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