fs/btrfs/inode.c

   1 /*
   2  * Copyright (C) 2007 Oracle.  All rights reserved.
   3  *
   4  * This program is free software; you can redistribute it and/or
   5  * modify it under the terms of the GNU General Public
   6  * License v2 as published by the Free Software Foundation.
   7  *
   8  * This program is distributed in the hope that it will be useful,
   9  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  11  * General Public License for more details.
  12  *
  13  * You should have received a copy of the GNU General Public
  14  * License along with this program; if not, write to the
  15  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  16  * Boston, MA 021110-1307, USA.
  17  */
  18
  19 #include <linux/kernel.h>
  20 #include <linux/bio.h>
  21 #include <linux/buffer_head.h>
  22 #include <linux/file.h>
  23 #include <linux/fs.h>
  24 #include <linux/pagemap.h>
  25 #include <linux/highmem.h>
  26 #include <linux/time.h>
  27 #include <linux/init.h>
  28 #include <linux/string.h>
  29 #include <linux/backing-dev.h>
  30 #include <linux/mpage.h>
  31 #include <linux/swap.h>
  32 #include <linux/writeback.h>
  33 #include <linux/compat.h>
  34 #include <linux/bit_spinlock.h>
  35 #include <linux/xattr.h>
  36 #include <linux/posix_acl.h>
  37 #include <linux/falloc.h>
  38 #include <linux/slab.h>
  39 #include <linux/ratelimit.h>
  40 #include <linux/mount.h>
  41 #include <linux/btrfs.h>
  42 #include <linux/blkdev.h>
  43 #include <linux/posix_acl_xattr.h>
  44 #include <linux/uio.h>
  45 #include "ctree.h"
  46 #include "disk-io.h"
  47 #include "transaction.h"
  48 #include "btrfs_inode.h"
  49 #include "print-tree.h"
  50 #include "ordered-data.h"
  51 #include "xattr.h"
  52 #include "tree-log.h"
  53 #include "volumes.h"
  54 #include "compression.h"
  55 #include "locking.h"
  56 #include "free-space-cache.h"
  57 #include "inode-map.h"
  58 #include "backref.h"
  59 #include "hash.h"
  60 #include "props.h"
  61 #include "qgroup.h"
  62 #include "dedupe.h"
  63
  64 struct btrfs_iget_args {
  65         struct btrfs_key *location;
  66         struct btrfs_root *root;
  67 };
  68
  69 struct btrfs_dio_data {
  70         u64 outstanding_extents;
  71         u64 reserve;
  72         u64 unsubmitted_oe_range_start;
  73         u64 unsubmitted_oe_range_end;
  74         int overwrite;
  75 };
  76
  77 static const struct inode_operations btrfs_dir_inode_operations;
  78 static const struct inode_operations btrfs_symlink_inode_operations;
  79 static const struct inode_operations btrfs_dir_ro_inode_operations;
  80 static const struct inode_operations btrfs_special_inode_operations;
  81 static const struct inode_operations btrfs_file_inode_operations;
  82 static const struct address_space_operations btrfs_aops;
  83 static const struct address_space_operations btrfs_symlink_aops;
  84 static const struct file_operations btrfs_dir_file_operations;
  85 static const struct extent_io_ops btrfs_extent_io_ops;
  86
  87 static struct kmem_cache *btrfs_inode_cachep;
  88 struct kmem_cache *btrfs_trans_handle_cachep;
  89 struct kmem_cache *btrfs_transaction_cachep;
  90 struct kmem_cache *btrfs_path_cachep;
  91 struct kmem_cache *btrfs_free_space_cachep;
  92
  93 #define S_SHIFT 12
  94 static const unsigned char btrfs_type_by_mode[S_IFMT >> S_SHIFT] = {
  95         [S_IFREG >> S_SHIFT]    = BTRFS_FT_REG_FILE,
  96         [S_IFDIR >> S_SHIFT]    = BTRFS_FT_DIR,
  97         [S_IFCHR >> S_SHIFT]    = BTRFS_FT_CHRDEV,
  98         [S_IFBLK >> S_SHIFT]    = BTRFS_FT_BLKDEV,
  99         [S_IFIFO >> S_SHIFT]    = BTRFS_FT_FIFO,
 100         [S_IFSOCK >> S_SHIFT]   = BTRFS_FT_SOCK,
 101         [S_IFLNK >> S_SHIFT]    = BTRFS_FT_SYMLINK,
 102 };
 103
 104 static int btrfs_setsize(struct inode *inode, struct iattr *attr);
 105 static int btrfs_truncate(struct inode *inode);
 106 static int btrfs_finish_ordered_io(struct btrfs_ordered_extent *ordered_extent);
 107 static noinline int cow_file_range(struct inode *inode,
 108                                    struct page *locked_page,
 109                                    u64 start, u64 end, u64 delalloc_end,
 110                                    int *page_started, unsigned long *nr_written,
 111                                    int unlock, struct btrfs_dedupe_hash *hash);
 112 static struct extent_map *create_io_em(struct inode *inode, u64 start, u64 len,
 113                                        u64 orig_start, u64 block_start,
 114                                        u64 block_len, u64 orig_block_len,
 115                                        u64 ram_bytes, int compress_type,
 116                                        int type);
 117
 118 static int btrfs_dirty_inode(struct inode *inode);
 119
 120 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
 121 void btrfs_test_inode_set_ops(struct inode *inode)
 122 {
 123         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
 124 }
 125 #endif
 126
 127 static int btrfs_init_inode_security(struct btrfs_trans_handle *trans,
 128                                      struct inode *inode,  struct inode *dir,
 129                                      const struct qstr *qstr)
 130 {
 131         int err;
 132
 133         err = btrfs_init_acl(trans, inode, dir);
 134         if (!err)
 135                 err = btrfs_xattr_security_init(trans, inode, dir, qstr);
 136         return err;
 137 }
 138
 139 /*
 140  * this does all the hard work for inserting an inline extent into
 141  * the btree.  The caller should have done a btrfs_drop_extents so that
 142  * no overlapping inline items exist in the btree
 143  */
 144 static int insert_inline_extent(struct btrfs_trans_handle *trans,
 145                                 struct btrfs_path *path, int extent_inserted,
 146                                 struct btrfs_root *root, struct inode *inode,
 147                                 u64 start, size_t size, size_t compressed_size,
 148                                 int compress_type,
 149                                 struct page **compressed_pages)
 150 {
 151         struct extent_buffer *leaf;
 152         struct page *page = NULL;
 153         char *kaddr;
 154         unsigned long ptr;
 155         struct btrfs_file_extent_item *ei;
 156         int err = 0;
 157         int ret;
 158         size_t cur_size = size;
 159         unsigned long offset;
 160
 161         if (compressed_size && compressed_pages)
 162                 cur_size = compressed_size;
 163
 164         inode_add_bytes(inode, size);
 165
 166         if (!extent_inserted) {
 167                 struct btrfs_key key;
 168                 size_t datasize;
 169
 170                 key.objectid = btrfs_ino(BTRFS_I(inode));
 171                 key.offset = start;
 172                 key.type = BTRFS_EXTENT_DATA_KEY;
 173
 174                 datasize = btrfs_file_extent_calc_inline_size(cur_size);
 175                 path->leave_spinning = 1;
 176                 ret = btrfs_insert_empty_item(trans, root, path, &key,
 177                                               datasize);
 178                 if (ret) {
 179                         err = ret;
 180                         goto fail;
 181                 }
 182         }
 183         leaf = path->nodes[0];
 184         ei = btrfs_item_ptr(leaf, path->slots[0],
 185                             struct btrfs_file_extent_item);
 186         btrfs_set_file_extent_generation(leaf, ei, trans->transid);
 187         btrfs_set_file_extent_type(leaf, ei, BTRFS_FILE_EXTENT_INLINE);
 188         btrfs_set_file_extent_encryption(leaf, ei, 0);
 189         btrfs_set_file_extent_other_encoding(leaf, ei, 0);
 190         btrfs_set_file_extent_ram_bytes(leaf, ei, size);
 191         ptr = btrfs_file_extent_inline_start(ei);
 192
 193         if (compress_type != BTRFS_COMPRESS_NONE) {
 194                 struct page *cpage;
 195                 int i = 0;
 196                 while (compressed_size > 0) {
 197                         cpage = compressed_pages[i];
 198                         cur_size = min_t(unsigned long, compressed_size,
 199                                        PAGE_SIZE);
 200
 201                         kaddr = kmap_atomic(cpage);
 202                         write_extent_buffer(leaf, kaddr, ptr, cur_size);
 203                         kunmap_atomic(kaddr);
 204
 205                         i++;
 206                         ptr += cur_size;
 207                         compressed_size -= cur_size;
 208                 }
 209                 btrfs_set_file_extent_compression(leaf, ei,
 210                                                   compress_type);
 211         } else {
 212                 page = find_get_page(inode->i_mapping,
 213                                      start >> PAGE_SHIFT);
 214                 btrfs_set_file_extent_compression(leaf, ei, 0);
 215                 kaddr = kmap_atomic(page);
 216                 offset = start & (PAGE_SIZE - 1);
 217                 write_extent_buffer(leaf, kaddr + offset, ptr, size);
 218                 kunmap_atomic(kaddr);
 219                 put_page(page);
 220         }
 221         btrfs_mark_buffer_dirty(leaf);
 222         btrfs_release_path(path);
 223
 224         /*
 225          * we're an inline extent, so nobody can
 226          * extend the file past i_size without locking
 227          * a page we already have locked.
 228          *
 229          * We must do any isize and inode updates
 230          * before we unlock the pages.  Otherwise we
 231          * could end up racing with unlink.
 232          */
 233         BTRFS_I(inode)->disk_i_size = inode->i_size;
 234         ret = btrfs_update_inode(trans, root, inode);
 235
 236         return ret;
 237 fail:
 238         return err;
 239 }
 240
 241
 242 /*
 243  * conditionally insert an inline extent into the file.  This
 244  * does the checks required to make sure the data is small enough
 245  * to fit as an inline extent.
 246  */
 247 static noinline int cow_file_range_inline(struct btrfs_root *root,
 248                                           struct inode *inode, u64 start,
 249                                           u64 end, size_t compressed_size,
 250                                           int compress_type,
 251                                           struct page **compressed_pages)
 252 {
 253         struct btrfs_fs_info *fs_info = root->fs_info;
 254         struct btrfs_trans_handle *trans;
 255         u64 isize = i_size_read(inode);
 256         u64 actual_end = min(end + 1, isize);
 257         u64 inline_len = actual_end - start;
 258         u64 aligned_end = ALIGN(end, fs_info->sectorsize);
 259         u64 data_len = inline_len;
 260         int ret;
 261         struct btrfs_path *path;
 262         int extent_inserted = 0;
 263         u32 extent_item_size;
 264
 265         if (compressed_size)
 266                 data_len = compressed_size;
 267
 268         if (start > 0 ||
 269             actual_end > fs_info->sectorsize ||
 270             data_len > BTRFS_MAX_INLINE_DATA_SIZE(fs_info) ||
 271             (!compressed_size &&
 272             (actual_end & (fs_info->sectorsize - 1)) == 0) ||
 273             end + 1 < isize ||
 274             data_len > fs_info->max_inline) {
 275                 return 1;
 276         }
 277
 278         path = btrfs_alloc_path();
 279         if (!path)
 280                 return -ENOMEM;
 281
 282         trans = btrfs_join_transaction(root);
 283         if (IS_ERR(trans)) {
 284                 btrfs_free_path(path);
 285                 return PTR_ERR(trans);
 286         }
 287         trans->block_rsv = &fs_info->delalloc_block_rsv;
 288
 289         if (compressed_size && compressed_pages)
 290                 extent_item_size = btrfs_file_extent_calc_inline_size(
 291                    compressed_size);
 292         else
 293                 extent_item_size = btrfs_file_extent_calc_inline_size(
 294                     inline_len);
 295
 296         ret = __btrfs_drop_extents(trans, root, inode, path,
 297                                    start, aligned_end, NULL,
 298                                    1, 1, extent_item_size, &extent_inserted);
 299         if (ret) {
 300                 btrfs_abort_transaction(trans, ret);
 301                 goto out;
 302         }
 303
 304         if (isize > actual_end)
 305                 inline_len = min_t(u64, isize, actual_end);
 306         ret = insert_inline_extent(trans, path, extent_inserted,
 307                                    root, inode, start,
 308                                    inline_len, compressed_size,
 309                                    compress_type, compressed_pages);
 310         if (ret && ret != -ENOSPC) {
 311                 btrfs_abort_transaction(trans, ret);
 312                 goto out;
 313         } else if (ret == -ENOSPC) {
 314                 ret = 1;
 315                 goto out;
 316         }
 317
 318         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &BTRFS_I(inode)->runtime_flags);
 319         btrfs_delalloc_release_metadata(BTRFS_I(inode), end + 1 - start);
 320         btrfs_drop_extent_cache(BTRFS_I(inode), start, aligned_end - 1, 0);
 321 out:
 322         /*
 323          * Don't forget to free the reserved space, as for inlined extent
 324          * it won't count as data extent, free them directly here.
 325          * And at reserve time, it's always aligned to page size, so
 326          * just free one page here.
 327          */
 328         btrfs_qgroup_free_data(inode, 0, PAGE_SIZE);
 329         btrfs_free_path(path);
 330         btrfs_end_transaction(trans);
 331         return ret;
 332 }
 333
 334 struct async_extent {
 335         u64 start;
 336         u64 ram_size;
 337         u64 compressed_size;
 338         struct page **pages;
 339         unsigned long nr_pages;
 340         int compress_type;
 341         struct list_head list;
 342 };
 343
 344 struct async_cow {
 345         struct inode *inode;
 346         struct btrfs_root *root;
 347         struct page *locked_page;
 348         u64 start;
 349         u64 end;
 350         struct list_head extents;
 351         struct btrfs_work work;
 352 };
 353
 354 static noinline int add_async_extent(struct async_cow *cow,
 355                                      u64 start, u64 ram_size,
 356                                      u64 compressed_size,
 357                                      struct page **pages,
 358                                      unsigned long nr_pages,
 359                                      int compress_type)
 360 {
 361         struct async_extent *async_extent;
 362
 363         async_extent = kmalloc(sizeof(*async_extent), GFP_NOFS);
 364         BUG_ON(!async_extent); /* -ENOMEM */
 365         async_extent->start = start;
 366         async_extent->ram_size = ram_size;
 367         async_extent->compressed_size = compressed_size;
 368         async_extent->pages = pages;
 369         async_extent->nr_pages = nr_pages;
 370         async_extent->compress_type = compress_type;
 371         list_add_tail(&async_extent->list, &cow->extents);
 372         return 0;
 373 }
 374
 375 static inline int inode_need_compress(struct inode *inode)
 376 {
 377         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
 378
 379         /* force compress */
 380         if (btrfs_test_opt(fs_info, FORCE_COMPRESS))
 381                 return 1;
 382         /* bad compression ratios */
 383         if (BTRFS_I(inode)->flags & BTRFS_INODE_NOCOMPRESS)
 384                 return 0;
 385         if (btrfs_test_opt(fs_info, COMPRESS) ||
 386             BTRFS_I(inode)->flags & BTRFS_INODE_COMPRESS ||
 387             BTRFS_I(inode)->force_compress)
 388                 return 1;
 389         return 0;
 390 }
 391
 392 static inline void inode_should_defrag(struct btrfs_inode *inode,
 393                 u64 start, u64 end, u64 num_bytes, u64 small_write)
 394 {
 395         /* If this is a small write inside eof, kick off a defrag */
 396         if (num_bytes < small_write &&
 397             (start > 0 || end + 1 < inode->disk_i_size))
 398                 btrfs_add_inode_defrag(NULL, inode);
 399 }
 400
 401 /*
 402  * we create compressed extents in two phases.  The first
 403  * phase compresses a range of pages that have already been
 404  * locked (both pages and state bits are locked).
 405  *
 406  * This is done inside an ordered work queue, and the compression
 407  * is spread across many cpus.  The actual IO submission is step
 408  * two, and the ordered work queue takes care of making sure that
 409  * happens in the same order things were put onto the queue by
 410  * writepages and friends.
 411  *
 412  * If this code finds it can't get good compression, it puts an
 413  * entry onto the work queue to write the uncompressed bytes.  This
 414  * makes sure that both compressed inodes and uncompressed inodes
 415  * are written in the same order that the flusher thread sent them
 416  * down.
 417  */
 418 static noinline void compress_file_range(struct inode *inode,
 419                                         struct page *locked_page,
 420                                         u64 start, u64 end,
 421                                         struct async_cow *async_cow,
 422                                         int *num_added)
 423 {
 424         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
 425         struct btrfs_root *root = BTRFS_I(inode)->root;
 426         u64 num_bytes;
 427         u64 blocksize = fs_info->sectorsize;
 428         u64 actual_end;
 429         u64 isize = i_size_read(inode);
 430         int ret = 0;
 431         struct page **pages = NULL;
 432         unsigned long nr_pages;
 433         unsigned long total_compressed = 0;
 434         unsigned long total_in = 0;
 435         int i;
 436         int will_compress;
 437         int compress_type = fs_info->compress_type;
 438         int redirty = 0;
 439
 440         inode_should_defrag(BTRFS_I(inode), start, end, end - start + 1,
 441                         SZ_16K);
 442
 443         actual_end = min_t(u64, isize, end + 1);
 444 again:
 445         will_compress = 0;
 446         nr_pages = (end >> PAGE_SHIFT) - (start >> PAGE_SHIFT) + 1;
 447         BUILD_BUG_ON((BTRFS_MAX_COMPRESSED % PAGE_SIZE) != 0);
 448         nr_pages = min_t(unsigned long, nr_pages,
 449                         BTRFS_MAX_COMPRESSED / PAGE_SIZE);
 450
 451         /*
 452          * we don't want to send crud past the end of i_size through
 453          * compression, that's just a waste of CPU time.  So, if the
 454          * end of the file is before the start of our current
 455          * requested range of bytes, we bail out to the uncompressed
 456          * cleanup code that can deal with all of this.
 457          *
 458          * It isn't really the fastest way to fix things, but this is a
 459          * very uncommon corner.
 460          */
 461         if (actual_end <= start)
 462                 goto cleanup_and_bail_uncompressed;
 463
 464         total_compressed = actual_end - start;
 465
 466         /*
 467          * skip compression for a small file range(<=blocksize) that
 468          * isn't an inline extent, since it doesn't save disk space at all.
 469          */
 470         if (total_compressed <= blocksize &&
 471            (start > 0 || end + 1 < BTRFS_I(inode)->disk_i_size))
 472                 goto cleanup_and_bail_uncompressed;
 473
 474         total_compressed = min_t(unsigned long, total_compressed,
 475                         BTRFS_MAX_UNCOMPRESSED);
 476         num_bytes = ALIGN(end - start + 1, blocksize);
 477         num_bytes = max(blocksize,  num_bytes);
 478         total_in = 0;
 479         ret = 0;
 480
 481         /*
 482          * we do compression for mount -o compress and when the
 483          * inode has not been flagged as nocompress.  This flag can
 484          * change at any time if we discover bad compression ratios.
 485          */
 486         if (inode_need_compress(inode)) {
 487                 WARN_ON(pages);
 488                 pages = kcalloc(nr_pages, sizeof(struct page *), GFP_NOFS);
 489                 if (!pages) {
 490                         /* just bail out to the uncompressed code */
 491                         goto cont;
 492                 }
 493
 494                 if (BTRFS_I(inode)->force_compress)
 495                         compress_type = BTRFS_I(inode)->force_compress;
 496
 497                 /*
 498                  * we need to call clear_page_dirty_for_io on each
 499                  * page in the range.  Otherwise applications with the file
 500                  * mmap'd can wander in and change the page contents while
 501                  * we are compressing them.
 502                  *
 503                  * If the compression fails for any reason, we set the pages
 504                  * dirty again later on.
 505                  */
 506                 extent_range_clear_dirty_for_io(inode, start, end);
 507                 redirty = 1;
 508                 ret = btrfs_compress_pages(compress_type,
 509                                            inode->i_mapping, start,
 510                                            pages,
 511                                            &nr_pages,
 512                                            &total_in,
 513                                            &total_compressed);
 514
 515                 if (!ret) {
 516                         unsigned long offset = total_compressed &
 517                                 (PAGE_SIZE - 1);
 518                         struct page *page = pages[nr_pages - 1];
 519                         char *kaddr;
 520
 521                         /* zero the tail end of the last page, we might be
 522                          * sending it down to disk
 523                          */
 524                         if (offset) {
 525                                 kaddr = kmap_atomic(page);
 526                                 memset(kaddr + offset, 0,
 527                                        PAGE_SIZE - offset);
 528                                 kunmap_atomic(kaddr);
 529                         }
 530                         will_compress = 1;
 531                 }
 532         }
 533 cont:
 534         if (start == 0) {
 535                 /* lets try to make an inline extent */
 536                 if (ret || total_in < (actual_end - start)) {
 537                         /* we didn't compress the entire range, try
 538                          * to make an uncompressed inline extent.
 539                          */
 540                         ret = cow_file_range_inline(root, inode, start, end,
 541                                             0, BTRFS_COMPRESS_NONE, NULL);
 542                 } else {
 543                         /* try making a compressed inline extent */
 544                         ret = cow_file_range_inline(root, inode, start, end,
 545                                                     total_compressed,
 546                                                     compress_type, pages);
 547                 }
 548                 if (ret <= 0) {
 549                         unsigned long clear_flags = EXTENT_DELALLOC |
 550                                 EXTENT_DEFRAG;
 551                         unsigned long page_error_op;
 552
 553                         clear_flags |= (ret < 0) ? EXTENT_DO_ACCOUNTING : 0;
 554                         page_error_op = ret < 0 ? PAGE_SET_ERROR : 0;
 555
 556                         /*
 557                          * inline extent creation worked or returned error,
 558                          * we don't need to create any more async work items.
 559                          * Unlock and free up our temp pages.
 560                          */
 561                         extent_clear_unlock_delalloc(inode, start, end, end,
 562                                                      NULL, clear_flags,
 563                                                      PAGE_UNLOCK |
 564                                                      PAGE_CLEAR_DIRTY |
 565                                                      PAGE_SET_WRITEBACK |
 566                                                      page_error_op |
 567                                                      PAGE_END_WRITEBACK);
 568                         btrfs_free_reserved_data_space_noquota(inode, start,
 569                                                 end - start + 1);
 570                         goto free_pages_out;
 571                 }
 572         }
 573
 574         if (will_compress) {
 575                 /*
 576                  * we aren't doing an inline extent round the compressed size
 577                  * up to a block size boundary so the allocator does sane
 578                  * things
 579                  */
 580                 total_compressed = ALIGN(total_compressed, blocksize);
 581
 582                 /*
 583                  * one last check to make sure the compression is really a
 584                  * win, compare the page count read with the blocks on disk
 585                  */
 586                 total_in = ALIGN(total_in, PAGE_SIZE);
 587                 if (total_compressed >= total_in) {
 588                         will_compress = 0;
 589                 } else {
 590                         num_bytes = total_in;
 591                         *num_added += 1;
 592
 593                         /*
 594                          * The async work queues will take care of doing actual
 595                          * allocation on disk for these compressed pages, and
 596                          * will submit them to the elevator.
 597                          */
 598                         add_async_extent(async_cow, start, num_bytes,
 599                                         total_compressed, pages, nr_pages,
 600                                         compress_type);
 601
 602                         if (start + num_bytes < end) {
 603                                 start += num_bytes;
 604                                 pages = NULL;
 605                                 cond_resched();
 606                                 goto again;
 607                         }
 608                         return;
 609                 }
 610         }
 611         if (pages) {
 612                 /*
 613                  * the compression code ran but failed to make things smaller,
 614                  * free any pages it allocated and our page pointer array
 615                  */
 616                 for (i = 0; i < nr_pages; i++) {
 617                         WARN_ON(pages[i]->mapping);
 618                         put_page(pages[i]);
 619                 }
 620                 kfree(pages);
 621                 pages = NULL;
 622                 total_compressed = 0;
 623                 nr_pages = 0;
 624
 625                 /* flag the file so we don't compress in the future */
 626                 if (!btrfs_test_opt(fs_info, FORCE_COMPRESS) &&
 627                     !(BTRFS_I(inode)->force_compress)) {
 628                         BTRFS_I(inode)->flags |= BTRFS_INODE_NOCOMPRESS;
 629                 }
 630         }
 631 cleanup_and_bail_uncompressed:
 632         /*
 633          * No compression, but we still need to write the pages in the file
 634          * we've been given so far.  redirty the locked page if it corresponds
 635          * to our extent and set things up for the async work queue to run
 636          * cow_file_range to do the normal delalloc dance.
 637          */
 638         if (page_offset(locked_page) >= start &&
 639             page_offset(locked_page) <= end)
 640                 __set_page_dirty_nobuffers(locked_page);
 641                 /* unlocked later on in the async handlers */
 642
 643         if (redirty)
 644                 extent_range_redirty_for_io(inode, start, end);
 645         add_async_extent(async_cow, start, end - start + 1, 0, NULL, 0,
 646                          BTRFS_COMPRESS_NONE);
 647         *num_added += 1;
 648
 649         return;
 650
 651 free_pages_out:
 652         for (i = 0; i < nr_pages; i++) {
 653                 WARN_ON(pages[i]->mapping);
 654                 put_page(pages[i]);
 655         }
 656         kfree(pages);
 657 }
 658
 659 static void free_async_extent_pages(struct async_extent *async_extent)
 660 {
 661         int i;
 662
 663         if (!async_extent->pages)
 664                 return;
 665
 666         for (i = 0; i < async_extent->nr_pages; i++) {
 667                 WARN_ON(async_extent->pages[i]->mapping);
 668                 put_page(async_extent->pages[i]);
 669         }
 670         kfree(async_extent->pages);
 671         async_extent->nr_pages = 0;
 672         async_extent->pages = NULL;
 673 }
 674
 675 /*
 676  * phase two of compressed writeback.  This is the ordered portion
 677  * of the code, which only gets called in the order the work was
 678  * queued.  We walk all the async extents created by compress_file_range
 679  * and send them down to the disk.
 680  */
 681 static noinline void submit_compressed_extents(struct inode *inode,
 682                                               struct async_cow *async_cow)
 683 {
 684         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
 685         struct async_extent *async_extent;
 686         u64 alloc_hint = 0;
 687         struct btrfs_key ins;
 688         struct extent_map *em;
 689         struct btrfs_root *root = BTRFS_I(inode)->root;
 690         struct extent_io_tree *io_tree;
 691         int ret = 0;
 692
 693 again:
 694         while (!list_empty(&async_cow->extents)) {
 695                 async_extent = list_entry(async_cow->extents.next,
 696                                           struct async_extent, list);
 697                 list_del(&async_extent->list);
 698
 699                 io_tree = &BTRFS_I(inode)->io_tree;
 700
 701 retry:
 702                 /* did the compression code fall back to uncompressed IO? */
 703                 if (!async_extent->pages) {
 704                         int page_started = 0;
 705                         unsigned long nr_written = 0;
 706
 707                         lock_extent(io_tree, async_extent->start,
 708                                          async_extent->start +
 709                                          async_extent->ram_size - 1);
 710
 711                         /* allocate blocks */
 712                         ret = cow_file_range(inode, async_cow->locked_page,
 713                                              async_extent->start,
 714                                              async_extent->start +
 715                                              async_extent->ram_size - 1,
 716                                              async_extent->start +
 717                                              async_extent->ram_size - 1,
 718                                              &page_started, &nr_written, 0,
 719                                              NULL);
 720
 721                         /* JDM XXX */
 722
 723                         /*
 724                          * if page_started, cow_file_range inserted an
 725                          * inline extent and took care of all the unlocking
 726                          * and IO for us.  Otherwise, we need to submit
 727                          * all those pages down to the drive.
 728                          */
 729                         if (!page_started && !ret)
 730                                 extent_write_locked_range(io_tree,
 731                                                   inode, async_extent->start,
 732                                                   async_extent->start +
 733                                                   async_extent->ram_size - 1,
 734                                                   btrfs_get_extent,
 735                                                   WB_SYNC_ALL);
 736                         else if (ret)
 737                                 unlock_page(async_cow->locked_page);
 738                         kfree(async_extent);
 739                         cond_resched();
 740                         continue;
 741                 }
 742
 743                 lock_extent(io_tree, async_extent->start,
 744                             async_extent->start + async_extent->ram_size - 1);
 745
 746                 ret = btrfs_reserve_extent(root, async_extent->ram_size,
 747                                            async_extent->compressed_size,
 748                                            async_extent->compressed_size,
 749                                            0, alloc_hint, &ins, 1, 1);
 750                 if (ret) {
 751                         free_async_extent_pages(async_extent);
 752
 753                         if (ret == -ENOSPC) {
 754                                 unlock_extent(io_tree, async_extent->start,
 755                                               async_extent->start +
 756                                               async_extent->ram_size - 1);
 757
 758                                 /*
 759                                  * we need to redirty the pages if we decide to
 760                                  * fallback to uncompressed IO, otherwise we
 761                                  * will not submit these pages down to lower
 762                                  * layers.
 763                                  */
 764                                 extent_range_redirty_for_io(inode,
 765                                                 async_extent->start,
 766                                                 async_extent->start +
 767                                                 async_extent->ram_size - 1);
 768
 769                                 goto retry;
 770                         }
 771                         goto out_free;
 772                 }
 773                 /*
 774                  * here we're doing allocation and writeback of the
 775                  * compressed pages
 776                  */
 777                 em = create_io_em(inode, async_extent->start,
 778                                   async_extent->ram_size, /* len */
 779                                   async_extent->start, /* orig_start */
 780                                   ins.objectid, /* block_start */
 781                                   ins.offset, /* block_len */
 782                                   ins.offset, /* orig_block_len */
 783                                   async_extent->ram_size, /* ram_bytes */
 784                                   async_extent->compress_type,
 785                                   BTRFS_ORDERED_COMPRESSED);
 786                 if (IS_ERR(em))
 787                         /* ret value is not necessary due to void function */
 788                         goto out_free_reserve;
 789                 free_extent_map(em);
 790
 791                 ret = btrfs_add_ordered_extent_compress(inode,
 792                                                 async_extent->start,
 793                                                 ins.objectid,
 794                                                 async_extent->ram_size,
 795                                                 ins.offset,
 796                                                 BTRFS_ORDERED_COMPRESSED,
 797                                                 async_extent->compress_type);
 798                 if (ret) {
 799                         btrfs_drop_extent_cache(BTRFS_I(inode),
 800                                                 async_extent->start,
 801                                                 async_extent->start +
 802                                                 async_extent->ram_size - 1, 0);
 803                         goto out_free_reserve;
 804                 }
 805                 btrfs_dec_block_group_reservations(fs_info, ins.objectid);
 806
 807                 /*
 808                  * clear dirty, set writeback and unlock the pages.
 809                  */
 810                 extent_clear_unlock_delalloc(inode, async_extent->start,
 811                                 async_extent->start +
 812                                 async_extent->ram_size - 1,
 813                                 async_extent->start +
 814                                 async_extent->ram_size - 1,
 815                                 NULL, EXTENT_LOCKED | EXTENT_DELALLOC,
 816                                 PAGE_UNLOCK | PAGE_CLEAR_DIRTY |
 817                                 PAGE_SET_WRITEBACK);
 818                 ret = btrfs_submit_compressed_write(inode,
 819                                     async_extent->start,
 820                                     async_extent->ram_size,
 821                                     ins.objectid,
 822                                     ins.offset, async_extent->pages,
 823                                     async_extent->nr_pages);
 824                 if (ret) {
 825                         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
 826                         struct page *p = async_extent->pages[0];
 827                         const u64 start = async_extent->start;
 828                         const u64 end = start + async_extent->ram_size - 1;
 829
 830                         p->mapping = inode->i_mapping;
 831                         tree->ops->writepage_end_io_hook(p, start, end,
 832                                                          NULL, 0);
 833                         p->mapping = NULL;
 834                         extent_clear_unlock_delalloc(inode, start, end, end,
 835                                                      NULL, 0,
 836                                                      PAGE_END_WRITEBACK |
 837                                                      PAGE_SET_ERROR);
 838                         free_async_extent_pages(async_extent);
 839                 }
 840                 alloc_hint = ins.objectid + ins.offset;
 841                 kfree(async_extent);
 842                 cond_resched();
 843         }
 844         return;
 845 out_free_reserve:
 846         btrfs_dec_block_group_reservations(fs_info, ins.objectid);
 847         btrfs_free_reserved_extent(fs_info, ins.objectid, ins.offset, 1);
 848 out_free:
 849         extent_clear_unlock_delalloc(inode, async_extent->start,
 850                                      async_extent->start +
 851                                      async_extent->ram_size - 1,
 852                                      async_extent->start +
 853                                      async_extent->ram_size - 1,
 854                                      NULL, EXTENT_LOCKED | EXTENT_DELALLOC |
 855                                      EXTENT_DEFRAG | EXTENT_DO_ACCOUNTING,
 856                                      PAGE_UNLOCK | PAGE_CLEAR_DIRTY |
 857                                      PAGE_SET_WRITEBACK | PAGE_END_WRITEBACK |
 858                                      PAGE_SET_ERROR);
 859         free_async_extent_pages(async_extent);
 860         kfree(async_extent);
 861         goto again;
 862 }
 863
 864 static u64 get_extent_allocation_hint(struct inode *inode, u64 start,
 865                                       u64 num_bytes)
 866 {
 867         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
 868         struct extent_map *em;
 869         u64 alloc_hint = 0;
 870
 871         read_lock(&em_tree->lock);
 872         em = search_extent_mapping(em_tree, start, num_bytes);
 873         if (em) {
 874                 /*
 875                  * if block start isn't an actual block number then find the
 876                  * first block in this inode and use that as a hint.  If that
 877                  * block is also bogus then just don't worry about it.
 878                  */
 879                 if (em->block_start >= EXTENT_MAP_LAST_BYTE) {
 880                         free_extent_map(em);
 881                         em = search_extent_mapping(em_tree, 0, 0);
 882                         if (em && em->block_start < EXTENT_MAP_LAST_BYTE)
 883                                 alloc_hint = em->block_start;
 884                         if (em)
 885                                 free_extent_map(em);
 886                 } else {
 887                         alloc_hint = em->block_start;
 888                         free_extent_map(em);
 889                 }
 890         }
 891         read_unlock(&em_tree->lock);
 892
 893         return alloc_hint;
 894 }
 895
 896 /*
 897  * when extent_io.c finds a delayed allocation range in the file,
 898  * the call backs end up in this code.  The basic idea is to
 899  * allocate extents on disk for the range, and create ordered data structs
 900  * in ram to track those extents.
 901  *
 902  * locked_page is the page that writepage had locked already.  We use
 903  * it to make sure we don't do extra locks or unlocks.
 904  *
 905  * *page_started is set to one if we unlock locked_page and do everything
 906  * required to start IO on it.  It may be clean and already done with
 907  * IO when we return.
 908  */
 909 static noinline int cow_file_range(struct inode *inode,
 910                                    struct page *locked_page,
 911                                    u64 start, u64 end, u64 delalloc_end,
 912                                    int *page_started, unsigned long *nr_written,
 913                                    int unlock, struct btrfs_dedupe_hash *hash)
 914 {
 915         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
 916         struct btrfs_root *root = BTRFS_I(inode)->root;
 917         u64 alloc_hint = 0;
 918         u64 num_bytes;
 919         unsigned long ram_size;
 920         u64 disk_num_bytes;
 921         u64 cur_alloc_size;
 922         u64 blocksize = fs_info->sectorsize;
 923         struct btrfs_key ins;
 924         struct extent_map *em;
 925         int ret = 0;
 926
 927         if (btrfs_is_free_space_inode(BTRFS_I(inode))) {
 928                 WARN_ON_ONCE(1);
 929                 ret = -EINVAL;
 930                 goto out_unlock;
 931         }
 932
 933         num_bytes = ALIGN(end - start + 1, blocksize);
 934         num_bytes = max(blocksize,  num_bytes);
 935         disk_num_bytes = num_bytes;
 936
 937         inode_should_defrag(BTRFS_I(inode), start, end, num_bytes, SZ_64K);
 938
 939         if (start == 0) {
 940                 /* lets try to make an inline extent */
 941                 ret = cow_file_range_inline(root, inode, start, end, 0,
 942                                         BTRFS_COMPRESS_NONE, NULL);
 943                 if (ret == 0) {
 944                         extent_clear_unlock_delalloc(inode, start, end,
 945                                      delalloc_end, NULL,
 946                                      EXTENT_LOCKED | EXTENT_DELALLOC |
 947                                      EXTENT_DEFRAG, PAGE_UNLOCK |
 948                                      PAGE_CLEAR_DIRTY | PAGE_SET_WRITEBACK |
 949                                      PAGE_END_WRITEBACK);
 950                         btrfs_free_reserved_data_space_noquota(inode, start,
 951                                                 end - start + 1);
 952                         *nr_written = *nr_written +
 953                              (end - start + PAGE_SIZE) / PAGE_SIZE;
 954                         *page_started = 1;
 955                         goto out;
 956                 } else if (ret < 0) {
 957                         goto out_unlock;
 958                 }
 959         }
 960
 961         BUG_ON(disk_num_bytes >
 962                btrfs_super_total_bytes(fs_info->super_copy));
 963
 964         alloc_hint = get_extent_allocation_hint(inode, start, num_bytes);
 965         btrfs_drop_extent_cache(BTRFS_I(inode), start,
 966                         start + num_bytes - 1, 0);
 967
 968         while (disk_num_bytes > 0) {
 969                 unsigned long op;
 970
 971                 cur_alloc_size = disk_num_bytes;
 972                 ret = btrfs_reserve_extent(root, cur_alloc_size, cur_alloc_size,
 973                                            fs_info->sectorsize, 0, alloc_hint,
 974                                            &ins, 1, 1);
 975                 if (ret < 0)
 976                         goto out_unlock;
 977
 978                 ram_size = ins.offset;
 979                 em = create_io_em(inode, start, ins.offset, /* len */
 980                                   start, /* orig_start */
 981                                   ins.objectid, /* block_start */
 982                                   ins.offset, /* block_len */
 983                                   ins.offset, /* orig_block_len */
 984                                   ram_size, /* ram_bytes */
 985                                   BTRFS_COMPRESS_NONE, /* compress_type */
 986                                   BTRFS_ORDERED_REGULAR /* type */);
 987                 if (IS_ERR(em))
 988                         goto out_reserve;
 989                 free_extent_map(em);
 990
 991                 cur_alloc_size = ins.offset;
 992                 ret = btrfs_add_ordered_extent(inode, start, ins.objectid,
 993                                                ram_size, cur_alloc_size, 0);
 994                 if (ret)
 995                         goto out_drop_extent_cache;
 996
 997                 if (root->root_key.objectid ==
 998                     BTRFS_DATA_RELOC_TREE_OBJECTID) {
 999                         ret = btrfs_reloc_clone_csums(inode, start,
1000                                                       cur_alloc_size);
1001                         if (ret)
1002                                 goto out_drop_extent_cache;
1003                 }
1004
1005                 btrfs_dec_block_group_reservations(fs_info, ins.objectid);
1006
1007                 if (disk_num_bytes < cur_alloc_size)
1008                         break;
1009
1010                 /* we're not doing compressed IO, don't unlock the first
1011                  * page (which the caller expects to stay locked), don't
1012                  * clear any dirty bits and don't set any writeback bits
1013                  *
1014                  * Do set the Private2 bit so we know this page was properly
1015                  * setup for writepage
1016                  */
1017                 op = unlock ? PAGE_UNLOCK : 0;
1018                 op |= PAGE_SET_PRIVATE2;
1019
1020                 extent_clear_unlock_delalloc(inode, start,
1021                                              start + ram_size - 1,
1022                                              delalloc_end, locked_page,
1023                                              EXTENT_LOCKED | EXTENT_DELALLOC,
1024                                              op);
1025                 disk_num_bytes -= cur_alloc_size;
1026                 num_bytes -= cur_alloc_size;
1027                 alloc_hint = ins.objectid + ins.offset;
1028                 start += cur_alloc_size;
1029         }
1030 out:
1031         return ret;
1032
1033 out_drop_extent_cache:
1034         btrfs_drop_extent_cache(BTRFS_I(inode), start, start + ram_size - 1, 0);
1035 out_reserve:
1036         btrfs_dec_block_group_reservations(fs_info, ins.objectid);
1037         btrfs_free_reserved_extent(fs_info, ins.objectid, ins.offset, 1);
1038 out_unlock:
1039         extent_clear_unlock_delalloc(inode, start, end, delalloc_end,
1040                                      locked_page,
1041                                      EXTENT_LOCKED | EXTENT_DO_ACCOUNTING |
1042                                      EXTENT_DELALLOC | EXTENT_DEFRAG,
1043                                      PAGE_UNLOCK | PAGE_CLEAR_DIRTY |
1044                                      PAGE_SET_WRITEBACK | PAGE_END_WRITEBACK);
1045         goto out;
1046 }
1047
1048 /*
1049  * work queue call back to started compression on a file and pages
1050  */
1051 static noinline void async_cow_start(struct btrfs_work *work)
1052 {
1053         struct async_cow *async_cow;
1054         int num_added = 0;
1055         async_cow = container_of(work, struct async_cow, work);
1056
1057         compress_file_range(async_cow->inode, async_cow->locked_page,
1058                             async_cow->start, async_cow->end, async_cow,
1059                             &num_added);
1060         if (num_added == 0) {
1061                 btrfs_add_delayed_iput(async_cow->inode);
1062                 async_cow->inode = NULL;
1063         }
1064 }
1065
1066 /*
1067  * work queue call back to submit previously compressed pages
1068  */
1069 static noinline void async_cow_submit(struct btrfs_work *work)
1070 {
1071         struct btrfs_fs_info *fs_info;
1072         struct async_cow *async_cow;
1073         struct btrfs_root *root;
1074         unsigned long nr_pages;
1075
1076         async_cow = container_of(work, struct async_cow, work);
1077
1078         root = async_cow->root;
1079         fs_info = root->fs_info;
1080         nr_pages = (async_cow->end - async_cow->start + PAGE_SIZE) >>
1081                 PAGE_SHIFT;
1082
1083         /*
1084          * atomic_sub_return implies a barrier for waitqueue_active
1085          */
1086         if (atomic_sub_return(nr_pages, &fs_info->async_delalloc_pages) <
1087             5 * SZ_1M &&
1088             waitqueue_active(&fs_info->async_submit_wait))
1089                 wake_up(&fs_info->async_submit_wait);
1090
1091         if (async_cow->inode)
1092                 submit_compressed_extents(async_cow->inode, async_cow);
1093 }
1094
1095 static noinline void async_cow_free(struct btrfs_work *work)
1096 {
1097         struct async_cow *async_cow;
1098         async_cow = container_of(work, struct async_cow, work);
1099         if (async_cow->inode)
1100                 btrfs_add_delayed_iput(async_cow->inode);
1101         kfree(async_cow);
1102 }
1103
1104 static int cow_file_range_async(struct inode *inode, struct page *locked_page,
1105                                 u64 start, u64 end, int *page_started,
1106                                 unsigned long *nr_written)
1107 {
1108         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1109         struct async_cow *async_cow;
1110         struct btrfs_root *root = BTRFS_I(inode)->root;
1111         unsigned long nr_pages;
1112         u64 cur_end;
1113
1114         clear_extent_bit(&BTRFS_I(inode)->io_tree, start, end, EXTENT_LOCKED,
1115                          1, 0, NULL, GFP_NOFS);
1116         while (start < end) {
1117                 async_cow = kmalloc(sizeof(*async_cow), GFP_NOFS);
1118                 BUG_ON(!async_cow); /* -ENOMEM */
1119                 async_cow->inode = igrab(inode);
1120                 async_cow->root = root;
1121                 async_cow->locked_page = locked_page;
1122                 async_cow->start = start;
1123
1124                 if (BTRFS_I(inode)->flags & BTRFS_INODE_NOCOMPRESS &&
1125                     !btrfs_test_opt(fs_info, FORCE_COMPRESS))
1126                         cur_end = end;
1127                 else
1128                         cur_end = min(end, start + SZ_512K - 1);
1129
1130                 async_cow->end = cur_end;
1131                 INIT_LIST_HEAD(&async_cow->extents);
1132
1133                 btrfs_init_work(&async_cow->work,
1134                                 btrfs_delalloc_helper,
1135                                 async_cow_start, async_cow_submit,
1136                                 async_cow_free);
1137
1138                 nr_pages = (cur_end - start + PAGE_SIZE) >>
1139                         PAGE_SHIFT;
1140                 atomic_add(nr_pages, &fs_info->async_delalloc_pages);
1141
1142                 btrfs_queue_work(fs_info->delalloc_workers, &async_cow->work);
1143
1144                 while (atomic_read(&fs_info->async_submit_draining) &&
1145                        atomic_read(&fs_info->async_delalloc_pages)) {
1146                         wait_event(fs_info->async_submit_wait,
1147                                    (atomic_read(&fs_info->async_delalloc_pages) ==
1148                                     0));
1149                 }
1150
1151                 *nr_written += nr_pages;
1152                 start = cur_end + 1;
1153         }
1154         *page_started = 1;
1155         return 0;
1156 }
1157
1158 static noinline int csum_exist_in_range(struct btrfs_fs_info *fs_info,
1159                                         u64 bytenr, u64 num_bytes)
1160 {
1161         int ret;
1162         struct btrfs_ordered_sum *sums;
1163         LIST_HEAD(list);
1164
1165         ret = btrfs_lookup_csums_range(fs_info->csum_root, bytenr,
1166                                        bytenr + num_bytes - 1, &list, 0);
1167         if (ret == 0 && list_empty(&list))
1168                 return 0;
1169
1170         while (!list_empty(&list)) {
1171                 sums = list_entry(list.next, struct btrfs_ordered_sum, list);
1172                 list_del(&sums->list);
1173                 kfree(sums);
1174         }
1175         return 1;
1176 }
1177
1178 /*
1179  * when nowcow writeback call back.  This checks for snapshots or COW copies
1180  * of the extents that exist in the file, and COWs the file as required.
1181  *
1182  * If no cow copies or snapshots exist, we write directly to the existing
1183  * blocks on disk
1184  */
1185 static noinline int run_delalloc_nocow(struct inode *inode,
1186                                        struct page *locked_page,
1187                               u64 start, u64 end, int *page_started, int force,
1188                               unsigned long *nr_written)
1189 {
1190         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1191         struct btrfs_root *root = BTRFS_I(inode)->root;
1192         struct extent_buffer *leaf;
1193         struct btrfs_path *path;
1194         struct btrfs_file_extent_item *fi;
1195         struct btrfs_key found_key;
1196         struct extent_map *em;
1197         u64 cow_start;
1198         u64 cur_offset;
1199         u64 extent_end;
1200         u64 extent_offset;
1201         u64 disk_bytenr;
1202         u64 num_bytes;
1203         u64 disk_num_bytes;
1204         u64 ram_bytes;
1205         int extent_type;
1206         int ret, err;
1207         int type;
1208         int nocow;
1209         int check_prev = 1;
1210         bool nolock;
1211         u64 ino = btrfs_ino(BTRFS_I(inode));
1212
1213         path = btrfs_alloc_path();
1214         if (!path) {
1215                 extent_clear_unlock_delalloc(inode, start, end, end,
1216                                              locked_page,
1217                                              EXTENT_LOCKED | EXTENT_DELALLOC |
1218                                              EXTENT_DO_ACCOUNTING |
1219                                              EXTENT_DEFRAG, PAGE_UNLOCK |
1220                                              PAGE_CLEAR_DIRTY |
1221                                              PAGE_SET_WRITEBACK |
1222                                              PAGE_END_WRITEBACK);
1223                 return -ENOMEM;
1224         }
1225
1226         nolock = btrfs_is_free_space_inode(BTRFS_I(inode));
1227
1228         cow_start = (u64)-1;
1229         cur_offset = start;
1230         while (1) {
1231                 ret = btrfs_lookup_file_extent(NULL, root, path, ino,
1232                                                cur_offset, 0);
1233                 if (ret < 0)
1234                         goto error;
1235                 if (ret > 0 && path->slots[0] > 0 && check_prev) {
1236                         leaf = path->nodes[0];
1237                         btrfs_item_key_to_cpu(leaf, &found_key,
1238                                               path->slots[0] - 1);
1239                         if (found_key.objectid == ino &&
1240                             found_key.type == BTRFS_EXTENT_DATA_KEY)
1241                                 path->slots[0]--;
1242                 }
1243                 check_prev = 0;
1244 next_slot:
1245                 leaf = path->nodes[0];
1246                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
1247                         ret = btrfs_next_leaf(root, path);
1248                         if (ret < 0)
1249                                 goto error;
1250                         if (ret > 0)
1251                                 break;
1252                         leaf = path->nodes[0];
1253                 }
1254
1255                 nocow = 0;
1256                 disk_bytenr = 0;
1257                 num_bytes = 0;
1258                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
1259
1260                 if (found_key.objectid > ino)
1261                         break;
1262                 if (WARN_ON_ONCE(found_key.objectid < ino) ||
1263                     found_key.type < BTRFS_EXTENT_DATA_KEY) {
1264                         path->slots[0]++;
1265                         goto next_slot;
1266                 }
1267                 if (found_key.type > BTRFS_EXTENT_DATA_KEY ||
1268                     found_key.offset > end)
1269                         break;
1270
1271                 if (found_key.offset > cur_offset) {
1272                         extent_end = found_key.offset;
1273                         extent_type = 0;
1274                         goto out_check;
1275                 }
1276
1277                 fi = btrfs_item_ptr(leaf, path->slots[0],
1278                                     struct btrfs_file_extent_item);
1279                 extent_type = btrfs_file_extent_type(leaf, fi);
1280
1281                 ram_bytes = btrfs_file_extent_ram_bytes(leaf, fi);
1282                 if (extent_type == BTRFS_FILE_EXTENT_REG ||
1283                     extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
1284                         disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1285                         extent_offset = btrfs_file_extent_offset(leaf, fi);
1286                         extent_end = found_key.offset +
1287                                 btrfs_file_extent_num_bytes(leaf, fi);
1288                         disk_num_bytes =
1289                                 btrfs_file_extent_disk_num_bytes(leaf, fi);
1290                         if (extent_end <= start) {
1291                                 path->slots[0]++;
1292                                 goto next_slot;
1293                         }
1294                         if (disk_bytenr == 0)
1295                                 goto out_check;
1296                         if (btrfs_file_extent_compression(leaf, fi) ||
1297                             btrfs_file_extent_encryption(leaf, fi) ||
1298                             btrfs_file_extent_other_encoding(leaf, fi))
1299                                 goto out_check;
1300                         if (extent_type == BTRFS_FILE_EXTENT_REG && !force)
1301                                 goto out_check;
1302                         if (btrfs_extent_readonly(fs_info, disk_bytenr))
1303                                 goto out_check;
1304                         if (btrfs_cross_ref_exist(root, ino,
1305                                                   found_key.offset -
1306                                                   extent_offset, disk_bytenr))
1307                                 goto out_check;
1308                         disk_bytenr += extent_offset;
1309                         disk_bytenr += cur_offset - found_key.offset;
1310                         num_bytes = min(end + 1, extent_end) - cur_offset;
1311                         /*
1312                          * if there are pending snapshots for this root,
1313                          * we fall into common COW way.
1314                          */
1315                         if (!nolock) {
1316                                 err = btrfs_start_write_no_snapshoting(root);
1317                                 if (!err)
1318                                         goto out_check;
1319                         }
1320                         /*
1321                          * force cow if csum exists in the range.
1322                          * this ensure that csum for a given extent are
1323                          * either valid or do not exist.
1324                          */
1325                         if (csum_exist_in_range(fs_info, disk_bytenr,
1326                                                 num_bytes)) {
1327                                 if (!nolock)
1328                                         btrfs_end_write_no_snapshoting(root);
1329                                 goto out_check;
1330                         }
1331                         if (!btrfs_inc_nocow_writers(fs_info, disk_bytenr)) {
1332                                 if (!nolock)
1333                                         btrfs_end_write_no_snapshoting(root);
1334                                 goto out_check;
1335                         }
1336                         nocow = 1;
1337                 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
1338                         extent_end = found_key.offset +
1339                                 btrfs_file_extent_inline_len(leaf,
1340                                                      path->slots[0], fi);
1341                         extent_end = ALIGN(extent_end,
1342                                            fs_info->sectorsize);
1343                 } else {
1344                         BUG_ON(1);
1345                 }
1346 out_check:
1347                 if (extent_end <= start) {
1348                         path->slots[0]++;
1349                         if (!nolock && nocow)
1350                                 btrfs_end_write_no_snapshoting(root);
1351                         if (nocow)
1352                                 btrfs_dec_nocow_writers(fs_info, disk_bytenr);
1353                         goto next_slot;
1354                 }
1355                 if (!nocow) {
1356                         if (cow_start == (u64)-1)
1357                                 cow_start = cur_offset;
1358                         cur_offset = extent_end;
1359                         if (cur_offset > end)
1360                                 break;
1361                         path->slots[0]++;
1362                         goto next_slot;
1363                 }
1364
1365                 btrfs_release_path(path);
1366                 if (cow_start != (u64)-1) {
1367                         ret = cow_file_range(inode, locked_page,
1368                                              cow_start, found_key.offset - 1,
1369                                              end, page_started, nr_written, 1,
1370                                              NULL);
1371                         if (ret) {
1372                                 if (!nolock && nocow)
1373                                         btrfs_end_write_no_snapshoting(root);
1374                                 if (nocow)
1375                                         btrfs_dec_nocow_writers(fs_info,
1376                                                                 disk_bytenr);
1377                                 goto error;
1378                         }
1379                         cow_start = (u64)-1;
1380                 }
1381
1382                 if (extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
1383                         u64 orig_start = found_key.offset - extent_offset;
1384
1385                         em = create_io_em(inode, cur_offset, num_bytes,
1386                                           orig_start,
1387                                           disk_bytenr, /* block_start */
1388                                           num_bytes, /* block_len */
1389                                           disk_num_bytes, /* orig_block_len */
1390                                           ram_bytes, BTRFS_COMPRESS_NONE,
1391                                           BTRFS_ORDERED_PREALLOC);
1392                         if (IS_ERR(em)) {
1393                                 if (!nolock && nocow)
1394                                         btrfs_end_write_no_snapshoting(root);
1395                                 if (nocow)
1396                                         btrfs_dec_nocow_writers(fs_info,
1397                                                                 disk_bytenr);
1398                                 ret = PTR_ERR(em);
1399                                 goto error;
1400                         }
1401                         free_extent_map(em);
1402                 }
1403
1404                 if (extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
1405                         type = BTRFS_ORDERED_PREALLOC;
1406                 } else {
1407                         type = BTRFS_ORDERED_NOCOW;
1408                 }
1409
1410                 ret = btrfs_add_ordered_extent(inode, cur_offset, disk_bytenr,
1411                                                num_bytes, num_bytes, type);
1412                 if (nocow)
1413                         btrfs_dec_nocow_writers(fs_info, disk_bytenr);
1414                 BUG_ON(ret); /* -ENOMEM */
1415
1416                 if (root->root_key.objectid ==
1417                     BTRFS_DATA_RELOC_TREE_OBJECTID) {
1418                         ret = btrfs_reloc_clone_csums(inode, cur_offset,
1419                                                       num_bytes);
1420                         if (ret) {
1421                                 if (!nolock && nocow)
1422                                         btrfs_end_write_no_snapshoting(root);
1423                                 goto error;
1424                         }
1425                 }
1426
1427                 extent_clear_unlock_delalloc(inode, cur_offset,
1428                                              cur_offset + num_bytes - 1, end,
1429                                              locked_page, EXTENT_LOCKED |
1430                                              EXTENT_DELALLOC |
1431                                              EXTENT_CLEAR_DATA_RESV,
1432                                              PAGE_UNLOCK | PAGE_SET_PRIVATE2);
1433
1434                 if (!nolock && nocow)
1435                         btrfs_end_write_no_snapshoting(root);
1436                 cur_offset = extent_end;
1437                 if (cur_offset > end)
1438                         break;
1439         }
1440         btrfs_release_path(path);
1441
1442         if (cur_offset <= end && cow_start == (u64)-1) {
1443                 cow_start = cur_offset;
1444                 cur_offset = end;
1445         }
1446
1447         if (cow_start != (u64)-1) {
1448                 ret = cow_file_range(inode, locked_page, cow_start, end, end,
1449                                      page_started, nr_written, 1, NULL);
1450                 if (ret)
1451                         goto error;
1452         }
1453
1454 error:
1455         if (ret && cur_offset < end)
1456                 extent_clear_unlock_delalloc(inode, cur_offset, end, end,
1457                                              locked_page, EXTENT_LOCKED |
1458                                              EXTENT_DELALLOC | EXTENT_DEFRAG |
1459                                              EXTENT_DO_ACCOUNTING, PAGE_UNLOCK |
1460                                              PAGE_CLEAR_DIRTY |
1461                                              PAGE_SET_WRITEBACK |
1462                                              PAGE_END_WRITEBACK);
1463         btrfs_free_path(path);
1464         return ret;
1465 }
1466
1467 static inline int need_force_cow(struct inode *inode, u64 start, u64 end)
1468 {
1469
1470         if (!(BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW) &&
1471             !(BTRFS_I(inode)->flags & BTRFS_INODE_PREALLOC))
1472                 return 0;
1473
1474         /*
1475          * @defrag_bytes is a hint value, no spinlock held here,
1476          * if is not zero, it means the file is defragging.
1477          * Force cow if given extent needs to be defragged.
1478          */
1479         if (BTRFS_I(inode)->defrag_bytes &&
1480             test_range_bit(&BTRFS_I(inode)->io_tree, start, end,
1481                            EXTENT_DEFRAG, 0, NULL))
1482                 return 1;
1483
1484         return 0;
1485 }
1486
1487 /*
1488  * extent_io.c call back to do delayed allocation processing
1489  */
1490 static int run_delalloc_range(struct inode *inode, struct page *locked_page,
1491                               u64 start, u64 end, int *page_started,
1492                               unsigned long *nr_written)
1493 {
1494         int ret;
1495         int force_cow = need_force_cow(inode, start, end);
1496
1497         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW && !force_cow) {
1498                 ret = run_delalloc_nocow(inode, locked_page, start, end,
1499                                          page_started, 1, nr_written);
1500         } else if (BTRFS_I(inode)->flags & BTRFS_INODE_PREALLOC && !force_cow) {
1501                 ret = run_delalloc_nocow(inode, locked_page, start, end,
1502                                          page_started, 0, nr_written);
1503         } else if (!inode_need_compress(inode)) {
1504                 ret = cow_file_range(inode, locked_page, start, end, end,
1505                                       page_started, nr_written, 1, NULL);
1506         } else {
1507                 set_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
1508                         &BTRFS_I(inode)->runtime_flags);
1509                 ret = cow_file_range_async(inode, locked_page, start, end,
1510                                            page_started, nr_written);
1511         }
1512         return ret;
1513 }
1514
1515 static void btrfs_split_extent_hook(struct inode *inode,
1516                                     struct extent_state *orig, u64 split)
1517 {
1518         u64 size;
1519
1520         /* not delalloc, ignore it */
1521         if (!(orig->state & EXTENT_DELALLOC))
1522                 return;
1523
1524         size = orig->end - orig->start + 1;
1525         if (size > BTRFS_MAX_EXTENT_SIZE) {
1526                 u32 num_extents;
1527                 u64 new_size;
1528
1529                 /*
1530                  * See the explanation in btrfs_merge_extent_hook, the same
1531                  * applies here, just in reverse.
1532                  */
1533                 new_size = orig->end - split + 1;
1534                 num_extents = count_max_extents(new_size);
1535                 new_size = split - orig->start;
1536                 num_extents += count_max_extents(new_size);
1537                 if (count_max_extents(size) >= num_extents)
1538                         return;
1539         }
1540
1541         spin_lock(&BTRFS_I(inode)->lock);
1542         BTRFS_I(inode)->outstanding_extents++;
1543         spin_unlock(&BTRFS_I(inode)->lock);
1544 }
1545
1546 /*
1547  * extent_io.c merge_extent_hook, used to track merged delayed allocation
1548  * extents so we can keep track of new extents that are just merged onto old
1549  * extents, such as when we are doing sequential writes, so we can properly
1550  * account for the metadata space we'll need.
1551  */
1552 static void btrfs_merge_extent_hook(struct inode *inode,
1553                                     struct extent_state *new,
1554                                     struct extent_state *other)
1555 {
1556         u64 new_size, old_size;
1557         u32 num_extents;
1558
1559         /* not delalloc, ignore it */
1560         if (!(other->state & EXTENT_DELALLOC))
1561                 return;
1562
1563         if (new->start > other->start)
1564                 new_size = new->end - other->start + 1;
1565         else
1566                 new_size = other->end - new->start + 1;
1567
1568         /* we're not bigger than the max, unreserve the space and go */
1569         if (new_size <= BTRFS_MAX_EXTENT_SIZE) {
1570                 spin_lock(&BTRFS_I(inode)->lock);
1571                 BTRFS_I(inode)->outstanding_extents--;
1572                 spin_unlock(&BTRFS_I(inode)->lock);
1573                 return;
1574         }
1575
1576         /*
1577          * We have to add up either side to figure out how many extents were
1578          * accounted for before we merged into one big extent.  If the number of
1579          * extents we accounted for is <= the amount we need for the new range
1580          * then we can return, otherwise drop.  Think of it like this
1581          *
1582          * [ 4k][MAX_SIZE]
1583          *
1584          * So we've grown the extent by a MAX_SIZE extent, this would mean we
1585          * need 2 outstanding extents, on one side we have 1 and the other side
1586          * we have 1 so they are == and we can return.  But in this case
1587          *
1588          * [MAX_SIZE+4k][MAX_SIZE+4k]
1589          *
1590          * Each range on their own accounts for 2 extents, but merged together
1591          * they are only 3 extents worth of accounting, so we need to drop in
1592          * this case.
1593          */
1594         old_size = other->end - other->start + 1;
1595         num_extents = count_max_extents(old_size);
1596         old_size = new->end - new->start + 1;
1597         num_extents += count_max_extents(old_size);
1598         if (count_max_extents(new_size) >= num_extents)
1599                 return;
1600
1601         spin_lock(&BTRFS_I(inode)->lock);
1602         BTRFS_I(inode)->outstanding_extents--;
1603         spin_unlock(&BTRFS_I(inode)->lock);
1604 }
1605
1606 static void btrfs_add_delalloc_inodes(struct btrfs_root *root,
1607                                       struct inode *inode)
1608 {
1609         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1610
1611         spin_lock(&root->delalloc_lock);
1612         if (list_empty(&BTRFS_I(inode)->delalloc_inodes)) {
1613                 list_add_tail(&BTRFS_I(inode)->delalloc_inodes,
1614                               &root->delalloc_inodes);
1615                 set_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1616                         &BTRFS_I(inode)->runtime_flags);
1617                 root->nr_delalloc_inodes++;
1618                 if (root->nr_delalloc_inodes == 1) {
1619                         spin_lock(&fs_info->delalloc_root_lock);
1620                         BUG_ON(!list_empty(&root->delalloc_root));
1621                         list_add_tail(&root->delalloc_root,
1622                                       &fs_info->delalloc_roots);
1623                         spin_unlock(&fs_info->delalloc_root_lock);
1624                 }
1625         }
1626         spin_unlock(&root->delalloc_lock);
1627 }
1628
1629 static void btrfs_del_delalloc_inode(struct btrfs_root *root,
1630                                      struct btrfs_inode *inode)
1631 {
1632         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
1633
1634         spin_lock(&root->delalloc_lock);
1635         if (!list_empty(&inode->delalloc_inodes)) {
1636                 list_del_init(&inode->delalloc_inodes);
1637                 clear_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1638                           &inode->runtime_flags);
1639                 root->nr_delalloc_inodes--;
1640                 if (!root->nr_delalloc_inodes) {
1641                         spin_lock(&fs_info->delalloc_root_lock);
1642                         BUG_ON(list_empty(&root->delalloc_root));
1643                         list_del_init(&root->delalloc_root);
1644                         spin_unlock(&fs_info->delalloc_root_lock);
1645                 }
1646         }
1647         spin_unlock(&root->delalloc_lock);
1648 }
1649
1650 /*
1651  * extent_io.c set_bit_hook, used to track delayed allocation
1652  * bytes in this file, and to maintain the list of inodes that
1653  * have pending delalloc work to be done.
1654  */
1655 static void btrfs_set_bit_hook(struct inode *inode,
1656                                struct extent_state *state, unsigned *bits)
1657 {
1658
1659         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1660
1661         if ((*bits & EXTENT_DEFRAG) && !(*bits & EXTENT_DELALLOC))
1662                 WARN_ON(1);
1663         /*
1664          * set_bit and clear bit hooks normally require _irqsave/restore
1665          * but in this case, we are only testing for the DELALLOC
1666          * bit, which is only set or cleared with irqs on
1667          */
1668         if (!(state->state & EXTENT_DELALLOC) && (*bits & EXTENT_DELALLOC)) {
1669                 struct btrfs_root *root = BTRFS_I(inode)->root;
1670                 u64 len = state->end + 1 - state->start;
1671                 bool do_list = !btrfs_is_free_space_inode(BTRFS_I(inode));
1672
1673                 if (*bits & EXTENT_FIRST_DELALLOC) {
1674                         *bits &= ~EXTENT_FIRST_DELALLOC;
1675                 } else {
1676                         spin_lock(&BTRFS_I(inode)->lock);
1677                         BTRFS_I(inode)->outstanding_extents++;
1678                         spin_unlock(&BTRFS_I(inode)->lock);
1679                 }
1680
1681                 /* For sanity tests */
1682                 if (btrfs_is_testing(fs_info))
1683                         return;
1684
1685                 __percpu_counter_add(&fs_info->delalloc_bytes, len,
1686                                      fs_info->delalloc_batch);
1687                 spin_lock(&BTRFS_I(inode)->lock);
1688                 BTRFS_I(inode)->delalloc_bytes += len;
1689                 if (*bits & EXTENT_DEFRAG)
1690                         BTRFS_I(inode)->defrag_bytes += len;
1691                 if (do_list && !test_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1692                                          &BTRFS_I(inode)->runtime_flags))
1693                         btrfs_add_delalloc_inodes(root, inode);
1694                 spin_unlock(&BTRFS_I(inode)->lock);
1695         }
1696 }
1697
1698 /*
1699  * extent_io.c clear_bit_hook, see set_bit_hook for why
1700  */
1701 static void btrfs_clear_bit_hook(struct btrfs_inode *inode,
1702                                  struct extent_state *state,
1703                                  unsigned *bits)
1704 {
1705         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
1706         u64 len = state->end + 1 - state->start;
1707         u32 num_extents = count_max_extents(len);
1708
1709         spin_lock(&inode->lock);
1710         if ((state->state & EXTENT_DEFRAG) && (*bits & EXTENT_DEFRAG))
1711                 inode->defrag_bytes -= len;
1712         spin_unlock(&inode->lock);
1713
1714         /*
1715          * set_bit and clear bit hooks normally require _irqsave/restore
1716          * but in this case, we are only testing for the DELALLOC
1717          * bit, which is only set or cleared with irqs on
1718          */
1719         if ((state->state & EXTENT_DELALLOC) && (*bits & EXTENT_DELALLOC)) {
1720                 struct btrfs_root *root = inode->root;
1721                 bool do_list = !btrfs_is_free_space_inode(inode);
1722
1723                 if (*bits & EXTENT_FIRST_DELALLOC) {
1724                         *bits &= ~EXTENT_FIRST_DELALLOC;
1725                 } else if (!(*bits & EXTENT_DO_ACCOUNTING)) {
1726                         spin_lock(&inode->lock);
1727                         inode->outstanding_extents -= num_extents;
1728                         spin_unlock(&inode->lock);
1729                 }
1730
1731                 /*
1732                  * We don't reserve metadata space for space cache inodes so we
1733                  * don't need to call dellalloc_release_metadata if there is an
1734                  * error.
1735                  */
1736                 if (*bits & EXTENT_DO_ACCOUNTING &&
1737                     root != fs_info->tree_root)
1738                         btrfs_delalloc_release_metadata(inode, len);
1739
1740                 /* For sanity tests. */
1741                 if (btrfs_is_testing(fs_info))
1742                         return;
1743
1744                 if (root->root_key.objectid != BTRFS_DATA_RELOC_TREE_OBJECTID
1745                     && do_list && !(state->state & EXTENT_NORESERVE)
1746                     && (*bits & (EXTENT_DO_ACCOUNTING |
1747                     EXTENT_CLEAR_DATA_RESV)))
1748                         btrfs_free_reserved_data_space_noquota(
1749                                         &inode->vfs_inode,
1750                                         state->start, len);
1751
1752                 __percpu_counter_add(&fs_info->delalloc_bytes, -len,
1753                                      fs_info->delalloc_batch);
1754                 spin_lock(&inode->lock);
1755                 inode->delalloc_bytes -= len;
1756                 if (do_list && inode->delalloc_bytes == 0 &&
1757                     test_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1758                                         &inode->runtime_flags))
1759                         btrfs_del_delalloc_inode(root, inode);
1760                 spin_unlock(&inode->lock);
1761         }
1762 }
1763
1764 /*
1765  * extent_io.c merge_bio_hook, this must check the chunk tree to make sure
1766  * we don't create bios that span stripes or chunks
1767  *
1768  * return 1 if page cannot be merged to bio
1769  * return 0 if page can be merged to bio
1770  * return error otherwise
1771  */
1772 int btrfs_merge_bio_hook(struct page *page, unsigned long offset,
1773                          size_t size, struct bio *bio,
1774                          unsigned long bio_flags)
1775 {
1776         struct inode *inode = page->mapping->host;
1777         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1778         u64 logical = (u64)bio->bi_iter.bi_sector << 9;
1779         u64 length = 0;
1780         u64 map_length;
1781         int ret;
1782
1783         if (bio_flags & EXTENT_BIO_COMPRESSED)
1784                 return 0;
1785
1786         length = bio->bi_iter.bi_size;
1787         map_length = length;
1788         ret = btrfs_map_block(fs_info, btrfs_op(bio), logical, &map_length,
1789                               NULL, 0);
1790         if (ret < 0)
1791                 return ret;
1792         if (map_length < length + size)
1793                 return 1;
1794         return 0;
1795 }
1796
1797 /*
1798  * in order to insert checksums into the metadata in large chunks,
1799  * we wait until bio submission time.   All the pages in the bio are
1800  * checksummed and sums are attached onto the ordered extent record.
1801  *
1802  * At IO completion time the cums attached on the ordered extent record
1803  * are inserted into the btree
1804  */
1805 static int __btrfs_submit_bio_start(struct inode *inode, struct bio *bio,
1806                                     int mirror_num, unsigned long bio_flags,
1807                                     u64 bio_offset)
1808 {
1809         int ret = 0;
1810
1811         ret = btrfs_csum_one_bio(inode, bio, 0, 0);
1812         BUG_ON(ret); /* -ENOMEM */
1813         return 0;
1814 }
1815
1816 /*
1817  * in order to insert checksums into the metadata in large chunks,
1818  * we wait until bio submission time.   All the pages in the bio are
1819  * checksummed and sums are attached onto the ordered extent record.
1820  *
1821  * At IO completion time the cums attached on the ordered extent record
1822  * are inserted into the btree
1823  */
1824 static int __btrfs_submit_bio_done(struct inode *inode, struct bio *bio,
1825                           int mirror_num, unsigned long bio_flags,
1826                           u64 bio_offset)
1827 {
1828         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1829         int ret;
1830
1831         ret = btrfs_map_bio(fs_info, bio, mirror_num, 1);
1832         if (ret) {
1833                 bio->bi_error = ret;
1834                 bio_endio(bio);
1835         }
1836         return ret;
1837 }
1838
1839 /*
1840  * extent_io.c submission hook. This does the right thing for csum calculation
1841  * on write, or reading the csums from the tree before a read
1842  */
1843 static int btrfs_submit_bio_hook(struct inode *inode, struct bio *bio,
1844                           int mirror_num, unsigned long bio_flags,
1845                           u64 bio_offset)
1846 {
1847         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1848         struct btrfs_root *root = BTRFS_I(inode)->root;
1849         enum btrfs_wq_endio_type metadata = BTRFS_WQ_ENDIO_DATA;
1850         int ret = 0;
1851         int skip_sum;
1852         int async = !atomic_read(&BTRFS_I(inode)->sync_writers);
1853
1854         skip_sum = BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM;
1855
1856         if (btrfs_is_free_space_inode(BTRFS_I(inode)))
1857                 metadata = BTRFS_WQ_ENDIO_FREE_SPACE;
1858
1859         if (bio_op(bio) != REQ_OP_WRITE) {
1860                 ret = btrfs_bio_wq_end_io(fs_info, bio, metadata);
1861                 if (ret)
1862                         goto out;
1863
1864                 if (bio_flags & EXTENT_BIO_COMPRESSED) {
1865                         ret = btrfs_submit_compressed_read(inode, bio,
1866                                                            mirror_num,
1867                                                            bio_flags);
1868                         goto out;
1869                 } else if (!skip_sum) {
1870                         ret = btrfs_lookup_bio_sums(inode, bio, NULL);
1871                         if (ret)
1872                                 goto out;
1873                 }
1874                 goto mapit;
1875         } else if (async && !skip_sum) {
1876                 /* csum items have already been cloned */
1877                 if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
1878                         goto mapit;
1879                 /* we're doing a write, do the async checksumming */
1880                 ret = btrfs_wq_submit_bio(fs_info, inode, bio, mirror_num,
1881                                           bio_flags, bio_offset,
1882                                           __btrfs_submit_bio_start,
1883                                           __btrfs_submit_bio_done);
1884                 goto out;
1885         } else if (!skip_sum) {
1886                 ret = btrfs_csum_one_bio(inode, bio, 0, 0);
1887                 if (ret)
1888                         goto out;
1889         }
1890
1891 mapit:
1892         ret = btrfs_map_bio(fs_info, bio, mirror_num, 0);
1893
1894 out:
1895         if (ret < 0) {
1896                 bio->bi_error = ret;
1897                 bio_endio(bio);
1898         }
1899         return ret;
1900 }
1901
1902 /*
1903  * given a list of ordered sums record them in the inode.  This happens
1904  * at IO completion time based on sums calculated at bio submission time.
1905  */
1906 static noinline int add_pending_csums(struct btrfs_trans_handle *trans,
1907                              struct inode *inode, struct list_head *list)
1908 {
1909         struct btrfs_ordered_sum *sum;
1910
1911         list_for_each_entry(sum, list, list) {
1912                 trans->adding_csums = 1;
1913                 btrfs_csum_file_blocks(trans,
1914                        BTRFS_I(inode)->root->fs_info->csum_root, sum);
1915                 trans->adding_csums = 0;
1916         }
1917         return 0;
1918 }
1919
1920 int btrfs_set_extent_delalloc(struct inode *inode, u64 start, u64 end,
1921                               struct extent_state **cached_state, int dedupe)
1922 {
1923         WARN_ON((end & (PAGE_SIZE - 1)) == 0);
1924         return set_extent_delalloc(&BTRFS_I(inode)->io_tree, start, end,
1925                                    cached_state);
1926 }
1927
1928 /* see btrfs_writepage_start_hook for details on why this is required */
1929 struct btrfs_writepage_fixup {
1930         struct page *page;
1931         struct btrfs_work work;
1932 };
1933
1934 static void btrfs_writepage_fixup_worker(struct btrfs_work *work)
1935 {
1936         struct btrfs_writepage_fixup *fixup;
1937         struct btrfs_ordered_extent *ordered;
1938         struct extent_state *cached_state = NULL;
1939         struct page *page;
1940         struct inode *inode;
1941         u64 page_start;
1942         u64 page_end;
1943         int ret;
1944
1945         fixup = container_of(work, struct btrfs_writepage_fixup, work);
1946         page = fixup->page;
1947 again:
1948         lock_page(page);
1949         if (!page->mapping || !PageDirty(page) || !PageChecked(page)) {
1950                 ClearPageChecked(page);
1951                 goto out_page;
1952         }
1953
1954         inode = page->mapping->host;
1955         page_start = page_offset(page);
1956         page_end = page_offset(page) + PAGE_SIZE - 1;
1957
1958         lock_extent_bits(&BTRFS_I(inode)->io_tree, page_start, page_end,
1959                          &cached_state);
1960
1961         /* already ordered? We're done */
1962         if (PagePrivate2(page))
1963                 goto out;
1964
1965         ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), page_start,
1966                                         PAGE_SIZE);
1967         if (ordered) {
1968                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, page_start,
1969                                      page_end, &cached_state, GFP_NOFS);
1970                 unlock_page(page);
1971                 btrfs_start_ordered_extent(inode, ordered, 1);
1972                 btrfs_put_ordered_extent(ordered);
1973                 goto again;
1974         }
1975
1976         ret = btrfs_delalloc_reserve_space(inode, page_start,
1977                                            PAGE_SIZE);
1978         if (ret) {
1979                 mapping_set_error(page->mapping, ret);
1980                 end_extent_writepage(page, ret, page_start, page_end);
1981                 ClearPageChecked(page);
1982                 goto out;
1983          }
1984
1985         btrfs_set_extent_delalloc(inode, page_start, page_end, &cached_state,
1986                                   0);
1987         ClearPageChecked(page);
1988         set_page_dirty(page);
1989 out:
1990         unlock_extent_cached(&BTRFS_I(inode)->io_tree, page_start, page_end,
1991                              &cached_state, GFP_NOFS);
1992 out_page:
1993         unlock_page(page);
1994         put_page(page);
1995         kfree(fixup);
1996 }
1997
1998 /*
1999  * There are a few paths in the higher layers of the kernel that directly
2000  * set the page dirty bit without asking the filesystem if it is a
2001  * good idea.  This causes problems because we want to make sure COW
2002  * properly happens and the data=ordered rules are followed.
2003  *
2004  * In our case any range that doesn't have the ORDERED bit set
2005  * hasn't been properly setup for IO.  We kick off an async process
2006  * to fix it up.  The async helper will wait for ordered extents, set
2007  * the delalloc bit and make it safe to write the page.
2008  */
2009 static int btrfs_writepage_start_hook(struct page *page, u64 start, u64 end)
2010 {
2011         struct inode *inode = page->mapping->host;
2012         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2013         struct btrfs_writepage_fixup *fixup;
2014
2015         /* this page is properly in the ordered list */
2016         if (TestClearPagePrivate2(page))
2017                 return 0;
2018
2019         if (PageChecked(page))
2020                 return -EAGAIN;
2021
2022         fixup = kzalloc(sizeof(*fixup), GFP_NOFS);
2023         if (!fixup)
2024                 return -EAGAIN;
2025
2026         SetPageChecked(page);
2027         get_page(page);
2028         btrfs_init_work(&fixup->work, btrfs_fixup_helper,
2029                         btrfs_writepage_fixup_worker, NULL, NULL);
2030         fixup->page = page;
2031         btrfs_queue_work(fs_info->fixup_workers, &fixup->work);
2032         return -EBUSY;
2033 }
2034
2035 static int insert_reserved_file_extent(struct btrfs_trans_handle *trans,
2036                                        struct inode *inode, u64 file_pos,
2037                                        u64 disk_bytenr, u64 disk_num_bytes,
2038                                        u64 num_bytes, u64 ram_bytes,
2039                                        u8 compression, u8 encryption,
2040                                        u16 other_encoding, int extent_type)
2041 {
2042         struct btrfs_root *root = BTRFS_I(inode)->root;
2043         struct btrfs_file_extent_item *fi;
2044         struct btrfs_path *path;
2045         struct extent_buffer *leaf;
2046         struct btrfs_key ins;
2047         int extent_inserted = 0;
2048         int ret;
2049
2050         path = btrfs_alloc_path();
2051         if (!path)
2052                 return -ENOMEM;
2053
2054         /*
2055          * we may be replacing one extent in the tree with another.
2056          * The new extent is pinned in the extent map, and we don't want
2057          * to drop it from the cache until it is completely in the btree.
2058          *
2059          * So, tell btrfs_drop_extents to leave this extent in the cache.
2060          * the caller is expected to unpin it and allow it to be merged
2061          * with the others.
2062          */
2063         ret = __btrfs_drop_extents(trans, root, inode, path, file_pos,
2064                                    file_pos + num_bytes, NULL, 0,
2065                                    1, sizeof(*fi), &extent_inserted);
2066         if (ret)
2067                 goto out;
2068
2069         if (!extent_inserted) {
2070                 ins.objectid = btrfs_ino(BTRFS_I(inode));
2071                 ins.offset = file_pos;
2072                 ins.type = BTRFS_EXTENT_DATA_KEY;
2073
2074                 path->leave_spinning = 1;
2075                 ret = btrfs_insert_empty_item(trans, root, path, &ins,
2076                                               sizeof(*fi));
2077                 if (ret)
2078                         goto out;
2079         }
2080         leaf = path->nodes[0];
2081         fi = btrfs_item_ptr(leaf, path->slots[0],
2082                             struct btrfs_file_extent_item);
2083         btrfs_set_file_extent_generation(leaf, fi, trans->transid);
2084         btrfs_set_file_extent_type(leaf, fi, extent_type);
2085         btrfs_set_file_extent_disk_bytenr(leaf, fi, disk_bytenr);
2086         btrfs_set_file_extent_disk_num_bytes(leaf, fi, disk_num_bytes);
2087         btrfs_set_file_extent_offset(leaf, fi, 0);
2088         btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2089         btrfs_set_file_extent_ram_bytes(leaf, fi, ram_bytes);
2090         btrfs_set_file_extent_compression(leaf, fi, compression);
2091         btrfs_set_file_extent_encryption(leaf, fi, encryption);
2092         btrfs_set_file_extent_other_encoding(leaf, fi, other_encoding);
2093
2094         btrfs_mark_buffer_dirty(leaf);
2095         btrfs_release_path(path);
2096
2097         inode_add_bytes(inode, num_bytes);
2098
2099         ins.objectid = disk_bytenr;
2100         ins.offset = disk_num_bytes;
2101         ins.type = BTRFS_EXTENT_ITEM_KEY;
2102         ret = btrfs_alloc_reserved_file_extent(trans, root->root_key.objectid,
2103                         btrfs_ino(BTRFS_I(inode)), file_pos, ram_bytes, &ins);
2104         /*
2105          * Release the reserved range from inode dirty range map, as it is
2106          * already moved into delayed_ref_head
2107          */
2108         btrfs_qgroup_release_data(inode, file_pos, ram_bytes);
2109 out:
2110         btrfs_free_path(path);
2111
2112         return ret;
2113 }
2114
2115 /* snapshot-aware defrag */
2116 struct sa_defrag_extent_backref {
2117         struct rb_node node;
2118         struct old_sa_defrag_extent *old;
2119         u64 root_id;
2120         u64 inum;
2121         u64 file_pos;
2122         u64 extent_offset;
2123         u64 num_bytes;
2124         u64 generation;
2125 };
2126
2127 struct old_sa_defrag_extent {
2128         struct list_head list;
2129         struct new_sa_defrag_extent *new;
2130
2131         u64 extent_offset;
2132         u64 bytenr;
2133         u64 offset;
2134         u64 len;
2135         int count;
2136 };
2137
2138 struct new_sa_defrag_extent {
2139         struct rb_root root;
2140         struct list_head head;
2141         struct btrfs_path *path;
2142         struct inode *inode;
2143         u64 file_pos;
2144         u64 len;
2145         u64 bytenr;
2146         u64 disk_len;
2147         u8 compress_type;
2148 };
2149
2150 static int backref_comp(struct sa_defrag_extent_backref *b1,
2151                         struct sa_defrag_extent_backref *b2)
2152 {
2153         if (b1->root_id < b2->root_id)
2154                 return -1;
2155         else if (b1->root_id > b2->root_id)
2156                 return 1;
2157
2158         if (b1->inum < b2->inum)
2159                 return -1;
2160         else if (b1->inum > b2->inum)
2161                 return 1;
2162
2163         if (b1->file_pos < b2->file_pos)
2164                 return -1;
2165         else if (b1->file_pos > b2->file_pos)
2166                 return 1;
2167
2168         /*
2169          * [------------------------------] ===> (a range of space)
2170          *     |<--->|   |<---->| =============> (fs/file tree A)
2171          * |<---------------------------->| ===> (fs/file tree B)
2172          *
2173          * A range of space can refer to two file extents in one tree while
2174          * refer to only one file extent in another tree.
2175          *
2176          * So we may process a disk offset more than one time(two extents in A)
2177          * and locate at the same extent(one extent in B), then insert two same
2178          * backrefs(both refer to the extent in B).
2179          */
2180         return 0;
2181 }
2182
2183 static void backref_insert(struct rb_root *root,
2184                            struct sa_defrag_extent_backref *backref)
2185 {
2186         struct rb_node **p = &root->rb_node;
2187         struct rb_node *parent = NULL;
2188         struct sa_defrag_extent_backref *entry;
2189         int ret;
2190
2191         while (*p) {
2192                 parent = *p;
2193                 entry = rb_entry(parent, struct sa_defrag_extent_backref, node);
2194
2195                 ret = backref_comp(backref, entry);
2196                 if (ret < 0)
2197                         p = &(*p)->rb_left;
2198                 else
2199                         p = &(*p)->rb_right;
2200         }
2201
2202         rb_link_node(&backref->node, parent, p);
2203         rb_insert_color(&backref->node, root);
2204 }
2205
2206 /*
2207  * Note the backref might has changed, and in this case we just return 0.
2208  */
2209 static noinline int record_one_backref(u64 inum, u64 offset, u64 root_id,
2210                                        void *ctx)
2211 {
2212         struct btrfs_file_extent_item *extent;
2213         struct old_sa_defrag_extent *old = ctx;
2214         struct new_sa_defrag_extent *new = old->new;
2215         struct btrfs_path *path = new->path;
2216         struct btrfs_key key;
2217         struct btrfs_root *root;
2218         struct sa_defrag_extent_backref *backref;
2219         struct extent_buffer *leaf;
2220         struct inode *inode = new->inode;
2221         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2222         int slot;
2223         int ret;
2224         u64 extent_offset;
2225         u64 num_bytes;
2226
2227         if (BTRFS_I(inode)->root->root_key.objectid == root_id &&
2228             inum == btrfs_ino(BTRFS_I(inode)))
2229                 return 0;
2230
2231         key.objectid = root_id;
2232         key.type = BTRFS_ROOT_ITEM_KEY;
2233         key.offset = (u64)-1;
2234
2235         root = btrfs_read_fs_root_no_name(fs_info, &key);
2236         if (IS_ERR(root)) {
2237                 if (PTR_ERR(root) == -ENOENT)
2238                         return 0;
2239                 WARN_ON(1);
2240                 btrfs_debug(fs_info, "inum=%llu, offset=%llu, root_id=%llu",
2241                          inum, offset, root_id);
2242                 return PTR_ERR(root);
2243         }
2244
2245         key.objectid = inum;
2246         key.type = BTRFS_EXTENT_DATA_KEY;
2247         if (offset > (u64)-1 << 32)
2248                 key.offset = 0;
2249         else
2250                 key.offset = offset;
2251
2252         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2253         if (WARN_ON(ret < 0))
2254                 return ret;
2255         ret = 0;
2256
2257         while (1) {
2258                 cond_resched();
2259
2260                 leaf = path->nodes[0];
2261                 slot = path->slots[0];
2262
2263                 if (slot >= btrfs_header_nritems(leaf)) {
2264                         ret = btrfs_next_leaf(root, path);
2265                         if (ret < 0) {
2266                                 goto out;
2267                         } else if (ret > 0) {
2268                                 ret = 0;
2269                                 goto out;
2270                         }
2271                         continue;
2272                 }
2273
2274                 path->slots[0]++;
2275
2276                 btrfs_item_key_to_cpu(leaf, &key, slot);
2277
2278                 if (key.objectid > inum)
2279                         goto out;
2280
2281                 if (key.objectid < inum || key.type != BTRFS_EXTENT_DATA_KEY)
2282                         continue;
2283
2284                 extent = btrfs_item_ptr(leaf, slot,
2285                                         struct btrfs_file_extent_item);
2286
2287                 if (btrfs_file_extent_disk_bytenr(leaf, extent) != old->bytenr)
2288                         continue;
2289
2290                 /*
2291                  * 'offset' refers to the exact key.offset,
2292                  * NOT the 'offset' field in btrfs_extent_data_ref, ie.
2293                  * (key.offset - extent_offset).
2294                  */
2295                 if (key.offset != offset)
2296                         continue;
2297
2298                 extent_offset = btrfs_file_extent_offset(leaf, extent);
2299                 num_bytes = btrfs_file_extent_num_bytes(leaf, extent);
2300
2301                 if (extent_offset >= old->extent_offset + old->offset +
2302                     old->len || extent_offset + num_bytes <=
2303                     old->extent_offset + old->offset)
2304                         continue;
2305                 break;
2306         }
2307
2308         backref = kmalloc(sizeof(*backref), GFP_NOFS);
2309         if (!backref) {
2310                 ret = -ENOENT;
2311                 goto out;
2312         }
2313
2314         backref->root_id = root_id;
2315         backref->inum = inum;
2316         backref->file_pos = offset;
2317         backref->num_bytes = num_bytes;
2318         backref->extent_offset = extent_offset;
2319         backref->generation = btrfs_file_extent_generation(leaf, extent);
2320         backref->old = old;
2321         backref_insert(&new->root, backref);
2322         old->count++;
2323 out:
2324         btrfs_release_path(path);
2325         WARN_ON(ret);
2326         return ret;
2327 }
2328
2329 static noinline bool record_extent_backrefs(struct btrfs_path *path,
2330                                    struct new_sa_defrag_extent *new)
2331 {
2332         struct btrfs_fs_info *fs_info = btrfs_sb(new->inode->i_sb);
2333         struct old_sa_defrag_extent *old, *tmp;
2334         int ret;
2335
2336         new->path = path;
2337
2338         list_for_each_entry_safe(old, tmp, &new->head, list) {
2339                 ret = iterate_inodes_from_logical(old->bytenr +
2340                                                   old->extent_offset, fs_info,
2341                                                   path, record_one_backref,
2342                                                   old);
2343                 if (ret < 0 && ret != -ENOENT)
2344                         return false;
2345
2346                 /* no backref to be processed for this extent */
2347                 if (!old->count) {
2348                         list_del(&old->list);
2349                         kfree(old);
2350                 }
2351         }
2352
2353         if (list_empty(&new->head))
2354                 return false;
2355
2356         return true;
2357 }
2358
2359 static int relink_is_mergable(struct extent_buffer *leaf,
2360                               struct btrfs_file_extent_item *fi,
2361                               struct new_sa_defrag_extent *new)
2362 {
2363         if (btrfs_file_extent_disk_bytenr(leaf, fi) != new->bytenr)
2364                 return 0;
2365
2366         if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG)
2367                 return 0;
2368
2369         if (btrfs_file_extent_compression(leaf, fi) != new->compress_type)
2370                 return 0;
2371
2372         if (btrfs_file_extent_encryption(leaf, fi) ||
2373             btrfs_file_extent_other_encoding(leaf, fi))
2374                 return 0;
2375
2376         return 1;
2377 }
2378
2379 /*
2380  * Note the backref might has changed, and in this case we just return 0.
2381  */
2382 static noinline int relink_extent_backref(struct btrfs_path *path,
2383                                  struct sa_defrag_extent_backref *prev,
2384                                  struct sa_defrag_extent_backref *backref)
2385 {
2386         struct btrfs_file_extent_item *extent;
2387         struct btrfs_file_extent_item *item;
2388         struct btrfs_ordered_extent *ordered;
2389         struct btrfs_trans_handle *trans;
2390         struct btrfs_root *root;
2391         struct btrfs_key key;
2392         struct extent_buffer *leaf;
2393         struct old_sa_defrag_extent *old = backref->old;
2394         struct new_sa_defrag_extent *new = old->new;
2395         struct btrfs_fs_info *fs_info = btrfs_sb(new->inode->i_sb);
2396         struct inode *inode;
2397         struct extent_state *cached = NULL;
2398         int ret = 0;
2399         u64 start;
2400         u64 len;
2401         u64 lock_start;
2402         u64 lock_end;
2403         bool merge = false;
2404         int index;
2405
2406         if (prev && prev->root_id == backref->root_id &&
2407             prev->inum == backref->inum &&
2408             prev->file_pos + prev->num_bytes == backref->file_pos)
2409                 merge = true;
2410
2411         /* step 1: get root */
2412         key.objectid = backref->root_id;
2413         key.type = BTRFS_ROOT_ITEM_KEY;
2414         key.offset = (u64)-1;
2415
2416         index = srcu_read_lock(&fs_info->subvol_srcu);
2417
2418         root = btrfs_read_fs_root_no_name(fs_info, &key);
2419         if (IS_ERR(root)) {
2420                 srcu_read_unlock(&fs_info->subvol_srcu, index);
2421                 if (PTR_ERR(root) == -ENOENT)
2422                         return 0;
2423                 return PTR_ERR(root);
2424         }
2425
2426         if (btrfs_root_readonly(root)) {
2427                 srcu_read_unlock(&fs_info->subvol_srcu, index);
2428                 return 0;
2429         }
2430
2431         /* step 2: get inode */
2432         key.objectid = backref->inum;
2433         key.type = BTRFS_INODE_ITEM_KEY;
2434         key.offset = 0;
2435
2436         inode = btrfs_iget(fs_info->sb, &key, root, NULL);
2437         if (IS_ERR(inode)) {
2438                 srcu_read_unlock(&fs_info->subvol_srcu, index);
2439                 return 0;
2440         }
2441
2442         srcu_read_unlock(&fs_info->subvol_srcu, index);
2443
2444         /* step 3: relink backref */
2445         lock_start = backref->file_pos;
2446         lock_end = backref->file_pos + backref->num_bytes - 1;
2447         lock_extent_bits(&BTRFS_I(inode)->io_tree, lock_start, lock_end,
2448                          &cached);
2449
2450         ordered = btrfs_lookup_first_ordered_extent(inode, lock_end);
2451         if (ordered) {
2452                 btrfs_put_ordered_extent(ordered);
2453                 goto out_unlock;
2454         }
2455
2456         trans = btrfs_join_transaction(root);
2457         if (IS_ERR(trans)) {
2458                 ret = PTR_ERR(trans);
2459                 goto out_unlock;
2460         }
2461
2462         key.objectid = backref->inum;
2463         key.type = BTRFS_EXTENT_DATA_KEY;
2464         key.offset = backref->file_pos;
2465
2466         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2467         if (ret < 0) {
2468                 goto out_free_path;
2469         } else if (ret > 0) {
2470                 ret = 0;
2471                 goto out_free_path;
2472         }
2473
2474         extent = btrfs_item_ptr(path->nodes[0], path->slots[0],
2475                                 struct btrfs_file_extent_item);
2476
2477         if (btrfs_file_extent_generation(path->nodes[0], extent) !=
2478             backref->generation)
2479                 goto out_free_path;
2480
2481         btrfs_release_path(path);
2482
2483         start = backref->file_pos;
2484         if (backref->extent_offset < old->extent_offset + old->offset)
2485                 start += old->extent_offset + old->offset -
2486                          backref->extent_offset;
2487
2488         len = min(backref->extent_offset + backref->num_bytes,
2489                   old->extent_offset + old->offset + old->len);
2490         len -= max(backref->extent_offset, old->extent_offset + old->offset);
2491
2492         ret = btrfs_drop_extents(trans, root, inode, start,
2493                                  start + len, 1);
2494         if (ret)
2495                 goto out_free_path;
2496 again:
2497         key.objectid = btrfs_ino(BTRFS_I(inode));
2498         key.type = BTRFS_EXTENT_DATA_KEY;
2499         key.offset = start;
2500
2501         path->leave_spinning = 1;
2502         if (merge) {
2503                 struct btrfs_file_extent_item *fi;
2504                 u64 extent_len;
2505                 struct btrfs_key found_key;
2506
2507                 ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
2508                 if (ret < 0)
2509                         goto out_free_path;
2510
2511                 path->slots[0]--;
2512                 leaf = path->nodes[0];
2513                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
2514
2515                 fi = btrfs_item_ptr(leaf, path->slots[0],
2516                                     struct btrfs_file_extent_item);
2517                 extent_len = btrfs_file_extent_num_bytes(leaf, fi);
2518
2519                 if (extent_len + found_key.offset == start &&
2520                     relink_is_mergable(leaf, fi, new)) {
2521                         btrfs_set_file_extent_num_bytes(leaf, fi,
2522                                                         extent_len + len);
2523                         btrfs_mark_buffer_dirty(leaf);
2524                         inode_add_bytes(inode, len);
2525
2526                         ret = 1;
2527                         goto out_free_path;
2528                 } else {
2529                         merge = false;
2530                         btrfs_release_path(path);
2531                         goto again;
2532                 }
2533         }
2534
2535         ret = btrfs_insert_empty_item(trans, root, path, &key,
2536                                         sizeof(*extent));
2537         if (ret) {
2538                 btrfs_abort_transaction(trans, ret);
2539                 goto out_free_path;
2540         }
2541
2542         leaf = path->nodes[0];
2543         item = btrfs_item_ptr(leaf, path->slots[0],
2544                                 struct btrfs_file_extent_item);
2545         btrfs_set_file_extent_disk_bytenr(leaf, item, new->bytenr);
2546         btrfs_set_file_extent_disk_num_bytes(leaf, item, new->disk_len);
2547         btrfs_set_file_extent_offset(leaf, item, start - new->file_pos);
2548         btrfs_set_file_extent_num_bytes(leaf, item, len);
2549         btrfs_set_file_extent_ram_bytes(leaf, item, new->len);
2550         btrfs_set_file_extent_generation(leaf, item, trans->transid);
2551         btrfs_set_file_extent_type(leaf, item, BTRFS_FILE_EXTENT_REG);
2552         btrfs_set_file_extent_compression(leaf, item, new->compress_type);
2553         btrfs_set_file_extent_encryption(leaf, item, 0);
2554         btrfs_set_file_extent_other_encoding(leaf, item, 0);
2555
2556         btrfs_mark_buffer_dirty(leaf);
2557         inode_add_bytes(inode, len);
2558         btrfs_release_path(path);
2559
2560         ret = btrfs_inc_extent_ref(trans, fs_info, new->bytenr,
2561                         new->disk_len, 0,
2562                         backref->root_id, backref->inum,
2563                         new->file_pos); /* start - extent_offset */
2564         if (ret) {
2565                 btrfs_abort_transaction(trans, ret);
2566                 goto out_free_path;
2567         }
2568
2569         ret = 1;
2570 out_free_path:
2571         btrfs_release_path(path);
2572         path->leave_spinning = 0;
2573         btrfs_end_transaction(trans);
2574 out_unlock:
2575         unlock_extent_cached(&BTRFS_I(inode)->io_tree, lock_start, lock_end,
2576                              &cached, GFP_NOFS);
2577         iput(inode);
2578         return ret;
2579 }
2580
2581 static void free_sa_defrag_extent(struct new_sa_defrag_extent *new)
2582 {
2583         struct old_sa_defrag_extent *old, *tmp;
2584
2585         if (!new)
2586                 return;
2587
2588         list_for_each_entry_safe(old, tmp, &new->head, list) {
2589                 kfree(old);
2590         }
2591         kfree(new);
2592 }
2593
2594 static void relink_file_extents(struct new_sa_defrag_extent *new)
2595 {
2596         struct btrfs_fs_info *fs_info = btrfs_sb(new->inode->i_sb);
2597         struct btrfs_path *path;
2598         struct sa_defrag_extent_backref *backref;
2599         struct sa_defrag_extent_backref *prev = NULL;
2600         struct inode *inode;
2601         struct btrfs_root *root;
2602         struct rb_node *node;
2603         int ret;
2604
2605         inode = new->inode;
2606         root = BTRFS_I(inode)->root;
2607
2608         path = btrfs_alloc_path();
2609         if (!path)
2610                 return;
2611
2612         if (!record_extent_backrefs(path, new)) {
2613                 btrfs_free_path(path);
2614                 goto out;
2615         }
2616         btrfs_release_path(path);
2617
2618         while (1) {
2619                 node = rb_first(&new->root);
2620                 if (!node)
2621                         break;
2622                 rb_erase(node, &new->root);
2623
2624                 backref = rb_entry(node, struct sa_defrag_extent_backref, node);
2625
2626                 ret = relink_extent_backref(path, prev, backref);
2627                 WARN_ON(ret < 0);
2628
2629                 kfree(prev);
2630
2631                 if (ret == 1)
2632                         prev = backref;
2633                 else
2634                         prev = NULL;
2635                 cond_resched();
2636         }
2637         kfree(prev);
2638
2639         btrfs_free_path(path);
2640 out:
2641         free_sa_defrag_extent(new);
2642
2643         atomic_dec(&fs_info->defrag_running);
2644         wake_up(&fs_info->transaction_wait);
2645 }
2646
2647 static struct new_sa_defrag_extent *
2648 record_old_file_extents(struct inode *inode,
2649                         struct btrfs_ordered_extent *ordered)
2650 {
2651         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2652         struct btrfs_root *root = BTRFS_I(inode)->root;
2653         struct btrfs_path *path;
2654         struct btrfs_key key;
2655         struct old_sa_defrag_extent *old;
2656         struct new_sa_defrag_extent *new;
2657         int ret;
2658
2659         new = kmalloc(sizeof(*new), GFP_NOFS);
2660         if (!new)
2661                 return NULL;
2662
2663         new->inode = inode;
2664         new->file_pos = ordered->file_offset;
2665         new->len = ordered->len;
2666         new->bytenr = ordered->start;
2667         new->disk_len = ordered->disk_len;
2668         new->compress_type = ordered->compress_type;
2669         new->root = RB_ROOT;
2670         INIT_LIST_HEAD(&new->head);
2671
2672         path = btrfs_alloc_path();
2673         if (!path)
2674                 goto out_kfree;
2675
2676         key.objectid = btrfs_ino(BTRFS_I(inode));
2677         key.type = BTRFS_EXTENT_DATA_KEY;
2678         key.offset = new->file_pos;
2679
2680         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2681         if (ret < 0)
2682                 goto out_free_path;
2683         if (ret > 0 && path->slots[0] > 0)
2684                 path->slots[0]--;
2685
2686         /* find out all the old extents for the file range */
2687         while (1) {
2688                 struct btrfs_file_extent_item *extent;
2689                 struct extent_buffer *l;
2690                 int slot;
2691                 u64 num_bytes;
2692                 u64 offset;
2693                 u64 end;
2694                 u64 disk_bytenr;
2695                 u64 extent_offset;
2696
2697                 l = path->nodes[0];
2698                 slot = path->slots[0];
2699
2700                 if (slot >= btrfs_header_nritems(l)) {
2701                         ret = btrfs_next_leaf(root, path);
2702                         if (ret < 0)
2703                                 goto out_free_path;
2704                         else if (ret > 0)
2705                                 break;
2706                         continue;
2707                 }
2708
2709                 btrfs_item_key_to_cpu(l, &key, slot);
2710
2711                 if (key.objectid != btrfs_ino(BTRFS_I(inode)))
2712                         break;
2713                 if (key.type != BTRFS_EXTENT_DATA_KEY)
2714                         break;
2715                 if (key.offset >= new->file_pos + new->len)
2716                         break;
2717
2718                 extent = btrfs_item_ptr(l, slot, struct btrfs_file_extent_item);
2719
2720                 num_bytes = btrfs_file_extent_num_bytes(l, extent);
2721                 if (key.offset + num_bytes < new->file_pos)
2722                         goto next;
2723
2724                 disk_bytenr = btrfs_file_extent_disk_bytenr(l, extent);
2725                 if (!disk_bytenr)
2726                         goto next;
2727
2728                 extent_offset = btrfs_file_extent_offset(l, extent);
2729
2730                 old = kmalloc(sizeof(*old), GFP_NOFS);
2731                 if (!old)
2732                         goto out_free_path;
2733
2734                 offset = max(new->file_pos, key.offset);
2735                 end = min(new->file_pos + new->len, key.offset + num_bytes);
2736
2737                 old->bytenr = disk_bytenr;
2738                 old->extent_offset = extent_offset;
2739                 old->offset = offset - key.offset;
2740                 old->len = end - offset;
2741                 old->new = new;
2742                 old->count = 0;
2743                 list_add_tail(&old->list, &new->head);
2744 next:
2745                 path->slots[0]++;
2746                 cond_resched();
2747         }
2748
2749         btrfs_free_path(path);
2750         atomic_inc(&fs_info->defrag_running);
2751
2752         return new;
2753
2754 out_free_path:
2755         btrfs_free_path(path);
2756 out_kfree:
2757         free_sa_defrag_extent(new);
2758         return NULL;
2759 }
2760
2761 static void btrfs_release_delalloc_bytes(struct btrfs_fs_info *fs_info,
2762                                          u64 start, u64 len)
2763 {
2764         struct btrfs_block_group_cache *cache;
2765
2766         cache = btrfs_lookup_block_group(fs_info, start);
2767         ASSERT(cache);
2768
2769         spin_lock(&cache->lock);
2770         cache->delalloc_bytes -= len;
2771         spin_unlock(&cache->lock);
2772
2773         btrfs_put_block_group(cache);
2774 }
2775
2776 /* as ordered data IO finishes, this gets called so we can finish
2777  * an ordered extent if the range of bytes in the file it covers are
2778  * fully written.
2779  */
2780 static int btrfs_finish_ordered_io(struct btrfs_ordered_extent *ordered_extent)
2781 {
2782         struct inode *inode = ordered_extent->inode;
2783         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2784         struct btrfs_root *root = BTRFS_I(inode)->root;
2785         struct btrfs_trans_handle *trans = NULL;
2786         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
2787         struct extent_state *cached_state = NULL;
2788         struct new_sa_defrag_extent *new = NULL;
2789         int compress_type = 0;
2790         int ret = 0;
2791         u64 logical_len = ordered_extent->len;
2792         bool nolock;
2793         bool truncated = false;
2794
2795         nolock = btrfs_is_free_space_inode(BTRFS_I(inode));
2796
2797         if (test_bit(BTRFS_ORDERED_IOERR, &ordered_extent->flags)) {
2798                 ret = -EIO;
2799                 goto out;
2800         }
2801
2802         btrfs_free_io_failure_record(BTRFS_I(inode),
2803                         ordered_extent->file_offset,
2804                         ordered_extent->file_offset +
2805                         ordered_extent->len - 1);
2806
2807         if (test_bit(BTRFS_ORDERED_TRUNCATED, &ordered_extent->flags)) {
2808                 truncated = true;
2809                 logical_len = ordered_extent->truncated_len;
2810                 /* Truncated the entire extent, don't bother adding */
2811                 if (!logical_len)
2812                         goto out;
2813         }
2814
2815         if (test_bit(BTRFS_ORDERED_NOCOW, &ordered_extent->flags)) {
2816                 BUG_ON(!list_empty(&ordered_extent->list)); /* Logic error */
2817
2818                 /*
2819                  * For mwrite(mmap + memset to write) case, we still reserve
2820                  * space for NOCOW range.
2821                  * As NOCOW won't cause a new delayed ref, just free the space
2822                  */
2823                 btrfs_qgroup_free_data(inode, ordered_extent->file_offset,
2824                                        ordered_extent->len);
2825                 btrfs_ordered_update_i_size(inode, 0, ordered_extent);
2826                 if (nolock)
2827                         trans = btrfs_join_transaction_nolock(root);
2828                 else
2829                         trans = btrfs_join_transaction(root);
2830                 if (IS_ERR(trans)) {
2831                         ret = PTR_ERR(trans);
2832                         trans = NULL;
2833                         goto out;
2834                 }
2835                 trans->block_rsv = &fs_info->delalloc_block_rsv;
2836                 ret = btrfs_update_inode_fallback(trans, root, inode);
2837                 if (ret) /* -ENOMEM or corruption */
2838                         btrfs_abort_transaction(trans, ret);
2839                 goto out;
2840         }
2841
2842         lock_extent_bits(io_tree, ordered_extent->file_offset,
2843                          ordered_extent->file_offset + ordered_extent->len - 1,
2844                          &cached_state);
2845
2846         ret = test_range_bit(io_tree, ordered_extent->file_offset,
2847                         ordered_extent->file_offset + ordered_extent->len - 1,
2848                         EXTENT_DEFRAG, 1, cached_state);
2849         if (ret) {
2850                 u64 last_snapshot = btrfs_root_last_snapshot(&root->root_item);
2851                 if (0 && last_snapshot >= BTRFS_I(inode)->generation)
2852                         /* the inode is shared */
2853                         new = record_old_file_extents(inode, ordered_extent);
2854
2855                 clear_extent_bit(io_tree, ordered_extent->file_offset,
2856                         ordered_extent->file_offset + ordered_extent->len - 1,
2857                         EXTENT_DEFRAG, 0, 0, &cached_state, GFP_NOFS);
2858         }
2859
2860         if (nolock)
2861                 trans = btrfs_join_transaction_nolock(root);
2862         else
2863                 trans = btrfs_join_transaction(root);
2864         if (IS_ERR(trans)) {
2865                 ret = PTR_ERR(trans);
2866                 trans = NULL;
2867                 goto out_unlock;
2868         }
2869
2870         trans->block_rsv = &fs_info->delalloc_block_rsv;
2871
2872         if (test_bit(BTRFS_ORDERED_COMPRESSED, &ordered_extent->flags))
2873                 compress_type = ordered_extent->compress_type;
2874         if (test_bit(BTRFS_ORDERED_PREALLOC, &ordered_extent->flags)) {
2875                 BUG_ON(compress_type);
2876                 ret = btrfs_mark_extent_written(trans, BTRFS_I(inode),
2877                                                 ordered_extent->file_offset,
2878                                                 ordered_extent->file_offset +
2879                                                 logical_len);
2880         } else {
2881                 BUG_ON(root == fs_info->tree_root);
2882                 ret = insert_reserved_file_extent(trans, inode,
2883                                                 ordered_extent->file_offset,
2884                                                 ordered_extent->start,
2885                                                 ordered_extent->disk_len,
2886                                                 logical_len, logical_len,
2887                                                 compress_type, 0, 0,
2888                                                 BTRFS_FILE_EXTENT_REG);
2889                 if (!ret)
2890                         btrfs_release_delalloc_bytes(fs_info,
2891                                                      ordered_extent->start,
2892                                                      ordered_extent->disk_len);
2893         }
2894         unpin_extent_cache(&BTRFS_I(inode)->extent_tree,
2895                            ordered_extent->file_offset, ordered_extent->len,
2896                            trans->transid);
2897         if (ret < 0) {
2898                 btrfs_abort_transaction(trans, ret);
2899                 goto out_unlock;
2900         }
2901
2902         add_pending_csums(trans, inode, &ordered_extent->list);
2903
2904         btrfs_ordered_update_i_size(inode, 0, ordered_extent);
2905         ret = btrfs_update_inode_fallback(trans, root, inode);
2906         if (ret) { /* -ENOMEM or corruption */
2907                 btrfs_abort_transaction(trans, ret);
2908                 goto out_unlock;
2909         }
2910         ret = 0;
2911 out_unlock:
2912         unlock_extent_cached(io_tree, ordered_extent->file_offset,
2913                              ordered_extent->file_offset +
2914                              ordered_extent->len - 1, &cached_state, GFP_NOFS);
2915 out:
2916         if (root != fs_info->tree_root)
2917                 btrfs_delalloc_release_metadata(BTRFS_I(inode),
2918                                 ordered_extent->len);
2919         if (trans)
2920                 btrfs_end_transaction(trans);
2921
2922         if (ret || truncated) {
2923                 u64 start, end;
2924
2925                 if (truncated)
2926                         start = ordered_extent->file_offset + logical_len;
2927                 else
2928                         start = ordered_extent->file_offset;
2929                 end = ordered_extent->file_offset + ordered_extent->len - 1;
2930                 clear_extent_uptodate(io_tree, start, end, NULL, GFP_NOFS);
2931
2932                 /* Drop the cache for the part of the extent we didn't write. */
2933                 btrfs_drop_extent_cache(BTRFS_I(inode), start, end, 0);
2934
2935                 /*
2936                  * If the ordered extent had an IOERR or something else went
2937                  * wrong we need to return the space for this ordered extent
2938                  * back to the allocator.  We only free the extent in the
2939                  * truncated case if we didn't write out the extent at all.
2940                  */
2941                 if ((ret || !logical_len) &&
2942                     !test_bit(BTRFS_ORDERED_NOCOW, &ordered_extent->flags) &&
2943                     !test_bit(BTRFS_ORDERED_PREALLOC, &ordered_extent->flags))
2944                         btrfs_free_reserved_extent(fs_info,
2945                                                    ordered_extent->start,
2946                                                    ordered_extent->disk_len, 1);
2947         }
2948
2949
2950         /*
2951          * This needs to be done to make sure anybody waiting knows we are done
2952          * updating everything for this ordered extent.
2953          */
2954         btrfs_remove_ordered_extent(inode, ordered_extent);
2955
2956         /* for snapshot-aware defrag */
2957         if (new) {
2958                 if (ret) {
2959                         free_sa_defrag_extent(new);
2960                         atomic_dec(&fs_info->defrag_running);
2961                 } else {
2962                         relink_file_extents(new);
2963                 }
2964         }
2965
2966         /* once for us */
2967         btrfs_put_ordered_extent(ordered_extent);
2968         /* once for the tree */
2969         btrfs_put_ordered_extent(ordered_extent);
2970
2971         return ret;
2972 }
2973
2974 static void finish_ordered_fn(struct btrfs_work *work)
2975 {
2976         struct btrfs_ordered_extent *ordered_extent;
2977         ordered_extent = container_of(work, struct btrfs_ordered_extent, work);
2978         btrfs_finish_ordered_io(ordered_extent);
2979 }
2980
2981 static void btrfs_writepage_end_io_hook(struct page *page, u64 start, u64 end,
2982                                 struct extent_state *state, int uptodate)
2983 {
2984         struct inode *inode = page->mapping->host;
2985         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2986         struct btrfs_ordered_extent *ordered_extent = NULL;
2987         struct btrfs_workqueue *wq;
2988         btrfs_work_func_t func;
2989
2990         trace_btrfs_writepage_end_io_hook(page, start, end, uptodate);
2991
2992         ClearPagePrivate2(page);
2993         if (!btrfs_dec_test_ordered_pending(inode, &ordered_extent, start,
2994                                             end - start + 1, uptodate))
2995                 return;
2996
2997         if (btrfs_is_free_space_inode(BTRFS_I(inode))) {
2998                 wq = fs_info->endio_freespace_worker;
2999                 func = btrfs_freespace_write_helper;
3000         } else {
3001                 wq = fs_info->endio_write_workers;
3002                 func = btrfs_endio_write_helper;
3003         }
3004
3005         btrfs_init_work(&ordered_extent->work, func, finish_ordered_fn, NULL,
3006                         NULL);
3007         btrfs_queue_work(wq, &ordered_extent->work);
3008 }
3009
3010 static int __readpage_endio_check(struct inode *inode,
3011                                   struct btrfs_io_bio *io_bio,
3012                                   int icsum, struct page *page,
3013                                   int pgoff, u64 start, size_t len)
3014 {
3015         char *kaddr;
3016         u32 csum_expected;
3017         u32 csum = ~(u32)0;
3018
3019         csum_expected = *(((u32 *)io_bio->csum) + icsum);
3020
3021         kaddr = kmap_atomic(page);
3022         csum = btrfs_csum_data(kaddr + pgoff, csum,  len);
3023         btrfs_csum_final(csum, (u8 *)&csum);
3024         if (csum != csum_expected)
3025                 goto zeroit;
3026
3027         kunmap_atomic(kaddr);
3028         return 0;
3029 zeroit:
3030         btrfs_print_data_csum_error(BTRFS_I(inode), start, csum, csum_expected,
3031                                     io_bio->mirror_num);
3032         memset(kaddr + pgoff, 1, len);
3033         flush_dcache_page(page);
3034         kunmap_atomic(kaddr);
3035         if (csum_expected == 0)
3036                 return 0;
3037         return -EIO;
3038 }
3039
3040 /*
3041  * when reads are done, we need to check csums to verify the data is correct
3042  * if there's a match, we allow the bio to finish.  If not, the code in
3043  * extent_io.c will try to find good copies for us.
3044  */
3045 static int btrfs_readpage_end_io_hook(struct btrfs_io_bio *io_bio,
3046                                       u64 phy_offset, struct page *page,
3047                                       u64 start, u64 end, int mirror)
3048 {
3049         size_t offset = start - page_offset(page);
3050         struct inode *inode = page->mapping->host;
3051         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
3052         struct btrfs_root *root = BTRFS_I(inode)->root;
3053
3054         if (PageChecked(page)) {
3055                 ClearPageChecked(page);
3056                 return 0;
3057         }
3058
3059         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)
3060                 return 0;
3061
3062         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID &&
3063             test_range_bit(io_tree, start, end, EXTENT_NODATASUM, 1, NULL)) {
3064                 clear_extent_bits(io_tree, start, end, EXTENT_NODATASUM);
3065                 return 0;
3066         }
3067
3068         phy_offset >>= inode->i_sb->s_blocksize_bits;
3069         return __readpage_endio_check(inode, io_bio, phy_offset, page, offset,
3070                                       start, (size_t)(end - start + 1));
3071 }
3072
3073 void btrfs_add_delayed_iput(struct inode *inode)
3074 {
3075         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
3076         struct btrfs_inode *binode = BTRFS_I(inode);
3077
3078         if (atomic_add_unless(&inode->i_count, -1, 1))
3079                 return;
3080
3081         spin_lock(&fs_info->delayed_iput_lock);
3082         if (binode->delayed_iput_count == 0) {
3083                 ASSERT(list_empty(&binode->delayed_iput));
3084                 list_add_tail(&binode->delayed_iput, &fs_info->delayed_iputs);
3085         } else {
3086                 binode->delayed_iput_count++;
3087         }
3088         spin_unlock(&fs_info->delayed_iput_lock);
3089 }
3090
3091 void btrfs_run_delayed_iputs(struct btrfs_fs_info *fs_info)
3092 {
3093
3094         spin_lock(&fs_info->delayed_iput_lock);
3095         while (!list_empty(&fs_info->delayed_iputs)) {
3096                 struct btrfs_inode *inode;
3097
3098                 inode = list_first_entry(&fs_info->delayed_iputs,
3099                                 struct btrfs_inode, delayed_iput);
3100                 if (inode->delayed_iput_count) {
3101                         inode->delayed_iput_count--;
3102                         list_move_tail(&inode->delayed_iput,
3103                                         &fs_info->delayed_iputs);
3104                 } else {
3105                         list_del_init(&inode->delayed_iput);
3106                 }
3107                 spin_unlock(&fs_info->delayed_iput_lock);
3108                 iput(&inode->vfs_inode);
3109                 spin_lock(&fs_info->delayed_iput_lock);
3110         }
3111         spin_unlock(&fs_info->delayed_iput_lock);
3112 }
3113
3114 /*
3115  * This is called in transaction commit time. If there are no orphan
3116  * files in the subvolume, it removes orphan item and frees block_rsv
3117  * structure.
3118  */
3119 void btrfs_orphan_commit_root(struct btrfs_trans_handle *trans,
3120                               struct btrfs_root *root)
3121 {
3122         struct btrfs_fs_info *fs_info = root->fs_info;
3123         struct btrfs_block_rsv *block_rsv;
3124         int ret;
3125
3126         if (atomic_read(&root->orphan_inodes) ||
3127             root->orphan_cleanup_state != ORPHAN_CLEANUP_DONE)
3128                 return;
3129
3130         spin_lock(&root->orphan_lock);
3131         if (atomic_read(&root->orphan_inodes)) {
3132                 spin_unlock(&root->orphan_lock);
3133                 return;
3134         }
3135
3136         if (root->orphan_cleanup_state != ORPHAN_CLEANUP_DONE) {
3137                 spin_unlock(&root->orphan_lock);
3138                 return;
3139         }
3140
3141         block_rsv = root->orphan_block_rsv;
3142         root->orphan_block_rsv = NULL;
3143         spin_unlock(&root->orphan_lock);
3144
3145         if (test_bit(BTRFS_ROOT_ORPHAN_ITEM_INSERTED, &root->state) &&
3146             btrfs_root_refs(&root->root_item) > 0) {
3147                 ret = btrfs_del_orphan_item(trans, fs_info->tree_root,
3148                                             root->root_key.objectid);
3149                 if (ret)
3150                         btrfs_abort_transaction(trans, ret);
3151                 else
3152                         clear_bit(BTRFS_ROOT_ORPHAN_ITEM_INSERTED,
3153                                   &root->state);
3154         }
3155
3156         if (block_rsv) {
3157                 WARN_ON(block_rsv->size > 0);
3158                 btrfs_free_block_rsv(fs_info, block_rsv);
3159         }
3160 }
3161
3162 /*
3163  * This creates an orphan entry for the given inode in case something goes
3164  * wrong in the middle of an unlink/truncate.
3165  *
3166  * NOTE: caller of this function should reserve 5 units of metadata for
3167  *       this function.
3168  */
3169 int btrfs_orphan_add(struct btrfs_trans_handle *trans,
3170                 struct btrfs_inode *inode)
3171 {
3172         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
3173         struct btrfs_root *root = inode->root;
3174         struct btrfs_block_rsv *block_rsv = NULL;
3175         int reserve = 0;
3176         int insert = 0;
3177         int ret;
3178
3179         if (!root->orphan_block_rsv) {
3180                 block_rsv = btrfs_alloc_block_rsv(fs_info,
3181                                                   BTRFS_BLOCK_RSV_TEMP);
3182                 if (!block_rsv)
3183                         return -ENOMEM;
3184         }
3185
3186         spin_lock(&root->orphan_lock);
3187         if (!root->orphan_block_rsv) {
3188                 root->orphan_block_rsv = block_rsv;
3189         } else if (block_rsv) {
3190                 btrfs_free_block_rsv(fs_info, block_rsv);
3191                 block_rsv = NULL;
3192         }
3193
3194         if (!test_and_set_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
3195                               &inode->runtime_flags)) {
3196 #if 0
3197                 /*
3198                  * For proper ENOSPC handling, we should do orphan
3199                  * cleanup when mounting. But this introduces backward
3200                  * compatibility issue.
3201                  */
3202                 if (!xchg(&root->orphan_item_inserted, 1))
3203                         insert = 2;
3204                 else
3205                         insert = 1;
3206 #endif
3207                 insert = 1;
3208                 atomic_inc(&root->orphan_inodes);
3209         }
3210
3211         if (!test_and_set_bit(BTRFS_INODE_ORPHAN_META_RESERVED,
3212                               &inode->runtime_flags))
3213                 reserve = 1;
3214         spin_unlock(&root->orphan_lock);
3215
3216         /* grab metadata reservation from transaction handle */
3217         if (reserve) {
3218                 ret = btrfs_orphan_reserve_metadata(trans, inode);
3219                 ASSERT(!ret);
3220                 if (ret) {
3221                         atomic_dec(&root->orphan_inodes);
3222                         clear_bit(BTRFS_INODE_ORPHAN_META_RESERVED,
3223                                   &inode->runtime_flags);
3224                         if (insert)
3225                                 clear_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
3226                                           &inode->runtime_flags);
3227                         return ret;
3228                 }
3229         }
3230
3231         /* insert an orphan item to track this unlinked/truncated file */
3232         if (insert >= 1) {
3233                 ret = btrfs_insert_orphan_item(trans, root, btrfs_ino(inode));
3234                 if (ret) {
3235                         atomic_dec(&root->orphan_inodes);
3236                         if (reserve) {
3237                                 clear_bit(BTRFS_INODE_ORPHAN_META_RESERVED,
3238                                           &inode->runtime_flags);
3239                                 btrfs_orphan_release_metadata(inode);
3240                         }
3241                         if (ret != -EEXIST) {
3242                                 clear_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
3243                                           &inode->runtime_flags);
3244                                 btrfs_abort_transaction(trans, ret);
3245                                 return ret;
3246                         }
3247                 }
3248                 ret = 0;
3249         }
3250
3251         /* insert an orphan item to track subvolume contains orphan files */
3252         if (insert >= 2) {
3253                 ret = btrfs_insert_orphan_item(trans, fs_info->tree_root,
3254                                                root->root_key.objectid);
3255                 if (ret && ret != -EEXIST) {
3256                         btrfs_abort_transaction(trans, ret);
3257                         return ret;
3258                 }
3259         }
3260         return 0;
3261 }
3262
3263 /*
3264  * We have done the truncate/delete so we can go ahead and remove the orphan
3265  * item for this particular inode.
3266  */
3267 static int btrfs_orphan_del(struct btrfs_trans_handle *trans,
3268                             struct btrfs_inode *inode)
3269 {
3270         struct btrfs_root *root = inode->root;
3271         int delete_item = 0;
3272         int release_rsv = 0;
3273         int ret = 0;
3274
3275         spin_lock(&root->orphan_lock);
3276         if (test_and_clear_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
3277                                &inode->runtime_flags))
3278                 delete_item = 1;
3279
3280         if (test_and_clear_bit(BTRFS_INODE_ORPHAN_META_RESERVED,
3281                                &inode->runtime_flags))
3282                 release_rsv = 1;
3283         spin_unlock(&root->orphan_lock);
3284
3285         if (delete_item) {
3286                 atomic_dec(&root->orphan_inodes);
3287                 if (trans)
3288                         ret = btrfs_del_orphan_item(trans, root,
3289                                                     btrfs_ino(inode));
3290         }
3291
3292         if (release_rsv)
3293                 btrfs_orphan_release_metadata(inode);
3294
3295         return ret;
3296 }
3297
3298 /*
3299  * this cleans up any orphans that may be left on the list from the last use
3300  * of this root.
3301  */
3302 int btrfs_orphan_cleanup(struct btrfs_root *root)
3303 {
3304         struct btrfs_fs_info *fs_info = root->fs_info;
3305         struct btrfs_path *path;
3306         struct extent_buffer *leaf;
3307         struct btrfs_key key, found_key;
3308         struct btrfs_trans_handle *trans;
3309         struct inode *inode;
3310         u64 last_objectid = 0;
3311         int ret = 0, nr_unlink = 0, nr_truncate = 0;
3312
3313         if (cmpxchg(&root->orphan_cleanup_state, 0, ORPHAN_CLEANUP_STARTED))
3314                 return 0;
3315
3316         path = btrfs_alloc_path();
3317         if (!path) {
3318                 ret = -ENOMEM;
3319                 goto out;
3320         }
3321         path->reada = READA_BACK;
3322
3323         key.objectid = BTRFS_ORPHAN_OBJECTID;
3324         key.type = BTRFS_ORPHAN_ITEM_KEY;
3325         key.offset = (u64)-1;
3326
3327         while (1) {
3328                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3329                 if (ret < 0)
3330                         goto out;
3331
3332                 /*
3333                  * if ret == 0 means we found what we were searching for, which
3334                  * is weird, but possible, so only screw with path if we didn't
3335                  * find the key and see if we have stuff that matches
3336                  */
3337                 if (ret > 0) {
3338                         ret = 0;
3339                         if (path->slots[0] == 0)
3340                                 break;
3341                         path->slots[0]--;
3342                 }
3343
3344                 /* pull out the item */
3345                 leaf = path->nodes[0];
3346                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
3347
3348                 /* make sure the item matches what we want */
3349                 if (found_key.objectid != BTRFS_ORPHAN_OBJECTID)
3350                         break;
3351                 if (found_key.type != BTRFS_ORPHAN_ITEM_KEY)
3352                         break;
3353
3354                 /* release the path since we're done with it */
3355                 btrfs_release_path(path);
3356
3357                 /*
3358                  * this is where we are basically btrfs_lookup, without the
3359                  * crossing root thing.  we store the inode number in the
3360                  * offset of the orphan item.
3361                  */
3362
3363                 if (found_key.offset == last_objectid) {
3364                         btrfs_err(fs_info,
3365                                   "Error removing orphan entry, stopping orphan cleanup");
3366                         ret = -EINVAL;
3367                         goto out;
3368                 }
3369
3370                 last_objectid = found_key.offset;
3371
3372                 found_key.objectid = found_key.offset;
3373                 found_key.type = BTRFS_INODE_ITEM_KEY;
3374                 found_key.offset = 0;
3375                 inode = btrfs_iget(fs_info->sb, &found_key, root, NULL);
3376                 ret = PTR_ERR_OR_ZERO(inode);
3377                 if (ret && ret != -ENOENT)
3378                         goto out;
3379
3380                 if (ret == -ENOENT && root == fs_info->tree_root) {
3381                         struct btrfs_root *dead_root;
3382                         struct btrfs_fs_info *fs_info = root->fs_info;
3383                         int is_dead_root = 0;
3384
3385                         /*
3386                          * this is an orphan in the tree root. Currently these
3387                          * could come from 2 sources:
3388                          *  a) a snapshot deletion in progress
3389                          *  b) a free space cache inode
3390                          * We need to distinguish those two, as the snapshot
3391                          * orphan must not get deleted.
3392                          * find_dead_roots already ran before us, so if this
3393                          * is a snapshot deletion, we should find the root
3394                          * in the dead_roots list
3395                          */
3396                         spin_lock(&fs_info->trans_lock);
3397                         list_for_each_entry(dead_root, &fs_info->dead_roots,
3398                                             root_list) {
3399                                 if (dead_root->root_key.objectid ==
3400                                     found_key.objectid) {
3401                                         is_dead_root = 1;
3402                                         break;
3403                                 }
3404                         }
3405                         spin_unlock(&fs_info->trans_lock);
3406                         if (is_dead_root) {
3407                                 /* prevent this orphan from being found again */
3408                                 key.offset = found_key.objectid - 1;
3409                                 continue;
3410                         }
3411                 }
3412                 /*
3413                  * Inode is already gone but the orphan item is still there,
3414                  * kill the orphan item.
3415                  */
3416                 if (ret == -ENOENT) {
3417                         trans = btrfs_start_transaction(root, 1);
3418                         if (IS_ERR(trans)) {
3419                                 ret = PTR_ERR(trans);
3420                                 goto out;
3421                         }
3422                         btrfs_debug(fs_info, "auto deleting %Lu",
3423                                     found_key.objectid);
3424                         ret = btrfs_del_orphan_item(trans, root,
3425                                                     found_key.objectid);
3426                         btrfs_end_transaction(trans);
3427                         if (ret)
3428                                 goto out;
3429                         continue;
3430                 }
3431
3432                 /*
3433                  * add this inode to the orphan list so btrfs_orphan_del does
3434                  * the proper thing when we hit it
3435                  */
3436                 set_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
3437                         &BTRFS_I(inode)->runtime_flags);
3438                 atomic_inc(&root->orphan_inodes);
3439
3440                 /* if we have links, this was a truncate, lets do that */
3441                 if (inode->i_nlink) {
3442                         if (WARN_ON(!S_ISREG(inode->i_mode))) {
3443                                 iput(inode);
3444                                 continue;
3445                         }
3446                         nr_truncate++;
3447
3448                         /* 1 for the orphan item deletion. */
3449                         trans = btrfs_start_transaction(root, 1);
3450                         if (IS_ERR(trans)) {
3451                                 iput(inode);
3452                                 ret = PTR_ERR(trans);
3453                                 goto out;
3454                         }
3455                         ret = btrfs_orphan_add(trans, BTRFS_I(inode));
3456                         btrfs_end_transaction(trans);
3457                         if (ret) {
3458                                 iput(inode);
3459                                 goto out;
3460                         }
3461
3462                         ret = btrfs_truncate(inode);
3463                         if (ret)
3464                                 btrfs_orphan_del(NULL, BTRFS_I(inode));
3465                 } else {
3466                         nr_unlink++;
3467                 }
3468
3469                 /* this will do delete_inode and everything for us */
3470                 iput(inode);
3471                 if (ret)
3472                         goto out;
3473         }
3474         /* release the path since we're done with it */
3475         btrfs_release_path(path);
3476
3477         root->orphan_cleanup_state = ORPHAN_CLEANUP_DONE;
3478
3479         if (root->orphan_block_rsv)
3480                 btrfs_block_rsv_release(fs_info, root->orphan_block_rsv,
3481                                         (u64)-1);
3482
3483         if (root->orphan_block_rsv ||
3484             test_bit(BTRFS_ROOT_ORPHAN_ITEM_INSERTED, &root->state)) {
3485                 trans = btrfs_join_transaction(root);
3486                 if (!IS_ERR(trans))
3487                         btrfs_end_transaction(trans);
3488         }
3489
3490         if (nr_unlink)
3491                 btrfs_debug(fs_info, "unlinked %d orphans", nr_unlink);
3492         if (nr_truncate)
3493                 btrfs_debug(fs_info, "truncated %d orphans", nr_truncate);
3494
3495 out:
3496         if (ret)
3497                 btrfs_err(fs_info, "could not do orphan cleanup %d", ret);
3498         btrfs_free_path(path);
3499         return ret;
3500 }
3501
3502 /*
3503  * very simple check to peek ahead in the leaf looking for xattrs.  If we
3504  * don't find any xattrs, we know there can't be any acls.
3505  *
3506  * slot is the slot the inode is in, objectid is the objectid of the inode
3507  */
3508 static noinline int acls_after_inode_item(struct extent_buffer *leaf,
3509                                           int slot, u64 objectid,
3510                                           int *first_xattr_slot)
3511 {
3512         u32 nritems = btrfs_header_nritems(leaf);
3513         struct btrfs_key found_key;
3514         static u64 xattr_access = 0;
3515         static u64 xattr_default = 0;
3516         int scanned = 0;
3517
3518         if (!xattr_access) {
3519                 xattr_access = btrfs_name_hash(XATTR_NAME_POSIX_ACL_ACCESS,
3520                                         strlen(XATTR_NAME_POSIX_ACL_ACCESS));
3521                 xattr_default = btrfs_name_hash(XATTR_NAME_POSIX_ACL_DEFAULT,
3522                                         strlen(XATTR_NAME_POSIX_ACL_DEFAULT));
3523         }
3524
3525         slot++;
3526         *first_xattr_slot = -1;
3527         while (slot < nritems) {
3528                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
3529
3530                 /* we found a different objectid, there must not be acls */
3531                 if (found_key.objectid != objectid)
3532                         return 0;
3533
3534                 /* we found an xattr, assume we've got an acl */
3535                 if (found_key.type == BTRFS_XATTR_ITEM_KEY) {
3536                         if (*first_xattr_slot == -1)
3537                                 *first_xattr_slot = slot;
3538                         if (found_key.offset == xattr_access ||
3539                             found_key.offset == xattr_default)
3540                                 return 1;
3541                 }
3542
3543                 /*
3544                  * we found a key greater than an xattr key, there can't
3545                  * be any acls later on
3546                  */
3547                 if (found_key.type > BTRFS_XATTR_ITEM_KEY)
3548                         return 0;
3549
3550                 slot++;
3551                 scanned++;
3552
3553                 /*
3554                  * it goes inode, inode backrefs, xattrs, extents,
3555                  * so if there are a ton of hard links to an inode there can
3556                  * be a lot of backrefs.  Don't waste time searching too hard,
3557                  * this is just an optimization
3558                  */
3559                 if (scanned >= 8)
3560                         break;
3561         }
3562         /* we hit the end of the leaf before we found an xattr or
3563          * something larger than an xattr.  We have to assume the inode
3564          * has acls
3565          */
3566         if (*first_xattr_slot == -1)
3567                 *first_xattr_slot = slot;
3568         return 1;
3569 }
3570
3571 /*
3572  * read an inode from the btree into the in-memory inode
3573  */
3574 static int btrfs_read_locked_inode(struct inode *inode)
3575 {
3576         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
3577         struct btrfs_path *path;
3578         struct extent_buffer *leaf;
3579         struct btrfs_inode_item *inode_item;
3580         struct btrfs_root *root = BTRFS_I(inode)->root;
3581         struct btrfs_key location;
3582         unsigned long ptr;
3583         int maybe_acls;
3584         u32 rdev;
3585         int ret;
3586         bool filled = false;
3587         int first_xattr_slot;
3588
3589         ret = btrfs_fill_inode(inode, &rdev);
3590         if (!ret)
3591                 filled = true;
3592
3593         path = btrfs_alloc_path();
3594         if (!path) {
3595                 ret = -ENOMEM;
3596                 goto make_bad;
3597         }
3598
3599         memcpy(&location, &BTRFS_I(inode)->location, sizeof(location));
3600
3601         ret = btrfs_lookup_inode(NULL, root, path, &location, 0);
3602         if (ret) {
3603                 if (ret > 0)
3604                         ret = -ENOENT;
3605                 goto make_bad;
3606         }
3607
3608         leaf = path->nodes[0];
3609
3610         if (filled)
3611                 goto cache_index;
3612
3613         inode_item = btrfs_item_ptr(leaf, path->slots[0],
3614                                     struct btrfs_inode_item);
3615         inode->i_mode = btrfs_inode_mode(leaf, inode_item);
3616         set_nlink(inode, btrfs_inode_nlink(leaf, inode_item));
3617         i_uid_write(inode, btrfs_inode_uid(leaf, inode_item));
3618         i_gid_write(inode, btrfs_inode_gid(leaf, inode_item));
3619         btrfs_i_size_write(BTRFS_I(inode), btrfs_inode_size(leaf, inode_item));
3620
3621         inode->i_atime.tv_sec = btrfs_timespec_sec(leaf, &inode_item->atime);
3622         inode->i_atime.tv_nsec = btrfs_timespec_nsec(leaf, &inode_item->atime);
3623
3624         inode->i_mtime.tv_sec = btrfs_timespec_sec(leaf, &inode_item->mtime);
3625         inode->i_mtime.tv_nsec = btrfs_timespec_nsec(leaf, &inode_item->mtime);
3626
3627         inode->i_ctime.tv_sec = btrfs_timespec_sec(leaf, &inode_item->ctime);
3628         inode->i_ctime.tv_nsec = btrfs_timespec_nsec(leaf, &inode_item->ctime);
3629
3630         BTRFS_I(inode)->i_otime.tv_sec =
3631                 btrfs_timespec_sec(leaf, &inode_item->otime);
3632         BTRFS_I(inode)->i_otime.tv_nsec =
3633                 btrfs_timespec_nsec(leaf, &inode_item->otime);
3634
3635         inode_set_bytes(inode, btrfs_inode_nbytes(leaf, inode_item));
3636         BTRFS_I(inode)->generation = btrfs_inode_generation(leaf, inode_item);
3637         BTRFS_I(inode)->last_trans = btrfs_inode_transid(leaf, inode_item);
3638
3639         inode->i_version = btrfs_inode_sequence(leaf, inode_item);
3640         inode->i_generation = BTRFS_I(inode)->generation;
3641         inode->i_rdev = 0;
3642         rdev = btrfs_inode_rdev(leaf, inode_item);
3643
3644         BTRFS_I(inode)->index_cnt = (u64)-1;
3645         BTRFS_I(inode)->flags = btrfs_inode_flags(leaf, inode_item);
3646
3647 cache_index:
3648         /*
3649          * If we were modified in the current generation and evicted from memory
3650          * and then re-read we need to do a full sync since we don't have any
3651          * idea about which extents were modified before we were evicted from
3652          * cache.
3653          *
3654          * This is required for both inode re-read from disk and delayed inode
3655          * in delayed_nodes_tree.
3656          */
3657         if (BTRFS_I(inode)->last_trans == fs_info->generation)
3658                 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
3659                         &BTRFS_I(inode)->runtime_flags);
3660
3661         /*
3662          * We don't persist the id of the transaction where an unlink operation
3663          * against the inode was last made. So here we assume the inode might
3664          * have been evicted, and therefore the exact value of last_unlink_trans
3665          * lost, and set it to last_trans to avoid metadata inconsistencies
3666          * between the inode and its parent if the inode is fsync'ed and the log
3667          * replayed. For example, in the scenario:
3668          *
3669          * touch mydir/foo
3670          * ln mydir/foo mydir/bar
3671          * sync
3672          * unlink mydir/bar
3673          * echo 2 > /proc/sys/vm/drop_caches   # evicts inode
3674          * xfs_io -c fsync mydir/foo
3675          * <power failure>
3676          * mount fs, triggers fsync log replay
3677          *
3678          * We must make sure that when we fsync our inode foo we also log its
3679          * parent inode, otherwise after log replay the parent still has the
3680          * dentry with the "bar" name but our inode foo has a link count of 1
3681          * and doesn't have an inode ref with the name "bar" anymore.
3682          *
3683          * Setting last_unlink_trans to last_trans is a pessimistic approach,
3684          * but it guarantees correctness at the expense of occasional full
3685          * transaction commits on fsync if our inode is a directory, or if our
3686          * inode is not a directory, logging its parent unnecessarily.
3687          */
3688         BTRFS_I(inode)->last_unlink_trans = BTRFS_I(inode)->last_trans;
3689
3690         path->slots[0]++;
3691         if (inode->i_nlink != 1 ||
3692             path->slots[0] >= btrfs_header_nritems(leaf))
3693                 goto cache_acl;
3694
3695         btrfs_item_key_to_cpu(leaf, &location, path->slots[0]);
3696         if (location.objectid != btrfs_ino(BTRFS_I(inode)))
3697                 goto cache_acl;
3698
3699         ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
3700         if (location.type == BTRFS_INODE_REF_KEY) {
3701                 struct btrfs_inode_ref *ref;
3702
3703                 ref = (struct btrfs_inode_ref *)ptr;
3704                 BTRFS_I(inode)->dir_index = btrfs_inode_ref_index(leaf, ref);
3705         } else if (location.type == BTRFS_INODE_EXTREF_KEY) {
3706                 struct btrfs_inode_extref *extref;
3707
3708                 extref = (struct btrfs_inode_extref *)ptr;
3709                 BTRFS_I(inode)->dir_index = btrfs_inode_extref_index(leaf,
3710                                                                      extref);
3711         }
3712 cache_acl:
3713         /*
3714          * try to precache a NULL acl entry for files that don't have
3715          * any xattrs or acls
3716          */
3717         maybe_acls = acls_after_inode_item(leaf, path->slots[0],
3718                         btrfs_ino(BTRFS_I(inode)), &first_xattr_slot);
3719         if (first_xattr_slot != -1) {
3720                 path->slots[0] = first_xattr_slot;
3721                 ret = btrfs_load_inode_props(inode, path);
3722                 if (ret)
3723                         btrfs_err(fs_info,
3724                                   "error loading props for ino %llu (root %llu): %d",
3725                                   btrfs_ino(BTRFS_I(inode)),
3726                                   root->root_key.objectid, ret);
3727         }
3728         btrfs_free_path(path);
3729
3730         if (!maybe_acls)
3731                 cache_no_acl(inode);
3732
3733         switch (inode->i_mode & S_IFMT) {
3734         case S_IFREG:
3735                 inode->i_mapping->a_ops = &btrfs_aops;
3736                 BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
3737                 inode->i_fop = &btrfs_file_operations;
3738                 inode->i_op = &btrfs_file_inode_operations;
3739                 break;
3740         case S_IFDIR:
3741                 inode->i_fop = &btrfs_dir_file_operations;
3742                 inode->i_op = &btrfs_dir_inode_operations;
3743                 break;
3744         case S_IFLNK:
3745                 inode->i_op = &btrfs_symlink_inode_operations;
3746                 inode_nohighmem(inode);
3747                 inode->i_mapping->a_ops = &btrfs_symlink_aops;
3748                 break;
3749         default:
3750                 inode->i_op = &btrfs_special_inode_operations;
3751                 init_special_inode(inode, inode->i_mode, rdev);
3752                 break;
3753         }
3754
3755         btrfs_update_iflags(inode);
3756         return 0;
3757
3758 make_bad:
3759         btrfs_free_path(path);
3760         make_bad_inode(inode);
3761         return ret;
3762 }
3763
3764 /*
3765  * given a leaf and an inode, copy the inode fields into the leaf
3766  */
3767 static void fill_inode_item(struct btrfs_trans_handle *trans,
3768                             struct extent_buffer *leaf,
3769                             struct btrfs_inode_item *item,
3770                             struct inode *inode)
3771 {
3772         struct btrfs_map_token token;
3773
3774         btrfs_init_map_token(&token);
3775
3776         btrfs_set_token_inode_uid(leaf, item, i_uid_read(inode), &token);
3777         btrfs_set_token_inode_gid(leaf, item, i_gid_read(inode), &token);
3778         btrfs_set_token_inode_size(leaf, item, BTRFS_I(inode)->disk_i_size,
3779                                    &token);
3780         btrfs_set_token_inode_mode(leaf, item, inode->i_mode, &token);
3781         btrfs_set_token_inode_nlink(leaf, item, inode->i_nlink, &token);
3782
3783         btrfs_set_token_timespec_sec(leaf, &item->atime,
3784                                      inode->i_atime.tv_sec, &token);
3785         btrfs_set_token_timespec_nsec(leaf, &item->atime,
3786                                       inode->i_atime.tv_nsec, &token);
3787
3788         btrfs_set_token_timespec_sec(leaf, &item->mtime,
3789                                      inode->i_mtime.tv_sec, &token);
3790         btrfs_set_token_timespec_nsec(leaf, &item->mtime,
3791                                       inode->i_mtime.tv_nsec, &token);
3792
3793         btrfs_set_token_timespec_sec(leaf, &item->ctime,
3794                                      inode->i_ctime.tv_sec, &token);
3795         btrfs_set_token_timespec_nsec(leaf, &item->ctime,
3796                                       inode->i_ctime.tv_nsec, &token);
3797
3798         btrfs_set_token_timespec_sec(leaf, &item->otime,
3799                                      BTRFS_I(inode)->i_otime.tv_sec, &token);
3800         btrfs_set_token_timespec_nsec(leaf, &item->otime,
3801                                       BTRFS_I(inode)->i_otime.tv_nsec, &token);
3802
3803         btrfs_set_token_inode_nbytes(leaf, item, inode_get_bytes(inode),
3804                                      &token);
3805         btrfs_set_token_inode_generation(leaf, item, BTRFS_I(inode)->generation,
3806                                          &token);
3807         btrfs_set_token_inode_sequence(leaf, item, inode->i_version, &token);
3808         btrfs_set_token_inode_transid(leaf, item, trans->transid, &token);
3809         btrfs_set_token_inode_rdev(leaf, item, inode->i_rdev, &token);
3810         btrfs_set_token_inode_flags(leaf, item, BTRFS_I(inode)->flags, &token);
3811         btrfs_set_token_inode_block_group(leaf, item, 0, &token);
3812 }
3813
3814 /*
3815  * copy everything in the in-memory inode into the btree.
3816  */
3817 static noinline int btrfs_update_inode_item(struct btrfs_trans_handle *trans,
3818                                 struct btrfs_root *root, struct inode *inode)
3819 {
3820         struct btrfs_inode_item *inode_item;
3821         struct btrfs_path *path;
3822         struct extent_buffer *leaf;
3823         int ret;
3824
3825         path = btrfs_alloc_path();
3826         if (!path)
3827                 return -ENOMEM;
3828
3829         path->leave_spinning = 1;
3830         ret = btrfs_lookup_inode(trans, root, path, &BTRFS_I(inode)->location,
3831                                  1);
3832         if (ret) {
3833                 if (ret > 0)
3834                         ret = -ENOENT;
3835                 goto failed;
3836         }
3837
3838         leaf = path->nodes[0];
3839         inode_item = btrfs_item_ptr(leaf, path->slots[0],
3840                                     struct btrfs_inode_item);
3841
3842         fill_inode_item(trans, leaf, inode_item, inode);
3843         btrfs_mark_buffer_dirty(leaf);
3844         btrfs_set_inode_last_trans(trans, inode);
3845         ret = 0;
3846 failed:
3847         btrfs_free_path(path);
3848         return ret;
3849 }
3850
3851 /*
3852  * copy everything in the in-memory inode into the btree.
3853  */
3854 noinline int btrfs_update_inode(struct btrfs_trans_handle *trans,
3855                                 struct btrfs_root *root, struct inode *inode)
3856 {
3857         struct btrfs_fs_info *fs_info = root->fs_info;
3858         int ret;
3859
3860         /*
3861          * If the inode is a free space inode, we can deadlock during commit
3862          * if we put it into the delayed code.
3863          *
3864          * The data relocation inode should also be directly updated
3865          * without delay
3866          */
3867         if (!btrfs_is_free_space_inode(BTRFS_I(inode))
3868             && root->root_key.objectid != BTRFS_DATA_RELOC_TREE_OBJECTID
3869             && !test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags)) {
3870                 btrfs_update_root_times(trans, root);
3871
3872                 ret = btrfs_delayed_update_inode(trans, root, inode);
3873                 if (!ret)
3874                         btrfs_set_inode_last_trans(trans, inode);
3875                 return ret;
3876         }
3877
3878         return btrfs_update_inode_item(trans, root, inode);
3879 }
3880
3881 noinline int btrfs_update_inode_fallback(struct btrfs_trans_handle *trans,
3882                                          struct btrfs_root *root,
3883                                          struct inode *inode)
3884 {
3885         int ret;
3886
3887         ret = btrfs_update_inode(trans, root, inode);
3888         if (ret == -ENOSPC)
3889                 return btrfs_update_inode_item(trans, root, inode);
3890         return ret;
3891 }
3892
3893 /*
3894  * unlink helper that gets used here in inode.c and in the tree logging
3895  * recovery code.  It remove a link in a directory with a given name, and
3896  * also drops the back refs in the inode to the directory
3897  */
3898 static int __btrfs_unlink_inode(struct btrfs_trans_handle *trans,
3899                                 struct btrfs_root *root,
3900                                 struct btrfs_inode *dir,
3901                                 struct btrfs_inode *inode,
3902                                 const char *name, int name_len)
3903 {
3904         struct btrfs_fs_info *fs_info = root->fs_info;
3905         struct btrfs_path *path;
3906         int ret = 0;
3907         struct extent_buffer *leaf;
3908         struct btrfs_dir_item *di;
3909         struct btrfs_key key;
3910         u64 index;
3911         u64 ino = btrfs_ino(inode);
3912         u64 dir_ino = btrfs_ino(dir);
3913
3914         path = btrfs_alloc_path();
3915         if (!path) {
3916                 ret = -ENOMEM;
3917                 goto out;
3918         }
3919
3920         path->leave_spinning = 1;
3921         di = btrfs_lookup_dir_item(trans, root, path, dir_ino,
3922                                     name, name_len, -1);
3923         if (IS_ERR(di)) {
3924                 ret = PTR_ERR(di);
3925                 goto err;
3926         }
3927         if (!di) {
3928                 ret = -ENOENT;
3929                 goto err;
3930         }
3931         leaf = path->nodes[0];
3932         btrfs_dir_item_key_to_cpu(leaf, di, &key);
3933         ret = btrfs_delete_one_dir_name(trans, root, path, di);
3934         if (ret)
3935                 goto err;
3936         btrfs_release_path(path);
3937
3938         /*
3939          * If we don't have dir index, we have to get it by looking up
3940          * the inode ref, since we get the inode ref, remove it directly,
3941          * it is unnecessary to do delayed deletion.
3942          *
3943          * But if we have dir index, needn't search inode ref to get it.
3944          * Since the inode ref is close to the inode item, it is better
3945          * that we delay to delete it, and just do this deletion when
3946          * we update the inode item.
3947          */
3948         if (inode->dir_index) {
3949                 ret = btrfs_delayed_delete_inode_ref(inode);
3950                 if (!ret) {
3951                         index = inode->dir_index;
3952                         goto skip_backref;
3953                 }
3954         }
3955
3956         ret = btrfs_del_inode_ref(trans, root, name, name_len, ino,
3957                                   dir_ino, &index);
3958         if (ret) {
3959                 btrfs_info(fs_info,
3960                         "failed to delete reference to %.*s, inode %llu parent %llu",
3961                         name_len, name, ino, dir_ino);
3962                 btrfs_abort_transaction(trans, ret);
3963                 goto err;
3964         }
3965 skip_backref:
3966         ret = btrfs_delete_delayed_dir_index(trans, fs_info, dir, index);
3967         if (ret) {
3968                 btrfs_abort_transaction(trans, ret);
3969                 goto err;
3970         }
3971
3972         ret = btrfs_del_inode_ref_in_log(trans, root, name, name_len, inode,
3973                         dir_ino);
3974         if (ret != 0 && ret != -ENOENT) {
3975                 btrfs_abort_transaction(trans, ret);
3976                 goto err;
3977         }
3978
3979         ret = btrfs_del_dir_entries_in_log(trans, root, name, name_len, dir,
3980                         index);
3981         if (ret == -ENOENT)
3982                 ret = 0;
3983         else if (ret)
3984                 btrfs_abort_transaction(trans, ret);
3985 err:
3986         btrfs_free_path(path);
3987         if (ret)
3988                 goto out;
3989
3990         btrfs_i_size_write(dir, dir->vfs_inode.i_size - name_len * 2);
3991         inode_inc_iversion(&inode->vfs_inode);
3992         inode_inc_iversion(&dir->vfs_inode);
3993         inode->vfs_inode.i_ctime = dir->vfs_inode.i_mtime =
3994                 dir->vfs_inode.i_ctime = current_time(&inode->vfs_inode);
3995         ret = btrfs_update_inode(trans, root, &dir->vfs_inode);
3996 out:
3997         return ret;
3998 }
3999
4000 int btrfs_unlink_inode(struct btrfs_trans_handle *trans,
4001                        struct btrfs_root *root,
4002                        struct btrfs_inode *dir, struct btrfs_inode *inode,
4003                        const char *name, int name_len)
4004 {
4005         int ret;
4006         ret = __btrfs_unlink_inode(trans, root, dir, inode, name, name_len);
4007         if (!ret) {
4008                 drop_nlink(&inode->vfs_inode);
4009                 ret = btrfs_update_inode(trans, root, &inode->vfs_inode);
4010         }
4011         return ret;
4012 }
4013
4014 /*
4015  * helper to start transaction for unlink and rmdir.
4016  *
4017  * unlink and rmdir are special in btrfs, they do not always free space, so
4018  * if we cannot make our reservations the normal way try and see if there is
4019  * plenty of slack room in the global reserve to migrate, otherwise we cannot
4020  * allow the unlink to occur.
4021  */
4022 static struct btrfs_trans_handle *__unlink_start_trans(struct inode *dir)
4023 {
4024         struct btrfs_root *root = BTRFS_I(dir)->root;
4025
4026         /*
4027          * 1 for the possible orphan item
4028          * 1 for the dir item
4029          * 1 for the dir index
4030          * 1 for the inode ref
4031          * 1 for the inode
4032          */
4033         return btrfs_start_transaction_fallback_global_rsv(root, 5, 5);
4034 }
4035
4036 static int btrfs_unlink(struct inode *dir, struct dentry *dentry)
4037 {
4038         struct btrfs_root *root = BTRFS_I(dir)->root;
4039         struct btrfs_trans_handle *trans;
4040         struct inode *inode = d_inode(dentry);
4041         int ret;
4042
4043         trans = __unlink_start_trans(dir);
4044         if (IS_ERR(trans))
4045                 return PTR_ERR(trans);
4046
4047         btrfs_record_unlink_dir(trans, BTRFS_I(dir), BTRFS_I(d_inode(dentry)),
4048                         0);
4049
4050         ret = btrfs_unlink_inode(trans, root, BTRFS_I(dir),
4051                         BTRFS_I(d_inode(dentry)), dentry->d_name.name,
4052                         dentry->d_name.len);
4053         if (ret)
4054                 goto out;
4055
4056         if (inode->i_nlink == 0) {
4057                 ret = btrfs_orphan_add(trans, BTRFS_I(inode));
4058                 if (ret)
4059                         goto out;
4060         }
4061
4062 out:
4063         btrfs_end_transaction(trans);
4064         btrfs_btree_balance_dirty(root->fs_info);
4065         return ret;
4066 }
4067
4068 int btrfs_unlink_subvol(struct btrfs_trans_handle *trans,
4069                         struct btrfs_root *root,
4070                         struct inode *dir, u64 objectid,
4071                         const char *name, int name_len)
4072 {
4073         struct btrfs_fs_info *fs_info = root->fs_info;
4074         struct btrfs_path *path;
4075         struct extent_buffer *leaf;
4076         struct btrfs_dir_item *di;
4077         struct btrfs_key key;
4078         u64 index;
4079         int ret;
4080         u64 dir_ino = btrfs_ino(BTRFS_I(dir));
4081
4082         path = btrfs_alloc_path();
4083         if (!path)
4084                 return -ENOMEM;
4085
4086         di = btrfs_lookup_dir_item(trans, root, path, dir_ino,
4087                                    name, name_len, -1);
4088         if (IS_ERR_OR_NULL(di)) {
4089                 if (!di)
4090                         ret = -ENOENT;
4091                 else
4092                         ret = PTR_ERR(di);
4093                 goto out;
4094         }
4095
4096         leaf = path->nodes[0];
4097         btrfs_dir_item_key_to_cpu(leaf, di, &key);
4098         WARN_ON(key.type != BTRFS_ROOT_ITEM_KEY || key.objectid != objectid);
4099         ret = btrfs_delete_one_dir_name(trans, root, path, di);
4100         if (ret) {
4101                 btrfs_abort_transaction(trans, ret);
4102                 goto out;
4103         }
4104         btrfs_release_path(path);
4105
4106         ret = btrfs_del_root_ref(trans, fs_info, objectid,
4107                                  root->root_key.objectid, dir_ino,
4108                                  &index, name, name_len);
4109         if (ret < 0) {
4110                 if (ret != -ENOENT) {
4111                         btrfs_abort_transaction(trans, ret);
4112                         goto out;
4113                 }
4114                 di = btrfs_search_dir_index_item(root, path, dir_ino,
4115                                                  name, name_len);
4116                 if (IS_ERR_OR_NULL(di)) {
4117                         if (!di)
4118                                 ret = -ENOENT;
4119                         else
4120                                 ret = PTR_ERR(di);
4121                         btrfs_abort_transaction(trans, ret);
4122                         goto out;
4123                 }
4124
4125                 leaf = path->nodes[0];
4126                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
4127                 btrfs_release_path(path);
4128                 index = key.offset;
4129         }
4130         btrfs_release_path(path);
4131
4132         ret = btrfs_delete_delayed_dir_index(trans, fs_info, BTRFS_I(dir), index);
4133         if (ret) {
4134                 btrfs_abort_transaction(trans, ret);
4135                 goto out;
4136         }
4137
4138         btrfs_i_size_write(BTRFS_I(dir), dir->i_size - name_len * 2);
4139         inode_inc_iversion(dir);
4140         dir->i_mtime = dir->i_ctime = current_time(dir);
4141         ret = btrfs_update_inode_fallback(trans, root, dir);
4142         if (ret)
4143                 btrfs_abort_transaction(trans, ret);
4144 out:
4145         btrfs_free_path(path);
4146         return ret;
4147 }
4148
4149 static int btrfs_rmdir(struct inode *dir, struct dentry *dentry)
4150 {
4151         struct inode *inode = d_inode(dentry);
4152         int err = 0;
4153         struct btrfs_root *root = BTRFS_I(dir)->root;
4154         struct btrfs_trans_handle *trans;
4155         u64 last_unlink_trans;
4156
4157         if (inode->i_size > BTRFS_EMPTY_DIR_SIZE)
4158                 return -ENOTEMPTY;
4159         if (btrfs_ino(BTRFS_I(inode)) == BTRFS_FIRST_FREE_OBJECTID)
4160                 return -EPERM;
4161
4162         trans = __unlink_start_trans(dir);
4163         if (IS_ERR(trans))
4164                 return PTR_ERR(trans);
4165
4166         if (unlikely(btrfs_ino(BTRFS_I(inode)) == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)) {
4167                 err = btrfs_unlink_subvol(trans, root, dir,
4168                                           BTRFS_I(inode)->location.objectid,
4169                                           dentry->d_name.name,
4170                                           dentry->d_name.len);
4171                 goto out;
4172         }
4173
4174         err = btrfs_orphan_add(trans, BTRFS_I(inode));
4175         if (err)
4176                 goto out;
4177
4178         last_unlink_trans = BTRFS_I(inode)->last_unlink_trans;
4179
4180         /* now the directory is empty */
4181         err = btrfs_unlink_inode(trans, root, BTRFS_I(dir),
4182                         BTRFS_I(d_inode(dentry)), dentry->d_name.name,
4183                         dentry->d_name.len);
4184         if (!err) {
4185                 btrfs_i_size_write(BTRFS_I(inode), 0);
4186                 /*
4187                  * Propagate the last_unlink_trans value of the deleted dir to
4188                  * its parent directory. This is to prevent an unrecoverable
4189                  * log tree in the case we do something like this:
4190                  * 1) create dir foo
4191                  * 2) create snapshot under dir foo
4192                  * 3) delete the snapshot
4193                  * 4) rmdir foo
4194                  * 5) mkdir foo
4195                  * 6) fsync foo or some file inside foo
4196                  */
4197                 if (last_unlink_trans >= trans->transid)
4198                         BTRFS_I(dir)->last_unlink_trans = last_unlink_trans;
4199         }
4200 out:
4201         btrfs_end_transaction(trans);
4202         btrfs_btree_balance_dirty(root->fs_info);
4203
4204         return err;
4205 }
4206
4207 static int truncate_space_check(struct btrfs_trans_handle *trans,
4208                                 struct btrfs_root *root,
4209                                 u64 bytes_deleted)
4210 {
4211         struct btrfs_fs_info *fs_info = root->fs_info;
4212         int ret;
4213
4214         /*
4215          * This is only used to apply pressure to the enospc system, we don't
4216          * intend to use this reservation at all.
4217          */
4218         bytes_deleted = btrfs_csum_bytes_to_leaves(fs_info, bytes_deleted);
4219         bytes_deleted *= fs_info->nodesize;
4220         ret = btrfs_block_rsv_add(root, &fs_info->trans_block_rsv,
4221                                   bytes_deleted, BTRFS_RESERVE_NO_FLUSH);
4222         if (!ret) {
4223                 trace_btrfs_space_reservation(fs_info, "transaction",
4224                                               trans->transid,
4225                                               bytes_deleted, 1);
4226                 trans->bytes_reserved += bytes_deleted;
4227         }
4228         return ret;
4229
4230 }
4231
4232 static int truncate_inline_extent(struct inode *inode,
4233                                   struct btrfs_path *path,
4234                                   struct btrfs_key *found_key,
4235                                   const u64 item_end,
4236                                   const u64 new_size)
4237 {
4238         struct extent_buffer *leaf = path->nodes[0];
4239         int slot = path->slots[0];
4240         struct btrfs_file_extent_item *fi;
4241         u32 size = (u32)(new_size - found_key->offset);
4242         struct btrfs_root *root = BTRFS_I(inode)->root;
4243
4244         fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
4245
4246         if (btrfs_file_extent_compression(leaf, fi) != BTRFS_COMPRESS_NONE) {
4247                 loff_t offset = new_size;
4248                 loff_t page_end = ALIGN(offset, PAGE_SIZE);
4249
4250                 /*
4251                  * Zero out the remaining of the last page of our inline extent,
4252                  * instead of directly truncating our inline extent here - that
4253                  * would be much more complex (decompressing all the data, then
4254                  * compressing the truncated data, which might be bigger than
4255                  * the size of the inline extent, resize the extent, etc).
4256                  * We release the path because to get the page we might need to
4257                  * read the extent item from disk (data not in the page cache).
4258                  */
4259                 btrfs_release_path(path);
4260                 return btrfs_truncate_block(inode, offset, page_end - offset,
4261                                         0);
4262         }
4263
4264         btrfs_set_file_extent_ram_bytes(leaf, fi, size);
4265         size = btrfs_file_extent_calc_inline_size(size);
4266         btrfs_truncate_item(root->fs_info, path, size, 1);
4267
4268         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state))
4269                 inode_sub_bytes(inode, item_end + 1 - new_size);
4270
4271         return 0;
4272 }
4273
4274 /*
4275  * this can truncate away extent items, csum items and directory items.
4276  * It starts at a high offset and removes keys until it can't find
4277  * any higher than new_size
4278  *
4279  * csum items that cross the new i_size are truncated to the new size
4280  * as well.
4281  *
4282  * min_type is the minimum key type to truncate down to.  If set to 0, this
4283  * will kill all the items on this inode, including the INODE_ITEM_KEY.
4284  */
4285 int btrfs_truncate_inode_items(struct btrfs_trans_handle *trans,
4286                                struct btrfs_root *root,
4287                                struct inode *inode,
4288                                u64 new_size, u32 min_type)
4289 {
4290         struct btrfs_fs_info *fs_info = root->fs_info;
4291         struct btrfs_path *path;
4292         struct extent_buffer *leaf;
4293         struct btrfs_file_extent_item *fi;
4294         struct btrfs_key key;
4295         struct btrfs_key found_key;
4296         u64 extent_start = 0;
4297         u64 extent_num_bytes = 0;
4298         u64 extent_offset = 0;
4299         u64 item_end = 0;
4300         u64 last_size = new_size;
4301         u32 found_type = (u8)-1;
4302         int found_extent;
4303         int del_item;
4304         int pending_del_nr = 0;
4305         int pending_del_slot = 0;
4306         int extent_type = -1;
4307         int ret;
4308         int err = 0;
4309         u64 ino = btrfs_ino(BTRFS_I(inode));
4310         u64 bytes_deleted = 0;
4311         bool be_nice = 0;
4312         bool should_throttle = 0;
4313         bool should_end = 0;
4314
4315         BUG_ON(new_size > 0 && min_type != BTRFS_EXTENT_DATA_KEY);
4316
4317         /*
4318          * for non-free space inodes and ref cows, we want to back off from
4319          * time to time
4320          */
4321         if (!btrfs_is_free_space_inode(BTRFS_I(inode)) &&
4322             test_bit(BTRFS_ROOT_REF_COWS, &root->state))
4323                 be_nice = 1;
4324
4325         path = btrfs_alloc_path();
4326         if (!path)
4327                 return -ENOMEM;
4328         path->reada = READA_BACK;
4329
4330         /*
4331          * We want to drop from the next block forward in case this new size is
4332          * not block aligned since we will be keeping the last block of the
4333          * extent just the way it is.
4334          */
4335         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
4336             root == fs_info->tree_root)
4337                 btrfs_drop_extent_cache(BTRFS_I(inode), ALIGN(new_size,
4338                                         fs_info->sectorsize),
4339                                         (u64)-1, 0);
4340
4341         /*
4342          * This function is also used to drop the items in the log tree before
4343          * we relog the inode, so if root != BTRFS_I(inode)->root, it means
4344          * it is used to drop the loged items. So we shouldn't kill the delayed
4345          * items.
4346          */
4347         if (min_type == 0 && root == BTRFS_I(inode)->root)
4348                 btrfs_kill_delayed_inode_items(BTRFS_I(inode));
4349
4350         key.objectid = ino;
4351         key.offset = (u64)-1;
4352         key.type = (u8)-1;
4353
4354 search_again:
4355         /*
4356          * with a 16K leaf size and 128MB extents, you can actually queue
4357          * up a huge file in a single leaf.  Most of the time that
4358          * bytes_deleted is > 0, it will be huge by the time we get here
4359          */
4360         if (be_nice && bytes_deleted > SZ_32M) {
4361                 if (btrfs_should_end_transaction(trans)) {
4362                         err = -EAGAIN;
4363                         goto error;
4364                 }
4365         }
4366
4367
4368         path->leave_spinning = 1;
4369         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
4370         if (ret < 0) {
4371                 err = ret;
4372                 goto out;
4373         }
4374
4375         if (ret > 0) {
4376                 /* there are no items in the tree for us to truncate, we're
4377                  * done
4378                  */
4379                 if (path->slots[0] == 0)
4380                         goto out;
4381                 path->slots[0]--;
4382         }
4383
4384         while (1) {
4385                 fi = NULL;
4386                 leaf = path->nodes[0];
4387                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
4388                 found_type = found_key.type;
4389
4390                 if (found_key.objectid != ino)
4391                         break;
4392
4393                 if (found_type < min_type)
4394                         break;
4395
4396                 item_end = found_key.offset;
4397                 if (found_type == BTRFS_EXTENT_DATA_KEY) {
4398                         fi = btrfs_item_ptr(leaf, path->slots[0],
4399                                             struct btrfs_file_extent_item);
4400                         extent_type = btrfs_file_extent_type(leaf, fi);
4401                         if (extent_type != BTRFS_FILE_EXTENT_INLINE) {
4402                                 item_end +=
4403                                     btrfs_file_extent_num_bytes(leaf, fi);
4404                         } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
4405                                 item_end += btrfs_file_extent_inline_len(leaf,
4406                                                          path->slots[0], fi);
4407                         }
4408                         item_end--;
4409                 }
4410                 if (found_type > min_type) {
4411                         del_item = 1;
4412                 } else {
4413                         if (item_end < new_size)
4414                                 break;
4415                         if (found_key.offset >= new_size)
4416                                 del_item = 1;
4417                         else
4418                                 del_item = 0;
4419                 }
4420                 found_extent = 0;
4421                 /* FIXME, shrink the extent if the ref count is only 1 */
4422                 if (found_type != BTRFS_EXTENT_DATA_KEY)
4423                         goto delete;
4424
4425                 if (del_item)
4426                         last_size = found_key.offset;
4427                 else
4428                         last_size = new_size;
4429
4430                 if (extent_type != BTRFS_FILE_EXTENT_INLINE) {
4431                         u64 num_dec;
4432                         extent_start = btrfs_file_extent_disk_bytenr(leaf, fi);
4433                         if (!del_item) {
4434                                 u64 orig_num_bytes =
4435                                         btrfs_file_extent_num_bytes(leaf, fi);
4436                                 extent_num_bytes = ALIGN(new_size -
4437                                                 found_key.offset,
4438                                                 fs_info->sectorsize);
4439                                 btrfs_set_file_extent_num_bytes(leaf, fi,
4440                                                          extent_num_bytes);
4441                                 num_dec = (orig_num_bytes -
4442                                            extent_num_bytes);
4443                                 if (test_bit(BTRFS_ROOT_REF_COWS,
4444                                              &root->state) &&
4445                                     extent_start != 0)
4446                                         inode_sub_bytes(inode, num_dec);
4447                                 btrfs_mark_buffer_dirty(leaf);
4448                         } else {
4449                                 extent_num_bytes =
4450                                         btrfs_file_extent_disk_num_bytes(leaf,
4451                                                                          fi);
4452                                 extent_offset = found_key.offset -
4453                                         btrfs_file_extent_offset(leaf, fi);
4454
4455                                 /* FIXME blocksize != 4096 */
4456                                 num_dec = btrfs_file_extent_num_bytes(leaf, fi);
4457                                 if (extent_start != 0) {
4458                                         found_extent = 1;
4459                                         if (test_bit(BTRFS_ROOT_REF_COWS,
4460                                                      &root->state))
4461                                                 inode_sub_bytes(inode, num_dec);
4462                                 }
4463                         }
4464                 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
4465                         /*
4466                          * we can't truncate inline items that have had
4467                          * special encodings
4468                          */
4469                         if (!del_item &&
4470                             btrfs_file_extent_encryption(leaf, fi) == 0 &&
4471                             btrfs_file_extent_other_encoding(leaf, fi) == 0) {
4472
4473                                 /*
4474                                  * Need to release path in order to truncate a
4475                                  * compressed extent. So delete any accumulated
4476                                  * extent items so far.
4477                                  */
4478                                 if (btrfs_file_extent_compression(leaf, fi) !=
4479                                     BTRFS_COMPRESS_NONE && pending_del_nr) {
4480                                         err = btrfs_del_items(trans, root, path,
4481                                                               pending_del_slot,
4482                                                               pending_del_nr);
4483                                         if (err) {
4484                                                 btrfs_abort_transaction(trans,
4485                                                                         err);
4486                                                 goto error;
4487                                         }
4488                                         pending_del_nr = 0;
4489                                 }
4490
4491                                 err = truncate_inline_extent(inode, path,
4492                                                              &found_key,
4493                                                              item_end,
4494                                                              new_size);
4495                                 if (err) {
4496                                         btrfs_abort_transaction(trans, err);
4497                                         goto error;
4498                                 }
4499                         } else if (test_bit(BTRFS_ROOT_REF_COWS,
4500                                             &root->state)) {
4501                                 inode_sub_bytes(inode, item_end + 1 - new_size);
4502                         }
4503                 }
4504 delete:
4505                 if (del_item) {
4506                         if (!pending_del_nr) {
4507                                 /* no pending yet, add ourselves */
4508                                 pending_del_slot = path->slots[0];
4509                                 pending_del_nr = 1;
4510                         } else if (pending_del_nr &&
4511                                    path->slots[0] + 1 == pending_del_slot) {
4512                                 /* hop on the pending chunk */
4513                                 pending_del_nr++;
4514                                 pending_del_slot = path->slots[0];
4515                         } else {
4516                                 BUG();
4517                         }
4518                 } else {
4519                         break;
4520                 }
4521                 should_throttle = 0;
4522
4523                 if (found_extent &&
4524                     (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
4525                      root == fs_info->tree_root)) {
4526                         btrfs_set_path_blocking(path);
4527                         bytes_deleted += extent_num_bytes;
4528                         ret = btrfs_free_extent(trans, fs_info, extent_start,
4529                                                 extent_num_bytes, 0,
4530                                                 btrfs_header_owner(leaf),
4531                                                 ino, extent_offset);
4532                         BUG_ON(ret);
4533                         if (btrfs_should_throttle_delayed_refs(trans, fs_info))
4534                                 btrfs_async_run_delayed_refs(fs_info,
4535                                         trans->delayed_ref_updates * 2,
4536                                         trans->transid, 0);
4537                         if (be_nice) {
4538                                 if (truncate_space_check(trans, root,
4539                                                          extent_num_bytes)) {
4540                                         should_end = 1;
4541                                 }
4542                                 if (btrfs_should_throttle_delayed_refs(trans,
4543                                                                        fs_info))
4544                                         should_throttle = 1;
4545                         }
4546                 }
4547
4548                 if (found_type == BTRFS_INODE_ITEM_KEY)
4549                         break;
4550
4551                 if (path->slots[0] == 0 ||
4552                     path->slots[0] != pending_del_slot ||
4553                     should_throttle || should_end) {
4554                         if (pending_del_nr) {
4555                                 ret = btrfs_del_items(trans, root, path,
4556                                                 pending_del_slot,
4557                                                 pending_del_nr);
4558                                 if (ret) {
4559                                         btrfs_abort_transaction(trans, ret);
4560                                         goto error;
4561                                 }
4562                                 pending_del_nr = 0;
4563                         }
4564                         btrfs_release_path(path);
4565                         if (should_throttle) {
4566                                 unsigned long updates = trans->delayed_ref_updates;
4567                                 if (updates) {
4568                                         trans->delayed_ref_updates = 0;
4569                                         ret = btrfs_run_delayed_refs(trans,
4570                                                                    fs_info,
4571                                                                    updates * 2);
4572                                         if (ret && !err)
4573                                                 err = ret;
4574                                 }
4575                         }
4576                         /*
4577                          * if we failed to refill our space rsv, bail out
4578                          * and let the transaction restart
4579                          */
4580                         if (should_end) {
4581                                 err = -EAGAIN;
4582                                 goto error;
4583                         }
4584                         goto search_again;
4585                 } else {
4586                         path->slots[0]--;
4587                 }
4588         }
4589 out:
4590         if (pending_del_nr) {
4591                 ret = btrfs_del_items(trans, root, path, pending_del_slot,
4592                                       pending_del_nr);
4593                 if (ret)
4594                         btrfs_abort_transaction(trans, ret);
4595         }
4596 error:
4597         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
4598                 ASSERT(last_size >= new_size);
4599                 if (!err && last_size > new_size)
4600                         last_size = new_size;
4601                 btrfs_ordered_update_i_size(inode, last_size, NULL);
4602         }
4603
4604         btrfs_free_path(path);
4605
4606         if (err == 0) {
4607                 /* only inline file may have last_size != new_size */
4608                 if (new_size >= fs_info->sectorsize ||
4609                     new_size > fs_info->max_inline)
4610                         ASSERT(last_size == new_size);
4611         }
4612
4613         if (be_nice && bytes_deleted > SZ_32M) {
4614                 unsigned long updates = trans->delayed_ref_updates;
4615                 if (updates) {
4616                         trans->delayed_ref_updates = 0;
4617                         ret = btrfs_run_delayed_refs(trans, fs_info,
4618                                                      updates * 2);
4619                         if (ret && !err)
4620                                 err = ret;
4621                 }
4622         }
4623         return err;
4624 }
4625
4626 /*
4627  * btrfs_truncate_block - read, zero a chunk and write a block
4628  * @inode - inode that we're zeroing
4629  * @from - the offset to start zeroing
4630  * @len - the length to zero, 0 to zero the entire range respective to the
4631  *      offset
4632  * @front - zero up to the offset instead of from the offset on
4633  *
4634  * This will find the block for the "from" offset and cow the block and zero the
4635  * part we want to zero.  This is used with truncate and hole punching.
4636  */
4637 int btrfs_truncate_block(struct inode *inode, loff_t from, loff_t len,
4638                         int front)
4639 {
4640         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
4641         struct address_space *mapping = inode->i_mapping;
4642         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
4643         struct btrfs_ordered_extent *ordered;
4644         struct extent_state *cached_state = NULL;
4645         char *kaddr;
4646         u32 blocksize = fs_info->sectorsize;
4647         pgoff_t index = from >> PAGE_SHIFT;
4648         unsigned offset = from & (blocksize - 1);
4649         struct page *page;
4650         gfp_t mask = btrfs_alloc_write_mask(mapping);
4651         int ret = 0;
4652         u64 block_start;
4653         u64 block_end;
4654
4655         if ((offset & (blocksize - 1)) == 0 &&
4656             (!len || ((len & (blocksize - 1)) == 0)))
4657                 goto out;
4658
4659         ret = btrfs_delalloc_reserve_space(inode,
4660                         round_down(from, blocksize), blocksize);
4661         if (ret)
4662                 goto out;
4663
4664 again:
4665         page = find_or_create_page(mapping, index, mask);
4666         if (!page) {
4667                 btrfs_delalloc_release_space(inode,
4668                                 round_down(from, blocksize),
4669                                 blocksize);
4670                 ret = -ENOMEM;
4671                 goto out;
4672         }
4673
4674         block_start = round_down(from, blocksize);
4675         block_end = block_start + blocksize - 1;
4676
4677         if (!PageUptodate(page)) {
4678                 ret = btrfs_readpage(NULL, page);
4679                 lock_page(page);
4680                 if (page->mapping != mapping) {
4681                         unlock_page(page);
4682                         put_page(page);
4683                         goto again;
4684                 }
4685                 if (!PageUptodate(page)) {
4686                         ret = -EIO;
4687                         goto out_unlock;
4688                 }
4689         }
4690         wait_on_page_writeback(page);
4691
4692         lock_extent_bits(io_tree, block_start, block_end, &cached_state);
4693         set_page_extent_mapped(page);
4694
4695         ordered = btrfs_lookup_ordered_extent(inode, block_start);
4696         if (ordered) {
4697                 unlock_extent_cached(io_tree, block_start, block_end,
4698                                      &cached_state, GFP_NOFS);
4699                 unlock_page(page);
4700                 put_page(page);
4701                 btrfs_start_ordered_extent(inode, ordered, 1);
4702                 btrfs_put_ordered_extent(ordered);
4703                 goto again;
4704         }
4705
4706         clear_extent_bit(&BTRFS_I(inode)->io_tree, block_start, block_end,
4707                           EXTENT_DIRTY | EXTENT_DELALLOC |
4708                           EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG,
4709                           0, 0, &cached_state, GFP_NOFS);
4710
4711         ret = btrfs_set_extent_delalloc(inode, block_start, block_end,
4712                                         &cached_state, 0);
4713         if (ret) {
4714                 unlock_extent_cached(io_tree, block_start, block_end,
4715                                      &cached_state, GFP_NOFS);
4716                 goto out_unlock;
4717         }
4718
4719         if (offset != blocksize) {
4720                 if (!len)
4721                         len = blocksize - offset;
4722                 kaddr = kmap(page);
4723                 if (front)
4724                         memset(kaddr + (block_start - page_offset(page)),
4725                                 0, offset);
4726                 else
4727                         memset(kaddr + (block_start - page_offset(page)) +  offset,
4728                                 0, len);
4729                 flush_dcache_page(page);
4730                 kunmap(page);
4731         }
4732         ClearPageChecked(page);
4733         set_page_dirty(page);
4734         unlock_extent_cached(io_tree, block_start, block_end, &cached_state,
4735                              GFP_NOFS);
4736
4737 out_unlock:
4738         if (ret)
4739                 btrfs_delalloc_release_space(inode, block_start,
4740                                              blocksize);
4741         unlock_page(page);
4742         put_page(page);
4743 out:
4744         return ret;
4745 }
4746
4747 static int maybe_insert_hole(struct btrfs_root *root, struct inode *inode,
4748                              u64 offset, u64 len)
4749 {
4750         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
4751         struct btrfs_trans_handle *trans;
4752         int ret;
4753
4754         /*
4755          * Still need to make sure the inode looks like it's been updated so
4756          * that any holes get logged if we fsync.
4757          */
4758         if (btrfs_fs_incompat(fs_info, NO_HOLES)) {
4759                 BTRFS_I(inode)->last_trans = fs_info->generation;
4760                 BTRFS_I(inode)->last_sub_trans = root->log_transid;
4761                 BTRFS_I(inode)->last_log_commit = root->last_log_commit;
4762                 return 0;
4763         }
4764
4765         /*
4766          * 1 - for the one we're dropping
4767          * 1 - for the one we're adding
4768          * 1 - for updating the inode.
4769          */
4770         trans = btrfs_start_transaction(root, 3);
4771         if (IS_ERR(trans))
4772                 return PTR_ERR(trans);
4773
4774         ret = btrfs_drop_extents(trans, root, inode, offset, offset + len, 1);
4775         if (ret) {
4776                 btrfs_abort_transaction(trans, ret);
4777                 btrfs_end_transaction(trans);
4778                 return ret;
4779         }
4780
4781         ret = btrfs_insert_file_extent(trans, root, btrfs_ino(BTRFS_I(inode)),
4782                         offset, 0, 0, len, 0, len, 0, 0, 0);
4783         if (ret)
4784                 btrfs_abort_transaction(trans, ret);
4785         else
4786                 btrfs_update_inode(trans, root, inode);
4787         btrfs_end_transaction(trans);
4788         return ret;
4789 }
4790
4791 /*
4792  * This function puts in dummy file extents for the area we're creating a hole
4793  * for.  So if we are truncating this file to a larger size we need to insert
4794  * these file extents so that btrfs_get_extent will return a EXTENT_MAP_HOLE for
4795  * the range between oldsize and size
4796  */
4797 int btrfs_cont_expand(struct inode *inode, loff_t oldsize, loff_t size)
4798 {
4799         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
4800         struct btrfs_root *root = BTRFS_I(inode)->root;
4801         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
4802         struct extent_map *em = NULL;
4803         struct extent_state *cached_state = NULL;
4804         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
4805         u64 hole_start = ALIGN(oldsize, fs_info->sectorsize);
4806         u64 block_end = ALIGN(size, fs_info->sectorsize);
4807         u64 last_byte;
4808         u64 cur_offset;
4809         u64 hole_size;
4810         int err = 0;
4811
4812         /*
4813          * If our size started in the middle of a block we need to zero out the
4814          * rest of the block before we expand the i_size, otherwise we could
4815          * expose stale data.
4816          */
4817         err = btrfs_truncate_block(inode, oldsize, 0, 0);
4818         if (err)
4819                 return err;
4820
4821         if (size <= hole_start)
4822                 return 0;
4823
4824         while (1) {
4825                 struct btrfs_ordered_extent *ordered;
4826
4827                 lock_extent_bits(io_tree, hole_start, block_end - 1,
4828                                  &cached_state);
4829                 ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), hole_start,
4830                                                      block_end - hole_start);
4831                 if (!ordered)
4832                         break;
4833                 unlock_extent_cached(io_tree, hole_start, block_end - 1,
4834                                      &cached_state, GFP_NOFS);
4835                 btrfs_start_ordered_extent(inode, ordered, 1);
4836                 btrfs_put_ordered_extent(ordered);
4837         }
4838
4839         cur_offset = hole_start;
4840         while (1) {
4841                 em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, cur_offset,
4842                                 block_end - cur_offset, 0);
4843                 if (IS_ERR(em)) {
4844                         err = PTR_ERR(em);
4845                         em = NULL;
4846                         break;
4847                 }
4848                 last_byte = min(extent_map_end(em), block_end);
4849                 last_byte = ALIGN(last_byte, fs_info->sectorsize);
4850                 if (!test_bit(EXTENT_FLAG_PREALLOC, &em->flags)) {
4851                         struct extent_map *hole_em;
4852                         hole_size = last_byte - cur_offset;
4853
4854                         err = maybe_insert_hole(root, inode, cur_offset,
4855                                                 hole_size);
4856                         if (err)
4857                                 break;
4858                         btrfs_drop_extent_cache(BTRFS_I(inode), cur_offset,
4859                                                 cur_offset + hole_size - 1, 0);
4860                         hole_em = alloc_extent_map();
4861                         if (!hole_em) {
4862                                 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
4863                                         &BTRFS_I(inode)->runtime_flags);
4864                                 goto next;
4865                         }
4866                         hole_em->start = cur_offset;
4867                         hole_em->len = hole_size;
4868                         hole_em->orig_start = cur_offset;
4869
4870                         hole_em->block_start = EXTENT_MAP_HOLE;
4871                         hole_em->block_len = 0;
4872                         hole_em->orig_block_len = 0;
4873                         hole_em->ram_bytes = hole_size;
4874                         hole_em->bdev = fs_info->fs_devices->latest_bdev;
4875                         hole_em->compress_type = BTRFS_COMPRESS_NONE;
4876                         hole_em->generation = fs_info->generation;
4877
4878                         while (1) {
4879                                 write_lock(&em_tree->lock);
4880                                 err = add_extent_mapping(em_tree, hole_em, 1);
4881                                 write_unlock(&em_tree->lock);
4882                                 if (err != -EEXIST)
4883                                         break;
4884                                 btrfs_drop_extent_cache(BTRFS_I(inode),
4885                                                         cur_offset,
4886                                                         cur_offset +
4887                                                         hole_size - 1, 0);
4888                         }
4889                         free_extent_map(hole_em);
4890                 }
4891 next:
4892                 free_extent_map(em);
4893                 em = NULL;
4894                 cur_offset = last_byte;
4895                 if (cur_offset >= block_end)
4896                         break;
4897         }
4898         free_extent_map(em);
4899         unlock_extent_cached(io_tree, hole_start, block_end - 1, &cached_state,
4900                              GFP_NOFS);
4901         return err;
4902 }
4903
4904 static int btrfs_setsize(struct inode *inode, struct iattr *attr)
4905 {
4906         struct btrfs_root *root = BTRFS_I(inode)->root;
4907         struct btrfs_trans_handle *trans;
4908         loff_t oldsize = i_size_read(inode);
4909         loff_t newsize = attr->ia_size;
4910         int mask = attr->ia_valid;
4911         int ret;
4912
4913         /*
4914          * The regular truncate() case without ATTR_CTIME and ATTR_MTIME is a
4915          * special case where we need to update the times despite not having
4916          * these flags set.  For all other operations the VFS set these flags
4917          * explicitly if it wants a timestamp update.
4918          */
4919         if (newsize != oldsize) {
4920                 inode_inc_iversion(inode);
4921                 if (!(mask & (ATTR_CTIME | ATTR_MTIME)))
4922                         inode->i_ctime = inode->i_mtime =
4923                                 current_time(inode);
4924         }
4925
4926         if (newsize > oldsize) {
4927                 /*
4928                  * Don't do an expanding truncate while snapshoting is ongoing.
4929                  * This is to ensure the snapshot captures a fully consistent
4930                  * state of this file - if the snapshot captures this expanding
4931                  * truncation, it must capture all writes that happened before
4932                  * this truncation.
4933                  */
4934                 btrfs_wait_for_snapshot_creation(root);
4935                 ret = btrfs_cont_expand(inode, oldsize, newsize);
4936                 if (ret) {
4937                         btrfs_end_write_no_snapshoting(root);
4938                         return ret;
4939                 }
4940
4941                 trans = btrfs_start_transaction(root, 1);
4942                 if (IS_ERR(trans)) {
4943                         btrfs_end_write_no_snapshoting(root);
4944                         return PTR_ERR(trans);
4945                 }
4946
4947                 i_size_write(inode, newsize);
4948                 btrfs_ordered_update_i_size(inode, i_size_read(inode), NULL);
4949                 pagecache_isize_extended(inode, oldsize, newsize);
4950                 ret = btrfs_update_inode(trans, root, inode);
4951                 btrfs_end_write_no_snapshoting(root);
4952                 btrfs_end_transaction(trans);
4953         } else {
4954
4955                 /*
4956                  * We're truncating a file that used to have good data down to
4957                  * zero. Make sure it gets into the ordered flush list so that
4958                  * any new writes get down to disk quickly.
4959                  */
4960                 if (newsize == 0)
4961                         set_bit(BTRFS_INODE_ORDERED_DATA_CLOSE,
4962                                 &BTRFS_I(inode)->runtime_flags);
4963
4964                 /*
4965                  * 1 for the orphan item we're going to add
4966                  * 1 for the orphan item deletion.
4967                  */
4968                 trans = btrfs_start_transaction(root, 2);
4969                 if (IS_ERR(trans))
4970                         return PTR_ERR(trans);
4971
4972                 /*
4973                  * We need to do this in case we fail at _any_ point during the
4974                  * actual truncate.  Once we do the truncate_setsize we could
4975                  * invalidate pages which forces any outstanding ordered io to
4976                  * be instantly completed which will give us extents that need
4977                  * to be truncated.  If we fail to get an orphan inode down we
4978                  * could have left over extents that were never meant to live,
4979                  * so we need to guarantee from this point on that everything
4980                  * will be consistent.
4981                  */
4982                 ret = btrfs_orphan_add(trans, BTRFS_I(inode));
4983                 btrfs_end_transaction(trans);
4984                 if (ret)
4985                         return ret;
4986
4987                 /* we don't support swapfiles, so vmtruncate shouldn't fail */
4988                 truncate_setsize(inode, newsize);
4989
4990                 /* Disable nonlocked read DIO to avoid the end less truncate */
4991                 btrfs_inode_block_unlocked_dio(BTRFS_I(inode));
4992                 inode_dio_wait(inode);
4993                 btrfs_inode_resume_unlocked_dio(BTRFS_I(inode));
4994
4995                 ret = btrfs_truncate(inode);
4996                 if (ret && inode->i_nlink) {
4997                         int err;
4998
4999                         /* To get a stable disk_i_size */
5000                         err = btrfs_wait_ordered_range(inode, 0, (u64)-1);
5001                         if (err) {
5002                                 btrfs_orphan_del(NULL, BTRFS_I(inode));
5003                                 return err;
5004                         }
5005
5006                         /*
5007                          * failed to truncate, disk_i_size is only adjusted down
5008                          * as we remove extents, so it should represent the true
5009                          * size of the inode, so reset the in memory size and
5010                          * delete our orphan entry.
5011                          */
5012                         trans = btrfs_join_transaction(root);
5013                         if (IS_ERR(trans)) {
5014                                 btrfs_orphan_del(NULL, BTRFS_I(inode));
5015                                 return ret;
5016                         }
5017                         i_size_write(inode, BTRFS_I(inode)->disk_i_size);
5018                         err = btrfs_orphan_del(trans, BTRFS_I(inode));
5019                         if (err)
5020                                 btrfs_abort_transaction(trans, err);
5021                         btrfs_end_transaction(trans);
5022                 }
5023         }
5024
5025         return ret;
5026 }
5027
5028 static int btrfs_setattr(struct dentry *dentry, struct iattr *attr)
5029 {
5030         struct inode *inode = d_inode(dentry);
5031         struct btrfs_root *root = BTRFS_I(inode)->root;
5032         int err;
5033
5034         if (btrfs_root_readonly(root))
5035                 return -EROFS;
5036
5037         err = setattr_prepare(dentry, attr);
5038         if (err)
5039                 return err;
5040
5041         if (S_ISREG(inode->i_mode) && (attr->ia_valid & ATTR_SIZE)) {
5042                 err = btrfs_setsize(inode, attr);
5043                 if (err)
5044                         return err;
5045         }
5046
5047         if (attr->ia_valid) {
5048                 setattr_copy(inode, attr);
5049                 inode_inc_iversion(inode);
5050                 err = btrfs_dirty_inode(inode);
5051
5052                 if (!err && attr->ia_valid & ATTR_MODE)
5053                         err = posix_acl_chmod(inode, inode->i_mode);
5054         }
5055
5056         return err;
5057 }
5058
5059 /*
5060  * While truncating the inode pages during eviction, we get the VFS calling
5061  * btrfs_invalidatepage() against each page of the inode. This is slow because
5062  * the calls to btrfs_invalidatepage() result in a huge amount of calls to
5063  * lock_extent_bits() and clear_extent_bit(), which keep merging and splitting
5064  * extent_state structures over and over, wasting lots of time.
5065  *
5066  * Therefore if the inode is being evicted, let btrfs_invalidatepage() skip all
5067  * those expensive operations on a per page basis and do only the ordered io
5068  * finishing, while we release here the extent_map and extent_state structures,
5069  * without the excessive merging and splitting.
5070  */
5071 static void evict_inode_truncate_pages(struct inode *inode)
5072 {
5073         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
5074         struct extent_map_tree *map_tree = &BTRFS_I(inode)->extent_tree;
5075         struct rb_node *node;
5076
5077         ASSERT(inode->i_state & I_FREEING);
5078         truncate_inode_pages_final(&inode->i_data);
5079
5080         write_lock(&map_tree->lock);
5081         while (!RB_EMPTY_ROOT(&map_tree->map)) {
5082                 struct extent_map *em;
5083
5084                 node = rb_first(&map_tree->map);
5085                 em = rb_entry(node, struct extent_map, rb_node);
5086                 clear_bit(EXTENT_FLAG_PINNED, &em->flags);
5087                 clear_bit(EXTENT_FLAG_LOGGING, &em->flags);
5088                 remove_extent_mapping(map_tree, em);
5089                 free_extent_map(em);
5090                 if (need_resched()) {
5091                         write_unlock(&map_tree->lock);
5092                         cond_resched();
5093                         write_lock(&map_tree->lock);
5094                 }
5095         }
5096         write_unlock(&map_tree->lock);
5097
5098         /*
5099          * Keep looping until we have no more ranges in the io tree.
5100          * We can have ongoing bios started by readpages (called from readahead)
5101          * that have their endio callback (extent_io.c:end_bio_extent_readpage)
5102          * still in progress (unlocked the pages in the bio but did not yet
5103          * unlocked the ranges in the io tree). Therefore this means some
5104          * ranges can still be locked and eviction started because before
5105          * submitting those bios, which are executed by a separate task (work
5106          * queue kthread), inode references (inode->i_count) were not taken
5107          * (which would be dropped in the end io callback of each bio).
5108          * Therefore here we effectively end up waiting for those bios and
5109          * anyone else holding locked ranges without having bumped the inode's
5110          * reference count - if we don't do it, when they access the inode's
5111          * io_tree to unlock a range it may be too late, leading to an
5112          * use-after-free issue.
5113          */
5114         spin_lock(&io_tree->lock);
5115         while (!RB_EMPTY_ROOT(&io_tree->state)) {
5116                 struct extent_state *state;
5117                 struct extent_state *cached_state = NULL;
5118                 u64 start;
5119                 u64 end;
5120
5121                 node = rb_first(&io_tree->state);
5122                 state = rb_entry(node, struct extent_state, rb_node);
5123                 start = state->start;
5124                 end = state->end;
5125                 spin_unlock(&io_tree->lock);
5126
5127                 lock_extent_bits(io_tree, start, end, &cached_state);
5128
5129                 /*
5130                  * If still has DELALLOC flag, the extent didn't reach disk,
5131                  * and its reserved space won't be freed by delayed_ref.
5132                  * So we need to free its reserved space here.
5133                  * (Refer to comment in btrfs_invalidatepage, case 2)
5134                  *
5135                  * Note, end is the bytenr of last byte, so we need + 1 here.
5136                  */
5137                 if (state->state & EXTENT_DELALLOC)
5138                         btrfs_qgroup_free_data(inode, start, end - start + 1);
5139
5140                 clear_extent_bit(io_tree, start, end,
5141                                  EXTENT_LOCKED | EXTENT_DIRTY |
5142                                  EXTENT_DELALLOC | EXTENT_DO_ACCOUNTING |
5143                                  EXTENT_DEFRAG, 1, 1,
5144                                  &cached_state, GFP_NOFS);
5145
5146                 cond_resched();
5147                 spin_lock(&io_tree->lock);
5148         }
5149         spin_unlock(&io_tree->lock);
5150 }
5151
5152 void btrfs_evict_inode(struct inode *inode)
5153 {
5154         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
5155         struct btrfs_trans_handle *trans;
5156         struct btrfs_root *root = BTRFS_I(inode)->root;
5157         struct btrfs_block_rsv *rsv, *global_rsv;
5158         int steal_from_global = 0;
5159         u64 min_size;
5160         int ret;
5161
5162         trace_btrfs_inode_evict(inode);
5163
5164         if (!root) {
5165                 kmem_cache_free(btrfs_inode_cachep, BTRFS_I(inode));
5166                 return;
5167         }
5168
5169         min_size = btrfs_calc_trunc_metadata_size(fs_info, 1);
5170
5171         evict_inode_truncate_pages(inode);
5172
5173         if (inode->i_nlink &&
5174             ((btrfs_root_refs(&root->root_item) != 0 &&
5175               root->root_key.objectid != BTRFS_ROOT_TREE_OBJECTID) ||
5176              btrfs_is_free_space_inode(BTRFS_I(inode))))
5177                 goto no_delete;
5178
5179         if (is_bad_inode(inode)) {
5180                 btrfs_orphan_del(NULL, BTRFS_I(inode));
5181                 goto no_delete;
5182         }
5183         /* do we really want it for ->i_nlink > 0 and zero btrfs_root_refs? */
5184         if (!special_file(inode->i_mode))
5185                 btrfs_wait_ordered_range(inode, 0, (u64)-1);
5186
5187         btrfs_free_io_failure_record(BTRFS_I(inode), 0, (u64)-1);
5188
5189         if (test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags)) {
5190                 BUG_ON(test_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
5191                                  &BTRFS_I(inode)->runtime_flags));
5192                 goto no_delete;
5193         }
5194
5195         if (inode->i_nlink > 0) {
5196                 BUG_ON(btrfs_root_refs(&root->root_item) != 0 &&
5197                        root->root_key.objectid != BTRFS_ROOT_TREE_OBJECTID);
5198                 goto no_delete;
5199         }
5200
5201         ret = btrfs_commit_inode_delayed_inode(BTRFS_I(inode));
5202         if (ret) {
5203                 btrfs_orphan_del(NULL, BTRFS_I(inode));
5204                 goto no_delete;
5205         }
5206
5207         rsv = btrfs_alloc_block_rsv(fs_info, BTRFS_BLOCK_RSV_TEMP);
5208         if (!rsv) {
5209                 btrfs_orphan_del(NULL, BTRFS_I(inode));
5210                 goto no_delete;
5211         }
5212         rsv->size = min_size;
5213         rsv->failfast = 1;
5214         global_rsv = &fs_info->global_block_rsv;
5215
5216         btrfs_i_size_write(BTRFS_I(inode), 0);
5217
5218         /*
5219          * This is a bit simpler than btrfs_truncate since we've already
5220          * reserved our space for our orphan item in the unlink, so we just
5221          * need to reserve some slack space in case we add bytes and update
5222          * inode item when doing the truncate.
5223          */
5224         while (1) {
5225                 ret = btrfs_block_rsv_refill(root, rsv, min_size,
5226                                              BTRFS_RESERVE_FLUSH_LIMIT);
5227
5228                 /*
5229                  * Try and steal from the global reserve since we will
5230                  * likely not use this space anyway, we want to try as
5231                  * hard as possible to get this to work.
5232                  */
5233                 if (ret)
5234                         steal_from_global++;
5235                 else
5236                         steal_from_global = 0;
5237                 ret = 0;
5238
5239                 /*
5240                  * steal_from_global == 0: we reserved stuff, hooray!
5241                  * steal_from_global == 1: we didn't reserve stuff, boo!
5242                  * steal_from_global == 2: we've committed, still not a lot of
5243                  * room but maybe we'll have room in the global reserve this
5244                  * time.
5245                  * steal_from_global == 3: abandon all hope!
5246                  */
5247                 if (steal_from_global > 2) {
5248                         btrfs_warn(fs_info,
5249                                    "Could not get space for a delete, will truncate on mount %d",
5250                                    ret);
5251                         btrfs_orphan_del(NULL, BTRFS_I(inode));
5252                         btrfs_free_block_rsv(fs_info, rsv);
5253                         goto no_delete;
5254                 }
5255
5256                 trans = btrfs_join_transaction(root);
5257                 if (IS_ERR(trans)) {
5258                         btrfs_orphan_del(NULL, BTRFS_I(inode));
5259                         btrfs_free_block_rsv(fs_info, rsv);
5260                         goto no_delete;
5261                 }
5262
5263                 /*
5264                  * We can't just steal from the global reserve, we need to make
5265                  * sure there is room to do it, if not we need to commit and try
5266                  * again.
5267                  */
5268                 if (steal_from_global) {
5269                         if (!btrfs_check_space_for_delayed_refs(trans, fs_info))
5270                                 ret = btrfs_block_rsv_migrate(global_rsv, rsv,
5271                                                               min_size, 0);
5272                         else
5273                                 ret = -ENOSPC;
5274                 }
5275
5276                 /*
5277                  * Couldn't steal from the global reserve, we have too much
5278                  * pending stuff built up, commit the transaction and try it
5279                  * again.
5280                  */
5281                 if (ret) {
5282                         ret = btrfs_commit_transaction(trans);
5283                         if (ret) {
5284                                 btrfs_orphan_del(NULL, BTRFS_I(inode));
5285                                 btrfs_free_block_rsv(fs_info, rsv);
5286                                 goto no_delete;
5287                         }
5288                         continue;
5289                 } else {
5290                         steal_from_global = 0;
5291                 }
5292
5293                 trans->block_rsv = rsv;
5294
5295                 ret = btrfs_truncate_inode_items(trans, root, inode, 0, 0);
5296                 if (ret != -ENOSPC && ret != -EAGAIN)
5297                         break;
5298
5299                 trans->block_rsv = &fs_info->trans_block_rsv;
5300                 btrfs_end_transaction(trans);
5301                 trans = NULL;
5302                 btrfs_btree_balance_dirty(fs_info);
5303         }
5304
5305         btrfs_free_block_rsv(fs_info, rsv);
5306
5307         /*
5308          * Errors here aren't a big deal, it just means we leave orphan items
5309          * in the tree.  They will be cleaned up on the next mount.
5310          */
5311         if (ret == 0) {
5312                 trans->block_rsv = root->orphan_block_rsv;
5313                 btrfs_orphan_del(trans, BTRFS_I(inode));
5314         } else {
5315                 btrfs_orphan_del(NULL, BTRFS_I(inode));
5316         }
5317
5318         trans->block_rsv = &fs_info->trans_block_rsv;
5319         if (!(root == fs_info->tree_root ||
5320               root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID))
5321                 btrfs_return_ino(root, btrfs_ino(BTRFS_I(inode)));
5322
5323         btrfs_end_transaction(trans);
5324         btrfs_btree_balance_dirty(fs_info);
5325 no_delete:
5326         btrfs_remove_delayed_node(BTRFS_I(inode));
5327         clear_inode(inode);
5328 }
5329
5330 /*
5331  * this returns the key found in the dir entry in the location pointer.
5332  * If no dir entries were found, location->objectid is 0.
5333  */
5334 static int btrfs_inode_by_name(struct inode *dir, struct dentry *dentry,
5335                                struct btrfs_key *location)
5336 {
5337         const char *name = dentry->d_name.name;
5338         int namelen = dentry->d_name.len;
5339         struct btrfs_dir_item *di;
5340         struct btrfs_path *path;
5341         struct btrfs_root *root = BTRFS_I(dir)->root;
5342         int ret = 0;
5343
5344         path = btrfs_alloc_path();
5345         if (!path)
5346                 return -ENOMEM;
5347
5348         di = btrfs_lookup_dir_item(NULL, root, path, btrfs_ino(BTRFS_I(dir)),
5349                         name, namelen, 0);
5350         if (IS_ERR(di))
5351                 ret = PTR_ERR(di);
5352
5353         if (IS_ERR_OR_NULL(di))
5354                 goto out_err;
5355
5356         btrfs_dir_item_key_to_cpu(path->nodes[0], di, location);
5357 out:
5358         btrfs_free_path(path);
5359         return ret;
5360 out_err:
5361         location->objectid = 0;
5362         goto out;
5363 }
5364
5365 /*
5366  * when we hit a tree root in a directory, the btrfs part of the inode
5367  * needs to be changed to reflect the root directory of the tree root.  This
5368  * is kind of like crossing a mount point.
5369  */
5370 static int fixup_tree_root_location(struct btrfs_fs_info *fs_info,
5371                                     struct inode *dir,
5372                                     struct dentry *dentry,
5373                                     struct btrfs_key *location,
5374                                     struct btrfs_root **sub_root)
5375 {
5376         struct btrfs_path *path;
5377         struct btrfs_root *new_root;
5378         struct btrfs_root_ref *ref;
5379         struct extent_buffer *leaf;
5380         struct btrfs_key key;
5381         int ret;
5382         int err = 0;
5383
5384         path = btrfs_alloc_path();
5385         if (!path) {
5386                 err = -ENOMEM;
5387                 goto out;
5388         }
5389
5390         err = -ENOENT;
5391         key.objectid = BTRFS_I(dir)->root->root_key.objectid;
5392         key.type = BTRFS_ROOT_REF_KEY;
5393         key.offset = location->objectid;
5394
5395         ret = btrfs_search_slot(NULL, fs_info->tree_root, &key, path, 0, 0);
5396         if (ret) {
5397                 if (ret < 0)
5398                         err = ret;
5399                 goto out;
5400         }
5401
5402         leaf = path->nodes[0];
5403         ref = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_root_ref);
5404         if (btrfs_root_ref_dirid(leaf, ref) != btrfs_ino(BTRFS_I(dir)) ||
5405             btrfs_root_ref_name_len(leaf, ref) != dentry->d_name.len)
5406                 goto out;
5407
5408         ret = memcmp_extent_buffer(leaf, dentry->d_name.name,
5409                                    (unsigned long)(ref + 1),
5410                                    dentry->d_name.len);
5411         if (ret)
5412                 goto out;
5413
5414         btrfs_release_path(path);
5415
5416         new_root = btrfs_read_fs_root_no_name(fs_info, location);
5417         if (IS_ERR(new_root)) {
5418                 err = PTR_ERR(new_root);
5419                 goto out;
5420         }
5421
5422         *sub_root = new_root;
5423         location->objectid = btrfs_root_dirid(&new_root->root_item);
5424         location->type = BTRFS_INODE_ITEM_KEY;
5425         location->offset = 0;
5426         err = 0;
5427 out:
5428         btrfs_free_path(path);
5429         return err;
5430 }
5431
5432 static void inode_tree_add(struct inode *inode)
5433 {
5434         struct btrfs_root *root = BTRFS_I(inode)->root;
5435         struct btrfs_inode *entry;
5436         struct rb_node **p;
5437         struct rb_node *parent;
5438         struct rb_node *new = &BTRFS_I(inode)->rb_node;
5439         u64 ino = btrfs_ino(BTRFS_I(inode));
5440
5441         if (inode_unhashed(inode))
5442                 return;
5443         parent = NULL;
5444         spin_lock(&root->inode_lock);
5445         p = &root->inode_tree.rb_node;
5446         while (*p) {
5447                 parent = *p;
5448                 entry = rb_entry(parent, struct btrfs_inode, rb_node);
5449
5450                 if (ino < btrfs_ino(BTRFS_I(&entry->vfs_inode)))
5451                         p = &parent->rb_left;
5452                 else if (ino > btrfs_ino(BTRFS_I(&entry->vfs_inode)))
5453                         p = &parent->rb_right;
5454                 else {
5455                         WARN_ON(!(entry->vfs_inode.i_state &
5456                                   (I_WILL_FREE | I_FREEING)));
5457                         rb_replace_node(parent, new, &root->inode_tree);
5458                         RB_CLEAR_NODE(parent);
5459                         spin_unlock(&root->inode_lock);
5460                         return;
5461                 }
5462         }
5463         rb_link_node(new, parent, p);
5464         rb_insert_color(new, &root->inode_tree);
5465         spin_unlock(&root->inode_lock);
5466 }
5467
5468 static void inode_tree_del(struct inode *inode)
5469 {
5470         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
5471         struct btrfs_root *root = BTRFS_I(inode)->root;
5472         int empty = 0;
5473
5474         spin_lock(&root->inode_lock);
5475         if (!RB_EMPTY_NODE(&BTRFS_I(inode)->rb_node)) {
5476                 rb_erase(&BTRFS_I(inode)->rb_node, &root->inode_tree);
5477                 RB_CLEAR_NODE(&BTRFS_I(inode)->rb_node);
5478                 empty = RB_EMPTY_ROOT(&root->inode_tree);
5479         }
5480         spin_unlock(&root->inode_lock);
5481
5482         if (empty && btrfs_root_refs(&root->root_item) == 0) {
5483                 synchronize_srcu(&fs_info->subvol_srcu);
5484                 spin_lock(&root->inode_lock);
5485                 empty = RB_EMPTY_ROOT(&root->inode_tree);
5486                 spin_unlock(&root->inode_lock);
5487                 if (empty)
5488                         btrfs_add_dead_root(root);
5489         }
5490 }
5491
5492 void btrfs_invalidate_inodes(struct btrfs_root *root)
5493 {
5494         struct btrfs_fs_info *fs_info = root->fs_info;
5495         struct rb_node *node;
5496         struct rb_node *prev;
5497         struct btrfs_inode *entry;
5498         struct inode *inode;
5499         u64 objectid = 0;
5500
5501         if (!test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state))
5502                 WARN_ON(btrfs_root_refs(&root->root_item) != 0);
5503
5504         spin_lock(&root->inode_lock);
5505 again:
5506         node = root->inode_tree.rb_node;
5507         prev = NULL;
5508         while (node) {
5509                 prev = node;
5510                 entry = rb_entry(node, struct btrfs_inode, rb_node);
5511
5512                 if (objectid < btrfs_ino(BTRFS_I(&entry->vfs_inode)))
5513                         node = node->rb_left;
5514                 else if (objectid > btrfs_ino(BTRFS_I(&entry->vfs_inode)))
5515                         node = node->rb_right;
5516                 else
5517                         break;
5518         }
5519         if (!node) {
5520                 while (prev) {
5521                         entry = rb_entry(prev, struct btrfs_inode, rb_node);
5522                         if (objectid <= btrfs_ino(BTRFS_I(&entry->vfs_inode))) {
5523                                 node = prev;
5524                                 break;
5525                         }
5526                         prev = rb_next(prev);
5527                 }
5528         }
5529         while (node) {
5530                 entry = rb_entry(node, struct btrfs_inode, rb_node);
5531                 objectid = btrfs_ino(BTRFS_I(&entry->vfs_inode)) + 1;
5532                 inode = igrab(&entry->vfs_inode);
5533                 if (inode) {
5534                         spin_unlock(&root->inode_lock);
5535                         if (atomic_read(&inode->i_count) > 1)
5536                                 d_prune_aliases(inode);
5537                         /*
5538                          * btrfs_drop_inode will have it removed from
5539                          * the inode cache when its usage count
5540                          * hits zero.
5541                          */
5542                         iput(inode);
5543                         cond_resched();
5544                         spin_lock(&root->inode_lock);
5545                         goto again;
5546                 }
5547
5548                 if (cond_resched_lock(&root->inode_lock))
5549                         goto again;
5550
5551                 node = rb_next(node);
5552         }
5553         spin_unlock(&root->inode_lock);
5554 }
5555
5556 static int btrfs_init_locked_inode(struct inode *inode, void *p)
5557 {
5558         struct btrfs_iget_args *args = p;
5559         inode->i_ino = args->location->objectid;
5560         memcpy(&BTRFS_I(inode)->location, args->location,
5561                sizeof(*args->location));
5562         BTRFS_I(inode)->root = args->root;
5563         return 0;
5564 }
5565
5566 static int btrfs_find_actor(struct inode *inode, void *opaque)
5567 {
5568         struct btrfs_iget_args *args = opaque;
5569         return args->location->objectid == BTRFS_I(inode)->location.objectid &&
5570                 args->root == BTRFS_I(inode)->root;
5571 }
5572
5573 static struct inode *btrfs_iget_locked(struct super_block *s,
5574                                        struct btrfs_key *location,
5575                                        struct btrfs_root *root)
5576 {
5577         struct inode *inode;
5578         struct btrfs_iget_args args;
5579         unsigned long hashval = btrfs_inode_hash(location->objectid, root);
5580
5581         args.location = location;
5582         args.root = root;
5583
5584         inode = iget5_locked(s, hashval, btrfs_find_actor,
5585                              btrfs_init_locked_inode,
5586                              (void *)&args);
5587         return inode;
5588 }
5589
5590 /* Get an inode object given its location and corresponding root.
5591  * Returns in *is_new if the inode was read from disk
5592  */
5593 struct inode *btrfs_iget(struct super_block *s, struct btrfs_key *location,
5594                          struct btrfs_root *root, int *new)
5595 {
5596         struct inode *inode;
5597
5598         inode = btrfs_iget_locked(s, location, root);
5599         if (!inode)
5600                 return ERR_PTR(-ENOMEM);
5601
5602         if (inode->i_state & I_NEW) {
5603                 int ret;
5604
5605                 ret = btrfs_read_locked_inode(inode);
5606                 if (!is_bad_inode(inode)) {
5607                         inode_tree_add(inode);
5608                         unlock_new_inode(inode);
5609                         if (new)
5610                                 *new = 1;
5611                 } else {
5612                         unlock_new_inode(inode);
5613                         iput(inode);
5614                         ASSERT(ret < 0);
5615                         inode = ERR_PTR(ret < 0 ? ret : -ESTALE);
5616                 }
5617         }
5618
5619         return inode;
5620 }
5621
5622 static struct inode *new_simple_dir(struct super_block *s,
5623                                     struct btrfs_key *key,
5624                                     struct btrfs_root *root)
5625 {
5626         struct inode *inode = new_inode(s);
5627
5628         if (!inode)
5629                 return ERR_PTR(-ENOMEM);
5630
5631         BTRFS_I(inode)->root = root;
5632         memcpy(&BTRFS_I(inode)->location, key, sizeof(*key));
5633         set_bit(BTRFS_INODE_DUMMY, &BTRFS_I(inode)->runtime_flags);
5634
5635         inode->i_ino = BTRFS_EMPTY_SUBVOL_DIR_OBJECTID;
5636         inode->i_op = &btrfs_dir_ro_inode_operations;
5637         inode->i_opflags &= ~IOP_XATTR;
5638         inode->i_fop = &simple_dir_operations;
5639         inode->i_mode = S_IFDIR | S_IRUGO | S_IWUSR | S_IXUGO;
5640         inode->i_mtime = current_time(inode);
5641         inode->i_atime = inode->i_mtime;
5642         inode->i_ctime = inode->i_mtime;
5643         BTRFS_I(inode)->i_otime = inode->i_mtime;
5644
5645         return inode;
5646 }
5647
5648 struct inode *btrfs_lookup_dentry(struct inode *dir, struct dentry *dentry)
5649 {
5650         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
5651         struct inode *inode;
5652         struct btrfs_root *root = BTRFS_I(dir)->root;
5653         struct btrfs_root *sub_root = root;
5654         struct btrfs_key location;
5655         int index;
5656         int ret = 0;
5657
5658         if (dentry->d_name.len > BTRFS_NAME_LEN)
5659                 return ERR_PTR(-ENAMETOOLONG);
5660
5661         ret = btrfs_inode_by_name(dir, dentry, &location);
5662         if (ret < 0)
5663                 return ERR_PTR(ret);
5664
5665         if (location.objectid == 0)
5666                 return ERR_PTR(-ENOENT);
5667
5668         if (location.type == BTRFS_INODE_ITEM_KEY) {
5669                 inode = btrfs_iget(dir->i_sb, &location, root, NULL);
5670                 return inode;
5671         }
5672
5673         BUG_ON(location.type != BTRFS_ROOT_ITEM_KEY);
5674
5675         index = srcu_read_lock(&fs_info->subvol_srcu);
5676         ret = fixup_tree_root_location(fs_info, dir, dentry,
5677                                        &location, &sub_root);
5678         if (ret < 0) {
5679                 if (ret != -ENOENT)
5680                         inode = ERR_PTR(ret);
5681                 else
5682                         inode = new_simple_dir(dir->i_sb, &location, sub_root);
5683         } else {
5684                 inode = btrfs_iget(dir->i_sb, &location, sub_root, NULL);
5685         }
5686         srcu_read_unlock(&fs_info->subvol_srcu, index);
5687
5688         if (!IS_ERR(inode) && root != sub_root) {
5689                 down_read(&fs_info->cleanup_work_sem);
5690                 if (!(inode->i_sb->s_flags & MS_RDONLY))
5691                         ret = btrfs_orphan_cleanup(sub_root);
5692                 up_read(&fs_info->cleanup_work_sem);
5693                 if (ret) {
5694                         iput(inode);
5695                         inode = ERR_PTR(ret);
5696                 }
5697         }
5698
5699         return inode;
5700 }
5701
5702 static int btrfs_dentry_delete(const struct dentry *dentry)
5703 {
5704         struct btrfs_root *root;
5705         struct inode *inode = d_inode(dentry);
5706
5707         if (!inode && !IS_ROOT(dentry))
5708                 inode = d_inode(dentry->d_parent);
5709
5710         if (inode) {
5711                 root = BTRFS_I(inode)->root;
5712                 if (btrfs_root_refs(&root->root_item) == 0)
5713                         return 1;
5714
5715                 if (btrfs_ino(BTRFS_I(inode)) == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)
5716                         return 1;
5717         }
5718         return 0;
5719 }
5720
5721 static void btrfs_dentry_release(struct dentry *dentry)
5722 {
5723         kfree(dentry->d_fsdata);
5724 }
5725
5726 static struct dentry *btrfs_lookup(struct inode *dir, struct dentry *dentry,
5727                                    unsigned int flags)
5728 {
5729         struct inode *inode;
5730
5731         inode = btrfs_lookup_dentry(dir, dentry);
5732         if (IS_ERR(inode)) {
5733                 if (PTR_ERR(inode) == -ENOENT)
5734                         inode = NULL;
5735                 else
5736                         return ERR_CAST(inode);
5737         }
5738
5739         return d_splice_alias(inode, dentry);
5740 }
5741
5742 unsigned char btrfs_filetype_table[] = {
5743         DT_UNKNOWN, DT_REG, DT_DIR, DT_CHR, DT_BLK, DT_FIFO, DT_SOCK, DT_LNK
5744 };
5745
5746 static int btrfs_real_readdir(struct file *file, struct dir_context *ctx)
5747 {
5748         struct inode *inode = file_inode(file);
5749         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
5750         struct btrfs_root *root = BTRFS_I(inode)->root;
5751         struct btrfs_item *item;
5752         struct btrfs_dir_item *di;
5753         struct btrfs_key key;
5754         struct btrfs_key found_key;
5755         struct btrfs_path *path;
5756         struct list_head ins_list;
5757         struct list_head del_list;
5758         int ret;
5759         struct extent_buffer *leaf;
5760         int slot;
5761         unsigned char d_type;
5762         int over = 0;
5763         char tmp_name[32];
5764         char *name_ptr;
5765         int name_len;
5766         bool put = false;
5767         struct btrfs_key location;
5768
5769         if (!dir_emit_dots(file, ctx))
5770                 return 0;
5771
5772         path = btrfs_alloc_path();
5773         if (!path)
5774                 return -ENOMEM;
5775
5776         path->reada = READA_FORWARD;
5777
5778         INIT_LIST_HEAD(&ins_list);
5779         INIT_LIST_HEAD(&del_list);
5780         put = btrfs_readdir_get_delayed_items(inode, &ins_list, &del_list);
5781
5782         key.type = BTRFS_DIR_INDEX_KEY;
5783         key.offset = ctx->pos;
5784         key.objectid = btrfs_ino(BTRFS_I(inode));
5785
5786         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
5787         if (ret < 0)
5788                 goto err;
5789
5790         while (1) {
5791                 leaf = path->nodes[0];
5792                 slot = path->slots[0];
5793                 if (slot >= btrfs_header_nritems(leaf)) {
5794                         ret = btrfs_next_leaf(root, path);
5795                         if (ret < 0)
5796                                 goto err;
5797                         else if (ret > 0)
5798                                 break;
5799                         continue;
5800                 }
5801
5802                 item = btrfs_item_nr(slot);
5803                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
5804
5805                 if (found_key.objectid != key.objectid)
5806                         break;
5807                 if (found_key.type != BTRFS_DIR_INDEX_KEY)
5808                         break;
5809                 if (found_key.offset < ctx->pos)
5810                         goto next;
5811                 if (btrfs_should_delete_dir_index(&del_list, found_key.offset))
5812                         goto next;
5813
5814                 ctx->pos = found_key.offset;
5815
5816                 di = btrfs_item_ptr(leaf, slot, struct btrfs_dir_item);
5817                 if (verify_dir_item(fs_info, leaf, di))
5818                         goto next;
5819
5820                 name_len = btrfs_dir_name_len(leaf, di);
5821                 if (name_len <= sizeof(tmp_name)) {
5822                         name_ptr = tmp_name;
5823                 } else {
5824                         name_ptr = kmalloc(name_len, GFP_KERNEL);
5825                         if (!name_ptr) {
5826                                 ret = -ENOMEM;
5827                                 goto err;
5828                         }
5829                 }
5830                 read_extent_buffer(leaf, name_ptr, (unsigned long)(di + 1),
5831                                    name_len);
5832
5833                 d_type = btrfs_filetype_table[btrfs_dir_type(leaf, di)];
5834                 btrfs_dir_item_key_to_cpu(leaf, di, &location);
5835
5836                 over = !dir_emit(ctx, name_ptr, name_len, location.objectid,
5837                                  d_type);
5838
5839                 if (name_ptr != tmp_name)
5840                         kfree(name_ptr);
5841
5842                 if (over)
5843                         goto nopos;
5844                 ctx->pos++;
5845 next:
5846                 path->slots[0]++;
5847         }
5848
5849         ret = btrfs_readdir_delayed_dir_index(ctx, &ins_list);
5850         if (ret)
5851                 goto nopos;
5852
5853         /*
5854          * Stop new entries from being returned after we return the last
5855          * entry.
5856          *
5857          * New directory entries are assigned a strictly increasing
5858          * offset.  This means that new entries created during readdir
5859          * are *guaranteed* to be seen in the future by that readdir.
5860          * This has broken buggy programs which operate on names as
5861          * they're returned by readdir.  Until we re-use freed offsets
5862          * we have this hack to stop new entries from being returned
5863          * under the assumption that they'll never reach this huge
5864          * offset.
5865          *
5866          * This is being careful not to overflow 32bit loff_t unless the
5867          * last entry requires it because doing so has broken 32bit apps
5868          * in the past.
5869          */
5870         if (ctx->pos >= INT_MAX)
5871                 ctx->pos = LLONG_MAX;
5872         else
5873                 ctx->pos = INT_MAX;
5874 nopos:
5875         ret = 0;
5876 err:
5877         if (put)
5878                 btrfs_readdir_put_delayed_items(inode, &ins_list, &del_list);
5879         btrfs_free_path(path);
5880         return ret;
5881 }
5882
5883 int btrfs_write_inode(struct inode *inode, struct writeback_control *wbc)
5884 {
5885         struct btrfs_root *root = BTRFS_I(inode)->root;
5886         struct btrfs_trans_handle *trans;
5887         int ret = 0;
5888         bool nolock = false;
5889
5890         if (test_bit(BTRFS_INODE_DUMMY, &BTRFS_I(inode)->runtime_flags))
5891                 return 0;
5892
5893         if (btrfs_fs_closing(root->fs_info) &&
5894                         btrfs_is_free_space_inode(BTRFS_I(inode)))
5895                 nolock = true;
5896
5897         if (wbc->sync_mode == WB_SYNC_ALL) {
5898                 if (nolock)
5899                         trans = btrfs_join_transaction_nolock(root);
5900                 else
5901                         trans = btrfs_join_transaction(root);
5902                 if (IS_ERR(trans))
5903                         return PTR_ERR(trans);
5904                 ret = btrfs_commit_transaction(trans);
5905         }
5906         return ret;
5907 }
5908
5909 /*
5910  * This is somewhat expensive, updating the tree every time the
5911  * inode changes.  But, it is most likely to find the inode in cache.
5912  * FIXME, needs more benchmarking...there are no reasons other than performance
5913  * to keep or drop this code.
5914  */
5915 static int btrfs_dirty_inode(struct inode *inode)
5916 {
5917         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
5918         struct btrfs_root *root = BTRFS_I(inode)->root;
5919         struct btrfs_trans_handle *trans;
5920         int ret;
5921
5922         if (test_bit(BTRFS_INODE_DUMMY, &BTRFS_I(inode)->runtime_flags))
5923                 return 0;
5924
5925         trans = btrfs_join_transaction(root);
5926         if (IS_ERR(trans))
5927                 return PTR_ERR(trans);
5928
5929         ret = btrfs_update_inode(trans, root, inode);
5930         if (ret && ret == -ENOSPC) {
5931                 /* whoops, lets try again with the full transaction */
5932                 btrfs_end_transaction(trans);
5933                 trans = btrfs_start_transaction(root, 1);
5934                 if (IS_ERR(trans))
5935                         return PTR_ERR(trans);
5936
5937                 ret = btrfs_update_inode(trans, root, inode);
5938         }
5939         btrfs_end_transaction(trans);
5940         if (BTRFS_I(inode)->delayed_node)
5941                 btrfs_balance_delayed_items(fs_info);
5942
5943         return ret;
5944 }
5945
5946 /*
5947  * This is a copy of file_update_time.  We need this so we can return error on
5948  * ENOSPC for updating the inode in the case of file write and mmap writes.
5949  */
5950 static int btrfs_update_time(struct inode *inode, struct timespec *now,
5951                              int flags)
5952 {
5953         struct btrfs_root *root = BTRFS_I(inode)->root;
5954
5955         if (btrfs_root_readonly(root))
5956                 return -EROFS;
5957
5958         if (flags & S_VERSION)
5959                 inode_inc_iversion(inode);
5960         if (flags & S_CTIME)
5961                 inode->i_ctime = *now;
5962         if (flags & S_MTIME)
5963                 inode->i_mtime = *now;
5964         if (flags & S_ATIME)
5965                 inode->i_atime = *now;
5966         return btrfs_dirty_inode(inode);
5967 }
5968
5969 /*
5970  * find the highest existing sequence number in a directory
5971  * and then set the in-memory index_cnt variable to reflect
5972  * free sequence numbers
5973  */
5974 static int btrfs_set_inode_index_count(struct btrfs_inode *inode)
5975 {
5976         struct btrfs_root *root = inode->root;
5977         struct btrfs_key key, found_key;
5978         struct btrfs_path *path;
5979         struct extent_buffer *leaf;
5980         int ret;
5981
5982         key.objectid = btrfs_ino(inode);
5983         key.type = BTRFS_DIR_INDEX_KEY;
5984         key.offset = (u64)-1;
5985
5986         path = btrfs_alloc_path();
5987         if (!path)
5988                 return -ENOMEM;
5989
5990         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
5991         if (ret < 0)
5992                 goto out;
5993         /* FIXME: we should be able to handle this */
5994         if (ret == 0)
5995                 goto out;
5996         ret = 0;
5997
5998         /*
5999          * MAGIC NUMBER EXPLANATION:
6000          * since we search a directory based on f_pos we have to start at 2
6001          * since '.' and '..' have f_pos of 0 and 1 respectively, so everybody
6002          * else has to start at 2
6003          */
6004         if (path->slots[0] == 0) {
6005                 inode->index_cnt = 2;
6006                 goto out;
6007         }
6008
6009         path->slots[0]--;
6010
6011         leaf = path->nodes[0];
6012         btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
6013
6014         if (found_key.objectid != btrfs_ino(inode) ||
6015             found_key.type != BTRFS_DIR_INDEX_KEY) {
6016                 inode->index_cnt = 2;
6017                 goto out;
6018         }
6019
6020         inode->index_cnt = found_key.offset + 1;
6021 out:
6022         btrfs_free_path(path);
6023         return ret;
6024 }
6025
6026 /*
6027  * helper to find a free sequence number in a given directory.  This current
6028  * code is very simple, later versions will do smarter things in the btree
6029  */
6030 int btrfs_set_inode_index(struct btrfs_inode *dir, u64 *index)
6031 {
6032         int ret = 0;
6033
6034         if (dir->index_cnt == (u64)-1) {
6035                 ret = btrfs_inode_delayed_dir_index_count(dir);
6036                 if (ret) {
6037                         ret = btrfs_set_inode_index_count(dir);
6038                         if (ret)
6039                                 return ret;
6040                 }
6041         }
6042
6043         *index = dir->index_cnt;
6044         dir->index_cnt++;
6045
6046         return ret;
6047 }
6048
6049 static int btrfs_insert_inode_locked(struct inode *inode)
6050 {
6051         struct btrfs_iget_args args;
6052         args.location = &BTRFS_I(inode)->location;
6053         args.root = BTRFS_I(inode)->root;
6054
6055         return insert_inode_locked4(inode,
6056                    btrfs_inode_hash(inode->i_ino, BTRFS_I(inode)->root),
6057                    btrfs_find_actor, &args);
6058 }
6059
6060 static struct inode *btrfs_new_inode(struct btrfs_trans_handle *trans,
6061                                      struct btrfs_root *root,
6062                                      struct inode *dir,
6063                                      const char *name, int name_len,
6064                                      u64 ref_objectid, u64 objectid,
6065                                      umode_t mode, u64 *index)
6066 {
6067         struct btrfs_fs_info *fs_info = root->fs_info;
6068         struct inode *inode;
6069         struct btrfs_inode_item *inode_item;
6070         struct btrfs_key *location;
6071         struct btrfs_path *path;
6072         struct btrfs_inode_ref *ref;
6073         struct btrfs_key key[2];
6074         u32 sizes[2];
6075         int nitems = name ? 2 : 1;
6076         unsigned long ptr;
6077         int ret;
6078
6079         path = btrfs_alloc_path();
6080         if (!path)
6081                 return ERR_PTR(-ENOMEM);
6082
6083         inode = new_inode(fs_info->sb);
6084         if (!inode) {
6085                 btrfs_free_path(path);
6086                 return ERR_PTR(-ENOMEM);
6087         }
6088
6089         /*
6090          * O_TMPFILE, set link count to 0, so that after this point,
6091          * we fill in an inode item with the correct link count.
6092          */
6093         if (!name)
6094                 set_nlink(inode, 0);
6095
6096         /*
6097          * we have to initialize this early, so we can reclaim the inode
6098          * number if we fail afterwards in this function.
6099          */
6100         inode->i_ino = objectid;
6101
6102         if (dir && name) {
6103                 trace_btrfs_inode_request(dir);
6104
6105                 ret = btrfs_set_inode_index(BTRFS_I(dir), index);
6106                 if (ret) {
6107                         btrfs_free_path(path);
6108                         iput(inode);
6109                         return ERR_PTR(ret);
6110                 }
6111         } else if (dir) {
6112                 *index = 0;
6113         }
6114         /*
6115          * index_cnt is ignored for everything but a dir,
6116          * btrfs_get_inode_index_count has an explanation for the magic
6117          * number
6118          */
6119         BTRFS_I(inode)->index_cnt = 2;
6120         BTRFS_I(inode)->dir_index = *index;
6121         BTRFS_I(inode)->root = root;
6122         BTRFS_I(inode)->generation = trans->transid;
6123         inode->i_generation = BTRFS_I(inode)->generation;
6124
6125         /*
6126          * We could have gotten an inode number from somebody who was fsynced
6127          * and then removed in this same transaction, so let's just set full
6128          * sync since it will be a full sync anyway and this will blow away the
6129          * old info in the log.
6130          */
6131         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &BTRFS_I(inode)->runtime_flags);
6132
6133         key[0].objectid = objectid;
6134         key[0].type = BTRFS_INODE_ITEM_KEY;
6135         key[0].offset = 0;
6136
6137         sizes[0] = sizeof(struct btrfs_inode_item);
6138
6139         if (name) {
6140                 /*
6141                  * Start new inodes with an inode_ref. This is slightly more
6142                  * efficient for small numbers of hard links since they will
6143                  * be packed into one item. Extended refs will kick in if we
6144                  * add more hard links than can fit in the ref item.
6145                  */
6146                 key[1].objectid = objectid;
6147                 key[1].type = BTRFS_INODE_REF_KEY;
6148                 key[1].offset = ref_objectid;
6149
6150                 sizes[1] = name_len + sizeof(*ref);
6151         }
6152
6153         location = &BTRFS_I(inode)->location;
6154         location->objectid = objectid;
6155         location->offset = 0;
6156         location->type = BTRFS_INODE_ITEM_KEY;
6157
6158         ret = btrfs_insert_inode_locked(inode);
6159         if (ret < 0)
6160                 goto fail;
6161
6162         path->leave_spinning = 1;
6163         ret = btrfs_insert_empty_items(trans, root, path, key, sizes, nitems);
6164         if (ret != 0)
6165                 goto fail_unlock;
6166
6167         inode_init_owner(inode, dir, mode);
6168         inode_set_bytes(inode, 0);
6169
6170         inode->i_mtime = current_time(inode);
6171         inode->i_atime = inode->i_mtime;
6172         inode->i_ctime = inode->i_mtime;
6173         BTRFS_I(inode)->i_otime = inode->i_mtime;
6174
6175         inode_item = btrfs_item_ptr(path->nodes[0], path->slots[0],
6176                                   struct btrfs_inode_item);
6177         memzero_extent_buffer(path->nodes[0], (unsigned long)inode_item,
6178                              sizeof(*inode_item));
6179         fill_inode_item(trans, path->nodes[0], inode_item, inode);
6180
6181         if (name) {
6182                 ref = btrfs_item_ptr(path->nodes[0], path->slots[0] + 1,
6183                                      struct btrfs_inode_ref);
6184                 btrfs_set_inode_ref_name_len(path->nodes[0], ref, name_len);
6185                 btrfs_set_inode_ref_index(path->nodes[0], ref, *index);
6186                 ptr = (unsigned long)(ref + 1);
6187                 write_extent_buffer(path->nodes[0], name, ptr, name_len);
6188         }
6189
6190         btrfs_mark_buffer_dirty(path->nodes[0]);
6191         btrfs_free_path(path);
6192
6193         btrfs_inherit_iflags(inode, dir);
6194
6195         if (S_ISREG(mode)) {
6196                 if (btrfs_test_opt(fs_info, NODATASUM))
6197                         BTRFS_I(inode)->flags |= BTRFS_INODE_NODATASUM;
6198                 if (btrfs_test_opt(fs_info, NODATACOW))
6199                         BTRFS_I(inode)->flags |= BTRFS_INODE_NODATACOW |
6200                                 BTRFS_INODE_NODATASUM;
6201         }
6202
6203         inode_tree_add(inode);
6204
6205         trace_btrfs_inode_new(inode);
6206         btrfs_set_inode_last_trans(trans, inode);
6207
6208         btrfs_update_root_times(trans, root);
6209
6210         ret = btrfs_inode_inherit_props(trans, inode, dir);
6211         if (ret)
6212                 btrfs_err(fs_info,
6213                           "error inheriting props for ino %llu (root %llu): %d",
6214                         btrfs_ino(BTRFS_I(inode)), root->root_key.objectid, ret);
6215
6216         return inode;
6217
6218 fail_unlock:
6219         unlock_new_inode(inode);
6220 fail:
6221         if (dir && name)
6222                 BTRFS_I(dir)->index_cnt--;
6223         btrfs_free_path(path);
6224         iput(inode);
6225         return ERR_PTR(ret);
6226 }
6227
6228 static inline u8 btrfs_inode_type(struct inode *inode)
6229 {
6230         return btrfs_type_by_mode[(inode->i_mode & S_IFMT) >> S_SHIFT];
6231 }
6232
6233 /*
6234  * utility function to add 'inode' into 'parent_inode' with
6235  * a give name and a given sequence number.
6236  * if 'add_backref' is true, also insert a backref from the
6237  * inode to the parent directory.
6238  */
6239 int btrfs_add_link(struct btrfs_trans_handle *trans,
6240                    struct btrfs_inode *parent_inode, struct btrfs_inode *inode,
6241                    const char *name, int name_len, int add_backref, u64 index)
6242 {
6243         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
6244         int ret = 0;
6245         struct btrfs_key key;
6246         struct btrfs_root *root = parent_inode->root;
6247         u64 ino = btrfs_ino(inode);
6248         u64 parent_ino = btrfs_ino(parent_inode);
6249
6250         if (unlikely(ino == BTRFS_FIRST_FREE_OBJECTID)) {
6251                 memcpy(&key, &inode->root->root_key, sizeof(key));
6252         } else {
6253                 key.objectid = ino;
6254                 key.type = BTRFS_INODE_ITEM_KEY;
6255                 key.offset = 0;
6256         }
6257
6258         if (unlikely(ino == BTRFS_FIRST_FREE_OBJECTID)) {
6259                 ret = btrfs_add_root_ref(trans, fs_info, key.objectid,
6260                                          root->root_key.objectid, parent_ino,
6261                                          index, name, name_len);
6262         } else if (add_backref) {
6263                 ret = btrfs_insert_inode_ref(trans, root, name, name_len, ino,
6264                                              parent_ino, index);
6265         }
6266
6267         /* Nothing to clean up yet */
6268         if (ret)
6269                 return ret;
6270
6271         ret = btrfs_insert_dir_item(trans, root, name, name_len,
6272                                     parent_inode, &key,
6273                                     btrfs_inode_type(&inode->vfs_inode), index);
6274         if (ret == -EEXIST || ret == -EOVERFLOW)
6275                 goto fail_dir_item;
6276         else if (ret) {
6277                 btrfs_abort_transaction(trans, ret);
6278                 return ret;
6279         }
6280
6281         btrfs_i_size_write(parent_inode, parent_inode->vfs_inode.i_size +
6282                            name_len * 2);
6283         inode_inc_iversion(&parent_inode->vfs_inode);
6284         parent_inode->vfs_inode.i_mtime = parent_inode->vfs_inode.i_ctime =
6285                 current_time(&parent_inode->vfs_inode);
6286         ret = btrfs_update_inode(trans, root, &parent_inode->vfs_inode);
6287         if (ret)
6288                 btrfs_abort_transaction(trans, ret);
6289         return ret;
6290
6291 fail_dir_item:
6292         if (unlikely(ino == BTRFS_FIRST_FREE_OBJECTID)) {
6293                 u64 local_index;
6294                 int err;
6295                 err = btrfs_del_root_ref(trans, fs_info, key.objectid,
6296                                          root->root_key.objectid, parent_ino,
6297                                          &local_index, name, name_len);
6298
6299         } else if (add_backref) {
6300                 u64 local_index;
6301                 int err;
6302
6303                 err = btrfs_del_inode_ref(trans, root, name, name_len,
6304                                           ino, parent_ino, &local_index);
6305         }
6306         return ret;
6307 }
6308
6309 static int btrfs_add_nondir(struct btrfs_trans_handle *trans,
6310                             struct btrfs_inode *dir, struct dentry *dentry,
6311                             struct btrfs_inode *inode, int backref, u64 index)
6312 {
6313         int err = btrfs_add_link(trans, dir, inode,
6314                                  dentry->d_name.name, dentry->d_name.len,
6315                                  backref, index);
6316         if (err > 0)
6317                 err = -EEXIST;
6318         return err;
6319 }
6320
6321 static int btrfs_mknod(struct inode *dir, struct dentry *dentry,
6322                         umode_t mode, dev_t rdev)
6323 {
6324         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
6325         struct btrfs_trans_handle *trans;
6326         struct btrfs_root *root = BTRFS_I(dir)->root;
6327         struct inode *inode = NULL;
6328         int err;
6329         int drop_inode = 0;
6330         u64 objectid;
6331         u64 index = 0;
6332
6333         /*
6334          * 2 for inode item and ref
6335          * 2 for dir items
6336          * 1 for xattr if selinux is on
6337          */
6338         trans = btrfs_start_transaction(root, 5);
6339         if (IS_ERR(trans))
6340                 return PTR_ERR(trans);
6341
6342         err = btrfs_find_free_ino(root, &objectid);
6343         if (err)
6344                 goto out_unlock;
6345
6346         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
6347                         dentry->d_name.len, btrfs_ino(BTRFS_I(dir)), objectid,
6348                         mode, &index);
6349         if (IS_ERR(inode)) {
6350                 err = PTR_ERR(inode);
6351                 goto out_unlock;
6352         }
6353
6354         /*
6355         * If the active LSM wants to access the inode during
6356         * d_instantiate it needs these. Smack checks to see
6357         * if the filesystem supports xattrs by looking at the
6358         * ops vector.
6359         */
6360         inode->i_op = &btrfs_special_inode_operations;
6361         init_special_inode(inode, inode->i_mode, rdev);
6362
6363         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
6364         if (err)
6365                 goto out_unlock_inode;
6366
6367         err = btrfs_add_nondir(trans, BTRFS_I(dir), dentry, BTRFS_I(inode),
6368                         0, index);
6369         if (err) {
6370                 goto out_unlock_inode;
6371         } else {
6372                 btrfs_update_inode(trans, root, inode);
6373                 unlock_new_inode(inode);
6374                 d_instantiate(dentry, inode);
6375         }
6376
6377 out_unlock:
6378         btrfs_end_transaction(trans);
6379         btrfs_balance_delayed_items(fs_info);
6380         btrfs_btree_balance_dirty(fs_info);
6381         if (drop_inode) {
6382                 inode_dec_link_count(inode);
6383                 iput(inode);
6384         }
6385         return err;
6386
6387 out_unlock_inode:
6388         drop_inode = 1;
6389         unlock_new_inode(inode);
6390         goto out_unlock;
6391
6392 }
6393
6394 static int btrfs_create(struct inode *dir, struct dentry *dentry,
6395                         umode_t mode, bool excl)
6396 {
6397         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
6398         struct btrfs_trans_handle *trans;
6399         struct btrfs_root *root = BTRFS_I(dir)->root;
6400         struct inode *inode = NULL;
6401         int drop_inode_on_err = 0;
6402         int err;
6403         u64 objectid;
6404         u64 index = 0;
6405
6406         /*
6407          * 2 for inode item and ref
6408          * 2 for dir items
6409          * 1 for xattr if selinux is on
6410          */
6411         trans = btrfs_start_transaction(root, 5);
6412         if (IS_ERR(trans))
6413                 return PTR_ERR(trans);
6414
6415         err = btrfs_find_free_ino(root, &objectid);
6416         if (err)
6417                 goto out_unlock;
6418
6419         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
6420                         dentry->d_name.len, btrfs_ino(BTRFS_I(dir)), objectid,
6421                         mode, &index);
6422         if (IS_ERR(inode)) {
6423                 err = PTR_ERR(inode);
6424                 goto out_unlock;
6425         }
6426         drop_inode_on_err = 1;
6427         /*
6428         * If the active LSM wants to access the inode during
6429         * d_instantiate it needs these. Smack checks to see
6430         * if the filesystem supports xattrs by looking at the
6431         * ops vector.
6432         */
6433         inode->i_fop = &btrfs_file_operations;
6434         inode->i_op = &btrfs_file_inode_operations;
6435         inode->i_mapping->a_ops = &btrfs_aops;
6436
6437         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
6438         if (err)
6439                 goto out_unlock_inode;
6440
6441         err = btrfs_update_inode(trans, root, inode);
6442         if (err)
6443                 goto out_unlock_inode;
6444
6445         err = btrfs_add_nondir(trans, BTRFS_I(dir), dentry, BTRFS_I(inode),
6446                         0, index);
6447         if (err)
6448                 goto out_unlock_inode;
6449
6450         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
6451         unlock_new_inode(inode);
6452         d_instantiate(dentry, inode);
6453
6454 out_unlock:
6455         btrfs_end_transaction(trans);
6456         if (err && drop_inode_on_err) {
6457                 inode_dec_link_count(inode);
6458                 iput(inode);
6459         }
6460         btrfs_balance_delayed_items(fs_info);
6461         btrfs_btree_balance_dirty(fs_info);
6462         return err;
6463
6464 out_unlock_inode:
6465         unlock_new_inode(inode);
6466         goto out_unlock;
6467
6468 }
6469
6470 static int btrfs_link(struct dentry *old_dentry, struct inode *dir,
6471                       struct dentry *dentry)
6472 {
6473         struct btrfs_trans_handle *trans = NULL;
6474         struct btrfs_root *root = BTRFS_I(dir)->root;
6475         struct inode *inode = d_inode(old_dentry);
6476         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
6477         u64 index;
6478         int err;
6479         int drop_inode = 0;
6480
6481         /* do not allow sys_link's with other subvols of the same device */
6482         if (root->objectid != BTRFS_I(inode)->root->objectid)
6483                 return -EXDEV;
6484
6485         if (inode->i_nlink >= BTRFS_LINK_MAX)
6486                 return -EMLINK;
6487
6488         err = btrfs_set_inode_index(BTRFS_I(dir), &index);
6489         if (err)
6490                 goto fail;
6491
6492         /*
6493          * 2 items for inode and inode ref
6494          * 2 items for dir items
6495          * 1 item for parent inode
6496          */
6497         trans = btrfs_start_transaction(root, 5);
6498         if (IS_ERR(trans)) {
6499                 err = PTR_ERR(trans);
6500                 trans = NULL;
6501                 goto fail;
6502         }
6503
6504         /* There are several dir indexes for this inode, clear the cache. */
6505         BTRFS_I(inode)->dir_index = 0ULL;
6506         inc_nlink(inode);
6507         inode_inc_iversion(inode);
6508         inode->i_ctime = current_time(inode);
6509         ihold(inode);
6510         set_bit(BTRFS_INODE_COPY_EVERYTHING, &BTRFS_I(inode)->runtime_flags);
6511
6512         err = btrfs_add_nondir(trans, BTRFS_I(dir), dentry, BTRFS_I(inode),
6513                         1, index);
6514
6515         if (err) {
6516                 drop_inode = 1;
6517         } else {
6518                 struct dentry *parent = dentry->d_parent;
6519                 err = btrfs_update_inode(trans, root, inode);
6520                 if (err)
6521                         goto fail;
6522                 if (inode->i_nlink == 1) {
6523                         /*
6524                          * If new hard link count is 1, it's a file created
6525                          * with open(2) O_TMPFILE flag.
6526                          */
6527                         err = btrfs_orphan_del(trans, BTRFS_I(inode));
6528                         if (err)
6529                                 goto fail;
6530                 }
6531                 d_instantiate(dentry, inode);
6532                 btrfs_log_new_name(trans, BTRFS_I(inode), NULL, parent);
6533         }
6534
6535         btrfs_balance_delayed_items(fs_info);
6536 fail:
6537         if (trans)
6538                 btrfs_end_transaction(trans);
6539         if (drop_inode) {
6540                 inode_dec_link_count(inode);
6541                 iput(inode);
6542         }
6543         btrfs_btree_balance_dirty(fs_info);
6544         return err;
6545 }
6546
6547 static int btrfs_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode)
6548 {
6549         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
6550         struct inode *inode = NULL;
6551         struct btrfs_trans_handle *trans;
6552         struct btrfs_root *root = BTRFS_I(dir)->root;
6553         int err = 0;
6554         int drop_on_err = 0;
6555         u64 objectid = 0;
6556         u64 index = 0;
6557
6558         /*
6559          * 2 items for inode and ref
6560          * 2 items for dir items
6561          * 1 for xattr if selinux is on
6562          */
6563         trans = btrfs_start_transaction(root, 5);
6564         if (IS_ERR(trans))
6565                 return PTR_ERR(trans);
6566
6567         err = btrfs_find_free_ino(root, &objectid);
6568         if (err)
6569                 goto out_fail;
6570
6571         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
6572                         dentry->d_name.len, btrfs_ino(BTRFS_I(dir)), objectid,
6573                         S_IFDIR | mode, &index);
6574         if (IS_ERR(inode)) {
6575                 err = PTR_ERR(inode);
6576                 goto out_fail;
6577         }
6578
6579         drop_on_err = 1;
6580         /* these must be set before we unlock the inode */
6581         inode->i_op = &btrfs_dir_inode_operations;
6582         inode->i_fop = &btrfs_dir_file_operations;
6583
6584         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
6585         if (err)
6586                 goto out_fail_inode;
6587
6588         btrfs_i_size_write(BTRFS_I(inode), 0);
6589         err = btrfs_update_inode(trans, root, inode);
6590         if (err)
6591                 goto out_fail_inode;
6592
6593         err = btrfs_add_link(trans, BTRFS_I(dir), BTRFS_I(inode),
6594                         dentry->d_name.name,
6595                         dentry->d_name.len, 0, index);
6596         if (err)
6597                 goto out_fail_inode;
6598
6599         d_instantiate(dentry, inode);
6600         /*
6601          * mkdir is special.  We're unlocking after we call d_instantiate
6602          * to avoid a race with nfsd calling d_instantiate.
6603          */
6604         unlock_new_inode(inode);
6605         drop_on_err = 0;
6606
6607 out_fail:
6608         btrfs_end_transaction(trans);
6609         if (drop_on_err) {
6610                 inode_dec_link_count(inode);
6611                 iput(inode);
6612         }
6613         btrfs_balance_delayed_items(fs_info);
6614         btrfs_btree_balance_dirty(fs_info);
6615         return err;
6616
6617 out_fail_inode:
6618         unlock_new_inode(inode);
6619         goto out_fail;
6620 }
6621
6622 /* Find next extent map of a given extent map, caller needs to ensure locks */
6623 static struct extent_map *next_extent_map(struct extent_map *em)
6624 {
6625         struct rb_node *next;
6626
6627         next = rb_next(&em->rb_node);
6628         if (!next)
6629                 return NULL;
6630         return container_of(next, struct extent_map, rb_node);
6631 }
6632
6633 static struct extent_map *prev_extent_map(struct extent_map *em)
6634 {
6635         struct rb_node *prev;
6636
6637         prev = rb_prev(&em->rb_node);
6638         if (!prev)
6639                 return NULL;
6640         return container_of(prev, struct extent_map, rb_node);
6641 }
6642
6643 /* helper for btfs_get_extent.  Given an existing extent in the tree,
6644  * the existing extent is the nearest extent to map_start,
6645  * and an extent that you want to insert, deal with overlap and insert
6646  * the best fitted new extent into the tree.
6647  */
6648 static int merge_extent_mapping(struct extent_map_tree *em_tree,
6649                                 struct extent_map *existing,
6650                                 struct extent_map *em,
6651                                 u64 map_start)
6652 {
6653         struct extent_map *prev;
6654         struct extent_map *next;
6655         u64 start;
6656         u64 end;
6657         u64 start_diff;
6658
6659         BUG_ON(map_start < em->start || map_start >= extent_map_end(em));
6660
6661         if (existing->start > map_start) {
6662                 next = existing;
6663                 prev = prev_extent_map(next);
6664         } else {
6665                 prev = existing;
6666                 next = next_extent_map(prev);
6667         }
6668
6669         start = prev ? extent_map_end(prev) : em->start;
6670         start = max_t(u64, start, em->start);
6671         end = next ? next->start : extent_map_end(em);
6672         end = min_t(u64, end, extent_map_end(em));
6673         start_diff = start - em->start;
6674         em->start = start;
6675         em->len = end - start;
6676         if (em->block_start < EXTENT_MAP_LAST_BYTE &&
6677             !test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
6678                 em->block_start += start_diff;
6679                 em->block_len -= start_diff;
6680         }
6681         return add_extent_mapping(em_tree, em, 0);
6682 }
6683
6684 static noinline int uncompress_inline(struct btrfs_path *path,
6685                                       struct page *page,
6686                                       size_t pg_offset, u64 extent_offset,
6687                                       struct btrfs_file_extent_item *item)
6688 {
6689         int ret;
6690         struct extent_buffer *leaf = path->nodes[0];
6691         char *tmp;
6692         size_t max_size;
6693         unsigned long inline_size;
6694         unsigned long ptr;
6695         int compress_type;
6696
6697         WARN_ON(pg_offset != 0);
6698         compress_type = btrfs_file_extent_compression(leaf, item);
6699         max_size = btrfs_file_extent_ram_bytes(leaf, item);
6700         inline_size = btrfs_file_extent_inline_item_len(leaf,
6701                                         btrfs_item_nr(path->slots[0]));
6702         tmp = kmalloc(inline_size, GFP_NOFS);
6703         if (!tmp)
6704                 return -ENOMEM;
6705         ptr = btrfs_file_extent_inline_start(item);
6706
6707         read_extent_buffer(leaf, tmp, ptr, inline_size);
6708
6709         max_size = min_t(unsigned long, PAGE_SIZE, max_size);
6710         ret = btrfs_decompress(compress_type, tmp, page,
6711                                extent_offset, inline_size, max_size);
6712
6713         /*
6714          * decompression code contains a memset to fill in any space between the end
6715          * of the uncompressed data and the end of max_size in case the decompressed
6716          * data ends up shorter than ram_bytes.  That doesn't cover the hole between
6717          * the end of an inline extent and the beginning of the next block, so we
6718          * cover that region here.
6719          */
6720
6721         if (max_size + pg_offset < PAGE_SIZE) {
6722                 char *map = kmap(page);
6723                 memset(map + pg_offset + max_size, 0, PAGE_SIZE - max_size - pg_offset);
6724                 kunmap(page);
6725         }
6726         kfree(tmp);
6727         return ret;
6728 }
6729
6730 /*
6731  * a bit scary, this does extent mapping from logical file offset to the disk.
6732  * the ugly parts come from merging extents from the disk with the in-ram
6733  * representation.  This gets more complex because of the data=ordered code,
6734  * where the in-ram extents might be locked pending data=ordered completion.
6735  *
6736  * This also copies inline extents directly into the page.
6737  */
6738
6739 struct extent_map *btrfs_get_extent(struct btrfs_inode *inode,
6740                 struct page *page,
6741             size_t pg_offset, u64 start, u64 len,
6742                 int create)
6743 {
6744         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
6745         int ret;
6746         int err = 0;
6747         u64 extent_start = 0;
6748         u64 extent_end = 0;
6749         u64 objectid = btrfs_ino(inode);
6750         u32 found_type;
6751         struct btrfs_path *path = NULL;
6752         struct btrfs_root *root = inode->root;
6753         struct btrfs_file_extent_item *item;
6754         struct extent_buffer *leaf;
6755         struct btrfs_key found_key;
6756         struct extent_map *em = NULL;
6757         struct extent_map_tree *em_tree = &inode->extent_tree;
6758         struct extent_io_tree *io_tree = &inode->io_tree;
6759         struct btrfs_trans_handle *trans = NULL;
6760         const bool new_inline = !page || create;
6761
6762 again:
6763         read_lock(&em_tree->lock);
6764         em = lookup_extent_mapping(em_tree, start, len);
6765         if (em)
6766                 em->bdev = fs_info->fs_devices->latest_bdev;
6767         read_unlock(&em_tree->lock);
6768
6769         if (em) {
6770                 if (em->start > start || em->start + em->len <= start)
6771                         free_extent_map(em);
6772                 else if (em->block_start == EXTENT_MAP_INLINE && page)
6773                         free_extent_map(em);
6774                 else
6775                         goto out;
6776         }
6777         em = alloc_extent_map();
6778         if (!em) {
6779                 err = -ENOMEM;
6780                 goto out;
6781         }
6782         em->bdev = fs_info->fs_devices->latest_bdev;
6783         em->start = EXTENT_MAP_HOLE;
6784         em->orig_start = EXTENT_MAP_HOLE;
6785         em->len = (u64)-1;
6786         em->block_len = (u64)-1;
6787
6788         if (!path) {
6789                 path = btrfs_alloc_path();
6790                 if (!path) {
6791                         err = -ENOMEM;
6792                         goto out;
6793                 }
6794                 /*
6795                  * Chances are we'll be called again, so go ahead and do
6796                  * readahead
6797                  */
6798                 path->reada = READA_FORWARD;
6799         }
6800
6801         ret = btrfs_lookup_file_extent(trans, root, path,
6802                                        objectid, start, trans != NULL);
6803         if (ret < 0) {
6804                 err = ret;
6805                 goto out;
6806         }
6807
6808         if (ret != 0) {
6809                 if (path->slots[0] == 0)
6810                         goto not_found;
6811                 path->slots[0]--;
6812         }
6813
6814         leaf = path->nodes[0];
6815         item = btrfs_item_ptr(leaf, path->slots[0],
6816                               struct btrfs_file_extent_item);
6817         /* are we inside the extent that was found? */
6818         btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
6819         found_type = found_key.type;
6820         if (found_key.objectid != objectid ||
6821             found_type != BTRFS_EXTENT_DATA_KEY) {
6822                 /*
6823                  * If we backup past the first extent we want to move forward
6824                  * and see if there is an extent in front of us, otherwise we'll
6825                  * say there is a hole for our whole search range which can
6826                  * cause problems.
6827                  */
6828                 extent_end = start;
6829                 goto next;
6830         }
6831
6832         found_type = btrfs_file_extent_type(leaf, item);
6833         extent_start = found_key.offset;
6834         if (found_type == BTRFS_FILE_EXTENT_REG ||
6835             found_type == BTRFS_FILE_EXTENT_PREALLOC) {
6836                 extent_end = extent_start +
6837                        btrfs_file_extent_num_bytes(leaf, item);
6838         } else if (found_type == BTRFS_FILE_EXTENT_INLINE) {
6839                 size_t size;
6840                 size = btrfs_file_extent_inline_len(leaf, path->slots[0], item);
6841                 extent_end = ALIGN(extent_start + size,
6842                                    fs_info->sectorsize);
6843         }
6844 next:
6845         if (start >= extent_end) {
6846                 path->slots[0]++;
6847                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
6848                         ret = btrfs_next_leaf(root, path);
6849                         if (ret < 0) {
6850                                 err = ret;
6851                                 goto out;
6852                         }
6853                         if (ret > 0)
6854                                 goto not_found;
6855                         leaf = path->nodes[0];
6856                 }
6857                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
6858                 if (found_key.objectid != objectid ||
6859                     found_key.type != BTRFS_EXTENT_DATA_KEY)
6860                         goto not_found;
6861                 if (start + len <= found_key.offset)
6862                         goto not_found;
6863                 if (start > found_key.offset)
6864                         goto next;
6865                 em->start = start;
6866                 em->orig_start = start;
6867                 em->len = found_key.offset - start;
6868                 goto not_found_em;
6869         }
6870
6871         btrfs_extent_item_to_extent_map(inode, path, item,
6872                         new_inline, em);
6873
6874         if (found_type == BTRFS_FILE_EXTENT_REG ||
6875             found_type == BTRFS_FILE_EXTENT_PREALLOC) {
6876                 goto insert;
6877         } else if (found_type == BTRFS_FILE_EXTENT_INLINE) {
6878                 unsigned long ptr;
6879                 char *map;
6880                 size_t size;
6881                 size_t extent_offset;
6882                 size_t copy_size;
6883
6884                 if (new_inline)
6885                         goto out;
6886
6887                 size = btrfs_file_extent_inline_len(leaf, path->slots[0], item);
6888                 extent_offset = page_offset(page) + pg_offset - extent_start;
6889                 copy_size = min_t(u64, PAGE_SIZE - pg_offset,
6890                                   size - extent_offset);
6891                 em->start = extent_start + extent_offset;
6892                 em->len = ALIGN(copy_size, fs_info->sectorsize);
6893                 em->orig_block_len = em->len;
6894                 em->orig_start = em->start;
6895                 ptr = btrfs_file_extent_inline_start(item) + extent_offset;
6896                 if (create == 0 && !PageUptodate(page)) {
6897                         if (btrfs_file_extent_compression(leaf, item) !=
6898                             BTRFS_COMPRESS_NONE) {
6899                                 ret = uncompress_inline(path, page, pg_offset,
6900                                                         extent_offset, item);
6901                                 if (ret) {
6902                                         err = ret;
6903                                         goto out;
6904                                 }
6905                         } else {
6906                                 map = kmap(page);
6907                                 read_extent_buffer(leaf, map + pg_offset, ptr,
6908                                                    copy_size);
6909                                 if (pg_offset + copy_size < PAGE_SIZE) {
6910                                         memset(map + pg_offset + copy_size, 0,
6911                                                PAGE_SIZE - pg_offset -
6912                                                copy_size);
6913                                 }
6914                                 kunmap(page);
6915                         }
6916                         flush_dcache_page(page);
6917                 } else if (create && PageUptodate(page)) {
6918                         BUG();
6919                         if (!trans) {
6920                                 kunmap(page);
6921                                 free_extent_map(em);
6922                                 em = NULL;
6923
6924                                 btrfs_release_path(path);
6925                                 trans = btrfs_join_transaction(root);
6926
6927                                 if (IS_ERR(trans))
6928                                         return ERR_CAST(trans);
6929                                 goto again;
6930                         }
6931                         map = kmap(page);
6932                         write_extent_buffer(leaf, map + pg_offset, ptr,
6933                                             copy_size);
6934                         kunmap(page);
6935                         btrfs_mark_buffer_dirty(leaf);
6936                 }
6937                 set_extent_uptodate(io_tree, em->start,
6938                                     extent_map_end(em) - 1, NULL, GFP_NOFS);
6939                 goto insert;
6940         }
6941 not_found:
6942         em->start = start;
6943         em->orig_start = start;
6944         em->len = len;
6945 not_found_em:
6946         em->block_start = EXTENT_MAP_HOLE;
6947         set_bit(EXTENT_FLAG_VACANCY, &em->flags);
6948 insert:
6949         btrfs_release_path(path);
6950         if (em->start > start || extent_map_end(em) <= start) {
6951                 btrfs_err(fs_info,
6952                           "bad extent! em: [%llu %llu] passed [%llu %llu]",
6953                           em->start, em->len, start, len);
6954                 err = -EIO;
6955                 goto out;
6956         }
6957
6958         err = 0;
6959         write_lock(&em_tree->lock);
6960         ret = add_extent_mapping(em_tree, em, 0);
6961         /* it is possible that someone inserted the extent into the tree
6962          * while we had the lock dropped.  It is also possible that
6963          * an overlapping map exists in the tree
6964          */
6965         if (ret == -EEXIST) {
6966                 struct extent_map *existing;
6967
6968                 ret = 0;
6969
6970                 existing = search_extent_mapping(em_tree, start, len);
6971                 /*
6972                  * existing will always be non-NULL, since there must be
6973                  * extent causing the -EEXIST.
6974                  */
6975                 if (existing->start == em->start &&
6976                     extent_map_end(existing) >= extent_map_end(em) &&
6977                     em->block_start == existing->block_start) {
6978                         /*
6979                          * The existing extent map already encompasses the
6980                          * entire extent map we tried to add.
6981                          */
6982                         free_extent_map(em);
6983                         em = existing;
6984                         err = 0;
6985
6986                 } else if (start >= extent_map_end(existing) ||
6987                     start <= existing->start) {
6988                         /*
6989                          * The existing extent map is the one nearest to
6990                          * the [start, start + len) range which overlaps
6991                          */
6992                         err = merge_extent_mapping(em_tree, existing,
6993                                                    em, start);
6994                         free_extent_map(existing);
6995                         if (err) {
6996                                 free_extent_map(em);
6997                                 em = NULL;
6998                         }
6999                 } else {
7000                         free_extent_map(em);
7001                         em = existing;
7002                         err = 0;
7003                 }
7004         }
7005         write_unlock(&em_tree->lock);
7006 out:
7007
7008         trace_btrfs_get_extent(root, inode, em);
7009
7010         btrfs_free_path(path);
7011         if (trans) {
7012                 ret = btrfs_end_transaction(trans);
7013                 if (!err)
7014                         err = ret;
7015         }
7016         if (err) {
7017                 free_extent_map(em);
7018                 return ERR_PTR(err);
7019         }
7020         BUG_ON(!em); /* Error is always set */
7021         return em;
7022 }
7023
7024 struct extent_map *btrfs_get_extent_fiemap(struct btrfs_inode *inode,
7025                 struct page *page,
7026                 size_t pg_offset, u64 start, u64 len,
7027                 int create)
7028 {
7029         struct extent_map *em;
7030         struct extent_map *hole_em = NULL;
7031         u64 range_start = start;
7032         u64 end;
7033         u64 found;
7034         u64 found_end;
7035         int err = 0;
7036
7037         em = btrfs_get_extent(inode, page, pg_offset, start, len, create);
7038         if (IS_ERR(em))
7039                 return em;
7040         if (em) {
7041                 /*
7042                  * if our em maps to
7043                  * -  a hole or
7044                  * -  a pre-alloc extent,
7045                  * there might actually be delalloc bytes behind it.
7046                  */
7047                 if (em->block_start != EXTENT_MAP_HOLE &&
7048                     !test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
7049                         return em;
7050                 else
7051                         hole_em = em;
7052         }
7053
7054         /* check to see if we've wrapped (len == -1 or similar) */
7055         end = start + len;
7056         if (end < start)
7057                 end = (u64)-1;
7058         else
7059                 end -= 1;
7060
7061         em = NULL;
7062
7063         /* ok, we didn't find anything, lets look for delalloc */
7064         found = count_range_bits(&inode->io_tree, &range_start,
7065                                  end, len, EXTENT_DELALLOC, 1);
7066         found_end = range_start + found;
7067         if (found_end < range_start)
7068                 found_end = (u64)-1;
7069
7070         /*
7071          * we didn't find anything useful, return
7072          * the original results from get_extent()
7073          */
7074         if (range_start > end || found_end <= start) {
7075                 em = hole_em;
7076                 hole_em = NULL;
7077                 goto out;
7078         }
7079
7080         /* adjust the range_start to make sure it doesn't
7081          * go backwards from the start they passed in
7082          */
7083         range_start = max(start, range_start);
7084         found = found_end - range_start;
7085
7086         if (found > 0) {
7087                 u64 hole_start = start;
7088                 u64 hole_len = len;
7089
7090                 em = alloc_extent_map();
7091                 if (!em) {
7092                         err = -ENOMEM;
7093                         goto out;
7094                 }
7095                 /*
7096                  * when btrfs_get_extent can't find anything it
7097                  * returns one huge hole
7098                  *
7099                  * make sure what it found really fits our range, and
7100                  * adjust to make sure it is based on the start from
7101                  * the caller
7102                  */
7103                 if (hole_em) {
7104                         u64 calc_end = extent_map_end(hole_em);
7105
7106                         if (calc_end <= start || (hole_em->start > end)) {
7107                                 free_extent_map(hole_em);
7108                                 hole_em = NULL;
7109                         } else {
7110                                 hole_start = max(hole_em->start, start);
7111                                 hole_len = calc_end - hole_start;
7112                         }
7113                 }
7114                 em->bdev = NULL;
7115                 if (hole_em && range_start > hole_start) {
7116                         /* our hole starts before our delalloc, so we
7117                          * have to return just the parts of the hole
7118                          * that go until  the delalloc starts
7119                          */
7120                         em->len = min(hole_len,
7121                                       range_start - hole_start);
7122                         em->start = hole_start;
7123                         em->orig_start = hole_start;
7124                         /*
7125                          * don't adjust block start at all,
7126                          * it is fixed at EXTENT_MAP_HOLE
7127                          */
7128                         em->block_start = hole_em->block_start;
7129                         em->block_len = hole_len;
7130                         if (test_bit(EXTENT_FLAG_PREALLOC, &hole_em->flags))
7131                                 set_bit(EXTENT_FLAG_PREALLOC, &em->flags);
7132                 } else {
7133                         em->start = range_start;
7134                         em->len = found;
7135                         em->orig_start = range_start;
7136                         em->block_start = EXTENT_MAP_DELALLOC;
7137                         em->block_len = found;
7138                 }
7139         } else if (hole_em) {
7140                 return hole_em;
7141         }
7142 out:
7143
7144         free_extent_map(hole_em);
7145         if (err) {
7146                 free_extent_map(em);
7147                 return ERR_PTR(err);
7148         }
7149         return em;
7150 }
7151
7152 static struct extent_map *btrfs_create_dio_extent(struct inode *inode,
7153                                                   const u64 start,
7154                                                   const u64 len,
7155                                                   const u64 orig_start,
7156                                                   const u64 block_start,
7157                                                   const u64 block_len,
7158                                                   const u64 orig_block_len,
7159                                                   const u64 ram_bytes,
7160                                                   const int type)
7161 {
7162         struct extent_map *em = NULL;
7163         int ret;
7164
7165         if (type != BTRFS_ORDERED_NOCOW) {
7166                 em = create_io_em(inode, start, len, orig_start,
7167                                   block_start, block_len, orig_block_len,
7168                                   ram_bytes,
7169                                   BTRFS_COMPRESS_NONE, /* compress_type */
7170                                   type);
7171                 if (IS_ERR(em))
7172                         goto out;
7173         }
7174         ret = btrfs_add_ordered_extent_dio(inode, start, block_start,
7175                                            len, block_len, type);
7176         if (ret) {
7177                 if (em) {
7178                         free_extent_map(em);
7179                         btrfs_drop_extent_cache(BTRFS_I(inode), start,
7180                                                 start + len - 1, 0);
7181                 }
7182                 em = ERR_PTR(ret);
7183         }
7184  out:
7185
7186         return em;
7187 }
7188
7189 static struct extent_map *btrfs_new_extent_direct(struct inode *inode,
7190                                                   u64 start, u64 len)
7191 {
7192         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7193         struct btrfs_root *root = BTRFS_I(inode)->root;
7194         struct extent_map *em;
7195         struct btrfs_key ins;
7196         u64 alloc_hint;
7197         int ret;
7198
7199         alloc_hint = get_extent_allocation_hint(inode, start, len);
7200         ret = btrfs_reserve_extent(root, len, len, fs_info->sectorsize,
7201                                    0, alloc_hint, &ins, 1, 1);
7202         if (ret)
7203                 return ERR_PTR(ret);
7204
7205         em = btrfs_create_dio_extent(inode, start, ins.offset, start,
7206                                      ins.objectid, ins.offset, ins.offset,
7207                                      ins.offset, BTRFS_ORDERED_REGULAR);
7208         btrfs_dec_block_group_reservations(fs_info, ins.objectid);
7209         if (IS_ERR(em))
7210                 btrfs_free_reserved_extent(fs_info, ins.objectid,
7211                                            ins.offset, 1);
7212
7213         return em;
7214 }
7215
7216 /*
7217  * returns 1 when the nocow is safe, < 1 on error, 0 if the
7218  * block must be cow'd
7219  */
7220 noinline int can_nocow_extent(struct inode *inode, u64 offset, u64 *len,
7221                               u64 *orig_start, u64 *orig_block_len,
7222                               u64 *ram_bytes)
7223 {
7224         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7225         struct btrfs_path *path;
7226         int ret;
7227         struct extent_buffer *leaf;
7228         struct btrfs_root *root = BTRFS_I(inode)->root;
7229         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
7230         struct btrfs_file_extent_item *fi;
7231         struct btrfs_key key;
7232         u64 disk_bytenr;
7233         u64 backref_offset;
7234         u64 extent_end;
7235         u64 num_bytes;
7236         int slot;
7237         int found_type;
7238         bool nocow = (BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW);
7239
7240         path = btrfs_alloc_path();
7241         if (!path)
7242                 return -ENOMEM;
7243
7244         ret = btrfs_lookup_file_extent(NULL, root, path,
7245                         btrfs_ino(BTRFS_I(inode)), offset, 0);
7246         if (ret < 0)
7247                 goto out;
7248
7249         slot = path->slots[0];
7250         if (ret == 1) {
7251                 if (slot == 0) {
7252                         /* can't find the item, must cow */
7253                         ret = 0;
7254                         goto out;
7255                 }
7256                 slot--;
7257         }
7258         ret = 0;
7259         leaf = path->nodes[0];
7260         btrfs_item_key_to_cpu(leaf, &key, slot);
7261         if (key.objectid != btrfs_ino(BTRFS_I(inode)) ||
7262             key.type != BTRFS_EXTENT_DATA_KEY) {
7263                 /* not our file or wrong item type, must cow */
7264                 goto out;
7265         }
7266
7267         if (key.offset > offset) {
7268                 /* Wrong offset, must cow */
7269                 goto out;
7270         }
7271
7272         fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
7273         found_type = btrfs_file_extent_type(leaf, fi);
7274         if (found_type != BTRFS_FILE_EXTENT_REG &&
7275             found_type != BTRFS_FILE_EXTENT_PREALLOC) {
7276                 /* not a regular extent, must cow */
7277                 goto out;
7278         }
7279
7280         if (!nocow && found_type == BTRFS_FILE_EXTENT_REG)
7281                 goto out;
7282
7283         extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
7284         if (extent_end <= offset)
7285                 goto out;
7286
7287         disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
7288         if (disk_bytenr == 0)
7289                 goto out;
7290
7291         if (btrfs_file_extent_compression(leaf, fi) ||
7292             btrfs_file_extent_encryption(leaf, fi) ||
7293             btrfs_file_extent_other_encoding(leaf, fi))
7294                 goto out;
7295
7296         backref_offset = btrfs_file_extent_offset(leaf, fi);
7297
7298         if (orig_start) {
7299                 *orig_start = key.offset - backref_offset;
7300                 *orig_block_len = btrfs_file_extent_disk_num_bytes(leaf, fi);
7301                 *ram_bytes = btrfs_file_extent_ram_bytes(leaf, fi);
7302         }
7303
7304         if (btrfs_extent_readonly(fs_info, disk_bytenr))
7305                 goto out;
7306
7307         num_bytes = min(offset + *len, extent_end) - offset;
7308         if (!nocow && found_type == BTRFS_FILE_EXTENT_PREALLOC) {
7309                 u64 range_end;
7310
7311                 range_end = round_up(offset + num_bytes,
7312                                      root->fs_info->sectorsize) - 1;
7313                 ret = test_range_bit(io_tree, offset, range_end,
7314                                      EXTENT_DELALLOC, 0, NULL);
7315                 if (ret) {
7316                         ret = -EAGAIN;
7317                         goto out;
7318                 }
7319         }
7320
7321         btrfs_release_path(path);
7322
7323         /*
7324          * look for other files referencing this extent, if we
7325          * find any we must cow
7326          */
7327
7328         ret = btrfs_cross_ref_exist(root, btrfs_ino(BTRFS_I(inode)),
7329                                     key.offset - backref_offset, disk_bytenr);
7330         if (ret) {
7331                 ret = 0;
7332                 goto out;
7333         }
7334
7335         /*
7336          * adjust disk_bytenr and num_bytes to cover just the bytes
7337          * in this extent we are about to write.  If there
7338          * are any csums in that range we have to cow in order
7339          * to keep the csums correct
7340          */
7341         disk_bytenr += backref_offset;
7342         disk_bytenr += offset - key.offset;
7343         if (csum_exist_in_range(fs_info, disk_bytenr, num_bytes))
7344                 goto out;
7345         /*
7346          * all of the above have passed, it is safe to overwrite this extent
7347          * without cow
7348          */
7349         *len = num_bytes;
7350         ret = 1;
7351 out:
7352         btrfs_free_path(path);
7353         return ret;
7354 }
7355
7356 bool btrfs_page_exists_in_range(struct inode *inode, loff_t start, loff_t end)
7357 {
7358         struct radix_tree_root *root = &inode->i_mapping->page_tree;
7359         int found = false;
7360         void **pagep = NULL;
7361         struct page *page = NULL;
7362         int start_idx;
7363         int end_idx;
7364
7365         start_idx = start >> PAGE_SHIFT;
7366
7367         /*
7368          * end is the last byte in the last page.  end == start is legal
7369          */
7370         end_idx = end >> PAGE_SHIFT;
7371
7372         rcu_read_lock();
7373
7374         /* Most of the code in this while loop is lifted from
7375          * find_get_page.  It's been modified to begin searching from a
7376          * page and return just the first page found in that range.  If the
7377          * found idx is less than or equal to the end idx then we know that
7378          * a page exists.  If no pages are found or if those pages are
7379          * outside of the range then we're fine (yay!) */
7380         while (page == NULL &&
7381                radix_tree_gang_lookup_slot(root, &pagep, NULL, start_idx, 1)) {
7382                 page = radix_tree_deref_slot(pagep);
7383                 if (unlikely(!page))
7384                         break;
7385
7386                 if (radix_tree_exception(page)) {
7387                         if (radix_tree_deref_retry(page)) {
7388                                 page = NULL;
7389                                 continue;
7390                         }
7391                         /*
7392                          * Otherwise, shmem/tmpfs must be storing a swap entry
7393                          * here as an exceptional entry: so return it without
7394                          * attempting to raise page count.
7395                          */
7396                         page = NULL;
7397                         break; /* TODO: Is this relevant for this use case? */
7398                 }
7399
7400                 if (!page_cache_get_speculative(page)) {
7401                         page = NULL;
7402                         continue;
7403                 }
7404
7405                 /*
7406                  * Has the page moved?
7407                  * This is part of the lockless pagecache protocol. See
7408                  * include/linux/pagemap.h for details.
7409                  */
7410                 if (unlikely(page != *pagep)) {
7411                         put_page(page);
7412                         page = NULL;
7413                 }
7414         }
7415
7416         if (page) {
7417                 if (page->index <= end_idx)
7418                         found = true;
7419                 put_page(page);
7420         }
7421
7422         rcu_read_unlock();
7423         return found;
7424 }
7425
7426 static int lock_extent_direct(struct inode *inode, u64 lockstart, u64 lockend,
7427                               struct extent_state **cached_state, int writing)
7428 {
7429         struct btrfs_ordered_extent *ordered;
7430         int ret = 0;
7431
7432         while (1) {
7433                 lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend,
7434                                  cached_state);
7435                 /*
7436                  * We're concerned with the entire range that we're going to be
7437                  * doing DIO to, so we need to make sure there's no ordered
7438                  * extents in this range.
7439                  */
7440                 ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), lockstart,
7441                                                      lockend - lockstart + 1);
7442
7443                 /*
7444                  * We need to make sure there are no buffered pages in this
7445                  * range either, we could have raced between the invalidate in
7446                  * generic_file_direct_write and locking the extent.  The
7447                  * invalidate needs to happen so that reads after a write do not
7448                  * get stale data.
7449                  */
7450                 if (!ordered &&
7451                     (!writing ||
7452                      !btrfs_page_exists_in_range(inode, lockstart, lockend)))
7453                         break;
7454
7455                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
7456                                      cached_state, GFP_NOFS);
7457
7458                 if (ordered) {
7459                         /*
7460                          * If we are doing a DIO read and the ordered extent we
7461                          * found is for a buffered write, we can not wait for it
7462                          * to complete and retry, because if we do so we can
7463                          * deadlock with concurrent buffered writes on page
7464                          * locks. This happens only if our DIO read covers more
7465                          * than one extent map, if at this point has already
7466                          * created an ordered extent for a previous extent map
7467                          * and locked its range in the inode's io tree, and a
7468                          * concurrent write against that previous extent map's
7469                          * range and this range started (we unlock the ranges
7470                          * in the io tree only when the bios complete and
7471                          * buffered writes always lock pages before attempting
7472                          * to lock range in the io tree).
7473                          */
7474                         if (writing ||
7475                             test_bit(BTRFS_ORDERED_DIRECT, &ordered->flags))
7476                                 btrfs_start_ordered_extent(inode, ordered, 1);
7477                         else
7478                                 ret = -ENOTBLK;
7479                         btrfs_put_ordered_extent(ordered);
7480                 } else {
7481                         /*
7482                          * We could trigger writeback for this range (and wait
7483                          * for it to complete) and then invalidate the pages for
7484                          * this range (through invalidate_inode_pages2_range()),
7485                          * but that can lead us to a deadlock with a concurrent
7486                          * call to readpages() (a buffered read or a defrag call
7487                          * triggered a readahead) on a page lock due to an
7488                          * ordered dio extent we created before but did not have
7489                          * yet a corresponding bio submitted (whence it can not
7490                          * complete), which makes readpages() wait for that
7491                          * ordered extent to complete while holding a lock on
7492                          * that page.
7493                          */
7494                         ret = -ENOTBLK;
7495                 }
7496
7497                 if (ret)
7498                         break;
7499
7500                 cond_resched();
7501         }
7502
7503         return ret;
7504 }
7505
7506 /* The callers of this must take lock_extent() */
7507 static struct extent_map *create_io_em(struct inode *inode, u64 start, u64 len,
7508                                        u64 orig_start, u64 block_start,
7509                                        u64 block_len, u64 orig_block_len,
7510                                        u64 ram_bytes, int compress_type,
7511                                        int type)
7512 {
7513         struct extent_map_tree *em_tree;
7514         struct extent_map *em;
7515         struct btrfs_root *root = BTRFS_I(inode)->root;
7516         int ret;
7517
7518         ASSERT(type == BTRFS_ORDERED_PREALLOC ||
7519                type == BTRFS_ORDERED_COMPRESSED ||
7520                type == BTRFS_ORDERED_NOCOW ||
7521                type == BTRFS_ORDERED_REGULAR);
7522
7523         em_tree = &BTRFS_I(inode)->extent_tree;
7524         em = alloc_extent_map();
7525         if (!em)
7526                 return ERR_PTR(-ENOMEM);
7527
7528         em->start = start;
7529         em->orig_start = orig_start;
7530         em->len = len;
7531         em->block_len = block_len;
7532         em->block_start = block_start;
7533         em->bdev = root->fs_info->fs_devices->latest_bdev;
7534         em->orig_block_len = orig_block_len;
7535         em->ram_bytes = ram_bytes;
7536         em->generation = -1;
7537         set_bit(EXTENT_FLAG_PINNED, &em->flags);
7538         if (type == BTRFS_ORDERED_PREALLOC) {
7539                 set_bit(EXTENT_FLAG_FILLING, &em->flags);
7540         } else if (type == BTRFS_ORDERED_COMPRESSED) {
7541                 set_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
7542                 em->compress_type = compress_type;
7543         }
7544
7545         do {
7546                 btrfs_drop_extent_cache(BTRFS_I(inode), em->start,
7547                                 em->start + em->len - 1, 0);
7548                 write_lock(&em_tree->lock);
7549                 ret = add_extent_mapping(em_tree, em, 1);
7550                 write_unlock(&em_tree->lock);
7551                 /*
7552                  * The caller has taken lock_extent(), who could race with us
7553                  * to add em?
7554                  */
7555         } while (ret == -EEXIST);
7556
7557         if (ret) {
7558                 free_extent_map(em);
7559                 return ERR_PTR(ret);
7560         }
7561
7562         /* em got 2 refs now, callers needs to do free_extent_map once. */
7563         return em;
7564 }
7565
7566 static void adjust_dio_outstanding_extents(struct inode *inode,
7567                                            struct btrfs_dio_data *dio_data,
7568                                            const u64 len)
7569 {
7570         unsigned num_extents = count_max_extents(len);
7571
7572         /*
7573          * If we have an outstanding_extents count still set then we're
7574          * within our reservation, otherwise we need to adjust our inode
7575          * counter appropriately.
7576          */
7577         if (dio_data->outstanding_extents >= num_extents) {
7578                 dio_data->outstanding_extents -= num_extents;
7579         } else {
7580                 /*
7581                  * If dio write length has been split due to no large enough
7582                  * contiguous space, we need to compensate our inode counter
7583                  * appropriately.
7584                  */
7585                 u64 num_needed = num_extents - dio_data->outstanding_extents;
7586
7587                 spin_lock(&BTRFS_I(inode)->lock);
7588                 BTRFS_I(inode)->outstanding_extents += num_needed;
7589                 spin_unlock(&BTRFS_I(inode)->lock);
7590         }
7591 }
7592
7593 static int btrfs_get_blocks_direct(struct inode *inode, sector_t iblock,
7594                                    struct buffer_head *bh_result, int create)
7595 {
7596         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7597         struct extent_map *em;
7598         struct extent_state *cached_state = NULL;
7599         struct btrfs_dio_data *dio_data = NULL;
7600         u64 start = iblock << inode->i_blkbits;
7601         u64 lockstart, lockend;
7602         u64 len = bh_result->b_size;
7603         int unlock_bits = EXTENT_LOCKED;
7604         int ret = 0;
7605
7606         if (create)
7607                 unlock_bits |= EXTENT_DIRTY;
7608         else
7609                 len = min_t(u64, len, fs_info->sectorsize);
7610
7611         lockstart = start;
7612         lockend = start + len - 1;
7613
7614         if (current->journal_info) {
7615                 /*
7616                  * Need to pull our outstanding extents and set journal_info to NULL so
7617                  * that anything that needs to check if there's a transaction doesn't get
7618                  * confused.
7619                  */
7620                 dio_data = current->journal_info;
7621                 current->journal_info = NULL;
7622         }
7623
7624         /*
7625          * If this errors out it's because we couldn't invalidate pagecache for
7626          * this range and we need to fallback to buffered.
7627          */
7628         if (lock_extent_direct(inode, lockstart, lockend, &cached_state,
7629                                create)) {
7630                 ret = -ENOTBLK;
7631                 goto err;
7632         }
7633
7634         em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, start, len, 0);
7635         if (IS_ERR(em)) {
7636                 ret = PTR_ERR(em);
7637                 goto unlock_err;
7638         }
7639
7640         /*
7641          * Ok for INLINE and COMPRESSED extents we need to fallback on buffered
7642          * io.  INLINE is special, and we could probably kludge it in here, but
7643          * it's still buffered so for safety lets just fall back to the generic
7644          * buffered path.
7645          *
7646          * For COMPRESSED we _have_ to read the entire extent in so we can
7647          * decompress it, so there will be buffering required no matter what we
7648          * do, so go ahead and fallback to buffered.
7649          *
7650          * We return -ENOTBLK because that's what makes DIO go ahead and go back
7651          * to buffered IO.  Don't blame me, this is the price we pay for using
7652          * the generic code.
7653          */
7654         if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags) ||
7655             em->block_start == EXTENT_MAP_INLINE) {
7656                 free_extent_map(em);
7657                 ret = -ENOTBLK;
7658                 goto unlock_err;
7659         }
7660
7661         /* Just a good old fashioned hole, return */
7662         if (!create && (em->block_start == EXTENT_MAP_HOLE ||
7663                         test_bit(EXTENT_FLAG_PREALLOC, &em->flags))) {
7664                 free_extent_map(em);
7665                 goto unlock_err;
7666         }
7667
7668         /*
7669          * We don't allocate a new extent in the following cases
7670          *
7671          * 1) The inode is marked as NODATACOW.  In this case we'll just use the
7672          * existing extent.
7673          * 2) The extent is marked as PREALLOC.  We're good to go here and can
7674          * just use the extent.
7675          *
7676          */
7677         if (!create) {
7678                 len = min(len, em->len - (start - em->start));
7679                 lockstart = start + len;
7680                 goto unlock;
7681         }
7682
7683         if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags) ||
7684             ((BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW) &&
7685              em->block_start != EXTENT_MAP_HOLE)) {
7686                 int type;
7687                 u64 block_start, orig_start, orig_block_len, ram_bytes;
7688
7689                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
7690                         type = BTRFS_ORDERED_PREALLOC;
7691                 else
7692                         type = BTRFS_ORDERED_NOCOW;
7693                 len = min(len, em->len - (start - em->start));
7694                 block_start = em->block_start + (start - em->start);
7695
7696                 if (can_nocow_extent(inode, start, &len, &orig_start,
7697                                      &orig_block_len, &ram_bytes) == 1 &&
7698                     btrfs_inc_nocow_writers(fs_info, block_start)) {
7699                         struct extent_map *em2;
7700
7701                         em2 = btrfs_create_dio_extent(inode, start, len,
7702                                                       orig_start, block_start,
7703                                                       len, orig_block_len,
7704                                                       ram_bytes, type);
7705                         btrfs_dec_nocow_writers(fs_info, block_start);
7706                         if (type == BTRFS_ORDERED_PREALLOC) {
7707                                 free_extent_map(em);
7708                                 em = em2;
7709                         }
7710                         if (em2 && IS_ERR(em2)) {
7711                                 ret = PTR_ERR(em2);
7712                                 goto unlock_err;
7713                         }
7714                         /*
7715                          * For inode marked NODATACOW or extent marked PREALLOC,
7716                          * use the existing or preallocated extent, so does not
7717                          * need to adjust btrfs_space_info's bytes_may_use.
7718                          */
7719                         btrfs_free_reserved_data_space_noquota(inode,
7720                                         start, len);
7721                         goto unlock;
7722                 }
7723         }
7724
7725         /*
7726          * this will cow the extent, reset the len in case we changed
7727          * it above
7728          */
7729         len = bh_result->b_size;
7730         free_extent_map(em);
7731         em = btrfs_new_extent_direct(inode, start, len);
7732         if (IS_ERR(em)) {
7733                 ret = PTR_ERR(em);
7734                 goto unlock_err;
7735         }
7736         len = min(len, em->len - (start - em->start));
7737 unlock:
7738         bh_result->b_blocknr = (em->block_start + (start - em->start)) >>
7739                 inode->i_blkbits;
7740         bh_result->b_size = len;
7741         bh_result->b_bdev = em->bdev;
7742         set_buffer_mapped(bh_result);
7743         if (create) {
7744                 if (!test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
7745                         set_buffer_new(bh_result);
7746
7747                 /*
7748                  * Need to update the i_size under the extent lock so buffered
7749                  * readers will get the updated i_size when we unlock.
7750                  */
7751                 if (!dio_data->overwrite && start + len > i_size_read(inode))
7752                         i_size_write(inode, start + len);
7753
7754                 adjust_dio_outstanding_extents(inode, dio_data, len);
7755                 WARN_ON(dio_data->reserve < len);
7756                 dio_data->reserve -= len;
7757                 dio_data->unsubmitted_oe_range_end = start + len;
7758                 current->journal_info = dio_data;
7759         }
7760
7761         /*
7762          * In the case of write we need to clear and unlock the entire range,
7763          * in the case of read we need to unlock only the end area that we
7764          * aren't using if there is any left over space.
7765          */
7766         if (lockstart < lockend) {
7767                 clear_extent_bit(&BTRFS_I(inode)->io_tree, lockstart,
7768                                  lockend, unlock_bits, 1, 0,
7769                                  &cached_state, GFP_NOFS);
7770         } else {
7771                 free_extent_state(cached_state);
7772         }
7773
7774         free_extent_map(em);
7775
7776         return 0;
7777
7778 unlock_err:
7779         clear_extent_bit(&BTRFS_I(inode)->io_tree, lockstart, lockend,
7780                          unlock_bits, 1, 0, &cached_state, GFP_NOFS);
7781 err:
7782         if (dio_data)
7783                 current->journal_info = dio_data;
7784         /*
7785          * Compensate the delalloc release we do in btrfs_direct_IO() when we
7786          * write less data then expected, so that we don't underflow our inode's
7787          * outstanding extents counter.
7788          */
7789         if (create && dio_data)
7790                 adjust_dio_outstanding_extents(inode, dio_data, len);
7791
7792         return ret;
7793 }
7794
7795 static inline int submit_dio_repair_bio(struct inode *inode, struct bio *bio,
7796                                         int mirror_num)
7797 {
7798         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7799         int ret;
7800
7801         BUG_ON(bio_op(bio) == REQ_OP_WRITE);
7802
7803         bio_get(bio);
7804
7805         ret = btrfs_bio_wq_end_io(fs_info, bio, BTRFS_WQ_ENDIO_DIO_REPAIR);
7806         if (ret)
7807                 goto err;
7808
7809         ret = btrfs_map_bio(fs_info, bio, mirror_num, 0);
7810 err:
7811         bio_put(bio);
7812         return ret;
7813 }
7814
7815 static int btrfs_check_dio_repairable(struct inode *inode,
7816                                       struct bio *failed_bio,
7817                                       struct io_failure_record *failrec,
7818                                       int failed_mirror)
7819 {
7820         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7821         int num_copies;
7822
7823         num_copies = btrfs_num_copies(fs_info, failrec->logical, failrec->len);
7824         if (num_copies == 1) {
7825                 /*
7826                  * we only have a single copy of the data, so don't bother with
7827                  * all the retry and error correction code that follows. no
7828                  * matter what the error is, it is very likely to persist.
7829                  */
7830                 btrfs_debug(fs_info,
7831                         "Check DIO Repairable: cannot repair, num_copies=%d, next_mirror %d, failed_mirror %d",
7832                         num_copies, failrec->this_mirror, failed_mirror);
7833                 return 0;
7834         }
7835
7836         failrec->failed_mirror = failed_mirror;
7837         failrec->this_mirror++;
7838         if (failrec->this_mirror == failed_mirror)
7839                 failrec->this_mirror++;
7840
7841         if (failrec->this_mirror > num_copies) {
7842                 btrfs_debug(fs_info,
7843                         "Check DIO Repairable: (fail) num_copies=%d, next_mirror %d, failed_mirror %d",
7844                         num_copies, failrec->this_mirror, failed_mirror);
7845                 return 0;
7846         }
7847
7848         return 1;
7849 }
7850
7851 static int dio_read_error(struct inode *inode, struct bio *failed_bio,
7852                         struct page *page, unsigned int pgoff,
7853                         u64 start, u64 end, int failed_mirror,
7854                         bio_end_io_t *repair_endio, void *repair_arg)
7855 {
7856         struct io_failure_record *failrec;
7857         struct bio *bio;
7858         int isector;
7859         int read_mode = 0;
7860         int ret;
7861
7862         BUG_ON(bio_op(failed_bio) == REQ_OP_WRITE);
7863
7864         ret = btrfs_get_io_failure_record(inode, start, end, &failrec);
7865         if (ret)
7866                 return ret;
7867
7868         ret = btrfs_check_dio_repairable(inode, failed_bio, failrec,
7869                                          failed_mirror);
7870         if (!ret) {
7871                 free_io_failure(BTRFS_I(inode), failrec);
7872                 return -EIO;
7873         }
7874
7875         if ((failed_bio->bi_vcnt > 1)
7876                 || (failed_bio->bi_io_vec->bv_len
7877                         > btrfs_inode_sectorsize(inode)))
7878                 read_mode |= REQ_FAILFAST_DEV;
7879
7880         isector = start - btrfs_io_bio(failed_bio)->logical;
7881         isector >>= inode->i_sb->s_blocksize_bits;
7882         bio = btrfs_create_repair_bio(inode, failed_bio, failrec, page,
7883                                 pgoff, isector, repair_endio, repair_arg);
7884         if (!bio) {
7885                 free_io_failure(BTRFS_I(inode), failrec);
7886                 return -EIO;
7887         }
7888         bio_set_op_attrs(bio, REQ_OP_READ, read_mode);
7889
7890         btrfs_debug(BTRFS_I(inode)->root->fs_info,
7891                     "Repair DIO Read Error: submitting new dio read[%#x] to this_mirror=%d, in_validation=%d\n",
7892                     read_mode, failrec->this_mirror, failrec->in_validation);
7893
7894         ret = submit_dio_repair_bio(inode, bio, failrec->this_mirror);
7895         if (ret) {
7896                 free_io_failure(BTRFS_I(inode), failrec);
7897                 bio_put(bio);
7898         }
7899
7900         return ret;
7901 }
7902
7903 struct btrfs_retry_complete {
7904         struct completion done;
7905         struct inode *inode;
7906         u64 start;
7907         int uptodate;
7908 };
7909
7910 static void btrfs_retry_endio_nocsum(struct bio *bio)
7911 {
7912         struct btrfs_retry_complete *done = bio->bi_private;
7913         struct bio_vec *bvec;
7914         int i;
7915
7916         if (bio->bi_error)
7917                 goto end;
7918
7919         ASSERT(bio->bi_vcnt == 1);
7920         ASSERT(bio->bi_io_vec->bv_len == btrfs_inode_sectorsize(done->inode));
7921
7922         done->uptodate = 1;
7923         bio_for_each_segment_all(bvec, bio, i)
7924                 clean_io_failure(BTRFS_I(done->inode), done->start,
7925                                  bvec->bv_page, 0);
7926 end:
7927         complete(&done->done);
7928         bio_put(bio);
7929 }
7930
7931 static int __btrfs_correct_data_nocsum(struct inode *inode,
7932                                        struct btrfs_io_bio *io_bio)
7933 {
7934         struct btrfs_fs_info *fs_info;
7935         struct bio_vec *bvec;
7936         struct btrfs_retry_complete done;
7937         u64 start;
7938         unsigned int pgoff;
7939         u32 sectorsize;
7940         int nr_sectors;
7941         int i;
7942         int ret;
7943
7944         fs_info = BTRFS_I(inode)->root->fs_info;
7945         sectorsize = fs_info->sectorsize;
7946
7947         start = io_bio->logical;
7948         done.inode = inode;
7949
7950         bio_for_each_segment_all(bvec, &io_bio->bio, i) {
7951                 nr_sectors = BTRFS_BYTES_TO_BLKS(fs_info, bvec->bv_len);
7952                 pgoff = bvec->bv_offset;
7953
7954 next_block_or_try_again:
7955                 done.uptodate = 0;
7956                 done.start = start;
7957                 init_completion(&done.done);
7958
7959                 ret = dio_read_error(inode, &io_bio->bio, bvec->bv_page,
7960                                 pgoff, start, start + sectorsize - 1,
7961                                 io_bio->mirror_num,
7962                                 btrfs_retry_endio_nocsum, &done);
7963                 if (ret)
7964                         return ret;
7965
7966                 wait_for_completion(&done.done);
7967
7968                 if (!done.uptodate) {
7969                         /* We might have another mirror, so try again */
7970                         goto next_block_or_try_again;
7971                 }
7972
7973                 start += sectorsize;
7974
7975                 nr_sectors--;
7976                 if (nr_sectors) {
7977                         pgoff += sectorsize;
7978                         ASSERT(pgoff < PAGE_SIZE);
7979                         goto next_block_or_try_again;
7980                 }
7981         }
7982
7983         return 0;
7984 }
7985
7986 static void btrfs_retry_endio(struct bio *bio)
7987 {
7988         struct btrfs_retry_complete *done = bio->bi_private;
7989         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
7990         struct bio_vec *bvec;
7991         int uptodate;
7992         int ret;
7993         int i;
7994
7995         if (bio->bi_error)
7996                 goto end;
7997
7998         uptodate = 1;
7999
8000         ASSERT(bio->bi_vcnt == 1);
8001         ASSERT(bio->bi_io_vec->bv_len == btrfs_inode_sectorsize(done->inode));
8002
8003         bio_for_each_segment_all(bvec, bio, i) {
8004                 ret = __readpage_endio_check(done->inode, io_bio, i,
8005                                         bvec->bv_page, bvec->bv_offset,
8006                                         done->start, bvec->bv_len);
8007                 if (!ret)
8008                         clean_io_failure(BTRFS_I(done->inode), done->start,
8009                                         bvec->bv_page, bvec->bv_offset);
8010                 else
8011                         uptodate = 0;
8012         }
8013
8014         done->uptodate = uptodate;
8015 end:
8016         complete(&done->done);
8017         bio_put(bio);
8018 }
8019
8020 static int __btrfs_subio_endio_read(struct inode *inode,
8021                                     struct btrfs_io_bio *io_bio, int err)
8022 {
8023         struct btrfs_fs_info *fs_info;
8024         struct bio_vec *bvec;
8025         struct btrfs_retry_complete done;
8026         u64 start;
8027         u64 offset = 0;
8028         u32 sectorsize;
8029         int nr_sectors;
8030         unsigned int pgoff;
8031         int csum_pos;
8032         int i;
8033         int ret;
8034
8035         fs_info = BTRFS_I(inode)->root->fs_info;
8036         sectorsize = fs_info->sectorsize;
8037
8038         err = 0;
8039         start = io_bio->logical;
8040         done.inode = inode;
8041
8042         bio_for_each_segment_all(bvec, &io_bio->bio, i) {
8043                 nr_sectors = BTRFS_BYTES_TO_BLKS(fs_info, bvec->bv_len);
8044
8045                 pgoff = bvec->bv_offset;
8046 next_block:
8047                 csum_pos = BTRFS_BYTES_TO_BLKS(fs_info, offset);
8048                 ret = __readpage_endio_check(inode, io_bio, csum_pos,
8049                                         bvec->bv_page, pgoff, start,
8050                                         sectorsize);
8051                 if (likely(!ret))
8052                         goto next;
8053 try_again:
8054                 done.uptodate = 0;
8055                 done.start = start;
8056                 init_completion(&done.done);
8057
8058                 ret = dio_read_error(inode, &io_bio->bio, bvec->bv_page,
8059                                 pgoff, start, start + sectorsize - 1,
8060                                 io_bio->mirror_num,
8061                                 btrfs_retry_endio, &done);
8062                 if (ret) {
8063                         err = ret;
8064                         goto next;
8065                 }
8066
8067                 wait_for_completion(&done.done);
8068
8069                 if (!done.uptodate) {
8070                         /* We might have another mirror, so try again */
8071                         goto try_again;
8072                 }
8073 next:
8074                 offset += sectorsize;
8075                 start += sectorsize;
8076
8077                 ASSERT(nr_sectors);
8078
8079                 nr_sectors--;
8080                 if (nr_sectors) {
8081                         pgoff += sectorsize;
8082                         ASSERT(pgoff < PAGE_SIZE);
8083                         goto next_block;
8084                 }
8085         }
8086
8087         return err;
8088 }
8089
8090 static int btrfs_subio_endio_read(struct inode *inode,
8091                                   struct btrfs_io_bio *io_bio, int err)
8092 {
8093         bool skip_csum = BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM;
8094
8095         if (skip_csum) {
8096                 if (unlikely(err))
8097                         return __btrfs_correct_data_nocsum(inode, io_bio);
8098                 else
8099                         return 0;
8100         } else {
8101                 return __btrfs_subio_endio_read(inode, io_bio, err);
8102         }
8103 }
8104
8105 static void btrfs_endio_direct_read(struct bio *bio)
8106 {
8107         struct btrfs_dio_private *dip = bio->bi_private;
8108         struct inode *inode = dip->inode;
8109         struct bio *dio_bio;
8110         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
8111         int err = bio->bi_error;
8112
8113         if (dip->flags & BTRFS_DIO_ORIG_BIO_SUBMITTED)
8114                 err = btrfs_subio_endio_read(inode, io_bio, err);
8115
8116         unlock_extent(&BTRFS_I(inode)->io_tree, dip->logical_offset,
8117                       dip->logical_offset + dip->bytes - 1);
8118         dio_bio = dip->dio_bio;
8119
8120         kfree(dip);
8121
8122         dio_bio->bi_error = bio->bi_error;
8123         dio_end_io(dio_bio, bio->bi_error);
8124
8125         if (io_bio->end_io)
8126                 io_bio->end_io(io_bio, err);
8127         bio_put(bio);
8128 }
8129
8130 static void btrfs_endio_direct_write_update_ordered(struct inode *inode,
8131                                                     const u64 offset,
8132                                                     const u64 bytes,
8133                                                     const int uptodate)
8134 {
8135         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
8136         struct btrfs_ordered_extent *ordered = NULL;
8137         u64 ordered_offset = offset;
8138         u64 ordered_bytes = bytes;
8139         int ret;
8140
8141 again:
8142         ret = btrfs_dec_test_first_ordered_pending(inode, &ordered,
8143                                                    &ordered_offset,
8144                                                    ordered_bytes,
8145                                                    uptodate);
8146         if (!ret)
8147                 goto out_test;
8148
8149         btrfs_init_work(&ordered->work, btrfs_endio_write_helper,
8150                         finish_ordered_fn, NULL, NULL);
8151         btrfs_queue_work(fs_info->endio_write_workers, &ordered->work);
8152 out_test:
8153         /*
8154          * our bio might span multiple ordered extents.  If we haven't
8155          * completed the accounting for the whole dio, go back and try again
8156          */
8157         if (ordered_offset < offset + bytes) {
8158                 ordered_bytes = offset + bytes - ordered_offset;
8159                 ordered = NULL;
8160                 goto again;
8161         }
8162 }
8163
8164 static void btrfs_endio_direct_write(struct bio *bio)
8165 {
8166         struct btrfs_dio_private *dip = bio->bi_private;
8167         struct bio *dio_bio = dip->dio_bio;
8168
8169         btrfs_endio_direct_write_update_ordered(dip->inode,
8170                                                 dip->logical_offset,
8171                                                 dip->bytes,
8172                                                 !bio->bi_error);
8173
8174         kfree(dip);
8175
8176         dio_bio->bi_error = bio->bi_error;
8177         dio_end_io(dio_bio, bio->bi_error);
8178         bio_put(bio);
8179 }
8180
8181 static int __btrfs_submit_bio_start_direct_io(struct inode *inode,
8182                                     struct bio *bio, int mirror_num,
8183                                     unsigned long bio_flags, u64 offset)
8184 {
8185         int ret;
8186         ret = btrfs_csum_one_bio(inode, bio, offset, 1);
8187         BUG_ON(ret); /* -ENOMEM */
8188         return 0;
8189 }
8190
8191 static void btrfs_end_dio_bio(struct bio *bio)
8192 {
8193         struct btrfs_dio_private *dip = bio->bi_private;
8194         int err = bio->bi_error;
8195
8196         if (err)
8197                 btrfs_warn(BTRFS_I(dip->inode)->root->fs_info,
8198                            "direct IO failed ino %llu rw %d,%u sector %#Lx len %u err no %d",
8199                            btrfs_ino(BTRFS_I(dip->inode)), bio_op(bio),
8200                            bio->bi_opf,
8201                            (unsigned long long)bio->bi_iter.bi_sector,
8202                            bio->bi_iter.bi_size, err);
8203
8204         if (dip->subio_endio)
8205                 err = dip->subio_endio(dip->inode, btrfs_io_bio(bio), err);
8206
8207         if (err) {
8208                 dip->errors = 1;
8209
8210                 /*
8211                  * before atomic variable goto zero, we must make sure
8212                  * dip->errors is perceived to be set.
8213                  */
8214                 smp_mb__before_atomic();
8215         }
8216
8217         /* if there are more bios still pending for this dio, just exit */
8218         if (!atomic_dec_and_test(&dip->pending_bios))
8219                 goto out;
8220
8221         if (dip->errors) {
8222                 bio_io_error(dip->orig_bio);
8223         } else {
8224                 dip->dio_bio->bi_error = 0;
8225                 bio_endio(dip->orig_bio);
8226         }
8227 out:
8228         bio_put(bio);
8229 }
8230
8231 static struct bio *btrfs_dio_bio_alloc(struct block_device *bdev,
8232                                        u64 first_sector, gfp_t gfp_flags)
8233 {
8234         struct bio *bio;
8235         bio = btrfs_bio_alloc(bdev, first_sector, BIO_MAX_PAGES, gfp_flags);
8236         if (bio)
8237                 bio_associate_current(bio);
8238         return bio;
8239 }
8240
8241 static inline int btrfs_lookup_and_bind_dio_csum(struct inode *inode,
8242                                                  struct btrfs_dio_private *dip,
8243                                                  struct bio *bio,
8244                                                  u64 file_offset)
8245 {
8246         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
8247         struct btrfs_io_bio *orig_io_bio = btrfs_io_bio(dip->orig_bio);
8248         int ret;
8249
8250         /*
8251          * We load all the csum data we need when we submit
8252          * the first bio to reduce the csum tree search and
8253          * contention.
8254          */
8255         if (dip->logical_offset == file_offset) {
8256                 ret = btrfs_lookup_bio_sums_dio(inode, dip->orig_bio,
8257                                                 file_offset);
8258                 if (ret)
8259                         return ret;
8260         }
8261
8262         if (bio == dip->orig_bio)
8263                 return 0;
8264
8265         file_offset -= dip->logical_offset;
8266         file_offset >>= inode->i_sb->s_blocksize_bits;
8267         io_bio->csum = (u8 *)(((u32 *)orig_io_bio->csum) + file_offset);
8268
8269         return 0;
8270 }
8271
8272 static inline int __btrfs_submit_dio_bio(struct bio *bio, struct inode *inode,
8273                                          u64 file_offset, int skip_sum,
8274                                          int async_submit)
8275 {
8276         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
8277         struct btrfs_dio_private *dip = bio->bi_private;
8278         bool write = bio_op(bio) == REQ_OP_WRITE;
8279         int ret;
8280
8281         if (async_submit)
8282                 async_submit = !atomic_read(&BTRFS_I(inode)->sync_writers);
8283
8284         bio_get(bio);
8285
8286         if (!write) {
8287                 ret = btrfs_bio_wq_end_io(fs_info, bio, BTRFS_WQ_ENDIO_DATA);
8288                 if (ret)
8289                         goto err;
8290         }
8291
8292         if (skip_sum)
8293                 goto map;
8294
8295         if (write && async_submit) {
8296                 ret = btrfs_wq_submit_bio(fs_info, inode, bio, 0, 0,
8297                                           file_offset,
8298                                           __btrfs_submit_bio_start_direct_io,
8299                                           __btrfs_submit_bio_done);
8300                 goto err;
8301         } else if (write) {
8302                 /*
8303                  * If we aren't doing async submit, calculate the csum of the
8304                  * bio now.
8305                  */
8306                 ret = btrfs_csum_one_bio(inode, bio, file_offset, 1);
8307                 if (ret)
8308                         goto err;
8309         } else {
8310                 ret = btrfs_lookup_and_bind_dio_csum(inode, dip, bio,
8311                                                      file_offset);
8312                 if (ret)
8313                         goto err;
8314         }
8315 map:
8316         ret = btrfs_map_bio(fs_info, bio, 0, async_submit);
8317 err:
8318         bio_put(bio);
8319         return ret;
8320 }
8321
8322 static int btrfs_submit_direct_hook(struct btrfs_dio_private *dip,
8323                                     int skip_sum)
8324 {
8325         struct inode *inode = dip->inode;
8326         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
8327         struct btrfs_root *root = BTRFS_I(inode)->root;
8328         struct bio *bio;
8329         struct bio *orig_bio = dip->orig_bio;
8330         struct bio_vec *bvec;
8331         u64 start_sector = orig_bio->bi_iter.bi_sector;
8332         u64 file_offset = dip->logical_offset;
8333         u64 submit_len = 0;
8334         u64 map_length;
8335         u32 blocksize = fs_info->sectorsize;
8336         int async_submit = 0;
8337         int nr_sectors;
8338         int ret;
8339         int i, j;
8340
8341         map_length = orig_bio->bi_iter.bi_size;
8342         ret = btrfs_map_block(fs_info, btrfs_op(orig_bio), start_sector << 9,
8343                               &map_length, NULL, 0);
8344         if (ret)
8345                 return -EIO;
8346
8347         if (map_length >= orig_bio->bi_iter.bi_size) {
8348                 bio = orig_bio;
8349                 dip->flags |= BTRFS_DIO_ORIG_BIO_SUBMITTED;
8350                 goto submit;
8351         }
8352
8353         /* async crcs make it difficult to collect full stripe writes. */
8354         if (btrfs_get_alloc_profile(root, 1) & BTRFS_BLOCK_GROUP_RAID56_MASK)
8355                 async_submit = 0;
8356         else
8357                 async_submit = 1;
8358
8359         bio = btrfs_dio_bio_alloc(orig_bio->bi_bdev, start_sector, GFP_NOFS);
8360         if (!bio)
8361                 return -ENOMEM;
8362
8363         bio->bi_opf = orig_bio->bi_opf;
8364         bio->bi_private = dip;
8365         bio->bi_end_io = btrfs_end_dio_bio;
8366         btrfs_io_bio(bio)->logical = file_offset;
8367         atomic_inc(&dip->pending_bios);
8368
8369         bio_for_each_segment_all(bvec, orig_bio, j) {
8370                 nr_sectors = BTRFS_BYTES_TO_BLKS(fs_info, bvec->bv_len);
8371                 i = 0;
8372 next_block:
8373                 if (unlikely(map_length < submit_len + blocksize ||
8374                     bio_add_page(bio, bvec->bv_page, blocksize,
8375                             bvec->bv_offset + (i * blocksize)) < blocksize)) {
8376                         /*
8377                          * inc the count before we submit the bio so
8378                          * we know the end IO handler won't happen before
8379                          * we inc the count. Otherwise, the dip might get freed
8380                          * before we're done setting it up
8381                          */
8382                         atomic_inc(&dip->pending_bios);
8383                         ret = __btrfs_submit_dio_bio(bio, inode,
8384                                                      file_offset, skip_sum,
8385                                                      async_submit);
8386                         if (ret) {
8387                                 bio_put(bio);
8388                                 atomic_dec(&dip->pending_bios);
8389                                 goto out_err;
8390                         }
8391
8392                         start_sector += submit_len >> 9;
8393                         file_offset += submit_len;
8394
8395                         submit_len = 0;
8396
8397                         bio = btrfs_dio_bio_alloc(orig_bio->bi_bdev,
8398                                                   start_sector, GFP_NOFS);
8399                         if (!bio)
8400                                 goto out_err;
8401                         bio->bi_opf = orig_bio->bi_opf;
8402                         bio->bi_private = dip;
8403                         bio->bi_end_io = btrfs_end_dio_bio;
8404                         btrfs_io_bio(bio)->logical = file_offset;
8405
8406                         map_length = orig_bio->bi_iter.bi_size;
8407                         ret = btrfs_map_block(fs_info, btrfs_op(orig_bio),
8408                                               start_sector << 9,
8409                                               &map_length, NULL, 0);
8410                         if (ret) {
8411                                 bio_put(bio);
8412                                 goto out_err;
8413                         }
8414
8415                         goto next_block;
8416                 } else {
8417                         submit_len += blocksize;
8418                         if (--nr_sectors) {
8419                                 i++;
8420                                 goto next_block;
8421                         }
8422                 }
8423         }
8424
8425 submit:
8426         ret = __btrfs_submit_dio_bio(bio, inode, file_offset, skip_sum,
8427                                      async_submit);
8428         if (!ret)
8429                 return 0;
8430
8431         bio_put(bio);
8432 out_err:
8433         dip->errors = 1;
8434         /*
8435          * before atomic variable goto zero, we must
8436          * make sure dip->errors is perceived to be set.
8437          */
8438         smp_mb__before_atomic();
8439         if (atomic_dec_and_test(&dip->pending_bios))
8440                 bio_io_error(dip->orig_bio);
8441
8442         /* bio_end_io() will handle error, so we needn't return it */
8443         return 0;
8444 }
8445
8446 static void btrfs_submit_direct(struct bio *dio_bio, struct inode *inode,
8447                                 loff_t file_offset)
8448 {
8449         struct btrfs_dio_private *dip = NULL;
8450         struct bio *io_bio = NULL;
8451         struct btrfs_io_bio *btrfs_bio;
8452         int skip_sum;
8453         bool write = (bio_op(dio_bio) == REQ_OP_WRITE);
8454         int ret = 0;
8455
8456         skip_sum = BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM;
8457
8458         io_bio = btrfs_bio_clone(dio_bio, GFP_NOFS);
8459         if (!io_bio) {
8460                 ret = -ENOMEM;
8461                 goto free_ordered;
8462         }
8463
8464         dip = kzalloc(sizeof(*dip), GFP_NOFS);
8465         if (!dip) {
8466                 ret = -ENOMEM;
8467                 goto free_ordered;
8468         }
8469
8470         dip->private = dio_bio->bi_private;
8471         dip->inode = inode;
8472         dip->logical_offset = file_offset;
8473         dip->bytes = dio_bio->bi_iter.bi_size;
8474         dip->disk_bytenr = (u64)dio_bio->bi_iter.bi_sector << 9;
8475         io_bio->bi_private = dip;
8476         dip->orig_bio = io_bio;
8477         dip->dio_bio = dio_bio;
8478         atomic_set(&dip->pending_bios, 0);
8479         btrfs_bio = btrfs_io_bio(io_bio);
8480         btrfs_bio->logical = file_offset;
8481
8482         if (write) {
8483                 io_bio->bi_end_io = btrfs_endio_direct_write;
8484         } else {
8485                 io_bio->bi_end_io = btrfs_endio_direct_read;
8486                 dip->subio_endio = btrfs_subio_endio_read;
8487         }
8488
8489         /*
8490          * Reset the range for unsubmitted ordered extents (to a 0 length range)
8491          * even if we fail to submit a bio, because in such case we do the
8492          * corresponding error handling below and it must not be done a second
8493          * time by btrfs_direct_IO().
8494          */
8495         if (write) {
8496                 struct btrfs_dio_data *dio_data = current->journal_info;
8497
8498                 dio_data->unsubmitted_oe_range_end = dip->logical_offset +
8499                         dip->bytes;
8500                 dio_data->unsubmitted_oe_range_start =
8501                         dio_data->unsubmitted_oe_range_end;
8502         }
8503
8504         ret = btrfs_submit_direct_hook(dip, skip_sum);
8505         if (!ret)
8506                 return;
8507
8508         if (btrfs_bio->end_io)
8509                 btrfs_bio->end_io(btrfs_bio, ret);
8510
8511 free_ordered:
8512         /*
8513          * If we arrived here it means either we failed to submit the dip
8514          * or we either failed to clone the dio_bio or failed to allocate the
8515          * dip. If we cloned the dio_bio and allocated the dip, we can just
8516          * call bio_endio against our io_bio so that we get proper resource
8517          * cleanup if we fail to submit the dip, otherwise, we must do the
8518          * same as btrfs_endio_direct_[write|read] because we can't call these
8519          * callbacks - they require an allocated dip and a clone of dio_bio.
8520          */
8521         if (io_bio && dip) {
8522                 io_bio->bi_error = -EIO;
8523                 bio_endio(io_bio);
8524                 /*
8525                  * The end io callbacks free our dip, do the final put on io_bio
8526                  * and all the cleanup and final put for dio_bio (through
8527                  * dio_end_io()).
8528                  */
8529                 dip = NULL;
8530                 io_bio = NULL;
8531         } else {
8532                 if (write)
8533                         btrfs_endio_direct_write_update_ordered(inode,
8534                                                 file_offset,
8535                                                 dio_bio->bi_iter.bi_size,
8536                                                 0);
8537                 else
8538                         unlock_extent(&BTRFS_I(inode)->io_tree, file_offset,
8539                               file_offset + dio_bio->bi_iter.bi_size - 1);
8540
8541                 dio_bio->bi_error = -EIO;
8542                 /*
8543                  * Releases and cleans up our dio_bio, no need to bio_put()
8544                  * nor bio_endio()/bio_io_error() against dio_bio.
8545                  */
8546                 dio_end_io(dio_bio, ret);
8547         }
8548         if (io_bio)
8549                 bio_put(io_bio);
8550         kfree(dip);
8551 }
8552
8553 static ssize_t check_direct_IO(struct btrfs_fs_info *fs_info,
8554                                struct kiocb *iocb,
8555                                const struct iov_iter *iter, loff_t offset)
8556 {
8557         int seg;
8558         int i;
8559         unsigned int blocksize_mask = fs_info->sectorsize - 1;
8560         ssize_t retval = -EINVAL;
8561
8562         if (offset & blocksize_mask)
8563                 goto out;
8564
8565         if (iov_iter_alignment(iter) & blocksize_mask)
8566                 goto out;
8567
8568         /* If this is a write we don't need to check anymore */
8569         if (iov_iter_rw(iter) != READ || !iter_is_iovec(iter))
8570                 return 0;
8571         /*
8572          * Check to make sure we don't have duplicate iov_base's in this
8573          * iovec, if so return EINVAL, otherwise we'll get csum errors
8574          * when reading back.
8575          */
8576         for (seg = 0; seg < iter->nr_segs; seg++) {
8577                 for (i = seg + 1; i < iter->nr_segs; i++) {
8578                         if (iter->iov[seg].iov_base == iter->iov[i].iov_base)
8579                                 goto out;
8580                 }
8581         }
8582         retval = 0;
8583 out:
8584         return retval;
8585 }
8586
8587 static ssize_t btrfs_direct_IO(struct kiocb *iocb, struct iov_iter *iter)
8588 {
8589         struct file *file = iocb->ki_filp;
8590         struct inode *inode = file->f_mapping->host;
8591         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
8592         struct btrfs_dio_data dio_data = { 0 };
8593         loff_t offset = iocb->ki_pos;
8594         size_t count = 0;
8595         int flags = 0;
8596         bool wakeup = true;
8597         bool relock = false;
8598         ssize_t ret;
8599
8600         if (check_direct_IO(fs_info, iocb, iter, offset))
8601                 return 0;
8602
8603         inode_dio_begin(inode);
8604         smp_mb__after_atomic();
8605
8606         /*
8607          * The generic stuff only does filemap_write_and_wait_range, which
8608          * isn't enough if we've written compressed pages to this area, so
8609          * we need to flush the dirty pages again to make absolutely sure
8610          * that any outstanding dirty pages are on disk.
8611          */
8612         count = iov_iter_count(iter);
8613         if (test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
8614                      &BTRFS_I(inode)->runtime_flags))
8615                 filemap_fdatawrite_range(inode->i_mapping, offset,
8616                                          offset + count - 1);
8617
8618         if (iov_iter_rw(iter) == WRITE) {
8619                 /*
8620                  * If the write DIO is beyond the EOF, we need update
8621                  * the isize, but it is protected by i_mutex. So we can
8622                  * not unlock the i_mutex at this case.
8623                  */
8624                 if (offset + count <= inode->i_size) {
8625                         dio_data.overwrite = 1;
8626                         inode_unlock(inode);
8627                         relock = true;
8628                 }
8629                 ret = btrfs_delalloc_reserve_space(inode, offset, count);
8630                 if (ret)
8631                         goto out;
8632                 dio_data.outstanding_extents = count_max_extents(count);
8633
8634                 /*
8635                  * We need to know how many extents we reserved so that we can
8636                  * do the accounting properly if we go over the number we
8637                  * originally calculated.  Abuse current->journal_info for this.
8638                  */
8639                 dio_data.reserve = round_up(count,
8640                                             fs_info->sectorsize);
8641                 dio_data.unsubmitted_oe_range_start = (u64)offset;
8642                 dio_data.unsubmitted_oe_range_end = (u64)offset;
8643                 current->journal_info = &dio_data;
8644                 down_read(&BTRFS_I(inode)->dio_sem);
8645         } else if (test_bit(BTRFS_INODE_READDIO_NEED_LOCK,
8646                                      &BTRFS_I(inode)->runtime_flags)) {
8647                 inode_dio_end(inode);
8648                 flags = DIO_LOCKING | DIO_SKIP_HOLES;
8649                 wakeup = false;
8650         }
8651
8652         ret = __blockdev_direct_IO(iocb, inode,
8653                                    fs_info->fs_devices->latest_bdev,
8654                                    iter, btrfs_get_blocks_direct, NULL,
8655                                    btrfs_submit_direct, flags);
8656         if (iov_iter_rw(iter) == WRITE) {
8657                 up_read(&BTRFS_I(inode)->dio_sem);
8658                 current->journal_info = NULL;
8659                 if (ret < 0 && ret != -EIOCBQUEUED) {
8660                         if (dio_data.reserve)
8661                                 btrfs_delalloc_release_space(inode, offset,
8662                                                              dio_data.reserve);
8663                         /*
8664                          * On error we might have left some ordered extents
8665                          * without submitting corresponding bios for them, so
8666                          * cleanup them up to avoid other tasks getting them
8667                          * and waiting for them to complete forever.
8668                          */
8669                         if (dio_data.unsubmitted_oe_range_start <
8670                             dio_data.unsubmitted_oe_range_end)
8671                                 btrfs_endio_direct_write_update_ordered(inode,
8672                                         dio_data.unsubmitted_oe_range_start,
8673                                         dio_data.unsubmitted_oe_range_end -
8674                                         dio_data.unsubmitted_oe_range_start,
8675                                         0);
8676                 } else if (ret >= 0 && (size_t)ret < count)
8677                         btrfs_delalloc_release_space(inode, offset,
8678                                                      count - (size_t)ret);
8679         }
8680 out:
8681         if (wakeup)
8682                 inode_dio_end(inode);
8683         if (relock)
8684                 inode_lock(inode);
8685
8686         return ret;
8687 }
8688
8689 #define BTRFS_FIEMAP_FLAGS      (FIEMAP_FLAG_SYNC)
8690
8691 static int btrfs_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
8692                 __u64 start, __u64 len)
8693 {
8694         int     ret;
8695
8696         ret = fiemap_check_flags(fieinfo, BTRFS_FIEMAP_FLAGS);
8697         if (ret)
8698                 return ret;
8699
8700         return extent_fiemap(inode, fieinfo, start, len, btrfs_get_extent_fiemap);
8701 }
8702
8703 int btrfs_readpage(struct file *file, struct page *page)
8704 {
8705         struct extent_io_tree *tree;
8706         tree = &BTRFS_I(page->mapping->host)->io_tree;
8707         return extent_read_full_page(tree, page, btrfs_get_extent, 0);
8708 }
8709
8710 static int btrfs_writepage(struct page *page, struct writeback_control *wbc)
8711 {
8712         struct extent_io_tree *tree;
8713         struct inode *inode = page->mapping->host;
8714         int ret;
8715
8716         if (current->flags & PF_MEMALLOC) {
8717                 redirty_page_for_writepage(wbc, page);
8718                 unlock_page(page);
8719                 return 0;
8720         }
8721
8722         /*
8723          * If we are under memory pressure we will call this directly from the
8724          * VM, we need to make sure we have the inode referenced for the ordered
8725          * extent.  If not just return like we didn't do anything.
8726          */
8727         if (!igrab(inode)) {
8728                 redirty_page_for_writepage(wbc, page);
8729                 return AOP_WRITEPAGE_ACTIVATE;
8730         }
8731         tree = &BTRFS_I(page->mapping->host)->io_tree;
8732         ret = extent_write_full_page(tree, page, btrfs_get_extent, wbc);
8733         btrfs_add_delayed_iput(inode);
8734         return ret;
8735 }
8736
8737 static int btrfs_writepages(struct address_space *mapping,
8738                             struct writeback_control *wbc)
8739 {
8740         struct extent_io_tree *tree;
8741
8742         tree = &BTRFS_I(mapping->host)->io_tree;
8743         return extent_writepages(tree, mapping, btrfs_get_extent, wbc);
8744 }
8745
8746 static int
8747 btrfs_readpages(struct file *file, struct address_space *mapping,
8748                 struct list_head *pages, unsigned nr_pages)
8749 {
8750         struct extent_io_tree *tree;
8751         tree = &BTRFS_I(mapping->host)->io_tree;
8752         return extent_readpages(tree, mapping, pages, nr_pages,
8753                                 btrfs_get_extent);
8754 }
8755 static int __btrfs_releasepage(struct page *page, gfp_t gfp_flags)
8756 {
8757         struct extent_io_tree *tree;
8758         struct extent_map_tree *map;
8759         int ret;
8760
8761         tree = &BTRFS_I(page->mapping->host)->io_tree;
8762         map = &BTRFS_I(page->mapping->host)->extent_tree;
8763         ret = try_release_extent_mapping(map, tree, page, gfp_flags);
8764         if (ret == 1) {
8765                 ClearPagePrivate(page);
8766                 set_page_private(page, 0);
8767                 put_page(page);
8768         }
8769         return ret;
8770 }
8771
8772 static int btrfs_releasepage(struct page *page, gfp_t gfp_flags)
8773 {
8774         if (PageWriteback(page) || PageDirty(page))
8775                 return 0;
8776         return __btrfs_releasepage(page, gfp_flags);
8777 }
8778
8779 static void btrfs_invalidatepage(struct page *page, unsigned int offset,
8780                                  unsigned int length)
8781 {
8782         struct inode *inode = page->mapping->host;
8783         struct extent_io_tree *tree;
8784         struct btrfs_ordered_extent *ordered;
8785         struct extent_state *cached_state = NULL;
8786         u64 page_start = page_offset(page);
8787         u64 page_end = page_start + PAGE_SIZE - 1;
8788         u64 start;
8789         u64 end;
8790         int inode_evicting = inode->i_state & I_FREEING;
8791
8792         /*
8793          * we have the page locked, so new writeback can't start,
8794          * and the dirty bit won't be cleared while we are here.
8795          *
8796          * Wait for IO on this page so that we can safely clear
8797          * the PagePrivate2 bit and do ordered accounting
8798          */
8799         wait_on_page_writeback(page);
8800
8801         tree = &BTRFS_I(inode)->io_tree;
8802         if (offset) {
8803                 btrfs_releasepage(page, GFP_NOFS);
8804                 return;
8805         }
8806
8807         if (!inode_evicting)
8808                 lock_extent_bits(tree, page_start, page_end, &cached_state);
8809 again:
8810         start = page_start;
8811         ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), start,
8812                                         page_end - start + 1);
8813         if (ordered) {
8814                 end = min(page_end, ordered->file_offset + ordered->len - 1);
8815                 /*
8816                  * IO on this page will never be started, so we need
8817                  * to account for any ordered extents now
8818                  */
8819                 if (!inode_evicting)
8820                         clear_extent_bit(tree, start, end,
8821                                          EXTENT_DIRTY | EXTENT_DELALLOC |
8822                                          EXTENT_LOCKED | EXTENT_DO_ACCOUNTING |
8823                                          EXTENT_DEFRAG, 1, 0, &cached_state,
8824                                          GFP_NOFS);
8825                 /*
8826                  * whoever cleared the private bit is responsible
8827                  * for the finish_ordered_io
8828                  */
8829                 if (TestClearPagePrivate2(page)) {
8830                         struct btrfs_ordered_inode_tree *tree;
8831                         u64 new_len;
8832
8833                         tree = &BTRFS_I(inode)->ordered_tree;
8834
8835                         spin_lock_irq(&tree->lock);
8836                         set_bit(BTRFS_ORDERED_TRUNCATED, &ordered->flags);
8837                         new_len = start - ordered->file_offset;
8838                         if (new_len < ordered->truncated_len)
8839                                 ordered->truncated_len = new_len;
8840                         spin_unlock_irq(&tree->lock);
8841
8842                         if (btrfs_dec_test_ordered_pending(inode, &ordered,
8843                                                            start,
8844                                                            end - start + 1, 1))
8845                                 btrfs_finish_ordered_io(ordered);
8846                 }
8847                 btrfs_put_ordered_extent(ordered);
8848                 if (!inode_evicting) {
8849                         cached_state = NULL;
8850                         lock_extent_bits(tree, start, end,
8851                                          &cached_state);
8852                 }
8853
8854                 start = end + 1;
8855                 if (start < page_end)
8856                         goto again;
8857         }
8858
8859         /*
8860          * Qgroup reserved space handler
8861          * Page here will be either
8862          * 1) Already written to disk
8863          *    In this case, its reserved space is released from data rsv map
8864          *    and will be freed by delayed_ref handler finally.
8865          *    So even we call qgroup_free_data(), it won't decrease reserved
8866          *    space.
8867          * 2) Not written to disk
8868          *    This means the reserved space should be freed here. However,
8869          *    if a truncate invalidates the page (by clearing PageDirty)
8870          *    and the page is accounted for while allocating extent
8871          *    in btrfs_check_data_free_space() we let delayed_ref to
8872          *    free the entire extent.
8873          */
8874         if (PageDirty(page))
8875                 btrfs_qgroup_free_data(inode, page_start, PAGE_SIZE);
8876         if (!inode_evicting) {
8877                 clear_extent_bit(tree, page_start, page_end,
8878                                  EXTENT_LOCKED | EXTENT_DIRTY |
8879                                  EXTENT_DELALLOC | EXTENT_DO_ACCOUNTING |
8880                                  EXTENT_DEFRAG, 1, 1,
8881                                  &cached_state, GFP_NOFS);
8882
8883                 __btrfs_releasepage(page, GFP_NOFS);
8884         }
8885
8886         ClearPageChecked(page);
8887         if (PagePrivate(page)) {
8888                 ClearPagePrivate(page);
8889                 set_page_private(page, 0);
8890                 put_page(page);
8891         }
8892 }
8893
8894 /*
8895  * btrfs_page_mkwrite() is not allowed to change the file size as it gets
8896  * called from a page fault handler when a page is first dirtied. Hence we must
8897  * be careful to check for EOF conditions here. We set the page up correctly
8898  * for a written page which means we get ENOSPC checking when writing into
8899  * holes and correct delalloc and unwritten extent mapping on filesystems that
8900  * support these features.
8901  *
8902  * We are not allowed to take the i_mutex here so we have to play games to
8903  * protect against truncate races as the page could now be beyond EOF.  Because
8904  * vmtruncate() writes the inode size before removing pages, once we have the
8905  * page lock we can determine safely if the page is beyond EOF. If it is not
8906  * beyond EOF, then the page is guaranteed safe against truncation until we
8907  * unlock the page.
8908  */
8909 int btrfs_page_mkwrite(struct vm_fault *vmf)
8910 {
8911         struct page *page = vmf->page;
8912         struct inode *inode = file_inode(vmf->vma->vm_file);
8913         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
8914         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
8915         struct btrfs_ordered_extent *ordered;
8916         struct extent_state *cached_state = NULL;
8917         char *kaddr;
8918         unsigned long zero_start;
8919         loff_t size;
8920         int ret;
8921         int reserved = 0;
8922         u64 reserved_space;
8923         u64 page_start;
8924         u64 page_end;
8925         u64 end;
8926
8927         reserved_space = PAGE_SIZE;
8928
8929         sb_start_pagefault(inode->i_sb);
8930         page_start = page_offset(page);
8931         page_end = page_start + PAGE_SIZE - 1;
8932         end = page_end;
8933
8934         /*
8935          * Reserving delalloc space after obtaining the page lock can lead to
8936          * deadlock. For example, if a dirty page is locked by this function
8937          * and the call to btrfs_delalloc_reserve_space() ends up triggering
8938          * dirty page write out, then the btrfs_writepage() function could
8939          * end up waiting indefinitely to get a lock on the page currently
8940          * being processed by btrfs_page_mkwrite() function.
8941          */
8942         ret = btrfs_delalloc_reserve_space(inode, page_start,
8943                                            reserved_space);
8944         if (!ret) {
8945                 ret = file_update_time(vmf->vma->vm_file);
8946                 reserved = 1;
8947         }
8948         if (ret) {
8949                 if (ret == -ENOMEM)
8950                         ret = VM_FAULT_OOM;
8951                 else /* -ENOSPC, -EIO, etc */
8952                         ret = VM_FAULT_SIGBUS;
8953                 if (reserved)
8954                         goto out;
8955                 goto out_noreserve;
8956         }
8957
8958         ret = VM_FAULT_NOPAGE; /* make the VM retry the fault */
8959 again:
8960         lock_page(page);
8961         size = i_size_read(inode);
8962
8963         if ((page->mapping != inode->i_mapping) ||
8964             (page_start >= size)) {
8965                 /* page got truncated out from underneath us */
8966                 goto out_unlock;
8967         }
8968         wait_on_page_writeback(page);
8969
8970         lock_extent_bits(io_tree, page_start, page_end, &cached_state);
8971         set_page_extent_mapped(page);
8972
8973         /*
8974          * we can't set the delalloc bits if there are pending ordered
8975          * extents.  Drop our locks and wait for them to finish
8976          */
8977         ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), page_start,
8978                         PAGE_SIZE);
8979         if (ordered) {
8980                 unlock_extent_cached(io_tree, page_start, page_end,
8981                                      &cached_state, GFP_NOFS);
8982                 unlock_page(page);
8983                 btrfs_start_ordered_extent(inode, ordered, 1);
8984                 btrfs_put_ordered_extent(ordered);
8985                 goto again;
8986         }
8987
8988         if (page->index == ((size - 1) >> PAGE_SHIFT)) {
8989                 reserved_space = round_up(size - page_start,
8990                                           fs_info->sectorsize);
8991                 if (reserved_space < PAGE_SIZE) {
8992                         end = page_start + reserved_space - 1;
8993                         spin_lock(&BTRFS_I(inode)->lock);
8994                         BTRFS_I(inode)->outstanding_extents++;
8995                         spin_unlock(&BTRFS_I(inode)->lock);
8996                         btrfs_delalloc_release_space(inode, page_start,
8997                                                 PAGE_SIZE - reserved_space);
8998                 }
8999         }
9000
9001         /*
9002          * page_mkwrite gets called when the page is firstly dirtied after it's
9003          * faulted in, but write(2) could also dirty a page and set delalloc
9004          * bits, thus in this case for space account reason, we still need to
9005          * clear any delalloc bits within this page range since we have to
9006          * reserve data&meta space before lock_page() (see above comments).
9007          */
9008         clear_extent_bit(&BTRFS_I(inode)->io_tree, page_start, end,
9009                           EXTENT_DIRTY | EXTENT_DELALLOC |
9010                           EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG,
9011                           0, 0, &cached_state, GFP_NOFS);
9012
9013         ret = btrfs_set_extent_delalloc(inode, page_start, end,
9014                                         &cached_state, 0);
9015         if (ret) {
9016                 unlock_extent_cached(io_tree, page_start, page_end,
9017                                      &cached_state, GFP_NOFS);
9018                 ret = VM_FAULT_SIGBUS;
9019                 goto out_unlock;
9020         }
9021         ret = 0;
9022
9023         /* page is wholly or partially inside EOF */
9024         if (page_start + PAGE_SIZE > size)
9025                 zero_start = size & ~PAGE_MASK;
9026         else
9027                 zero_start = PAGE_SIZE;
9028
9029         if (zero_start != PAGE_SIZE) {
9030                 kaddr = kmap(page);
9031                 memset(kaddr + zero_start, 0, PAGE_SIZE - zero_start);
9032                 flush_dcache_page(page);
9033                 kunmap(page);
9034         }
9035         ClearPageChecked(page);
9036         set_page_dirty(page);
9037         SetPageUptodate(page);
9038
9039         BTRFS_I(inode)->last_trans = fs_info->generation;
9040         BTRFS_I(inode)->last_sub_trans = BTRFS_I(inode)->root->log_transid;
9041         BTRFS_I(inode)->last_log_commit = BTRFS_I(inode)->root->last_log_commit;
9042
9043         unlock_extent_cached(io_tree, page_start, page_end, &cached_state, GFP_NOFS);
9044
9045 out_unlock:
9046         if (!ret) {
9047                 sb_end_pagefault(inode->i_sb);
9048                 return VM_FAULT_LOCKED;
9049         }
9050         unlock_page(page);
9051 out:
9052         btrfs_delalloc_release_space(inode, page_start, reserved_space);
9053 out_noreserve:
9054         sb_end_pagefault(inode->i_sb);
9055         return ret;
9056 }
9057
9058 static int btrfs_truncate(struct inode *inode)
9059 {
9060         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
9061         struct btrfs_root *root = BTRFS_I(inode)->root;
9062         struct btrfs_block_rsv *rsv;
9063         int ret = 0;
9064         int err = 0;
9065         struct btrfs_trans_handle *trans;
9066         u64 mask = fs_info->sectorsize - 1;
9067         u64 min_size = btrfs_calc_trunc_metadata_size(fs_info, 1);
9068
9069         ret = btrfs_wait_ordered_range(inode, inode->i_size & (~mask),
9070                                        (u64)-1);
9071         if (ret)
9072                 return ret;
9073
9074         /*
9075          * Yes ladies and gentlemen, this is indeed ugly.  The fact is we have
9076          * 3 things going on here
9077          *
9078          * 1) We need to reserve space for our orphan item and the space to
9079          * delete our orphan item.  Lord knows we don't want to have a dangling
9080          * orphan item because we didn't reserve space to remove it.
9081          *
9082          * 2) We need to reserve space to update our inode.
9083          *
9084          * 3) We need to have something to cache all the space that is going to
9085          * be free'd up by the truncate operation, but also have some slack
9086          * space reserved in case it uses space during the truncate (thank you
9087          * very much snapshotting).
9088          *
9089          * And we need these to all be separate.  The fact is we can use a lot of
9090          * space doing the truncate, and we have no earthly idea how much space
9091          * we will use, so we need the truncate reservation to be separate so it
9092          * doesn't end up using space reserved for updating the inode or
9093          * removing the orphan item.  We also need to be able to stop the
9094          * transaction and start a new one, which means we need to be able to
9095          * update the inode several times, and we have no idea of knowing how
9096          * many times that will be, so we can't just reserve 1 item for the
9097          * entirety of the operation, so that has to be done separately as well.
9098          * Then there is the orphan item, which does indeed need to be held on
9099          * to for the whole operation, and we need nobody to touch this reserved
9100          * space except the orphan code.
9101          *
9102          * So that leaves us with
9103          *
9104          * 1) root->orphan_block_rsv - for the orphan deletion.
9105          * 2) rsv - for the truncate reservation, which we will steal from the
9106          * transaction reservation.
9107          * 3) fs_info->trans_block_rsv - this will have 1 items worth left for
9108          * updating the inode.
9109          */
9110         rsv = btrfs_alloc_block_rsv(fs_info, BTRFS_BLOCK_RSV_TEMP);
9111         if (!rsv)
9112                 return -ENOMEM;
9113         rsv->size = min_size;
9114         rsv->failfast = 1;
9115
9116         /*
9117          * 1 for the truncate slack space
9118          * 1 for updating the inode.
9119          */
9120         trans = btrfs_start_transaction(root, 2);
9121         if (IS_ERR(trans)) {
9122                 err = PTR_ERR(trans);
9123                 goto out;
9124         }
9125
9126         /* Migrate the slack space for the truncate to our reserve */
9127         ret = btrfs_block_rsv_migrate(&fs_info->trans_block_rsv, rsv,
9128                                       min_size, 0);
9129         BUG_ON(ret);
9130
9131         /*
9132          * So if we truncate and then write and fsync we normally would just
9133          * write the extents that changed, which is a problem if we need to
9134          * first truncate that entire inode.  So set this flag so we write out
9135          * all of the extents in the inode to the sync log so we're completely
9136          * safe.
9137          */
9138         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &BTRFS_I(inode)->runtime_flags);
9139         trans->block_rsv = rsv;
9140
9141         while (1) {
9142                 ret = btrfs_truncate_inode_items(trans, root, inode,
9143                                                  inode->i_size,
9144                                                  BTRFS_EXTENT_DATA_KEY);
9145                 if (ret != -ENOSPC && ret != -EAGAIN) {
9146                         err = ret;
9147                         break;
9148                 }
9149
9150                 trans->block_rsv = &fs_info->trans_block_rsv;
9151                 ret = btrfs_update_inode(trans, root, inode);
9152                 if (ret) {
9153                         err = ret;
9154                         break;
9155                 }
9156
9157                 btrfs_end_transaction(trans);
9158                 btrfs_btree_balance_dirty(fs_info);
9159
9160                 trans = btrfs_start_transaction(root, 2);
9161                 if (IS_ERR(trans)) {
9162                         ret = err = PTR_ERR(trans);
9163                         trans = NULL;
9164                         break;
9165                 }
9166
9167                 btrfs_block_rsv_release(fs_info, rsv, -1);
9168                 ret = btrfs_block_rsv_migrate(&fs_info->trans_block_rsv,
9169                                               rsv, min_size, 0);
9170                 BUG_ON(ret);    /* shouldn't happen */
9171                 trans->block_rsv = rsv;
9172         }
9173
9174         if (ret == 0 && inode->i_nlink > 0) {
9175                 trans->block_rsv = root->orphan_block_rsv;
9176                 ret = btrfs_orphan_del(trans, BTRFS_I(inode));
9177                 if (ret)
9178                         err = ret;
9179         }
9180
9181         if (trans) {
9182                 trans->block_rsv = &fs_info->trans_block_rsv;
9183                 ret = btrfs_update_inode(trans, root, inode);
9184                 if (ret && !err)
9185                         err = ret;
9186
9187                 ret = btrfs_end_transaction(trans);
9188                 btrfs_btree_balance_dirty(fs_info);
9189         }
9190 out:
9191         btrfs_free_block_rsv(fs_info, rsv);
9192
9193         if (ret && !err)
9194                 err = ret;
9195
9196         return err;
9197 }
9198
9199 /*
9200  * create a new subvolume directory/inode (helper for the ioctl).
9201  */
9202 int btrfs_create_subvol_root(struct btrfs_trans_handle *trans,
9203                              struct btrfs_root *new_root,
9204                              struct btrfs_root *parent_root,
9205                              u64 new_dirid)
9206 {
9207         struct inode *inode;
9208         int err;
9209         u64 index = 0;
9210
9211         inode = btrfs_new_inode(trans, new_root, NULL, "..", 2,
9212                                 new_dirid, new_dirid,
9213                                 S_IFDIR | (~current_umask() & S_IRWXUGO),
9214                                 &index);
9215         if (IS_ERR(inode))
9216                 return PTR_ERR(inode);
9217         inode->i_op = &btrfs_dir_inode_operations;
9218         inode->i_fop = &btrfs_dir_file_operations;
9219
9220         set_nlink(inode, 1);
9221         btrfs_i_size_write(BTRFS_I(inode), 0);
9222         unlock_new_inode(inode);
9223
9224         err = btrfs_subvol_inherit_props(trans, new_root, parent_root);
9225         if (err)
9226                 btrfs_err(new_root->fs_info,
9227                           "error inheriting subvolume %llu properties: %d",
9228                           new_root->root_key.objectid, err);
9229
9230         err = btrfs_update_inode(trans, new_root, inode);
9231
9232         iput(inode);
9233         return err;
9234 }
9235
9236 struct inode *btrfs_alloc_inode(struct super_block *sb)
9237 {
9238         struct btrfs_inode *ei;
9239         struct inode *inode;
9240
9241         ei = kmem_cache_alloc(btrfs_inode_cachep, GFP_NOFS);
9242         if (!ei)
9243                 return NULL;
9244
9245         ei->root = NULL;
9246         ei->generation = 0;
9247         ei->last_trans = 0;
9248         ei->last_sub_trans = 0;
9249         ei->logged_trans = 0;
9250         ei->delalloc_bytes = 0;
9251         ei->defrag_bytes = 0;
9252         ei->disk_i_size = 0;
9253         ei->flags = 0;
9254         ei->csum_bytes = 0;
9255         ei->index_cnt = (u64)-1;
9256         ei->dir_index = 0;
9257         ei->last_unlink_trans = 0;
9258         ei->last_log_commit = 0;
9259         ei->delayed_iput_count = 0;
9260
9261         spin_lock_init(&ei->lock);
9262         ei->outstanding_extents = 0;
9263         ei->reserved_extents = 0;
9264
9265         ei->runtime_flags = 0;
9266         ei->force_compress = BTRFS_COMPRESS_NONE;
9267
9268         ei->delayed_node = NULL;
9269
9270         ei->i_otime.tv_sec = 0;
9271         ei->i_otime.tv_nsec = 0;
9272
9273         inode = &ei->vfs_inode;
9274         extent_map_tree_init(&ei->extent_tree);
9275         extent_io_tree_init(&ei->io_tree, &inode->i_data);
9276         extent_io_tree_init(&ei->io_failure_tree, &inode->i_data);
9277         ei->io_tree.track_uptodate = 1;
9278         ei->io_failure_tree.track_uptodate = 1;
9279         atomic_set(&ei->sync_writers, 0);
9280         mutex_init(&ei->log_mutex);
9281         mutex_init(&ei->delalloc_mutex);
9282         btrfs_ordered_inode_tree_init(&ei->ordered_tree);
9283         INIT_LIST_HEAD(&ei->delalloc_inodes);
9284         INIT_LIST_HEAD(&ei->delayed_iput);
9285         RB_CLEAR_NODE(&ei->rb_node);
9286         init_rwsem(&ei->dio_sem);
9287
9288         return inode;
9289 }
9290
9291 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
9292 void btrfs_test_destroy_inode(struct inode *inode)
9293 {
9294         btrfs_drop_extent_cache(BTRFS_I(inode), 0, (u64)-1, 0);
9295         kmem_cache_free(btrfs_inode_cachep, BTRFS_I(inode));
9296 }
9297 #endif
9298
9299 static void btrfs_i_callback(struct rcu_head *head)
9300 {
9301         struct inode *inode = container_of(head, struct inode, i_rcu);
9302         kmem_cache_free(btrfs_inode_cachep, BTRFS_I(inode));
9303 }
9304
9305 void btrfs_destroy_inode(struct inode *inode)
9306 {
9307         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
9308         struct btrfs_ordered_extent *ordered;
9309         struct btrfs_root *root = BTRFS_I(inode)->root;
9310
9311         WARN_ON(!hlist_empty(&inode->i_dentry));
9312         WARN_ON(inode->i_data.nrpages);
9313         WARN_ON(BTRFS_I(inode)->outstanding_extents);
9314         WARN_ON(BTRFS_I(inode)->reserved_extents);
9315         WARN_ON(BTRFS_I(inode)->delalloc_bytes);
9316         WARN_ON(BTRFS_I(inode)->csum_bytes);
9317         WARN_ON(BTRFS_I(inode)->defrag_bytes);
9318
9319         /*
9320          * This can happen where we create an inode, but somebody else also
9321          * created the same inode and we need to destroy the one we already
9322          * created.
9323          */
9324         if (!root)
9325                 goto free;
9326
9327         if (test_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
9328                      &BTRFS_I(inode)->runtime_flags)) {
9329                 btrfs_info(fs_info, "inode %llu still on the orphan list",
9330                            btrfs_ino(BTRFS_I(inode)));
9331                 atomic_dec(&root->orphan_inodes);
9332         }
9333
9334         while (1) {
9335                 ordered = btrfs_lookup_first_ordered_extent(inode, (u64)-1);
9336                 if (!ordered)
9337                         break;
9338                 else {
9339                         btrfs_err(fs_info,
9340                                   "found ordered extent %llu %llu on inode cleanup",
9341                                   ordered->file_offset, ordered->len);
9342                         btrfs_remove_ordered_extent(inode, ordered);
9343                         btrfs_put_ordered_extent(ordered);
9344                         btrfs_put_ordered_extent(ordered);
9345                 }
9346         }
9347         btrfs_qgroup_check_reserved_leak(inode);
9348         inode_tree_del(inode);
9349         btrfs_drop_extent_cache(BTRFS_I(inode), 0, (u64)-1, 0);
9350 free:
9351         call_rcu(&inode->i_rcu, btrfs_i_callback);
9352 }
9353
9354 int btrfs_drop_inode(struct inode *inode)
9355 {
9356         struct btrfs_root *root = BTRFS_I(inode)->root;
9357
9358         if (root == NULL)
9359                 return 1;
9360
9361         /* the snap/subvol tree is on deleting */
9362         if (btrfs_root_refs(&root->root_item) == 0)
9363                 return 1;
9364         else
9365                 return generic_drop_inode(inode);
9366 }
9367
9368 static void init_once(void *foo)
9369 {
9370         struct btrfs_inode *ei = (struct btrfs_inode *) foo;
9371
9372         inode_init_once(&ei->vfs_inode);
9373 }
9374
9375 void btrfs_destroy_cachep(void)
9376 {
9377         /*
9378          * Make sure all delayed rcu free inodes are flushed before we
9379          * destroy cache.
9380          */
9381         rcu_barrier();
9382         kmem_cache_destroy(btrfs_inode_cachep);
9383         kmem_cache_destroy(btrfs_trans_handle_cachep);
9384         kmem_cache_destroy(btrfs_transaction_cachep);
9385         kmem_cache_destroy(btrfs_path_cachep);
9386         kmem_cache_destroy(btrfs_free_space_cachep);
9387 }
9388
9389 int btrfs_init_cachep(void)
9390 {
9391         btrfs_inode_cachep = kmem_cache_create("btrfs_inode",
9392                         sizeof(struct btrfs_inode), 0,
9393                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD | SLAB_ACCOUNT,
9394                         init_once);
9395         if (!btrfs_inode_cachep)
9396                 goto fail;
9397
9398         btrfs_trans_handle_cachep = kmem_cache_create("btrfs_trans_handle",
9399                         sizeof(struct btrfs_trans_handle), 0,
9400                         SLAB_TEMPORARY | SLAB_MEM_SPREAD, NULL);
9401         if (!btrfs_trans_handle_cachep)
9402                 goto fail;
9403
9404         btrfs_transaction_cachep = kmem_cache_create("btrfs_transaction",
9405                         sizeof(struct btrfs_transaction), 0,
9406                         SLAB_TEMPORARY | SLAB_MEM_SPREAD, NULL);
9407         if (!btrfs_transaction_cachep)
9408                 goto fail;
9409
9410         btrfs_path_cachep = kmem_cache_create("btrfs_path",
9411                         sizeof(struct btrfs_path), 0,
9412                         SLAB_MEM_SPREAD, NULL);
9413         if (!btrfs_path_cachep)
9414                 goto fail;
9415
9416         btrfs_free_space_cachep = kmem_cache_create("btrfs_free_space",
9417                         sizeof(struct btrfs_free_space), 0,
9418                         SLAB_MEM_SPREAD, NULL);
9419         if (!btrfs_free_space_cachep)
9420                 goto fail;
9421
9422         return 0;
9423 fail:
9424         btrfs_destroy_cachep();
9425         return -ENOMEM;
9426 }
9427
9428 static int btrfs_getattr(const struct path *path, struct kstat *stat,
9429                          u32 request_mask, unsigned int flags)
9430 {
9431         u64 delalloc_bytes;
9432         struct inode *inode = d_inode(path->dentry);
9433         u32 blocksize = inode->i_sb->s_blocksize;
9434
9435         generic_fillattr(inode, stat);
9436         stat->dev = BTRFS_I(inode)->root->anon_dev;
9437
9438         spin_lock(&BTRFS_I(inode)->lock);
9439         delalloc_bytes = BTRFS_I(inode)->delalloc_bytes;
9440         spin_unlock(&BTRFS_I(inode)->lock);
9441         stat->blocks = (ALIGN(inode_get_bytes(inode), blocksize) +
9442                         ALIGN(delalloc_bytes, blocksize)) >> 9;
9443         return 0;
9444 }
9445
9446 static int btrfs_rename_exchange(struct inode *old_dir,
9447                               struct dentry *old_dentry,
9448                               struct inode *new_dir,
9449                               struct dentry *new_dentry)
9450 {
9451         struct btrfs_fs_info *fs_info = btrfs_sb(old_dir->i_sb);
9452         struct btrfs_trans_handle *trans;
9453         struct btrfs_root *root = BTRFS_I(old_dir)->root;
9454         struct btrfs_root *dest = BTRFS_I(new_dir)->root;
9455         struct inode *new_inode = new_dentry->d_inode;
9456         struct inode *old_inode = old_dentry->d_inode;
9457         struct timespec ctime = current_time(old_inode);
9458         struct dentry *parent;
9459         u64 old_ino = btrfs_ino(BTRFS_I(old_inode));
9460         u64 new_ino = btrfs_ino(BTRFS_I(new_inode));
9461         u64 old_idx = 0;
9462         u64 new_idx = 0;
9463         u64 root_objectid;
9464         int ret;
9465         bool root_log_pinned = false;
9466         bool dest_log_pinned = false;
9467
9468         /* we only allow rename subvolume link between subvolumes */
9469         if (old_ino != BTRFS_FIRST_FREE_OBJECTID && root != dest)
9470                 return -EXDEV;
9471
9472         /* close the race window with snapshot create/destroy ioctl */
9473         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
9474                 down_read(&fs_info->subvol_sem);
9475         if (new_ino == BTRFS_FIRST_FREE_OBJECTID)
9476                 down_read(&fs_info->subvol_sem);
9477
9478         /*
9479          * We want to reserve the absolute worst case amount of items.  So if
9480          * both inodes are subvols and we need to unlink them then that would
9481          * require 4 item modifications, but if they are both normal inodes it
9482          * would require 5 item modifications, so we'll assume their normal
9483          * inodes.  So 5 * 2 is 10, plus 2 for the new links, so 12 total items
9484          * should cover the worst case number of items we'll modify.
9485          */
9486         trans = btrfs_start_transaction(root, 12);
9487         if (IS_ERR(trans)) {
9488                 ret = PTR_ERR(trans);
9489                 goto out_notrans;
9490         }
9491
9492         /*
9493          * We need to find a free sequence number both in the source and
9494          * in the destination directory for the exchange.
9495          */
9496         ret = btrfs_set_inode_index(BTRFS_I(new_dir), &old_idx);
9497         if (ret)
9498                 goto out_fail;
9499         ret = btrfs_set_inode_index(BTRFS_I(old_dir), &new_idx);
9500         if (ret)
9501                 goto out_fail;
9502
9503         BTRFS_I(old_inode)->dir_index = 0ULL;
9504         BTRFS_I(new_inode)->dir_index = 0ULL;
9505
9506         /* Reference for the source. */
9507         if (old_ino == BTRFS_FIRST_FREE_OBJECTID) {
9508                 /* force full log commit if subvolume involved. */
9509                 btrfs_set_log_full_commit(fs_info, trans);
9510         } else {
9511                 btrfs_pin_log_trans(root);
9512                 root_log_pinned = true;
9513                 ret = btrfs_insert_inode_ref(trans, dest,
9514                                              new_dentry->d_name.name,
9515                                              new_dentry->d_name.len,
9516                                              old_ino,
9517                                              btrfs_ino(BTRFS_I(new_dir)),
9518                                              old_idx);
9519                 if (ret)
9520                         goto out_fail;
9521         }
9522
9523         /* And now for the dest. */
9524         if (new_ino == BTRFS_FIRST_FREE_OBJECTID) {
9525                 /* force full log commit if subvolume involved. */
9526                 btrfs_set_log_full_commit(fs_info, trans);
9527         } else {
9528                 btrfs_pin_log_trans(dest);
9529                 dest_log_pinned = true;
9530                 ret = btrfs_insert_inode_ref(trans, root,
9531                                              old_dentry->d_name.name,
9532                                              old_dentry->d_name.len,
9533                                              new_ino,
9534                                              btrfs_ino(BTRFS_I(old_dir)),
9535                                              new_idx);
9536                 if (ret)
9537                         goto out_fail;
9538         }
9539
9540         /* Update inode version and ctime/mtime. */
9541         inode_inc_iversion(old_dir);
9542         inode_inc_iversion(new_dir);
9543         inode_inc_iversion(old_inode);
9544         inode_inc_iversion(new_inode);
9545         old_dir->i_ctime = old_dir->i_mtime = ctime;
9546         new_dir->i_ctime = new_dir->i_mtime = ctime;
9547         old_inode->i_ctime = ctime;
9548         new_inode->i_ctime = ctime;
9549
9550         if (old_dentry->d_parent != new_dentry->d_parent) {
9551                 btrfs_record_unlink_dir(trans, BTRFS_I(old_dir),
9552                                 BTRFS_I(old_inode), 1);
9553                 btrfs_record_unlink_dir(trans, BTRFS_I(new_dir),
9554                                 BTRFS_I(new_inode), 1);
9555         }
9556
9557         /* src is a subvolume */
9558         if (old_ino == BTRFS_FIRST_FREE_OBJECTID) {
9559                 root_objectid = BTRFS_I(old_inode)->root->root_key.objectid;
9560                 ret = btrfs_unlink_subvol(trans, root, old_dir,
9561                                           root_objectid,
9562                                           old_dentry->d_name.name,
9563                                           old_dentry->d_name.len);
9564         } else { /* src is an inode */
9565                 ret = __btrfs_unlink_inode(trans, root, BTRFS_I(old_dir),
9566                                            BTRFS_I(old_dentry->d_inode),
9567                                            old_dentry->d_name.name,
9568                                            old_dentry->d_name.len);
9569                 if (!ret)
9570                         ret = btrfs_update_inode(trans, root, old_inode);
9571         }
9572         if (ret) {
9573                 btrfs_abort_transaction(trans, ret);
9574                 goto out_fail;
9575         }
9576
9577         /* dest is a subvolume */
9578         if (new_ino == BTRFS_FIRST_FREE_OBJECTID) {
9579                 root_objectid = BTRFS_I(new_inode)->root->root_key.objectid;
9580                 ret = btrfs_unlink_subvol(trans, dest, new_dir,
9581                                           root_objectid,
9582                                           new_dentry->d_name.name,
9583                                           new_dentry->d_name.len);
9584         } else { /* dest is an inode */
9585                 ret = __btrfs_unlink_inode(trans, dest, BTRFS_I(new_dir),
9586                                            BTRFS_I(new_dentry->d_inode),
9587                                            new_dentry->d_name.name,
9588                                            new_dentry->d_name.len);
9589                 if (!ret)
9590                         ret = btrfs_update_inode(trans, dest, new_inode);
9591         }
9592         if (ret) {
9593                 btrfs_abort_transaction(trans, ret);
9594                 goto out_fail;
9595         }
9596
9597         ret = btrfs_add_link(trans, BTRFS_I(new_dir), BTRFS_I(old_inode),
9598                              new_dentry->d_name.name,
9599                              new_dentry->d_name.len, 0, old_idx);
9600         if (ret) {
9601                 btrfs_abort_transaction(trans, ret);
9602                 goto out_fail;
9603         }
9604
9605         ret = btrfs_add_link(trans, BTRFS_I(old_dir), BTRFS_I(new_inode),
9606                              old_dentry->d_name.name,
9607                              old_dentry->d_name.len, 0, new_idx);
9608         if (ret) {
9609                 btrfs_abort_transaction(trans, ret);
9610                 goto out_fail;
9611         }
9612
9613         if (old_inode->i_nlink == 1)
9614                 BTRFS_I(old_inode)->dir_index = old_idx;
9615         if (new_inode->i_nlink == 1)
9616                 BTRFS_I(new_inode)->dir_index = new_idx;
9617
9618         if (root_log_pinned) {
9619                 parent = new_dentry->d_parent;
9620                 btrfs_log_new_name(trans, BTRFS_I(old_inode), BTRFS_I(old_dir),
9621                                 parent);
9622                 btrfs_end_log_trans(root);
9623                 root_log_pinned = false;
9624         }
9625         if (dest_log_pinned) {
9626                 parent = old_dentry->d_parent;
9627                 btrfs_log_new_name(trans, BTRFS_I(new_inode), BTRFS_I(new_dir),
9628                                 parent);
9629                 btrfs_end_log_trans(dest);
9630                 dest_log_pinned = false;
9631         }
9632 out_fail:
9633         /*
9634          * If we have pinned a log and an error happened, we unpin tasks
9635          * trying to sync the log and force them to fallback to a transaction
9636          * commit if the log currently contains any of the inodes involved in
9637          * this rename operation (to ensure we do not persist a log with an
9638          * inconsistent state for any of these inodes or leading to any
9639          * inconsistencies when replayed). If the transaction was aborted, the
9640          * abortion reason is propagated to userspace when attempting to commit
9641          * the transaction. If the log does not contain any of these inodes, we
9642          * allow the tasks to sync it.
9643          */
9644         if (ret && (root_log_pinned || dest_log_pinned)) {
9645                 if (btrfs_inode_in_log(BTRFS_I(old_dir), fs_info->generation) ||
9646                     btrfs_inode_in_log(BTRFS_I(new_dir), fs_info->generation) ||
9647                     btrfs_inode_in_log(BTRFS_I(old_inode), fs_info->generation) ||
9648                     (new_inode &&
9649                      btrfs_inode_in_log(BTRFS_I(new_inode), fs_info->generation)))
9650                         btrfs_set_log_full_commit(fs_info, trans);
9651
9652                 if (root_log_pinned) {
9653                         btrfs_end_log_trans(root);
9654                         root_log_pinned = false;
9655                 }
9656                 if (dest_log_pinned) {
9657                         btrfs_end_log_trans(dest);
9658                         dest_log_pinned = false;
9659                 }
9660         }
9661         ret = btrfs_end_transaction(trans);
9662 out_notrans:
9663         if (new_ino == BTRFS_FIRST_FREE_OBJECTID)
9664                 up_read(&fs_info->subvol_sem);
9665         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
9666                 up_read(&fs_info->subvol_sem);
9667
9668         return ret;
9669 }
9670
9671 static int btrfs_whiteout_for_rename(struct btrfs_trans_handle *trans,
9672                                      struct btrfs_root *root,
9673                                      struct inode *dir,
9674                                      struct dentry *dentry)
9675 {
9676         int ret;
9677         struct inode *inode;
9678         u64 objectid;
9679         u64 index;
9680
9681         ret = btrfs_find_free_ino(root, &objectid);
9682         if (ret)
9683                 return ret;
9684
9685         inode = btrfs_new_inode(trans, root, dir,
9686                                 dentry->d_name.name,
9687                                 dentry->d_name.len,
9688                                 btrfs_ino(BTRFS_I(dir)),
9689                                 objectid,
9690                                 S_IFCHR | WHITEOUT_MODE,
9691                                 &index);
9692
9693         if (IS_ERR(inode)) {
9694                 ret = PTR_ERR(inode);
9695                 return ret;
9696         }
9697
9698         inode->i_op = &btrfs_special_inode_operations;
9699         init_special_inode(inode, inode->i_mode,
9700                 WHITEOUT_DEV);
9701
9702         ret = btrfs_init_inode_security(trans, inode, dir,
9703                                 &dentry->d_name);
9704         if (ret)
9705                 goto out;
9706
9707         ret = btrfs_add_nondir(trans, BTRFS_I(dir), dentry,
9708                                 BTRFS_I(inode), 0, index);
9709         if (ret)
9710                 goto out;
9711
9712         ret = btrfs_update_inode(trans, root, inode);
9713 out:
9714         unlock_new_inode(inode);
9715         if (ret)
9716                 inode_dec_link_count(inode);
9717         iput(inode);
9718
9719         return ret;
9720 }
9721
9722 static int btrfs_rename(struct inode *old_dir, struct dentry *old_dentry,
9723                            struct inode *new_dir, struct dentry *new_dentry,
9724                            unsigned int flags)
9725 {
9726         struct btrfs_fs_info *fs_info = btrfs_sb(old_dir->i_sb);
9727         struct btrfs_trans_handle *trans;
9728         unsigned int trans_num_items;
9729         struct btrfs_root *root = BTRFS_I(old_dir)->root;
9730         struct btrfs_root *dest = BTRFS_I(new_dir)->root;
9731         struct inode *new_inode = d_inode(new_dentry);
9732         struct inode *old_inode = d_inode(old_dentry);
9733         u64 index = 0;
9734         u64 root_objectid;
9735         int ret;
9736         u64 old_ino = btrfs_ino(BTRFS_I(old_inode));
9737         bool log_pinned = false;
9738
9739         if (btrfs_ino(BTRFS_I(new_dir)) == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)
9740                 return -EPERM;
9741
9742         /* we only allow rename subvolume link between subvolumes */
9743         if (old_ino != BTRFS_FIRST_FREE_OBJECTID && root != dest)
9744                 return -EXDEV;
9745
9746         if (old_ino == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID ||
9747             (new_inode && btrfs_ino(BTRFS_I(new_inode)) == BTRFS_FIRST_FREE_OBJECTID))
9748                 return -ENOTEMPTY;
9749
9750         if (S_ISDIR(old_inode->i_mode) && new_inode &&
9751             new_inode->i_size > BTRFS_EMPTY_DIR_SIZE)
9752                 return -ENOTEMPTY;
9753
9754
9755         /* check for collisions, even if the  name isn't there */
9756         ret = btrfs_check_dir_item_collision(dest, new_dir->i_ino,
9757                              new_dentry->d_name.name,
9758                              new_dentry->d_name.len);
9759
9760         if (ret) {
9761                 if (ret == -EEXIST) {
9762                         /* we shouldn't get
9763                          * eexist without a new_inode */
9764                         if (WARN_ON(!new_inode)) {
9765                                 return ret;
9766                         }
9767                 } else {
9768                         /* maybe -EOVERFLOW */
9769                         return ret;
9770                 }
9771         }
9772         ret = 0;
9773
9774         /*
9775          * we're using rename to replace one file with another.  Start IO on it
9776          * now so  we don't add too much work to the end of the transaction
9777          */
9778         if (new_inode && S_ISREG(old_inode->i_mode) && new_inode->i_size)
9779                 filemap_flush(old_inode->i_mapping);
9780
9781         /* close the racy window with snapshot create/destroy ioctl */
9782         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
9783                 down_read(&fs_info->subvol_sem);
9784         /*
9785          * We want to reserve the absolute worst case amount of items.  So if
9786          * both inodes are subvols and we need to unlink them then that would
9787          * require 4 item modifications, but if they are both normal inodes it
9788          * would require 5 item modifications, so we'll assume they are normal
9789          * inodes.  So 5 * 2 is 10, plus 1 for the new link, so 11 total items
9790          * should cover the worst case number of items we'll modify.
9791          * If our rename has the whiteout flag, we need more 5 units for the
9792          * new inode (1 inode item, 1 inode ref, 2 dir items and 1 xattr item
9793          * when selinux is enabled).
9794          */
9795         trans_num_items = 11;
9796         if (flags & RENAME_WHITEOUT)
9797                 trans_num_items += 5;
9798         trans = btrfs_start_transaction(root, trans_num_items);
9799         if (IS_ERR(trans)) {
9800                 ret = PTR_ERR(trans);
9801                 goto out_notrans;
9802         }
9803
9804         if (dest != root)
9805                 btrfs_record_root_in_trans(trans, dest);
9806
9807         ret = btrfs_set_inode_index(BTRFS_I(new_dir), &index);
9808         if (ret)
9809                 goto out_fail;
9810
9811         BTRFS_I(old_inode)->dir_index = 0ULL;
9812         if (unlikely(old_ino == BTRFS_FIRST_FREE_OBJECTID)) {
9813                 /* force full log commit if subvolume involved. */
9814                 btrfs_set_log_full_commit(fs_info, trans);
9815         } else {
9816                 btrfs_pin_log_trans(root);
9817                 log_pinned = true;
9818                 ret = btrfs_insert_inode_ref(trans, dest,
9819                                              new_dentry->d_name.name,
9820                                              new_dentry->d_name.len,
9821                                              old_ino,
9822                                              btrfs_ino(BTRFS_I(new_dir)), index);
9823                 if (ret)
9824                         goto out_fail;
9825         }
9826
9827         inode_inc_iversion(old_dir);
9828         inode_inc_iversion(new_dir);
9829         inode_inc_iversion(old_inode);
9830         old_dir->i_ctime = old_dir->i_mtime =
9831         new_dir->i_ctime = new_dir->i_mtime =
9832         old_inode->i_ctime = current_time(old_dir);
9833
9834         if (old_dentry->d_parent != new_dentry->d_parent)
9835                 btrfs_record_unlink_dir(trans, BTRFS_I(old_dir),
9836                                 BTRFS_I(old_inode), 1);
9837
9838         if (unlikely(old_ino == BTRFS_FIRST_FREE_OBJECTID)) {
9839                 root_objectid = BTRFS_I(old_inode)->root->root_key.objectid;
9840                 ret = btrfs_unlink_subvol(trans, root, old_dir, root_objectid,
9841                                         old_dentry->d_name.name,
9842                                         old_dentry->d_name.len);
9843         } else {
9844                 ret = __btrfs_unlink_inode(trans, root, BTRFS_I(old_dir),
9845                                         BTRFS_I(d_inode(old_dentry)),
9846                                         old_dentry->d_name.name,
9847                                         old_dentry->d_name.len);
9848                 if (!ret)
9849                         ret = btrfs_update_inode(trans, root, old_inode);
9850         }
9851         if (ret) {
9852                 btrfs_abort_transaction(trans, ret);
9853                 goto out_fail;
9854         }
9855
9856         if (new_inode) {
9857                 inode_inc_iversion(new_inode);
9858                 new_inode->i_ctime = current_time(new_inode);
9859                 if (unlikely(btrfs_ino(BTRFS_I(new_inode)) ==
9860                              BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)) {
9861                         root_objectid = BTRFS_I(new_inode)->location.objectid;
9862                         ret = btrfs_unlink_subvol(trans, dest, new_dir,
9863                                                 root_objectid,
9864                                                 new_dentry->d_name.name,
9865                                                 new_dentry->d_name.len);
9866                         BUG_ON(new_inode->i_nlink == 0);
9867                 } else {
9868                         ret = btrfs_unlink_inode(trans, dest, BTRFS_I(new_dir),
9869                                                  BTRFS_I(d_inode(new_dentry)),
9870                                                  new_dentry->d_name.name,
9871                                                  new_dentry->d_name.len);
9872                 }
9873                 if (!ret && new_inode->i_nlink == 0)
9874                         ret = btrfs_orphan_add(trans,
9875                                         BTRFS_I(d_inode(new_dentry)));
9876                 if (ret) {
9877                         btrfs_abort_transaction(trans, ret);
9878                         goto out_fail;
9879                 }
9880         }
9881
9882         ret = btrfs_add_link(trans, BTRFS_I(new_dir), BTRFS_I(old_inode),
9883                              new_dentry->d_name.name,
9884                              new_dentry->d_name.len, 0, index);
9885         if (ret) {
9886                 btrfs_abort_transaction(trans, ret);
9887                 goto out_fail;
9888         }
9889
9890         if (old_inode->i_nlink == 1)
9891                 BTRFS_I(old_inode)->dir_index = index;
9892
9893         if (log_pinned) {
9894                 struct dentry *parent = new_dentry->d_parent;
9895
9896                 btrfs_log_new_name(trans, BTRFS_I(old_inode), BTRFS_I(old_dir),
9897                                 parent);
9898                 btrfs_end_log_trans(root);
9899                 log_pinned = false;
9900         }
9901
9902         if (flags & RENAME_WHITEOUT) {
9903                 ret = btrfs_whiteout_for_rename(trans, root, old_dir,
9904                                                 old_dentry);
9905
9906                 if (ret) {
9907                         btrfs_abort_transaction(trans, ret);
9908                         goto out_fail;
9909                 }
9910         }
9911 out_fail:
9912         /*
9913          * If we have pinned the log and an error happened, we unpin tasks
9914          * trying to sync the log and force them to fallback to a transaction
9915          * commit if the log currently contains any of the inodes involved in
9916          * this rename operation (to ensure we do not persist a log with an
9917          * inconsistent state for any of these inodes or leading to any
9918          * inconsistencies when replayed). If the transaction was aborted, the
9919          * abortion reason is propagated to userspace when attempting to commit
9920          * the transaction. If the log does not contain any of these inodes, we
9921          * allow the tasks to sync it.
9922          */
9923         if (ret && log_pinned) {
9924                 if (btrfs_inode_in_log(BTRFS_I(old_dir), fs_info->generation) ||
9925                     btrfs_inode_in_log(BTRFS_I(new_dir), fs_info->generation) ||
9926                     btrfs_inode_in_log(BTRFS_I(old_inode), fs_info->generation) ||
9927                     (new_inode &&
9928                      btrfs_inode_in_log(BTRFS_I(new_inode), fs_info->generation)))
9929                         btrfs_set_log_full_commit(fs_info, trans);
9930
9931                 btrfs_end_log_trans(root);
9932                 log_pinned = false;
9933         }
9934         btrfs_end_transaction(trans);
9935 out_notrans:
9936         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
9937                 up_read(&fs_info->subvol_sem);
9938
9939         return ret;
9940 }
9941
9942 static int btrfs_rename2(struct inode *old_dir, struct dentry *old_dentry,
9943                          struct inode *new_dir, struct dentry *new_dentry,
9944                          unsigned int flags)
9945 {
9946         if (flags & ~(RENAME_NOREPLACE | RENAME_EXCHANGE | RENAME_WHITEOUT))
9947                 return -EINVAL;
9948
9949         if (flags & RENAME_EXCHANGE)
9950                 return btrfs_rename_exchange(old_dir, old_dentry, new_dir,
9951                                           new_dentry);
9952
9953         return btrfs_rename(old_dir, old_dentry, new_dir, new_dentry, flags);
9954 }
9955
9956 static void btrfs_run_delalloc_work(struct btrfs_work *work)
9957 {
9958         struct btrfs_delalloc_work *delalloc_work;
9959         struct inode *inode;
9960
9961         delalloc_work = container_of(work, struct btrfs_delalloc_work,
9962                                      work);
9963         inode = delalloc_work->inode;
9964         filemap_flush(inode->i_mapping);
9965         if (test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
9966                                 &BTRFS_I(inode)->runtime_flags))
9967                 filemap_flush(inode->i_mapping);
9968
9969         if (delalloc_work->delay_iput)
9970                 btrfs_add_delayed_iput(inode);
9971         else
9972                 iput(inode);
9973         complete(&delalloc_work->completion);
9974 }
9975
9976 struct btrfs_delalloc_work *btrfs_alloc_delalloc_work(struct inode *inode,
9977                                                     int delay_iput)
9978 {
9979         struct btrfs_delalloc_work *work;
9980
9981         work = kmalloc(sizeof(*work), GFP_NOFS);
9982         if (!work)
9983                 return NULL;
9984
9985         init_completion(&work->completion);
9986         INIT_LIST_HEAD(&work->list);
9987         work->inode = inode;
9988         work->delay_iput = delay_iput;
9989         WARN_ON_ONCE(!inode);
9990         btrfs_init_work(&work->work, btrfs_flush_delalloc_helper,
9991                         btrfs_run_delalloc_work, NULL, NULL);
9992
9993         return work;
9994 }
9995
9996 void btrfs_wait_and_free_delalloc_work(struct btrfs_delalloc_work *work)
9997 {
9998         wait_for_completion(&work->completion);
9999         kfree(work);
10000 }
10001
10002 /*
10003  * some fairly slow code that needs optimization. This walks the list
10004  * of all the inodes with pending delalloc and forces them to disk.
10005  */
10006 static int __start_delalloc_inodes(struct btrfs_root *root, int delay_iput,
10007                                    int nr)
10008 {
10009         struct btrfs_inode *binode;
10010         struct inode *inode;
10011         struct btrfs_delalloc_work *work, *next;
10012         struct list_head works;
10013         struct list_head splice;
10014         int ret = 0;
10015
10016         INIT_LIST_HEAD(&works);
10017         INIT_LIST_HEAD(&splice);
10018
10019         mutex_lock(&root->delalloc_mutex);
10020         spin_lock(&root->delalloc_lock);
10021         list_splice_init(&root->delalloc_inodes, &splice);
10022         while (!list_empty(&splice)) {
10023                 binode = list_entry(splice.next, struct btrfs_inode,
10024                                     delalloc_inodes);
10025
10026                 list_move_tail(&binode->delalloc_inodes,
10027                                &root->delalloc_inodes);
10028                 inode = igrab(&binode->vfs_inode);
10029                 if (!inode) {
10030                         cond_resched_lock(&root->delalloc_lock);
10031                         continue;
10032                 }
10033                 spin_unlock(&root->delalloc_lock);
10034
10035                 work = btrfs_alloc_delalloc_work(inode, delay_iput);
10036                 if (!work) {
10037                         if (delay_iput)
10038                                 btrfs_add_delayed_iput(inode);
10039                         else
10040                                 iput(inode);
10041                         ret = -ENOMEM;
10042                         goto out;
10043                 }
10044                 list_add_tail(&work->list, &works);
10045                 btrfs_queue_work(root->fs_info->flush_workers,
10046                                  &work->work);
10047                 ret++;
10048                 if (nr != -1 && ret >= nr)
10049                         goto out;
10050                 cond_resched();
10051                 spin_lock(&root->delalloc_lock);
10052         }
10053         spin_unlock(&root->delalloc_lock);
10054
10055 out:
10056         list_for_each_entry_safe(work, next, &works, list) {
10057                 list_del_init(&work->list);
10058                 btrfs_wait_and_free_delalloc_work(work);
10059         }
10060
10061         if (!list_empty_careful(&splice)) {
10062                 spin_lock(&root->delalloc_lock);
10063                 list_splice_tail(&splice, &root->delalloc_inodes);
10064                 spin_unlock(&root->delalloc_lock);
10065         }
10066         mutex_unlock(&root->delalloc_mutex);
10067         return ret;
10068 }
10069
10070 int btrfs_start_delalloc_inodes(struct btrfs_root *root, int delay_iput)
10071 {
10072         struct btrfs_fs_info *fs_info = root->fs_info;
10073         int ret;
10074
10075         if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state))
10076                 return -EROFS;
10077
10078         ret = __start_delalloc_inodes(root, delay_iput, -1);
10079         if (ret > 0)
10080                 ret = 0;
10081         /*
10082          * the filemap_flush will queue IO into the worker threads, but
10083          * we have to make sure the IO is actually started and that
10084          * ordered extents get created before we return
10085          */
10086         atomic_inc(&fs_info->async_submit_draining);
10087         while (atomic_read(&fs_info->nr_async_submits) ||
10088                atomic_read(&fs_info->async_delalloc_pages)) {
10089                 wait_event(fs_info->async_submit_wait,
10090                            (atomic_read(&fs_info->nr_async_submits) == 0 &&
10091                             atomic_read(&fs_info->async_delalloc_pages) == 0));
10092         }
10093         atomic_dec(&fs_info->async_submit_draining);
10094         return ret;
10095 }
10096
10097 int btrfs_start_delalloc_roots(struct btrfs_fs_info *fs_info, int delay_iput,
10098                                int nr)
10099 {
10100         struct btrfs_root *root;
10101         struct list_head splice;
10102         int ret;
10103
10104         if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state))
10105                 return -EROFS;
10106
10107         INIT_LIST_HEAD(&splice);
10108
10109         mutex_lock(&fs_info->delalloc_root_mutex);
10110         spin_lock(&fs_info->delalloc_root_lock);
10111         list_splice_init(&fs_info->delalloc_roots, &splice);
10112         while (!list_empty(&splice) && nr) {
10113                 root = list_first_entry(&splice, struct btrfs_root,
10114                                         delalloc_root);
10115                 root = btrfs_grab_fs_root(root);
10116                 BUG_ON(!root);
10117                 list_move_tail(&root->delalloc_root,
10118                                &fs_info->delalloc_roots);
10119                 spin_unlock(&fs_info->delalloc_root_lock);
10120
10121                 ret = __start_delalloc_inodes(root, delay_iput, nr);
10122                 btrfs_put_fs_root(root);
10123                 if (ret < 0)
10124                         goto out;
10125
10126                 if (nr != -1) {
10127                         nr -= ret;
10128                         WARN_ON(nr < 0);
10129                 }
10130                 spin_lock(&fs_info->delalloc_root_lock);
10131         }
10132         spin_unlock(&fs_info->delalloc_root_lock);
10133
10134         ret = 0;
10135         atomic_inc(&fs_info->async_submit_draining);
10136         while (atomic_read(&fs_info->nr_async_submits) ||
10137               atomic_read(&fs_info->async_delalloc_pages)) {
10138                 wait_event(fs_info->async_submit_wait,
10139                    (atomic_read(&fs_info->nr_async_submits) == 0 &&
10140                     atomic_read(&fs_info->async_delalloc_pages) == 0));
10141         }
10142         atomic_dec(&fs_info->async_submit_draining);
10143 out:
10144         if (!list_empty_careful(&splice)) {
10145                 spin_lock(&fs_info->delalloc_root_lock);
10146                 list_splice_tail(&splice, &fs_info->delalloc_roots);
10147                 spin_unlock(&fs_info->delalloc_root_lock);
10148         }
10149         mutex_unlock(&fs_info->delalloc_root_mutex);
10150         return ret;
10151 }
10152
10153 static int btrfs_symlink(struct inode *dir, struct dentry *dentry,
10154                          const char *symname)
10155 {
10156         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
10157         struct btrfs_trans_handle *trans;
10158         struct btrfs_root *root = BTRFS_I(dir)->root;
10159         struct btrfs_path *path;
10160         struct btrfs_key key;
10161         struct inode *inode = NULL;
10162         int err;
10163         int drop_inode = 0;
10164         u64 objectid;
10165         u64 index = 0;
10166         int name_len;
10167         int datasize;
10168         unsigned long ptr;
10169         struct btrfs_file_extent_item *ei;
10170         struct extent_buffer *leaf;
10171
10172         name_len = strlen(symname);
10173         if (name_len > BTRFS_MAX_INLINE_DATA_SIZE(fs_info))
10174                 return -ENAMETOOLONG;
10175
10176         /*
10177          * 2 items for inode item and ref
10178          * 2 items for dir items
10179          * 1 item for updating parent inode item
10180          * 1 item for the inline extent item
10181          * 1 item for xattr if selinux is on
10182          */
10183         trans = btrfs_start_transaction(root, 7);
10184         if (IS_ERR(trans))
10185                 return PTR_ERR(trans);
10186
10187         err = btrfs_find_free_ino(root, &objectid);
10188         if (err)
10189                 goto out_unlock;
10190
10191         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
10192                                 dentry->d_name.len, btrfs_ino(BTRFS_I(dir)),
10193                                 objectid, S_IFLNK|S_IRWXUGO, &index);
10194         if (IS_ERR(inode)) {
10195                 err = PTR_ERR(inode);
10196                 goto out_unlock;
10197         }
10198
10199         /*
10200         * If the active LSM wants to access the inode during
10201         * d_instantiate it needs these. Smack checks to see
10202         * if the filesystem supports xattrs by looking at the
10203         * ops vector.
10204         */
10205         inode->i_fop = &btrfs_file_operations;
10206         inode->i_op = &btrfs_file_inode_operations;
10207         inode->i_mapping->a_ops = &btrfs_aops;
10208         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
10209
10210         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
10211         if (err)
10212                 goto out_unlock_inode;
10213
10214         path = btrfs_alloc_path();
10215         if (!path) {
10216                 err = -ENOMEM;
10217                 goto out_unlock_inode;
10218         }
10219         key.objectid = btrfs_ino(BTRFS_I(inode));
10220         key.offset = 0;
10221         key.type = BTRFS_EXTENT_DATA_KEY;
10222         datasize = btrfs_file_extent_calc_inline_size(name_len);
10223         err = btrfs_insert_empty_item(trans, root, path, &key,
10224                                       datasize);
10225         if (err) {
10226                 btrfs_free_path(path);
10227                 goto out_unlock_inode;
10228         }
10229         leaf = path->nodes[0];
10230         ei = btrfs_item_ptr(leaf, path->slots[0],
10231                             struct btrfs_file_extent_item);
10232         btrfs_set_file_extent_generation(leaf, ei, trans->transid);
10233         btrfs_set_file_extent_type(leaf, ei,
10234                                    BTRFS_FILE_EXTENT_INLINE);
10235         btrfs_set_file_extent_encryption(leaf, ei, 0);
10236         btrfs_set_file_extent_compression(leaf, ei, 0);
10237         btrfs_set_file_extent_other_encoding(leaf, ei, 0);
10238         btrfs_set_file_extent_ram_bytes(leaf, ei, name_len);
10239
10240         ptr = btrfs_file_extent_inline_start(ei);
10241         write_extent_buffer(leaf, symname, ptr, name_len);
10242         btrfs_mark_buffer_dirty(leaf);
10243         btrfs_free_path(path);
10244
10245         inode->i_op = &btrfs_symlink_inode_operations;
10246         inode_nohighmem(inode);
10247         inode->i_mapping->a_ops = &btrfs_symlink_aops;
10248         inode_set_bytes(inode, name_len);
10249         btrfs_i_size_write(BTRFS_I(inode), name_len);
10250         err = btrfs_update_inode(trans, root, inode);
10251         /*
10252          * Last step, add directory indexes for our symlink inode. This is the
10253          * last step to avoid extra cleanup of these indexes if an error happens
10254          * elsewhere above.
10255          */
10256         if (!err)
10257                 err = btrfs_add_nondir(trans, BTRFS_I(dir), dentry,
10258                                 BTRFS_I(inode), 0, index);
10259         if (err) {
10260                 drop_inode = 1;
10261                 goto out_unlock_inode;
10262         }
10263
10264         unlock_new_inode(inode);
10265         d_instantiate(dentry, inode);
10266
10267 out_unlock:
10268         btrfs_end_transaction(trans);
10269         if (drop_inode) {
10270                 inode_dec_link_count(inode);
10271                 iput(inode);
10272         }
10273         btrfs_btree_balance_dirty(fs_info);
10274         return err;
10275
10276 out_unlock_inode:
10277         drop_inode = 1;
10278         unlock_new_inode(inode);
10279         goto out_unlock;
10280 }
10281
10282 static int __btrfs_prealloc_file_range(struct inode *inode, int mode,
10283                                        u64 start, u64 num_bytes, u64 min_size,
10284                                        loff_t actual_len, u64 *alloc_hint,
10285                                        struct btrfs_trans_handle *trans)
10286 {
10287         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
10288         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
10289         struct extent_map *em;
10290         struct btrfs_root *root = BTRFS_I(inode)->root;
10291         struct btrfs_key ins;
10292         u64 cur_offset = start;
10293         u64 i_size;
10294         u64 cur_bytes;
10295         u64 last_alloc = (u64)-1;
10296         int ret = 0;
10297         bool own_trans = true;
10298         u64 end = start + num_bytes - 1;
10299
10300         if (trans)
10301                 own_trans = false;
10302         while (num_bytes > 0) {
10303                 if (own_trans) {
10304                         trans = btrfs_start_transaction(root, 3);
10305                         if (IS_ERR(trans)) {
10306                                 ret = PTR_ERR(trans);
10307                                 break;
10308                         }
10309                 }
10310
10311                 cur_bytes = min_t(u64, num_bytes, SZ_256M);
10312                 cur_bytes = max(cur_bytes, min_size);
10313                 /*
10314                  * If we are severely fragmented we could end up with really
10315                  * small allocations, so if the allocator is returning small
10316                  * chunks lets make its job easier by only searching for those
10317                  * sized chunks.
10318                  */
10319                 cur_bytes = min(cur_bytes, last_alloc);
10320                 ret = btrfs_reserve_extent(root, cur_bytes, cur_bytes,
10321                                 min_size, 0, *alloc_hint, &ins, 1, 0);
10322                 if (ret) {
10323                         if (own_trans)
10324                                 btrfs_end_transaction(trans);
10325                         break;
10326                 }
10327                 btrfs_dec_block_group_reservations(fs_info, ins.objectid);
10328
10329                 last_alloc = ins.offset;
10330                 ret = insert_reserved_file_extent(trans, inode,
10331                                                   cur_offset, ins.objectid,
10332                                                   ins.offset, ins.offset,
10333                                                   ins.offset, 0, 0, 0,
10334                                                   BTRFS_FILE_EXTENT_PREALLOC);
10335                 if (ret) {
10336                         btrfs_free_reserved_extent(fs_info, ins.objectid,
10337                                                    ins.offset, 0);
10338                         btrfs_abort_transaction(trans, ret);
10339                         if (own_trans)
10340                                 btrfs_end_transaction(trans);
10341                         break;
10342                 }
10343
10344                 btrfs_drop_extent_cache(BTRFS_I(inode), cur_offset,
10345                                         cur_offset + ins.offset -1, 0);
10346
10347                 em = alloc_extent_map();
10348                 if (!em) {
10349                         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
10350                                 &BTRFS_I(inode)->runtime_flags);
10351                         goto next;
10352                 }
10353
10354                 em->start = cur_offset;
10355                 em->orig_start = cur_offset;
10356                 em->len = ins.offset;
10357                 em->block_start = ins.objectid;
10358                 em->block_len = ins.offset;
10359                 em->orig_block_len = ins.offset;
10360                 em->ram_bytes = ins.offset;
10361                 em->bdev = fs_info->fs_devices->latest_bdev;
10362                 set_bit(EXTENT_FLAG_PREALLOC, &em->flags);
10363                 em->generation = trans->transid;
10364
10365                 while (1) {
10366                         write_lock(&em_tree->lock);
10367                         ret = add_extent_mapping(em_tree, em, 1);
10368                         write_unlock(&em_tree->lock);
10369                         if (ret != -EEXIST)
10370                                 break;
10371                         btrfs_drop_extent_cache(BTRFS_I(inode), cur_offset,
10372                                                 cur_offset + ins.offset - 1,
10373                                                 0);
10374                 }
10375                 free_extent_map(em);
10376 next:
10377                 num_bytes -= ins.offset;
10378                 cur_offset += ins.offset;
10379                 *alloc_hint = ins.objectid + ins.offset;
10380
10381                 inode_inc_iversion(inode);
10382                 inode->i_ctime = current_time(inode);
10383                 BTRFS_I(inode)->flags |= BTRFS_INODE_PREALLOC;
10384                 if (!(mode & FALLOC_FL_KEEP_SIZE) &&
10385                     (actual_len > inode->i_size) &&
10386                     (cur_offset > inode->i_size)) {
10387                         if (cur_offset > actual_len)
10388                                 i_size = actual_len;
10389                         else
10390                                 i_size = cur_offset;
10391                         i_size_write(inode, i_size);
10392                         btrfs_ordered_update_i_size(inode, i_size, NULL);
10393                 }
10394
10395                 ret = btrfs_update_inode(trans, root, inode);
10396
10397                 if (ret) {
10398                         btrfs_abort_transaction(trans, ret);
10399                         if (own_trans)
10400                                 btrfs_end_transaction(trans);
10401                         break;
10402                 }
10403
10404                 if (own_trans)
10405                         btrfs_end_transaction(trans);
10406         }
10407         if (cur_offset < end)
10408                 btrfs_free_reserved_data_space(inode, cur_offset,
10409                         end - cur_offset + 1);
10410         return ret;
10411 }
10412
10413 int btrfs_prealloc_file_range(struct inode *inode, int mode,
10414                               u64 start, u64 num_bytes, u64 min_size,
10415                               loff_t actual_len, u64 *alloc_hint)
10416 {
10417         return __btrfs_prealloc_file_range(inode, mode, start, num_bytes,
10418                                            min_size, actual_len, alloc_hint,
10419                                            NULL);
10420 }
10421
10422 int btrfs_prealloc_file_range_trans(struct inode *inode,
10423                                     struct btrfs_trans_handle *trans, int mode,
10424                                     u64 start, u64 num_bytes, u64 min_size,
10425                                     loff_t actual_len, u64 *alloc_hint)
10426 {
10427         return __btrfs_prealloc_file_range(inode, mode, start, num_bytes,
10428                                            min_size, actual_len, alloc_hint, trans);
10429 }
10430
10431 static int btrfs_set_page_dirty(struct page *page)
10432 {
10433         return __set_page_dirty_nobuffers(page);
10434 }
10435
10436 static int btrfs_permission(struct inode *inode, int mask)
10437 {
10438         struct btrfs_root *root = BTRFS_I(inode)->root;
10439         umode_t mode = inode->i_mode;
10440
10441         if (mask & MAY_WRITE &&
10442             (S_ISREG(mode) || S_ISDIR(mode) || S_ISLNK(mode))) {
10443                 if (btrfs_root_readonly(root))
10444                         return -EROFS;
10445                 if (BTRFS_I(inode)->flags & BTRFS_INODE_READONLY)
10446                         return -EACCES;
10447         }
10448         return generic_permission(inode, mask);
10449 }
10450
10451 static int btrfs_tmpfile(struct inode *dir, struct dentry *dentry, umode_t mode)
10452 {
10453         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
10454         struct btrfs_trans_handle *trans;
10455         struct btrfs_root *root = BTRFS_I(dir)->root;
10456         struct inode *inode = NULL;
10457         u64 objectid;
10458         u64 index;
10459         int ret = 0;
10460
10461         /*
10462          * 5 units required for adding orphan entry
10463          */
10464         trans = btrfs_start_transaction(root, 5);
10465         if (IS_ERR(trans))
10466                 return PTR_ERR(trans);
10467
10468         ret = btrfs_find_free_ino(root, &objectid);
10469         if (ret)
10470                 goto out;
10471
10472         inode = btrfs_new_inode(trans, root, dir, NULL, 0,
10473                         btrfs_ino(BTRFS_I(dir)), objectid, mode, &index);
10474         if (IS_ERR(inode)) {
10475                 ret = PTR_ERR(inode);
10476                 inode = NULL;
10477                 goto out;
10478         }
10479
10480         inode->i_fop = &btrfs_file_operations;
10481         inode->i_op = &btrfs_file_inode_operations;
10482
10483         inode->i_mapping->a_ops = &btrfs_aops;
10484         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
10485
10486         ret = btrfs_init_inode_security(trans, inode, dir, NULL);
10487         if (ret)
10488                 goto out_inode;
10489
10490         ret = btrfs_update_inode(trans, root, inode);
10491         if (ret)
10492                 goto out_inode;
10493         ret = btrfs_orphan_add(trans, BTRFS_I(inode));
10494         if (ret)
10495                 goto out_inode;
10496
10497         /*
10498          * We set number of links to 0 in btrfs_new_inode(), and here we set
10499          * it to 1 because d_tmpfile() will issue a warning if the count is 0,
10500          * through:
10501          *
10502          *    d_tmpfile() -> inode_dec_link_count() -> drop_nlink()
10503          */
10504         set_nlink(inode, 1);
10505         unlock_new_inode(inode);
10506         d_tmpfile(dentry, inode);
10507         mark_inode_dirty(inode);
10508
10509 out:
10510         btrfs_end_transaction(trans);
10511         if (ret)
10512                 iput(inode);
10513         btrfs_balance_delayed_items(fs_info);
10514         btrfs_btree_balance_dirty(fs_info);
10515         return ret;
10516
10517 out_inode:
10518         unlock_new_inode(inode);
10519         goto out;
10520
10521 }
10522
10523 __attribute__((const))
10524 static int btrfs_readpage_io_failed_hook(struct page *page, int failed_mirror)
10525 {
10526         return -EAGAIN;
10527 }
10528
10529 static const struct inode_operations btrfs_dir_inode_operations = {
10530         .getattr        = btrfs_getattr,
10531         .lookup         = btrfs_lookup,
10532         .create         = btrfs_create,
10533         .unlink         = btrfs_unlink,
10534         .link           = btrfs_link,
10535         .mkdir          = btrfs_mkdir,
10536         .rmdir          = btrfs_rmdir,
10537         .rename         = btrfs_rename2,
10538         .symlink        = btrfs_symlink,
10539         .setattr        = btrfs_setattr,
10540         .mknod          = btrfs_mknod,
10541         .listxattr      = btrfs_listxattr,
10542         .permission     = btrfs_permission,
10543         .get_acl        = btrfs_get_acl,
10544         .set_acl        = btrfs_set_acl,
10545         .update_time    = btrfs_update_time,
10546         .tmpfile        = btrfs_tmpfile,
10547 };
10548 static const struct inode_operations btrfs_dir_ro_inode_operations = {
10549         .lookup         = btrfs_lookup,
10550         .permission     = btrfs_permission,
10551         .update_time    = btrfs_update_time,
10552 };
10553
10554 static const struct file_operations btrfs_dir_file_operations = {
10555         .llseek         = generic_file_llseek,
10556         .read           = generic_read_dir,
10557         .iterate_shared = btrfs_real_readdir,
10558         .unlocked_ioctl = btrfs_ioctl,
10559 #ifdef CONFIG_COMPAT
10560         .compat_ioctl   = btrfs_compat_ioctl,
10561 #endif
10562         .release        = btrfs_release_file,
10563         .fsync          = btrfs_sync_file,
10564 };
10565
10566 static const struct extent_io_ops btrfs_extent_io_ops = {
10567         /* mandatory callbacks */
10568         .submit_bio_hook = btrfs_submit_bio_hook,
10569         .readpage_end_io_hook = btrfs_readpage_end_io_hook,
10570         .merge_bio_hook = btrfs_merge_bio_hook,
10571         .readpage_io_failed_hook = btrfs_readpage_io_failed_hook,
10572
10573         /* optional callbacks */
10574         .fill_delalloc = run_delalloc_range,
10575         .writepage_end_io_hook = btrfs_writepage_end_io_hook,
10576         .writepage_start_hook = btrfs_writepage_start_hook,
10577         .set_bit_hook = btrfs_set_bit_hook,
10578         .clear_bit_hook = btrfs_clear_bit_hook,
10579         .merge_extent_hook = btrfs_merge_extent_hook,
10580         .split_extent_hook = btrfs_split_extent_hook,
10581 };
10582
10583 /*
10584  * btrfs doesn't support the bmap operation because swapfiles
10585  * use bmap to make a mapping of extents in the file.  They assume
10586  * these extents won't change over the life of the file and they
10587  * use the bmap result to do IO directly to the drive.
10588  *
10589  * the btrfs bmap call would return logical addresses that aren't
10590  * suitable for IO and they also will change frequently as COW
10591  * operations happen.  So, swapfile + btrfs == corruption.
10592  *
10593  * For now we're avoiding this by dropping bmap.
10594  */
10595 static const struct address_space_operations btrfs_aops = {
10596         .readpage       = btrfs_readpage,
10597         .writepage      = btrfs_writepage,
10598         .writepages     = btrfs_writepages,
10599         .readpages      = btrfs_readpages,
10600         .direct_IO      = btrfs_direct_IO,
10601         .invalidatepage = btrfs_invalidatepage,
10602         .releasepage    = btrfs_releasepage,
10603         .set_page_dirty = btrfs_set_page_dirty,
10604         .error_remove_page = generic_error_remove_page,
10605 };
10606
10607 static const struct address_space_operations btrfs_symlink_aops = {
10608         .readpage       = btrfs_readpage,
10609         .writepage      = btrfs_writepage,
10610         .invalidatepage = btrfs_invalidatepage,
10611         .releasepage    = btrfs_releasepage,
10612 };
10613
10614 static const struct inode_operations btrfs_file_inode_operations = {
10615         .getattr        = btrfs_getattr,
10616         .setattr        = btrfs_setattr,
10617         .listxattr      = btrfs_listxattr,
10618         .permission     = btrfs_permission,
10619         .fiemap         = btrfs_fiemap,
10620         .get_acl        = btrfs_get_acl,
10621         .set_acl        = btrfs_set_acl,
10622         .update_time    = btrfs_update_time,
10623 };
10624 static const struct inode_operations btrfs_special_inode_operations = {
10625         .getattr        = btrfs_getattr,
10626         .setattr        = btrfs_setattr,
10627         .permission     = btrfs_permission,
10628         .listxattr      = btrfs_listxattr,
10629         .get_acl        = btrfs_get_acl,
10630         .set_acl        = btrfs_set_acl,
10631         .update_time    = btrfs_update_time,
10632 };
10633 static const struct inode_operations btrfs_symlink_inode_operations = {
10634         .get_link       = page_get_link,
10635         .getattr        = btrfs_getattr,
10636         .setattr        = btrfs_setattr,
10637         .permission     = btrfs_permission,
10638         .listxattr      = btrfs_listxattr,
10639         .update_time    = btrfs_update_time,
10640 };
10641
10642 const struct dentry_operations btrfs_dentry_operations = {
10643         .d_delete       = btrfs_dentry_delete,
10644         .d_release      = btrfs_dentry_release,
10645 };