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