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