fs/btrfs/file.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Copyright (C) 2007 Oracle.  All rights reserved.
   4  */
   5
   6 #include <linux/fs.h>
   7 #include <linux/pagemap.h>
   8 #include <linux/highmem.h>
   9 #include <linux/time.h>
  10 #include <linux/init.h>
  11 #include <linux/string.h>
  12 #include <linux/backing-dev.h>
  13 #include <linux/mpage.h>
  14 #include <linux/falloc.h>
  15 #include <linux/swap.h>
  16 #include <linux/writeback.h>
  17 #include <linux/compat.h>
  18 #include <linux/slab.h>
  19 #include <linux/btrfs.h>
  20 #include <linux/uio.h>
  21 #include <linux/iversion.h>
  22 #include "ctree.h"
  23 #include "disk-io.h"
  24 #include "transaction.h"
  25 #include "btrfs_inode.h"
  26 #include "print-tree.h"
  27 #include "tree-log.h"
  28 #include "locking.h"
  29 #include "volumes.h"
  30 #include "qgroup.h"
  31 #include "compression.h"
  32
  33 static struct kmem_cache *btrfs_inode_defrag_cachep;
  34 /*
  35  * when auto defrag is enabled we
  36  * queue up these defrag structs to remember which
  37  * inodes need defragging passes
  38  */
  39 struct inode_defrag {
  40         struct rb_node rb_node;
  41         /* objectid */
  42         u64 ino;
  43         /*
  44          * transid where the defrag was added, we search for
  45          * extents newer than this
  46          */
  47         u64 transid;
  48
  49         /* root objectid */
  50         u64 root;
  51
  52         /* last offset we were able to defrag */
  53         u64 last_offset;
  54
  55         /* if we've wrapped around back to zero once already */
  56         int cycled;
  57 };
  58
  59 static int __compare_inode_defrag(struct inode_defrag *defrag1,
  60                                   struct inode_defrag *defrag2)
  61 {
  62         if (defrag1->root > defrag2->root)
  63                 return 1;
  64         else if (defrag1->root < defrag2->root)
  65                 return -1;
  66         else if (defrag1->ino > defrag2->ino)
  67                 return 1;
  68         else if (defrag1->ino < defrag2->ino)
  69                 return -1;
  70         else
  71                 return 0;
  72 }
  73
  74 /* pop a record for an inode into the defrag tree.  The lock
  75  * must be held already
  76  *
  77  * If you're inserting a record for an older transid than an
  78  * existing record, the transid already in the tree is lowered
  79  *
  80  * If an existing record is found the defrag item you
  81  * pass in is freed
  82  */
  83 static int __btrfs_add_inode_defrag(struct btrfs_inode *inode,
  84                                     struct inode_defrag *defrag)
  85 {
  86         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
  87         struct inode_defrag *entry;
  88         struct rb_node **p;
  89         struct rb_node *parent = NULL;
  90         int ret;
  91
  92         p = &fs_info->defrag_inodes.rb_node;
  93         while (*p) {
  94                 parent = *p;
  95                 entry = rb_entry(parent, struct inode_defrag, rb_node);
  96
  97                 ret = __compare_inode_defrag(defrag, entry);
  98                 if (ret < 0)
  99                         p = &parent->rb_left;
 100                 else if (ret > 0)
 101                         p = &parent->rb_right;
 102                 else {
 103                         /* if we're reinserting an entry for
 104                          * an old defrag run, make sure to
 105                          * lower the transid of our existing record
 106                          */
 107                         if (defrag->transid < entry->transid)
 108                                 entry->transid = defrag->transid;
 109                         if (defrag->last_offset > entry->last_offset)
 110                                 entry->last_offset = defrag->last_offset;
 111                         return -EEXIST;
 112                 }
 113         }
 114         set_bit(BTRFS_INODE_IN_DEFRAG, &inode->runtime_flags);
 115         rb_link_node(&defrag->rb_node, parent, p);
 116         rb_insert_color(&defrag->rb_node, &fs_info->defrag_inodes);
 117         return 0;
 118 }
 119
 120 static inline int __need_auto_defrag(struct btrfs_fs_info *fs_info)
 121 {
 122         if (!btrfs_test_opt(fs_info, AUTO_DEFRAG))
 123                 return 0;
 124
 125         if (btrfs_fs_closing(fs_info))
 126                 return 0;
 127
 128         return 1;
 129 }
 130
 131 /*
 132  * insert a defrag record for this inode if auto defrag is
 133  * enabled
 134  */
 135 int btrfs_add_inode_defrag(struct btrfs_trans_handle *trans,
 136                            struct btrfs_inode *inode)
 137 {
 138         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
 139         struct btrfs_root *root = inode->root;
 140         struct inode_defrag *defrag;
 141         u64 transid;
 142         int ret;
 143
 144         if (!__need_auto_defrag(fs_info))
 145                 return 0;
 146
 147         if (test_bit(BTRFS_INODE_IN_DEFRAG, &inode->runtime_flags))
 148                 return 0;
 149
 150         if (trans)
 151                 transid = trans->transid;
 152         else
 153                 transid = inode->root->last_trans;
 154
 155         defrag = kmem_cache_zalloc(btrfs_inode_defrag_cachep, GFP_NOFS);
 156         if (!defrag)
 157                 return -ENOMEM;
 158
 159         defrag->ino = btrfs_ino(inode);
 160         defrag->transid = transid;
 161         defrag->root = root->root_key.objectid;
 162
 163         spin_lock(&fs_info->defrag_inodes_lock);
 164         if (!test_bit(BTRFS_INODE_IN_DEFRAG, &inode->runtime_flags)) {
 165                 /*
 166                  * If we set IN_DEFRAG flag and evict the inode from memory,
 167                  * and then re-read this inode, this new inode doesn't have
 168                  * IN_DEFRAG flag. At the case, we may find the existed defrag.
 169                  */
 170                 ret = __btrfs_add_inode_defrag(inode, defrag);
 171                 if (ret)
 172                         kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
 173         } else {
 174                 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
 175         }
 176         spin_unlock(&fs_info->defrag_inodes_lock);
 177         return 0;
 178 }
 179
 180 /*
 181  * Requeue the defrag object. If there is a defrag object that points to
 182  * the same inode in the tree, we will merge them together (by
 183  * __btrfs_add_inode_defrag()) and free the one that we want to requeue.
 184  */
 185 static void btrfs_requeue_inode_defrag(struct btrfs_inode *inode,
 186                                        struct inode_defrag *defrag)
 187 {
 188         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
 189         int ret;
 190
 191         if (!__need_auto_defrag(fs_info))
 192                 goto out;
 193
 194         /*
 195          * Here we don't check the IN_DEFRAG flag, because we need merge
 196          * them together.
 197          */
 198         spin_lock(&fs_info->defrag_inodes_lock);
 199         ret = __btrfs_add_inode_defrag(inode, defrag);
 200         spin_unlock(&fs_info->defrag_inodes_lock);
 201         if (ret)
 202                 goto out;
 203         return;
 204 out:
 205         kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
 206 }
 207
 208 /*
 209  * pick the defragable inode that we want, if it doesn't exist, we will get
 210  * the next one.
 211  */
 212 static struct inode_defrag *
 213 btrfs_pick_defrag_inode(struct btrfs_fs_info *fs_info, u64 root, u64 ino)
 214 {
 215         struct inode_defrag *entry = NULL;
 216         struct inode_defrag tmp;
 217         struct rb_node *p;
 218         struct rb_node *parent = NULL;
 219         int ret;
 220
 221         tmp.ino = ino;
 222         tmp.root = root;
 223
 224         spin_lock(&fs_info->defrag_inodes_lock);
 225         p = fs_info->defrag_inodes.rb_node;
 226         while (p) {
 227                 parent = p;
 228                 entry = rb_entry(parent, struct inode_defrag, rb_node);
 229
 230                 ret = __compare_inode_defrag(&tmp, entry);
 231                 if (ret < 0)
 232                         p = parent->rb_left;
 233                 else if (ret > 0)
 234                         p = parent->rb_right;
 235                 else
 236                         goto out;
 237         }
 238
 239         if (parent && __compare_inode_defrag(&tmp, entry) > 0) {
 240                 parent = rb_next(parent);
 241                 if (parent)
 242                         entry = rb_entry(parent, struct inode_defrag, rb_node);
 243                 else
 244                         entry = NULL;
 245         }
 246 out:
 247         if (entry)
 248                 rb_erase(parent, &fs_info->defrag_inodes);
 249         spin_unlock(&fs_info->defrag_inodes_lock);
 250         return entry;
 251 }
 252
 253 void btrfs_cleanup_defrag_inodes(struct btrfs_fs_info *fs_info)
 254 {
 255         struct inode_defrag *defrag;
 256         struct rb_node *node;
 257
 258         spin_lock(&fs_info->defrag_inodes_lock);
 259         node = rb_first(&fs_info->defrag_inodes);
 260         while (node) {
 261                 rb_erase(node, &fs_info->defrag_inodes);
 262                 defrag = rb_entry(node, struct inode_defrag, rb_node);
 263                 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
 264
 265                 cond_resched_lock(&fs_info->defrag_inodes_lock);
 266
 267                 node = rb_first(&fs_info->defrag_inodes);
 268         }
 269         spin_unlock(&fs_info->defrag_inodes_lock);
 270 }
 271
 272 #define BTRFS_DEFRAG_BATCH      1024
 273
 274 static int __btrfs_run_defrag_inode(struct btrfs_fs_info *fs_info,
 275                                     struct inode_defrag *defrag)
 276 {
 277         struct btrfs_root *inode_root;
 278         struct inode *inode;
 279         struct btrfs_key key;
 280         struct btrfs_ioctl_defrag_range_args range;
 281         int num_defrag;
 282         int index;
 283         int ret;
 284
 285         /* get the inode */
 286         key.objectid = defrag->root;
 287         key.type = BTRFS_ROOT_ITEM_KEY;
 288         key.offset = (u64)-1;
 289
 290         index = srcu_read_lock(&fs_info->subvol_srcu);
 291
 292         inode_root = btrfs_read_fs_root_no_name(fs_info, &key);
 293         if (IS_ERR(inode_root)) {
 294                 ret = PTR_ERR(inode_root);
 295                 goto cleanup;
 296         }
 297
 298         key.objectid = defrag->ino;
 299         key.type = BTRFS_INODE_ITEM_KEY;
 300         key.offset = 0;
 301         inode = btrfs_iget(fs_info->sb, &key, inode_root, NULL);
 302         if (IS_ERR(inode)) {
 303                 ret = PTR_ERR(inode);
 304                 goto cleanup;
 305         }
 306         srcu_read_unlock(&fs_info->subvol_srcu, index);
 307
 308         /* do a chunk of defrag */
 309         clear_bit(BTRFS_INODE_IN_DEFRAG, &BTRFS_I(inode)->runtime_flags);
 310         memset(&range, 0, sizeof(range));
 311         range.len = (u64)-1;
 312         range.start = defrag->last_offset;
 313
 314         sb_start_write(fs_info->sb);
 315         num_defrag = btrfs_defrag_file(inode, NULL, &range, defrag->transid,
 316                                        BTRFS_DEFRAG_BATCH);
 317         sb_end_write(fs_info->sb);
 318         /*
 319          * if we filled the whole defrag batch, there
 320          * must be more work to do.  Queue this defrag
 321          * again
 322          */
 323         if (num_defrag == BTRFS_DEFRAG_BATCH) {
 324                 defrag->last_offset = range.start;
 325                 btrfs_requeue_inode_defrag(BTRFS_I(inode), defrag);
 326         } else if (defrag->last_offset && !defrag->cycled) {
 327                 /*
 328                  * we didn't fill our defrag batch, but
 329                  * we didn't start at zero.  Make sure we loop
 330                  * around to the start of the file.
 331                  */
 332                 defrag->last_offset = 0;
 333                 defrag->cycled = 1;
 334                 btrfs_requeue_inode_defrag(BTRFS_I(inode), defrag);
 335         } else {
 336                 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
 337         }
 338
 339         iput(inode);
 340         return 0;
 341 cleanup:
 342         srcu_read_unlock(&fs_info->subvol_srcu, index);
 343         kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
 344         return ret;
 345 }
 346
 347 /*
 348  * run through the list of inodes in the FS that need
 349  * defragging
 350  */
 351 int btrfs_run_defrag_inodes(struct btrfs_fs_info *fs_info)
 352 {
 353         struct inode_defrag *defrag;
 354         u64 first_ino = 0;
 355         u64 root_objectid = 0;
 356
 357         atomic_inc(&fs_info->defrag_running);
 358         while (1) {
 359                 /* Pause the auto defragger. */
 360                 if (test_bit(BTRFS_FS_STATE_REMOUNTING,
 361                              &fs_info->fs_state))
 362                         break;
 363
 364                 if (!__need_auto_defrag(fs_info))
 365                         break;
 366
 367                 /* find an inode to defrag */
 368                 defrag = btrfs_pick_defrag_inode(fs_info, root_objectid,
 369                                                  first_ino);
 370                 if (!defrag) {
 371                         if (root_objectid || first_ino) {
 372                                 root_objectid = 0;
 373                                 first_ino = 0;
 374                                 continue;
 375                         } else {
 376                                 break;
 377                         }
 378                 }
 379
 380                 first_ino = defrag->ino + 1;
 381                 root_objectid = defrag->root;
 382
 383                 __btrfs_run_defrag_inode(fs_info, defrag);
 384         }
 385         atomic_dec(&fs_info->defrag_running);
 386
 387         /*
 388          * during unmount, we use the transaction_wait queue to
 389          * wait for the defragger to stop
 390          */
 391         wake_up(&fs_info->transaction_wait);
 392         return 0;
 393 }
 394
 395 /* simple helper to fault in pages and copy.  This should go away
 396  * and be replaced with calls into generic code.
 397  */
 398 static noinline int btrfs_copy_from_user(loff_t pos, size_t write_bytes,
 399                                          struct page **prepared_pages,
 400                                          struct iov_iter *i)
 401 {
 402         size_t copied = 0;
 403         size_t total_copied = 0;
 404         int pg = 0;
 405         int offset = pos & (PAGE_SIZE - 1);
 406
 407         while (write_bytes > 0) {
 408                 size_t count = min_t(size_t,
 409                                      PAGE_SIZE - offset, write_bytes);
 410                 struct page *page = prepared_pages[pg];
 411                 /*
 412                  * Copy data from userspace to the current page
 413                  */
 414                 copied = iov_iter_copy_from_user_atomic(page, i, offset, count);
 415
 416                 /* Flush processor's dcache for this page */
 417                 flush_dcache_page(page);
 418
 419                 /*
 420                  * if we get a partial write, we can end up with
 421                  * partially up to date pages.  These add
 422                  * a lot of complexity, so make sure they don't
 423                  * happen by forcing this copy to be retried.
 424                  *
 425                  * The rest of the btrfs_file_write code will fall
 426                  * back to page at a time copies after we return 0.
 427                  */
 428                 if (!PageUptodate(page) && copied < count)
 429                         copied = 0;
 430
 431                 iov_iter_advance(i, copied);
 432                 write_bytes -= copied;
 433                 total_copied += copied;
 434
 435                 /* Return to btrfs_file_write_iter to fault page */
 436                 if (unlikely(copied == 0))
 437                         break;
 438
 439                 if (copied < PAGE_SIZE - offset) {
 440                         offset += copied;
 441                 } else {
 442                         pg++;
 443                         offset = 0;
 444                 }
 445         }
 446         return total_copied;
 447 }
 448
 449 /*
 450  * unlocks pages after btrfs_file_write is done with them
 451  */
 452 static void btrfs_drop_pages(struct page **pages, size_t num_pages)
 453 {
 454         size_t i;
 455         for (i = 0; i < num_pages; i++) {
 456                 /* page checked is some magic around finding pages that
 457                  * have been modified without going through btrfs_set_page_dirty
 458                  * clear it here. There should be no need to mark the pages
 459                  * accessed as prepare_pages should have marked them accessed
 460                  * in prepare_pages via find_or_create_page()
 461                  */
 462                 ClearPageChecked(pages[i]);
 463                 unlock_page(pages[i]);
 464                 put_page(pages[i]);
 465         }
 466 }
 467
 468 static int btrfs_find_new_delalloc_bytes(struct btrfs_inode *inode,
 469                                          const u64 start,
 470                                          const u64 len,
 471                                          struct extent_state **cached_state)
 472 {
 473         u64 search_start = start;
 474         const u64 end = start + len - 1;
 475
 476         while (search_start < end) {
 477                 const u64 search_len = end - search_start + 1;
 478                 struct extent_map *em;
 479                 u64 em_len;
 480                 int ret = 0;
 481
 482                 em = btrfs_get_extent(inode, NULL, 0, search_start,
 483                                       search_len, 0);
 484                 if (IS_ERR(em))
 485                         return PTR_ERR(em);
 486
 487                 if (em->block_start != EXTENT_MAP_HOLE)
 488                         goto next;
 489
 490                 em_len = em->len;
 491                 if (em->start < search_start)
 492                         em_len -= search_start - em->start;
 493                 if (em_len > search_len)
 494                         em_len = search_len;
 495
 496                 ret = set_extent_bit(&inode->io_tree, search_start,
 497                                      search_start + em_len - 1,
 498                                      EXTENT_DELALLOC_NEW,
 499                                      NULL, cached_state, GFP_NOFS);
 500 next:
 501                 search_start = extent_map_end(em);
 502                 free_extent_map(em);
 503                 if (ret)
 504                         return ret;
 505         }
 506         return 0;
 507 }
 508
 509 /*
 510  * after copy_from_user, pages need to be dirtied and we need to make
 511  * sure holes are created between the current EOF and the start of
 512  * any next extents (if required).
 513  *
 514  * this also makes the decision about creating an inline extent vs
 515  * doing real data extents, marking pages dirty and delalloc as required.
 516  */
 517 int btrfs_dirty_pages(struct inode *inode, struct page **pages,
 518                       size_t num_pages, loff_t pos, size_t write_bytes,
 519                       struct extent_state **cached)
 520 {
 521         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
 522         int err = 0;
 523         int i;
 524         u64 num_bytes;
 525         u64 start_pos;
 526         u64 end_of_last_block;
 527         u64 end_pos = pos + write_bytes;
 528         loff_t isize = i_size_read(inode);
 529         unsigned int extra_bits = 0;
 530
 531         start_pos = pos & ~((u64) fs_info->sectorsize - 1);
 532         num_bytes = round_up(write_bytes + pos - start_pos,
 533                              fs_info->sectorsize);
 534
 535         end_of_last_block = start_pos + num_bytes - 1;
 536
 537         if (!btrfs_is_free_space_inode(BTRFS_I(inode))) {
 538                 if (start_pos >= isize &&
 539                     !(BTRFS_I(inode)->flags & BTRFS_INODE_PREALLOC)) {
 540                         /*
 541                          * There can't be any extents following eof in this case
 542                          * so just set the delalloc new bit for the range
 543                          * directly.
 544                          */
 545                         extra_bits |= EXTENT_DELALLOC_NEW;
 546                 } else {
 547                         err = btrfs_find_new_delalloc_bytes(BTRFS_I(inode),
 548                                                             start_pos,
 549                                                             num_bytes, cached);
 550                         if (err)
 551                                 return err;
 552                 }
 553         }
 554
 555         err = btrfs_set_extent_delalloc(inode, start_pos, end_of_last_block,
 556                                         extra_bits, cached, 0);
 557         if (err)
 558                 return err;
 559
 560         for (i = 0; i < num_pages; i++) {
 561                 struct page *p = pages[i];
 562                 SetPageUptodate(p);
 563                 ClearPageChecked(p);
 564                 set_page_dirty(p);
 565         }
 566
 567         /*
 568          * we've only changed i_size in ram, and we haven't updated
 569          * the disk i_size.  There is no need to log the inode
 570          * at this time.
 571          */
 572         if (end_pos > isize)
 573                 i_size_write(inode, end_pos);
 574         return 0;
 575 }
 576
 577 /*
 578  * this drops all the extents in the cache that intersect the range
 579  * [start, end].  Existing extents are split as required.
 580  */
 581 void btrfs_drop_extent_cache(struct btrfs_inode *inode, u64 start, u64 end,
 582                              int skip_pinned)
 583 {
 584         struct extent_map *em;
 585         struct extent_map *split = NULL;
 586         struct extent_map *split2 = NULL;
 587         struct extent_map_tree *em_tree = &inode->extent_tree;
 588         u64 len = end - start + 1;
 589         u64 gen;
 590         int ret;
 591         int testend = 1;
 592         unsigned long flags;
 593         int compressed = 0;
 594         bool modified;
 595
 596         WARN_ON(end < start);
 597         if (end == (u64)-1) {
 598                 len = (u64)-1;
 599                 testend = 0;
 600         }
 601         while (1) {
 602                 int no_splits = 0;
 603
 604                 modified = false;
 605                 if (!split)
 606                         split = alloc_extent_map();
 607                 if (!split2)
 608                         split2 = alloc_extent_map();
 609                 if (!split || !split2)
 610                         no_splits = 1;
 611
 612                 write_lock(&em_tree->lock);
 613                 em = lookup_extent_mapping(em_tree, start, len);
 614                 if (!em) {
 615                         write_unlock(&em_tree->lock);
 616                         break;
 617                 }
 618                 flags = em->flags;
 619                 gen = em->generation;
 620                 if (skip_pinned && test_bit(EXTENT_FLAG_PINNED, &em->flags)) {
 621                         if (testend && em->start + em->len >= start + len) {
 622                                 free_extent_map(em);
 623                                 write_unlock(&em_tree->lock);
 624                                 break;
 625                         }
 626                         start = em->start + em->len;
 627                         if (testend)
 628                                 len = start + len - (em->start + em->len);
 629                         free_extent_map(em);
 630                         write_unlock(&em_tree->lock);
 631                         continue;
 632                 }
 633                 compressed = test_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
 634                 clear_bit(EXTENT_FLAG_PINNED, &em->flags);
 635                 clear_bit(EXTENT_FLAG_LOGGING, &flags);
 636                 modified = !list_empty(&em->list);
 637                 if (no_splits)
 638                         goto next;
 639
 640                 if (em->start < start) {
 641                         split->start = em->start;
 642                         split->len = start - em->start;
 643
 644                         if (em->block_start < EXTENT_MAP_LAST_BYTE) {
 645                                 split->orig_start = em->orig_start;
 646                                 split->block_start = em->block_start;
 647
 648                                 if (compressed)
 649                                         split->block_len = em->block_len;
 650                                 else
 651                                         split->block_len = split->len;
 652                                 split->orig_block_len = max(split->block_len,
 653                                                 em->orig_block_len);
 654                                 split->ram_bytes = em->ram_bytes;
 655                         } else {
 656                                 split->orig_start = split->start;
 657                                 split->block_len = 0;
 658                                 split->block_start = em->block_start;
 659                                 split->orig_block_len = 0;
 660                                 split->ram_bytes = split->len;
 661                         }
 662
 663                         split->generation = gen;
 664                         split->bdev = em->bdev;
 665                         split->flags = flags;
 666                         split->compress_type = em->compress_type;
 667                         replace_extent_mapping(em_tree, em, split, modified);
 668                         free_extent_map(split);
 669                         split = split2;
 670                         split2 = NULL;
 671                 }
 672                 if (testend && em->start + em->len > start + len) {
 673                         u64 diff = start + len - em->start;
 674
 675                         split->start = start + len;
 676                         split->len = em->start + em->len - (start + len);
 677                         split->bdev = em->bdev;
 678                         split->flags = flags;
 679                         split->compress_type = em->compress_type;
 680                         split->generation = gen;
 681
 682                         if (em->block_start < EXTENT_MAP_LAST_BYTE) {
 683                                 split->orig_block_len = max(em->block_len,
 684                                                     em->orig_block_len);
 685
 686                                 split->ram_bytes = em->ram_bytes;
 687                                 if (compressed) {
 688                                         split->block_len = em->block_len;
 689                                         split->block_start = em->block_start;
 690                                         split->orig_start = em->orig_start;
 691                                 } else {
 692                                         split->block_len = split->len;
 693                                         split->block_start = em->block_start
 694                                                 + diff;
 695                                         split->orig_start = em->orig_start;
 696                                 }
 697                         } else {
 698                                 split->ram_bytes = split->len;
 699                                 split->orig_start = split->start;
 700                                 split->block_len = 0;
 701                                 split->block_start = em->block_start;
 702                                 split->orig_block_len = 0;
 703                         }
 704
 705                         if (extent_map_in_tree(em)) {
 706                                 replace_extent_mapping(em_tree, em, split,
 707                                                        modified);
 708                         } else {
 709                                 ret = add_extent_mapping(em_tree, split,
 710                                                          modified);
 711                                 ASSERT(ret == 0); /* Logic error */
 712                         }
 713                         free_extent_map(split);
 714                         split = NULL;
 715                 }
 716 next:
 717                 if (extent_map_in_tree(em))
 718                         remove_extent_mapping(em_tree, em);
 719                 write_unlock(&em_tree->lock);
 720
 721                 /* once for us */
 722                 free_extent_map(em);
 723                 /* once for the tree*/
 724                 free_extent_map(em);
 725         }
 726         if (split)
 727                 free_extent_map(split);
 728         if (split2)
 729                 free_extent_map(split2);
 730 }
 731
 732 /*
 733  * this is very complex, but the basic idea is to drop all extents
 734  * in the range start - end.  hint_block is filled in with a block number
 735  * that would be a good hint to the block allocator for this file.
 736  *
 737  * If an extent intersects the range but is not entirely inside the range
 738  * it is either truncated or split.  Anything entirely inside the range
 739  * is deleted from the tree.
 740  */
 741 int __btrfs_drop_extents(struct btrfs_trans_handle *trans,
 742                          struct btrfs_root *root, struct inode *inode,
 743                          struct btrfs_path *path, u64 start, u64 end,
 744                          u64 *drop_end, int drop_cache,
 745                          int replace_extent,
 746                          u32 extent_item_size,
 747                          int *key_inserted)
 748 {
 749         struct btrfs_fs_info *fs_info = root->fs_info;
 750         struct extent_buffer *leaf;
 751         struct btrfs_file_extent_item *fi;
 752         struct btrfs_key key;
 753         struct btrfs_key new_key;
 754         u64 ino = btrfs_ino(BTRFS_I(inode));
 755         u64 search_start = start;
 756         u64 disk_bytenr = 0;
 757         u64 num_bytes = 0;
 758         u64 extent_offset = 0;
 759         u64 extent_end = 0;
 760         u64 last_end = start;
 761         int del_nr = 0;
 762         int del_slot = 0;
 763         int extent_type;
 764         int recow;
 765         int ret;
 766         int modify_tree = -1;
 767         int update_refs;
 768         int found = 0;
 769         int leafs_visited = 0;
 770
 771         if (drop_cache)
 772                 btrfs_drop_extent_cache(BTRFS_I(inode), start, end - 1, 0);
 773
 774         if (start >= BTRFS_I(inode)->disk_i_size && !replace_extent)
 775                 modify_tree = 0;
 776
 777         update_refs = (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
 778                        root == fs_info->tree_root);
 779         while (1) {
 780                 recow = 0;
 781                 ret = btrfs_lookup_file_extent(trans, root, path, ino,
 782                                                search_start, modify_tree);
 783                 if (ret < 0)
 784                         break;
 785                 if (ret > 0 && path->slots[0] > 0 && search_start == start) {
 786                         leaf = path->nodes[0];
 787                         btrfs_item_key_to_cpu(leaf, &key, path->slots[0] - 1);
 788                         if (key.objectid == ino &&
 789                             key.type == BTRFS_EXTENT_DATA_KEY)
 790                                 path->slots[0]--;
 791                 }
 792                 ret = 0;
 793                 leafs_visited++;
 794 next_slot:
 795                 leaf = path->nodes[0];
 796                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
 797                         BUG_ON(del_nr > 0);
 798                         ret = btrfs_next_leaf(root, path);
 799                         if (ret < 0)
 800                                 break;
 801                         if (ret > 0) {
 802                                 ret = 0;
 803                                 break;
 804                         }
 805                         leafs_visited++;
 806                         leaf = path->nodes[0];
 807                         recow = 1;
 808                 }
 809
 810                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 811
 812                 if (key.objectid > ino)
 813                         break;
 814                 if (WARN_ON_ONCE(key.objectid < ino) ||
 815                     key.type < BTRFS_EXTENT_DATA_KEY) {
 816                         ASSERT(del_nr == 0);
 817                         path->slots[0]++;
 818                         goto next_slot;
 819                 }
 820                 if (key.type > BTRFS_EXTENT_DATA_KEY || key.offset >= end)
 821                         break;
 822
 823                 fi = btrfs_item_ptr(leaf, path->slots[0],
 824                                     struct btrfs_file_extent_item);
 825                 extent_type = btrfs_file_extent_type(leaf, fi);
 826
 827                 if (extent_type == BTRFS_FILE_EXTENT_REG ||
 828                     extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
 829                         disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
 830                         num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
 831                         extent_offset = btrfs_file_extent_offset(leaf, fi);
 832                         extent_end = key.offset +
 833                                 btrfs_file_extent_num_bytes(leaf, fi);
 834                 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
 835                         extent_end = key.offset +
 836                                 btrfs_file_extent_inline_len(leaf,
 837                                                      path->slots[0], fi);
 838                 } else {
 839                         /* can't happen */
 840                         BUG();
 841                 }
 842
 843                 /*
 844                  * Don't skip extent items representing 0 byte lengths. They
 845                  * used to be created (bug) if while punching holes we hit
 846                  * -ENOSPC condition. So if we find one here, just ensure we
 847                  * delete it, otherwise we would insert a new file extent item
 848                  * with the same key (offset) as that 0 bytes length file
 849                  * extent item in the call to setup_items_for_insert() later
 850                  * in this function.
 851                  */
 852                 if (extent_end == key.offset && extent_end >= search_start) {
 853                         last_end = extent_end;
 854                         goto delete_extent_item;
 855                 }
 856
 857                 if (extent_end <= search_start) {
 858                         path->slots[0]++;
 859                         goto next_slot;
 860                 }
 861
 862                 found = 1;
 863                 search_start = max(key.offset, start);
 864                 if (recow || !modify_tree) {
 865                         modify_tree = -1;
 866                         btrfs_release_path(path);
 867                         continue;
 868                 }
 869
 870                 /*
 871                  *     | - range to drop - |
 872                  *  | -------- extent -------- |
 873                  */
 874                 if (start > key.offset && end < extent_end) {
 875                         BUG_ON(del_nr > 0);
 876                         if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
 877                                 ret = -EOPNOTSUPP;
 878                                 break;
 879                         }
 880
 881                         memcpy(&new_key, &key, sizeof(new_key));
 882                         new_key.offset = start;
 883                         ret = btrfs_duplicate_item(trans, root, path,
 884                                                    &new_key);
 885                         if (ret == -EAGAIN) {
 886                                 btrfs_release_path(path);
 887                                 continue;
 888                         }
 889                         if (ret < 0)
 890                                 break;
 891
 892                         leaf = path->nodes[0];
 893                         fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
 894                                             struct btrfs_file_extent_item);
 895                         btrfs_set_file_extent_num_bytes(leaf, fi,
 896                                                         start - key.offset);
 897
 898                         fi = btrfs_item_ptr(leaf, path->slots[0],
 899                                             struct btrfs_file_extent_item);
 900
 901                         extent_offset += start - key.offset;
 902                         btrfs_set_file_extent_offset(leaf, fi, extent_offset);
 903                         btrfs_set_file_extent_num_bytes(leaf, fi,
 904                                                         extent_end - start);
 905                         btrfs_mark_buffer_dirty(leaf);
 906
 907                         if (update_refs && disk_bytenr > 0) {
 908                                 ret = btrfs_inc_extent_ref(trans, root,
 909                                                 disk_bytenr, num_bytes, 0,
 910                                                 root->root_key.objectid,
 911                                                 new_key.objectid,
 912                                                 start - extent_offset);
 913                                 BUG_ON(ret); /* -ENOMEM */
 914                         }
 915                         key.offset = start;
 916                 }
 917                 /*
 918                  * From here on out we will have actually dropped something, so
 919                  * last_end can be updated.
 920                  */
 921                 last_end = extent_end;
 922
 923                 /*
 924                  *  | ---- range to drop ----- |
 925                  *      | -------- extent -------- |
 926                  */
 927                 if (start <= key.offset && end < extent_end) {
 928                         if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
 929                                 ret = -EOPNOTSUPP;
 930                                 break;
 931                         }
 932
 933                         memcpy(&new_key, &key, sizeof(new_key));
 934                         new_key.offset = end;
 935                         btrfs_set_item_key_safe(fs_info, path, &new_key);
 936
 937                         extent_offset += end - key.offset;
 938                         btrfs_set_file_extent_offset(leaf, fi, extent_offset);
 939                         btrfs_set_file_extent_num_bytes(leaf, fi,
 940                                                         extent_end - end);
 941                         btrfs_mark_buffer_dirty(leaf);
 942                         if (update_refs && disk_bytenr > 0)
 943                                 inode_sub_bytes(inode, end - key.offset);
 944                         break;
 945                 }
 946
 947                 search_start = extent_end;
 948                 /*
 949                  *       | ---- range to drop ----- |
 950                  *  | -------- extent -------- |
 951                  */
 952                 if (start > key.offset && end >= extent_end) {
 953                         BUG_ON(del_nr > 0);
 954                         if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
 955                                 ret = -EOPNOTSUPP;
 956                                 break;
 957                         }
 958
 959                         btrfs_set_file_extent_num_bytes(leaf, fi,
 960                                                         start - key.offset);
 961                         btrfs_mark_buffer_dirty(leaf);
 962                         if (update_refs && disk_bytenr > 0)
 963                                 inode_sub_bytes(inode, extent_end - start);
 964                         if (end == extent_end)
 965                                 break;
 966
 967                         path->slots[0]++;
 968                         goto next_slot;
 969                 }
 970
 971                 /*
 972                  *  | ---- range to drop ----- |
 973                  *    | ------ extent ------ |
 974                  */
 975                 if (start <= key.offset && end >= extent_end) {
 976 delete_extent_item:
 977                         if (del_nr == 0) {
 978                                 del_slot = path->slots[0];
 979                                 del_nr = 1;
 980                         } else {
 981                                 BUG_ON(del_slot + del_nr != path->slots[0]);
 982                                 del_nr++;
 983                         }
 984
 985                         if (update_refs &&
 986                             extent_type == BTRFS_FILE_EXTENT_INLINE) {
 987                                 inode_sub_bytes(inode,
 988                                                 extent_end - key.offset);
 989                                 extent_end = ALIGN(extent_end,
 990                                                    fs_info->sectorsize);
 991                         } else if (update_refs && disk_bytenr > 0) {
 992                                 ret = btrfs_free_extent(trans, root,
 993                                                 disk_bytenr, num_bytes, 0,
 994                                                 root->root_key.objectid,
 995                                                 key.objectid, key.offset -
 996                                                 extent_offset);
 997                                 BUG_ON(ret); /* -ENOMEM */
 998                                 inode_sub_bytes(inode,
 999                                                 extent_end - key.offset);
1000                         }
1001
1002                         if (end == extent_end)
1003                                 break;
1004
1005                         if (path->slots[0] + 1 < btrfs_header_nritems(leaf)) {
1006                                 path->slots[0]++;
1007                                 goto next_slot;
1008                         }
1009
1010                         ret = btrfs_del_items(trans, root, path, del_slot,
1011                                               del_nr);
1012                         if (ret) {
1013                                 btrfs_abort_transaction(trans, ret);
1014                                 break;
1015                         }
1016
1017                         del_nr = 0;
1018                         del_slot = 0;
1019
1020                         btrfs_release_path(path);
1021                         continue;
1022                 }
1023
1024                 BUG_ON(1);
1025         }
1026
1027         if (!ret && del_nr > 0) {
1028                 /*
1029                  * Set path->slots[0] to first slot, so that after the delete
1030                  * if items are move off from our leaf to its immediate left or
1031                  * right neighbor leafs, we end up with a correct and adjusted
1032                  * path->slots[0] for our insertion (if replace_extent != 0).
1033                  */
1034                 path->slots[0] = del_slot;
1035                 ret = btrfs_del_items(trans, root, path, del_slot, del_nr);
1036                 if (ret)
1037                         btrfs_abort_transaction(trans, ret);
1038         }
1039
1040         leaf = path->nodes[0];
1041         /*
1042          * If btrfs_del_items() was called, it might have deleted a leaf, in
1043          * which case it unlocked our path, so check path->locks[0] matches a
1044          * write lock.
1045          */
1046         if (!ret && replace_extent && leafs_visited == 1 &&
1047             (path->locks[0] == BTRFS_WRITE_LOCK_BLOCKING ||
1048              path->locks[0] == BTRFS_WRITE_LOCK) &&
1049             btrfs_leaf_free_space(fs_info, leaf) >=
1050             sizeof(struct btrfs_item) + extent_item_size) {
1051
1052                 key.objectid = ino;
1053                 key.type = BTRFS_EXTENT_DATA_KEY;
1054                 key.offset = start;
1055                 if (!del_nr && path->slots[0] < btrfs_header_nritems(leaf)) {
1056                         struct btrfs_key slot_key;
1057
1058                         btrfs_item_key_to_cpu(leaf, &slot_key, path->slots[0]);
1059                         if (btrfs_comp_cpu_keys(&key, &slot_key) > 0)
1060                                 path->slots[0]++;
1061                 }
1062                 setup_items_for_insert(root, path, &key,
1063                                        &extent_item_size,
1064                                        extent_item_size,
1065                                        sizeof(struct btrfs_item) +
1066                                        extent_item_size, 1);
1067                 *key_inserted = 1;
1068         }
1069
1070         if (!replace_extent || !(*key_inserted))
1071                 btrfs_release_path(path);
1072         if (drop_end)
1073                 *drop_end = found ? min(end, last_end) : end;
1074         return ret;
1075 }
1076
1077 int btrfs_drop_extents(struct btrfs_trans_handle *trans,
1078                        struct btrfs_root *root, struct inode *inode, u64 start,
1079                        u64 end, int drop_cache)
1080 {
1081         struct btrfs_path *path;
1082         int ret;
1083
1084         path = btrfs_alloc_path();
1085         if (!path)
1086                 return -ENOMEM;
1087         ret = __btrfs_drop_extents(trans, root, inode, path, start, end, NULL,
1088                                    drop_cache, 0, 0, NULL);
1089         btrfs_free_path(path);
1090         return ret;
1091 }
1092
1093 static int extent_mergeable(struct extent_buffer *leaf, int slot,
1094                             u64 objectid, u64 bytenr, u64 orig_offset,
1095                             u64 *start, u64 *end)
1096 {
1097         struct btrfs_file_extent_item *fi;
1098         struct btrfs_key key;
1099         u64 extent_end;
1100
1101         if (slot < 0 || slot >= btrfs_header_nritems(leaf))
1102                 return 0;
1103
1104         btrfs_item_key_to_cpu(leaf, &key, slot);
1105         if (key.objectid != objectid || key.type != BTRFS_EXTENT_DATA_KEY)
1106                 return 0;
1107
1108         fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
1109         if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG ||
1110             btrfs_file_extent_disk_bytenr(leaf, fi) != bytenr ||
1111             btrfs_file_extent_offset(leaf, fi) != key.offset - orig_offset ||
1112             btrfs_file_extent_compression(leaf, fi) ||
1113             btrfs_file_extent_encryption(leaf, fi) ||
1114             btrfs_file_extent_other_encoding(leaf, fi))
1115                 return 0;
1116
1117         extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
1118         if ((*start && *start != key.offset) || (*end && *end != extent_end))
1119                 return 0;
1120
1121         *start = key.offset;
1122         *end = extent_end;
1123         return 1;
1124 }
1125
1126 /*
1127  * Mark extent in the range start - end as written.
1128  *
1129  * This changes extent type from 'pre-allocated' to 'regular'. If only
1130  * part of extent is marked as written, the extent will be split into
1131  * two or three.
1132  */
1133 int btrfs_mark_extent_written(struct btrfs_trans_handle *trans,
1134                               struct btrfs_inode *inode, u64 start, u64 end)
1135 {
1136         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
1137         struct btrfs_root *root = inode->root;
1138         struct extent_buffer *leaf;
1139         struct btrfs_path *path;
1140         struct btrfs_file_extent_item *fi;
1141         struct btrfs_key key;
1142         struct btrfs_key new_key;
1143         u64 bytenr;
1144         u64 num_bytes;
1145         u64 extent_end;
1146         u64 orig_offset;
1147         u64 other_start;
1148         u64 other_end;
1149         u64 split;
1150         int del_nr = 0;
1151         int del_slot = 0;
1152         int recow;
1153         int ret;
1154         u64 ino = btrfs_ino(inode);
1155
1156         path = btrfs_alloc_path();
1157         if (!path)
1158                 return -ENOMEM;
1159 again:
1160         recow = 0;
1161         split = start;
1162         key.objectid = ino;
1163         key.type = BTRFS_EXTENT_DATA_KEY;
1164         key.offset = split;
1165
1166         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1167         if (ret < 0)
1168                 goto out;
1169         if (ret > 0 && path->slots[0] > 0)
1170                 path->slots[0]--;
1171
1172         leaf = path->nodes[0];
1173         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1174         if (key.objectid != ino ||
1175             key.type != BTRFS_EXTENT_DATA_KEY) {
1176                 ret = -EINVAL;
1177                 btrfs_abort_transaction(trans, ret);
1178                 goto out;
1179         }
1180         fi = btrfs_item_ptr(leaf, path->slots[0],
1181                             struct btrfs_file_extent_item);
1182         if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_PREALLOC) {
1183                 ret = -EINVAL;
1184                 btrfs_abort_transaction(trans, ret);
1185                 goto out;
1186         }
1187         extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
1188         if (key.offset > start || extent_end < end) {
1189                 ret = -EINVAL;
1190                 btrfs_abort_transaction(trans, ret);
1191                 goto out;
1192         }
1193
1194         bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1195         num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
1196         orig_offset = key.offset - btrfs_file_extent_offset(leaf, fi);
1197         memcpy(&new_key, &key, sizeof(new_key));
1198
1199         if (start == key.offset && end < extent_end) {
1200                 other_start = 0;
1201                 other_end = start;
1202                 if (extent_mergeable(leaf, path->slots[0] - 1,
1203                                      ino, bytenr, orig_offset,
1204                                      &other_start, &other_end)) {
1205                         new_key.offset = end;
1206                         btrfs_set_item_key_safe(fs_info, path, &new_key);
1207                         fi = btrfs_item_ptr(leaf, path->slots[0],
1208                                             struct btrfs_file_extent_item);
1209                         btrfs_set_file_extent_generation(leaf, fi,
1210                                                          trans->transid);
1211                         btrfs_set_file_extent_num_bytes(leaf, fi,
1212                                                         extent_end - end);
1213                         btrfs_set_file_extent_offset(leaf, fi,
1214                                                      end - orig_offset);
1215                         fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
1216                                             struct btrfs_file_extent_item);
1217                         btrfs_set_file_extent_generation(leaf, fi,
1218                                                          trans->transid);
1219                         btrfs_set_file_extent_num_bytes(leaf, fi,
1220                                                         end - other_start);
1221                         btrfs_mark_buffer_dirty(leaf);
1222                         goto out;
1223                 }
1224         }
1225
1226         if (start > key.offset && end == extent_end) {
1227                 other_start = end;
1228                 other_end = 0;
1229                 if (extent_mergeable(leaf, path->slots[0] + 1,
1230                                      ino, bytenr, orig_offset,
1231                                      &other_start, &other_end)) {
1232                         fi = btrfs_item_ptr(leaf, path->slots[0],
1233                                             struct btrfs_file_extent_item);
1234                         btrfs_set_file_extent_num_bytes(leaf, fi,
1235                                                         start - key.offset);
1236                         btrfs_set_file_extent_generation(leaf, fi,
1237                                                          trans->transid);
1238                         path->slots[0]++;
1239                         new_key.offset = start;
1240                         btrfs_set_item_key_safe(fs_info, path, &new_key);
1241
1242                         fi = btrfs_item_ptr(leaf, path->slots[0],
1243                                             struct btrfs_file_extent_item);
1244                         btrfs_set_file_extent_generation(leaf, fi,
1245                                                          trans->transid);
1246                         btrfs_set_file_extent_num_bytes(leaf, fi,
1247                                                         other_end - start);
1248                         btrfs_set_file_extent_offset(leaf, fi,
1249                                                      start - orig_offset);
1250                         btrfs_mark_buffer_dirty(leaf);
1251                         goto out;
1252                 }
1253         }
1254
1255         while (start > key.offset || end < extent_end) {
1256                 if (key.offset == start)
1257                         split = end;
1258
1259                 new_key.offset = split;
1260                 ret = btrfs_duplicate_item(trans, root, path, &new_key);
1261                 if (ret == -EAGAIN) {
1262                         btrfs_release_path(path);
1263                         goto again;
1264                 }
1265                 if (ret < 0) {
1266                         btrfs_abort_transaction(trans, ret);
1267                         goto out;
1268                 }
1269
1270                 leaf = path->nodes[0];
1271                 fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
1272                                     struct btrfs_file_extent_item);
1273                 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1274                 btrfs_set_file_extent_num_bytes(leaf, fi,
1275                                                 split - key.offset);
1276
1277                 fi = btrfs_item_ptr(leaf, path->slots[0],
1278                                     struct btrfs_file_extent_item);
1279
1280                 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1281                 btrfs_set_file_extent_offset(leaf, fi, split - orig_offset);
1282                 btrfs_set_file_extent_num_bytes(leaf, fi,
1283                                                 extent_end - split);
1284                 btrfs_mark_buffer_dirty(leaf);
1285
1286                 ret = btrfs_inc_extent_ref(trans, root, bytenr, num_bytes,
1287                                            0, root->root_key.objectid,
1288                                            ino, orig_offset);
1289                 if (ret) {
1290                         btrfs_abort_transaction(trans, ret);
1291                         goto out;
1292                 }
1293
1294                 if (split == start) {
1295                         key.offset = start;
1296                 } else {
1297                         if (start != key.offset) {
1298                                 ret = -EINVAL;
1299                                 btrfs_abort_transaction(trans, ret);
1300                                 goto out;
1301                         }
1302                         path->slots[0]--;
1303                         extent_end = end;
1304                 }
1305                 recow = 1;
1306         }
1307
1308         other_start = end;
1309         other_end = 0;
1310         if (extent_mergeable(leaf, path->slots[0] + 1,
1311                              ino, bytenr, orig_offset,
1312                              &other_start, &other_end)) {
1313                 if (recow) {
1314                         btrfs_release_path(path);
1315                         goto again;
1316                 }
1317                 extent_end = other_end;
1318                 del_slot = path->slots[0] + 1;
1319                 del_nr++;
1320                 ret = btrfs_free_extent(trans, root, bytenr, num_bytes,
1321                                         0, root->root_key.objectid,
1322                                         ino, orig_offset);
1323                 if (ret) {
1324                         btrfs_abort_transaction(trans, ret);
1325                         goto out;
1326                 }
1327         }
1328         other_start = 0;
1329         other_end = start;
1330         if (extent_mergeable(leaf, path->slots[0] - 1,
1331                              ino, bytenr, orig_offset,
1332                              &other_start, &other_end)) {
1333                 if (recow) {
1334                         btrfs_release_path(path);
1335                         goto again;
1336                 }
1337                 key.offset = other_start;
1338                 del_slot = path->slots[0];
1339                 del_nr++;
1340                 ret = btrfs_free_extent(trans, root, bytenr, num_bytes,
1341                                         0, root->root_key.objectid,
1342                                         ino, orig_offset);
1343                 if (ret) {
1344                         btrfs_abort_transaction(trans, ret);
1345                         goto out;
1346                 }
1347         }
1348         if (del_nr == 0) {
1349                 fi = btrfs_item_ptr(leaf, path->slots[0],
1350                            struct btrfs_file_extent_item);
1351                 btrfs_set_file_extent_type(leaf, fi,
1352                                            BTRFS_FILE_EXTENT_REG);
1353                 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1354                 btrfs_mark_buffer_dirty(leaf);
1355         } else {
1356                 fi = btrfs_item_ptr(leaf, del_slot - 1,
1357                            struct btrfs_file_extent_item);
1358                 btrfs_set_file_extent_type(leaf, fi,
1359                                            BTRFS_FILE_EXTENT_REG);
1360                 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1361                 btrfs_set_file_extent_num_bytes(leaf, fi,
1362                                                 extent_end - key.offset);
1363                 btrfs_mark_buffer_dirty(leaf);
1364
1365                 ret = btrfs_del_items(trans, root, path, del_slot, del_nr);
1366                 if (ret < 0) {
1367                         btrfs_abort_transaction(trans, ret);
1368                         goto out;
1369                 }
1370         }
1371 out:
1372         btrfs_free_path(path);
1373         return 0;
1374 }
1375
1376 /*
1377  * on error we return an unlocked page and the error value
1378  * on success we return a locked page and 0
1379  */
1380 static int prepare_uptodate_page(struct inode *inode,
1381                                  struct page *page, u64 pos,
1382                                  bool force_uptodate)
1383 {
1384         int ret = 0;
1385
1386         if (((pos & (PAGE_SIZE - 1)) || force_uptodate) &&
1387             !PageUptodate(page)) {
1388                 ret = btrfs_readpage(NULL, page);
1389                 if (ret)
1390                         return ret;
1391                 lock_page(page);
1392                 if (!PageUptodate(page)) {
1393                         unlock_page(page);
1394                         return -EIO;
1395                 }
1396                 if (page->mapping != inode->i_mapping) {
1397                         unlock_page(page);
1398                         return -EAGAIN;
1399                 }
1400         }
1401         return 0;
1402 }
1403
1404 /*
1405  * this just gets pages into the page cache and locks them down.
1406  */
1407 static noinline int prepare_pages(struct inode *inode, struct page **pages,
1408                                   size_t num_pages, loff_t pos,
1409                                   size_t write_bytes, bool force_uptodate)
1410 {
1411         int i;
1412         unsigned long index = pos >> PAGE_SHIFT;
1413         gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping);
1414         int err = 0;
1415         int faili;
1416
1417         for (i = 0; i < num_pages; i++) {
1418 again:
1419                 pages[i] = find_or_create_page(inode->i_mapping, index + i,
1420                                                mask | __GFP_WRITE);
1421                 if (!pages[i]) {
1422                         faili = i - 1;
1423                         err = -ENOMEM;
1424                         goto fail;
1425                 }
1426
1427                 if (i == 0)
1428                         err = prepare_uptodate_page(inode, pages[i], pos,
1429                                                     force_uptodate);
1430                 if (!err && i == num_pages - 1)
1431                         err = prepare_uptodate_page(inode, pages[i],
1432                                                     pos + write_bytes, false);
1433                 if (err) {
1434                         put_page(pages[i]);
1435                         if (err == -EAGAIN) {
1436                                 err = 0;
1437                                 goto again;
1438                         }
1439                         faili = i - 1;
1440                         goto fail;
1441                 }
1442                 wait_on_page_writeback(pages[i]);
1443         }
1444
1445         return 0;
1446 fail:
1447         while (faili >= 0) {
1448                 unlock_page(pages[faili]);
1449                 put_page(pages[faili]);
1450                 faili--;
1451         }
1452         return err;
1453
1454 }
1455
1456 /*
1457  * This function locks the extent and properly waits for data=ordered extents
1458  * to finish before allowing the pages to be modified if need.
1459  *
1460  * The return value:
1461  * 1 - the extent is locked
1462  * 0 - the extent is not locked, and everything is OK
1463  * -EAGAIN - need re-prepare the pages
1464  * the other < 0 number - Something wrong happens
1465  */
1466 static noinline int
1467 lock_and_cleanup_extent_if_need(struct btrfs_inode *inode, struct page **pages,
1468                                 size_t num_pages, loff_t pos,
1469                                 size_t write_bytes,
1470                                 u64 *lockstart, u64 *lockend,
1471                                 struct extent_state **cached_state)
1472 {
1473         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
1474         u64 start_pos;
1475         u64 last_pos;
1476         int i;
1477         int ret = 0;
1478
1479         start_pos = round_down(pos, fs_info->sectorsize);
1480         last_pos = start_pos
1481                 + round_up(pos + write_bytes - start_pos,
1482                            fs_info->sectorsize) - 1;
1483
1484         if (start_pos < inode->vfs_inode.i_size) {
1485                 struct btrfs_ordered_extent *ordered;
1486
1487                 lock_extent_bits(&inode->io_tree, start_pos, last_pos,
1488                                 cached_state);
1489                 ordered = btrfs_lookup_ordered_range(inode, start_pos,
1490                                                      last_pos - start_pos + 1);
1491                 if (ordered &&
1492                     ordered->file_offset + ordered->len > start_pos &&
1493                     ordered->file_offset <= last_pos) {
1494                         unlock_extent_cached(&inode->io_tree, start_pos,
1495                                         last_pos, cached_state);
1496                         for (i = 0; i < num_pages; i++) {
1497                                 unlock_page(pages[i]);
1498                                 put_page(pages[i]);
1499                         }
1500                         btrfs_start_ordered_extent(&inode->vfs_inode,
1501                                         ordered, 1);
1502                         btrfs_put_ordered_extent(ordered);
1503                         return -EAGAIN;
1504                 }
1505                 if (ordered)
1506                         btrfs_put_ordered_extent(ordered);
1507                 clear_extent_bit(&inode->io_tree, start_pos, last_pos,
1508                                  EXTENT_DIRTY | EXTENT_DELALLOC |
1509                                  EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG,
1510                                  0, 0, cached_state);
1511                 *lockstart = start_pos;
1512                 *lockend = last_pos;
1513                 ret = 1;
1514         }
1515
1516         for (i = 0; i < num_pages; i++) {
1517                 if (clear_page_dirty_for_io(pages[i]))
1518                         account_page_redirty(pages[i]);
1519                 set_page_extent_mapped(pages[i]);
1520                 WARN_ON(!PageLocked(pages[i]));
1521         }
1522
1523         return ret;
1524 }
1525
1526 static noinline int check_can_nocow(struct btrfs_inode *inode, loff_t pos,
1527                                     size_t *write_bytes)
1528 {
1529         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
1530         struct btrfs_root *root = inode->root;
1531         struct btrfs_ordered_extent *ordered;
1532         u64 lockstart, lockend;
1533         u64 num_bytes;
1534         int ret;
1535
1536         ret = btrfs_start_write_no_snapshotting(root);
1537         if (!ret)
1538                 return -ENOSPC;
1539
1540         lockstart = round_down(pos, fs_info->sectorsize);
1541         lockend = round_up(pos + *write_bytes,
1542                            fs_info->sectorsize) - 1;
1543
1544         while (1) {
1545                 lock_extent(&inode->io_tree, lockstart, lockend);
1546                 ordered = btrfs_lookup_ordered_range(inode, lockstart,
1547                                                      lockend - lockstart + 1);
1548                 if (!ordered) {
1549                         break;
1550                 }
1551                 unlock_extent(&inode->io_tree, lockstart, lockend);
1552                 btrfs_start_ordered_extent(&inode->vfs_inode, ordered, 1);
1553                 btrfs_put_ordered_extent(ordered);
1554         }
1555
1556         num_bytes = lockend - lockstart + 1;
1557         ret = can_nocow_extent(&inode->vfs_inode, lockstart, &num_bytes,
1558                         NULL, NULL, NULL);
1559         if (ret <= 0) {
1560                 ret = 0;
1561                 btrfs_end_write_no_snapshotting(root);
1562         } else {
1563                 *write_bytes = min_t(size_t, *write_bytes ,
1564                                      num_bytes - pos + lockstart);
1565         }
1566
1567         unlock_extent(&inode->io_tree, lockstart, lockend);
1568
1569         return ret;
1570 }
1571
1572 static noinline ssize_t __btrfs_buffered_write(struct file *file,
1573                                                struct iov_iter *i,
1574                                                loff_t pos)
1575 {
1576         struct inode *inode = file_inode(file);
1577         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1578         struct btrfs_root *root = BTRFS_I(inode)->root;
1579         struct page **pages = NULL;
1580         struct extent_state *cached_state = NULL;
1581         struct extent_changeset *data_reserved = NULL;
1582         u64 release_bytes = 0;
1583         u64 lockstart;
1584         u64 lockend;
1585         size_t num_written = 0;
1586         int nrptrs;
1587         int ret = 0;
1588         bool only_release_metadata = false;
1589         bool force_page_uptodate = false;
1590
1591         nrptrs = min(DIV_ROUND_UP(iov_iter_count(i), PAGE_SIZE),
1592                         PAGE_SIZE / (sizeof(struct page *)));
1593         nrptrs = min(nrptrs, current->nr_dirtied_pause - current->nr_dirtied);
1594         nrptrs = max(nrptrs, 8);
1595         pages = kmalloc_array(nrptrs, sizeof(struct page *), GFP_KERNEL);
1596         if (!pages)
1597                 return -ENOMEM;
1598
1599         while (iov_iter_count(i) > 0) {
1600                 size_t offset = pos & (PAGE_SIZE - 1);
1601                 size_t sector_offset;
1602                 size_t write_bytes = min(iov_iter_count(i),
1603                                          nrptrs * (size_t)PAGE_SIZE -
1604                                          offset);
1605                 size_t num_pages = DIV_ROUND_UP(write_bytes + offset,
1606                                                 PAGE_SIZE);
1607                 size_t reserve_bytes;
1608                 size_t dirty_pages;
1609                 size_t copied;
1610                 size_t dirty_sectors;
1611                 size_t num_sectors;
1612                 int extents_locked;
1613
1614                 WARN_ON(num_pages > nrptrs);
1615
1616                 /*
1617                  * Fault pages before locking them in prepare_pages
1618                  * to avoid recursive lock
1619                  */
1620                 if (unlikely(iov_iter_fault_in_readable(i, write_bytes))) {
1621                         ret = -EFAULT;
1622                         break;
1623                 }
1624
1625                 sector_offset = pos & (fs_info->sectorsize - 1);
1626                 reserve_bytes = round_up(write_bytes + sector_offset,
1627                                 fs_info->sectorsize);
1628
1629                 extent_changeset_release(data_reserved);
1630                 ret = btrfs_check_data_free_space(inode, &data_reserved, pos,
1631                                                   write_bytes);
1632                 if (ret < 0) {
1633                         if ((BTRFS_I(inode)->flags & (BTRFS_INODE_NODATACOW |
1634                                                       BTRFS_INODE_PREALLOC)) &&
1635                             check_can_nocow(BTRFS_I(inode), pos,
1636                                         &write_bytes) > 0) {
1637                                 /*
1638                                  * For nodata cow case, no need to reserve
1639                                  * data space.
1640                                  */
1641                                 only_release_metadata = true;
1642                                 /*
1643                                  * our prealloc extent may be smaller than
1644                                  * write_bytes, so scale down.
1645                                  */
1646                                 num_pages = DIV_ROUND_UP(write_bytes + offset,
1647                                                          PAGE_SIZE);
1648                                 reserve_bytes = round_up(write_bytes +
1649                                                          sector_offset,
1650                                                          fs_info->sectorsize);
1651                         } else {
1652                                 break;
1653                         }
1654                 }
1655
1656                 WARN_ON(reserve_bytes == 0);
1657                 ret = btrfs_delalloc_reserve_metadata(BTRFS_I(inode),
1658                                 reserve_bytes);
1659                 if (ret) {
1660                         if (!only_release_metadata)
1661                                 btrfs_free_reserved_data_space(inode,
1662                                                 data_reserved, pos,
1663                                                 write_bytes);
1664                         else
1665                                 btrfs_end_write_no_snapshotting(root);
1666                         break;
1667                 }
1668
1669                 release_bytes = reserve_bytes;
1670 again:
1671                 /*
1672                  * This is going to setup the pages array with the number of
1673                  * pages we want, so we don't really need to worry about the
1674                  * contents of pages from loop to loop
1675                  */
1676                 ret = prepare_pages(inode, pages, num_pages,
1677                                     pos, write_bytes,
1678                                     force_page_uptodate);
1679                 if (ret) {
1680                         btrfs_delalloc_release_extents(BTRFS_I(inode),
1681                                                        reserve_bytes, true);
1682                         break;
1683                 }
1684
1685                 extents_locked = lock_and_cleanup_extent_if_need(
1686                                 BTRFS_I(inode), pages,
1687                                 num_pages, pos, write_bytes, &lockstart,
1688                                 &lockend, &cached_state);
1689                 if (extents_locked < 0) {
1690                         if (extents_locked == -EAGAIN)
1691                                 goto again;
1692                         btrfs_delalloc_release_extents(BTRFS_I(inode),
1693                                                        reserve_bytes, true);
1694                         ret = extents_locked;
1695                         break;
1696                 }
1697
1698                 copied = btrfs_copy_from_user(pos, write_bytes, pages, i);
1699
1700                 num_sectors = BTRFS_BYTES_TO_BLKS(fs_info, reserve_bytes);
1701                 dirty_sectors = round_up(copied + sector_offset,
1702                                         fs_info->sectorsize);
1703                 dirty_sectors = BTRFS_BYTES_TO_BLKS(fs_info, dirty_sectors);
1704
1705                 /*
1706                  * if we have trouble faulting in the pages, fall
1707                  * back to one page at a time
1708                  */
1709                 if (copied < write_bytes)
1710                         nrptrs = 1;
1711
1712                 if (copied == 0) {
1713                         force_page_uptodate = true;
1714                         dirty_sectors = 0;
1715                         dirty_pages = 0;
1716                 } else {
1717                         force_page_uptodate = false;
1718                         dirty_pages = DIV_ROUND_UP(copied + offset,
1719                                                    PAGE_SIZE);
1720                 }
1721
1722                 if (num_sectors > dirty_sectors) {
1723                         /* release everything except the sectors we dirtied */
1724                         release_bytes -= dirty_sectors <<
1725                                                 fs_info->sb->s_blocksize_bits;
1726                         if (only_release_metadata) {
1727                                 btrfs_delalloc_release_metadata(BTRFS_I(inode),
1728                                                         release_bytes, true);
1729                         } else {
1730                                 u64 __pos;
1731
1732                                 __pos = round_down(pos,
1733                                                    fs_info->sectorsize) +
1734                                         (dirty_pages << PAGE_SHIFT);
1735                                 btrfs_delalloc_release_space(inode,
1736                                                 data_reserved, __pos,
1737                                                 release_bytes, true);
1738                         }
1739                 }
1740
1741                 release_bytes = round_up(copied + sector_offset,
1742                                         fs_info->sectorsize);
1743
1744                 if (copied > 0)
1745                         ret = btrfs_dirty_pages(inode, pages, dirty_pages,
1746                                                 pos, copied, &cached_state);
1747                 if (extents_locked)
1748                         unlock_extent_cached(&BTRFS_I(inode)->io_tree,
1749                                              lockstart, lockend, &cached_state);
1750                 btrfs_delalloc_release_extents(BTRFS_I(inode), reserve_bytes,
1751                                                true);
1752                 if (ret) {
1753                         btrfs_drop_pages(pages, num_pages);
1754                         break;
1755                 }
1756
1757                 release_bytes = 0;
1758                 if (only_release_metadata)
1759                         btrfs_end_write_no_snapshotting(root);
1760
1761                 if (only_release_metadata && copied > 0) {
1762                         lockstart = round_down(pos,
1763                                                fs_info->sectorsize);
1764                         lockend = round_up(pos + copied,
1765                                            fs_info->sectorsize) - 1;
1766
1767                         set_extent_bit(&BTRFS_I(inode)->io_tree, lockstart,
1768                                        lockend, EXTENT_NORESERVE, NULL,
1769                                        NULL, GFP_NOFS);
1770                         only_release_metadata = false;
1771                 }
1772
1773                 btrfs_drop_pages(pages, num_pages);
1774
1775                 cond_resched();
1776
1777                 balance_dirty_pages_ratelimited(inode->i_mapping);
1778                 if (dirty_pages < (fs_info->nodesize >> PAGE_SHIFT) + 1)
1779                         btrfs_btree_balance_dirty(fs_info);
1780
1781                 pos += copied;
1782                 num_written += copied;
1783         }
1784
1785         kfree(pages);
1786
1787         if (release_bytes) {
1788                 if (only_release_metadata) {
1789                         btrfs_end_write_no_snapshotting(root);
1790                         btrfs_delalloc_release_metadata(BTRFS_I(inode),
1791                                         release_bytes, true);
1792                 } else {
1793                         btrfs_delalloc_release_space(inode, data_reserved,
1794                                         round_down(pos, fs_info->sectorsize),
1795                                         release_bytes, true);
1796                 }
1797         }
1798
1799         extent_changeset_free(data_reserved);
1800         return num_written ? num_written : ret;
1801 }
1802
1803 static ssize_t __btrfs_direct_write(struct kiocb *iocb, struct iov_iter *from)
1804 {
1805         struct file *file = iocb->ki_filp;
1806         struct inode *inode = file_inode(file);
1807         loff_t pos = iocb->ki_pos;
1808         ssize_t written;
1809         ssize_t written_buffered;
1810         loff_t endbyte;
1811         int err;
1812
1813         written = generic_file_direct_write(iocb, from);
1814
1815         if (written < 0 || !iov_iter_count(from))
1816                 return written;
1817
1818         pos += written;
1819         written_buffered = __btrfs_buffered_write(file, from, pos);
1820         if (written_buffered < 0) {
1821                 err = written_buffered;
1822                 goto out;
1823         }
1824         /*
1825          * Ensure all data is persisted. We want the next direct IO read to be
1826          * able to read what was just written.
1827          */
1828         endbyte = pos + written_buffered - 1;
1829         err = btrfs_fdatawrite_range(inode, pos, endbyte);
1830         if (err)
1831                 goto out;
1832         err = filemap_fdatawait_range(inode->i_mapping, pos, endbyte);
1833         if (err)
1834                 goto out;
1835         written += written_buffered;
1836         iocb->ki_pos = pos + written_buffered;
1837         invalidate_mapping_pages(file->f_mapping, pos >> PAGE_SHIFT,
1838                                  endbyte >> PAGE_SHIFT);
1839 out:
1840         return written ? written : err;
1841 }
1842
1843 static void update_time_for_write(struct inode *inode)
1844 {
1845         struct timespec64 now;
1846
1847         if (IS_NOCMTIME(inode))
1848                 return;
1849
1850         now = current_time(inode);
1851         if (!timespec64_equal(&inode->i_mtime, &now))
1852                 inode->i_mtime = now;
1853
1854         if (!timespec64_equal(&inode->i_ctime, &now))
1855                 inode->i_ctime = now;
1856
1857         if (IS_I_VERSION(inode))
1858                 inode_inc_iversion(inode);
1859 }
1860
1861 static ssize_t btrfs_file_write_iter(struct kiocb *iocb,
1862                                     struct iov_iter *from)
1863 {
1864         struct file *file = iocb->ki_filp;
1865         struct inode *inode = file_inode(file);
1866         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1867         struct btrfs_root *root = BTRFS_I(inode)->root;
1868         u64 start_pos;
1869         u64 end_pos;
1870         ssize_t num_written = 0;
1871         bool sync = (file->f_flags & O_DSYNC) || IS_SYNC(file->f_mapping->host);
1872         ssize_t err;
1873         loff_t pos;
1874         size_t count = iov_iter_count(from);
1875         loff_t oldsize;
1876         int clean_page = 0;
1877
1878         if (!(iocb->ki_flags & IOCB_DIRECT) &&
1879             (iocb->ki_flags & IOCB_NOWAIT))
1880                 return -EOPNOTSUPP;
1881
1882         if (!inode_trylock(inode)) {
1883                 if (iocb->ki_flags & IOCB_NOWAIT)
1884                         return -EAGAIN;
1885                 inode_lock(inode);
1886         }
1887
1888         err = generic_write_checks(iocb, from);
1889         if (err <= 0) {
1890                 inode_unlock(inode);
1891                 return err;
1892         }
1893
1894         pos = iocb->ki_pos;
1895         if (iocb->ki_flags & IOCB_NOWAIT) {
1896                 /*
1897                  * We will allocate space in case nodatacow is not set,
1898                  * so bail
1899                  */
1900                 if (!(BTRFS_I(inode)->flags & (BTRFS_INODE_NODATACOW |
1901                                               BTRFS_INODE_PREALLOC)) ||
1902                     check_can_nocow(BTRFS_I(inode), pos, &count) <= 0) {
1903                         inode_unlock(inode);
1904                         return -EAGAIN;
1905                 }
1906         }
1907
1908         current->backing_dev_info = inode_to_bdi(inode);
1909         err = file_remove_privs(file);
1910         if (err) {
1911                 inode_unlock(inode);
1912                 goto out;
1913         }
1914
1915         /*
1916          * If BTRFS flips readonly due to some impossible error
1917          * (fs_info->fs_state now has BTRFS_SUPER_FLAG_ERROR),
1918          * although we have opened a file as writable, we have
1919          * to stop this write operation to ensure FS consistency.
1920          */
1921         if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state)) {
1922                 inode_unlock(inode);
1923                 err = -EROFS;
1924                 goto out;
1925         }
1926
1927         /*
1928          * We reserve space for updating the inode when we reserve space for the
1929          * extent we are going to write, so we will enospc out there.  We don't
1930          * need to start yet another transaction to update the inode as we will
1931          * update the inode when we finish writing whatever data we write.
1932          */
1933         update_time_for_write(inode);
1934
1935         start_pos = round_down(pos, fs_info->sectorsize);
1936         oldsize = i_size_read(inode);
1937         if (start_pos > oldsize) {
1938                 /* Expand hole size to cover write data, preventing empty gap */
1939                 end_pos = round_up(pos + count,
1940                                    fs_info->sectorsize);
1941                 err = btrfs_cont_expand(inode, oldsize, end_pos);
1942                 if (err) {
1943                         inode_unlock(inode);
1944                         goto out;
1945                 }
1946                 if (start_pos > round_up(oldsize, fs_info->sectorsize))
1947                         clean_page = 1;
1948         }
1949
1950         if (sync)
1951                 atomic_inc(&BTRFS_I(inode)->sync_writers);
1952
1953         if (iocb->ki_flags & IOCB_DIRECT) {
1954                 num_written = __btrfs_direct_write(iocb, from);
1955         } else {
1956                 num_written = __btrfs_buffered_write(file, from, pos);
1957                 if (num_written > 0)
1958                         iocb->ki_pos = pos + num_written;
1959                 if (clean_page)
1960                         pagecache_isize_extended(inode, oldsize,
1961                                                 i_size_read(inode));
1962         }
1963
1964         inode_unlock(inode);
1965
1966         /*
1967          * We also have to set last_sub_trans to the current log transid,
1968          * otherwise subsequent syncs to a file that's been synced in this
1969          * transaction will appear to have already occurred.
1970          */
1971         spin_lock(&BTRFS_I(inode)->lock);
1972         BTRFS_I(inode)->last_sub_trans = root->log_transid;
1973         spin_unlock(&BTRFS_I(inode)->lock);
1974         if (num_written > 0)
1975                 num_written = generic_write_sync(iocb, num_written);
1976
1977         if (sync)
1978                 atomic_dec(&BTRFS_I(inode)->sync_writers);
1979 out:
1980         current->backing_dev_info = NULL;
1981         return num_written ? num_written : err;
1982 }
1983
1984 int btrfs_release_file(struct inode *inode, struct file *filp)
1985 {
1986         struct btrfs_file_private *private = filp->private_data;
1987
1988         if (private && private->filldir_buf)
1989                 kfree(private->filldir_buf);
1990         kfree(private);
1991         filp->private_data = NULL;
1992
1993         /*
1994          * ordered_data_close is set by settattr when we are about to truncate
1995          * a file from a non-zero size to a zero size.  This tries to
1996          * flush down new bytes that may have been written if the
1997          * application were using truncate to replace a file in place.
1998          */
1999         if (test_and_clear_bit(BTRFS_INODE_ORDERED_DATA_CLOSE,
2000                                &BTRFS_I(inode)->runtime_flags))
2001                         filemap_flush(inode->i_mapping);
2002         return 0;
2003 }
2004
2005 static int start_ordered_ops(struct inode *inode, loff_t start, loff_t end)
2006 {
2007         int ret;
2008         struct blk_plug plug;
2009
2010         /*
2011          * This is only called in fsync, which would do synchronous writes, so
2012          * a plug can merge adjacent IOs as much as possible.  Esp. in case of
2013          * multiple disks using raid profile, a large IO can be split to
2014          * several segments of stripe length (currently 64K).
2015          */
2016         blk_start_plug(&plug);
2017         atomic_inc(&BTRFS_I(inode)->sync_writers);
2018         ret = btrfs_fdatawrite_range(inode, start, end);
2019         atomic_dec(&BTRFS_I(inode)->sync_writers);
2020         blk_finish_plug(&plug);
2021
2022         return ret;
2023 }
2024
2025 /*
2026  * fsync call for both files and directories.  This logs the inode into
2027  * the tree log instead of forcing full commits whenever possible.
2028  *
2029  * It needs to call filemap_fdatawait so that all ordered extent updates are
2030  * in the metadata btree are up to date for copying to the log.
2031  *
2032  * It drops the inode mutex before doing the tree log commit.  This is an
2033  * important optimization for directories because holding the mutex prevents
2034  * new operations on the dir while we write to disk.
2035  */
2036 int btrfs_sync_file(struct file *file, loff_t start, loff_t end, int datasync)
2037 {
2038         struct dentry *dentry = file_dentry(file);
2039         struct inode *inode = d_inode(dentry);
2040         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2041         struct btrfs_root *root = BTRFS_I(inode)->root;
2042         struct btrfs_trans_handle *trans;
2043         struct btrfs_log_ctx ctx;
2044         int ret = 0, err;
2045         bool full_sync = false;
2046         u64 len;
2047
2048         /*
2049          * The range length can be represented by u64, we have to do the typecasts
2050          * to avoid signed overflow if it's [0, LLONG_MAX] eg. from fsync()
2051          */
2052         len = (u64)end - (u64)start + 1;
2053         trace_btrfs_sync_file(file, datasync);
2054
2055         btrfs_init_log_ctx(&ctx, inode);
2056
2057         /*
2058          * We write the dirty pages in the range and wait until they complete
2059          * out of the ->i_mutex. If so, we can flush the dirty pages by
2060          * multi-task, and make the performance up.  See
2061          * btrfs_wait_ordered_range for an explanation of the ASYNC check.
2062          */
2063         ret = start_ordered_ops(inode, start, end);
2064         if (ret)
2065                 goto out;
2066
2067         inode_lock(inode);
2068         atomic_inc(&root->log_batch);
2069         full_sync = test_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2070                              &BTRFS_I(inode)->runtime_flags);
2071         /*
2072          * We might have have had more pages made dirty after calling
2073          * start_ordered_ops and before acquiring the inode's i_mutex.
2074          */
2075         if (full_sync) {
2076                 /*
2077                  * For a full sync, we need to make sure any ordered operations
2078                  * start and finish before we start logging the inode, so that
2079                  * all extents are persisted and the respective file extent
2080                  * items are in the fs/subvol btree.
2081                  */
2082                 ret = btrfs_wait_ordered_range(inode, start, len);
2083         } else {
2084                 /*
2085                  * Start any new ordered operations before starting to log the
2086                  * inode. We will wait for them to finish in btrfs_sync_log().
2087                  *
2088                  * Right before acquiring the inode's mutex, we might have new
2089                  * writes dirtying pages, which won't immediately start the
2090                  * respective ordered operations - that is done through the
2091                  * fill_delalloc callbacks invoked from the writepage and
2092                  * writepages address space operations. So make sure we start
2093                  * all ordered operations before starting to log our inode. Not
2094                  * doing this means that while logging the inode, writeback
2095                  * could start and invoke writepage/writepages, which would call
2096                  * the fill_delalloc callbacks (cow_file_range,
2097                  * submit_compressed_extents). These callbacks add first an
2098                  * extent map to the modified list of extents and then create
2099                  * the respective ordered operation, which means in
2100                  * tree-log.c:btrfs_log_inode() we might capture all existing
2101                  * ordered operations (with btrfs_get_logged_extents()) before
2102                  * the fill_delalloc callback adds its ordered operation, and by
2103                  * the time we visit the modified list of extent maps (with
2104                  * btrfs_log_changed_extents()), we see and process the extent
2105                  * map they created. We then use the extent map to construct a
2106                  * file extent item for logging without waiting for the
2107                  * respective ordered operation to finish - this file extent
2108                  * item points to a disk location that might not have yet been
2109                  * written to, containing random data - so after a crash a log
2110                  * replay will make our inode have file extent items that point
2111                  * to disk locations containing invalid data, as we returned
2112                  * success to userspace without waiting for the respective
2113                  * ordered operation to finish, because it wasn't captured by
2114                  * btrfs_get_logged_extents().
2115                  */
2116                 ret = start_ordered_ops(inode, start, end);
2117         }
2118         if (ret) {
2119                 inode_unlock(inode);
2120                 goto out;
2121         }
2122         atomic_inc(&root->log_batch);
2123
2124         /*
2125          * If the last transaction that changed this file was before the current
2126          * transaction and we have the full sync flag set in our inode, we can
2127          * bail out now without any syncing.
2128          *
2129          * Note that we can't bail out if the full sync flag isn't set. This is
2130          * because when the full sync flag is set we start all ordered extents
2131          * and wait for them to fully complete - when they complete they update
2132          * the inode's last_trans field through:
2133          *
2134          *     btrfs_finish_ordered_io() ->
2135          *         btrfs_update_inode_fallback() ->
2136          *             btrfs_update_inode() ->
2137          *                 btrfs_set_inode_last_trans()
2138          *
2139          * So we are sure that last_trans is up to date and can do this check to
2140          * bail out safely. For the fast path, when the full sync flag is not
2141          * set in our inode, we can not do it because we start only our ordered
2142          * extents and don't wait for them to complete (that is when
2143          * btrfs_finish_ordered_io runs), so here at this point their last_trans
2144          * value might be less than or equals to fs_info->last_trans_committed,
2145          * and setting a speculative last_trans for an inode when a buffered
2146          * write is made (such as fs_info->generation + 1 for example) would not
2147          * be reliable since after setting the value and before fsync is called
2148          * any number of transactions can start and commit (transaction kthread
2149          * commits the current transaction periodically), and a transaction
2150          * commit does not start nor waits for ordered extents to complete.
2151          */
2152         smp_mb();
2153         if (btrfs_inode_in_log(BTRFS_I(inode), fs_info->generation) ||
2154             (full_sync && BTRFS_I(inode)->last_trans <=
2155              fs_info->last_trans_committed) ||
2156             (!btrfs_have_ordered_extents_in_range(inode, start, len) &&
2157              BTRFS_I(inode)->last_trans
2158              <= fs_info->last_trans_committed)) {
2159                 /*
2160                  * We've had everything committed since the last time we were
2161                  * modified so clear this flag in case it was set for whatever
2162                  * reason, it's no longer relevant.
2163                  */
2164                 clear_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2165                           &BTRFS_I(inode)->runtime_flags);
2166                 /*
2167                  * An ordered extent might have started before and completed
2168                  * already with io errors, in which case the inode was not
2169                  * updated and we end up here. So check the inode's mapping
2170                  * for any errors that might have happened since we last
2171                  * checked called fsync.
2172                  */
2173                 ret = filemap_check_wb_err(inode->i_mapping, file->f_wb_err);
2174                 inode_unlock(inode);
2175                 goto out;
2176         }
2177
2178         /*
2179          * We use start here because we will need to wait on the IO to complete
2180          * in btrfs_sync_log, which could require joining a transaction (for
2181          * example checking cross references in the nocow path).  If we use join
2182          * here we could get into a situation where we're waiting on IO to
2183          * happen that is blocked on a transaction trying to commit.  With start
2184          * we inc the extwriter counter, so we wait for all extwriters to exit
2185          * before we start blocking join'ers.  This comment is to keep somebody
2186          * from thinking they are super smart and changing this to
2187          * btrfs_join_transaction *cough*Josef*cough*.
2188          */
2189         trans = btrfs_start_transaction(root, 0);
2190         if (IS_ERR(trans)) {
2191                 ret = PTR_ERR(trans);
2192                 inode_unlock(inode);
2193                 goto out;
2194         }
2195         trans->sync = true;
2196
2197         ret = btrfs_log_dentry_safe(trans, dentry, start, end, &ctx);
2198         if (ret < 0) {
2199                 /* Fallthrough and commit/free transaction. */
2200                 ret = 1;
2201         }
2202
2203         /* we've logged all the items and now have a consistent
2204          * version of the file in the log.  It is possible that
2205          * someone will come in and modify the file, but that's
2206          * fine because the log is consistent on disk, and we
2207          * have references to all of the file's extents
2208          *
2209          * It is possible that someone will come in and log the
2210          * file again, but that will end up using the synchronization
2211          * inside btrfs_sync_log to keep things safe.
2212          */
2213         inode_unlock(inode);
2214
2215         /*
2216          * If any of the ordered extents had an error, just return it to user
2217          * space, so that the application knows some writes didn't succeed and
2218          * can take proper action (retry for e.g.). Blindly committing the
2219          * transaction in this case, would fool userspace that everything was
2220          * successful. And we also want to make sure our log doesn't contain
2221          * file extent items pointing to extents that weren't fully written to -
2222          * just like in the non fast fsync path, where we check for the ordered
2223          * operation's error flag before writing to the log tree and return -EIO
2224          * if any of them had this flag set (btrfs_wait_ordered_range) -
2225          * therefore we need to check for errors in the ordered operations,
2226          * which are indicated by ctx.io_err.
2227          */
2228         if (ctx.io_err) {
2229                 btrfs_end_transaction(trans);
2230                 ret = ctx.io_err;
2231                 goto out;
2232         }
2233
2234         if (ret != BTRFS_NO_LOG_SYNC) {
2235                 if (!ret) {
2236                         ret = btrfs_sync_log(trans, root, &ctx);
2237                         if (!ret) {
2238                                 ret = btrfs_end_transaction(trans);
2239                                 goto out;
2240                         }
2241                 }
2242                 if (!full_sync) {
2243                         ret = btrfs_wait_ordered_range(inode, start, len);
2244                         if (ret) {
2245                                 btrfs_end_transaction(trans);
2246                                 goto out;
2247                         }
2248                 }
2249                 ret = btrfs_commit_transaction(trans);
2250         } else {
2251                 ret = btrfs_end_transaction(trans);
2252         }
2253 out:
2254         ASSERT(list_empty(&ctx.list));
2255         err = file_check_and_advance_wb_err(file);
2256         if (!ret)
2257                 ret = err;
2258         return ret > 0 ? -EIO : ret;
2259 }
2260
2261 static const struct vm_operations_struct btrfs_file_vm_ops = {
2262         .fault          = filemap_fault,
2263         .map_pages      = filemap_map_pages,
2264         .page_mkwrite   = btrfs_page_mkwrite,
2265 };
2266
2267 static int btrfs_file_mmap(struct file  *filp, struct vm_area_struct *vma)
2268 {
2269         struct address_space *mapping = filp->f_mapping;
2270
2271         if (!mapping->a_ops->readpage)
2272                 return -ENOEXEC;
2273
2274         file_accessed(filp);
2275         vma->vm_ops = &btrfs_file_vm_ops;
2276
2277         return 0;
2278 }
2279
2280 static int hole_mergeable(struct btrfs_inode *inode, struct extent_buffer *leaf,
2281                           int slot, u64 start, u64 end)
2282 {
2283         struct btrfs_file_extent_item *fi;
2284         struct btrfs_key key;
2285
2286         if (slot < 0 || slot >= btrfs_header_nritems(leaf))
2287                 return 0;
2288
2289         btrfs_item_key_to_cpu(leaf, &key, slot);
2290         if (key.objectid != btrfs_ino(inode) ||
2291             key.type != BTRFS_EXTENT_DATA_KEY)
2292                 return 0;
2293
2294         fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
2295
2296         if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG)
2297                 return 0;
2298
2299         if (btrfs_file_extent_disk_bytenr(leaf, fi))
2300                 return 0;
2301
2302         if (key.offset == end)
2303                 return 1;
2304         if (key.offset + btrfs_file_extent_num_bytes(leaf, fi) == start)
2305                 return 1;
2306         return 0;
2307 }
2308
2309 static int fill_holes(struct btrfs_trans_handle *trans,
2310                 struct btrfs_inode *inode,
2311                 struct btrfs_path *path, u64 offset, u64 end)
2312 {
2313         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
2314         struct btrfs_root *root = inode->root;
2315         struct extent_buffer *leaf;
2316         struct btrfs_file_extent_item *fi;
2317         struct extent_map *hole_em;
2318         struct extent_map_tree *em_tree = &inode->extent_tree;
2319         struct btrfs_key key;
2320         int ret;
2321
2322         if (btrfs_fs_incompat(fs_info, NO_HOLES))
2323                 goto out;
2324
2325         key.objectid = btrfs_ino(inode);
2326         key.type = BTRFS_EXTENT_DATA_KEY;
2327         key.offset = offset;
2328
2329         ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
2330         if (ret <= 0) {
2331                 /*
2332                  * We should have dropped this offset, so if we find it then
2333                  * something has gone horribly wrong.
2334                  */
2335                 if (ret == 0)
2336                         ret = -EINVAL;
2337                 return ret;
2338         }
2339
2340         leaf = path->nodes[0];
2341         if (hole_mergeable(inode, leaf, path->slots[0] - 1, offset, end)) {
2342                 u64 num_bytes;
2343
2344                 path->slots[0]--;
2345                 fi = btrfs_item_ptr(leaf, path->slots[0],
2346                                     struct btrfs_file_extent_item);
2347                 num_bytes = btrfs_file_extent_num_bytes(leaf, fi) +
2348                         end - offset;
2349                 btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2350                 btrfs_set_file_extent_ram_bytes(leaf, fi, num_bytes);
2351                 btrfs_set_file_extent_offset(leaf, fi, 0);
2352                 btrfs_mark_buffer_dirty(leaf);
2353                 goto out;
2354         }
2355
2356         if (hole_mergeable(inode, leaf, path->slots[0], offset, end)) {
2357                 u64 num_bytes;
2358
2359                 key.offset = offset;
2360                 btrfs_set_item_key_safe(fs_info, path, &key);
2361                 fi = btrfs_item_ptr(leaf, path->slots[0],
2362                                     struct btrfs_file_extent_item);
2363                 num_bytes = btrfs_file_extent_num_bytes(leaf, fi) + end -
2364                         offset;
2365                 btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2366                 btrfs_set_file_extent_ram_bytes(leaf, fi, num_bytes);
2367                 btrfs_set_file_extent_offset(leaf, fi, 0);
2368                 btrfs_mark_buffer_dirty(leaf);
2369                 goto out;
2370         }
2371         btrfs_release_path(path);
2372
2373         ret = btrfs_insert_file_extent(trans, root, btrfs_ino(inode),
2374                         offset, 0, 0, end - offset, 0, end - offset, 0, 0, 0);
2375         if (ret)
2376                 return ret;
2377
2378 out:
2379         btrfs_release_path(path);
2380
2381         hole_em = alloc_extent_map();
2382         if (!hole_em) {
2383                 btrfs_drop_extent_cache(inode, offset, end - 1, 0);
2384                 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &inode->runtime_flags);
2385         } else {
2386                 hole_em->start = offset;
2387                 hole_em->len = end - offset;
2388                 hole_em->ram_bytes = hole_em->len;
2389                 hole_em->orig_start = offset;
2390
2391                 hole_em->block_start = EXTENT_MAP_HOLE;
2392                 hole_em->block_len = 0;
2393                 hole_em->orig_block_len = 0;
2394                 hole_em->bdev = fs_info->fs_devices->latest_bdev;
2395                 hole_em->compress_type = BTRFS_COMPRESS_NONE;
2396                 hole_em->generation = trans->transid;
2397
2398                 do {
2399                         btrfs_drop_extent_cache(inode, offset, end - 1, 0);
2400                         write_lock(&em_tree->lock);
2401                         ret = add_extent_mapping(em_tree, hole_em, 1);
2402                         write_unlock(&em_tree->lock);
2403                 } while (ret == -EEXIST);
2404                 free_extent_map(hole_em);
2405                 if (ret)
2406                         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2407                                         &inode->runtime_flags);
2408         }
2409
2410         return 0;
2411 }
2412
2413 /*
2414  * Find a hole extent on given inode and change start/len to the end of hole
2415  * extent.(hole/vacuum extent whose em->start <= start &&
2416  *         em->start + em->len > start)
2417  * When a hole extent is found, return 1 and modify start/len.
2418  */
2419 static int find_first_non_hole(struct inode *inode, u64 *start, u64 *len)
2420 {
2421         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2422         struct extent_map *em;
2423         int ret = 0;
2424
2425         em = btrfs_get_extent(BTRFS_I(inode), NULL, 0,
2426                               round_down(*start, fs_info->sectorsize),
2427                               round_up(*len, fs_info->sectorsize), 0);
2428         if (IS_ERR(em))
2429                 return PTR_ERR(em);
2430
2431         /* Hole or vacuum extent(only exists in no-hole mode) */
2432         if (em->block_start == EXTENT_MAP_HOLE) {
2433                 ret = 1;
2434                 *len = em->start + em->len > *start + *len ?
2435                        0 : *start + *len - em->start - em->len;
2436                 *start = em->start + em->len;
2437         }
2438         free_extent_map(em);
2439         return ret;
2440 }
2441
2442 static int btrfs_punch_hole_lock_range(struct inode *inode,
2443                                        const u64 lockstart,
2444                                        const u64 lockend,
2445                                        struct extent_state **cached_state)
2446 {
2447         while (1) {
2448                 struct btrfs_ordered_extent *ordered;
2449                 int ret;
2450
2451                 truncate_pagecache_range(inode, lockstart, lockend);
2452
2453                 lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2454                                  cached_state);
2455                 ordered = btrfs_lookup_first_ordered_extent(inode, lockend);
2456
2457                 /*
2458                  * We need to make sure we have no ordered extents in this range
2459                  * and nobody raced in and read a page in this range, if we did
2460                  * we need to try again.
2461                  */
2462                 if ((!ordered ||
2463                     (ordered->file_offset + ordered->len <= lockstart ||
2464                      ordered->file_offset > lockend)) &&
2465                      !filemap_range_has_page(inode->i_mapping,
2466                                              lockstart, lockend)) {
2467                         if (ordered)
2468                                 btrfs_put_ordered_extent(ordered);
2469                         break;
2470                 }
2471                 if (ordered)
2472                         btrfs_put_ordered_extent(ordered);
2473                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart,
2474                                      lockend, cached_state);
2475                 ret = btrfs_wait_ordered_range(inode, lockstart,
2476                                                lockend - lockstart + 1);
2477                 if (ret)
2478                         return ret;
2479         }
2480         return 0;
2481 }
2482
2483 static int btrfs_punch_hole(struct inode *inode, loff_t offset, loff_t len)
2484 {
2485         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2486         struct btrfs_root *root = BTRFS_I(inode)->root;
2487         struct extent_state *cached_state = NULL;
2488         struct btrfs_path *path;
2489         struct btrfs_block_rsv *rsv;
2490         struct btrfs_trans_handle *trans;
2491         u64 lockstart;
2492         u64 lockend;
2493         u64 tail_start;
2494         u64 tail_len;
2495         u64 orig_start = offset;
2496         u64 cur_offset;
2497         u64 min_size = btrfs_calc_trans_metadata_size(fs_info, 1);
2498         u64 drop_end;
2499         int ret = 0;
2500         int err = 0;
2501         unsigned int rsv_count;
2502         bool same_block;
2503         bool no_holes = btrfs_fs_incompat(fs_info, NO_HOLES);
2504         u64 ino_size;
2505         bool truncated_block = false;
2506         bool updated_inode = false;
2507
2508         ret = btrfs_wait_ordered_range(inode, offset, len);
2509         if (ret)
2510                 return ret;
2511
2512         inode_lock(inode);
2513         ino_size = round_up(inode->i_size, fs_info->sectorsize);
2514         ret = find_first_non_hole(inode, &offset, &len);
2515         if (ret < 0)
2516                 goto out_only_mutex;
2517         if (ret && !len) {
2518                 /* Already in a large hole */
2519                 ret = 0;
2520                 goto out_only_mutex;
2521         }
2522
2523         lockstart = round_up(offset, btrfs_inode_sectorsize(inode));
2524         lockend = round_down(offset + len,
2525                              btrfs_inode_sectorsize(inode)) - 1;
2526         same_block = (BTRFS_BYTES_TO_BLKS(fs_info, offset))
2527                 == (BTRFS_BYTES_TO_BLKS(fs_info, offset + len - 1));
2528         /*
2529          * We needn't truncate any block which is beyond the end of the file
2530          * because we are sure there is no data there.
2531          */
2532         /*
2533          * Only do this if we are in the same block and we aren't doing the
2534          * entire block.
2535          */
2536         if (same_block && len < fs_info->sectorsize) {
2537                 if (offset < ino_size) {
2538                         truncated_block = true;
2539                         ret = btrfs_truncate_block(inode, offset, len, 0);
2540                 } else {
2541                         ret = 0;
2542                 }
2543                 goto out_only_mutex;
2544         }
2545
2546         /* zero back part of the first block */
2547         if (offset < ino_size) {
2548                 truncated_block = true;
2549                 ret = btrfs_truncate_block(inode, offset, 0, 0);
2550                 if (ret) {
2551                         inode_unlock(inode);
2552                         return ret;
2553                 }
2554         }
2555
2556         /* Check the aligned pages after the first unaligned page,
2557          * if offset != orig_start, which means the first unaligned page
2558          * including several following pages are already in holes,
2559          * the extra check can be skipped */
2560         if (offset == orig_start) {
2561                 /* after truncate page, check hole again */
2562                 len = offset + len - lockstart;
2563                 offset = lockstart;
2564                 ret = find_first_non_hole(inode, &offset, &len);
2565                 if (ret < 0)
2566                         goto out_only_mutex;
2567                 if (ret && !len) {
2568                         ret = 0;
2569                         goto out_only_mutex;
2570                 }
2571                 lockstart = offset;
2572         }
2573
2574         /* Check the tail unaligned part is in a hole */
2575         tail_start = lockend + 1;
2576         tail_len = offset + len - tail_start;
2577         if (tail_len) {
2578                 ret = find_first_non_hole(inode, &tail_start, &tail_len);
2579                 if (unlikely(ret < 0))
2580                         goto out_only_mutex;
2581                 if (!ret) {
2582                         /* zero the front end of the last page */
2583                         if (tail_start + tail_len < ino_size) {
2584                                 truncated_block = true;
2585                                 ret = btrfs_truncate_block(inode,
2586                                                         tail_start + tail_len,
2587                                                         0, 1);
2588                                 if (ret)
2589                                         goto out_only_mutex;
2590                         }
2591                 }
2592         }
2593
2594         if (lockend < lockstart) {
2595                 ret = 0;
2596                 goto out_only_mutex;
2597         }
2598
2599         ret = btrfs_punch_hole_lock_range(inode, lockstart, lockend,
2600                                           &cached_state);
2601         if (ret) {
2602                 inode_unlock(inode);
2603                 goto out_only_mutex;
2604         }
2605
2606         path = btrfs_alloc_path();
2607         if (!path) {
2608                 ret = -ENOMEM;
2609                 goto out;
2610         }
2611
2612         rsv = btrfs_alloc_block_rsv(fs_info, BTRFS_BLOCK_RSV_TEMP);
2613         if (!rsv) {
2614                 ret = -ENOMEM;
2615                 goto out_free;
2616         }
2617         rsv->size = btrfs_calc_trans_metadata_size(fs_info, 1);
2618         rsv->failfast = 1;
2619
2620         /*
2621          * 1 - update the inode
2622          * 1 - removing the extents in the range
2623          * 1 - adding the hole extent if no_holes isn't set
2624          */
2625         rsv_count = no_holes ? 2 : 3;
2626         trans = btrfs_start_transaction(root, rsv_count);
2627         if (IS_ERR(trans)) {
2628                 err = PTR_ERR(trans);
2629                 goto out_free;
2630         }
2631
2632         ret = btrfs_block_rsv_migrate(&fs_info->trans_block_rsv, rsv,
2633                                       min_size, 0);
2634         BUG_ON(ret);
2635         trans->block_rsv = rsv;
2636
2637         cur_offset = lockstart;
2638         len = lockend - cur_offset;
2639         while (cur_offset < lockend) {
2640                 ret = __btrfs_drop_extents(trans, root, inode, path,
2641                                            cur_offset, lockend + 1,
2642                                            &drop_end, 1, 0, 0, NULL);
2643                 if (ret != -ENOSPC)
2644                         break;
2645
2646                 trans->block_rsv = &fs_info->trans_block_rsv;
2647
2648                 if (cur_offset < drop_end && cur_offset < ino_size) {
2649                         ret = fill_holes(trans, BTRFS_I(inode), path,
2650                                         cur_offset, drop_end);
2651                         if (ret) {
2652                                 /*
2653                                  * If we failed then we didn't insert our hole
2654                                  * entries for the area we dropped, so now the
2655                                  * fs is corrupted, so we must abort the
2656                                  * transaction.
2657                                  */
2658                                 btrfs_abort_transaction(trans, ret);
2659                                 err = ret;
2660                                 break;
2661                         }
2662                 }
2663
2664                 cur_offset = drop_end;
2665
2666                 ret = btrfs_update_inode(trans, root, inode);
2667                 if (ret) {
2668                         err = ret;
2669                         break;
2670                 }
2671
2672                 btrfs_end_transaction(trans);
2673                 btrfs_btree_balance_dirty(fs_info);
2674
2675                 trans = btrfs_start_transaction(root, rsv_count);
2676                 if (IS_ERR(trans)) {
2677                         ret = PTR_ERR(trans);
2678                         trans = NULL;
2679                         break;
2680                 }
2681
2682                 ret = btrfs_block_rsv_migrate(&fs_info->trans_block_rsv,
2683                                               rsv, min_size, 0);
2684                 BUG_ON(ret);    /* shouldn't happen */
2685                 trans->block_rsv = rsv;
2686
2687                 ret = find_first_non_hole(inode, &cur_offset, &len);
2688                 if (unlikely(ret < 0))
2689                         break;
2690                 if (ret && !len) {
2691                         ret = 0;
2692                         break;
2693                 }
2694         }
2695
2696         if (ret) {
2697                 err = ret;
2698                 goto out_trans;
2699         }
2700
2701         trans->block_rsv = &fs_info->trans_block_rsv;
2702         /*
2703          * If we are using the NO_HOLES feature we might have had already an
2704          * hole that overlaps a part of the region [lockstart, lockend] and
2705          * ends at (or beyond) lockend. Since we have no file extent items to
2706          * represent holes, drop_end can be less than lockend and so we must
2707          * make sure we have an extent map representing the existing hole (the
2708          * call to __btrfs_drop_extents() might have dropped the existing extent
2709          * map representing the existing hole), otherwise the fast fsync path
2710          * will not record the existence of the hole region
2711          * [existing_hole_start, lockend].
2712          */
2713         if (drop_end <= lockend)
2714                 drop_end = lockend + 1;
2715         /*
2716          * Don't insert file hole extent item if it's for a range beyond eof
2717          * (because it's useless) or if it represents a 0 bytes range (when
2718          * cur_offset == drop_end).
2719          */
2720         if (cur_offset < ino_size && cur_offset < drop_end) {
2721                 ret = fill_holes(trans, BTRFS_I(inode), path,
2722                                 cur_offset, drop_end);
2723                 if (ret) {
2724                         /* Same comment as above. */
2725                         btrfs_abort_transaction(trans, ret);
2726                         err = ret;
2727                         goto out_trans;
2728                 }
2729         }
2730
2731 out_trans:
2732         if (!trans)
2733                 goto out_free;
2734
2735         inode_inc_iversion(inode);
2736         inode->i_mtime = inode->i_ctime = current_time(inode);
2737
2738         trans->block_rsv = &fs_info->trans_block_rsv;
2739         ret = btrfs_update_inode(trans, root, inode);
2740         updated_inode = true;
2741         btrfs_end_transaction(trans);
2742         btrfs_btree_balance_dirty(fs_info);
2743 out_free:
2744         btrfs_free_path(path);
2745         btrfs_free_block_rsv(fs_info, rsv);
2746 out:
2747         unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2748                              &cached_state);
2749 out_only_mutex:
2750         if (!updated_inode && truncated_block && !ret && !err) {
2751                 /*
2752                  * If we only end up zeroing part of a page, we still need to
2753                  * update the inode item, so that all the time fields are
2754                  * updated as well as the necessary btrfs inode in memory fields
2755                  * for detecting, at fsync time, if the inode isn't yet in the
2756                  * log tree or it's there but not up to date.
2757                  */
2758                 trans = btrfs_start_transaction(root, 1);
2759                 if (IS_ERR(trans)) {
2760                         err = PTR_ERR(trans);
2761                 } else {
2762                         err = btrfs_update_inode(trans, root, inode);
2763                         ret = btrfs_end_transaction(trans);
2764                 }
2765         }
2766         inode_unlock(inode);
2767         if (ret && !err)
2768                 err = ret;
2769         return err;
2770 }
2771
2772 /* Helper structure to record which range is already reserved */
2773 struct falloc_range {
2774         struct list_head list;
2775         u64 start;
2776         u64 len;
2777 };
2778
2779 /*
2780  * Helper function to add falloc range
2781  *
2782  * Caller should have locked the larger range of extent containing
2783  * [start, len)
2784  */
2785 static int add_falloc_range(struct list_head *head, u64 start, u64 len)
2786 {
2787         struct falloc_range *prev = NULL;
2788         struct falloc_range *range = NULL;
2789
2790         if (list_empty(head))
2791                 goto insert;
2792
2793         /*
2794          * As fallocate iterate by bytenr order, we only need to check
2795          * the last range.
2796          */
2797         prev = list_entry(head->prev, struct falloc_range, list);
2798         if (prev->start + prev->len == start) {
2799                 prev->len += len;
2800                 return 0;
2801         }
2802 insert:
2803         range = kmalloc(sizeof(*range), GFP_KERNEL);
2804         if (!range)
2805                 return -ENOMEM;
2806         range->start = start;
2807         range->len = len;
2808         list_add_tail(&range->list, head);
2809         return 0;
2810 }
2811
2812 static int btrfs_fallocate_update_isize(struct inode *inode,
2813                                         const u64 end,
2814                                         const int mode)
2815 {
2816         struct btrfs_trans_handle *trans;
2817         struct btrfs_root *root = BTRFS_I(inode)->root;
2818         int ret;
2819         int ret2;
2820
2821         if (mode & FALLOC_FL_KEEP_SIZE || end <= i_size_read(inode))
2822                 return 0;
2823
2824         trans = btrfs_start_transaction(root, 1);
2825         if (IS_ERR(trans))
2826                 return PTR_ERR(trans);
2827
2828         inode->i_ctime = current_time(inode);
2829         i_size_write(inode, end);
2830         btrfs_ordered_update_i_size(inode, end, NULL);
2831         ret = btrfs_update_inode(trans, root, inode);
2832         ret2 = btrfs_end_transaction(trans);
2833
2834         return ret ? ret : ret2;
2835 }
2836
2837 enum {
2838         RANGE_BOUNDARY_WRITTEN_EXTENT = 0,
2839         RANGE_BOUNDARY_PREALLOC_EXTENT = 1,
2840         RANGE_BOUNDARY_HOLE = 2,
2841 };
2842
2843 static int btrfs_zero_range_check_range_boundary(struct inode *inode,
2844                                                  u64 offset)
2845 {
2846         const u64 sectorsize = btrfs_inode_sectorsize(inode);
2847         struct extent_map *em;
2848         int ret;
2849
2850         offset = round_down(offset, sectorsize);
2851         em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, offset, sectorsize, 0);
2852         if (IS_ERR(em))
2853                 return PTR_ERR(em);
2854
2855         if (em->block_start == EXTENT_MAP_HOLE)
2856                 ret = RANGE_BOUNDARY_HOLE;
2857         else if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
2858                 ret = RANGE_BOUNDARY_PREALLOC_EXTENT;
2859         else
2860                 ret = RANGE_BOUNDARY_WRITTEN_EXTENT;
2861
2862         free_extent_map(em);
2863         return ret;
2864 }
2865
2866 static int btrfs_zero_range(struct inode *inode,
2867                             loff_t offset,
2868                             loff_t len,
2869                             const int mode)
2870 {
2871         struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
2872         struct extent_map *em;
2873         struct extent_changeset *data_reserved = NULL;
2874         int ret;
2875         u64 alloc_hint = 0;
2876         const u64 sectorsize = btrfs_inode_sectorsize(inode);
2877         u64 alloc_start = round_down(offset, sectorsize);
2878         u64 alloc_end = round_up(offset + len, sectorsize);
2879         u64 bytes_to_reserve = 0;
2880         bool space_reserved = false;
2881
2882         inode_dio_wait(inode);
2883
2884         em = btrfs_get_extent(BTRFS_I(inode), NULL, 0,
2885                               alloc_start, alloc_end - alloc_start, 0);
2886         if (IS_ERR(em)) {
2887                 ret = PTR_ERR(em);
2888                 goto out;
2889         }
2890
2891         /*
2892          * Avoid hole punching and extent allocation for some cases. More cases
2893          * could be considered, but these are unlikely common and we keep things
2894          * as simple as possible for now. Also, intentionally, if the target
2895          * range contains one or more prealloc extents together with regular
2896          * extents and holes, we drop all the existing extents and allocate a
2897          * new prealloc extent, so that we get a larger contiguous disk extent.
2898          */
2899         if (em->start <= alloc_start &&
2900             test_bit(EXTENT_FLAG_PREALLOC, &em->flags)) {
2901                 const u64 em_end = em->start + em->len;
2902
2903                 if (em_end >= offset + len) {
2904                         /*
2905                          * The whole range is already a prealloc extent,
2906                          * do nothing except updating the inode's i_size if
2907                          * needed.
2908                          */
2909                         free_extent_map(em);
2910                         ret = btrfs_fallocate_update_isize(inode, offset + len,
2911                                                            mode);
2912                         goto out;
2913                 }
2914                 /*
2915                  * Part of the range is already a prealloc extent, so operate
2916                  * only on the remaining part of the range.
2917                  */
2918                 alloc_start = em_end;
2919                 ASSERT(IS_ALIGNED(alloc_start, sectorsize));
2920                 len = offset + len - alloc_start;
2921                 offset = alloc_start;
2922                 alloc_hint = em->block_start + em->len;
2923         }
2924         free_extent_map(em);
2925
2926         if (BTRFS_BYTES_TO_BLKS(fs_info, offset) ==
2927             BTRFS_BYTES_TO_BLKS(fs_info, offset + len - 1)) {
2928                 em = btrfs_get_extent(BTRFS_I(inode), NULL, 0,
2929                                       alloc_start, sectorsize, 0);
2930                 if (IS_ERR(em)) {
2931                         ret = PTR_ERR(em);
2932                         goto out;
2933                 }
2934
2935                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags)) {
2936                         free_extent_map(em);
2937                         ret = btrfs_fallocate_update_isize(inode, offset + len,
2938                                                            mode);
2939                         goto out;
2940                 }
2941                 if (len < sectorsize && em->block_start != EXTENT_MAP_HOLE) {
2942                         free_extent_map(em);
2943                         ret = btrfs_truncate_block(inode, offset, len, 0);
2944                         if (!ret)
2945                                 ret = btrfs_fallocate_update_isize(inode,
2946                                                                    offset + len,
2947                                                                    mode);
2948                         return ret;
2949                 }
2950                 free_extent_map(em);
2951                 alloc_start = round_down(offset, sectorsize);
2952                 alloc_end = alloc_start + sectorsize;
2953                 goto reserve_space;
2954         }
2955
2956         alloc_start = round_up(offset, sectorsize);
2957         alloc_end = round_down(offset + len, sectorsize);
2958
2959         /*
2960          * For unaligned ranges, check the pages at the boundaries, they might
2961          * map to an extent, in which case we need to partially zero them, or
2962          * they might map to a hole, in which case we need our allocation range
2963          * to cover them.
2964          */
2965         if (!IS_ALIGNED(offset, sectorsize)) {
2966                 ret = btrfs_zero_range_check_range_boundary(inode, offset);
2967                 if (ret < 0)
2968                         goto out;
2969                 if (ret == RANGE_BOUNDARY_HOLE) {
2970                         alloc_start = round_down(offset, sectorsize);
2971                         ret = 0;
2972                 } else if (ret == RANGE_BOUNDARY_WRITTEN_EXTENT) {
2973                         ret = btrfs_truncate_block(inode, offset, 0, 0);
2974                         if (ret)
2975                                 goto out;
2976                 } else {
2977                         ret = 0;
2978                 }
2979         }
2980
2981         if (!IS_ALIGNED(offset + len, sectorsize)) {
2982                 ret = btrfs_zero_range_check_range_boundary(inode,
2983                                                             offset + len);
2984                 if (ret < 0)
2985                         goto out;
2986                 if (ret == RANGE_BOUNDARY_HOLE) {
2987                         alloc_end = round_up(offset + len, sectorsize);
2988                         ret = 0;
2989                 } else if (ret == RANGE_BOUNDARY_WRITTEN_EXTENT) {
2990                         ret = btrfs_truncate_block(inode, offset + len, 0, 1);
2991                         if (ret)
2992                                 goto out;
2993                 } else {
2994                         ret = 0;
2995                 }
2996         }
2997
2998 reserve_space:
2999         if (alloc_start < alloc_end) {
3000                 struct extent_state *cached_state = NULL;
3001                 const u64 lockstart = alloc_start;
3002                 const u64 lockend = alloc_end - 1;
3003
3004                 bytes_to_reserve = alloc_end - alloc_start;
3005                 ret = btrfs_alloc_data_chunk_ondemand(BTRFS_I(inode),
3006                                                       bytes_to_reserve);
3007                 if (ret < 0)
3008                         goto out;
3009                 space_reserved = true;
3010                 ret = btrfs_qgroup_reserve_data(inode, &data_reserved,
3011                                                 alloc_start, bytes_to_reserve);
3012                 if (ret)
3013                         goto out;
3014                 ret = btrfs_punch_hole_lock_range(inode, lockstart, lockend,
3015                                                   &cached_state);
3016                 if (ret)
3017                         goto out;
3018                 ret = btrfs_prealloc_file_range(inode, mode, alloc_start,
3019                                                 alloc_end - alloc_start,
3020                                                 i_blocksize(inode),
3021                                                 offset + len, &alloc_hint);
3022                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart,
3023                                      lockend, &cached_state);
3024                 /* btrfs_prealloc_file_range releases reserved space on error */
3025                 if (ret) {
3026                         space_reserved = false;
3027                         goto out;
3028                 }
3029         }
3030         ret = btrfs_fallocate_update_isize(inode, offset + len, mode);
3031  out:
3032         if (ret && space_reserved)
3033                 btrfs_free_reserved_data_space(inode, data_reserved,
3034                                                alloc_start, bytes_to_reserve);
3035         extent_changeset_free(data_reserved);
3036
3037         return ret;
3038 }
3039
3040 static long btrfs_fallocate(struct file *file, int mode,
3041                             loff_t offset, loff_t len)
3042 {
3043         struct inode *inode = file_inode(file);
3044         struct extent_state *cached_state = NULL;
3045         struct extent_changeset *data_reserved = NULL;
3046         struct falloc_range *range;
3047         struct falloc_range *tmp;
3048         struct list_head reserve_list;
3049         u64 cur_offset;
3050         u64 last_byte;
3051         u64 alloc_start;
3052         u64 alloc_end;
3053         u64 alloc_hint = 0;
3054         u64 locked_end;
3055         u64 actual_end = 0;
3056         struct extent_map *em;
3057         int blocksize = btrfs_inode_sectorsize(inode);
3058         int ret;
3059
3060         alloc_start = round_down(offset, blocksize);
3061         alloc_end = round_up(offset + len, blocksize);
3062         cur_offset = alloc_start;
3063
3064         /* Make sure we aren't being give some crap mode */
3065         if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE |
3066                      FALLOC_FL_ZERO_RANGE))
3067                 return -EOPNOTSUPP;
3068
3069         if (mode & FALLOC_FL_PUNCH_HOLE)
3070                 return btrfs_punch_hole(inode, offset, len);
3071
3072         /*
3073          * Only trigger disk allocation, don't trigger qgroup reserve
3074          *
3075          * For qgroup space, it will be checked later.
3076          */
3077         if (!(mode & FALLOC_FL_ZERO_RANGE)) {
3078                 ret = btrfs_alloc_data_chunk_ondemand(BTRFS_I(inode),
3079                                                       alloc_end - alloc_start);
3080                 if (ret < 0)
3081                         return ret;
3082         }
3083
3084         inode_lock(inode);
3085
3086         if (!(mode & FALLOC_FL_KEEP_SIZE) && offset + len > inode->i_size) {
3087                 ret = inode_newsize_ok(inode, offset + len);
3088                 if (ret)
3089                         goto out;
3090         }
3091
3092         /*
3093          * TODO: Move these two operations after we have checked
3094          * accurate reserved space, or fallocate can still fail but
3095          * with page truncated or size expanded.
3096          *
3097          * But that's a minor problem and won't do much harm BTW.
3098          */
3099         if (alloc_start > inode->i_size) {
3100                 ret = btrfs_cont_expand(inode, i_size_read(inode),
3101                                         alloc_start);
3102                 if (ret)
3103                         goto out;
3104         } else if (offset + len > inode->i_size) {
3105                 /*
3106                  * If we are fallocating from the end of the file onward we
3107                  * need to zero out the end of the block if i_size lands in the
3108                  * middle of a block.
3109                  */
3110                 ret = btrfs_truncate_block(inode, inode->i_size, 0, 0);
3111                 if (ret)
3112                         goto out;
3113         }
3114
3115         /*
3116          * wait for ordered IO before we have any locks.  We'll loop again
3117          * below with the locks held.
3118          */
3119         ret = btrfs_wait_ordered_range(inode, alloc_start,
3120                                        alloc_end - alloc_start);
3121         if (ret)
3122                 goto out;
3123
3124         if (mode & FALLOC_FL_ZERO_RANGE) {
3125                 ret = btrfs_zero_range(inode, offset, len, mode);
3126                 inode_unlock(inode);
3127                 return ret;
3128         }
3129
3130         locked_end = alloc_end - 1;
3131         while (1) {
3132                 struct btrfs_ordered_extent *ordered;
3133
3134                 /* the extent lock is ordered inside the running
3135                  * transaction
3136                  */
3137                 lock_extent_bits(&BTRFS_I(inode)->io_tree, alloc_start,
3138                                  locked_end, &cached_state);
3139                 ordered = btrfs_lookup_first_ordered_extent(inode, locked_end);
3140
3141                 if (ordered &&
3142                     ordered->file_offset + ordered->len > alloc_start &&
3143                     ordered->file_offset < alloc_end) {
3144                         btrfs_put_ordered_extent(ordered);
3145                         unlock_extent_cached(&BTRFS_I(inode)->io_tree,
3146                                              alloc_start, locked_end,
3147                                              &cached_state);
3148                         /*
3149                          * we can't wait on the range with the transaction
3150                          * running or with the extent lock held
3151                          */
3152                         ret = btrfs_wait_ordered_range(inode, alloc_start,
3153                                                        alloc_end - alloc_start);
3154                         if (ret)
3155                                 goto out;
3156                 } else {
3157                         if (ordered)
3158                                 btrfs_put_ordered_extent(ordered);
3159                         break;
3160                 }
3161         }
3162
3163         /* First, check if we exceed the qgroup limit */
3164         INIT_LIST_HEAD(&reserve_list);
3165         while (cur_offset < alloc_end) {
3166                 em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, cur_offset,
3167                                       alloc_end - cur_offset, 0);
3168                 if (IS_ERR(em)) {
3169                         ret = PTR_ERR(em);
3170                         break;
3171                 }
3172                 last_byte = min(extent_map_end(em), alloc_end);
3173                 actual_end = min_t(u64, extent_map_end(em), offset + len);
3174                 last_byte = ALIGN(last_byte, blocksize);
3175                 if (em->block_start == EXTENT_MAP_HOLE ||
3176                     (cur_offset >= inode->i_size &&
3177                      !test_bit(EXTENT_FLAG_PREALLOC, &em->flags))) {
3178                         ret = add_falloc_range(&reserve_list, cur_offset,
3179                                                last_byte - cur_offset);
3180                         if (ret < 0) {
3181                                 free_extent_map(em);
3182                                 break;
3183                         }
3184                         ret = btrfs_qgroup_reserve_data(inode, &data_reserved,
3185                                         cur_offset, last_byte - cur_offset);
3186                         if (ret < 0) {
3187                                 free_extent_map(em);
3188                                 break;
3189                         }
3190                 } else {
3191                         /*
3192                          * Do not need to reserve unwritten extent for this
3193                          * range, free reserved data space first, otherwise
3194                          * it'll result in false ENOSPC error.
3195                          */
3196                         btrfs_free_reserved_data_space(inode, data_reserved,
3197                                         cur_offset, last_byte - cur_offset);
3198                 }
3199                 free_extent_map(em);
3200                 cur_offset = last_byte;
3201         }
3202
3203         /*
3204          * If ret is still 0, means we're OK to fallocate.
3205          * Or just cleanup the list and exit.
3206          */
3207         list_for_each_entry_safe(range, tmp, &reserve_list, list) {
3208                 if (!ret)
3209                         ret = btrfs_prealloc_file_range(inode, mode,
3210                                         range->start,
3211                                         range->len, i_blocksize(inode),
3212                                         offset + len, &alloc_hint);
3213                 else
3214                         btrfs_free_reserved_data_space(inode,
3215                                         data_reserved, range->start,
3216                                         range->len);
3217                 list_del(&range->list);
3218                 kfree(range);
3219         }
3220         if (ret < 0)
3221                 goto out_unlock;
3222
3223         /*
3224          * We didn't need to allocate any more space, but we still extended the
3225          * size of the file so we need to update i_size and the inode item.
3226          */
3227         ret = btrfs_fallocate_update_isize(inode, actual_end, mode);
3228 out_unlock:
3229         unlock_extent_cached(&BTRFS_I(inode)->io_tree, alloc_start, locked_end,
3230                              &cached_state);
3231 out:
3232         inode_unlock(inode);
3233         /* Let go of our reservation. */
3234         if (ret != 0 && !(mode & FALLOC_FL_ZERO_RANGE))
3235                 btrfs_free_reserved_data_space(inode, data_reserved,
3236                                 alloc_start, alloc_end - cur_offset);
3237         extent_changeset_free(data_reserved);
3238         return ret;
3239 }
3240
3241 static int find_desired_extent(struct inode *inode, loff_t *offset, int whence)
3242 {
3243         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
3244         struct extent_map *em = NULL;
3245         struct extent_state *cached_state = NULL;
3246         u64 lockstart;
3247         u64 lockend;
3248         u64 start;
3249         u64 len;
3250         int ret = 0;
3251
3252         if (inode->i_size == 0)
3253                 return -ENXIO;
3254
3255         /*
3256          * *offset can be negative, in this case we start finding DATA/HOLE from
3257          * the very start of the file.
3258          */
3259         start = max_t(loff_t, 0, *offset);
3260
3261         lockstart = round_down(start, fs_info->sectorsize);
3262         lockend = round_up(i_size_read(inode),
3263                            fs_info->sectorsize);
3264         if (lockend <= lockstart)
3265                 lockend = lockstart + fs_info->sectorsize;
3266         lockend--;
3267         len = lockend - lockstart + 1;
3268
3269         lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend,
3270                          &cached_state);
3271
3272         while (start < inode->i_size) {
3273                 em = btrfs_get_extent_fiemap(BTRFS_I(inode), NULL, 0,
3274                                 start, len, 0);
3275                 if (IS_ERR(em)) {
3276                         ret = PTR_ERR(em);
3277                         em = NULL;
3278                         break;
3279                 }
3280
3281                 if (whence == SEEK_HOLE &&
3282                     (em->block_start == EXTENT_MAP_HOLE ||
3283                      test_bit(EXTENT_FLAG_PREALLOC, &em->flags)))
3284                         break;
3285                 else if (whence == SEEK_DATA &&
3286                            (em->block_start != EXTENT_MAP_HOLE &&
3287                             !test_bit(EXTENT_FLAG_PREALLOC, &em->flags)))
3288                         break;
3289
3290                 start = em->start + em->len;
3291                 free_extent_map(em);
3292                 em = NULL;
3293                 cond_resched();
3294         }
3295         free_extent_map(em);
3296         if (!ret) {
3297                 if (whence == SEEK_DATA && start >= inode->i_size)
3298                         ret = -ENXIO;
3299                 else
3300                         *offset = min_t(loff_t, start, inode->i_size);
3301         }
3302         unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
3303                              &cached_state);
3304         return ret;
3305 }
3306
3307 static loff_t btrfs_file_llseek(struct file *file, loff_t offset, int whence)
3308 {
3309         struct inode *inode = file->f_mapping->host;
3310         int ret;
3311
3312         inode_lock(inode);
3313         switch (whence) {
3314         case SEEK_END:
3315         case SEEK_CUR:
3316                 offset = generic_file_llseek(file, offset, whence);
3317                 goto out;
3318         case SEEK_DATA:
3319         case SEEK_HOLE:
3320                 if (offset >= i_size_read(inode)) {
3321                         inode_unlock(inode);
3322                         return -ENXIO;
3323                 }
3324
3325                 ret = find_desired_extent(inode, &offset, whence);
3326                 if (ret) {
3327                         inode_unlock(inode);
3328                         return ret;
3329                 }
3330         }
3331
3332         offset = vfs_setpos(file, offset, inode->i_sb->s_maxbytes);
3333 out:
3334         inode_unlock(inode);
3335         return offset;
3336 }
3337
3338 static int btrfs_file_open(struct inode *inode, struct file *filp)
3339 {
3340         filp->f_mode |= FMODE_NOWAIT;
3341         return generic_file_open(inode, filp);
3342 }
3343
3344 const struct file_operations btrfs_file_operations = {
3345         .llseek         = btrfs_file_llseek,
3346         .read_iter      = generic_file_read_iter,
3347         .splice_read    = generic_file_splice_read,
3348         .write_iter     = btrfs_file_write_iter,
3349         .mmap           = btrfs_file_mmap,
3350         .open           = btrfs_file_open,
3351         .release        = btrfs_release_file,
3352         .fsync          = btrfs_sync_file,
3353         .fallocate      = btrfs_fallocate,
3354         .unlocked_ioctl = btrfs_ioctl,
3355 #ifdef CONFIG_COMPAT
3356         .compat_ioctl   = btrfs_compat_ioctl,
3357 #endif
3358         .clone_file_range = btrfs_clone_file_range,
3359         .dedupe_file_range = btrfs_dedupe_file_range,
3360 };
3361
3362 void __cold btrfs_auto_defrag_exit(void)
3363 {
3364         kmem_cache_destroy(btrfs_inode_defrag_cachep);
3365 }
3366
3367 int __init btrfs_auto_defrag_init(void)
3368 {
3369         btrfs_inode_defrag_cachep = kmem_cache_create("btrfs_inode_defrag",
3370                                         sizeof(struct inode_defrag), 0,
3371                                         SLAB_MEM_SPREAD,
3372                                         NULL);
3373         if (!btrfs_inode_defrag_cachep)
3374                 return -ENOMEM;
3375
3376         return 0;
3377 }
3378
3379 int btrfs_fdatawrite_range(struct inode *inode, loff_t start, loff_t end)
3380 {
3381         int ret;
3382
3383         /*
3384          * So with compression we will find and lock a dirty page and clear the
3385          * first one as dirty, setup an async extent, and immediately return
3386          * with the entire range locked but with nobody actually marked with
3387          * writeback.  So we can't just filemap_write_and_wait_range() and
3388          * expect it to work since it will just kick off a thread to do the
3389          * actual work.  So we need to call filemap_fdatawrite_range _again_
3390          * since it will wait on the page lock, which won't be unlocked until
3391          * after the pages have been marked as writeback and so we're good to go
3392          * from there.  We have to do this otherwise we'll miss the ordered
3393          * extents and that results in badness.  Please Josef, do not think you
3394          * know better and pull this out at some point in the future, it is
3395          * right and you are wrong.
3396          */
3397         ret = filemap_fdatawrite_range(inode->i_mapping, start, end);
3398         if (!ret && test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
3399                              &BTRFS_I(inode)->runtime_flags))
3400                 ret = filemap_fdatawrite_range(inode->i_mapping, start, end);
3401
3402         return ret;
3403 }