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