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/time.h>
   9 #include <linux/init.h>
  10 #include <linux/string.h>
  11 #include <linux/backing-dev.h>
  12 #include <linux/falloc.h>
  13 #include <linux/writeback.h>
  14 #include <linux/compat.h>
  15 #include <linux/slab.h>
  16 #include <linux/btrfs.h>
  17 #include <linux/uio.h>
  18 #include <linux/iversion.h>
  19 #include "ctree.h"
  20 #include "disk-io.h"
  21 #include "transaction.h"
  22 #include "btrfs_inode.h"
  23 #include "print-tree.h"
  24 #include "tree-log.h"
  25 #include "locking.h"
  26 #include "volumes.h"
  27 #include "qgroup.h"
  28 #include "compression.h"
  29 #include "delalloc-space.h"
  30 #include "reflink.h"
  31
  32 static struct kmem_cache *btrfs_inode_defrag_cachep;
  33 /*
  34  * when auto defrag is enabled we
  35  * queue up these defrag structs to remember which
  36  * inodes need defragging passes
  37  */
  38 struct inode_defrag {
  39         struct rb_node rb_node;
  40         /* objectid */
  41         u64 ino;
  42         /*
  43          * transid where the defrag was added, we search for
  44          * extents newer than this
  45          */
  46         u64 transid;
  47
  48         /* root objectid */
  49         u64 root;
  50
  51         /* last offset we were able to defrag */
  52         u64 last_offset;
  53
  54         /* if we've wrapped around back to zero once already */
  55         int cycled;
  56 };
  57
  58 static int __compare_inode_defrag(struct inode_defrag *defrag1,
  59                                   struct inode_defrag *defrag2)
  60 {
  61         if (defrag1->root > defrag2->root)
  62                 return 1;
  63         else if (defrag1->root < defrag2->root)
  64                 return -1;
  65         else if (defrag1->ino > defrag2->ino)
  66                 return 1;
  67         else if (defrag1->ino < defrag2->ino)
  68                 return -1;
  69         else
  70                 return 0;
  71 }
  72
  73 /* pop a record for an inode into the defrag tree.  The lock
  74  * must be held already
  75  *
  76  * If you're inserting a record for an older transid than an
  77  * existing record, the transid already in the tree is lowered
  78  *
  79  * If an existing record is found the defrag item you
  80  * pass in is freed
  81  */
  82 static int __btrfs_add_inode_defrag(struct btrfs_inode *inode,
  83                                     struct inode_defrag *defrag)
  84 {
  85         struct btrfs_fs_info *fs_info = inode->root->fs_info;
  86         struct inode_defrag *entry;
  87         struct rb_node **p;
  88         struct rb_node *parent = NULL;
  89         int ret;
  90
  91         p = &fs_info->defrag_inodes.rb_node;
  92         while (*p) {
  93                 parent = *p;
  94                 entry = rb_entry(parent, struct inode_defrag, rb_node);
  95
  96                 ret = __compare_inode_defrag(defrag, entry);
  97                 if (ret < 0)
  98                         p = &parent->rb_left;
  99                 else if (ret > 0)
 100                         p = &parent->rb_right;
 101                 else {
 102                         /* if we're reinserting an entry for
 103                          * an old defrag run, make sure to
 104                          * lower the transid of our existing record
 105                          */
 106                         if (defrag->transid < entry->transid)
 107                                 entry->transid = defrag->transid;
 108                         if (defrag->last_offset > entry->last_offset)
 109                                 entry->last_offset = defrag->last_offset;
 110                         return -EEXIST;
 111                 }
 112         }
 113         set_bit(BTRFS_INODE_IN_DEFRAG, &inode->runtime_flags);
 114         rb_link_node(&defrag->rb_node, parent, p);
 115         rb_insert_color(&defrag->rb_node, &fs_info->defrag_inodes);
 116         return 0;
 117 }
 118
 119 static inline int __need_auto_defrag(struct btrfs_fs_info *fs_info)
 120 {
 121         if (!btrfs_test_opt(fs_info, AUTO_DEFRAG))
 122                 return 0;
 123
 124         if (btrfs_fs_closing(fs_info))
 125                 return 0;
 126
 127         return 1;
 128 }
 129
 130 /*
 131  * insert a defrag record for this inode if auto defrag is
 132  * enabled
 133  */
 134 int btrfs_add_inode_defrag(struct btrfs_trans_handle *trans,
 135                            struct btrfs_inode *inode)
 136 {
 137         struct btrfs_root *root = inode->root;
 138         struct btrfs_fs_info *fs_info = root->fs_info;
 139         struct inode_defrag *defrag;
 140         u64 transid;
 141         int ret;
 142
 143         if (!__need_auto_defrag(fs_info))
 144                 return 0;
 145
 146         if (test_bit(BTRFS_INODE_IN_DEFRAG, &inode->runtime_flags))
 147                 return 0;
 148
 149         if (trans)
 150                 transid = trans->transid;
 151         else
 152                 transid = inode->root->last_trans;
 153
 154         defrag = kmem_cache_zalloc(btrfs_inode_defrag_cachep, GFP_NOFS);
 155         if (!defrag)
 156                 return -ENOMEM;
 157
 158         defrag->ino = btrfs_ino(inode);
 159         defrag->transid = transid;
 160         defrag->root = root->root_key.objectid;
 161
 162         spin_lock(&fs_info->defrag_inodes_lock);
 163         if (!test_bit(BTRFS_INODE_IN_DEFRAG, &inode->runtime_flags)) {
 164                 /*
 165                  * If we set IN_DEFRAG flag and evict the inode from memory,
 166                  * and then re-read this inode, this new inode doesn't have
 167                  * IN_DEFRAG flag. At the case, we may find the existed defrag.
 168                  */
 169                 ret = __btrfs_add_inode_defrag(inode, defrag);
 170                 if (ret)
 171                         kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
 172         } else {
 173                 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
 174         }
 175         spin_unlock(&fs_info->defrag_inodes_lock);
 176         return 0;
 177 }
 178
 179 /*
 180  * Requeue the defrag object. If there is a defrag object that points to
 181  * the same inode in the tree, we will merge them together (by
 182  * __btrfs_add_inode_defrag()) and free the one that we want to requeue.
 183  */
 184 static void btrfs_requeue_inode_defrag(struct btrfs_inode *inode,
 185                                        struct inode_defrag *defrag)
 186 {
 187         struct btrfs_fs_info *fs_info = inode->root->fs_info;
 188         int ret;
 189
 190         if (!__need_auto_defrag(fs_info))
 191                 goto out;
 192
 193         /*
 194          * Here we don't check the IN_DEFRAG flag, because we need merge
 195          * them together.
 196          */
 197         spin_lock(&fs_info->defrag_inodes_lock);
 198         ret = __btrfs_add_inode_defrag(inode, defrag);
 199         spin_unlock(&fs_info->defrag_inodes_lock);
 200         if (ret)
 201                 goto out;
 202         return;
 203 out:
 204         kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
 205 }
 206
 207 /*
 208  * pick the defragable inode that we want, if it doesn't exist, we will get
 209  * the next one.
 210  */
 211 static struct inode_defrag *
 212 btrfs_pick_defrag_inode(struct btrfs_fs_info *fs_info, u64 root, u64 ino)
 213 {
 214         struct inode_defrag *entry = NULL;
 215         struct inode_defrag tmp;
 216         struct rb_node *p;
 217         struct rb_node *parent = NULL;
 218         int ret;
 219
 220         tmp.ino = ino;
 221         tmp.root = root;
 222
 223         spin_lock(&fs_info->defrag_inodes_lock);
 224         p = fs_info->defrag_inodes.rb_node;
 225         while (p) {
 226                 parent = p;
 227                 entry = rb_entry(parent, struct inode_defrag, rb_node);
 228
 229                 ret = __compare_inode_defrag(&tmp, entry);
 230                 if (ret < 0)
 231                         p = parent->rb_left;
 232                 else if (ret > 0)
 233                         p = parent->rb_right;
 234                 else
 235                         goto out;
 236         }
 237
 238         if (parent && __compare_inode_defrag(&tmp, entry) > 0) {
 239                 parent = rb_next(parent);
 240                 if (parent)
 241                         entry = rb_entry(parent, struct inode_defrag, rb_node);
 242                 else
 243                         entry = NULL;
 244         }
 245 out:
 246         if (entry)
 247                 rb_erase(parent, &fs_info->defrag_inodes);
 248         spin_unlock(&fs_info->defrag_inodes_lock);
 249         return entry;
 250 }
 251
 252 void btrfs_cleanup_defrag_inodes(struct btrfs_fs_info *fs_info)
 253 {
 254         struct inode_defrag *defrag;
 255         struct rb_node *node;
 256
 257         spin_lock(&fs_info->defrag_inodes_lock);
 258         node = rb_first(&fs_info->defrag_inodes);
 259         while (node) {
 260                 rb_erase(node, &fs_info->defrag_inodes);
 261                 defrag = rb_entry(node, struct inode_defrag, rb_node);
 262                 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
 263
 264                 cond_resched_lock(&fs_info->defrag_inodes_lock);
 265
 266                 node = rb_first(&fs_info->defrag_inodes);
 267         }
 268         spin_unlock(&fs_info->defrag_inodes_lock);
 269 }
 270
 271 #define BTRFS_DEFRAG_BATCH      1024
 272
 273 static int __btrfs_run_defrag_inode(struct btrfs_fs_info *fs_info,
 274                                     struct inode_defrag *defrag)
 275 {
 276         struct btrfs_root *inode_root;
 277         struct inode *inode;
 278         struct btrfs_ioctl_defrag_range_args range;
 279         int num_defrag;
 280         int ret;
 281
 282         /* get the inode */
 283         inode_root = btrfs_get_fs_root(fs_info, defrag->root, true);
 284         if (IS_ERR(inode_root)) {
 285                 ret = PTR_ERR(inode_root);
 286                 goto cleanup;
 287         }
 288
 289         inode = btrfs_iget(fs_info->sb, defrag->ino, inode_root);
 290         btrfs_put_root(inode_root);
 291         if (IS_ERR(inode)) {
 292                 ret = PTR_ERR(inode);
 293                 goto cleanup;
 294         }
 295
 296         /* do a chunk of defrag */
 297         clear_bit(BTRFS_INODE_IN_DEFRAG, &BTRFS_I(inode)->runtime_flags);
 298         memset(&range, 0, sizeof(range));
 299         range.len = (u64)-1;
 300         range.start = defrag->last_offset;
 301
 302         sb_start_write(fs_info->sb);
 303         num_defrag = btrfs_defrag_file(inode, NULL, &range, defrag->transid,
 304                                        BTRFS_DEFRAG_BATCH);
 305         sb_end_write(fs_info->sb);
 306         /*
 307          * if we filled the whole defrag batch, there
 308          * must be more work to do.  Queue this defrag
 309          * again
 310          */
 311         if (num_defrag == BTRFS_DEFRAG_BATCH) {
 312                 defrag->last_offset = range.start;
 313                 btrfs_requeue_inode_defrag(BTRFS_I(inode), defrag);
 314         } else if (defrag->last_offset && !defrag->cycled) {
 315                 /*
 316                  * we didn't fill our defrag batch, but
 317                  * we didn't start at zero.  Make sure we loop
 318                  * around to the start of the file.
 319                  */
 320                 defrag->last_offset = 0;
 321                 defrag->cycled = 1;
 322                 btrfs_requeue_inode_defrag(BTRFS_I(inode), defrag);
 323         } else {
 324                 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
 325         }
 326
 327         iput(inode);
 328         return 0;
 329 cleanup:
 330         kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
 331         return ret;
 332 }
 333
 334 /*
 335  * run through the list of inodes in the FS that need
 336  * defragging
 337  */
 338 int btrfs_run_defrag_inodes(struct btrfs_fs_info *fs_info)
 339 {
 340         struct inode_defrag *defrag;
 341         u64 first_ino = 0;
 342         u64 root_objectid = 0;
 343
 344         atomic_inc(&fs_info->defrag_running);
 345         while (1) {
 346                 /* Pause the auto defragger. */
 347                 if (test_bit(BTRFS_FS_STATE_REMOUNTING,
 348                              &fs_info->fs_state))
 349                         break;
 350
 351                 if (!__need_auto_defrag(fs_info))
 352                         break;
 353
 354                 /* find an inode to defrag */
 355                 defrag = btrfs_pick_defrag_inode(fs_info, root_objectid,
 356                                                  first_ino);
 357                 if (!defrag) {
 358                         if (root_objectid || first_ino) {
 359                                 root_objectid = 0;
 360                                 first_ino = 0;
 361                                 continue;
 362                         } else {
 363                                 break;
 364                         }
 365                 }
 366
 367                 first_ino = defrag->ino + 1;
 368                 root_objectid = defrag->root;
 369
 370                 __btrfs_run_defrag_inode(fs_info, defrag);
 371         }
 372         atomic_dec(&fs_info->defrag_running);
 373
 374         /*
 375          * during unmount, we use the transaction_wait queue to
 376          * wait for the defragger to stop
 377          */
 378         wake_up(&fs_info->transaction_wait);
 379         return 0;
 380 }
 381
 382 /* simple helper to fault in pages and copy.  This should go away
 383  * and be replaced with calls into generic code.
 384  */
 385 static noinline int btrfs_copy_from_user(loff_t pos, size_t write_bytes,
 386                                          struct page **prepared_pages,
 387                                          struct iov_iter *i)
 388 {
 389         size_t copied = 0;
 390         size_t total_copied = 0;
 391         int pg = 0;
 392         int offset = offset_in_page(pos);
 393
 394         while (write_bytes > 0) {
 395                 size_t count = min_t(size_t,
 396                                      PAGE_SIZE - offset, write_bytes);
 397                 struct page *page = prepared_pages[pg];
 398                 /*
 399                  * Copy data from userspace to the current page
 400                  */
 401                 copied = iov_iter_copy_from_user_atomic(page, i, offset, count);
 402
 403                 /* Flush processor's dcache for this page */
 404                 flush_dcache_page(page);
 405
 406                 /*
 407                  * if we get a partial write, we can end up with
 408                  * partially up to date pages.  These add
 409                  * a lot of complexity, so make sure they don't
 410                  * happen by forcing this copy to be retried.
 411                  *
 412                  * The rest of the btrfs_file_write code will fall
 413                  * back to page at a time copies after we return 0.
 414                  */
 415                 if (!PageUptodate(page) && copied < count)
 416                         copied = 0;
 417
 418                 iov_iter_advance(i, copied);
 419                 write_bytes -= copied;
 420                 total_copied += copied;
 421
 422                 /* Return to btrfs_file_write_iter to fault page */
 423                 if (unlikely(copied == 0))
 424                         break;
 425
 426                 if (copied < PAGE_SIZE - offset) {
 427                         offset += copied;
 428                 } else {
 429                         pg++;
 430                         offset = 0;
 431                 }
 432         }
 433         return total_copied;
 434 }
 435
 436 /*
 437  * unlocks pages after btrfs_file_write is done with them
 438  */
 439 static void btrfs_drop_pages(struct page **pages, size_t num_pages)
 440 {
 441         size_t i;
 442         for (i = 0; i < num_pages; i++) {
 443                 /* page checked is some magic around finding pages that
 444                  * have been modified without going through btrfs_set_page_dirty
 445                  * clear it here. There should be no need to mark the pages
 446                  * accessed as prepare_pages should have marked them accessed
 447                  * in prepare_pages via find_or_create_page()
 448                  */
 449                 ClearPageChecked(pages[i]);
 450                 unlock_page(pages[i]);
 451                 put_page(pages[i]);
 452         }
 453 }
 454
 455 /*
 456  * after copy_from_user, pages need to be dirtied and we need to make
 457  * sure holes are created between the current EOF and the start of
 458  * any next extents (if required).
 459  *
 460  * this also makes the decision about creating an inline extent vs
 461  * doing real data extents, marking pages dirty and delalloc as required.
 462  */
 463 int btrfs_dirty_pages(struct btrfs_inode *inode, struct page **pages,
 464                       size_t num_pages, loff_t pos, size_t write_bytes,
 465                       struct extent_state **cached, bool noreserve)
 466 {
 467         struct btrfs_fs_info *fs_info = inode->root->fs_info;
 468         int err = 0;
 469         int i;
 470         u64 num_bytes;
 471         u64 start_pos;
 472         u64 end_of_last_block;
 473         u64 end_pos = pos + write_bytes;
 474         loff_t isize = i_size_read(&inode->vfs_inode);
 475         unsigned int extra_bits = 0;
 476
 477         if (write_bytes == 0)
 478                 return 0;
 479
 480         if (noreserve)
 481                 extra_bits |= EXTENT_NORESERVE;
 482
 483         start_pos = round_down(pos, fs_info->sectorsize);
 484         num_bytes = round_up(write_bytes + pos - start_pos,
 485                              fs_info->sectorsize);
 486
 487         end_of_last_block = start_pos + num_bytes - 1;
 488
 489         /*
 490          * The pages may have already been dirty, clear out old accounting so
 491          * we can set things up properly
 492          */
 493         clear_extent_bit(&inode->io_tree, start_pos, end_of_last_block,
 494                          EXTENT_DELALLOC | EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG,
 495                          0, 0, cached);
 496
 497         err = btrfs_set_extent_delalloc(inode, start_pos, end_of_last_block,
 498                                         extra_bits, cached);
 499         if (err)
 500                 return err;
 501
 502         for (i = 0; i < num_pages; i++) {
 503                 struct page *p = pages[i];
 504                 SetPageUptodate(p);
 505                 ClearPageChecked(p);
 506                 set_page_dirty(p);
 507         }
 508
 509         /*
 510          * we've only changed i_size in ram, and we haven't updated
 511          * the disk i_size.  There is no need to log the inode
 512          * at this time.
 513          */
 514         if (end_pos > isize)
 515                 i_size_write(&inode->vfs_inode, end_pos);
 516         return 0;
 517 }
 518
 519 /*
 520  * this drops all the extents in the cache that intersect the range
 521  * [start, end].  Existing extents are split as required.
 522  */
 523 void btrfs_drop_extent_cache(struct btrfs_inode *inode, u64 start, u64 end,
 524                              int skip_pinned)
 525 {
 526         struct extent_map *em;
 527         struct extent_map *split = NULL;
 528         struct extent_map *split2 = NULL;
 529         struct extent_map_tree *em_tree = &inode->extent_tree;
 530         u64 len = end - start + 1;
 531         u64 gen;
 532         int ret;
 533         int testend = 1;
 534         unsigned long flags;
 535         int compressed = 0;
 536         bool modified;
 537
 538         WARN_ON(end < start);
 539         if (end == (u64)-1) {
 540                 len = (u64)-1;
 541                 testend = 0;
 542         }
 543         while (1) {
 544                 int no_splits = 0;
 545
 546                 modified = false;
 547                 if (!split)
 548                         split = alloc_extent_map();
 549                 if (!split2)
 550                         split2 = alloc_extent_map();
 551                 if (!split || !split2)
 552                         no_splits = 1;
 553
 554                 write_lock(&em_tree->lock);
 555                 em = lookup_extent_mapping(em_tree, start, len);
 556                 if (!em) {
 557                         write_unlock(&em_tree->lock);
 558                         break;
 559                 }
 560                 flags = em->flags;
 561                 gen = em->generation;
 562                 if (skip_pinned && test_bit(EXTENT_FLAG_PINNED, &em->flags)) {
 563                         if (testend && em->start + em->len >= start + len) {
 564                                 free_extent_map(em);
 565                                 write_unlock(&em_tree->lock);
 566                                 break;
 567                         }
 568                         start = em->start + em->len;
 569                         if (testend)
 570                                 len = start + len - (em->start + em->len);
 571                         free_extent_map(em);
 572                         write_unlock(&em_tree->lock);
 573                         continue;
 574                 }
 575                 compressed = test_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
 576                 clear_bit(EXTENT_FLAG_PINNED, &em->flags);
 577                 clear_bit(EXTENT_FLAG_LOGGING, &flags);
 578                 modified = !list_empty(&em->list);
 579                 if (no_splits)
 580                         goto next;
 581
 582                 if (em->start < start) {
 583                         split->start = em->start;
 584                         split->len = start - em->start;
 585
 586                         if (em->block_start < EXTENT_MAP_LAST_BYTE) {
 587                                 split->orig_start = em->orig_start;
 588                                 split->block_start = em->block_start;
 589
 590                                 if (compressed)
 591                                         split->block_len = em->block_len;
 592                                 else
 593                                         split->block_len = split->len;
 594                                 split->orig_block_len = max(split->block_len,
 595                                                 em->orig_block_len);
 596                                 split->ram_bytes = em->ram_bytes;
 597                         } else {
 598                                 split->orig_start = split->start;
 599                                 split->block_len = 0;
 600                                 split->block_start = em->block_start;
 601                                 split->orig_block_len = 0;
 602                                 split->ram_bytes = split->len;
 603                         }
 604
 605                         split->generation = gen;
 606                         split->flags = flags;
 607                         split->compress_type = em->compress_type;
 608                         replace_extent_mapping(em_tree, em, split, modified);
 609                         free_extent_map(split);
 610                         split = split2;
 611                         split2 = NULL;
 612                 }
 613                 if (testend && em->start + em->len > start + len) {
 614                         u64 diff = start + len - em->start;
 615
 616                         split->start = start + len;
 617                         split->len = em->start + em->len - (start + len);
 618                         split->flags = flags;
 619                         split->compress_type = em->compress_type;
 620                         split->generation = gen;
 621
 622                         if (em->block_start < EXTENT_MAP_LAST_BYTE) {
 623                                 split->orig_block_len = max(em->block_len,
 624                                                     em->orig_block_len);
 625
 626                                 split->ram_bytes = em->ram_bytes;
 627                                 if (compressed) {
 628                                         split->block_len = em->block_len;
 629                                         split->block_start = em->block_start;
 630                                         split->orig_start = em->orig_start;
 631                                 } else {
 632                                         split->block_len = split->len;
 633                                         split->block_start = em->block_start
 634                                                 + diff;
 635                                         split->orig_start = em->orig_start;
 636                                 }
 637                         } else {
 638                                 split->ram_bytes = split->len;
 639                                 split->orig_start = split->start;
 640                                 split->block_len = 0;
 641                                 split->block_start = em->block_start;
 642                                 split->orig_block_len = 0;
 643                         }
 644
 645                         if (extent_map_in_tree(em)) {
 646                                 replace_extent_mapping(em_tree, em, split,
 647                                                        modified);
 648                         } else {
 649                                 ret = add_extent_mapping(em_tree, split,
 650                                                          modified);
 651                                 ASSERT(ret == 0); /* Logic error */
 652                         }
 653                         free_extent_map(split);
 654                         split = NULL;
 655                 }
 656 next:
 657                 if (extent_map_in_tree(em))
 658                         remove_extent_mapping(em_tree, em);
 659                 write_unlock(&em_tree->lock);
 660
 661                 /* once for us */
 662                 free_extent_map(em);
 663                 /* once for the tree*/
 664                 free_extent_map(em);
 665         }
 666         if (split)
 667                 free_extent_map(split);
 668         if (split2)
 669                 free_extent_map(split2);
 670 }
 671
 672 /*
 673  * this is very complex, but the basic idea is to drop all extents
 674  * in the range start - end.  hint_block is filled in with a block number
 675  * that would be a good hint to the block allocator for this file.
 676  *
 677  * If an extent intersects the range but is not entirely inside the range
 678  * it is either truncated or split.  Anything entirely inside the range
 679  * is deleted from the tree.
 680  */
 681 int __btrfs_drop_extents(struct btrfs_trans_handle *trans,
 682                          struct btrfs_root *root, struct btrfs_inode *inode,
 683                          struct btrfs_path *path, u64 start, u64 end,
 684                          u64 *drop_end, int drop_cache,
 685                          int replace_extent,
 686                          u32 extent_item_size,
 687                          int *key_inserted)
 688 {
 689         struct btrfs_fs_info *fs_info = root->fs_info;
 690         struct extent_buffer *leaf;
 691         struct btrfs_file_extent_item *fi;
 692         struct btrfs_ref ref = { 0 };
 693         struct btrfs_key key;
 694         struct btrfs_key new_key;
 695         struct inode *vfs_inode = &inode->vfs_inode;
 696         u64 ino = btrfs_ino(inode);
 697         u64 search_start = start;
 698         u64 disk_bytenr = 0;
 699         u64 num_bytes = 0;
 700         u64 extent_offset = 0;
 701         u64 extent_end = 0;
 702         u64 last_end = start;
 703         int del_nr = 0;
 704         int del_slot = 0;
 705         int extent_type;
 706         int recow;
 707         int ret;
 708         int modify_tree = -1;
 709         int update_refs;
 710         int found = 0;
 711         int leafs_visited = 0;
 712
 713         if (drop_cache)
 714                 btrfs_drop_extent_cache(inode, start, end - 1, 0);
 715
 716         if (start >= inode->disk_i_size && !replace_extent)
 717                 modify_tree = 0;
 718
 719         update_refs = (test_bit(BTRFS_ROOT_SHAREABLE, &root->state) ||
 720                        root == fs_info->tree_root);
 721         while (1) {
 722                 recow = 0;
 723                 ret = btrfs_lookup_file_extent(trans, root, path, ino,
 724                                                search_start, modify_tree);
 725                 if (ret < 0)
 726                         break;
 727                 if (ret > 0 && path->slots[0] > 0 && search_start == start) {
 728                         leaf = path->nodes[0];
 729                         btrfs_item_key_to_cpu(leaf, &key, path->slots[0] - 1);
 730                         if (key.objectid == ino &&
 731                             key.type == BTRFS_EXTENT_DATA_KEY)
 732                                 path->slots[0]--;
 733                 }
 734                 ret = 0;
 735                 leafs_visited++;
 736 next_slot:
 737                 leaf = path->nodes[0];
 738                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
 739                         BUG_ON(del_nr > 0);
 740                         ret = btrfs_next_leaf(root, path);
 741                         if (ret < 0)
 742                                 break;
 743                         if (ret > 0) {
 744                                 ret = 0;
 745                                 break;
 746                         }
 747                         leafs_visited++;
 748                         leaf = path->nodes[0];
 749                         recow = 1;
 750                 }
 751
 752                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 753
 754                 if (key.objectid > ino)
 755                         break;
 756                 if (WARN_ON_ONCE(key.objectid < ino) ||
 757                     key.type < BTRFS_EXTENT_DATA_KEY) {
 758                         ASSERT(del_nr == 0);
 759                         path->slots[0]++;
 760                         goto next_slot;
 761                 }
 762                 if (key.type > BTRFS_EXTENT_DATA_KEY || key.offset >= end)
 763                         break;
 764
 765                 fi = btrfs_item_ptr(leaf, path->slots[0],
 766                                     struct btrfs_file_extent_item);
 767                 extent_type = btrfs_file_extent_type(leaf, fi);
 768
 769                 if (extent_type == BTRFS_FILE_EXTENT_REG ||
 770                     extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
 771                         disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
 772                         num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
 773                         extent_offset = btrfs_file_extent_offset(leaf, fi);
 774                         extent_end = key.offset +
 775                                 btrfs_file_extent_num_bytes(leaf, fi);
 776                 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
 777                         extent_end = key.offset +
 778                                 btrfs_file_extent_ram_bytes(leaf, fi);
 779                 } else {
 780                         /* can't happen */
 781                         BUG();
 782                 }
 783
 784                 /*
 785                  * Don't skip extent items representing 0 byte lengths. They
 786                  * used to be created (bug) if while punching holes we hit
 787                  * -ENOSPC condition. So if we find one here, just ensure we
 788                  * delete it, otherwise we would insert a new file extent item
 789                  * with the same key (offset) as that 0 bytes length file
 790                  * extent item in the call to setup_items_for_insert() later
 791                  * in this function.
 792                  */
 793                 if (extent_end == key.offset && extent_end >= search_start) {
 794                         last_end = extent_end;
 795                         goto delete_extent_item;
 796                 }
 797
 798                 if (extent_end <= search_start) {
 799                         path->slots[0]++;
 800                         goto next_slot;
 801                 }
 802
 803                 found = 1;
 804                 search_start = max(key.offset, start);
 805                 if (recow || !modify_tree) {
 806                         modify_tree = -1;
 807                         btrfs_release_path(path);
 808                         continue;
 809                 }
 810
 811                 /*
 812                  *     | - range to drop - |
 813                  *  | -------- extent -------- |
 814                  */
 815                 if (start > key.offset && end < extent_end) {
 816                         BUG_ON(del_nr > 0);
 817                         if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
 818                                 ret = -EOPNOTSUPP;
 819                                 break;
 820                         }
 821
 822                         memcpy(&new_key, &key, sizeof(new_key));
 823                         new_key.offset = start;
 824                         ret = btrfs_duplicate_item(trans, root, path,
 825                                                    &new_key);
 826                         if (ret == -EAGAIN) {
 827                                 btrfs_release_path(path);
 828                                 continue;
 829                         }
 830                         if (ret < 0)
 831                                 break;
 832
 833                         leaf = path->nodes[0];
 834                         fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
 835                                             struct btrfs_file_extent_item);
 836                         btrfs_set_file_extent_num_bytes(leaf, fi,
 837                                                         start - key.offset);
 838
 839                         fi = btrfs_item_ptr(leaf, path->slots[0],
 840                                             struct btrfs_file_extent_item);
 841
 842                         extent_offset += start - key.offset;
 843                         btrfs_set_file_extent_offset(leaf, fi, extent_offset);
 844                         btrfs_set_file_extent_num_bytes(leaf, fi,
 845                                                         extent_end - start);
 846                         btrfs_mark_buffer_dirty(leaf);
 847
 848                         if (update_refs && disk_bytenr > 0) {
 849                                 btrfs_init_generic_ref(&ref,
 850                                                 BTRFS_ADD_DELAYED_REF,
 851                                                 disk_bytenr, num_bytes, 0);
 852                                 btrfs_init_data_ref(&ref,
 853                                                 root->root_key.objectid,
 854                                                 new_key.objectid,
 855                                                 start - extent_offset);
 856                                 ret = btrfs_inc_extent_ref(trans, &ref);
 857                                 BUG_ON(ret); /* -ENOMEM */
 858                         }
 859                         key.offset = start;
 860                 }
 861                 /*
 862                  * From here on out we will have actually dropped something, so
 863                  * last_end can be updated.
 864                  */
 865                 last_end = extent_end;
 866
 867                 /*
 868                  *  | ---- range to drop ----- |
 869                  *      | -------- extent -------- |
 870                  */
 871                 if (start <= key.offset && end < extent_end) {
 872                         if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
 873                                 ret = -EOPNOTSUPP;
 874                                 break;
 875                         }
 876
 877                         memcpy(&new_key, &key, sizeof(new_key));
 878                         new_key.offset = end;
 879                         btrfs_set_item_key_safe(fs_info, path, &new_key);
 880
 881                         extent_offset += end - key.offset;
 882                         btrfs_set_file_extent_offset(leaf, fi, extent_offset);
 883                         btrfs_set_file_extent_num_bytes(leaf, fi,
 884                                                         extent_end - end);
 885                         btrfs_mark_buffer_dirty(leaf);
 886                         if (update_refs && disk_bytenr > 0)
 887                                 inode_sub_bytes(vfs_inode, end - key.offset);
 888                         break;
 889                 }
 890
 891                 search_start = extent_end;
 892                 /*
 893                  *       | ---- range to drop ----- |
 894                  *  | -------- extent -------- |
 895                  */
 896                 if (start > key.offset && end >= extent_end) {
 897                         BUG_ON(del_nr > 0);
 898                         if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
 899                                 ret = -EOPNOTSUPP;
 900                                 break;
 901                         }
 902
 903                         btrfs_set_file_extent_num_bytes(leaf, fi,
 904                                                         start - key.offset);
 905                         btrfs_mark_buffer_dirty(leaf);
 906                         if (update_refs && disk_bytenr > 0)
 907                                 inode_sub_bytes(vfs_inode, extent_end - start);
 908                         if (end == extent_end)
 909                                 break;
 910
 911                         path->slots[0]++;
 912                         goto next_slot;
 913                 }
 914
 915                 /*
 916                  *  | ---- range to drop ----- |
 917                  *    | ------ extent ------ |
 918                  */
 919                 if (start <= key.offset && end >= extent_end) {
 920 delete_extent_item:
 921                         if (del_nr == 0) {
 922                                 del_slot = path->slots[0];
 923                                 del_nr = 1;
 924                         } else {
 925                                 BUG_ON(del_slot + del_nr != path->slots[0]);
 926                                 del_nr++;
 927                         }
 928
 929                         if (update_refs &&
 930                             extent_type == BTRFS_FILE_EXTENT_INLINE) {
 931                                 inode_sub_bytes(vfs_inode,
 932                                                 extent_end - key.offset);
 933                                 extent_end = ALIGN(extent_end,
 934                                                    fs_info->sectorsize);
 935                         } else if (update_refs && disk_bytenr > 0) {
 936                                 btrfs_init_generic_ref(&ref,
 937                                                 BTRFS_DROP_DELAYED_REF,
 938                                                 disk_bytenr, num_bytes, 0);
 939                                 btrfs_init_data_ref(&ref,
 940                                                 root->root_key.objectid,
 941                                                 key.objectid,
 942                                                 key.offset - extent_offset);
 943                                 ret = btrfs_free_extent(trans, &ref);
 944                                 BUG_ON(ret); /* -ENOMEM */
 945                                 inode_sub_bytes(vfs_inode,
 946                                                 extent_end - key.offset);
 947                         }
 948
 949                         if (end == extent_end)
 950                                 break;
 951
 952                         if (path->slots[0] + 1 < btrfs_header_nritems(leaf)) {
 953                                 path->slots[0]++;
 954                                 goto next_slot;
 955                         }
 956
 957                         ret = btrfs_del_items(trans, root, path, del_slot,
 958                                               del_nr);
 959                         if (ret) {
 960                                 btrfs_abort_transaction(trans, ret);
 961                                 break;
 962                         }
 963
 964                         del_nr = 0;
 965                         del_slot = 0;
 966
 967                         btrfs_release_path(path);
 968                         continue;
 969                 }
 970
 971                 BUG();
 972         }
 973
 974         if (!ret && del_nr > 0) {
 975                 /*
 976                  * Set path->slots[0] to first slot, so that after the delete
 977                  * if items are move off from our leaf to its immediate left or
 978                  * right neighbor leafs, we end up with a correct and adjusted
 979                  * path->slots[0] for our insertion (if replace_extent != 0).
 980                  */
 981                 path->slots[0] = del_slot;
 982                 ret = btrfs_del_items(trans, root, path, del_slot, del_nr);
 983                 if (ret)
 984                         btrfs_abort_transaction(trans, ret);
 985         }
 986
 987         leaf = path->nodes[0];
 988         /*
 989          * If btrfs_del_items() was called, it might have deleted a leaf, in
 990          * which case it unlocked our path, so check path->locks[0] matches a
 991          * write lock.
 992          */
 993         if (!ret && replace_extent && leafs_visited == 1 &&
 994             path->locks[0] == BTRFS_WRITE_LOCK &&
 995             btrfs_leaf_free_space(leaf) >=
 996             sizeof(struct btrfs_item) + extent_item_size) {
 997
 998                 key.objectid = ino;
 999                 key.type = BTRFS_EXTENT_DATA_KEY;
1000                 key.offset = start;
1001                 if (!del_nr && path->slots[0] < btrfs_header_nritems(leaf)) {
1002                         struct btrfs_key slot_key;
1003
1004                         btrfs_item_key_to_cpu(leaf, &slot_key, path->slots[0]);
1005                         if (btrfs_comp_cpu_keys(&key, &slot_key) > 0)
1006                                 path->slots[0]++;
1007                 }
1008                 setup_items_for_insert(root, path, &key, &extent_item_size, 1);
1009                 *key_inserted = 1;
1010         }
1011
1012         if (!replace_extent || !(*key_inserted))
1013                 btrfs_release_path(path);
1014         if (drop_end)
1015                 *drop_end = found ? min(end, last_end) : end;
1016         return ret;
1017 }
1018
1019 int btrfs_drop_extents(struct btrfs_trans_handle *trans,
1020                        struct btrfs_root *root, struct inode *inode, u64 start,
1021                        u64 end, int drop_cache)
1022 {
1023         struct btrfs_path *path;
1024         int ret;
1025
1026         path = btrfs_alloc_path();
1027         if (!path)
1028                 return -ENOMEM;
1029         ret = __btrfs_drop_extents(trans, root, BTRFS_I(inode), path, start,
1030                                    end, NULL, drop_cache, 0, 0, NULL);
1031         btrfs_free_path(path);
1032         return ret;
1033 }
1034
1035 static int extent_mergeable(struct extent_buffer *leaf, int slot,
1036                             u64 objectid, u64 bytenr, u64 orig_offset,
1037                             u64 *start, u64 *end)
1038 {
1039         struct btrfs_file_extent_item *fi;
1040         struct btrfs_key key;
1041         u64 extent_end;
1042
1043         if (slot < 0 || slot >= btrfs_header_nritems(leaf))
1044                 return 0;
1045
1046         btrfs_item_key_to_cpu(leaf, &key, slot);
1047         if (key.objectid != objectid || key.type != BTRFS_EXTENT_DATA_KEY)
1048                 return 0;
1049
1050         fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
1051         if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG ||
1052             btrfs_file_extent_disk_bytenr(leaf, fi) != bytenr ||
1053             btrfs_file_extent_offset(leaf, fi) != key.offset - orig_offset ||
1054             btrfs_file_extent_compression(leaf, fi) ||
1055             btrfs_file_extent_encryption(leaf, fi) ||
1056             btrfs_file_extent_other_encoding(leaf, fi))
1057                 return 0;
1058
1059         extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
1060         if ((*start && *start != key.offset) || (*end && *end != extent_end))
1061                 return 0;
1062
1063         *start = key.offset;
1064         *end = extent_end;
1065         return 1;
1066 }
1067
1068 /*
1069  * Mark extent in the range start - end as written.
1070  *
1071  * This changes extent type from 'pre-allocated' to 'regular'. If only
1072  * part of extent is marked as written, the extent will be split into
1073  * two or three.
1074  */
1075 int btrfs_mark_extent_written(struct btrfs_trans_handle *trans,
1076                               struct btrfs_inode *inode, u64 start, u64 end)
1077 {
1078         struct btrfs_fs_info *fs_info = trans->fs_info;
1079         struct btrfs_root *root = inode->root;
1080         struct extent_buffer *leaf;
1081         struct btrfs_path *path;
1082         struct btrfs_file_extent_item *fi;
1083         struct btrfs_ref ref = { 0 };
1084         struct btrfs_key key;
1085         struct btrfs_key new_key;
1086         u64 bytenr;
1087         u64 num_bytes;
1088         u64 extent_end;
1089         u64 orig_offset;
1090         u64 other_start;
1091         u64 other_end;
1092         u64 split;
1093         int del_nr = 0;
1094         int del_slot = 0;
1095         int recow;
1096         int ret;
1097         u64 ino = btrfs_ino(inode);
1098
1099         path = btrfs_alloc_path();
1100         if (!path)
1101                 return -ENOMEM;
1102 again:
1103         recow = 0;
1104         split = start;
1105         key.objectid = ino;
1106         key.type = BTRFS_EXTENT_DATA_KEY;
1107         key.offset = split;
1108
1109         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1110         if (ret < 0)
1111                 goto out;
1112         if (ret > 0 && path->slots[0] > 0)
1113                 path->slots[0]--;
1114
1115         leaf = path->nodes[0];
1116         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1117         if (key.objectid != ino ||
1118             key.type != BTRFS_EXTENT_DATA_KEY) {
1119                 ret = -EINVAL;
1120                 btrfs_abort_transaction(trans, ret);
1121                 goto out;
1122         }
1123         fi = btrfs_item_ptr(leaf, path->slots[0],
1124                             struct btrfs_file_extent_item);
1125         if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_PREALLOC) {
1126                 ret = -EINVAL;
1127                 btrfs_abort_transaction(trans, ret);
1128                 goto out;
1129         }
1130         extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
1131         if (key.offset > start || extent_end < end) {
1132                 ret = -EINVAL;
1133                 btrfs_abort_transaction(trans, ret);
1134                 goto out;
1135         }
1136
1137         bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1138         num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
1139         orig_offset = key.offset - btrfs_file_extent_offset(leaf, fi);
1140         memcpy(&new_key, &key, sizeof(new_key));
1141
1142         if (start == key.offset && end < extent_end) {
1143                 other_start = 0;
1144                 other_end = start;
1145                 if (extent_mergeable(leaf, path->slots[0] - 1,
1146                                      ino, bytenr, orig_offset,
1147                                      &other_start, &other_end)) {
1148                         new_key.offset = end;
1149                         btrfs_set_item_key_safe(fs_info, path, &new_key);
1150                         fi = btrfs_item_ptr(leaf, path->slots[0],
1151                                             struct btrfs_file_extent_item);
1152                         btrfs_set_file_extent_generation(leaf, fi,
1153                                                          trans->transid);
1154                         btrfs_set_file_extent_num_bytes(leaf, fi,
1155                                                         extent_end - end);
1156                         btrfs_set_file_extent_offset(leaf, fi,
1157                                                      end - orig_offset);
1158                         fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
1159                                             struct btrfs_file_extent_item);
1160                         btrfs_set_file_extent_generation(leaf, fi,
1161                                                          trans->transid);
1162                         btrfs_set_file_extent_num_bytes(leaf, fi,
1163                                                         end - other_start);
1164                         btrfs_mark_buffer_dirty(leaf);
1165                         goto out;
1166                 }
1167         }
1168
1169         if (start > key.offset && end == extent_end) {
1170                 other_start = end;
1171                 other_end = 0;
1172                 if (extent_mergeable(leaf, path->slots[0] + 1,
1173                                      ino, bytenr, orig_offset,
1174                                      &other_start, &other_end)) {
1175                         fi = btrfs_item_ptr(leaf, path->slots[0],
1176                                             struct btrfs_file_extent_item);
1177                         btrfs_set_file_extent_num_bytes(leaf, fi,
1178                                                         start - key.offset);
1179                         btrfs_set_file_extent_generation(leaf, fi,
1180                                                          trans->transid);
1181                         path->slots[0]++;
1182                         new_key.offset = start;
1183                         btrfs_set_item_key_safe(fs_info, path, &new_key);
1184
1185                         fi = btrfs_item_ptr(leaf, path->slots[0],
1186                                             struct btrfs_file_extent_item);
1187                         btrfs_set_file_extent_generation(leaf, fi,
1188                                                          trans->transid);
1189                         btrfs_set_file_extent_num_bytes(leaf, fi,
1190                                                         other_end - start);
1191                         btrfs_set_file_extent_offset(leaf, fi,
1192                                                      start - orig_offset);
1193                         btrfs_mark_buffer_dirty(leaf);
1194                         goto out;
1195                 }
1196         }
1197
1198         while (start > key.offset || end < extent_end) {
1199                 if (key.offset == start)
1200                         split = end;
1201
1202                 new_key.offset = split;
1203                 ret = btrfs_duplicate_item(trans, root, path, &new_key);
1204                 if (ret == -EAGAIN) {
1205                         btrfs_release_path(path);
1206                         goto again;
1207                 }
1208                 if (ret < 0) {
1209                         btrfs_abort_transaction(trans, ret);
1210                         goto out;
1211                 }
1212
1213                 leaf = path->nodes[0];
1214                 fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
1215                                     struct btrfs_file_extent_item);
1216                 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1217                 btrfs_set_file_extent_num_bytes(leaf, fi,
1218                                                 split - key.offset);
1219
1220                 fi = btrfs_item_ptr(leaf, path->slots[0],
1221                                     struct btrfs_file_extent_item);
1222
1223                 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1224                 btrfs_set_file_extent_offset(leaf, fi, split - orig_offset);
1225                 btrfs_set_file_extent_num_bytes(leaf, fi,
1226                                                 extent_end - split);
1227                 btrfs_mark_buffer_dirty(leaf);
1228
1229                 btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF, bytenr,
1230                                        num_bytes, 0);
1231                 btrfs_init_data_ref(&ref, root->root_key.objectid, ino,
1232                                     orig_offset);
1233                 ret = btrfs_inc_extent_ref(trans, &ref);
1234                 if (ret) {
1235                         btrfs_abort_transaction(trans, ret);
1236                         goto out;
1237                 }
1238
1239                 if (split == start) {
1240                         key.offset = start;
1241                 } else {
1242                         if (start != key.offset) {
1243                                 ret = -EINVAL;
1244                                 btrfs_abort_transaction(trans, ret);
1245                                 goto out;
1246                         }
1247                         path->slots[0]--;
1248                         extent_end = end;
1249                 }
1250                 recow = 1;
1251         }
1252
1253         other_start = end;
1254         other_end = 0;
1255         btrfs_init_generic_ref(&ref, BTRFS_DROP_DELAYED_REF, bytenr,
1256                                num_bytes, 0);
1257         btrfs_init_data_ref(&ref, root->root_key.objectid, ino, orig_offset);
1258         if (extent_mergeable(leaf, path->slots[0] + 1,
1259                              ino, bytenr, orig_offset,
1260                              &other_start, &other_end)) {
1261                 if (recow) {
1262                         btrfs_release_path(path);
1263                         goto again;
1264                 }
1265                 extent_end = other_end;
1266                 del_slot = path->slots[0] + 1;
1267                 del_nr++;
1268                 ret = btrfs_free_extent(trans, &ref);
1269                 if (ret) {
1270                         btrfs_abort_transaction(trans, ret);
1271                         goto out;
1272                 }
1273         }
1274         other_start = 0;
1275         other_end = start;
1276         if (extent_mergeable(leaf, path->slots[0] - 1,
1277                              ino, bytenr, orig_offset,
1278                              &other_start, &other_end)) {
1279                 if (recow) {
1280                         btrfs_release_path(path);
1281                         goto again;
1282                 }
1283                 key.offset = other_start;
1284                 del_slot = path->slots[0];
1285                 del_nr++;
1286                 ret = btrfs_free_extent(trans, &ref);
1287                 if (ret) {
1288                         btrfs_abort_transaction(trans, ret);
1289                         goto out;
1290                 }
1291         }
1292         if (del_nr == 0) {
1293                 fi = btrfs_item_ptr(leaf, path->slots[0],
1294                            struct btrfs_file_extent_item);
1295                 btrfs_set_file_extent_type(leaf, fi,
1296                                            BTRFS_FILE_EXTENT_REG);
1297                 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1298                 btrfs_mark_buffer_dirty(leaf);
1299         } else {
1300                 fi = btrfs_item_ptr(leaf, del_slot - 1,
1301                            struct btrfs_file_extent_item);
1302                 btrfs_set_file_extent_type(leaf, fi,
1303                                            BTRFS_FILE_EXTENT_REG);
1304                 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1305                 btrfs_set_file_extent_num_bytes(leaf, fi,
1306                                                 extent_end - key.offset);
1307                 btrfs_mark_buffer_dirty(leaf);
1308
1309                 ret = btrfs_del_items(trans, root, path, del_slot, del_nr);
1310                 if (ret < 0) {
1311                         btrfs_abort_transaction(trans, ret);
1312                         goto out;
1313                 }
1314         }
1315 out:
1316         btrfs_free_path(path);
1317         return 0;
1318 }
1319
1320 /*
1321  * on error we return an unlocked page and the error value
1322  * on success we return a locked page and 0
1323  */
1324 static int prepare_uptodate_page(struct inode *inode,
1325                                  struct page *page, u64 pos,
1326                                  bool force_uptodate)
1327 {
1328         int ret = 0;
1329
1330         if (((pos & (PAGE_SIZE - 1)) || force_uptodate) &&
1331             !PageUptodate(page)) {
1332                 ret = btrfs_readpage(NULL, page);
1333                 if (ret)
1334                         return ret;
1335                 lock_page(page);
1336                 if (!PageUptodate(page)) {
1337                         unlock_page(page);
1338                         return -EIO;
1339                 }
1340                 if (page->mapping != inode->i_mapping) {
1341                         unlock_page(page);
1342                         return -EAGAIN;
1343                 }
1344         }
1345         return 0;
1346 }
1347
1348 /*
1349  * this just gets pages into the page cache and locks them down.
1350  */
1351 static noinline int prepare_pages(struct inode *inode, struct page **pages,
1352                                   size_t num_pages, loff_t pos,
1353                                   size_t write_bytes, bool force_uptodate)
1354 {
1355         int i;
1356         unsigned long index = pos >> PAGE_SHIFT;
1357         gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping);
1358         int err = 0;
1359         int faili;
1360
1361         for (i = 0; i < num_pages; i++) {
1362 again:
1363                 pages[i] = find_or_create_page(inode->i_mapping, index + i,
1364                                                mask | __GFP_WRITE);
1365                 if (!pages[i]) {
1366                         faili = i - 1;
1367                         err = -ENOMEM;
1368                         goto fail;
1369                 }
1370
1371                 if (i == 0)
1372                         err = prepare_uptodate_page(inode, pages[i], pos,
1373                                                     force_uptodate);
1374                 if (!err && i == num_pages - 1)
1375                         err = prepare_uptodate_page(inode, pages[i],
1376                                                     pos + write_bytes, false);
1377                 if (err) {
1378                         put_page(pages[i]);
1379                         if (err == -EAGAIN) {
1380                                 err = 0;
1381                                 goto again;
1382                         }
1383                         faili = i - 1;
1384                         goto fail;
1385                 }
1386                 wait_on_page_writeback(pages[i]);
1387         }
1388
1389         return 0;
1390 fail:
1391         while (faili >= 0) {
1392                 unlock_page(pages[faili]);
1393                 put_page(pages[faili]);
1394                 faili--;
1395         }
1396         return err;
1397
1398 }
1399
1400 /*
1401  * This function locks the extent and properly waits for data=ordered extents
1402  * to finish before allowing the pages to be modified if need.
1403  *
1404  * The return value:
1405  * 1 - the extent is locked
1406  * 0 - the extent is not locked, and everything is OK
1407  * -EAGAIN - need re-prepare the pages
1408  * the other < 0 number - Something wrong happens
1409  */
1410 static noinline int
1411 lock_and_cleanup_extent_if_need(struct btrfs_inode *inode, struct page **pages,
1412                                 size_t num_pages, loff_t pos,
1413                                 size_t write_bytes,
1414                                 u64 *lockstart, u64 *lockend,
1415                                 struct extent_state **cached_state)
1416 {
1417         struct btrfs_fs_info *fs_info = inode->root->fs_info;
1418         u64 start_pos;
1419         u64 last_pos;
1420         int i;
1421         int ret = 0;
1422
1423         start_pos = round_down(pos, fs_info->sectorsize);
1424         last_pos = round_up(pos + write_bytes, fs_info->sectorsize) - 1;
1425
1426         if (start_pos < inode->vfs_inode.i_size) {
1427                 struct btrfs_ordered_extent *ordered;
1428
1429                 lock_extent_bits(&inode->io_tree, start_pos, last_pos,
1430                                 cached_state);
1431                 ordered = btrfs_lookup_ordered_range(inode, start_pos,
1432                                                      last_pos - start_pos + 1);
1433                 if (ordered &&
1434                     ordered->file_offset + ordered->num_bytes > start_pos &&
1435                     ordered->file_offset <= last_pos) {
1436                         unlock_extent_cached(&inode->io_tree, start_pos,
1437                                         last_pos, cached_state);
1438                         for (i = 0; i < num_pages; i++) {
1439                                 unlock_page(pages[i]);
1440                                 put_page(pages[i]);
1441                         }
1442                         btrfs_start_ordered_extent(ordered, 1);
1443                         btrfs_put_ordered_extent(ordered);
1444                         return -EAGAIN;
1445                 }
1446                 if (ordered)
1447                         btrfs_put_ordered_extent(ordered);
1448
1449                 *lockstart = start_pos;
1450                 *lockend = last_pos;
1451                 ret = 1;
1452         }
1453
1454         /*
1455          * It's possible the pages are dirty right now, but we don't want
1456          * to clean them yet because copy_from_user may catch a page fault
1457          * and we might have to fall back to one page at a time.  If that
1458          * happens, we'll unlock these pages and we'd have a window where
1459          * reclaim could sneak in and drop the once-dirty page on the floor
1460          * without writing it.
1461          *
1462          * We have the pages locked and the extent range locked, so there's
1463          * no way someone can start IO on any dirty pages in this range.
1464          *
1465          * We'll call btrfs_dirty_pages() later on, and that will flip around
1466          * delalloc bits and dirty the pages as required.
1467          */
1468         for (i = 0; i < num_pages; i++) {
1469                 set_page_extent_mapped(pages[i]);
1470                 WARN_ON(!PageLocked(pages[i]));
1471         }
1472
1473         return ret;
1474 }
1475
1476 static int check_can_nocow(struct btrfs_inode *inode, loff_t pos,
1477                            size_t *write_bytes, bool nowait)
1478 {
1479         struct btrfs_fs_info *fs_info = inode->root->fs_info;
1480         struct btrfs_root *root = inode->root;
1481         u64 lockstart, lockend;
1482         u64 num_bytes;
1483         int ret;
1484
1485         if (!(inode->flags & (BTRFS_INODE_NODATACOW | BTRFS_INODE_PREALLOC)))
1486                 return 0;
1487
1488         if (!nowait && !btrfs_drew_try_write_lock(&root->snapshot_lock))
1489                 return -EAGAIN;
1490
1491         lockstart = round_down(pos, fs_info->sectorsize);
1492         lockend = round_up(pos + *write_bytes,
1493                            fs_info->sectorsize) - 1;
1494         num_bytes = lockend - lockstart + 1;
1495
1496         if (nowait) {
1497                 struct btrfs_ordered_extent *ordered;
1498
1499                 if (!try_lock_extent(&inode->io_tree, lockstart, lockend))
1500                         return -EAGAIN;
1501
1502                 ordered = btrfs_lookup_ordered_range(inode, lockstart,
1503                                                      num_bytes);
1504                 if (ordered) {
1505                         btrfs_put_ordered_extent(ordered);
1506                         ret = -EAGAIN;
1507                         goto out_unlock;
1508                 }
1509         } else {
1510                 btrfs_lock_and_flush_ordered_range(inode, lockstart,
1511                                                    lockend, NULL);
1512         }
1513
1514         ret = can_nocow_extent(&inode->vfs_inode, lockstart, &num_bytes,
1515                         NULL, NULL, NULL, false);
1516         if (ret <= 0) {
1517                 ret = 0;
1518                 if (!nowait)
1519                         btrfs_drew_write_unlock(&root->snapshot_lock);
1520         } else {
1521                 *write_bytes = min_t(size_t, *write_bytes ,
1522                                      num_bytes - pos + lockstart);
1523         }
1524 out_unlock:
1525         unlock_extent(&inode->io_tree, lockstart, lockend);
1526
1527         return ret;
1528 }
1529
1530 static int check_nocow_nolock(struct btrfs_inode *inode, loff_t pos,
1531                               size_t *write_bytes)
1532 {
1533         return check_can_nocow(inode, pos, write_bytes, true);
1534 }
1535
1536 /*
1537  * Check if we can do nocow write into the range [@pos, @pos + @write_bytes)
1538  *
1539  * @pos:         File offset
1540  * @write_bytes: The length to write, will be updated to the nocow writeable
1541  *               range
1542  *
1543  * This function will flush ordered extents in the range to ensure proper
1544  * nocow checks.
1545  *
1546  * Return:
1547  * >0           and update @write_bytes if we can do nocow write
1548  *  0           if we can't do nocow write
1549  * -EAGAIN      if we can't get the needed lock or there are ordered extents
1550  *              for * (nowait == true) case
1551  * <0           if other error happened
1552  *
1553  * NOTE: Callers need to release the lock by btrfs_check_nocow_unlock().
1554  */
1555 int btrfs_check_nocow_lock(struct btrfs_inode *inode, loff_t pos,
1556                            size_t *write_bytes)
1557 {
1558         return check_can_nocow(inode, pos, write_bytes, false);
1559 }
1560
1561 void btrfs_check_nocow_unlock(struct btrfs_inode *inode)
1562 {
1563         btrfs_drew_write_unlock(&inode->root->snapshot_lock);
1564 }
1565
1566 static void update_time_for_write(struct inode *inode)
1567 {
1568         struct timespec64 now;
1569
1570         if (IS_NOCMTIME(inode))
1571                 return;
1572
1573         now = current_time(inode);
1574         if (!timespec64_equal(&inode->i_mtime, &now))
1575                 inode->i_mtime = now;
1576
1577         if (!timespec64_equal(&inode->i_ctime, &now))
1578                 inode->i_ctime = now;
1579
1580         if (IS_I_VERSION(inode))
1581                 inode_inc_iversion(inode);
1582 }
1583
1584 static int btrfs_write_check(struct kiocb *iocb, struct iov_iter *from,
1585                              size_t count)
1586 {
1587         struct file *file = iocb->ki_filp;
1588         struct inode *inode = file_inode(file);
1589         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1590         loff_t pos = iocb->ki_pos;
1591         int ret;
1592         loff_t oldsize;
1593         loff_t start_pos;
1594
1595         if (iocb->ki_flags & IOCB_NOWAIT) {
1596                 size_t nocow_bytes = count;
1597
1598                 /* We will allocate space in case nodatacow is not set, so bail */
1599                 if (check_nocow_nolock(BTRFS_I(inode), pos, &nocow_bytes) <= 0)
1600                         return -EAGAIN;
1601                 /*
1602                  * There are holes in the range or parts of the range that must
1603                  * be COWed (shared extents, RO block groups, etc), so just bail
1604                  * out.
1605                  */
1606                 if (nocow_bytes < count)
1607                         return -EAGAIN;
1608         }
1609
1610         current->backing_dev_info = inode_to_bdi(inode);
1611         ret = file_remove_privs(file);
1612         if (ret)
1613                 return ret;
1614
1615         /*
1616          * We reserve space for updating the inode when we reserve space for the
1617          * extent we are going to write, so we will enospc out there.  We don't
1618          * need to start yet another transaction to update the inode as we will
1619          * update the inode when we finish writing whatever data we write.
1620          */
1621         update_time_for_write(inode);
1622
1623         start_pos = round_down(pos, fs_info->sectorsize);
1624         oldsize = i_size_read(inode);
1625         if (start_pos > oldsize) {
1626                 /* Expand hole size to cover write data, preventing empty gap */
1627                 loff_t end_pos = round_up(pos + count, fs_info->sectorsize);
1628
1629                 ret = btrfs_cont_expand(inode, oldsize, end_pos);
1630                 if (ret) {
1631                         current->backing_dev_info = NULL;
1632                         return ret;
1633                 }
1634         }
1635
1636         return 0;
1637 }
1638
1639 static noinline ssize_t btrfs_buffered_write(struct kiocb *iocb,
1640                                                struct iov_iter *i)
1641 {
1642         struct file *file = iocb->ki_filp;
1643         loff_t pos;
1644         struct inode *inode = file_inode(file);
1645         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1646         struct page **pages = NULL;
1647         struct extent_changeset *data_reserved = NULL;
1648         u64 release_bytes = 0;
1649         u64 lockstart;
1650         u64 lockend;
1651         size_t num_written = 0;
1652         int nrptrs;
1653         ssize_t ret;
1654         bool only_release_metadata = false;
1655         bool force_page_uptodate = false;
1656         loff_t old_isize = i_size_read(inode);
1657         unsigned int ilock_flags = 0;
1658
1659         if (iocb->ki_flags & IOCB_NOWAIT)
1660                 ilock_flags |= BTRFS_ILOCK_TRY;
1661
1662         ret = btrfs_inode_lock(inode, ilock_flags);
1663         if (ret < 0)
1664                 return ret;
1665
1666         ret = generic_write_checks(iocb, i);
1667         if (ret <= 0)
1668                 goto out;
1669
1670         ret = btrfs_write_check(iocb, i, ret);
1671         if (ret < 0)
1672                 goto out;
1673
1674         pos = iocb->ki_pos;
1675         nrptrs = min(DIV_ROUND_UP(iov_iter_count(i), PAGE_SIZE),
1676                         PAGE_SIZE / (sizeof(struct page *)));
1677         nrptrs = min(nrptrs, current->nr_dirtied_pause - current->nr_dirtied);
1678         nrptrs = max(nrptrs, 8);
1679         pages = kmalloc_array(nrptrs, sizeof(struct page *), GFP_KERNEL);
1680         if (!pages) {
1681                 ret = -ENOMEM;
1682                 goto out;
1683         }
1684
1685         while (iov_iter_count(i) > 0) {
1686                 struct extent_state *cached_state = NULL;
1687                 size_t offset = offset_in_page(pos);
1688                 size_t sector_offset;
1689                 size_t write_bytes = min(iov_iter_count(i),
1690                                          nrptrs * (size_t)PAGE_SIZE -
1691                                          offset);
1692                 size_t num_pages;
1693                 size_t reserve_bytes;
1694                 size_t dirty_pages;
1695                 size_t copied;
1696                 size_t dirty_sectors;
1697                 size_t num_sectors;
1698                 int extents_locked;
1699
1700                 /*
1701                  * Fault pages before locking them in prepare_pages
1702                  * to avoid recursive lock
1703                  */
1704                 if (unlikely(iov_iter_fault_in_readable(i, write_bytes))) {
1705                         ret = -EFAULT;
1706                         break;
1707                 }
1708
1709                 only_release_metadata = false;
1710                 sector_offset = pos & (fs_info->sectorsize - 1);
1711
1712                 extent_changeset_release(data_reserved);
1713                 ret = btrfs_check_data_free_space(BTRFS_I(inode),
1714                                                   &data_reserved, pos,
1715                                                   write_bytes);
1716                 if (ret < 0) {
1717                         /*
1718                          * If we don't have to COW at the offset, reserve
1719                          * metadata only. write_bytes may get smaller than
1720                          * requested here.
1721                          */
1722                         if (btrfs_check_nocow_lock(BTRFS_I(inode), pos,
1723                                                    &write_bytes) > 0)
1724                                 only_release_metadata = true;
1725                         else
1726                                 break;
1727                 }
1728
1729                 num_pages = DIV_ROUND_UP(write_bytes + offset, PAGE_SIZE);
1730                 WARN_ON(num_pages > nrptrs);
1731                 reserve_bytes = round_up(write_bytes + sector_offset,
1732                                          fs_info->sectorsize);
1733                 WARN_ON(reserve_bytes == 0);
1734                 ret = btrfs_delalloc_reserve_metadata(BTRFS_I(inode),
1735                                 reserve_bytes);
1736                 if (ret) {
1737                         if (!only_release_metadata)
1738                                 btrfs_free_reserved_data_space(BTRFS_I(inode),
1739                                                 data_reserved, pos,
1740                                                 write_bytes);
1741                         else
1742                                 btrfs_check_nocow_unlock(BTRFS_I(inode));
1743                         break;
1744                 }
1745
1746                 release_bytes = reserve_bytes;
1747 again:
1748                 /*
1749                  * This is going to setup the pages array with the number of
1750                  * pages we want, so we don't really need to worry about the
1751                  * contents of pages from loop to loop
1752                  */
1753                 ret = prepare_pages(inode, pages, num_pages,
1754                                     pos, write_bytes,
1755                                     force_page_uptodate);
1756                 if (ret) {
1757                         btrfs_delalloc_release_extents(BTRFS_I(inode),
1758                                                        reserve_bytes);
1759                         break;
1760                 }
1761
1762                 extents_locked = lock_and_cleanup_extent_if_need(
1763                                 BTRFS_I(inode), pages,
1764                                 num_pages, pos, write_bytes, &lockstart,
1765                                 &lockend, &cached_state);
1766                 if (extents_locked < 0) {
1767                         if (extents_locked == -EAGAIN)
1768                                 goto again;
1769                         btrfs_delalloc_release_extents(BTRFS_I(inode),
1770                                                        reserve_bytes);
1771                         ret = extents_locked;
1772                         break;
1773                 }
1774
1775                 copied = btrfs_copy_from_user(pos, write_bytes, pages, i);
1776
1777                 num_sectors = BTRFS_BYTES_TO_BLKS(fs_info, reserve_bytes);
1778                 dirty_sectors = round_up(copied + sector_offset,
1779                                         fs_info->sectorsize);
1780                 dirty_sectors = BTRFS_BYTES_TO_BLKS(fs_info, dirty_sectors);
1781
1782                 /*
1783                  * if we have trouble faulting in the pages, fall
1784                  * back to one page at a time
1785                  */
1786                 if (copied < write_bytes)
1787                         nrptrs = 1;
1788
1789                 if (copied == 0) {
1790                         force_page_uptodate = true;
1791                         dirty_sectors = 0;
1792                         dirty_pages = 0;
1793                 } else {
1794                         force_page_uptodate = false;
1795                         dirty_pages = DIV_ROUND_UP(copied + offset,
1796                                                    PAGE_SIZE);
1797                 }
1798
1799                 if (num_sectors > dirty_sectors) {
1800                         /* release everything except the sectors we dirtied */
1801                         release_bytes -= dirty_sectors << fs_info->sectorsize_bits;
1802                         if (only_release_metadata) {
1803                                 btrfs_delalloc_release_metadata(BTRFS_I(inode),
1804                                                         release_bytes, true);
1805                         } else {
1806                                 u64 __pos;
1807
1808                                 __pos = round_down(pos,
1809                                                    fs_info->sectorsize) +
1810                                         (dirty_pages << PAGE_SHIFT);
1811                                 btrfs_delalloc_release_space(BTRFS_I(inode),
1812                                                 data_reserved, __pos,
1813                                                 release_bytes, true);
1814                         }
1815                 }
1816
1817                 release_bytes = round_up(copied + sector_offset,
1818                                         fs_info->sectorsize);
1819
1820                 ret = btrfs_dirty_pages(BTRFS_I(inode), pages,
1821                                         dirty_pages, pos, copied,
1822                                         &cached_state, only_release_metadata);
1823
1824                 /*
1825                  * If we have not locked the extent range, because the range's
1826                  * start offset is >= i_size, we might still have a non-NULL
1827                  * cached extent state, acquired while marking the extent range
1828                  * as delalloc through btrfs_dirty_pages(). Therefore free any
1829                  * possible cached extent state to avoid a memory leak.
1830                  */
1831                 if (extents_locked)
1832                         unlock_extent_cached(&BTRFS_I(inode)->io_tree,
1833                                              lockstart, lockend, &cached_state);
1834                 else
1835                         free_extent_state(cached_state);
1836
1837                 btrfs_delalloc_release_extents(BTRFS_I(inode), reserve_bytes);
1838                 if (ret) {
1839                         btrfs_drop_pages(pages, num_pages);
1840                         break;
1841                 }
1842
1843                 release_bytes = 0;
1844                 if (only_release_metadata)
1845                         btrfs_check_nocow_unlock(BTRFS_I(inode));
1846
1847                 btrfs_drop_pages(pages, num_pages);
1848
1849                 cond_resched();
1850
1851                 balance_dirty_pages_ratelimited(inode->i_mapping);
1852
1853                 pos += copied;
1854                 num_written += copied;
1855         }
1856
1857         kfree(pages);
1858
1859         if (release_bytes) {
1860                 if (only_release_metadata) {
1861                         btrfs_check_nocow_unlock(BTRFS_I(inode));
1862                         btrfs_delalloc_release_metadata(BTRFS_I(inode),
1863                                         release_bytes, true);
1864                 } else {
1865                         btrfs_delalloc_release_space(BTRFS_I(inode),
1866                                         data_reserved,
1867                                         round_down(pos, fs_info->sectorsize),
1868                                         release_bytes, true);
1869                 }
1870         }
1871
1872         extent_changeset_free(data_reserved);
1873         if (num_written > 0) {
1874                 pagecache_isize_extended(inode, old_isize, iocb->ki_pos);
1875                 iocb->ki_pos += num_written;
1876         }
1877 out:
1878         btrfs_inode_unlock(inode, ilock_flags);
1879         return num_written ? num_written : ret;
1880 }
1881
1882 static ssize_t check_direct_IO(struct btrfs_fs_info *fs_info,
1883                                const struct iov_iter *iter, loff_t offset)
1884 {
1885         const u32 blocksize_mask = fs_info->sectorsize - 1;
1886
1887         if (offset & blocksize_mask)
1888                 return -EINVAL;
1889
1890         if (iov_iter_alignment(iter) & blocksize_mask)
1891                 return -EINVAL;
1892
1893         return 0;
1894 }
1895
1896 static ssize_t btrfs_direct_write(struct kiocb *iocb, struct iov_iter *from)
1897 {
1898         struct file *file = iocb->ki_filp;
1899         struct inode *inode = file_inode(file);
1900         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1901         loff_t pos;
1902         ssize_t written = 0;
1903         ssize_t written_buffered;
1904         loff_t endbyte;
1905         ssize_t err;
1906         unsigned int ilock_flags = 0;
1907         struct iomap_dio *dio = NULL;
1908
1909         if (iocb->ki_flags & IOCB_NOWAIT)
1910                 ilock_flags |= BTRFS_ILOCK_TRY;
1911
1912         /* If the write DIO is within EOF, use a shared lock */
1913         if (iocb->ki_pos + iov_iter_count(from) <= i_size_read(inode))
1914                 ilock_flags |= BTRFS_ILOCK_SHARED;
1915
1916 relock:
1917         err = btrfs_inode_lock(inode, ilock_flags);
1918         if (err < 0)
1919                 return err;
1920
1921         err = generic_write_checks(iocb, from);
1922         if (err <= 0) {
1923                 btrfs_inode_unlock(inode, ilock_flags);
1924                 return err;
1925         }
1926
1927         err = btrfs_write_check(iocb, from, err);
1928         if (err < 0) {
1929                 btrfs_inode_unlock(inode, ilock_flags);
1930                 goto out;
1931         }
1932
1933         pos = iocb->ki_pos;
1934         /*
1935          * Re-check since file size may have changed just before taking the
1936          * lock or pos may have changed because of O_APPEND in generic_write_check()
1937          */
1938         if ((ilock_flags & BTRFS_ILOCK_SHARED) &&
1939             pos + iov_iter_count(from) > i_size_read(inode)) {
1940                 btrfs_inode_unlock(inode, ilock_flags);
1941                 ilock_flags &= ~BTRFS_ILOCK_SHARED;
1942                 goto relock;
1943         }
1944
1945         if (check_direct_IO(fs_info, from, pos)) {
1946                 btrfs_inode_unlock(inode, ilock_flags);
1947                 goto buffered;
1948         }
1949
1950         dio = __iomap_dio_rw(iocb, from, &btrfs_dio_iomap_ops,
1951                              &btrfs_dio_ops, is_sync_kiocb(iocb));
1952
1953         btrfs_inode_unlock(inode, ilock_flags);
1954
1955         if (IS_ERR_OR_NULL(dio)) {
1956                 err = PTR_ERR_OR_ZERO(dio);
1957                 if (err < 0 && err != -ENOTBLK)
1958                         goto out;
1959         } else {
1960                 written = iomap_dio_complete(dio);
1961         }
1962
1963         if (written < 0 || !iov_iter_count(from)) {
1964                 err = written;
1965                 goto out;
1966         }
1967
1968 buffered:
1969         pos = iocb->ki_pos;
1970         written_buffered = btrfs_buffered_write(iocb, from);
1971         if (written_buffered < 0) {
1972                 err = written_buffered;
1973                 goto out;
1974         }
1975         /*
1976          * Ensure all data is persisted. We want the next direct IO read to be
1977          * able to read what was just written.
1978          */
1979         endbyte = pos + written_buffered - 1;
1980         err = btrfs_fdatawrite_range(inode, pos, endbyte);
1981         if (err)
1982                 goto out;
1983         err = filemap_fdatawait_range(inode->i_mapping, pos, endbyte);
1984         if (err)
1985                 goto out;
1986         written += written_buffered;
1987         iocb->ki_pos = pos + written_buffered;
1988         invalidate_mapping_pages(file->f_mapping, pos >> PAGE_SHIFT,
1989                                  endbyte >> PAGE_SHIFT);
1990 out:
1991         return written ? written : err;
1992 }
1993
1994 static ssize_t btrfs_file_write_iter(struct kiocb *iocb,
1995                                     struct iov_iter *from)
1996 {
1997         struct file *file = iocb->ki_filp;
1998         struct inode *inode = file_inode(file);
1999         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2000         struct btrfs_root *root = BTRFS_I(inode)->root;
2001         ssize_t num_written = 0;
2002         const bool sync = iocb->ki_flags & IOCB_DSYNC;
2003
2004         /*
2005          * If the fs flips readonly due to some impossible error, although we
2006          * have opened a file as writable, we have to stop this write operation
2007          * to ensure consistency.
2008          */
2009         if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state))
2010                 return -EROFS;
2011
2012         if (!(iocb->ki_flags & IOCB_DIRECT) &&
2013             (iocb->ki_flags & IOCB_NOWAIT))
2014                 return -EOPNOTSUPP;
2015
2016         if (sync)
2017                 atomic_inc(&BTRFS_I(inode)->sync_writers);
2018
2019         if (iocb->ki_flags & IOCB_DIRECT)
2020                 num_written = btrfs_direct_write(iocb, from);
2021         else
2022                 num_written = btrfs_buffered_write(iocb, from);
2023
2024         /*
2025          * We also have to set last_sub_trans to the current log transid,
2026          * otherwise subsequent syncs to a file that's been synced in this
2027          * transaction will appear to have already occurred.
2028          */
2029         spin_lock(&BTRFS_I(inode)->lock);
2030         BTRFS_I(inode)->last_sub_trans = root->log_transid;
2031         spin_unlock(&BTRFS_I(inode)->lock);
2032         if (num_written > 0)
2033                 num_written = generic_write_sync(iocb, num_written);
2034
2035         if (sync)
2036                 atomic_dec(&BTRFS_I(inode)->sync_writers);
2037
2038         current->backing_dev_info = NULL;
2039         return num_written;
2040 }
2041
2042 int btrfs_release_file(struct inode *inode, struct file *filp)
2043 {
2044         struct btrfs_file_private *private = filp->private_data;
2045
2046         if (private && private->filldir_buf)
2047                 kfree(private->filldir_buf);
2048         kfree(private);
2049         filp->private_data = NULL;
2050
2051         /*
2052          * Set by setattr when we are about to truncate a file from a non-zero
2053          * size to a zero size.  This tries to flush down new bytes that may
2054          * have been written if the application were using truncate to replace
2055          * a file in place.
2056          */
2057         if (test_and_clear_bit(BTRFS_INODE_FLUSH_ON_CLOSE,
2058                                &BTRFS_I(inode)->runtime_flags))
2059                         filemap_flush(inode->i_mapping);
2060         return 0;
2061 }
2062
2063 static int start_ordered_ops(struct inode *inode, loff_t start, loff_t end)
2064 {
2065         int ret;
2066         struct blk_plug plug;
2067
2068         /*
2069          * This is only called in fsync, which would do synchronous writes, so
2070          * a plug can merge adjacent IOs as much as possible.  Esp. in case of
2071          * multiple disks using raid profile, a large IO can be split to
2072          * several segments of stripe length (currently 64K).
2073          */
2074         blk_start_plug(&plug);
2075         atomic_inc(&BTRFS_I(inode)->sync_writers);
2076         ret = btrfs_fdatawrite_range(inode, start, end);
2077         atomic_dec(&BTRFS_I(inode)->sync_writers);
2078         blk_finish_plug(&plug);
2079
2080         return ret;
2081 }
2082
2083 /*
2084  * fsync call for both files and directories.  This logs the inode into
2085  * the tree log instead of forcing full commits whenever possible.
2086  *
2087  * It needs to call filemap_fdatawait so that all ordered extent updates are
2088  * in the metadata btree are up to date for copying to the log.
2089  *
2090  * It drops the inode mutex before doing the tree log commit.  This is an
2091  * important optimization for directories because holding the mutex prevents
2092  * new operations on the dir while we write to disk.
2093  */
2094 int btrfs_sync_file(struct file *file, loff_t start, loff_t end, int datasync)
2095 {
2096         struct dentry *dentry = file_dentry(file);
2097         struct inode *inode = d_inode(dentry);
2098         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2099         struct btrfs_root *root = BTRFS_I(inode)->root;
2100         struct btrfs_trans_handle *trans;
2101         struct btrfs_log_ctx ctx;
2102         int ret = 0, err;
2103         u64 len;
2104         bool full_sync;
2105
2106         trace_btrfs_sync_file(file, datasync);
2107
2108         btrfs_init_log_ctx(&ctx, inode);
2109
2110         /*
2111          * Always set the range to a full range, otherwise we can get into
2112          * several problems, from missing file extent items to represent holes
2113          * when not using the NO_HOLES feature, to log tree corruption due to
2114          * races between hole detection during logging and completion of ordered
2115          * extents outside the range, to missing checksums due to ordered extents
2116          * for which we flushed only a subset of their pages.
2117          */
2118         start = 0;
2119         end = LLONG_MAX;
2120         len = (u64)LLONG_MAX + 1;
2121
2122         /*
2123          * We write the dirty pages in the range and wait until they complete
2124          * out of the ->i_mutex. If so, we can flush the dirty pages by
2125          * multi-task, and make the performance up.  See
2126          * btrfs_wait_ordered_range for an explanation of the ASYNC check.
2127          */
2128         ret = start_ordered_ops(inode, start, end);
2129         if (ret)
2130                 goto out;
2131
2132         inode_lock(inode);
2133
2134         atomic_inc(&root->log_batch);
2135
2136         /*
2137          * Always check for the full sync flag while holding the inode's lock,
2138          * to avoid races with other tasks. The flag must be either set all the
2139          * time during logging or always off all the time while logging.
2140          */
2141         full_sync = test_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2142                              &BTRFS_I(inode)->runtime_flags);
2143
2144         /*
2145          * Before we acquired the inode's lock, someone may have dirtied more
2146          * pages in the target range. We need to make sure that writeback for
2147          * any such pages does not start while we are logging the inode, because
2148          * if it does, any of the following might happen when we are not doing a
2149          * full inode sync:
2150          *
2151          * 1) We log an extent after its writeback finishes but before its
2152          *    checksums are added to the csum tree, leading to -EIO errors
2153          *    when attempting to read the extent after a log replay.
2154          *
2155          * 2) We can end up logging an extent before its writeback finishes.
2156          *    Therefore after the log replay we will have a file extent item
2157          *    pointing to an unwritten extent (and no data checksums as well).
2158          *
2159          * So trigger writeback for any eventual new dirty pages and then we
2160          * wait for all ordered extents to complete below.
2161          */
2162         ret = start_ordered_ops(inode, start, end);
2163         if (ret) {
2164                 inode_unlock(inode);
2165                 goto out;
2166         }
2167
2168         /*
2169          * We have to do this here to avoid the priority inversion of waiting on
2170          * IO of a lower priority task while holding a transaction open.
2171          *
2172          * For a full fsync we wait for the ordered extents to complete while
2173          * for a fast fsync we wait just for writeback to complete, and then
2174          * attach the ordered extents to the transaction so that a transaction
2175          * commit waits for their completion, to avoid data loss if we fsync,
2176          * the current transaction commits before the ordered extents complete
2177          * and a power failure happens right after that.
2178          */
2179         if (full_sync) {
2180                 ret = btrfs_wait_ordered_range(inode, start, len);
2181         } else {
2182                 /*
2183                  * Get our ordered extents as soon as possible to avoid doing
2184                  * checksum lookups in the csum tree, and use instead the
2185                  * checksums attached to the ordered extents.
2186                  */
2187                 btrfs_get_ordered_extents_for_logging(BTRFS_I(inode),
2188                                                       &ctx.ordered_extents);
2189                 ret = filemap_fdatawait_range(inode->i_mapping, start, end);
2190         }
2191
2192         if (ret)
2193                 goto out_release_extents;
2194
2195         atomic_inc(&root->log_batch);
2196
2197         /*
2198          * If we are doing a fast fsync we can not bail out if the inode's
2199          * last_trans is <= then the last committed transaction, because we only
2200          * update the last_trans of the inode during ordered extent completion,
2201          * and for a fast fsync we don't wait for that, we only wait for the
2202          * writeback to complete.
2203          */
2204         smp_mb();
2205         if (btrfs_inode_in_log(BTRFS_I(inode), fs_info->generation) ||
2206             (BTRFS_I(inode)->last_trans <= fs_info->last_trans_committed &&
2207              (full_sync || list_empty(&ctx.ordered_extents)))) {
2208                 /*
2209                  * We've had everything committed since the last time we were
2210                  * modified so clear this flag in case it was set for whatever
2211                  * reason, it's no longer relevant.
2212                  */
2213                 clear_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2214                           &BTRFS_I(inode)->runtime_flags);
2215                 /*
2216                  * An ordered extent might have started before and completed
2217                  * already with io errors, in which case the inode was not
2218                  * updated and we end up here. So check the inode's mapping
2219                  * for any errors that might have happened since we last
2220                  * checked called fsync.
2221                  */
2222                 ret = filemap_check_wb_err(inode->i_mapping, file->f_wb_err);
2223                 goto out_release_extents;
2224         }
2225
2226         /*
2227          * We use start here because we will need to wait on the IO to complete
2228          * in btrfs_sync_log, which could require joining a transaction (for
2229          * example checking cross references in the nocow path).  If we use join
2230          * here we could get into a situation where we're waiting on IO to
2231          * happen that is blocked on a transaction trying to commit.  With start
2232          * we inc the extwriter counter, so we wait for all extwriters to exit
2233          * before we start blocking joiners.  This comment is to keep somebody
2234          * from thinking they are super smart and changing this to
2235          * btrfs_join_transaction *cough*Josef*cough*.
2236          */
2237         trans = btrfs_start_transaction(root, 0);
2238         if (IS_ERR(trans)) {
2239                 ret = PTR_ERR(trans);
2240                 goto out_release_extents;
2241         }
2242
2243         ret = btrfs_log_dentry_safe(trans, dentry, &ctx);
2244         btrfs_release_log_ctx_extents(&ctx);
2245         if (ret < 0) {
2246                 /* Fallthrough and commit/free transaction. */
2247                 ret = 1;
2248         }
2249
2250         /* we've logged all the items and now have a consistent
2251          * version of the file in the log.  It is possible that
2252          * someone will come in and modify the file, but that's
2253          * fine because the log is consistent on disk, and we
2254          * have references to all of the file's extents
2255          *
2256          * It is possible that someone will come in and log the
2257          * file again, but that will end up using the synchronization
2258          * inside btrfs_sync_log to keep things safe.
2259          */
2260         inode_unlock(inode);
2261
2262         if (ret != BTRFS_NO_LOG_SYNC) {
2263                 if (!ret) {
2264                         ret = btrfs_sync_log(trans, root, &ctx);
2265                         if (!ret) {
2266                                 ret = btrfs_end_transaction(trans);
2267                                 goto out;
2268                         }
2269                 }
2270                 if (!full_sync) {
2271                         ret = btrfs_wait_ordered_range(inode, start, len);
2272                         if (ret) {
2273                                 btrfs_end_transaction(trans);
2274                                 goto out;
2275                         }
2276                 }
2277                 ret = btrfs_commit_transaction(trans);
2278         } else {
2279                 ret = btrfs_end_transaction(trans);
2280         }
2281 out:
2282         ASSERT(list_empty(&ctx.list));
2283         err = file_check_and_advance_wb_err(file);
2284         if (!ret)
2285                 ret = err;
2286         return ret > 0 ? -EIO : ret;
2287
2288 out_release_extents:
2289         btrfs_release_log_ctx_extents(&ctx);
2290         inode_unlock(inode);
2291         goto out;
2292 }
2293
2294 static const struct vm_operations_struct btrfs_file_vm_ops = {
2295         .fault          = filemap_fault,
2296         .map_pages      = filemap_map_pages,
2297         .page_mkwrite   = btrfs_page_mkwrite,
2298 };
2299
2300 static int btrfs_file_mmap(struct file  *filp, struct vm_area_struct *vma)
2301 {
2302         struct address_space *mapping = filp->f_mapping;
2303
2304         if (!mapping->a_ops->readpage)
2305                 return -ENOEXEC;
2306
2307         file_accessed(filp);
2308         vma->vm_ops = &btrfs_file_vm_ops;
2309
2310         return 0;
2311 }
2312
2313 static int hole_mergeable(struct btrfs_inode *inode, struct extent_buffer *leaf,
2314                           int slot, u64 start, u64 end)
2315 {
2316         struct btrfs_file_extent_item *fi;
2317         struct btrfs_key key;
2318
2319         if (slot < 0 || slot >= btrfs_header_nritems(leaf))
2320                 return 0;
2321
2322         btrfs_item_key_to_cpu(leaf, &key, slot);
2323         if (key.objectid != btrfs_ino(inode) ||
2324             key.type != BTRFS_EXTENT_DATA_KEY)
2325                 return 0;
2326
2327         fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
2328
2329         if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG)
2330                 return 0;
2331
2332         if (btrfs_file_extent_disk_bytenr(leaf, fi))
2333                 return 0;
2334
2335         if (key.offset == end)
2336                 return 1;
2337         if (key.offset + btrfs_file_extent_num_bytes(leaf, fi) == start)
2338                 return 1;
2339         return 0;
2340 }
2341
2342 static int fill_holes(struct btrfs_trans_handle *trans,
2343                 struct btrfs_inode *inode,
2344                 struct btrfs_path *path, u64 offset, u64 end)
2345 {
2346         struct btrfs_fs_info *fs_info = trans->fs_info;
2347         struct btrfs_root *root = inode->root;
2348         struct extent_buffer *leaf;
2349         struct btrfs_file_extent_item *fi;
2350         struct extent_map *hole_em;
2351         struct extent_map_tree *em_tree = &inode->extent_tree;
2352         struct btrfs_key key;
2353         int ret;
2354
2355         if (btrfs_fs_incompat(fs_info, NO_HOLES))
2356                 goto out;
2357
2358         key.objectid = btrfs_ino(inode);
2359         key.type = BTRFS_EXTENT_DATA_KEY;
2360         key.offset = offset;
2361
2362         ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
2363         if (ret <= 0) {
2364                 /*
2365                  * We should have dropped this offset, so if we find it then
2366                  * something has gone horribly wrong.
2367                  */
2368                 if (ret == 0)
2369                         ret = -EINVAL;
2370                 return ret;
2371         }
2372
2373         leaf = path->nodes[0];
2374         if (hole_mergeable(inode, leaf, path->slots[0] - 1, offset, end)) {
2375                 u64 num_bytes;
2376
2377                 path->slots[0]--;
2378                 fi = btrfs_item_ptr(leaf, path->slots[0],
2379                                     struct btrfs_file_extent_item);
2380                 num_bytes = btrfs_file_extent_num_bytes(leaf, fi) +
2381                         end - offset;
2382                 btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2383                 btrfs_set_file_extent_ram_bytes(leaf, fi, num_bytes);
2384                 btrfs_set_file_extent_offset(leaf, fi, 0);
2385                 btrfs_mark_buffer_dirty(leaf);
2386                 goto out;
2387         }
2388
2389         if (hole_mergeable(inode, leaf, path->slots[0], offset, end)) {
2390                 u64 num_bytes;
2391
2392                 key.offset = offset;
2393                 btrfs_set_item_key_safe(fs_info, path, &key);
2394                 fi = btrfs_item_ptr(leaf, path->slots[0],
2395                                     struct btrfs_file_extent_item);
2396                 num_bytes = btrfs_file_extent_num_bytes(leaf, fi) + end -
2397                         offset;
2398                 btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2399                 btrfs_set_file_extent_ram_bytes(leaf, fi, num_bytes);
2400                 btrfs_set_file_extent_offset(leaf, fi, 0);
2401                 btrfs_mark_buffer_dirty(leaf);
2402                 goto out;
2403         }
2404         btrfs_release_path(path);
2405
2406         ret = btrfs_insert_file_extent(trans, root, btrfs_ino(inode),
2407                         offset, 0, 0, end - offset, 0, end - offset, 0, 0, 0);
2408         if (ret)
2409                 return ret;
2410
2411 out:
2412         btrfs_release_path(path);
2413
2414         hole_em = alloc_extent_map();
2415         if (!hole_em) {
2416                 btrfs_drop_extent_cache(inode, offset, end - 1, 0);
2417                 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &inode->runtime_flags);
2418         } else {
2419                 hole_em->start = offset;
2420                 hole_em->len = end - offset;
2421                 hole_em->ram_bytes = hole_em->len;
2422                 hole_em->orig_start = offset;
2423
2424                 hole_em->block_start = EXTENT_MAP_HOLE;
2425                 hole_em->block_len = 0;
2426                 hole_em->orig_block_len = 0;
2427                 hole_em->compress_type = BTRFS_COMPRESS_NONE;
2428                 hole_em->generation = trans->transid;
2429
2430                 do {
2431                         btrfs_drop_extent_cache(inode, offset, end - 1, 0);
2432                         write_lock(&em_tree->lock);
2433                         ret = add_extent_mapping(em_tree, hole_em, 1);
2434                         write_unlock(&em_tree->lock);
2435                 } while (ret == -EEXIST);
2436                 free_extent_map(hole_em);
2437                 if (ret)
2438                         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2439                                         &inode->runtime_flags);
2440         }
2441
2442         return 0;
2443 }
2444
2445 /*
2446  * Find a hole extent on given inode and change start/len to the end of hole
2447  * extent.(hole/vacuum extent whose em->start <= start &&
2448  *         em->start + em->len > start)
2449  * When a hole extent is found, return 1 and modify start/len.
2450  */
2451 static int find_first_non_hole(struct inode *inode, u64 *start, u64 *len)
2452 {
2453         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2454         struct extent_map *em;
2455         int ret = 0;
2456
2457         em = btrfs_get_extent(BTRFS_I(inode), NULL, 0,
2458                               round_down(*start, fs_info->sectorsize),
2459                               round_up(*len, fs_info->sectorsize));
2460         if (IS_ERR(em))
2461                 return PTR_ERR(em);
2462
2463         /* Hole or vacuum extent(only exists in no-hole mode) */
2464         if (em->block_start == EXTENT_MAP_HOLE) {
2465                 ret = 1;
2466                 *len = em->start + em->len > *start + *len ?
2467                        0 : *start + *len - em->start - em->len;
2468                 *start = em->start + em->len;
2469         }
2470         free_extent_map(em);
2471         return ret;
2472 }
2473
2474 static int btrfs_punch_hole_lock_range(struct inode *inode,
2475                                        const u64 lockstart,
2476                                        const u64 lockend,
2477                                        struct extent_state **cached_state)
2478 {
2479         while (1) {
2480                 struct btrfs_ordered_extent *ordered;
2481                 int ret;
2482
2483                 truncate_pagecache_range(inode, lockstart, lockend);
2484
2485                 lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2486                                  cached_state);
2487                 ordered = btrfs_lookup_first_ordered_extent(BTRFS_I(inode),
2488                                                             lockend);
2489
2490                 /*
2491                  * We need to make sure we have no ordered extents in this range
2492                  * and nobody raced in and read a page in this range, if we did
2493                  * we need to try again.
2494                  */
2495                 if ((!ordered ||
2496                     (ordered->file_offset + ordered->num_bytes <= lockstart ||
2497                      ordered->file_offset > lockend)) &&
2498                      !filemap_range_has_page(inode->i_mapping,
2499                                              lockstart, lockend)) {
2500                         if (ordered)
2501                                 btrfs_put_ordered_extent(ordered);
2502                         break;
2503                 }
2504                 if (ordered)
2505                         btrfs_put_ordered_extent(ordered);
2506                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart,
2507                                      lockend, cached_state);
2508                 ret = btrfs_wait_ordered_range(inode, lockstart,
2509                                                lockend - lockstart + 1);
2510                 if (ret)
2511                         return ret;
2512         }
2513         return 0;
2514 }
2515
2516 static int btrfs_insert_replace_extent(struct btrfs_trans_handle *trans,
2517                                      struct inode *inode,
2518                                      struct btrfs_path *path,
2519                                      struct btrfs_replace_extent_info *extent_info,
2520                                      const u64 replace_len)
2521 {
2522         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2523         struct btrfs_root *root = BTRFS_I(inode)->root;
2524         struct btrfs_file_extent_item *extent;
2525         struct extent_buffer *leaf;
2526         struct btrfs_key key;
2527         int slot;
2528         struct btrfs_ref ref = { 0 };
2529         int ret;
2530
2531         if (replace_len == 0)
2532                 return 0;
2533
2534         if (extent_info->disk_offset == 0 &&
2535             btrfs_fs_incompat(fs_info, NO_HOLES))
2536                 return 0;
2537
2538         key.objectid = btrfs_ino(BTRFS_I(inode));
2539         key.type = BTRFS_EXTENT_DATA_KEY;
2540         key.offset = extent_info->file_offset;
2541         ret = btrfs_insert_empty_item(trans, root, path, &key,
2542                                       sizeof(struct btrfs_file_extent_item));
2543         if (ret)
2544                 return ret;
2545         leaf = path->nodes[0];
2546         slot = path->slots[0];
2547         write_extent_buffer(leaf, extent_info->extent_buf,
2548                             btrfs_item_ptr_offset(leaf, slot),
2549                             sizeof(struct btrfs_file_extent_item));
2550         extent = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
2551         ASSERT(btrfs_file_extent_type(leaf, extent) != BTRFS_FILE_EXTENT_INLINE);
2552         btrfs_set_file_extent_offset(leaf, extent, extent_info->data_offset);
2553         btrfs_set_file_extent_num_bytes(leaf, extent, replace_len);
2554         if (extent_info->is_new_extent)
2555                 btrfs_set_file_extent_generation(leaf, extent, trans->transid);
2556         btrfs_mark_buffer_dirty(leaf);
2557         btrfs_release_path(path);
2558
2559         ret = btrfs_inode_set_file_extent_range(BTRFS_I(inode),
2560                         extent_info->file_offset, replace_len);
2561         if (ret)
2562                 return ret;
2563
2564         /* If it's a hole, nothing more needs to be done. */
2565         if (extent_info->disk_offset == 0)
2566                 return 0;
2567
2568         inode_add_bytes(inode, replace_len);
2569
2570         if (extent_info->is_new_extent && extent_info->insertions == 0) {
2571                 key.objectid = extent_info->disk_offset;
2572                 key.type = BTRFS_EXTENT_ITEM_KEY;
2573                 key.offset = extent_info->disk_len;
2574                 ret = btrfs_alloc_reserved_file_extent(trans, root,
2575                                                        btrfs_ino(BTRFS_I(inode)),
2576                                                        extent_info->file_offset,
2577                                                        extent_info->qgroup_reserved,
2578                                                        &key);
2579         } else {
2580                 u64 ref_offset;
2581
2582                 btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF,
2583                                        extent_info->disk_offset,
2584                                        extent_info->disk_len, 0);
2585                 ref_offset = extent_info->file_offset - extent_info->data_offset;
2586                 btrfs_init_data_ref(&ref, root->root_key.objectid,
2587                                     btrfs_ino(BTRFS_I(inode)), ref_offset);
2588                 ret = btrfs_inc_extent_ref(trans, &ref);
2589         }
2590
2591         extent_info->insertions++;
2592
2593         return ret;
2594 }
2595
2596 /*
2597  * The respective range must have been previously locked, as well as the inode.
2598  * The end offset is inclusive (last byte of the range).
2599  * @extent_info is NULL for fallocate's hole punching and non-NULL when replacing
2600  * the file range with an extent.
2601  * When not punching a hole, we don't want to end up in a state where we dropped
2602  * extents without inserting a new one, so we must abort the transaction to avoid
2603  * a corruption.
2604  */
2605 int btrfs_replace_file_extents(struct inode *inode, struct btrfs_path *path,
2606                            const u64 start, const u64 end,
2607                            struct btrfs_replace_extent_info *extent_info,
2608                            struct btrfs_trans_handle **trans_out)
2609 {
2610         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2611         u64 min_size = btrfs_calc_insert_metadata_size(fs_info, 1);
2612         u64 ino_size = round_up(inode->i_size, fs_info->sectorsize);
2613         struct btrfs_root *root = BTRFS_I(inode)->root;
2614         struct btrfs_trans_handle *trans = NULL;
2615         struct btrfs_block_rsv *rsv;
2616         unsigned int rsv_count;
2617         u64 cur_offset;
2618         u64 drop_end;
2619         u64 len = end - start;
2620         int ret = 0;
2621
2622         if (end <= start)
2623                 return -EINVAL;
2624
2625         rsv = btrfs_alloc_block_rsv(fs_info, BTRFS_BLOCK_RSV_TEMP);
2626         if (!rsv) {
2627                 ret = -ENOMEM;
2628                 goto out;
2629         }
2630         rsv->size = btrfs_calc_insert_metadata_size(fs_info, 1);
2631         rsv->failfast = 1;
2632
2633         /*
2634          * 1 - update the inode
2635          * 1 - removing the extents in the range
2636          * 1 - adding the hole extent if no_holes isn't set or if we are
2637          *     replacing the range with a new extent
2638          */
2639         if (!btrfs_fs_incompat(fs_info, NO_HOLES) || extent_info)
2640                 rsv_count = 3;
2641         else
2642                 rsv_count = 2;
2643
2644         trans = btrfs_start_transaction(root, rsv_count);
2645         if (IS_ERR(trans)) {
2646                 ret = PTR_ERR(trans);
2647                 trans = NULL;
2648                 goto out_free;
2649         }
2650
2651         ret = btrfs_block_rsv_migrate(&fs_info->trans_block_rsv, rsv,
2652                                       min_size, false);
2653         BUG_ON(ret);
2654         trans->block_rsv = rsv;
2655
2656         cur_offset = start;
2657         while (cur_offset < end) {
2658                 ret = __btrfs_drop_extents(trans, root, BTRFS_I(inode), path,
2659                                            cur_offset, end + 1, &drop_end,
2660                                            1, 0, 0, NULL);
2661                 if (ret != -ENOSPC) {
2662                         /*
2663                          * When cloning we want to avoid transaction aborts when
2664                          * nothing was done and we are attempting to clone parts
2665                          * of inline extents, in such cases -EOPNOTSUPP is
2666                          * returned by __btrfs_drop_extents() without having
2667                          * changed anything in the file.
2668                          */
2669                         if (extent_info && !extent_info->is_new_extent &&
2670                             ret && ret != -EOPNOTSUPP)
2671                                 btrfs_abort_transaction(trans, ret);
2672                         break;
2673                 }
2674
2675                 trans->block_rsv = &fs_info->trans_block_rsv;
2676
2677                 if (!extent_info && cur_offset < drop_end &&
2678                     cur_offset < ino_size) {
2679                         ret = fill_holes(trans, BTRFS_I(inode), path,
2680                                         cur_offset, drop_end);
2681                         if (ret) {
2682                                 /*
2683                                  * If we failed then we didn't insert our hole
2684                                  * entries for the area we dropped, so now the
2685                                  * fs is corrupted, so we must abort the
2686                                  * transaction.
2687                                  */
2688                                 btrfs_abort_transaction(trans, ret);
2689                                 break;
2690                         }
2691                 } else if (!extent_info && cur_offset < drop_end) {
2692                         /*
2693                          * We are past the i_size here, but since we didn't
2694                          * insert holes we need to clear the mapped area so we
2695                          * know to not set disk_i_size in this area until a new
2696                          * file extent is inserted here.
2697                          */
2698                         ret = btrfs_inode_clear_file_extent_range(BTRFS_I(inode),
2699                                         cur_offset, drop_end - cur_offset);
2700                         if (ret) {
2701                                 /*
2702                                  * We couldn't clear our area, so we could
2703                                  * presumably adjust up and corrupt the fs, so
2704                                  * we need to abort.
2705                                  */
2706                                 btrfs_abort_transaction(trans, ret);
2707                                 break;
2708                         }
2709                 }
2710
2711                 if (extent_info && drop_end > extent_info->file_offset) {
2712                         u64 replace_len = drop_end - extent_info->file_offset;
2713
2714                         ret = btrfs_insert_replace_extent(trans, inode, path,
2715                                                         extent_info, replace_len);
2716                         if (ret) {
2717                                 btrfs_abort_transaction(trans, ret);
2718                                 break;
2719                         }
2720                         extent_info->data_len -= replace_len;
2721                         extent_info->data_offset += replace_len;
2722                         extent_info->file_offset += replace_len;
2723                 }
2724
2725                 cur_offset = drop_end;
2726
2727                 ret = btrfs_update_inode(trans, root, inode);
2728                 if (ret)
2729                         break;
2730
2731                 btrfs_end_transaction(trans);
2732                 btrfs_btree_balance_dirty(fs_info);
2733
2734                 trans = btrfs_start_transaction(root, rsv_count);
2735                 if (IS_ERR(trans)) {
2736                         ret = PTR_ERR(trans);
2737                         trans = NULL;
2738                         break;
2739                 }
2740
2741                 ret = btrfs_block_rsv_migrate(&fs_info->trans_block_rsv,
2742                                               rsv, min_size, false);
2743                 BUG_ON(ret);    /* shouldn't happen */
2744                 trans->block_rsv = rsv;
2745
2746                 if (!extent_info) {
2747                         ret = find_first_non_hole(inode, &cur_offset, &len);
2748                         if (unlikely(ret < 0))
2749                                 break;
2750                         if (ret && !len) {
2751                                 ret = 0;
2752                                 break;
2753                         }
2754                 }
2755         }
2756
2757         /*
2758          * If we were cloning, force the next fsync to be a full one since we
2759          * we replaced (or just dropped in the case of cloning holes when
2760          * NO_HOLES is enabled) extents and extent maps.
2761          * This is for the sake of simplicity, and cloning into files larger
2762          * than 16Mb would force the full fsync any way (when
2763          * try_release_extent_mapping() is invoked during page cache truncation.
2764          */
2765         if (extent_info && !extent_info->is_new_extent)
2766                 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2767                         &BTRFS_I(inode)->runtime_flags);
2768
2769         if (ret)
2770                 goto out_trans;
2771
2772         trans->block_rsv = &fs_info->trans_block_rsv;
2773         /*
2774          * If we are using the NO_HOLES feature we might have had already an
2775          * hole that overlaps a part of the region [lockstart, lockend] and
2776          * ends at (or beyond) lockend. Since we have no file extent items to
2777          * represent holes, drop_end can be less than lockend and so we must
2778          * make sure we have an extent map representing the existing hole (the
2779          * call to __btrfs_drop_extents() might have dropped the existing extent
2780          * map representing the existing hole), otherwise the fast fsync path
2781          * will not record the existence of the hole region
2782          * [existing_hole_start, lockend].
2783          */
2784         if (drop_end <= end)
2785                 drop_end = end + 1;
2786         /*
2787          * Don't insert file hole extent item if it's for a range beyond eof
2788          * (because it's useless) or if it represents a 0 bytes range (when
2789          * cur_offset == drop_end).
2790          */
2791         if (!extent_info && cur_offset < ino_size && cur_offset < drop_end) {
2792                 ret = fill_holes(trans, BTRFS_I(inode), path,
2793                                 cur_offset, drop_end);
2794                 if (ret) {
2795                         /* Same comment as above. */
2796                         btrfs_abort_transaction(trans, ret);
2797                         goto out_trans;
2798                 }
2799         } else if (!extent_info && cur_offset < drop_end) {
2800                 /* See the comment in the loop above for the reasoning here. */
2801                 ret = btrfs_inode_clear_file_extent_range(BTRFS_I(inode),
2802                                         cur_offset, drop_end - cur_offset);
2803                 if (ret) {
2804                         btrfs_abort_transaction(trans, ret);
2805                         goto out_trans;
2806                 }
2807
2808         }
2809         if (extent_info) {
2810                 ret = btrfs_insert_replace_extent(trans, inode, path, extent_info,
2811                                                 extent_info->data_len);
2812                 if (ret) {
2813                         btrfs_abort_transaction(trans, ret);
2814                         goto out_trans;
2815                 }
2816         }
2817
2818 out_trans:
2819         if (!trans)
2820                 goto out_free;
2821
2822         trans->block_rsv = &fs_info->trans_block_rsv;
2823         if (ret)
2824                 btrfs_end_transaction(trans);
2825         else
2826                 *trans_out = trans;
2827 out_free:
2828         btrfs_free_block_rsv(fs_info, rsv);
2829 out:
2830         return ret;
2831 }
2832
2833 static int btrfs_punch_hole(struct inode *inode, loff_t offset, loff_t len)
2834 {
2835         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2836         struct btrfs_root *root = BTRFS_I(inode)->root;
2837         struct extent_state *cached_state = NULL;
2838         struct btrfs_path *path;
2839         struct btrfs_trans_handle *trans = NULL;
2840         u64 lockstart;
2841         u64 lockend;
2842         u64 tail_start;
2843         u64 tail_len;
2844         u64 orig_start = offset;
2845         int ret = 0;
2846         bool same_block;
2847         u64 ino_size;
2848         bool truncated_block = false;
2849         bool updated_inode = false;
2850
2851         ret = btrfs_wait_ordered_range(inode, offset, len);
2852         if (ret)
2853                 return ret;
2854
2855         inode_lock(inode);
2856         ino_size = round_up(inode->i_size, fs_info->sectorsize);
2857         ret = find_first_non_hole(inode, &offset, &len);
2858         if (ret < 0)
2859                 goto out_only_mutex;
2860         if (ret && !len) {
2861                 /* Already in a large hole */
2862                 ret = 0;
2863                 goto out_only_mutex;
2864         }
2865
2866         lockstart = round_up(offset, btrfs_inode_sectorsize(BTRFS_I(inode)));
2867         lockend = round_down(offset + len,
2868                              btrfs_inode_sectorsize(BTRFS_I(inode))) - 1;
2869         same_block = (BTRFS_BYTES_TO_BLKS(fs_info, offset))
2870                 == (BTRFS_BYTES_TO_BLKS(fs_info, offset + len - 1));
2871         /*
2872          * We needn't truncate any block which is beyond the end of the file
2873          * because we are sure there is no data there.
2874          */
2875         /*
2876          * Only do this if we are in the same block and we aren't doing the
2877          * entire block.
2878          */
2879         if (same_block && len < fs_info->sectorsize) {
2880                 if (offset < ino_size) {
2881                         truncated_block = true;
2882                         ret = btrfs_truncate_block(inode, offset, len, 0);
2883                 } else {
2884                         ret = 0;
2885                 }
2886                 goto out_only_mutex;
2887         }
2888
2889         /* zero back part of the first block */
2890         if (offset < ino_size) {
2891                 truncated_block = true;
2892                 ret = btrfs_truncate_block(inode, offset, 0, 0);
2893                 if (ret) {
2894                         inode_unlock(inode);
2895                         return ret;
2896                 }
2897         }
2898
2899         /* Check the aligned pages after the first unaligned page,
2900          * if offset != orig_start, which means the first unaligned page
2901          * including several following pages are already in holes,
2902          * the extra check can be skipped */
2903         if (offset == orig_start) {
2904                 /* after truncate page, check hole again */
2905                 len = offset + len - lockstart;
2906                 offset = lockstart;
2907                 ret = find_first_non_hole(inode, &offset, &len);
2908                 if (ret < 0)
2909                         goto out_only_mutex;
2910                 if (ret && !len) {
2911                         ret = 0;
2912                         goto out_only_mutex;
2913                 }
2914                 lockstart = offset;
2915         }
2916
2917         /* Check the tail unaligned part is in a hole */
2918         tail_start = lockend + 1;
2919         tail_len = offset + len - tail_start;
2920         if (tail_len) {
2921                 ret = find_first_non_hole(inode, &tail_start, &tail_len);
2922                 if (unlikely(ret < 0))
2923                         goto out_only_mutex;
2924                 if (!ret) {
2925                         /* zero the front end of the last page */
2926                         if (tail_start + tail_len < ino_size) {
2927                                 truncated_block = true;
2928                                 ret = btrfs_truncate_block(inode,
2929                                                         tail_start + tail_len,
2930                                                         0, 1);
2931                                 if (ret)
2932                                         goto out_only_mutex;
2933                         }
2934                 }
2935         }
2936
2937         if (lockend < lockstart) {
2938                 ret = 0;
2939                 goto out_only_mutex;
2940         }
2941
2942         ret = btrfs_punch_hole_lock_range(inode, lockstart, lockend,
2943                                           &cached_state);
2944         if (ret)
2945                 goto out_only_mutex;
2946
2947         path = btrfs_alloc_path();
2948         if (!path) {
2949                 ret = -ENOMEM;
2950                 goto out;
2951         }
2952
2953         ret = btrfs_replace_file_extents(inode, path, lockstart, lockend, NULL,
2954                                      &trans);
2955         btrfs_free_path(path);
2956         if (ret)
2957                 goto out;
2958
2959         ASSERT(trans != NULL);
2960         inode_inc_iversion(inode);
2961         inode->i_mtime = inode->i_ctime = current_time(inode);
2962         ret = btrfs_update_inode(trans, root, inode);
2963         updated_inode = true;
2964         btrfs_end_transaction(trans);
2965         btrfs_btree_balance_dirty(fs_info);
2966 out:
2967         unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2968                              &cached_state);
2969 out_only_mutex:
2970         if (!updated_inode && truncated_block && !ret) {
2971                 /*
2972                  * If we only end up zeroing part of a page, we still need to
2973                  * update the inode item, so that all the time fields are
2974                  * updated as well as the necessary btrfs inode in memory fields
2975                  * for detecting, at fsync time, if the inode isn't yet in the
2976                  * log tree or it's there but not up to date.
2977                  */
2978                 struct timespec64 now = current_time(inode);
2979
2980                 inode_inc_iversion(inode);
2981                 inode->i_mtime = now;
2982                 inode->i_ctime = now;
2983                 trans = btrfs_start_transaction(root, 1);
2984                 if (IS_ERR(trans)) {
2985                         ret = PTR_ERR(trans);
2986                 } else {
2987                         int ret2;
2988
2989                         ret = btrfs_update_inode(trans, root, inode);
2990                         ret2 = btrfs_end_transaction(trans);
2991                         if (!ret)
2992                                 ret = ret2;
2993                 }
2994         }
2995         inode_unlock(inode);
2996         return ret;
2997 }
2998
2999 /* Helper structure to record which range is already reserved */
3000 struct falloc_range {
3001         struct list_head list;
3002         u64 start;
3003         u64 len;
3004 };
3005
3006 /*
3007  * Helper function to add falloc range
3008  *
3009  * Caller should have locked the larger range of extent containing
3010  * [start, len)
3011  */
3012 static int add_falloc_range(struct list_head *head, u64 start, u64 len)
3013 {
3014         struct falloc_range *prev = NULL;
3015         struct falloc_range *range = NULL;
3016
3017         if (list_empty(head))
3018                 goto insert;
3019
3020         /*
3021          * As fallocate iterate by bytenr order, we only need to check
3022          * the last range.
3023          */
3024         prev = list_entry(head->prev, struct falloc_range, list);
3025         if (prev->start + prev->len == start) {
3026                 prev->len += len;
3027                 return 0;
3028         }
3029 insert:
3030         range = kmalloc(sizeof(*range), GFP_KERNEL);
3031         if (!range)
3032                 return -ENOMEM;
3033         range->start = start;
3034         range->len = len;
3035         list_add_tail(&range->list, head);
3036         return 0;
3037 }
3038
3039 static int btrfs_fallocate_update_isize(struct inode *inode,
3040                                         const u64 end,
3041                                         const int mode)
3042 {
3043         struct btrfs_trans_handle *trans;
3044         struct btrfs_root *root = BTRFS_I(inode)->root;
3045         int ret;
3046         int ret2;
3047
3048         if (mode & FALLOC_FL_KEEP_SIZE || end <= i_size_read(inode))
3049                 return 0;
3050
3051         trans = btrfs_start_transaction(root, 1);
3052         if (IS_ERR(trans))
3053                 return PTR_ERR(trans);
3054
3055         inode->i_ctime = current_time(inode);
3056         i_size_write(inode, end);
3057         btrfs_inode_safe_disk_i_size_write(inode, 0);
3058         ret = btrfs_update_inode(trans, root, inode);
3059         ret2 = btrfs_end_transaction(trans);
3060
3061         return ret ? ret : ret2;
3062 }
3063
3064 enum {
3065         RANGE_BOUNDARY_WRITTEN_EXTENT,
3066         RANGE_BOUNDARY_PREALLOC_EXTENT,
3067         RANGE_BOUNDARY_HOLE,
3068 };
3069
3070 static int btrfs_zero_range_check_range_boundary(struct btrfs_inode *inode,
3071                                                  u64 offset)
3072 {
3073         const u64 sectorsize = btrfs_inode_sectorsize(inode);
3074         struct extent_map *em;
3075         int ret;
3076
3077         offset = round_down(offset, sectorsize);
3078         em = btrfs_get_extent(inode, NULL, 0, offset, sectorsize);
3079         if (IS_ERR(em))
3080                 return PTR_ERR(em);
3081
3082         if (em->block_start == EXTENT_MAP_HOLE)
3083                 ret = RANGE_BOUNDARY_HOLE;
3084         else if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
3085                 ret = RANGE_BOUNDARY_PREALLOC_EXTENT;
3086         else
3087                 ret = RANGE_BOUNDARY_WRITTEN_EXTENT;
3088
3089         free_extent_map(em);
3090         return ret;
3091 }
3092
3093 static int btrfs_zero_range(struct inode *inode,
3094                             loff_t offset,
3095                             loff_t len,
3096                             const int mode)
3097 {
3098         struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
3099         struct extent_map *em;
3100         struct extent_changeset *data_reserved = NULL;
3101         int ret;
3102         u64 alloc_hint = 0;
3103         const u64 sectorsize = btrfs_inode_sectorsize(BTRFS_I(inode));
3104         u64 alloc_start = round_down(offset, sectorsize);
3105         u64 alloc_end = round_up(offset + len, sectorsize);
3106         u64 bytes_to_reserve = 0;
3107         bool space_reserved = false;
3108
3109         inode_dio_wait(inode);
3110
3111         em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, alloc_start,
3112                               alloc_end - alloc_start);
3113         if (IS_ERR(em)) {
3114                 ret = PTR_ERR(em);
3115                 goto out;
3116         }
3117
3118         /*
3119          * Avoid hole punching and extent allocation for some cases. More cases
3120          * could be considered, but these are unlikely common and we keep things
3121          * as simple as possible for now. Also, intentionally, if the target
3122          * range contains one or more prealloc extents together with regular
3123          * extents and holes, we drop all the existing extents and allocate a
3124          * new prealloc extent, so that we get a larger contiguous disk extent.
3125          */
3126         if (em->start <= alloc_start &&
3127             test_bit(EXTENT_FLAG_PREALLOC, &em->flags)) {
3128                 const u64 em_end = em->start + em->len;
3129
3130                 if (em_end >= offset + len) {
3131                         /*
3132                          * The whole range is already a prealloc extent,
3133                          * do nothing except updating the inode's i_size if
3134                          * needed.
3135                          */
3136                         free_extent_map(em);
3137                         ret = btrfs_fallocate_update_isize(inode, offset + len,
3138                                                            mode);
3139                         goto out;
3140                 }
3141                 /*
3142                  * Part of the range is already a prealloc extent, so operate
3143                  * only on the remaining part of the range.
3144                  */
3145                 alloc_start = em_end;
3146                 ASSERT(IS_ALIGNED(alloc_start, sectorsize));
3147                 len = offset + len - alloc_start;
3148                 offset = alloc_start;
3149                 alloc_hint = em->block_start + em->len;
3150         }
3151         free_extent_map(em);
3152
3153         if (BTRFS_BYTES_TO_BLKS(fs_info, offset) ==
3154             BTRFS_BYTES_TO_BLKS(fs_info, offset + len - 1)) {
3155                 em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, alloc_start,
3156                                       sectorsize);
3157                 if (IS_ERR(em)) {
3158                         ret = PTR_ERR(em);
3159                         goto out;
3160                 }
3161
3162                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags)) {
3163                         free_extent_map(em);
3164                         ret = btrfs_fallocate_update_isize(inode, offset + len,
3165                                                            mode);
3166                         goto out;
3167                 }
3168                 if (len < sectorsize && em->block_start != EXTENT_MAP_HOLE) {
3169                         free_extent_map(em);
3170                         ret = btrfs_truncate_block(inode, offset, len, 0);
3171                         if (!ret)
3172                                 ret = btrfs_fallocate_update_isize(inode,
3173                                                                    offset + len,
3174                                                                    mode);
3175                         return ret;
3176                 }
3177                 free_extent_map(em);
3178                 alloc_start = round_down(offset, sectorsize);
3179                 alloc_end = alloc_start + sectorsize;
3180                 goto reserve_space;
3181         }
3182
3183         alloc_start = round_up(offset, sectorsize);
3184         alloc_end = round_down(offset + len, sectorsize);
3185
3186         /*
3187          * For unaligned ranges, check the pages at the boundaries, they might
3188          * map to an extent, in which case we need to partially zero them, or
3189          * they might map to a hole, in which case we need our allocation range
3190          * to cover them.
3191          */
3192         if (!IS_ALIGNED(offset, sectorsize)) {
3193                 ret = btrfs_zero_range_check_range_boundary(BTRFS_I(inode),
3194                                                             offset);
3195                 if (ret < 0)
3196                         goto out;
3197                 if (ret == RANGE_BOUNDARY_HOLE) {
3198                         alloc_start = round_down(offset, sectorsize);
3199                         ret = 0;
3200                 } else if (ret == RANGE_BOUNDARY_WRITTEN_EXTENT) {
3201                         ret = btrfs_truncate_block(inode, offset, 0, 0);
3202                         if (ret)
3203                                 goto out;
3204                 } else {
3205                         ret = 0;
3206                 }
3207         }
3208
3209         if (!IS_ALIGNED(offset + len, sectorsize)) {
3210                 ret = btrfs_zero_range_check_range_boundary(BTRFS_I(inode),
3211                                                             offset + len);
3212                 if (ret < 0)
3213                         goto out;
3214                 if (ret == RANGE_BOUNDARY_HOLE) {
3215                         alloc_end = round_up(offset + len, sectorsize);
3216                         ret = 0;
3217                 } else if (ret == RANGE_BOUNDARY_WRITTEN_EXTENT) {
3218                         ret = btrfs_truncate_block(inode, offset + len, 0, 1);
3219                         if (ret)
3220                                 goto out;
3221                 } else {
3222                         ret = 0;
3223                 }
3224         }
3225
3226 reserve_space:
3227         if (alloc_start < alloc_end) {
3228                 struct extent_state *cached_state = NULL;
3229                 const u64 lockstart = alloc_start;
3230                 const u64 lockend = alloc_end - 1;
3231
3232                 bytes_to_reserve = alloc_end - alloc_start;
3233                 ret = btrfs_alloc_data_chunk_ondemand(BTRFS_I(inode),
3234                                                       bytes_to_reserve);
3235                 if (ret < 0)
3236                         goto out;
3237                 space_reserved = true;
3238                 ret = btrfs_punch_hole_lock_range(inode, lockstart, lockend,
3239                                                   &cached_state);
3240                 if (ret)
3241                         goto out;
3242                 ret = btrfs_qgroup_reserve_data(BTRFS_I(inode), &data_reserved,
3243                                                 alloc_start, bytes_to_reserve);
3244                 if (ret)
3245                         goto out;
3246                 ret = btrfs_prealloc_file_range(inode, mode, alloc_start,
3247                                                 alloc_end - alloc_start,
3248                                                 i_blocksize(inode),
3249                                                 offset + len, &alloc_hint);
3250                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart,
3251                                      lockend, &cached_state);
3252                 /* btrfs_prealloc_file_range releases reserved space on error */
3253                 if (ret) {
3254                         space_reserved = false;
3255                         goto out;
3256                 }
3257         }
3258         ret = btrfs_fallocate_update_isize(inode, offset + len, mode);
3259  out:
3260         if (ret && space_reserved)
3261                 btrfs_free_reserved_data_space(BTRFS_I(inode), data_reserved,
3262                                                alloc_start, bytes_to_reserve);
3263         extent_changeset_free(data_reserved);
3264
3265         return ret;
3266 }
3267
3268 static long btrfs_fallocate(struct file *file, int mode,
3269                             loff_t offset, loff_t len)
3270 {
3271         struct inode *inode = file_inode(file);
3272         struct extent_state *cached_state = NULL;
3273         struct extent_changeset *data_reserved = NULL;
3274         struct falloc_range *range;
3275         struct falloc_range *tmp;
3276         struct list_head reserve_list;
3277         u64 cur_offset;
3278         u64 last_byte;
3279         u64 alloc_start;
3280         u64 alloc_end;
3281         u64 alloc_hint = 0;
3282         u64 locked_end;
3283         u64 actual_end = 0;
3284         struct extent_map *em;
3285         int blocksize = btrfs_inode_sectorsize(BTRFS_I(inode));
3286         int ret;
3287
3288         alloc_start = round_down(offset, blocksize);
3289         alloc_end = round_up(offset + len, blocksize);
3290         cur_offset = alloc_start;
3291
3292         /* Make sure we aren't being give some crap mode */
3293         if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE |
3294                      FALLOC_FL_ZERO_RANGE))
3295                 return -EOPNOTSUPP;
3296
3297         if (mode & FALLOC_FL_PUNCH_HOLE)
3298                 return btrfs_punch_hole(inode, offset, len);
3299
3300         /*
3301          * Only trigger disk allocation, don't trigger qgroup reserve
3302          *
3303          * For qgroup space, it will be checked later.
3304          */
3305         if (!(mode & FALLOC_FL_ZERO_RANGE)) {
3306                 ret = btrfs_alloc_data_chunk_ondemand(BTRFS_I(inode),
3307                                                       alloc_end - alloc_start);
3308                 if (ret < 0)
3309                         return ret;
3310         }
3311
3312         btrfs_inode_lock(inode, 0);
3313
3314         if (!(mode & FALLOC_FL_KEEP_SIZE) && offset + len > inode->i_size) {
3315                 ret = inode_newsize_ok(inode, offset + len);
3316                 if (ret)
3317                         goto out;
3318         }
3319
3320         /*
3321          * TODO: Move these two operations after we have checked
3322          * accurate reserved space, or fallocate can still fail but
3323          * with page truncated or size expanded.
3324          *
3325          * But that's a minor problem and won't do much harm BTW.
3326          */
3327         if (alloc_start > inode->i_size) {
3328                 ret = btrfs_cont_expand(inode, i_size_read(inode),
3329                                         alloc_start);
3330                 if (ret)
3331                         goto out;
3332         } else if (offset + len > inode->i_size) {
3333                 /*
3334                  * If we are fallocating from the end of the file onward we
3335                  * need to zero out the end of the block if i_size lands in the
3336                  * middle of a block.
3337                  */
3338                 ret = btrfs_truncate_block(inode, inode->i_size, 0, 0);
3339                 if (ret)
3340                         goto out;
3341         }
3342
3343         /*
3344          * wait for ordered IO before we have any locks.  We'll loop again
3345          * below with the locks held.
3346          */
3347         ret = btrfs_wait_ordered_range(inode, alloc_start,
3348                                        alloc_end - alloc_start);
3349         if (ret)
3350                 goto out;
3351
3352         if (mode & FALLOC_FL_ZERO_RANGE) {
3353                 ret = btrfs_zero_range(inode, offset, len, mode);
3354                 inode_unlock(inode);
3355                 return ret;
3356         }
3357
3358         locked_end = alloc_end - 1;
3359         while (1) {
3360                 struct btrfs_ordered_extent *ordered;
3361
3362                 /* the extent lock is ordered inside the running
3363                  * transaction
3364                  */
3365                 lock_extent_bits(&BTRFS_I(inode)->io_tree, alloc_start,
3366                                  locked_end, &cached_state);
3367                 ordered = btrfs_lookup_first_ordered_extent(BTRFS_I(inode),
3368                                                             locked_end);
3369
3370                 if (ordered &&
3371                     ordered->file_offset + ordered->num_bytes > alloc_start &&
3372                     ordered->file_offset < alloc_end) {
3373                         btrfs_put_ordered_extent(ordered);
3374                         unlock_extent_cached(&BTRFS_I(inode)->io_tree,
3375                                              alloc_start, locked_end,
3376                                              &cached_state);
3377                         /*
3378                          * we can't wait on the range with the transaction
3379                          * running or with the extent lock held
3380                          */
3381                         ret = btrfs_wait_ordered_range(inode, alloc_start,
3382                                                        alloc_end - alloc_start);
3383                         if (ret)
3384                                 goto out;
3385                 } else {
3386                         if (ordered)
3387                                 btrfs_put_ordered_extent(ordered);
3388                         break;
3389                 }
3390         }
3391
3392         /* First, check if we exceed the qgroup limit */
3393         INIT_LIST_HEAD(&reserve_list);
3394         while (cur_offset < alloc_end) {
3395                 em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, cur_offset,
3396                                       alloc_end - cur_offset);
3397                 if (IS_ERR(em)) {
3398                         ret = PTR_ERR(em);
3399                         break;
3400                 }
3401                 last_byte = min(extent_map_end(em), alloc_end);
3402                 actual_end = min_t(u64, extent_map_end(em), offset + len);
3403                 last_byte = ALIGN(last_byte, blocksize);
3404                 if (em->block_start == EXTENT_MAP_HOLE ||
3405                     (cur_offset >= inode->i_size &&
3406                      !test_bit(EXTENT_FLAG_PREALLOC, &em->flags))) {
3407                         ret = add_falloc_range(&reserve_list, cur_offset,
3408                                                last_byte - cur_offset);
3409                         if (ret < 0) {
3410                                 free_extent_map(em);
3411                                 break;
3412                         }
3413                         ret = btrfs_qgroup_reserve_data(BTRFS_I(inode),
3414                                         &data_reserved, cur_offset,
3415                                         last_byte - cur_offset);
3416                         if (ret < 0) {
3417                                 cur_offset = last_byte;
3418                                 free_extent_map(em);
3419                                 break;
3420                         }
3421                 } else {
3422                         /*
3423                          * Do not need to reserve unwritten extent for this
3424                          * range, free reserved data space first, otherwise
3425                          * it'll result in false ENOSPC error.
3426                          */
3427                         btrfs_free_reserved_data_space(BTRFS_I(inode),
3428                                 data_reserved, cur_offset,
3429                                 last_byte - cur_offset);
3430                 }
3431                 free_extent_map(em);
3432                 cur_offset = last_byte;
3433         }
3434
3435         /*
3436          * If ret is still 0, means we're OK to fallocate.
3437          * Or just cleanup the list and exit.
3438          */
3439         list_for_each_entry_safe(range, tmp, &reserve_list, list) {
3440                 if (!ret)
3441                         ret = btrfs_prealloc_file_range(inode, mode,
3442                                         range->start,
3443                                         range->len, i_blocksize(inode),
3444                                         offset + len, &alloc_hint);
3445                 else
3446                         btrfs_free_reserved_data_space(BTRFS_I(inode),
3447                                         data_reserved, range->start,
3448                                         range->len);
3449                 list_del(&range->list);
3450                 kfree(range);
3451         }
3452         if (ret < 0)
3453                 goto out_unlock;
3454
3455         /*
3456          * We didn't need to allocate any more space, but we still extended the
3457          * size of the file so we need to update i_size and the inode item.
3458          */
3459         ret = btrfs_fallocate_update_isize(inode, actual_end, mode);
3460 out_unlock:
3461         unlock_extent_cached(&BTRFS_I(inode)->io_tree, alloc_start, locked_end,
3462                              &cached_state);
3463 out:
3464         inode_unlock(inode);
3465         /* Let go of our reservation. */
3466         if (ret != 0 && !(mode & FALLOC_FL_ZERO_RANGE))
3467                 btrfs_free_reserved_data_space(BTRFS_I(inode), data_reserved,
3468                                 cur_offset, alloc_end - cur_offset);
3469         extent_changeset_free(data_reserved);
3470         return ret;
3471 }
3472
3473 static loff_t find_desired_extent(struct inode *inode, loff_t offset,
3474                                   int whence)
3475 {
3476         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
3477         struct extent_map *em = NULL;
3478         struct extent_state *cached_state = NULL;
3479         loff_t i_size = inode->i_size;
3480         u64 lockstart;
3481         u64 lockend;
3482         u64 start;
3483         u64 len;
3484         int ret = 0;
3485
3486         if (i_size == 0 || offset >= i_size)
3487                 return -ENXIO;
3488
3489         /*
3490          * offset can be negative, in this case we start finding DATA/HOLE from
3491          * the very start of the file.
3492          */
3493         start = max_t(loff_t, 0, offset);
3494
3495         lockstart = round_down(start, fs_info->sectorsize);
3496         lockend = round_up(i_size, fs_info->sectorsize);
3497         if (lockend <= lockstart)
3498                 lockend = lockstart + fs_info->sectorsize;
3499         lockend--;
3500         len = lockend - lockstart + 1;
3501
3502         lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend,
3503                          &cached_state);
3504
3505         while (start < i_size) {
3506                 em = btrfs_get_extent_fiemap(BTRFS_I(inode), start, len);
3507                 if (IS_ERR(em)) {
3508                         ret = PTR_ERR(em);
3509                         em = NULL;
3510                         break;
3511                 }
3512
3513                 if (whence == SEEK_HOLE &&
3514                     (em->block_start == EXTENT_MAP_HOLE ||
3515                      test_bit(EXTENT_FLAG_PREALLOC, &em->flags)))
3516                         break;
3517                 else if (whence == SEEK_DATA &&
3518                            (em->block_start != EXTENT_MAP_HOLE &&
3519                             !test_bit(EXTENT_FLAG_PREALLOC, &em->flags)))
3520                         break;
3521
3522                 start = em->start + em->len;
3523                 free_extent_map(em);
3524                 em = NULL;
3525                 cond_resched();
3526         }
3527         free_extent_map(em);
3528         unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
3529                              &cached_state);
3530         if (ret) {
3531                 offset = ret;
3532         } else {
3533                 if (whence == SEEK_DATA && start >= i_size)
3534                         offset = -ENXIO;
3535                 else
3536                         offset = min_t(loff_t, start, i_size);
3537         }
3538
3539         return offset;
3540 }
3541
3542 static loff_t btrfs_file_llseek(struct file *file, loff_t offset, int whence)
3543 {
3544         struct inode *inode = file->f_mapping->host;
3545
3546         switch (whence) {
3547         default:
3548                 return generic_file_llseek(file, offset, whence);
3549         case SEEK_DATA:
3550         case SEEK_HOLE:
3551                 btrfs_inode_lock(inode, BTRFS_ILOCK_SHARED);
3552                 offset = find_desired_extent(inode, offset, whence);
3553                 btrfs_inode_unlock(inode, BTRFS_ILOCK_SHARED);
3554                 break;
3555         }
3556
3557         if (offset < 0)
3558                 return offset;
3559
3560         return vfs_setpos(file, offset, inode->i_sb->s_maxbytes);
3561 }
3562
3563 static int btrfs_file_open(struct inode *inode, struct file *filp)
3564 {
3565         filp->f_mode |= FMODE_NOWAIT | FMODE_BUF_RASYNC;
3566         return generic_file_open(inode, filp);
3567 }
3568
3569 static int check_direct_read(struct btrfs_fs_info *fs_info,
3570                              const struct iov_iter *iter, loff_t offset)
3571 {
3572         int ret;
3573         int i, seg;
3574
3575         ret = check_direct_IO(fs_info, iter, offset);
3576         if (ret < 0)
3577                 return ret;
3578
3579         if (!iter_is_iovec(iter))
3580                 return 0;
3581
3582         for (seg = 0; seg < iter->nr_segs; seg++)
3583                 for (i = seg + 1; i < iter->nr_segs; i++)
3584                         if (iter->iov[seg].iov_base == iter->iov[i].iov_base)
3585                                 return -EINVAL;
3586         return 0;
3587 }
3588
3589 static ssize_t btrfs_direct_read(struct kiocb *iocb, struct iov_iter *to)
3590 {
3591         struct inode *inode = file_inode(iocb->ki_filp);
3592         ssize_t ret;
3593
3594         if (check_direct_read(btrfs_sb(inode->i_sb), to, iocb->ki_pos))
3595                 return 0;
3596
3597         btrfs_inode_lock(inode, BTRFS_ILOCK_SHARED);
3598         ret = iomap_dio_rw(iocb, to, &btrfs_dio_iomap_ops, &btrfs_dio_ops,
3599                            is_sync_kiocb(iocb));
3600         btrfs_inode_unlock(inode, BTRFS_ILOCK_SHARED);
3601         return ret;
3602 }
3603
3604 static ssize_t btrfs_file_read_iter(struct kiocb *iocb, struct iov_iter *to)
3605 {
3606         ssize_t ret = 0;
3607
3608         if (iocb->ki_flags & IOCB_DIRECT) {
3609                 ret = btrfs_direct_read(iocb, to);
3610                 if (ret < 0 || !iov_iter_count(to) ||
3611                     iocb->ki_pos >= i_size_read(file_inode(iocb->ki_filp)))
3612                         return ret;
3613         }
3614
3615         return generic_file_buffered_read(iocb, to, ret);
3616 }
3617
3618 const struct file_operations btrfs_file_operations = {
3619         .llseek         = btrfs_file_llseek,
3620         .read_iter      = btrfs_file_read_iter,
3621         .splice_read    = generic_file_splice_read,
3622         .write_iter     = btrfs_file_write_iter,
3623         .splice_write   = iter_file_splice_write,
3624         .mmap           = btrfs_file_mmap,
3625         .open           = btrfs_file_open,
3626         .release        = btrfs_release_file,
3627         .fsync          = btrfs_sync_file,
3628         .fallocate      = btrfs_fallocate,
3629         .unlocked_ioctl = btrfs_ioctl,
3630 #ifdef CONFIG_COMPAT
3631         .compat_ioctl   = btrfs_compat_ioctl,
3632 #endif
3633         .remap_file_range = btrfs_remap_file_range,
3634 };
3635
3636 void __cold btrfs_auto_defrag_exit(void)
3637 {
3638         kmem_cache_destroy(btrfs_inode_defrag_cachep);
3639 }
3640
3641 int __init btrfs_auto_defrag_init(void)
3642 {
3643         btrfs_inode_defrag_cachep = kmem_cache_create("btrfs_inode_defrag",
3644                                         sizeof(struct inode_defrag), 0,
3645                                         SLAB_MEM_SPREAD,
3646                                         NULL);
3647         if (!btrfs_inode_defrag_cachep)
3648                 return -ENOMEM;
3649
3650         return 0;
3651 }
3652
3653 int btrfs_fdatawrite_range(struct inode *inode, loff_t start, loff_t end)
3654 {
3655         int ret;
3656
3657         /*
3658          * So with compression we will find and lock a dirty page and clear the
3659          * first one as dirty, setup an async extent, and immediately return
3660          * with the entire range locked but with nobody actually marked with
3661          * writeback.  So we can't just filemap_write_and_wait_range() and
3662          * expect it to work since it will just kick off a thread to do the
3663          * actual work.  So we need to call filemap_fdatawrite_range _again_
3664          * since it will wait on the page lock, which won't be unlocked until
3665          * after the pages have been marked as writeback and so we're good to go
3666          * from there.  We have to do this otherwise we'll miss the ordered
3667          * extents and that results in badness.  Please Josef, do not think you
3668          * know better and pull this out at some point in the future, it is
3669          * right and you are wrong.
3670          */
3671         ret = filemap_fdatawrite_range(inode->i_mapping, start, end);
3672         if (!ret && test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
3673                              &BTRFS_I(inode)->runtime_flags))
3674                 ret = filemap_fdatawrite_range(inode->i_mapping, start, end);
3675
3676         return ret;
3677 }