsum(inode, io_bio);
8225                 else
8226                         return BLK_STS_OK;
8227         } else {
8228                 return __btrfs_subio_endio_read(inode, io_bio, err);
8229         }
8230 }
8231
8232 static void btrfs_endio_direct_read(struct bio *bio)
8233 {
8234         struct btrfs_dio_private *dip = bio->bi_private;
8235         struct inode *inode = dip->inode;
8236         struct bio *dio_bio;
8237         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
8238         blk_status_t err = bio->bi_status;
8239
8240         if (dip->flags & BTRFS_DIO_ORIG_BIO_SUBMITTED)
8241                 err = btrfs_subio_endio_read(inode, io_bio, err);
8242
8243         unlock_extent(&BTRFS_I(inode)->io_tree, dip->logical_offset,
8244                       dip->logical_offset + dip->bytes - 1);
8245         dio_bio = dip->dio_bio;
8246
8247         kfree(dip);
8248
8249         dio_bio->bi_status = err;
8250         dio_end_io(dio_bio);
8251         btrfs_io_bio_free_csum(io_bio);
8252         bio_put(bio);
8253 }
8254
8255 static void __endio_write_update_ordered(struct inode *inode,
8256                                          const u64 offset, const u64 bytes,
8257                                          const bool uptodate)
8258 {
8259         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
8260         struct btrfs_ordered_extent *ordered = NULL;
8261         struct btrfs_workqueue *wq;
8262         u64 ordered_offset = offset;
8263         u64 ordered_bytes = bytes;
8264         u64 last_offset;
8265
8266         if (btrfs_is_free_space_inode(BTRFS_I(inode)))
8267                 wq = fs_info->endio_freespace_worker;
8268         else
8269                 wq = fs_info->endio_write_workers;
8270
8271         while (ordered_offset < offset + bytes) {
8272                 last_offset = ordered_offset;
8273                 if (btrfs_dec_test_first_ordered_pending(inode, &ordered,
8274                                                            &ordered_offset,
8275                                                            ordered_bytes,
8276                                                            uptodate)) {
8277                         btrfs_init_work(&ordered->work, finish_ordered_fn, NULL,
8278                                         NULL);
8279                         btrfs_queue_work(wq, &ordered->work);
8280                 }
8281                 /*
8282                  * If btrfs_dec_test_ordered_pending does not find any ordered
8283                  * extent in the range, we can exit.
8284                  */
8285                 if (ordered_offset == last_offset)
8286                         return;
8287                 /*
8288                  * Our bio might span multiple ordered extents. In this case
8289                  * we keep going until we have accounted the whole dio.
8290                  */
8291                 if (ordered_offset < offset + bytes) {
8292                         ordered_bytes = offset + bytes - ordered_offset;
8293                         ordered = NULL;
8294                 }
8295         }
8296 }
8297
8298 static void btrfs_endio_direct_write(struct bio *bio)
8299 {
8300         struct btrfs_dio_private *dip = bio->bi_private;
8301         struct bio *dio_bio = dip->dio_bio;
8302
8303         __endio_write_update_ordered(dip->inode, dip->logical_offset,
8304                                      dip->bytes, !bio->bi_status);
8305
8306         kfree(dip);
8307
8308         dio_bio->bi_status = bio->bi_status;
8309         dio_end_io(dio_bio);
8310         bio_put(bio);
8311 }
8312
8313 static blk_status_t btrfs_submit_bio_start_direct_io(void *private_data,
8314                                     struct bio *bio, u64 offset)
8315 {
8316         struct inode *inode = private_data;
8317         blk_status_t ret;
8318         ret = btrfs_csum_one_bio(inode, bio, offset, 1);
8319         BUG_ON(ret); /* -ENOMEM */
8320         return 0;
8321 }
8322
8323 static void btrfs_end_dio_bio(struct bio *bio)
8324 {
8325         struct btrfs_dio_private *dip = bio->bi_private;
8326         blk_status_t err = bio->bi_status;
8327
8328         if (err)
8329                 btrfs_warn(BTRFS_I(dip->inode)->root->fs_info,
8330                            "direct IO failed ino %llu rw %d,%u sector %#Lx len %u err no %d",
8331                            btrfs_ino(BTRFS_I(dip->inode)), bio_op(bio),
8332                            bio->bi_opf,
8333                            (unsigned long long)bio->bi_iter.bi_sector,
8334                            bio->bi_iter.bi_size, err);
8335
8336         if (dip->subio_endio)
8337                 err = dip->subio_endio(dip->inode, btrfs_io_bio(bio), err);
8338
8339         if (err) {
8340                 /*
8341                  * We want to perceive the errors flag being set before
8342                  * decrementing the reference count. We don't need a barrier
8343                  * since atomic operations with a return value are fully
8344                  * ordered as per atomic_t.txt
8345                  */
8346                 dip->errors = 1;
8347         }
8348
8349         /* if there are more bios still pending for this dio, just exit */
8350         if (!atomic_dec_and_test(&dip->pending_bios))
8351                 goto out;
8352
8353         if (dip->errors) {
8354                 bio_io_error(dip->orig_bio);
8355         } else {
8356                 dip->dio_bio->bi_status = BLK_STS_OK;
8357                 bio_endio(dip->orig_bio);
8358         }
8359 out:
8360         bio_put(bio);
8361 }
8362
8363 static inline blk_status_t btrfs_lookup_and_bind_dio_csum(struct inode *inode,
8364                                                  struct btrfs_dio_private *dip,
8365                                                  struct bio *bio,
8366                                                  u64 file_offset)
8367 {
8368         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
8369         struct btrfs_io_bio *orig_io_bio = btrfs_io_bio(dip->orig_bio);
8370         blk_status_t ret;
8371
8372         /*
8373          * We load all the csum data we need when we submit
8374          * the first bio to reduce the csum tree search and
8375          * contention.
8376          */
8377         if (dip->logical_offset == file_offset) {
8378                 ret = btrfs_lookup_bio_sums_dio(inode, dip->orig_bio,
8379                                                 file_offset);
8380                 if (ret)
8381                         return ret;
8382         }
8383
8384         if (bio == dip->orig_bio)
8385                 return 0;
8386
8387         file_offset -= dip->logical_offset;
8388         file_offset >>= inode->i_sb->s_blocksize_bits;
8389         io_bio->csum = (u8 *)(((u32 *)orig_io_bio->csum) + file_offset);
8390
8391         return 0;
8392 }
8393
8394 static inline blk_status_t btrfs_submit_dio_bio(struct bio *bio,
8395                 struct inode *inode, u64 file_offset, int async_submit)
8396 {
8397         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
8398         struct btrfs_dio_private *dip = bio->bi_private;
8399         bool write = bio_op(bio) == REQ_OP_WRITE;
8400         blk_status_t ret;
8401
8402         /* Check btrfs_submit_bio_hook() for rules about async submit. */
8403         if (async_submit)
8404                 async_submit = !atomic_read(&BTRFS_I(inode)->sync_writers);
8405
8406         if (!write) {
8407                 ret = btrfs_bio_wq_end_io(fs_info, bio, BTRFS_WQ_ENDIO_DATA);
8408                 if (ret)
8409                         goto err;
8410         }
8411
8412         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)
8413                 goto map;
8414
8415         if (write && async_submit) {
8416                 ret = btrfs_wq_submit_bio(fs_info, bio, 0, 0,
8417                                           file_offset, inode,
8418                                           btrfs_submit_bio_start_direct_io);
8419                 goto err;
8420         } else if (write) {
8421                 /*
8422                  * If we aren't doing async submit, calculate the csum of the
8423                  * bio now.
8424                  */
8425                 ret = btrfs_csum_one_bio(inode, bio, file_offset, 1);
8426                 if (ret)
8427                         goto err;
8428         } else {
8429                 ret = btrfs_lookup_and_bind_dio_csum(inode, dip, bio,
8430                                                      file_offset);
8431                 if (ret)
8432                         goto err;
8433         }
8434 map:
8435         ret = btrfs_map_bio(fs_info, bio, 0);
8436 err:
8437         return ret;
8438 }
8439
8440 static int btrfs_submit_direct_hook(struct btrfs_dio_private *dip)
8441 {
8442         struct inode *inode = dip->inode;
8443         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
8444         struct bio *bio;
8445         struct bio *orig_bio = dip->orig_bio;
8446         u64 start_sector = orig_bio->bi_iter.bi_sector;
8447         u64 file_offset = dip->logical_offset;
8448         int async_submit = 0;
8449         u64 submit_len;
8450         int clone_offset = 0;
8451         int clone_len;
8452         int ret;
8453         blk_status_t status;
8454         struct btrfs_io_geometry geom;
8455
8456         submit_len = orig_bio->bi_iter.bi_size;
8457         ret = btrfs_get_io_geometry(fs_info, btrfs_op(orig_bio),
8458                                     start_sector << 9, submit_len, &geom);
8459         if (ret)
8460                 return -EIO;
8461
8462         if (geom.len >= submit_len) {
8463                 bio = orig_bio;
8464                 dip->flags |= BTRFS_DIO_ORIG_BIO_SUBMITTED;
8465                 goto submit;
8466         }
8467
8468         /* async crcs make it difficult to collect full stripe writes. */
8469         if (btrfs_data_alloc_profile(fs_info) & BTRFS_BLOCK_GROUP_RAID56_MASK)
8470                 async_submit = 0;
8471         else
8472                 async_submit = 1;
8473
8474         /* bio split */
8475         ASSERT(geom.len <= INT_MAX);
8476         atomic_inc(&dip->pending_bios);
8477         do {
8478                 clone_len = min_t(int, submit_len, geom.len);
8479
8480                 /*
8481                  * This will never fail as it's passing GPF_NOFS and
8482                  * the allocation is backed by btrfs_bioset.
8483                  */
8484                 bio = btrfs_bio_clone_partial(orig_bio, clone_offset,
8485                                               clone_len);
8486                 bio->bi_private = dip;
8487                 bio->bi_end_io = btrfs_end_dio_bio;
8488                 btrfs_io_bio(bio)->logical = file_offset;
8489
8490                 ASSERT(submit_len >= clone_len);
8491                 submit_len -= clone_len;
8492                 if (submit_len == 0)
8493                         break;
8494
8495                 /*
8496                  * Increase the count before we submit the bio so we know
8497                  * the end IO handler won't happen before we increase the
8498                  * count. Otherwise, the dip might get freed before we're
8499                  * done setting it up.
8500                  */
8501                 atomic_inc(&dip->pending_bios);
8502
8503                 status = btrfs_submit_dio_bio(bio, inode, file_offset,
8504                                                 async_submit);
8505                 if (status) {
8506                         bio_put(bio);
8507                         atomic_dec(&dip->pending_bios);
8508                         goto out_err;
8509                 }
8510
8511                 clone_offset += clone_len;
8512                 start_sector += clone_len >> 9;
8513                 file_offset += clone_len;
8514
8515                 ret = btrfs_get_io_geometry(fs_info, btrfs_op(orig_bio),
8516                                       start_sector << 9, submit_len, &geom);
8517                 if (ret)
8518                         goto out_err;
8519         } while (submit_len > 0);
8520
8521 submit:
8522         status = btrfs_submit_dio_bio(bio, inode, file_offset, async_submit);
8523         if (!status)
8524                 return 0;
8525
8526         bio_put(bio);
8527 out_err:
8528         dip->errors = 1;
8529         /*
8530          * Before atomic variable goto zero, we must  make sure dip->errors is
8531          * perceived to be set. This ordering is ensured by the fact that an
8532          * atomic operations with a return value are fully ordered as per
8533          * atomic_t.txt
8534          */
8535         if (atomic_dec_and_test(&dip->pending_bios))
8536                 bio_io_error(dip->orig_bio);
8537
8538         /* bio_end_io() will handle error, so we needn't return it */
8539         return 0;
8540 }
8541
8542 static void btrfs_submit_direct(struct bio *dio_bio, struct inode *inode,
8543                                 loff_t file_offset)
8544 {
8545         struct btrfs_dio_private *dip = NULL;
8546         struct bio *bio = NULL;
8547         struct btrfs_io_bio *io_bio;
8548         bool write = (bio_op(dio_bio) == REQ_OP_WRITE);
8549         int ret = 0;
8550
8551         bio = btrfs_bio_clone(dio_bio);
8552
8553         dip = kzalloc(sizeof(*dip), GFP_NOFS);
8554         if (!dip) {
8555                 ret = -ENOMEM;
8556                 goto free_ordered;
8557         }
8558
8559         dip->private = dio_bio->bi_private;
8560         dip->inode = inode;
8561         dip->logical_offset = file_offset;
8562         dip->bytes = dio_bio->bi_iter.bi_size;
8563         dip->disk_bytenr = (u64)dio_bio->bi_iter.bi_sector << 9;
8564         bio->bi_private = dip;
8565         dip->orig_bio = bio;
8566         dip->dio_bio = dio_bio;
8567         atomic_set(&dip->pending_bios, 0);
8568         io_bio = btrfs_io_bio(bio);
8569         io_bio->logical = file_offset;
8570
8571         if (write) {
8572                 bio->bi_end_io = btrfs_endio_direct_write;
8573         } else {
8574                 bio->bi_end_io = btrfs_endio_direct_read;
8575                 dip->subio_endio = btrfs_subio_endio_read;
8576         }
8577
8578         /*
8579          * Reset the range for unsubmitted ordered extents (to a 0 length range)
8580          * even if we fail to submit a bio, because in such case we do the
8581          * corresponding error handling below and it must not be done a second
8582          * time by btrfs_direct_IO().
8583          */
8584         if (write) {
8585                 struct btrfs_dio_data *dio_data = current->journal_info;
8586
8587                 dio_data->unsubmitted_oe_range_end = dip->logical_offset +
8588                         dip->bytes;
8589                 dio_data->unsubmitted_oe_range_start =
8590                         dio_data->unsubmitted_oe_range_end;
8591         }
8592
8593         ret = btrfs_submit_direct_hook(dip);
8594         if (!ret)
8595                 return;
8596
8597         btrfs_io_bio_free_csum(io_bio);
8598
8599 free_ordered:
8600         /*
8601          * If we arrived here it means either we failed to submit the dip
8602          * or we either failed to clone the dio_bio or failed to allocate the
8603          * dip. If we cloned the dio_bio and allocated the dip, we can just
8604          * call bio_endio against our io_bio so that we get proper resource
8605          * cleanup if we fail to submit the dip, otherwise, we must do the
8606          * same as btrfs_endio_direct_[write|read] because we can't call these
8607          * callbacks - they require an allocated dip and a clone of dio_bio.
8608          */
8609         if (bio && dip) {
8610                 bio_io_error(bio);
8611                 /*
8612                  * The end io callbacks free our dip, do the final put on bio
8613                  * and all the cleanup and final put for dio_bio (through
8614                  * dio_end_io()).
8615                  */
8616                 dip = NULL;
8617                 bio = NULL;
8618         } else {
8619                 if (write)
8620                         __endio_write_update_ordered(inode,
8621                                                 file_offset,
8622                                                 dio_bio->bi_iter.bi_size,
8623                                                 false);
8624                 else
8625                         unlock_extent(&BTRFS_I(inode)->io_tree, file_offset,
8626                               file_offset + dio_bio->bi_iter.bi_size - 1);
8627
8628                 dio_bio->bi_status = BLK_STS_IOERR;
8629                 /*
8630                  * Releases and cleans up our dio_bio, no need to bio_put()
8631                  * nor bio_endio()/bio_io_error() against dio_bio.
8632                  */
8633                 dio_end_io(dio_bio);
8634         }
8635         if (bio)
8636                 bio_put(bio);
8637         kfree(dip);
8638 }
8639
8640 static ssize_t check_direct_IO(struct btrfs_fs_info *fs_info,
8641                                const struct iov_iter *iter, loff_t offset)
8642 {
8643         int seg;
8644         int i;
8645         unsigned int blocksize_mask = fs_info->sectorsize - 1;
8646         ssize_t retval = -EINVAL;
8647
8648         if (offset & blocksize_mask)
8649                 goto out;
8650
8651         if (iov_iter_alignment(iter) & blocksize_mask)
8652                 goto out;
8653
8654         /* If this is a write we don't need to check anymore */
8655         if (iov_iter_rw(iter) != READ || !iter_is_iovec(iter))
8656                 return 0;
8657         /*
8658          * Check to make sure we don't have duplicate iov_base's in this
8659          * iovec, if so return EINVAL, otherwise we'll get csum errors
8660          * when reading back.
8661          */
8662         for (seg = 0; seg < iter->nr_segs; seg++) {
8663                 for (i = seg + 1; i < iter->nr_segs; i++) {
8664                         if (iter->iov[seg].iov_base == iter->iov[i].iov_base)
8665                                 goto out;
8666                 }
8667         }
8668         retval = 0;
8669 out:
8670         return retval;
8671 }
8672
8673 static ssize_t btrfs_direct_IO(struct kiocb *iocb, struct iov_iter *iter)
8674 {
8675         struct file *file = iocb->ki_filp;
8676         struct inode *inode = file->f_mapping->host;
8677         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
8678         struct btrfs_dio_data dio_data = { 0 };
8679         struct extent_changeset *data_reserved = NULL;
8680         loff_t offset = iocb->ki_pos;
8681         size_t count = 0;
8682         int flags = 0;
8683         bool wakeup = true;
8684         bool relock = false;
8685         ssize_t ret;
8686
8687         if (check_direct_IO(fs_info, iter, offset))
8688                 return 0;
8689
8690         inode_dio_begin(inode);
8691
8692         /*
8693          * The generic stuff only does filemap_write_and_wait_range, which
8694          * isn't enough if we've written compressed pages to this area, so
8695          * we need to flush the dirty pages again to make absolutely sure
8696          * that any outstanding dirty pages are on disk.
8697          */
8698         count = iov_iter_count(iter);
8699         if (test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
8700                      &BTRFS_I(inode)->runtime_flags))
8701                 filemap_fdatawrite_range(inode->i_mapping, offset,
8702                                          offset + count - 1);
8703
8704         if (iov_iter_rw(iter) == WRITE) {
8705                 /*
8706                  * If the write DIO is beyond the EOF, we need update
8707                  * the isize, but it is protected by i_mutex. So we can
8708                  * not unlock the i_mutex at this case.
8709                  */
8710                 if (offset + count <= inode->i_size) {
8711                         dio_data.overwrite = 1;
8712                         inode_unlock(inode);
8713                         relock = true;
8714                 } else if (iocb->ki_flags & IOCB_NOWAIT) {
8715                         ret = -EAGAIN;
8716                         goto out;
8717                 }
8718                 ret = btrfs_delalloc_reserve_space(inode, &data_reserved,
8719                                                    offset, count);
8720                 if (ret)
8721                         goto out;
8722
8723                 /*
8724                  * We need to know how many extents we reserved so that we can
8725                  * do the accounting properly if we go over the number we
8726                  * originally calculated.  Abuse current->journal_info for this.
8727                  */
8728                 dio_data.reserve = round_up(count,
8729                                             fs_info->sectorsize);
8730                 dio_data.unsubmitted_oe_range_start = (u64)offset;
8731                 dio_data.unsubmitted_oe_range_end = (u64)offset;
8732                 current->journal_info = &dio_data;
8733                 down_read(&BTRFS_I(inode)->dio_sem);
8734         } else if (test_bit(BTRFS_INODE_READDIO_NEED_LOCK,
8735                                      &BTRFS_I(inode)->runtime_flags)) {
8736                 inode_dio_end(inode);
8737                 flags = DIO_LOCKING | DIO_SKIP_HOLES;
8738                 wakeup = false;
8739         }
8740
8741         ret = __blockdev_direct_IO(iocb, inode,
8742                                    fs_info->fs_devices->latest_bdev,
8743                                    iter, btrfs_get_blocks_direct, NULL,
8744                                    btrfs_submit_direct, flags);
8745         if (iov_iter_rw(iter) == WRITE) {
8746                 up_read(&BTRFS_I(inode)->dio_sem);
8747                 current->journal_info = NULL;
8748                 if (ret < 0 && ret != -EIOCBQUEUED) {
8749                         if (dio_data.reserve)
8750                                 btrfs_delalloc_release_space(inode, data_reserved,
8751                                         offset, dio_data.reserve, true);
8752                         /*
8753                          * On error we might have left some ordered extents
8754                          * without submitting corresponding bios for them, so
8755                          * cleanup them up to avoid other tasks getting them
8756                          * and waiting for them to complete forever.
8757                          */
8758                         if (dio_data.unsubmitted_oe_range_start <
8759                             dio_data.unsubmitted_oe_range_end)
8760                                 __endio_write_update_ordered(inode,
8761                                         dio_data.unsubmitted_oe_range_start,
8762                                         dio_data.unsubmitted_oe_range_end -
8763                                         dio_data.unsubmitted_oe_range_start,
8764                                         false);
8765                 } else if (ret >= 0 && (size_t)ret < count)
8766                         btrfs_delalloc_release_space(inode, data_reserved,
8767                                         offset, count - (size_t)ret, true);
8768                 btrfs_delalloc_release_extents(BTRFS_I(inode), count);
8769         }
8770 out:
8771         if (wakeup)
8772                 inode_dio_end(inode);
8773         if (relock)
8774                 inode_lock(inode);
8775
8776         extent_changeset_free(data_reserved);
8777         return ret;
8778 }
8779
8780 #define BTRFS_FIEMAP_FLAGS      (FIEMAP_FLAG_SYNC)
8781
8782 static int btrfs_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
8783                 __u64 start, __u64 len)
8784 {
8785         int     ret;
8786
8787         ret = fiemap_check_flags(fieinfo, BTRFS_FIEMAP_FLAGS);
8788         if (ret)
8789                 return ret;
8790
8791         return extent_fiemap(inode, fieinfo, start, len);
8792 }
8793
8794 int btrfs_readpage(struct file *file, struct page *page)
8795 {
8796         struct extent_io_tree *tree;
8797         tree = &BTRFS_I(page->mapping->host)->io_tree;
8798         return extent_read_full_page(tree, page, btrfs_get_extent, 0);
8799 }
8800
8801 static int btrfs_writepage(struct page *page, struct writeback_control *wbc)
8802 {
8803         struct inode *inode = page->mapping->host;
8804         int ret;
8805
8806         if (current->flags & PF_MEMALLOC) {
8807                 redirty_page_for_writepage(wbc, page);
8808                 unlock_page(page);
8809                 return 0;
8810         }
8811
8812         /*
8813          * If we are under memory pressure we will call this directly from the
8814          * VM, we need to make sure we have the inode referenced for the ordered
8815          * extent.  If not just return like we didn't do anything.
8816          */
8817         if (!igrab(inode)) {
8818                 redirty_page_for_writepage(wbc, page);
8819                 return AOP_WRITEPAGE_ACTIVATE;
8820         }
8821         ret = extent_write_full_page(page, wbc);
8822         btrfs_add_delayed_iput(inode);
8823         return ret;
8824 }
8825
8826 static int btrfs_writepages(struct address_space *mapping,
8827                             struct writeback_control *wbc)
8828 {
8829         return extent_writepages(mapping, wbc);
8830 }
8831
8832 static int
8833 btrfs_readpages(struct file *file, struct address_space *mapping,
8834                 struct list_head *pages, unsigned nr_pages)
8835 {
8836         return extent_readpages(mapping, pages, nr_pages);
8837 }
8838
8839 static int __btrfs_releasepage(struct page *page, gfp_t gfp_flags)
8840 {
8841         int ret = try_release_extent_mapping(page, gfp_flags);
8842         if (ret == 1) {
8843                 ClearPagePrivate(page);
8844                 set_page_private(page, 0);
8845                 put_page(page);
8846         }
8847         return ret;
8848 }
8849
8850 static int btrfs_releasepage(struct page *page, gfp_t gfp_flags)
8851 {
8852         if (PageWriteback(page) || PageDirty(page))
8853                 return 0;
8854         return __btrfs_releasepage(page, gfp_flags);
8855 }
8856
8857 static void btrfs_invalidatepage(struct page *page, unsigned int offset,
8858                                  unsigned int length)
8859 {
8860         struct inode *inode = page->mapping->host;
8861         struct extent_io_tree *tree;
8862         struct btrfs_ordered_extent *ordered;
8863         struct extent_state *cached_state = NULL;
8864         u64 page_start = page_offset(page);
8865         u64 page_end = page_start + PAGE_SIZE - 1;
8866         u64 start;
8867         u64 end;
8868         int inode_evicting = inode->i_state & I_FREEING;
8869
8870         /*
8871          * we have the page locked, so new writeback can't start,
8872          * and the dirty bit won't be cleared while we are here.
8873          *
8874          * Wait for IO on this page so that we can safely clear
8875          * the PagePrivate2 bit and do ordered accounting
8876          */
8877         wait_on_page_writeback(page);
8878
8879         tree = &BTRFS_I(inode)->io_tree;
8880         if (offset) {
8881                 btrfs_releasepage(page, GFP_NOFS);
8882                 return;
8883         }
8884
8885         if (!inode_evicting)
8886                 lock_extent_bits(tree, page_start, page_end, &cached_state);
8887 again:
8888         start = page_start;
8889         ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), start,
8890                                         page_end - start + 1);
8891         if (ordered) {
8892                 end = min(page_end, ordered->file_offset + ordered->len - 1);
8893                 /*
8894                  * IO on this page will never be started, so we need
8895                  * to account for any ordered extents now
8896                  */
8897                 if (!inode_evicting)
8898                         clear_extent_bit(tree, start, end,
8899                                          EXTENT_DELALLOC | EXTENT_DELALLOC_NEW |
8900                                          EXTENT_LOCKED | EXTENT_DO_ACCOUNTING |
8901                                          EXTENT_DEFRAG, 1, 0, &cached_state);
8902                 /*
8903                  * whoever cleared the private bit is responsible
8904                  * for the finish_ordered_io
8905                  */
8906                 if (TestClearPagePrivate2(page)) {
8907                         struct btrfs_ordered_inode_tree *tree;
8908                         u64 new_len;
8909
8910                         tree = &BTRFS_I(inode)->ordered_tree;
8911
8912                         spin_lock_irq(&tree->lock);
8913                         set_bit(BTRFS_ORDERED_TRUNCATED, &ordered->flags);
8914                         new_len = start - ordered->file_offset;
8915                         if (new_len < ordered->truncated_len)
8916                                 ordered->truncated_len = new_len;
8917                         spin_unlock_irq(&tree->lock);
8918
8919                         if (btrfs_dec_test_ordered_pending(inode, &ordered,
8920                                                            start,
8921                                                            end - start + 1, 1))
8922                                 btrfs_finish_ordered_io(ordered);
8923                 }
8924                 btrfs_put_ordered_extent(ordered);
8925                 if (!inode_evicting) {
8926                         cached_state = NULL;
8927                         lock_extent_bits(tree, start, end,
8928                                          &cached_state);
8929                 }
8930
8931                 start = end + 1;
8932                 if (start < page_end)
8933                         goto again;
8934         }
8935
8936         /*
8937          * Qgroup reserved space handler
8938          * Page here will be either
8939          * 1) Already written to disk
8940          *    In this case, its reserved space is released from data rsv map
8941          *    and will be freed by delayed_ref handler finally.
8942          *    So even we call qgroup_free_data(), it won't decrease reserved
8943          *    space.
8944          * 2) Not written to disk
8945          *    This means the reserved space should be freed here. However,
8946          *    if a truncate invalidates the page (by clearing PageDirty)
8947          *    and the page is accounted for while allocating extent
8948          *    in btrfs_check_data_free_space() we let delayed_ref to
8949          *    free the entire extent.
8950          */
8951         if (PageDirty(page))
8952                 btrfs_qgroup_free_data(inode, NULL, page_start, PAGE_SIZE);
8953         if (!inode_evicting) {
8954                 clear_extent_bit(tree, page_start, page_end, EXTENT_LOCKED |
8955                                  EXTENT_DELALLOC | EXTENT_DELALLOC_NEW |
8956                                  EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG, 1, 1,
8957                                  &cached_state);
8958
8959                 __btrfs_releasepage(page, GFP_NOFS);
8960         }
8961
8962         ClearPageChecked(page);
8963         if (PagePrivate(page)) {
8964                 ClearPagePrivate(page);
8965                 set_page_private(page, 0);
8966                 put_page(page);
8967         }
8968 }
8969
8970 /*
8971  * btrfs_page_mkwrite() is not allowed to change the file size as it gets
8972  * called from a page fault handler when a page is first dirtied. Hence we must
8973  * be careful to check for EOF conditions here. We set the page up correctly
8974  * for a written page which means we get ENOSPC checking when writing into
8975  * holes and correct delalloc and unwritten extent mapping on filesystems that
8976  * support these features.
8977  *
8978  * We are not allowed to take the i_mutex here so we have to play games to
8979  * protect against truncate races as the page could now be beyond EOF.  Because
8980  * truncate_setsize() writes the inode size before removing pages, once we have
8981  * the page lock we can determine safely if the page is beyond EOF. If it is not
8982  * beyond EOF, then the page is guaranteed safe against truncation until we
8983  * unlock the page.
8984  */
8985 vm_fault_t btrfs_page_mkwrite(struct vm_fault *vmf)
8986 {
8987         struct page *page = vmf->page;
8988         struct inode *inode = file_inode(vmf->vma->vm_file);
8989         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
8990         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
8991         struct btrfs_ordered_extent *ordered;
8992         struct extent_state *cached_state = NULL;
8993         struct extent_changeset *data_reserved = NULL;
8994         char *kaddr;
8995         unsigned long zero_start;
8996         loff_t size;
8997         vm_fault_t ret;
8998         int ret2;
8999         int reserved = 0;
9000         u64 reserved_space;
9001         u64 page_start;
9002         u64 page_end;
9003         u64 end;
9004
9005         reserved_space = PAGE_SIZE;
9006
9007         sb_start_pagefault(inode->i_sb);
9008         page_start = page_offset(page);
9009         page_end = page_start + PAGE_SIZE - 1;
9010         end = page_end;
9011
9012         /*
9013          * Reserving delalloc space after obtaining the page lock can lead to
9014          * deadlock. For example, if a dirty page is locked by this function
9015          * and the call to btrfs_delalloc_reserve_space() ends up triggering
9016          * dirty page write out, then the btrfs_writepage() function could
9017          * end up waiting indefinitely to get a lock on the page currently
9018          * being processed by btrfs_page_mkwrite() function.
9019          */
9020         ret2 = btrfs_delalloc_reserve_space(inode, &data_reserved, page_start,
9021                                            reserved_space);
9022         if (!ret2) {
9023                 ret2 = file_update_time(vmf->vma->vm_file);
9024                 reserved = 1;
9025         }
9026         if (ret2) {
9027                 ret = vmf_error(ret2);
9028                 if (reserved)
9029                         goto out;
9030                 goto out_noreserve;
9031         }
9032
9033         ret = VM_FAULT_NOPAGE; /* make the VM retry the fault */
9034 again:
9035         lock_page(page);
9036         size = i_size_read(inode);
9037
9038         if ((page->mapping != inode->i_mapping) ||
9039             (page_start >= size)) {
9040                 /* page got truncated out from underneath us */
9041                 goto out_unlock;
9042         }
9043         wait_on_page_writeback(page);
9044
9045         lock_extent_bits(io_tree, page_start, page_end, &cached_state);
9046         set_page_extent_mapped(page);
9047
9048         /*
9049          * we can't set the delalloc bits if there are pending ordered
9050          * extents.  Drop our locks and wait for them to finish
9051          */
9052         ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), page_start,
9053                         PAGE_SIZE);
9054         if (ordered) {
9055                 unlock_extent_cached(io_tree, page_start, page_end,
9056                                      &cached_state);
9057                 unlock_page(page);
9058                 btrfs_start_ordered_extent(inode, ordered, 1);
9059                 btrfs_put_ordered_extent(ordered);
9060                 goto again;
9061         }
9062
9063         if (page->index == ((size - 1) >> PAGE_SHIFT)) {
9064                 reserved_space = round_up(size - page_start,
9065                                           fs_info->sectorsize);
9066                 if (reserved_space < PAGE_SIZE) {
9067                         end = page_start + reserved_space - 1;
9068                         btrfs_delalloc_release_space(inode, data_reserved,
9069                                         page_start, PAGE_SIZE - reserved_space,
9070                                         true);
9071                 }
9072         }
9073
9074         /*
9075          * page_mkwrite gets called when the page is firstly dirtied after it's
9076          * faulted in, but write(2) could also dirty a page and set delalloc
9077          * bits, thus in this case for space account reason, we still need to
9078          * clear any delalloc bits within this page range since we have to
9079          * reserve data&meta space before lock_page() (see above comments).
9080          */
9081         clear_extent_bit(&BTRFS_I(inode)->io_tree, page_start, end,
9082                           EXTENT_DELALLOC | EXTENT_DO_ACCOUNTING |
9083                           EXTENT_DEFRAG, 0, 0, &cached_state);
9084
9085         ret2 = btrfs_set_extent_delalloc(inode, page_start, end, 0,
9086                                         &cached_state);
9087         if (ret2) {
9088                 unlock_extent_cached(io_tree, page_start, page_end,
9089                                      &cached_state);
9090                 ret = VM_FAULT_SIGBUS;
9091                 goto out_unlock;
9092         }
9093         ret2 = 0;
9094
9095         /* page is wholly or partially inside EOF */
9096         if (page_start + PAGE_SIZE > size)
9097                 zero_start = offset_in_page(size);
9098         else
9099                 zero_start = PAGE_SIZE;
9100
9101         if (zero_start != PAGE_SIZE) {
9102                 kaddr = kmap(page);
9103                 memset(kaddr + zero_start, 0, PAGE_SIZE - zero_start);
9104                 flush_dcache_page(page);
9105                 kunmap(page);
9106         }
9107         ClearPageChecked(page);
9108         set_page_dirty(page);
9109         SetPageUptodate(page);
9110
9111         BTRFS_I(inode)->last_trans = fs_info->generation;
9112         BTRFS_I(inode)->last_sub_trans = BTRFS_I(inode)->root->log_transid;
9113         BTRFS_I(inode)->last_log_commit = BTRFS_I(inode)->root->last_log_commit;
9114
9115         unlock_extent_cached(io_tree, page_start, page_end, &cached_state);
9116
9117         if (!ret2) {
9118                 btrfs_delalloc_release_extents(BTRFS_I(inode), PAGE_SIZE);
9119                 sb_end_pagefault(inode->i_sb);
9120                 extent_changeset_free(data_reserved);
9121                 return VM_FAULT_LOCKED;
9122         }
9123
9124 out_unlock:
9125         unlock_page(page);
9126 out:
9127         btrfs_delalloc_release_extents(BTRFS_I(inode), PAGE_SIZE);
9128         btrfs_delalloc_release_space(inode, data_reserved, page_start,
9129                                      reserved_space, (ret != 0));
9130 out_noreserve:
9131         sb_end_pagefault(inode->i_sb);
9132         extent_changeset_free(data_reserved);
9133         return ret;
9134 }
9135
9136 static int btrfs_truncate(struct inode *inode, bool skip_writeback)
9137 {
9138         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
9139         struct btrfs_root *root = BTRFS_I(inode)->root;
9140         struct btrfs_block_rsv *rsv;
9141         int ret;
9142         struct btrfs_trans_handle *trans;
9143         u64 mask = fs_info->sectorsize - 1;
9144         u64 min_size = btrfs_calc_metadata_size(fs_info, 1);
9145
9146         if (!skip_writeback) {
9147                 ret = btrfs_wait_ordered_range(inode, inode->i_size & (~mask),
9148                                                (u64)-1);
9149                 if (ret)
9150                         return ret;
9151         }
9152
9153         /*
9154          * Yes ladies and gentlemen, this is indeed ugly.  We have a couple of
9155          * things going on here:
9156          *
9157          * 1) We need to reserve space to update our inode.
9158          *
9159          * 2) We need to have something to cache all the space that is going to
9160          * be free'd up by the truncate operation, but also have some slack
9161          * space reserved in case it uses space during the truncate (thank you
9162          * very much snapshotting).
9163          *
9164          * And we need these to be separate.  The fact is we can use a lot of
9165          * space doing the truncate, and we have no earthly idea how much space
9166          * we will use, so we need the truncate reservation to be separate so it
9167          * doesn't end up using space reserved for updating the inode.  We also
9168          * need to be able to stop the transaction and start a new one, which
9169          * means we need to be able to update the inode several times, and we
9170          * have no idea of knowing how many times that will be, so we can't just
9171          * reserve 1 item for the entirety of the operation, so that has to be
9172          * done separately as well.
9173          *
9174          * So that leaves us with
9175          *
9176          * 1) rsv - for the truncate reservation, which we will steal from the
9177          * transaction reservation.
9178          * 2) fs_info->trans_block_rsv - this will have 1 items worth left for
9179          * updating the inode.
9180          */
9181         rsv = btrfs_alloc_block_rsv(fs_info, BTRFS_BLOCK_RSV_TEMP);
9182         if (!rsv)
9183                 return -ENOMEM;
9184         rsv->size = min_size;
9185         rsv->failfast = 1;
9186
9187         /*
9188          * 1 for the truncate slack space
9189          * 1 for updating the inode.
9190          */
9191         trans = btrfs_start_transaction(root, 2);
9192         if (IS_ERR(trans)) {
9193                 ret = PTR_ERR(trans);
9194                 goto out;
9195         }
9196
9197         /* Migrate the slack space for the truncate to our reserve */
9198         ret = btrfs_block_rsv_migrate(&fs_info->trans_block_rsv, rsv,
9199                                       min_size, false);
9200         BUG_ON(ret);
9201
9202         /*
9203          * So if we truncate and then write and fsync we normally would just
9204          * write the extents that changed, which is a problem if we need to
9205          * first truncate that entire inode.  So set this flag so we write out
9206          * all of the extents in the inode to the sync log so we're completely
9207          * safe.
9208          */
9209         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &BTRFS_I(inode)->runtime_flags);
9210         trans->block_rsv = rsv;
9211
9212         while (1) {
9213                 ret = btrfs_truncate_inode_items(trans, root, inode,
9214                                                  inode->i_size,
9215                                                  BTRFS_EXTENT_DATA_KEY);
9216                 trans->block_rsv = &fs_info->trans_block_rsv;
9217                 if (ret != -ENOSPC && ret != -EAGAIN)
9218                         break;
9219
9220                 ret = btrfs_update_inode(trans, root, inode);
9221                 if (ret)
9222                         break;
9223
9224                 btrfs_end_transaction(trans);
9225                 btrfs_btree_balance_dirty(fs_info);
9226
9227                 trans = btrfs_start_transaction(root, 2);
9228                 if (IS_ERR(trans)) {
9229                         ret = PTR_ERR(trans);
9230                         trans = NULL;
9231                         break;
9232                 }
9233
9234                 btrfs_block_rsv_release(fs_info, rsv, -1);
9235                 ret = btrfs_block_rsv_migrate(&fs_info->trans_block_rsv,
9236                                               rsv, min_size, false);
9237                 BUG_ON(ret);    /* shouldn't happen */
9238                 trans->block_rsv = rsv;
9239         }
9240
9241         /*
9242          * We can't call btrfs_truncate_block inside a trans handle as we could
9243          * deadlock with freeze, if we got NEED_TRUNCATE_BLOCK then we know
9244          * we've truncated everything except the last little bit, and can do
9245          * btrfs_truncate_block and then update the disk_i_size.
9246          */
9247         if (ret == NEED_TRUNCATE_BLOCK) {
9248                 btrfs_end_transaction(trans);
9249                 btrfs_btree_balance_dirty(fs_info);
9250
9251                 ret = btrfs_truncate_block(inode, inode->i_size, 0, 0);
9252                 if (ret)
9253                         goto out;
9254                 trans = btrfs_start_transaction(root, 1);
9255                 if (IS_ERR(trans)) {
9256                         ret = PTR_ERR(trans);
9257                         goto out;
9258                 }
9259                 btrfs_ordered_update_i_size(inode, inode->i_size, NULL);
9260         }
9261
9262         if (trans) {
9263                 int ret2;
9264
9265                 trans->block_rsv = &fs_info->trans_block_rsv;
9266                 ret2 = btrfs_update_inode(trans, root, inode);
9267                 if (ret2 && !ret)
9268                         ret = ret2;
9269
9270                 ret2 = btrfs_end_transaction(trans);
9271                 if (ret2 && !ret)
9272                         ret = ret2;
9273                 btrfs_btree_balance_dirty(fs_info);
9274         }
9275 out:
9276         btrfs_free_block_rsv(fs_info, rsv);
9277
9278         return ret;
9279 }
9280
9281 /*
9282  * create a new subvolume directory/inode (helper for the ioctl).
9283  */
9284 int btrfs_create_subvol_root(struct btrfs_trans_handle *trans,
9285                              struct btrfs_root *new_root,
9286                              struct btrfs_root *parent_root,
9287                              u64 new_dirid)
9288 {
9289         struct inode *inode;
9290         int err;
9291         u64 index = 0;
9292
9293         inode = btrfs_new_inode(trans, new_root, NULL, "..", 2,
9294                                 new_dirid, new_dirid,
9295                                 S_IFDIR | (~current_umask() & S_IRWXUGO),
9296                                 &index);
9297         if (IS_ERR(inode))
9298                 return PTR_ERR(inode);
9299         inode->i_op = &btrfs_dir_inode_operations;
9300         inode->i_fop = &btrfs_dir_file_operations;
9301
9302         set_nlink(inode, 1);
9303         btrfs_i_size_write(BTRFS_I(inode), 0);
9304         unlock_new_inode(inode);
9305
9306         err = btrfs_subvol_inherit_props(trans, new_root, parent_root);
9307         if (err)
9308                 btrfs_err(new_root->fs_info,
9309                           "error inheriting subvolume %llu properties: %d",
9310                           new_root->root_key.objectid, err);
9311
9312         err = btrfs_update_inode(trans, new_root, inode);
9313
9314         iput(inode);
9315         return err;
9316 }
9317
9318 struct inode *btrfs_alloc_inode(struct super_block *sb)
9319 {
9320         struct btrfs_fs_info *fs_info = btrfs_sb(sb);
9321         struct btrfs_inode *ei;
9322         struct inode *inode;
9323
9324         ei = kmem_cache_alloc(btrfs_inode_cachep, GFP_KERNEL);
9325         if (!ei)
9326                 return NULL;
9327
9328         ei->root = NULL;
9329         ei->generation = 0;
9330         ei->last_trans = 0;
9331         ei->last_sub_trans = 0;
9332         ei->logged_trans = 0;
9333         ei->delalloc_bytes = 0;
9334         ei->new_delalloc_bytes = 0;
9335         ei->defrag_bytes = 0;
9336         ei->disk_i_size = 0;
9337         ei->flags = 0;
9338         ei->csum_bytes = 0;
9339         ei->index_cnt = (u64)-1;
9340         ei->dir_index = 0;
9341         ei->last_unlink_trans = 0;
9342         ei->last_log_commit = 0;
9343
9344         spin_lock_init(&ei->lock);
9345         ei->outstanding_extents = 0;
9346         if (sb->s_magic != BTRFS_TEST_MAGIC)
9347                 btrfs_init_metadata_block_rsv(fs_info, &ei->block_rsv,
9348                                               BTRFS_BLOCK_RSV_DELALLOC);
9349         ei->runtime_flags = 0;
9350         ei->prop_compress = BTRFS_COMPRESS_NONE;
9351         ei->defrag_compress = BTRFS_COMPRESS_NONE;
9352
9353         ei->delayed_node = NULL;
9354
9355         ei->i_otime.tv_sec = 0;
9356         ei->i_otime.tv_nsec = 0;
9357
9358         inode = &ei->vfs_inode;
9359         extent_map_tree_init(&ei->extent_tree);
9360         extent_io_tree_init(fs_info, &ei->io_tree, IO_TREE_INODE_IO, inode);
9361         extent_io_tree_init(fs_info, &ei->io_failure_tree,
9362                             IO_TREE_INODE_IO_FAILURE, inode);
9363         ei->io_tree.track_uptodate = true;
9364         ei->io_failure_tree.track_uptodate = true;
9365         atomic_set(&ei->sync_writers, 0);
9366         mutex_init(&ei->log_mutex);
9367         btrfs_ordered_inode_tree_init(&ei->ordered_tree);
9368         INIT_LIST_HEAD(&ei->delalloc_inodes);
9369         INIT_LIST_HEAD(&ei->delayed_iput);
9370         RB_CLEAR_NODE(&ei->rb_node);
9371         init_rwsem(&ei->dio_sem);
9372
9373         return inode;
9374 }
9375
9376 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
9377 void btrfs_test_destroy_inode(struct inode *inode)
9378 {
9379         btrfs_drop_extent_cache(BTRFS_I(inode), 0, (u64)-1, 0);
9380         kmem_cache_free(btrfs_inode_cachep, BTRFS_I(inode));
9381 }
9382 #endif
9383
9384 void btrfs_free_inode(struct inode *inode)
9385 {
9386         kmem_cache_free(btrfs_inode_cachep, BTRFS_I(inode));
9387 }
9388
9389 void btrfs_destroy_inode(struct inode *inode)
9390 {
9391         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
9392         struct btrfs_ordered_extent *ordered;
9393         struct btrfs_root *root = BTRFS_I(inode)->root;
9394
9395         WARN_ON(!hlist_empty(&inode->i_dentry));
9396         WARN_ON(inode->i_data.nrpages);
9397         WARN_ON(BTRFS_I(inode)->block_rsv.reserved);
9398         WARN_ON(BTRFS_I(inode)->block_rsv.size);
9399         WARN_ON(BTRFS_I(inode)->outstanding_extents);
9400         WARN_ON(BTRFS_I(inode)->delalloc_bytes);
9401         WARN_ON(BTRFS_I(inode)->new_delalloc_bytes);
9402         WARN_ON(BTRFS_I(inode)->csum_bytes);
9403         WARN_ON(BTRFS_I(inode)->defrag_bytes);
9404
9405         /*
9406          * This can happen where we create an inode, but somebody else also
9407          * created the same inode and we need to destroy the one we already
9408          * created.
9409          */
9410         if (!root)
9411                 return;
9412
9413         while (1) {
9414                 ordered = btrfs_lookup_first_ordered_extent(inode, (u64)-1);
9415                 if (!ordered)
9416                         break;
9417                 else {
9418                         btrfs_err(fs_info,
9419                                   "found ordered extent %llu %llu on inode cleanup",
9420                                   ordered->file_offset, ordered->len);
9421                         btrfs_remove_ordered_extent(inode, ordered);
9422                         btrfs_put_ordered_extent(ordered);
9423                         btrfs_put_ordered_extent(ordered);
9424                 }
9425         }
9426         btrfs_qgroup_check_reserved_leak(inode);
9427         inode_tree_del(inode);
9428         btrfs_drop_extent_cache(BTRFS_I(inode), 0, (u64)-1, 0);
9429 }
9430
9431 int btrfs_drop_inode(struct inode *inode)
9432 {
9433         struct btrfs_root *root = BTRFS_I(inode)->root;
9434
9435         if (root == NULL)
9436                 return 1;
9437
9438         /* the snap/subvol tree is on deleting */
9439         if (btrfs_root_refs(&root->root_item) == 0)
9440                 return 1;
9441         else
9442                 return generic_drop_inode(inode);
9443 }
9444
9445 static void init_once(void *foo)
9446 {
9447         struct btrfs_inode *ei = (struct btrfs_inode *) foo;
9448
9449         inode_init_once(&ei->vfs_inode);
9450 }
9451
9452 void __cold btrfs_destroy_cachep(void)
9453 {
9454         /*
9455          * Make sure all delayed rcu free inodes are flushed before we
9456          * destroy cache.
9457          */
9458         rcu_barrier();
9459         kmem_cache_destroy(btrfs_inode_cachep);
9460         kmem_cache_destroy(btrfs_trans_handle_cachep);
9461         kmem_cache_destroy(btrfs_path_cachep);
9462         kmem_cache_destroy(btrfs_free_space_cachep);
9463         kmem_cache_destroy(btrfs_free_space_bitmap_cachep);
9464 }
9465
9466 int __init btrfs_init_cachep(void)
9467 {
9468         btrfs_inode_cachep = kmem_cache_create("btrfs_inode",
9469                         sizeof(struct btrfs_inode), 0,
9470                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD | SLAB_ACCOUNT,
9471                         init_once);
9472         if (!btrfs_inode_cachep)
9473                 goto fail;
9474
9475         btrfs_trans_handle_cachep = kmem_cache_create("btrfs_trans_handle",
9476                         sizeof(struct btrfs_trans_handle), 0,
9477                         SLAB_TEMPORARY | SLAB_MEM_SPREAD, NULL);
9478         if (!btrfs_trans_handle_cachep)
9479                 goto fail;
9480
9481         btrfs_path_cachep = kmem_cache_create("btrfs_path",
9482                         sizeof(struct btrfs_path), 0,
9483                         SLAB_MEM_SPREAD, NULL);
9484         if (!btrfs_path_cachep)
9485                 goto fail;
9486
9487         btrfs_free_space_cachep = kmem_cache_create("btrfs_free_space",
9488                         sizeof(struct btrfs_free_space), 0,
9489                         SLAB_MEM_SPREAD, NULL);
9490         if (!btrfs_free_space_cachep)
9491                 goto fail;
9492
9493         btrfs_free_space_bitmap_cachep = kmem_cache_create("btrfs_free_space_bitmap",
9494                                                         PAGE_SIZE, PAGE_SIZE,
9495                                                         SLAB_RED_ZONE, NULL);
9496         if (!btrfs_free_space_bitmap_cachep)
9497                 goto fail;
9498
9499         return 0;
9500 fail:
9501         btrfs_destroy_cachep();
9502         return -ENOMEM;
9503 }
9504
9505 static int btrfs_getattr(const struct path *path, struct kstat *stat,
9506                          u32 request_mask, unsigned int flags)
9507 {
9508         u64 delalloc_bytes;
9509         struct inode *inode = d_inode(path->dentry);
9510         u32 blocksize = inode->i_sb->s_blocksize;
9511         u32 bi_flags = BTRFS_I(inode)->flags;
9512
9513         stat->result_mask |= STATX_BTIME;
9514         stat->btime.tv_sec = BTRFS_I(inode)->i_otime.tv_sec;
9515         stat->btime.tv_nsec = BTRFS_I(inode)->i_otime.tv_nsec;
9516         if (bi_flags & BTRFS_INODE_APPEND)
9517                 stat->attributes |= STATX_ATTR_APPEND;
9518         if (bi_flags & BTRFS_INODE_COMPRESS)
9519                 stat->attributes |= STATX_ATTR_COMPRESSED;
9520         if (bi_flags & BTRFS_INODE_IMMUTABLE)
9521                 stat->attributes |= STATX_ATTR_IMMUTABLE;
9522         if (bi_flags & BTRFS_INODE_NODUMP)
9523                 stat->attributes |= STATX_ATTR_NODUMP;
9524
9525         stat->attributes_mask |= (STATX_ATTR_APPEND |
9526                                   STATX_ATTR_COMPRESSED |
9527                                   STATX_ATTR_IMMUTABLE |
9528                                   STATX_ATTR_NODUMP);
9529
9530         generic_fillattr(inode, stat);
9531         stat->dev = BTRFS_I(inode)->root->anon_dev;
9532
9533         spin_lock(&BTRFS_I(inode)->lock);
9534         delalloc_bytes = BTRFS_I(inode)->new_delalloc_bytes;
9535         spin_unlock(&BTRFS_I(inode)->lock);
9536         stat->blocks = (ALIGN(inode_get_bytes(inode), blocksize) +
9537                         ALIGN(delalloc_bytes, blocksize)) >> 9;
9538         return 0;
9539 }
9540
9541 static int btrfs_rename_exchange(struct inode *old_dir,
9542                               struct dentry *old_dentry,
9543                               struct inode *new_dir,
9544                               struct dentry *new_dentry)
9545 {
9546         struct btrfs_fs_info *fs_info = btrfs_sb(old_dir->i_sb);
9547         struct btrfs_trans_handle *trans;
9548         struct btrfs_root *root = BTRFS_I(old_dir)->root;
9549         struct btrfs_root *dest = BTRFS_I(new_dir)->root;
9550         struct inode *new_inode = new_dentry->d_inode;
9551         struct inode *old_inode = old_dentry->d_inode;
9552         struct timespec64 ctime = current_time(old_inode);
9553         struct dentry *parent;
9554         u64 old_ino = btrfs_ino(BTRFS_I(old_inode));
9555         u64 new_ino = btrfs_ino(BTRFS_I(new_inode));
9556         u64 old_idx = 0;
9557         u64 new_idx = 0;
9558         int ret;
9559         bool root_log_pinned = false;
9560         bool dest_log_pinned = false;
9561         struct btrfs_log_ctx ctx_root;
9562         struct btrfs_log_ctx ctx_dest;
9563         bool sync_log_root = false;
9564         bool sync_log_dest = false;
9565         bool commit_transaction = false;
9566
9567         /* we only allow rename subvolume link between subvolumes */
9568         if (old_ino != BTRFS_FIRST_FREE_OBJECTID && root != dest)
9569                 return -EXDEV;
9570
9571         btrfs_init_log_ctx(&ctx_root, old_inode);
9572         btrfs_init_log_ctx(&ctx_dest, new_inode);
9573
9574         /* close the race window with snapshot create/destroy ioctl */
9575         if (old_ino == BTRFS_FIRST_FREE_OBJECTID ||
9576             new_ino == BTRFS_FIRST_FREE_OBJECTID)
9577                 down_read(&fs_info->subvol_sem);
9578
9579         /*
9580          * We want to reserve the absolute worst case amount of items.  So if
9581          * both inodes are subvols and we need to unlink them then that would
9582          * require 4 item modifications, but if they are both normal inodes it
9583          * would require 5 item modifications, so we'll assume their normal
9584          * inodes.  So 5 * 2 is 10, plus 2 for the new links, so 12 total items
9585          * should cover the worst case number of items we'll modify.
9586          */
9587         trans = btrfs_start_transaction(root, 12);
9588         if (IS_ERR(trans)) {
9589                 ret = PTR_ERR(trans);
9590                 goto out_notrans;
9591         }
9592
9593         if (dest != root)
9594                 btrfs_record_root_in_trans(trans, dest);
9595
9596         /*
9597          * We need to find a free sequence number both in the source and
9598          * in the destination directory for the exchange.
9599          */
9600         ret = btrfs_set_inode_index(BTRFS_I(new_dir), &old_idx);
9601         if (ret)
9602                 goto out_fail;
9603         ret = btrfs_set_inode_index(BTRFS_I(old_dir), &new_idx);
9604         if (ret)
9605                 goto out_fail;
9606
9607         BTRFS_I(old_inode)->dir_index = 0ULL;
9608         BTRFS_I(new_inode)->dir_index = 0ULL;
9609
9610         /* Reference for the source. */
9611         if (old_ino == BTRFS_FIRST_FREE_OBJECTID) {
9612                 /* force full log commit if subvolume involved. */
9613                 btrfs_set_log_full_commit(trans);
9614         } else {
9615                 btrfs_pin_log_trans(root);
9616                 root_log_pinned = true;
9617                 ret = btrfs_insert_inode_ref(trans, dest,
9618                                              new_dentry->d_name.name,
9619                                              new_dentry->d_name.len,
9620                                              old_ino,
9621                                              btrfs_ino(BTRFS_I(new_dir)),
9622                                              old_idx);
9623                 if (ret)
9624                         goto out_fail;
9625         }
9626
9627         /* And now for the dest. */
9628         if (new_ino == BTRFS_FIRST_FREE_OBJECTID) {
9629                 /* force full log commit if subvolume involved. */
9630                 btrfs_set_log_full_commit(trans);
9631         } else {
9632                 btrfs_pin_log_trans(dest);
9633                 dest_log_pinned = true;
9634                 ret = btrfs_insert_inode_ref(trans, root,
9635                                              old_dentry->d_name.name,
9636                                              old_dentry->d_name.len,
9637                                              new_ino,
9638                                              btrfs_ino(BTRFS_I(old_dir)),
9639                                              new_idx);
9640                 if (ret)
9641                         goto out_fail;
9642         }
9643
9644         /* Update inode version and ctime/mtime. */
9645         inode_inc_iversion(old_dir);
9646         inode_inc_iversion(new_dir);
9647         inode_inc_iversion(old_inode);
9648         inode_inc_iversion(new_inode);
9649         old_dir->i_ctime = old_dir->i_mtime = ctime;
9650         new_dir->i_ctime = new_dir->i_mtime = ctime;
9651         old_inode->i_ctime = ctime;
9652         new_inode->i_ctime = ctime;
9653
9654         if (old_dentry->d_parent != new_dentry->d_parent) {
9655                 btrfs_record_unlink_dir(trans, BTRFS_I(old_dir),
9656                                 BTRFS_I(old_inode), 1);
9657                 btrfs_record_unlink_dir(trans, BTRFS_I(new_dir),
9658                                 BTRFS_I(new_inode), 1);
9659         }
9660
9661         /* src is a subvolume */
9662         if (old_ino == BTRFS_FIRST_FREE_OBJECTID) {
9663                 ret = btrfs_unlink_subvol(trans, old_dir, old_dentry);
9664         } else { /* src is an inode */
9665                 ret = __btrfs_unlink_inode(trans, root, BTRFS_I(old_dir),
9666                                            BTRFS_I(old_dentry->d_inode),
9667                                            old_dentry->d_name.name,
9668                                            old_dentry->d_name.len);
9669                 if (!ret)
9670                         ret = btrfs_update_inode(trans, root, old_inode);
9671         }
9672         if (ret) {
9673                 btrfs_abort_transaction(trans, ret);
9674                 goto out_fail;
9675         }
9676
9677         /* dest is a subvolume */
9678         if (new_ino == BTRFS_FIRST_FREE_OBJECTID) {
9679                 ret = btrfs_unlink_subvol(trans, new_dir, new_dentry);
9680         } else { /* dest is an inode */
9681                 ret = __btrfs_unlink_inode(trans, dest, BTRFS_I(new_dir),
9682                                            BTRFS_I(new_dentry->d_inode),
9683                                            new_dentry->d_name.name,
9684                                            new_dentry->d_name.len);
9685                 if (!ret)
9686                         ret = btrfs_update_inode(trans, dest, new_inode);
9687         }
9688         if (ret) {
9689                 btrfs_abort_transaction(trans, ret);
9690                 goto out_fail;
9691         }
9692
9693         ret = btrfs_add_link(trans, BTRFS_I(new_dir), BTRFS_I(old_inode),
9694                              new_dentry->d_name.name,
9695                              new_dentry->d_name.len, 0, old_idx);
9696         if (ret) {
9697                 btrfs_abort_transaction(trans, ret);
9698                 goto out_fail;
9699         }
9700
9701         ret = btrfs_add_link(trans, BTRFS_I(old_dir), BTRFS_I(new_inode),
9702                              old_dentry->d_name.name,
9703                              old_dentry->d_name.len, 0, new_idx);
9704         if (ret) {
9705                 btrfs_abort_transaction(trans, ret);
9706                 goto out_fail;
9707         }
9708
9709         if (old_inode->i_nlink == 1)
9710                 BTRFS_I(old_inode)->dir_index = old_idx;
9711         if (new_inode->i_nlink == 1)
9712                 BTRFS_I(new_inode)->dir_index = new_idx;
9713
9714         if (root_log_pinned) {
9715                 parent = new_dentry->d_parent;
9716                 ret = btrfs_log_new_name(trans, BTRFS_I(old_inode),
9717                                          BTRFS_I(old_dir), parent,
9718                                          false, &ctx_root);
9719                 if (ret == BTRFS_NEED_LOG_SYNC)
9720                         sync_log_root = true;
9721                 else if (ret == BTRFS_NEED_TRANS_COMMIT)
9722                         commit_transaction = true;
9723                 ret = 0;
9724                 btrfs_end_log_trans(root);
9725                 root_log_pinned = false;
9726         }
9727         if (dest_log_pinned) {
9728                 if (!commit_transaction) {
9729                         parent = old_dentry->d_parent;
9730                         ret = btrfs_log_new_name(trans, BTRFS_I(new_inode),
9731                                                  BTRFS_I(new_dir), parent,
9732                                                  false, &ctx_dest);
9733                         if (ret == BTRFS_NEED_LOG_SYNC)
9734                                 sync_log_dest = true;
9735                         else if (ret == BTRFS_NEED_TRANS_COMMIT)
9736                                 commit_transaction = true;
9737                         ret = 0;
9738                 }
9739                 btrfs_end_log_trans(dest);
9740                 dest_log_pinned = false;
9741         }
9742 out_fail:
9743         /*
9744          * If we have pinned a log and an error happened, we unpin tasks
9745          * trying to sync the log and force them to fallback to a transaction
9746          * commit if the log currently contains any of the inodes involved in
9747          * this rename operation (to ensure we do not persist a log with an
9748          * inconsistent state for any of these inodes or leading to any
9749          * inconsistencies when replayed). If the transaction was aborted, the
9750          * abortion reason is propagated to userspace when attempting to commit
9751          * the transaction. If the log does not contain any of these inodes, we
9752          * allow the tasks to sync it.
9753          */
9754         if (ret && (root_log_pinned || dest_log_pinned)) {
9755                 if (btrfs_inode_in_log(BTRFS_I(old_dir), fs_info->generation) ||
9756                     btrfs_inode_in_log(BTRFS_I(new_dir), fs_info->generation) ||
9757                     btrfs_inode_in_log(BTRFS_I(old_inode), fs_info->generation) ||
9758                     (new_inode &&
9759                      btrfs_inode_in_log(BTRFS_I(new_inode), fs_info->generation)))
9760                         btrfs_set_log_full_commit(trans);
9761
9762                 if (root_log_pinned) {
9763                         btrfs_end_log_trans(root);
9764                         root_log_pinned = false;
9765                 }
9766                 if (dest_log_pinned) {
9767                         btrfs_end_log_trans(dest);
9768                         dest_log_pinned = false;
9769                 }
9770         }
9771         if (!ret && sync_log_root && !commit_transaction) {
9772                 ret = btrfs_sync_log(trans, BTRFS_I(old_inode)->root,
9773                                      &ctx_root);
9774                 if (ret)
9775                         commit_transaction = true;
9776         }
9777         if (!ret && sync_log_dest && !commit_transaction) {
9778                 ret = btrfs_sync_log(trans, BTRFS_I(new_inode)->root,
9779                                      &ctx_dest);
9780                 if (ret)
9781                         commit_transaction = true;
9782         }
9783         if (commit_transaction) {
9784                 /*
9785                  * We may have set commit_transaction when logging the new name
9786                  * in the destination root, in which case we left the source
9787                  * root context in the list of log contextes. So make sure we
9788                  * remove it to avoid invalid memory accesses, since the context
9789                  * was allocated in our stack frame.
9790                  */
9791                 if (sync_log_root) {
9792                         mutex_lock(&root->log_mutex);
9793                         list_del_init(&ctx_root.list);
9794                         mutex_unlock(&root->log_mutex);
9795                 }
9796                 ret = btrfs_commit_transaction(trans);
9797         } else {
9798                 int ret2;
9799
9800                 ret2 = btrfs_end_transaction(trans);
9801                 ret = ret ? ret : ret2;
9802         }
9803 out_notrans:
9804         if (new_ino == BTRFS_FIRST_FREE_OBJECTID ||
9805             old_ino == BTRFS_FIRST_FREE_OBJECTID)
9806                 up_read(&fs_info->subvol_sem);
9807
9808         ASSERT(list_empty(&ctx_root.list));
9809         ASSERT(list_empty(&ctx_dest.list));
9810
9811         return ret;
9812 }
9813
9814 static int btrfs_whiteout_for_rename(struct btrfs_trans_handle *trans,
9815                                      struct btrfs_root *root,
9816                                      struct inode *dir,
9817                                      struct dentry *dentry)
9818 {
9819         int ret;
9820         struct inode *inode;
9821         u64 objectid;
9822         u64 index;
9823
9824         ret = btrfs_find_free_ino(root, &objectid);
9825         if (ret)
9826                 return ret;
9827
9828         inode = btrfs_new_inode(trans, root, dir,
9829                                 dentry->d_name.name,
9830                                 dentry->d_name.len,
9831                                 btrfs_ino(BTRFS_I(dir)),
9832                                 objectid,
9833                                 S_IFCHR | WHITEOUT_MODE,
9834                                 &index);
9835
9836         if (IS_ERR(inode)) {
9837                 ret = PTR_ERR(inode);
9838                 return ret;
9839         }
9840
9841         inode->i_op = &btrfs_special_inode_operations;
9842         init_special_inode(inode, inode->i_mode,
9843                 WHITEOUT_DEV);
9844
9845         ret = btrfs_init_inode_security(trans, inode, dir,
9846                                 &dentry->d_name);
9847         if (ret)
9848                 goto out;
9849
9850         ret = btrfs_add_nondir(trans, BTRFS_I(dir), dentry,
9851                                 BTRFS_I(inode), 0, index);
9852         if (ret)
9853                 goto out;
9854
9855         ret = btrfs_update_inode(trans, root, inode);
9856 out:
9857         unlock_new_inode(inode);
9858         if (ret)
9859                 inode_dec_link_count(inode);
9860         iput(inode);
9861
9862         return ret;
9863 }
9864
9865 static int btrfs_rename(struct inode *old_dir, struct dentry *old_dentry,
9866                            struct inode *new_dir, struct dentry *new_dentry,
9867                            unsigned int flags)
9868 {
9869         struct btrfs_fs_info *fs_info = btrfs_sb(old_dir->i_sb);
9870         struct btrfs_trans_handle *trans;
9871         unsigned int trans_num_items;
9872         struct btrfs_root *root = BTRFS_I(old_dir)->root;
9873         struct btrfs_root *dest = BTRFS_I(new_dir)->root;
9874         struct inode *new_inode = d_inode(new_dentry);
9875         struct inode *old_inode = d_inode(old_dentry);
9876         u64 index = 0;
9877         int ret;
9878         u64 old_ino = btrfs_ino(BTRFS_I(old_inode));
9879         bool log_pinned = false;
9880         struct btrfs_log_ctx ctx;
9881         bool sync_log = false;
9882         bool commit_transaction = false;
9883
9884         if (btrfs_ino(BTRFS_I(new_dir)) == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)
9885                 return -EPERM;
9886
9887         /* we only allow rename subvolume link between subvolumes */
9888         if (old_ino != BTRFS_FIRST_FREE_OBJECTID && root != dest)
9889                 return -EXDEV;
9890
9891         if (old_ino == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID ||
9892             (new_inode && btrfs_ino(BTRFS_I(new_inode)) == BTRFS_FIRST_FREE_OBJECTID))
9893                 return -ENOTEMPTY;
9894
9895         if (S_ISDIR(old_inode->i_mode) && new_inode &&
9896             new_inode->i_size > BTRFS_EMPTY_DIR_SIZE)
9897                 return -ENOTEMPTY;
9898
9899
9900         /* check for collisions, even if the  name isn't there */
9901         ret = btrfs_check_dir_item_collision(dest, new_dir->i_ino,
9902                              new_dentry->d_name.name,
9903                              new_dentry->d_name.len);
9904
9905         if (ret) {
9906                 if (ret == -EEXIST) {
9907                         /* we shouldn't get
9908                          * eexist without a new_inode */
9909                         if (WARN_ON(!new_inode)) {
9910                                 return ret;
9911                         }
9912                 } else {
9913                         /* maybe -EOVERFLOW */
9914                         return ret;
9915                 }
9916         }
9917         ret = 0;
9918
9919         /*
9920          * we're using rename to replace one file with another.  Start IO on it
9921          * now so  we don't add too much work to the end of the transaction
9922          */
9923         if (new_inode && S_ISREG(old_inode->i_mode) && new_inode->i_size)
9924                 filemap_flush(old_inode->i_mapping);
9925
9926         /* close the racy window with snapshot create/destroy ioctl */
9927         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
9928                 down_read(&fs_info->subvol_sem);
9929         /*
9930          * We want to reserve the absolute worst case amount of items.  So if
9931          * both inodes are subvols and we need to unlink them then that would
9932          * require 4 item modifications, but if they are both normal inodes it
9933          * would require 5 item modifications, so we'll assume they are normal
9934          * inodes.  So 5 * 2 is 10, plus 1 for the new link, so 11 total items
9935          * should cover the worst case number of items we'll modify.
9936          * If our rename has the whiteout flag, we need more 5 units for the
9937          * new inode (1 inode item, 1 inode ref, 2 dir items and 1 xattr item
9938          * when selinux is enabled).
9939          */
9940         trans_num_items = 11;
9941         if (flags & RENAME_WHITEOUT)
9942                 trans_num_items += 5;
9943         trans = btrfs_start_transaction(root, trans_num_items);
9944         if (IS_ERR(trans)) {
9945                 ret = PTR_ERR(trans);
9946                 goto out_notrans;
9947         }
9948
9949         if (dest != root)
9950                 btrfs_record_root_in_trans(trans, dest);
9951
9952         ret = btrfs_set_inode_index(BTRFS_I(new_dir), &index);
9953         if (ret)
9954                 goto out_fail;
9955
9956         BTRFS_I(old_inode)->dir_index = 0ULL;
9957         if (unlikely(old_ino == BTRFS_FIRST_FREE_OBJECTID)) {
9958                 /* force full log commit if subvolume involved. */
9959                 btrfs_set_log_full_commit(trans);
9960         } else {
9961                 btrfs_pin_log_trans(root);
9962                 log_pinned = true;
9963                 ret = btrfs_insert_inode_ref(trans, dest,
9964                                              new_dentry->d_name.name,
9965                                              new_dentry->d_name.len,
9966                                              old_ino,
9967                                              btrfs_ino(BTRFS_I(new_dir)), index);
9968                 if (ret)
9969                         goto out_fail;
9970         }
9971
9972         inode_inc_iversion(old_dir);
9973         inode_inc_iversion(new_dir);
9974         inode_inc_iversion(old_inode);
9975         old_dir->i_ctime = old_dir->i_mtime =
9976         new_dir->i_ctime = new_dir->i_mtime =
9977         old_inode->i_ctime = current_time(old_dir);
9978
9979         if (old_dentry->d_parent != new_dentry->d_parent)
9980                 btrfs_record_unlink_dir(trans, BTRFS_I(old_dir),
9981                                 BTRFS_I(old_inode), 1);
9982
9983         if (unlikely(old_ino == BTRFS_FIRST_FREE_OBJECTID)) {
9984                 ret = btrfs_unlink_subvol(trans, old_dir, old_dentry);
9985         } else {
9986                 ret = __btrfs_unlink_inode(trans, root, BTRFS_I(old_dir),
9987                                         BTRFS_I(d_inode(old_dentry)),
9988                                         old_dentry->d_name.name,
9989                                         old_dentry->d_name.len);
9990                 if (!ret)
9991                         ret = btrfs_update_inode(trans, root, old_inode);
9992         }
9993         if (ret) {
9994                 btrfs_abort_transaction(trans, ret);
9995                 goto out_fail;
9996         }
9997
9998         if (new_inode) {
9999                 inode_inc_iversion(new_inode);
10000                 new_inode->i_ctime = current_time(new_inode);
10001                 if (unlikely(btrfs_ino(BTRFS_I(new_inode)) ==
10002                              BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)) {
10003                         ret = btrfs_unlink_subvol(trans, new_dir, new_dentry);
10004                         BUG_ON(new_inode->i_nlink == 0);
10005                 } else {
10006                         ret = btrfs_unlink_inode(trans, dest, BTRFS_I(new_dir),
10007                                                  BTRFS_I(d_inode(new_dentry)),
10008                                                  new_dentry->d_name.name,
10009                                                  new_dentry->d_name.len);
10010                 }
10011                 if (!ret && new_inode->i_nlink == 0)
10012                         ret = btrfs_orphan_add(trans,
10013                                         BTRFS_I(d_inode(new_dentry)));
10014                 if (ret) {
10015                         btrfs_abort_transaction(trans, ret);
10016                         goto out_fail;
10017                 }
10018         }
10019
10020         ret = btrfs_add_link(trans, BTRFS_I(new_dir), BTRFS_I(old_inode),
10021                              new_dentry->d_name.name,
10022                              new_dentry->d_name.len, 0, index);
10023         if (ret) {
10024                 btrfs_abort_transaction(trans, ret);
10025                 goto out_fail;
10026         }
10027
10028         if (old_inode->i_nlink == 1)
10029                 BTRFS_I(old_inode)->dir_index = index;
10030
10031         if (log_pinned) {
10032                 struct dentry *parent = new_dentry->d_parent;
10033
10034                 btrfs_init_log_ctx(&ctx, old_inode);
10035                 ret = btrfs_log_new_name(trans, BTRFS_I(old_inode),
10036                                          BTRFS_I(old_dir), parent,
10037                                          false, &ctx);
10038                 if (ret == BTRFS_NEED_LOG_SYNC)
10039                         sync_log = true;
10040                 else if (ret == BTRFS_NEED_TRANS_COMMIT)
10041                         commit_transaction = true;
10042                 ret = 0;
10043                 btrfs_end_log_trans(root);
10044                 log_pinned = false;
10045         }
10046
10047         if (flags & RENAME_WHITEOUT) {
10048                 ret = btrfs_whiteout_for_rename(trans, root, old_dir,
10049                                                 old_dentry);
10050
10051                 if (ret) {
10052                         btrfs_abort_transaction(trans, ret);
10053                         goto out_fail;
10054                 }
10055         }
10056 out_fail:
10057         /*
10058          * If we have pinned the log and an error happened, we unpin tasks
10059          * trying to sync the log and force them to fallback to a transaction
10060          * commit if the log currently contains any of the inodes involved in
10061          * this rename operation (to ensure we do not persist a log with an
10062          * inconsistent state for any of these inodes or leading to any
10063          * inconsistencies when replayed). If the transaction was aborted, the
10064          * abortion reason is propagated to userspace when attempting to commit
10065          * the transaction. If the log does not contain any of these inodes, we
10066          * allow the tasks to sync it.
10067          */
10068         if (ret && log_pinned) {
10069                 if (btrfs_inode_in_log(BTRFS_I(old_dir), fs_info->generation) ||
10070                     btrfs_inode_in_log(BTRFS_I(new_dir), fs_info->generation) ||
10071                     btrfs_inode_in_log(BTRFS_I(old_inode), fs_info->generation) ||
10072                     (new_inode &&
10073                      btrfs_inode_in_log(BTRFS_I(new_inode), fs_info->generation)))
10074                         btrfs_set_log_full_commit(trans);
10075
10076                 btrfs_end_log_trans(root);
10077                 log_pinned = false;
10078         }
10079         if (!ret && sync_log) {
10080                 ret = btrfs_sync_log(trans, BTRFS_I(old_inode)->root, &ctx);
10081                 if (ret)
10082                         commit_transaction = true;
10083         }
10084         if (commit_transaction) {
10085                 ret = btrfs_commit_transaction(trans);
10086         } else {
10087                 int ret2;
10088
10089                 ret2 = btrfs_end_transaction(trans);
10090                 ret = ret ? ret : ret2;
10091         }
10092 out_notrans:
10093         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
10094                 up_read(&fs_info->subvol_sem);
10095
10096         return ret;
10097 }
10098
10099 static int btrfs_rename2(struct inode *old_dir, struct dentry *old_dentry,
10100                          struct inode *new_dir, struct dentry *new_dentry,
10101                          unsigned int flags)
10102 {
10103         if (flags & ~(RENAME_NOREPLACE | RENAME_EXCHANGE | RENAME_WHITEOUT))
10104                 return -EINVAL;
10105
10106         if (flags & RENAME_EXCHANGE)
10107                 return btrfs_rename_exchange(old_dir, old_dentry, new_dir,
10108                                           new_dentry);
10109
10110         return btrfs_rename(old_dir, old_dentry, new_dir, new_dentry, flags);
10111 }
10112
10113 struct btrfs_delalloc_work {
10114         struct inode *inode;
10115         struct completion completion;
10116         struct list_head list;
10117         struct btrfs_work work;
10118 };
10119
10120 static void btrfs_run_delalloc_work(struct btrfs_work *work)
10121 {
10122         struct btrfs_delalloc_work *delalloc_work;
10123         struct inode *inode;
10124
10125         delalloc_work = container_of(work, struct btrfs_delalloc_work,
10126                                      work);
10127         inode = delalloc_work->inode;
10128         filemap_flush(inode->i_mapping);
10129         if (test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
10130                                 &BTRFS_I(inode)->runtime_flags))
10131                 filemap_flush(inode->i_mapping);
10132
10133         iput(inode);
10134         complete(&delalloc_work->completion);
10135 }
10136
10137 static struct btrfs_delalloc_work *btrfs_alloc_delalloc_work(struct inode *inode)
10138 {
10139         struct btrfs_delalloc_work *work;
10140
10141         work = kmalloc(sizeof(*work), GFP_NOFS);
10142         if (!work)
10143                 return NULL;
10144
10145         init_completion(&work->completion);
10146         INIT_LIST_HEAD(&work->list);
10147         work->inode = inode;
10148         btrfs_init_work(&work->work, btrfs_run_delalloc_work, NULL, NULL);
10149
10150         return work;
10151 }
10152
10153 /*
10154  * some fairly slow code that needs optimization. This walks the list
10155  * of all the inodes with pending delalloc and forces them to disk.
10156  */
10157 static int start_delalloc_inodes(struct btrfs_root *root, int nr, bool snapshot)
10158 {
10159         struct btrfs_inode *binode;
10160         struct inode *inode;
10161         struct btrfs_delalloc_work *work, *next;
10162         struct list_head works;
10163         struct list_head splice;
10164         int ret = 0;
10165
10166         INIT_LIST_HEAD(&works);
10167         INIT_LIST_HEAD(&splice);
10168
10169         mutex_lock(&root->delalloc_mutex);
10170         spin_lock(&root->delalloc_lock);
10171         list_splice_init(&root->delalloc_inodes, &splice);
10172         while (!list_empty(&splice)) {
10173                 binode = list_entry(splice.next, struct btrfs_inode,
10174                                     delalloc_inodes);
10175
10176                 list_move_tail(&binode->delalloc_inodes,
10177                                &root->delalloc_inodes);
10178                 inode = igrab(&binode->vfs_inode);
10179                 if (!inode) {
10180                         cond_resched_lock(&root->delalloc_lock);
10181                         continue;
10182                 }
10183                 spin_unlock(&root->delalloc_lock);
10184
10185                 if (snapshot)
10186                         set_bit(BTRFS_INODE_SNAPSHOT_FLUSH,
10187                                 &binode->runtime_flags);
10188                 work = btrfs_alloc_delalloc_work(inode);
10189                 if (!work) {
10190                         iput(inode);
10191                         ret = -ENOMEM;
10192                         goto out;
10193                 }
10194                 list_add_tail(&work->list, &works);
10195                 btrfs_queue_work(root->fs_info->flush_workers,
10196                                  &work->work);
10197                 ret++;
10198                 if (nr != -1 && ret >= nr)
10199                         goto out;
10200                 cond_resched();
10201                 spin_lock(&root->delalloc_lock);
10202         }
10203         spin_unlock(&root->delalloc_lock);
10204
10205 out:
10206         list_for_each_entry_safe(work, next, &works, list) {
10207                 list_del_init(&work->list);
10208                 wait_for_completion(&work->completion);
10209                 kfree(work);
10210         }
10211
10212         if (!list_empty(&splice)) {
10213                 spin_lock(&root->delalloc_lock);
10214                 list_splice_tail(&splice, &root->delalloc_inodes);
10215                 spin_unlock(&root->delalloc_lock);
10216         }
10217         mutex_unlock(&root->delalloc_mutex);
10218         return ret;
10219 }
10220
10221 int btrfs_start_delalloc_snapshot(struct btrfs_root *root)
10222 {
10223         struct btrfs_fs_info *fs_info = root->fs_info;
10224         int ret;
10225
10226         if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state))
10227                 return -EROFS;
10228
10229         ret = start_delalloc_inodes(root, -1, true);
10230         if (ret > 0)
10231                 ret = 0;
10232         return ret;
10233 }
10234
10235 int btrfs_start_delalloc_roots(struct btrfs_fs_info *fs_info, int nr)
10236 {
10237         struct btrfs_root *root;
10238         struct list_head splice;
10239         int ret;
10240
10241         if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state))
10242                 return -EROFS;
10243
10244         INIT_LIST_HEAD(&splice);
10245
10246         mutex_lock(&fs_info->delalloc_root_mutex);
10247         spin_lock(&fs_info->delalloc_root_lock);
10248         list_splice_init(&fs_info->delalloc_roots, &splice);
10249         while (!list_empty(&splice) && nr) {
10250                 root = list_first_entry(&splice, struct btrfs_root,
10251                                         delalloc_root);
10252                 root = btrfs_grab_fs_root(root);
10253                 BUG_ON(!root);
10254                 list_move_tail(&root->delalloc_root,
10255                                &fs_info->delalloc_roots);
10256                 spin_unlock(&fs_info->delalloc_root_lock);
10257
10258                 ret = start_delalloc_inodes(root, nr, false);
10259                 btrfs_put_fs_root(root);
10260                 if (ret < 0)
10261                         goto out;
10262
10263                 if (nr != -1) {
10264                         nr -= ret;
10265                         WARN_ON(nr < 0);
10266                 }
10267                 spin_lock(&fs_info->delalloc_root_lock);
10268         }
10269         spin_unlock(&fs_info->delalloc_root_lock);
10270
10271         ret = 0;
10272 out:
10273         if (!list_empty(&splice)) {
10274                 spin_lock(&fs_info->delalloc_root_lock);
10275                 list_splice_tail(&splice, &fs_info->delalloc_roots);
10276                 spin_unlock(&fs_info->delalloc_root_lock);
10277         }
10278         mutex_unlock(&fs_info->delalloc_root_mutex);
10279         return ret;
10280 }
10281
10282 static int btrfs_symlink(struct inode *dir, struct dentry *dentry,
10283                          const char *symname)
10284 {
10285         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
10286         struct btrfs_trans_handle *trans;
10287         struct btrfs_root *root = BTRFS_I(dir)->root;
10288         struct btrfs_path *path;
10289         struct btrfs_key key;
10290         struct inode *inode = NULL;
10291         int err;
10292         u64 objectid;
10293         u64 index = 0;
10294         int name_len;
10295         int datasize;
10296         unsigned long ptr;
10297         struct btrfs_file_extent_item *ei;
10298         struct extent_buffer *leaf;
10299
10300         name_len = strlen(symname);
10301         if (name_len > BTRFS_MAX_INLINE_DATA_SIZE(fs_info))
10302                 return -ENAMETOOLONG;
10303
10304         /*
10305          * 2 items for inode item and ref
10306          * 2 items for dir items
10307          * 1 item for updating parent inode item
10308          * 1 item for the inline extent item
10309          * 1 item for xattr if selinux is on
10310          */
10311         trans = btrfs_start_transaction(root, 7);
10312         if (IS_ERR(trans))
10313                 return PTR_ERR(trans);
10314
10315         err = btrfs_find_free_ino(root, &objectid);
10316         if (err)
10317                 goto out_unlock;
10318
10319         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
10320                                 dentry->d_name.len, btrfs_ino(BTRFS_I(dir)),
10321                                 objectid, S_IFLNK|S_IRWXUGO, &index);
10322         if (IS_ERR(inode)) {
10323                 err = PTR_ERR(inode);
10324                 inode = NULL;
10325                 goto out_unlock;
10326         }
10327
10328         /*
10329         * If the active LSM wants to access the inode during
10330         * d_instantiate it needs these. Smack checks to see
10331         * if the filesystem supports xattrs by looking at the
10332         * ops vector.
10333         */
10334         inode->i_fop = &btrfs_file_operations;
10335         inode->i_op = &btrfs_file_inode_operations;
10336         inode->i_mapping->a_ops = &btrfs_aops;
10337         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
10338
10339         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
10340         if (err)
10341                 goto out_unlock;
10342
10343         path = btrfs_alloc_path();
10344         if (!path) {
10345                 err = -ENOMEM;
10346                 goto out_unlock;
10347         }
10348         key.objectid = btrfs_ino(BTRFS_I(inode));
10349         key.offset = 0;
10350         key.type = BTRFS_EXTENT_DATA_KEY;
10351         datasize = btrfs_file_extent_calc_inline_size(name_len);
10352         err = btrfs_insert_empty_item(trans, root, path, &key,
10353                                       datasize);
10354         if (err) {
10355                 btrfs_free_path(path);
10356                 goto out_unlock;
10357         }
10358         leaf = path->nodes[0];
10359         ei = btrfs_item_ptr(leaf, path->slots[0],
10360                             struct btrfs_file_extent_item);
10361         btrfs_set_file_extent_generation(leaf, ei, trans->transid);
10362         btrfs_set_file_extent_type(leaf, ei,
10363                                    BTRFS_FILE_EXTENT_INLINE);
10364         btrfs_set_file_extent_encryption(leaf, ei, 0);
10365         btrfs_set_file_extent_compression(leaf, ei, 0);
10366         btrfs_set_file_extent_other_encoding(leaf, ei, 0);
10367         btrfs_set_file_extent_ram_bytes(leaf, ei, name_len);
10368
10369         ptr = btrfs_file_extent_inline_start(ei);
10370         write_extent_buffer(leaf, symname, ptr, name_len);
10371         btrfs_mark_buffer_dirty(leaf);
10372         btrfs_free_path(path);
10373
10374         inode->i_op = &btrfs_symlink_inode_operations;
10375         inode_nohighmem(inode);
10376         inode_set_bytes(inode, name_len);
10377         btrfs_i_size_write(BTRFS_I(inode), name_len);
10378         err = btrfs_update_inode(trans, root, inode);
10379         /*
10380          * Last step, add directory indexes for our symlink inode. This is the
10381          * last step to avoid extra cleanup of these indexes if an error happens
10382          * elsewhere above.
10383          */
10384         if (!err)
10385                 err = btrfs_add_nondir(trans, BTRFS_I(dir), dentry,
10386                                 BTRFS_I(inode), 0, index);
10387         if (err)
10388                 goto out_unlock;
10389
10390         d_instantiate_new(dentry, inode);
10391
10392 out_unlock:
10393         btrfs_end_transaction(trans);
10394         if (err && inode) {
10395                 inode_dec_link_count(inode);
10396                 discard_new_inode(inode);
10397         }
10398         btrfs_btree_balance_dirty(fs_info);
10399         return err;
10400 }
10401
10402 static int __btrfs_prealloc_file_range(struct inode *inode, int mode,
10403                                        u64 start, u64 num_bytes, u64 min_size,
10404                                        loff_t actual_len, u64 *alloc_hint,
10405                                        struct btrfs_trans_handle *trans)
10406 {
10407         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
10408         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
10409         struct extent_map *em;
10410         struct btrfs_root *root = BTRFS_I(inode)->root;
10411         struct btrfs_key ins;
10412         u64 cur_offset = start;
10413         u64 i_size;
10414         u64 cur_bytes;
10415         u64 last_alloc = (u64)-1;
10416         int ret = 0;
10417         bool own_trans = true;
10418         u64 end = start + num_bytes - 1;
10419
10420         if (trans)
10421                 own_trans = false;
10422         while (num_bytes > 0) {
10423                 if (own_trans) {
10424                         trans = btrfs_start_transaction(root, 3);
10425                         if (IS_ERR(trans)) {
10426                                 ret = PTR_ERR(trans);
10427                                 break;
10428                         }
10429                 }
10430
10431                 cur_bytes = min_t(u64, num_bytes, SZ_256M);
10432                 cur_bytes = max(cur_bytes, min_size);
10433                 /*
10434                  * If we are severely fragmented we could end up with really
10435                  * small allocations, so if the allocator is returning small
10436                  * chunks lets make its job easier by only searching for those
10437                  * sized chunks.
10438                  */
10439                 cur_bytes = min(cur_bytes, last_alloc);
10440                 ret = btrfs_reserve_extent(root, cur_bytes, cur_bytes,
10441                                 min_size, 0, *alloc_hint, &ins, 1, 0);
10442                 if (ret) {
10443                         if (own_trans)
10444                                 btrfs_end_transaction(trans);
10445                         break;
10446                 }
10447                 btrfs_dec_block_group_reservations(fs_info, ins.objectid);
10448
10449                 last_alloc = ins.offset;
10450                 ret = insert_reserved_file_extent(trans, inode,
10451                                                   cur_offset, ins.objectid,
10452                                                   ins.offset, ins.offset,
10453                                                   ins.offset, 0, 0, 0,
10454                                                   BTRFS_FILE_EXTENT_PREALLOC);
10455                 if (ret) {
10456                         btrfs_free_reserved_extent(fs_info, ins.objectid,
10457                                                    ins.offset, 0);
10458                         btrfs_abort_transaction(trans, ret);
10459                         if (own_trans)
10460                                 btrfs_end_transaction(trans);
10461                         break;
10462                 }
10463
10464                 btrfs_drop_extent_cache(BTRFS_I(inode), cur_offset,
10465                                         cur_offset + ins.offset -1, 0);
10466
10467                 em = alloc_extent_map();
10468                 if (!em) {
10469                         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
10470                                 &BTRFS_I(inode)->runtime_flags);
10471                         goto next;
10472                 }
10473
10474                 em->start = cur_offset;
10475                 em->orig_start = cur_offset;
10476                 em->len = ins.offset;
10477                 em->block_start = ins.objectid;
10478                 em->block_len = ins.offset;
10479                 em->orig_block_len = ins.offset;
10480                 em->ram_bytes = ins.offset;
10481                 set_bit(EXTENT_FLAG_PREALLOC, &em->flags);
10482                 em->generation = trans->transid;
10483
10484                 while (1) {
10485                         write_lock(&em_tree->lock);
10486                         ret = add_extent_mapping(em_tree, em, 1);
10487                         write_unlock(&em_tree->lock);
10488                         if (ret != -EEXIST)
10489                                 break;
10490                         btrfs_drop_extent_cache(BTRFS_I(inode), cur_offset,
10491                                                 cur_offset + ins.offset - 1,
10492                                                 0);
10493                 }
10494                 free_extent_map(em);
10495 next:
10496                 num_bytes -= ins.offset;
10497                 cur_offset += ins.offset;
10498                 *alloc_hint = ins.objectid + ins.offset;
10499
10500                 inode_inc_iversion(inode);
10501                 inode->i_ctime = current_time(inode);
10502                 BTRFS_I(inode)->flags |= BTRFS_INODE_PREALLOC;
10503                 if (!(mode & FALLOC_FL_KEEP_SIZE) &&
10504                     (actual_len > inode->i_size) &&
10505                     (cur_offset > inode->i_size)) {
10506                         if (cur_offset > actual_len)
10507                                 i_size = actual_len;
10508                         else
10509                                 i_size = cur_offset;
10510                         i_size_write(inode, i_size);
10511                         btrfs_ordered_update_i_size(inode, i_size, NULL);
10512                 }
10513
10514                 ret = btrfs_update_inode(trans, root, inode);
10515
10516                 if (ret) {
10517                         btrfs_abort_transaction(trans, ret);
10518                         if (own_trans)
10519                                 btrfs_end_transaction(trans);
10520                         break;
10521                 }
10522
10523                 if (own_trans)
10524                         btrfs_end_transaction(trans);
10525         }
10526         if (cur_offset < end)
10527                 btrfs_free_reserved_data_space(inode, NULL, cur_offset,
10528                         end - cur_offset + 1);
10529         return ret;
10530 }
10531
10532 int btrfs_prealloc_file_range(struct inode *inode, int mode,
10533                               u64 start, u64 num_bytes, u64 min_size,
10534                               loff_t actual_len, u64 *alloc_hint)
10535 {
10536         return __btrfs_prealloc_file_range(inode, mode, start, num_bytes,
10537                                            min_size, actual_len, alloc_hint,
10538                                            NULL);
10539 }
10540
10541 int btrfs_prealloc_file_range_trans(struct inode *inode,
10542                                     struct btrfs_trans_handle *trans, int mode,
10543                                     u64 start, u64 num_bytes, u64 min_size,
10544                                     loff_t actual_len, u64 *alloc_hint)
10545 {
10546         return __btrfs_prealloc_file_range(inode, mode, start, num_bytes,
10547                                            min_size, actual_len, alloc_hint, trans);
10548 }
10549
10550 static int btrfs_set_page_dirty(struct page *page)
10551 {
10552         return __set_page_dirty_nobuffers(page);
10553 }
10554
10555 static int btrfs_permission(struct inode *inode, int mask)
10556 {
10557         struct btrfs_root *root = BTRFS_I(inode)->root;
10558         umode_t mode = inode->i_mode;
10559
10560         if (mask & MAY_WRITE &&
10561             (S_ISREG(mode) || S_ISDIR(mode) || S_ISLNK(mode))) {
10562                 if (btrfs_root_readonly(root))
10563                         return -EROFS;
10564                 if (BTRFS_I(inode)->flags & BTRFS_INODE_READONLY)
10565                         return -EACCES;
10566         }
10567         return generic_permission(inode, mask);
10568 }
10569
10570 static int btrfs_tmpfile(struct inode *dir, struct dentry *dentry, umode_t mode)
10571 {
10572         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
10573         struct btrfs_trans_handle *trans;
10574         struct btrfs_root *root = BTRFS_I(dir)->root;
10575         struct inode *inode = NULL;
10576         u64 objectid;
10577         u64 index;
10578         int ret = 0;
10579
10580         /*
10581          * 5 units required for adding orphan entry
10582          */
10583         trans = btrfs_start_transaction(root, 5);
10584         if (IS_ERR(trans))
10585                 return PTR_ERR(trans);
10586
10587         ret = btrfs_find_free_ino(root, &objectid);
10588         if (ret)
10589                 goto out;
10590
10591         inode = btrfs_new_inode(trans, root, dir, NULL, 0,
10592                         btrfs_ino(BTRFS_I(dir)), objectid, mode, &index);
10593         if (IS_ERR(inode)) {
10594                 ret = PTR_ERR(inode);
10595                 inode = NULL;
10596                 goto out;
10597         }
10598
10599         inode->i_fop = &btrfs_file_operations;
10600         inode->i_op = &btrfs_file_inode_operations;
10601
10602         inode->i_mapping->a_ops = &btrfs_aops;
10603         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
10604
10605         ret = btrfs_init_inode_security(trans, inode, dir, NULL);
10606         if (ret)
10607                 goto out;
10608
10609         ret = btrfs_update_inode(trans, root, inode);
10610         if (ret)
10611                 goto out;
10612         ret = btrfs_orphan_add(trans, BTRFS_I(inode));
10613         if (ret)
10614                 goto out;
10615
10616         /*
10617          * We set number of links to 0 in btrfs_new_inode(), and here we set
10618          * it to 1 because d_tmpfile() will issue a warning if the count is 0,
10619          * through:
10620          *
10621          *    d_tmpfile() -> inode_dec_link_count() -> drop_nlink()
10622          */
10623         set_nlink(inode, 1);
10624         d_tmpfile(dentry, inode);
10625         unlock_new_inode(inode);
10626         mark_inode_dirty(inode);
10627 out:
10628         btrfs_end_transaction(trans);
10629         if (ret && inode)
10630                 discard_new_inode(inode);
10631         btrfs_btree_balance_dirty(fs_info);
10632         return ret;
10633 }
10634
10635 void btrfs_set_range_writeback(struct extent_io_tree *tree, u64 start, u64 end)
10636 {
10637         struct inode *inode = tree->private_data;
10638         unsigned long index = start >> PAGE_SHIFT;
10639         unsigned long end_index = end >> PAGE_SHIFT;
10640         struct page *page;
10641
10642         while (index <= end_index) {
10643                 page = find_get_page(inode->i_mapping, index);
10644                 ASSERT(page); /* Pages should be in the extent_io_tree */
10645                 set_page_writeback(page);
10646                 put_page(page);
10647                 index++;
10648         }
10649 }
10650
10651 #ifdef CONFIG_SWAP
10652 /*
10653  * Add an entry indicating a block group or device which is pinned by a
10654  * swapfile. Returns 0 on success, 1 if there is already an entry for it, or a
10655  * negative errno on failure.
10656  */
10657 static int btrfs_add_swapfile_pin(struct inode *inode, void *ptr,
10658                                   bool is_block_group)
10659 {
10660         struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
10661         struct btrfs_swapfile_pin *sp, *entry;
10662         struct rb_node **p;
10663         struct rb_node *parent = NULL;
10664
10665         sp = kmalloc(sizeof(*sp), GFP_NOFS);
10666         if (!sp)
10667                 return -ENOMEM;
10668         sp->ptr = ptr;
10669         sp->inode = inode;
10670         sp->is_block_group = is_block_group;
10671
10672         spin_lock(&fs_info->swapfile_pins_lock);
10673         p = &fs_info->swapfile_pins.rb_node;
10674         while (*p) {
10675                 parent = *p;
10676                 entry = rb_entry(parent, struct btrfs_swapfile_pin, node);
10677                 if (sp->ptr < entry->ptr ||
10678                     (sp->ptr == entry->ptr && sp->inode < entry->inode)) {
10679                         p = &(*p)->rb_left;
10680                 } else if (sp->ptr > entry->ptr ||
10681                            (sp->ptr == entry->ptr && sp->inode > entry->inode)) {
10682                         p = &(*p)->rb_right;
10683                 } else {
10684                         spin_unlock(&fs_info->swapfile_pins_lock);
10685                         kfree(sp);
10686                         return 1;
10687                 }
10688         }
10689         rb_link_node(&sp->node, parent, p);
10690         rb_insert_color(&sp->node, &fs_info->swapfile_pins);
10691         spin_unlock(&fs_info->swapfile_pins_lock);
10692         return 0;
10693 }
10694
10695 /* Free all of the entries pinned by this swapfile. */
10696 static void btrfs_free_swapfile_pins(struct inode *inode)
10697 {
10698         struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
10699         struct btrfs_swapfile_pin *sp;
10700         struct rb_node *node, *next;
10701
10702         spin_lock(&fs_info->swapfile_pins_lock);
10703         node = rb_first(&fs_info->swapfile_pins);
10704         while (node) {
10705                 next = rb_next(node);
10706                 sp = rb_entry(node, struct btrfs_swapfile_pin, node);
10707                 if (sp->inode == inode) {
10708                         rb_erase(&sp->node, &fs_info->swapfile_pins);
10709                         if (sp->is_block_group)
10710                                 btrfs_put_block_group(sp->ptr);
10711                         kfree(sp);
10712                 }
10713                 node = next;
10714         }
10715         spin_unlock(&fs_info->swapfile_pins_lock);
10716 }
10717
10718 struct btrfs_swap_info {
10719         u64 start;
10720         u64 block_start;
10721         u64 block_len;
10722         u64 lowest_ppage;
10723         u64 highest_ppage;
10724         unsigned long nr_pages;
10725         int nr_extents;
10726 };
10727
10728 static int btrfs_add_swap_extent(struct swap_info_struct *sis,
10729                                  struct btrfs_swap_info *bsi)
10730 {
10731         unsigned long nr_pages;
10732         u64 first_ppage, first_ppage_reported, next_ppage;
10733         int ret;
10734
10735         first_ppage = ALIGN(bsi->block_start, PAGE_SIZE) >> PAGE_SHIFT;
10736         next_ppage = ALIGN_DOWN(bsi->block_start + bsi->block_len,
10737                                 PAGE_SIZE) >> PAGE_SHIFT;
10738
10739         if (first_ppage >= next_ppage)
10740                 return 0;
10741         nr_pages = next_ppage - first_ppage;
10742
10743         first_ppage_reported = first_ppage;
10744         if (bsi->start == 0)
10745                 first_ppage_reported++;
10746         if (bsi->lowest_ppage > first_ppage_reported)
10747                 bsi->lowest_ppage = first_ppage_reported;
10748         if (bsi->highest_ppage < (next_ppage - 1))
10749                 bsi->highest_ppage = next_ppage - 1;
10750
10751         ret = add_swap_extent(sis, bsi->nr_pages, nr_pages, first_ppage);
10752         if (ret < 0)
10753                 return ret;
10754         bsi->nr_extents += ret;
10755         bsi->nr_pages += nr_pages;
10756         return 0;
10757 }
10758
10759 static void btrfs_swap_deactivate(struct file *file)
10760 {
10761         struct inode *inode = file_inode(file);
10762
10763         btrfs_free_swapfile_pins(inode);
10764         atomic_dec(&BTRFS_I(inode)->root->nr_swapfiles);
10765 }
10766
10767 static int btrfs_swap_activate(struct swap_info_struct *sis, struct file *file,
10768                                sector_t *span)
10769 {
10770         struct inode *inode = file_inode(file);
10771         struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
10772         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
10773         struct extent_state *cached_state = NULL;
10774         struct extent_map *em = NULL;
10775         struct btrfs_device *device = NULL;
10776         struct btrfs_swap_info bsi = {
10777                 .lowest_ppage = (sector_t)-1ULL,
10778         };
10779         int ret = 0;
10780         u64 isize;
10781         u64 start;
10782
10783         /*
10784          * If the swap file was just created, make sure delalloc is done. If the
10785          * file changes again after this, the user is doing something stupid and
10786          * we don't really care.
10787          */
10788         ret = btrfs_wait_ordered_range(inode, 0, (u64)-1);
10789         if (ret)
10790                 return ret;
10791
10792         /*
10793          * The inode is locked, so these flags won't change after we check them.
10794          */
10795         if (BTRFS_I(inode)->flags & BTRFS_INODE_COMPRESS) {
10796                 btrfs_warn(fs_info, "swapfile must not be compressed");
10797                 return -EINVAL;
10798         }
10799         if (!(BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW)) {
10800                 btrfs_warn(fs_info, "swapfile must not be copy-on-write");
10801                 return -EINVAL;
10802         }
10803         if (!(BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)) {
10804                 btrfs_warn(fs_info, "swapfile must not be checksummed");
10805                 return -EINVAL;
10806         }
10807
10808         /*
10809          * Balance or device remove/replace/resize can move stuff around from
10810          * under us. The EXCL_OP flag makes sure they aren't running/won't run
10811          * concurrently while we are mapping the swap extents, and
10812          * fs_info->swapfile_pins prevents them from running while the swap file
10813          * is active and moving the extents. Note that this also prevents a
10814          * concurrent device add which isn't actually necessary, but it's not
10815          * really worth the trouble to allow it.
10816          */
10817         if (test_and_set_bit(BTRFS_FS_EXCL_OP, &fs_info->flags)) {
10818                 btrfs_warn(fs_info,
10819            "cannot activate swapfile while exclusive operation is running");
10820                 return -EBUSY;
10821         }
10822         /*
10823          * Snapshots can create extents which require COW even if NODATACOW is
10824          * set. We use this counter to prevent snapshots. We must increment it
10825          * before walking the extents because we don't want a concurrent
10826          * snapshot to run after we've already checked the extents.
10827          */
10828         atomic_inc(&BTRFS_I(inode)->root->nr_swapfiles);
10829
10830         isize = ALIGN_DOWN(inode->i_size, fs_info->sectorsize);
10831
10832         lock_extent_bits(io_tree, 0, isize - 1, &cached_state);
10833         start = 0;
10834         while (start < isize) {
10835                 u64 logical_block_start, physical_block_start;
10836                 struct btrfs_block_group *bg;
10837                 u64 len = isize - start;
10838
10839                 em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, start, len, 0);
10840                 if (IS_ERR(em)) {
10841                         ret = PTR_ERR(em);
10842                         goto out;
10843                 }
10844
10845                 if (em->block_start == EXTENT_MAP_HOLE) {
10846                         btrfs_warn(fs_info, "swapfile must not have holes");
10847                         ret = -EINVAL;
10848                         goto out;
10849                 }
10850                 if (em->block_start == EXTENT_MAP_INLINE) {
10851                         /*
10852                          * It's unlikely we'll ever actually find ourselves
10853                          * here, as a file small enough to fit inline won't be
10854                          * big enough to store more than the swap header, but in
10855                          * case something changes in the future, let's catch it
10856                          * here rather than later.
10857                          */
10858                         btrfs_warn(fs_info, "swapfile must not be inline");
10859                         ret = -EINVAL;
10860                         goto out;
10861                 }
10862                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
10863                         btrfs_warn(fs_info, "swapfile must not be compressed");
10864                         ret = -EINVAL;
10865                         goto out;
10866                 }
10867
10868                 logical_block_start = em->block_start + (start - em->start);
10869                 len = min(len, em->len - (start - em->start));
10870                 free_extent_map(em);
10871                 em = NULL;
10872
10873                 ret = can_nocow_extent(inode, start, &len, NULL, NULL, NULL);
10874                 if (ret < 0) {
10875                         goto out;
10876                 } else if (ret) {
10877                         ret = 0;
10878                 } else {
10879                         btrfs_warn(fs_info,
10880                                    "swapfile must not be copy-on-write");
10881                         ret = -EINVAL;
10882                         goto out;
10883                 }
10884
10885                 em = btrfs_get_chunk_map(fs_info, logical_block_start, len);
10886                 if (IS_ERR(em)) {
10887                         ret = PTR_ERR(em);
10888                         goto out;
10889                 }
10890
10891                 if (em->map_lookup->type & BTRFS_BLOCK_GROUP_PROFILE_MASK) {
10892                         btrfs_warn(fs_info,
10893                                    "swapfile must have single data profile");
10894                         ret = -EINVAL;
10895                         goto out;
10896                 }
10897
10898                 if (device == NULL) {
10899                         device = em->map_lookup->stripes[0].dev;
10900                         ret = btrfs_add_swapfile_pin(inode, device, false);
10901                         if (ret == 1)
10902                                 ret = 0;
10903                         else if (ret)
10904                                 goto out;
10905                 } else if (device != em->map_lookup->stripes[0].dev) {
10906                         btrfs_warn(fs_info, "swapfile must be on one device");
10907                         ret = -EINVAL;
10908                         goto out;
10909                 }
10910
10911                 physical_block_start = (em->map_lookup->stripes[0].physical +
10912                                         (logical_block_start - em->start));
10913                 len = min(len, em->len - (logical_block_start - em->start));
10914                 free_extent_map(em);
10915                 em = NULL;
10916
10917                 bg = btrfs_lookup_block_group(fs_info, logical_block_start);
10918                 if (!bg) {
10919                         btrfs_warn(fs_info,
10920                            "could not find block group containing swapfile");
10921                         ret = -EINVAL;
10922                         goto out;
10923                 }
10924
10925                 ret = btrfs_add_swapfile_pin(inode, bg, true);
10926                 if (ret) {
10927                         btrfs_put_block_group(bg);
10928                         if (ret == 1)
10929                                 ret = 0;
10930                         else
10931                                 goto out;
10932                 }
10933
10934                 if (bsi.block_len &&
10935                     bsi.block_start + bsi.block_len == physical_block_start) {
10936                         bsi.block_len += len;
10937                 } else {
10938                         if (bsi.block_len) {
10939                                 ret = btrfs_add_swap_extent(sis, &bsi);
10940                                 if (ret)
10941                                         goto out;
10942                         }
10943                         bsi.start = start;
10944                         bsi.block_start = physical_block_start;
10945                         bsi.block_len = len;
10946                 }
10947
10948                 start += len;
10949         }
10950
10951         if (bsi.block_len)
10952                 ret = btrfs_add_swap_extent(sis, &bsi);
10953
10954 out:
10955         if (!IS_ERR_OR_NULL(em))
10956                 free_extent_map(em);
10957
10958         unlock_extent_cached(io_tree, 0, isize - 1, &cached_state);
10959
10960         if (ret)
10961                 btrfs_swap_deactivate(file);
10962
10963         clear_bit(BTRFS_FS_EXCL_OP, &fs_info->flags);
10964
10965         if (ret)
10966                 return ret;
10967
10968         if (device)
10969                 sis->bdev = device->bdev;
10970         *span = bsi.highest_ppage - bsi.lowest_ppage + 1;
10971         sis->max = bsi.nr_pages;
10972         sis->pages = bsi.nr_pages - 1;
10973         sis->highest_bit = bsi.nr_pages - 1;
10974         return bsi.nr_extents;
10975 }
10976 #else
10977 static void btrfs_swap_deactivate(struct file *file)
10978 {
10979 }
10980
10981 static int btrfs_swap_activate(struct swap_info_struct *sis, struct file *file,
10982                                sector_t *span)
10983 {
10984         return -EOPNOTSUPP;
10985 }
10986 #endif
10987
10988 static const struct inode_operations btrfs_dir_inode_operations = {
10989         .getattr        = btrfs_getattr,
10990         .lookup         = btrfs_lookup,
10991         .create         = btrfs_create,
10992         .unlink         = btrfs_unlink,
10993         .link           = btrfs_link,
10994         .mkdir          = btrfs_mkdir,
10995         .rmdir          = btrfs_rmdir,
10996         .rename         = btrfs_rename2,
10997         .symlink        = btrfs_symlink,
10998         .setattr        = btrfs_setattr,
10999         .mknod          = btrfs_mknod,
11000         .listxattr      = btrfs_listxattr,
11001         .permission     = btrfs_permission,
11002         .get_acl        = btrfs_get_acl,
11003         .set_acl        = btrfs_set_acl,
11004         .update_time    = btrfs_update_time,
11005         .tmpfile        = btrfs_tmpfile,
11006 };
11007 static const struct inode_operations btrfs_dir_ro_inode_operations = {
11008         .lookup         = btrfs_lookup,
11009         .permission     = btrfs_permission,
11010         .update_time    = btrfs_update_time,
11011 };
11012
11013 static const struct file_operations btrfs_dir_file_operations = {
11014         .llseek         = generic_file_llseek,
11015         .read           = generic_read_dir,
11016         .iterate_shared = btrfs_real_readdir,
11017         .open           = btrfs_opendir,
11018         .unlocked_ioctl = btrfs_ioctl,
11019 #ifdef CONFIG_COMPAT
11020         .compat_ioctl   = btrfs_compat_ioctl,
11021 #endif
11022         .release        = btrfs_release_file,
11023         .fsync          = btrfs_sync_file,
11024 };
11025
11026 static const struct extent_io_ops btrfs_extent_io_ops = {
11027         /* mandatory callbacks */
11028         .submit_bio_hook = btrfs_submit_bio_hook,
11029         .readpage_end_io_hook = btrfs_readpage_end_io_hook,
11030 };
11031
11032 /*
11033  * btrfs doesn't support the bmap operation because swapfiles
11034  * use bmap to make a mapping of extents in the file.  They assume
11035  * these extents won't change over the life of the file and they
11036  * use the bmap result to do IO directly to the drive.
11037  *
11038  * the btrfs bmap call would return logical addresses that aren't
11039  * suitable for IO and they also will change frequently as COW
11040  * operations happen.  So, swapfile + btrfs == corruption.
11041  *
11042  * For now we're avoiding this by dropping bmap.
11043  */
11044 static const struct address_space_operations btrfs_aops = {
11045         .readpage       = btrfs_readpage,
11046         .writepage      = btrfs_writepage,
11047         .writepages     = btrfs_writepages,
11048         .readpages      = btrfs_readpages,
11049         .direct_IO      = btrfs_direct_IO,
11050         .invalidatepage = btrfs_invalidatepage,
11051         .releasepage    = btrfs_releasepage,
11052         .set_page_dirty = btrfs_set_page_dirty,
11053         .error_remove_page = generic_error_remove_page,
11054         .swap_activate  = btrfs_swap_activate,
11055         .swap_deactivate = btrfs_swap_deactivate,
11056 };
11057
11058 static const struct inode_operations btrfs_file_inode_operations = {
11059         .getattr        = btrfs_getattr,
11060         .setattr        = btrfs_setattr,
11061         .listxattr      = btrfs_listxattr,
11062         .permission     = btrfs_permission,
11063         .fiemap         = btrfs_fiemap,
11064         .get_acl        = btrfs_get_acl,
11065         .set_acl        = btrfs_set_acl,
11066         .update_time    = btrfs_update_time,
11067 };
11068 static const struct inode_operations btrfs_special_inode_operations = {
11069         .getattr        = btrfs_getattr,
11070         .setattr        = btrfs_setattr,
11071         .permission     = btrfs_permission,
11072         .listxattr      = btrfs_listxattr,
11073         .get_acl        = btrfs_get_acl,
11074         .set_acl        = btrfs_set_acl,
11075         .update_time    = btrfs_update_time,
11076 };
11077 static const struct inode_operations btrfs_symlink_inode_operations = {
11078         .get_link       = page_get_link,
11079         .getattr        = btrfs_getattr,
11080         .setattr        = btrfs_setattr,
11081         .permission     = btrfs_permission,
11082         .listxattr      = btrfs_listxattr,
11083         .update_time    = btrfs_update_time,
11084 };
11085
11086 const struct dentry_operations btrfs_dentry_operations = {
11087         .d_delete       = btrfs_dentry_delete,
11088 };