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  * Note: the VFS' inode number of bytes is not updated, it's up to the caller
 682  * to deal with that. We set the field 'bytes_found' of the arguments structure
 683  * with the number of allocated bytes found in the target range, so that the
 684  * caller can update the inode's number of bytes in an atomic way when
 685  * replacing extents in a range to avoid races with stat(2).
 686  */
 687 int btrfs_drop_extents(struct btrfs_trans_handle *trans,
 688                        struct btrfs_root *root, struct btrfs_inode *inode,
 689                        struct btrfs_drop_extents_args *args)
 690 {
 691         struct btrfs_fs_info *fs_info = root->fs_info;
 692         struct extent_buffer *leaf;
 693         struct btrfs_file_extent_item *fi;
 694         struct btrfs_ref ref = { 0 };
 695         struct btrfs_key key;
 696         struct btrfs_key new_key;
 697         u64 ino = btrfs_ino(inode);
 698         u64 search_start = args->start;
 699         u64 disk_bytenr = 0;
 700         u64 num_bytes = 0;
 701         u64 extent_offset = 0;
 702         u64 extent_end = 0;
 703         u64 last_end = args->start;
 704         int del_nr = 0;
 705         int del_slot = 0;
 706         int extent_type;
 707         int recow;
 708         int ret;
 709         int modify_tree = -1;
 710         int update_refs;
 711         int found = 0;
 712         int leafs_visited = 0;
 713         struct btrfs_path *path = args->path;
 714
 715         args->bytes_found = 0;
 716         args->extent_inserted = false;
 717
 718         /* Must always have a path if ->replace_extent is true */
 719         ASSERT(!(args->replace_extent && !args->path));
 720
 721         if (!path) {
 722                 path = btrfs_alloc_path();
 723                 if (!path) {
 724                         ret = -ENOMEM;
 725                         goto out;
 726                 }
 727         }
 728
 729         if (args->drop_cache)
 730                 btrfs_drop_extent_cache(inode, args->start, args->end - 1, 0);
 731
 732         if (args->start >= inode->disk_i_size && !args->replace_extent)
 733                 modify_tree = 0;
 734
 735         update_refs = (test_bit(BTRFS_ROOT_SHAREABLE, &root->state) ||
 736                        root == fs_info->tree_root);
 737         while (1) {
 738                 recow = 0;
 739                 ret = btrfs_lookup_file_extent(trans, root, path, ino,
 740                                                search_start, modify_tree);
 741                 if (ret < 0)
 742                         break;
 743                 if (ret > 0 && path->slots[0] > 0 && search_start == args->start) {
 744                         leaf = path->nodes[0];
 745                         btrfs_item_key_to_cpu(leaf, &key, path->slots[0] - 1);
 746                         if (key.objectid == ino &&
 747                             key.type == BTRFS_EXTENT_DATA_KEY)
 748                                 path->slots[0]--;
 749                 }
 750                 ret = 0;
 751                 leafs_visited++;
 752 next_slot:
 753                 leaf = path->nodes[0];
 754                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
 755                         BUG_ON(del_nr > 0);
 756                         ret = btrfs_next_leaf(root, path);
 757                         if (ret < 0)
 758                                 break;
 759                         if (ret > 0) {
 760                                 ret = 0;
 761                                 break;
 762                         }
 763                         leafs_visited++;
 764                         leaf = path->nodes[0];
 765                         recow = 1;
 766                 }
 767
 768                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 769
 770                 if (key.objectid > ino)
 771                         break;
 772                 if (WARN_ON_ONCE(key.objectid < ino) ||
 773                     key.type < BTRFS_EXTENT_DATA_KEY) {
 774                         ASSERT(del_nr == 0);
 775                         path->slots[0]++;
 776                         goto next_slot;
 777                 }
 778                 if (key.type > BTRFS_EXTENT_DATA_KEY || key.offset >= args->end)
 779                         break;
 780
 781                 fi = btrfs_item_ptr(leaf, path->slots[0],
 782                                     struct btrfs_file_extent_item);
 783                 extent_type = btrfs_file_extent_type(leaf, fi);
 784
 785                 if (extent_type == BTRFS_FILE_EXTENT_REG ||
 786                     extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
 787                         disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
 788                         num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
 789                         extent_offset = btrfs_file_extent_offset(leaf, fi);
 790                         extent_end = key.offset +
 791                                 btrfs_file_extent_num_bytes(leaf, fi);
 792                 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
 793                         extent_end = key.offset +
 794                                 btrfs_file_extent_ram_bytes(leaf, fi);
 795                 } else {
 796                         /* can't happen */
 797                         BUG();
 798                 }
 799
 800                 /*
 801                  * Don't skip extent items representing 0 byte lengths. They
 802                  * used to be created (bug) if while punching holes we hit
 803                  * -ENOSPC condition. So if we find one here, just ensure we
 804                  * delete it, otherwise we would insert a new file extent item
 805                  * with the same key (offset) as that 0 bytes length file
 806                  * extent item in the call to setup_items_for_insert() later
 807                  * in this function.
 808                  */
 809                 if (extent_end == key.offset && extent_end >= search_start) {
 810                         last_end = extent_end;
 811                         goto delete_extent_item;
 812                 }
 813
 814                 if (extent_end <= search_start) {
 815                         path->slots[0]++;
 816                         goto next_slot;
 817                 }
 818
 819                 found = 1;
 820                 search_start = max(key.offset, args->start);
 821                 if (recow || !modify_tree) {
 822                         modify_tree = -1;
 823                         btrfs_release_path(path);
 824                         continue;
 825                 }
 826
 827                 /*
 828                  *     | - range to drop - |
 829                  *  | -------- extent -------- |
 830                  */
 831                 if (args->start > key.offset && args->end < extent_end) {
 832                         BUG_ON(del_nr > 0);
 833                         if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
 834                                 ret = -EOPNOTSUPP;
 835                                 break;
 836                         }
 837
 838                         memcpy(&new_key, &key, sizeof(new_key));
 839                         new_key.offset = args->start;
 840                         ret = btrfs_duplicate_item(trans, root, path,
 841                                                    &new_key);
 842                         if (ret == -EAGAIN) {
 843                                 btrfs_release_path(path);
 844                                 continue;
 845                         }
 846                         if (ret < 0)
 847                                 break;
 848
 849                         leaf = path->nodes[0];
 850                         fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
 851                                             struct btrfs_file_extent_item);
 852                         btrfs_set_file_extent_num_bytes(leaf, fi,
 853                                                         args->start - key.offset);
 854
 855                         fi = btrfs_item_ptr(leaf, path->slots[0],
 856                                             struct btrfs_file_extent_item);
 857
 858                         extent_offset += args->start - key.offset;
 859                         btrfs_set_file_extent_offset(leaf, fi, extent_offset);
 860                         btrfs_set_file_extent_num_bytes(leaf, fi,
 861                                                         extent_end - args->start);
 862                         btrfs_mark_buffer_dirty(leaf);
 863
 864                         if (update_refs && disk_bytenr > 0) {
 865                                 btrfs_init_generic_ref(&ref,
 866                                                 BTRFS_ADD_DELAYED_REF,
 867                                                 disk_bytenr, num_bytes, 0);
 868                                 btrfs_init_data_ref(&ref,
 869                                                 root->root_key.objectid,
 870                                                 new_key.objectid,
 871                                                 args->start - extent_offset);
 872                                 ret = btrfs_inc_extent_ref(trans, &ref);
 873                                 BUG_ON(ret); /* -ENOMEM */
 874                         }
 875                         key.offset = args->start;
 876                 }
 877                 /*
 878                  * From here on out we will have actually dropped something, so
 879                  * last_end can be updated.
 880                  */
 881                 last_end = extent_end;
 882
 883                 /*
 884                  *  | ---- range to drop ----- |
 885                  *      | -------- extent -------- |
 886                  */
 887                 if (args->start <= key.offset && args->end < extent_end) {
 888                         if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
 889                                 ret = -EOPNOTSUPP;
 890                                 break;
 891                         }
 892
 893                         memcpy(&new_key, &key, sizeof(new_key));
 894                         new_key.offset = args->end;
 895                         btrfs_set_item_key_safe(fs_info, path, &new_key);
 896
 897                         extent_offset += args->end - key.offset;
 898                         btrfs_set_file_extent_offset(leaf, fi, extent_offset);
 899                         btrfs_set_file_extent_num_bytes(leaf, fi,
 900                                                         extent_end - args->end);
 901                         btrfs_mark_buffer_dirty(leaf);
 902                         if (update_refs && disk_bytenr > 0)
 903                                 args->bytes_found += args->end - key.offset;
 904                         break;
 905                 }
 906
 907                 search_start = extent_end;
 908                 /*
 909                  *       | ---- range to drop ----- |
 910                  *  | -------- extent -------- |
 911                  */
 912                 if (args->start > key.offset && args->end >= extent_end) {
 913                         BUG_ON(del_nr > 0);
 914                         if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
 915                                 ret = -EOPNOTSUPP;
 916                                 break;
 917                         }
 918
 919                         btrfs_set_file_extent_num_bytes(leaf, fi,
 920                                                         args->start - key.offset);
 921                         btrfs_mark_buffer_dirty(leaf);
 922                         if (update_refs && disk_bytenr > 0)
 923                                 args->bytes_found += extent_end - args->start;
 924                         if (args->end == extent_end)
 925                                 break;
 926
 927                         path->slots[0]++;
 928                         goto next_slot;
 929                 }
 930
 931                 /*
 932                  *  | ---- range to drop ----- |
 933                  *    | ------ extent ------ |
 934                  */
 935                 if (args->start <= key.offset && args->end >= extent_end) {
 936 delete_extent_item:
 937                         if (del_nr == 0) {
 938                                 del_slot = path->slots[0];
 939                                 del_nr = 1;
 940                         } else {
 941                                 BUG_ON(del_slot + del_nr != path->slots[0]);
 942                                 del_nr++;
 943                         }
 944
 945                         if (update_refs &&
 946                             extent_type == BTRFS_FILE_EXTENT_INLINE) {
 947                                 args->bytes_found += extent_end - key.offset;
 948                                 extent_end = ALIGN(extent_end,
 949                                                    fs_info->sectorsize);
 950                         } else if (update_refs && disk_bytenr > 0) {
 951                                 btrfs_init_generic_ref(&ref,
 952                                                 BTRFS_DROP_DELAYED_REF,
 953                                                 disk_bytenr, num_bytes, 0);
 954                                 btrfs_init_data_ref(&ref,
 955                                                 root->root_key.objectid,
 956                                                 key.objectid,
 957                                                 key.offset - extent_offset);
 958                                 ret = btrfs_free_extent(trans, &ref);
 959                                 BUG_ON(ret); /* -ENOMEM */
 960                                 args->bytes_found += extent_end - key.offset;
 961                         }
 962
 963                         if (args->end == extent_end)
 964                                 break;
 965
 966                         if (path->slots[0] + 1 < btrfs_header_nritems(leaf)) {
 967                                 path->slots[0]++;
 968                                 goto next_slot;
 969                         }
 970
 971                         ret = btrfs_del_items(trans, root, path, del_slot,
 972                                               del_nr);
 973                         if (ret) {
 974                                 btrfs_abort_transaction(trans, ret);
 975                                 break;
 976                         }
 977
 978                         del_nr = 0;
 979                         del_slot = 0;
 980
 981                         btrfs_release_path(path);
 982                         continue;
 983                 }
 984
 985                 BUG();
 986         }
 987
 988         if (!ret && del_nr > 0) {
 989                 /*
 990                  * Set path->slots[0] to first slot, so that after the delete
 991                  * if items are move off from our leaf to its immediate left or
 992                  * right neighbor leafs, we end up with a correct and adjusted
 993                  * path->slots[0] for our insertion (if args->replace_extent).
 994                  */
 995                 path->slots[0] = del_slot;
 996                 ret = btrfs_del_items(trans, root, path, del_slot, del_nr);
 997                 if (ret)
 998                         btrfs_abort_transaction(trans, ret);
 999         }
1000
1001         leaf = path->nodes[0];
1002         /*
1003          * If btrfs_del_items() was called, it might have deleted a leaf, in
1004          * which case it unlocked our path, so check path->locks[0] matches a
1005          * write lock.
1006          */
1007         if (!ret && args->replace_extent && leafs_visited == 1 &&
1008             path->locks[0] == BTRFS_WRITE_LOCK &&
1009             btrfs_leaf_free_space(leaf) >=
1010             sizeof(struct btrfs_item) + args->extent_item_size) {
1011
1012                 key.objectid = ino;
1013                 key.type = BTRFS_EXTENT_DATA_KEY;
1014                 key.offset = args->start;
1015                 if (!del_nr && path->slots[0] < btrfs_header_nritems(leaf)) {
1016                         struct btrfs_key slot_key;
1017
1018                         btrfs_item_key_to_cpu(leaf, &slot_key, path->slots[0]);
1019                         if (btrfs_comp_cpu_keys(&key, &slot_key) > 0)
1020                                 path->slots[0]++;
1021                 }
1022                 setup_items_for_insert(root, path, &key,
1023                                        &args->extent_item_size, 1);
1024                 args->extent_inserted = true;
1025         }
1026
1027         if (!args->path)
1028                 btrfs_free_path(path);
1029         else if (!args->extent_inserted)
1030                 btrfs_release_path(path);
1031 out:
1032         args->drop_end = found ? min(args->end, last_end) : args->end;
1033
1034         return ret;
1035 }
1036
1037 static int extent_mergeable(struct extent_buffer *leaf, int slot,
1038                             u64 objectid, u64 bytenr, u64 orig_offset,
1039                             u64 *start, u64 *end)
1040 {
1041         struct btrfs_file_extent_item *fi;
1042         struct btrfs_key key;
1043         u64 extent_end;
1044
1045         if (slot < 0 || slot >= btrfs_header_nritems(leaf))
1046                 return 0;
1047
1048         btrfs_item_key_to_cpu(leaf, &key, slot);
1049         if (key.objectid != objectid || key.type != BTRFS_EXTENT_DATA_KEY)
1050                 return 0;
1051
1052         fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
1053         if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG ||
1054             btrfs_file_extent_disk_bytenr(leaf, fi) != bytenr ||
1055             btrfs_file_extent_offset(leaf, fi) != key.offset - orig_offset ||
1056             btrfs_file_extent_compression(leaf, fi) ||
1057             btrfs_file_extent_encryption(leaf, fi) ||
1058             btrfs_file_extent_other_encoding(leaf, fi))
1059                 return 0;
1060
1061         extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
1062         if ((*start && *start != key.offset) || (*end && *end != extent_end))
1063                 return 0;
1064
1065         *start = key.offset;
1066         *end = extent_end;
1067         return 1;
1068 }
1069
1070 /*
1071  * Mark extent in the range start - end as written.
1072  *
1073  * This changes extent type from 'pre-allocated' to 'regular'. If only
1074  * part of extent is marked as written, the extent will be split into
1075  * two or three.
1076  */
1077 int btrfs_mark_extent_written(struct btrfs_trans_handle *trans,
1078                               struct btrfs_inode *inode, u64 start, u64 end)
1079 {
1080         struct btrfs_fs_info *fs_info = trans->fs_info;
1081         struct btrfs_root *root = inode->root;
1082         struct extent_buffer *leaf;
1083         struct btrfs_path *path;
1084         struct btrfs_file_extent_item *fi;
1085         struct btrfs_ref ref = { 0 };
1086         struct btrfs_key key;
1087         struct btrfs_key new_key;
1088         u64 bytenr;
1089         u64 num_bytes;
1090         u64 extent_end;
1091         u64 orig_offset;
1092         u64 other_start;
1093         u64 other_end;
1094         u64 split;
1095         int del_nr = 0;
1096         int del_slot = 0;
1097         int recow;
1098         int ret;
1099         u64 ino = btrfs_ino(inode);
1100
1101         path = btrfs_alloc_path();
1102         if (!path)
1103                 return -ENOMEM;
1104 again:
1105         recow = 0;
1106         split = start;
1107         key.objectid = ino;
1108         key.type = BTRFS_EXTENT_DATA_KEY;
1109         key.offset = split;
1110
1111         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1112         if (ret < 0)
1113                 goto out;
1114         if (ret > 0 && path->slots[0] > 0)
1115                 path->slots[0]--;
1116
1117         leaf = path->nodes[0];
1118         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1119         if (key.objectid != ino ||
1120             key.type != BTRFS_EXTENT_DATA_KEY) {
1121                 ret = -EINVAL;
1122                 btrfs_abort_transaction(trans, ret);
1123                 goto out;
1124         }
1125         fi = btrfs_item_ptr(leaf, path->slots[0],
1126                             struct btrfs_file_extent_item);
1127         if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_PREALLOC) {
1128                 ret = -EINVAL;
1129                 btrfs_abort_transaction(trans, ret);
1130                 goto out;
1131         }
1132         extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
1133         if (key.offset > start || extent_end < end) {
1134                 ret = -EINVAL;
1135                 btrfs_abort_transaction(trans, ret);
1136                 goto out;
1137         }
1138
1139         bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1140         num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
1141         orig_offset = key.offset - btrfs_file_extent_offset(leaf, fi);
1142         memcpy(&new_key, &key, sizeof(new_key));
1143
1144         if (start == key.offset && end < extent_end) {
1145                 other_start = 0;
1146                 other_end = start;
1147                 if (extent_mergeable(leaf, path->slots[0] - 1,
1148                                      ino, bytenr, orig_offset,
1149                                      &other_start, &other_end)) {
1150                         new_key.offset = end;
1151                         btrfs_set_item_key_safe(fs_info, path, &new_key);
1152                         fi = btrfs_item_ptr(leaf, path->slots[0],
1153                                             struct btrfs_file_extent_item);
1154                         btrfs_set_file_extent_generation(leaf, fi,
1155                                                          trans->transid);
1156                         btrfs_set_file_extent_num_bytes(leaf, fi,
1157                                                         extent_end - end);
1158                         btrfs_set_file_extent_offset(leaf, fi,
1159                                                      end - orig_offset);
1160                         fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
1161                                             struct btrfs_file_extent_item);
1162                         btrfs_set_file_extent_generation(leaf, fi,
1163                                                          trans->transid);
1164                         btrfs_set_file_extent_num_bytes(leaf, fi,
1165                                                         end - other_start);
1166                         btrfs_mark_buffer_dirty(leaf);
1167                         goto out;
1168                 }
1169         }
1170
1171         if (start > key.offset && end == extent_end) {
1172                 other_start = end;
1173                 other_end = 0;
1174                 if (extent_mergeable(leaf, path->slots[0] + 1,
1175                                      ino, bytenr, orig_offset,
1176                                      &other_start, &other_end)) {
1177                         fi = btrfs_item_ptr(leaf, path->slots[0],
1178                                             struct btrfs_file_extent_item);
1179                         btrfs_set_file_extent_num_bytes(leaf, fi,
1180                                                         start - key.offset);
1181                         btrfs_set_file_extent_generation(leaf, fi,
1182                                                          trans->transid);
1183                         path->slots[0]++;
1184                         new_key.offset = start;
1185                         btrfs_set_item_key_safe(fs_info, path, &new_key);
1186
1187                         fi = btrfs_item_ptr(leaf, path->slots[0],
1188                                             struct btrfs_file_extent_item);
1189                         btrfs_set_file_extent_generation(leaf, fi,
1190                                                          trans->transid);
1191                         btrfs_set_file_extent_num_bytes(leaf, fi,
1192                                                         other_end - start);
1193                         btrfs_set_file_extent_offset(leaf, fi,
1194                                                      start - orig_offset);
1195                         btrfs_mark_buffer_dirty(leaf);
1196                         goto out;
1197                 }
1198         }
1199
1200         while (start > key.offset || end < extent_end) {
1201                 if (key.offset == start)
1202                         split = end;
1203
1204                 new_key.offset = split;
1205                 ret = btrfs_duplicate_item(trans, root, path, &new_key);
1206                 if (ret == -EAGAIN) {
1207                         btrfs_release_path(path);
1208                         goto again;
1209                 }
1210                 if (ret < 0) {
1211                         btrfs_abort_transaction(trans, ret);
1212                         goto out;
1213                 }
1214
1215                 leaf = path->nodes[0];
1216                 fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
1217                                     struct btrfs_file_extent_item);
1218                 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1219                 btrfs_set_file_extent_num_bytes(leaf, fi,
1220                                                 split - key.offset);
1221
1222                 fi = btrfs_item_ptr(leaf, path->slots[0],
1223                                     struct btrfs_file_extent_item);
1224
1225                 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1226                 btrfs_set_file_extent_offset(leaf, fi, split - orig_offset);
1227                 btrfs_set_file_extent_num_bytes(leaf, fi,
1228                                                 extent_end - split);
1229                 btrfs_mark_buffer_dirty(leaf);
1230
1231                 btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF, bytenr,
1232                                        num_bytes, 0);
1233                 btrfs_init_data_ref(&ref, root->root_key.objectid, ino,
1234                                     orig_offset);
1235                 ret = btrfs_inc_extent_ref(trans, &ref);
1236                 if (ret) {
1237                         btrfs_abort_transaction(trans, ret);
1238                         goto out;
1239                 }
1240
1241                 if (split == start) {
1242                         key.offset = start;
1243                 } else {
1244                         if (start != key.offset) {
1245                                 ret = -EINVAL;
1246                                 btrfs_abort_transaction(trans, ret);
1247                                 goto out;
1248                         }
1249                         path->slots[0]--;
1250                         extent_end = end;
1251                 }
1252                 recow = 1;
1253         }
1254
1255         other_start = end;
1256         other_end = 0;
1257         btrfs_init_generic_ref(&ref, BTRFS_DROP_DELAYED_REF, bytenr,
1258                                num_bytes, 0);
1259         btrfs_init_data_ref(&ref, root->root_key.objectid, ino, orig_offset);
1260         if (extent_mergeable(leaf, path->slots[0] + 1,
1261                              ino, bytenr, orig_offset,
1262                              &other_start, &other_end)) {
1263                 if (recow) {
1264                         btrfs_release_path(path);
1265                         goto again;
1266                 }
1267                 extent_end = other_end;
1268                 del_slot = path->slots[0] + 1;
1269                 del_nr++;
1270                 ret = btrfs_free_extent(trans, &ref);
1271                 if (ret) {
1272                         btrfs_abort_transaction(trans, ret);
1273                         goto out;
1274                 }
1275         }
1276         other_start = 0;
1277         other_end = start;
1278         if (extent_mergeable(leaf, path->slots[0] - 1,
1279                              ino, bytenr, orig_offset,
1280                              &other_start, &other_end)) {
1281                 if (recow) {
1282                         btrfs_release_path(path);
1283                         goto again;
1284                 }
1285                 key.offset = other_start;
1286                 del_slot = path->slots[0];
1287                 del_nr++;
1288                 ret = btrfs_free_extent(trans, &ref);
1289                 if (ret) {
1290                         btrfs_abort_transaction(trans, ret);
1291                         goto out;
1292                 }
1293         }
1294         if (del_nr == 0) {
1295                 fi = btrfs_item_ptr(leaf, path->slots[0],
1296                            struct btrfs_file_extent_item);
1297                 btrfs_set_file_extent_type(leaf, fi,
1298                                            BTRFS_FILE_EXTENT_REG);
1299                 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1300                 btrfs_mark_buffer_dirty(leaf);
1301         } else {
1302                 fi = btrfs_item_ptr(leaf, del_slot - 1,
1303                            struct btrfs_file_extent_item);
1304                 btrfs_set_file_extent_type(leaf, fi,
1305                                            BTRFS_FILE_EXTENT_REG);
1306                 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1307                 btrfs_set_file_extent_num_bytes(leaf, fi,
1308                                                 extent_end - key.offset);
1309                 btrfs_mark_buffer_dirty(leaf);
1310
1311                 ret = btrfs_del_items(trans, root, path, del_slot, del_nr);
1312                 if (ret < 0) {
1313                         btrfs_abort_transaction(trans, ret);
1314                         goto out;
1315                 }
1316         }
1317 out:
1318         btrfs_free_path(path);
1319         return 0;
1320 }
1321
1322 /*
1323  * on error we return an unlocked page and the error value
1324  * on success we return a locked page and 0
1325  */
1326 static int prepare_uptodate_page(struct inode *inode,
1327                                  struct page *page, u64 pos,
1328                                  bool force_uptodate)
1329 {
1330         int ret = 0;
1331
1332         if (((pos & (PAGE_SIZE - 1)) || force_uptodate) &&
1333             !PageUptodate(page)) {
1334                 ret = btrfs_readpage(NULL, page);
1335                 if (ret)
1336                         return ret;
1337                 lock_page(page);
1338                 if (!PageUptodate(page)) {
1339                         unlock_page(page);
1340                         return -EIO;
1341                 }
1342                 if (page->mapping != inode->i_mapping) {
1343                         unlock_page(page);
1344                         return -EAGAIN;
1345                 }
1346         }
1347         return 0;
1348 }
1349
1350 /*
1351  * this just gets pages into the page cache and locks them down.
1352  */
1353 static noinline int prepare_pages(struct inode *inode, struct page **pages,
1354                                   size_t num_pages, loff_t pos,
1355                                   size_t write_bytes, bool force_uptodate)
1356 {
1357         int i;
1358         unsigned long index = pos >> PAGE_SHIFT;
1359         gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping);
1360         int err = 0;
1361         int faili;
1362
1363         for (i = 0; i < num_pages; i++) {
1364 again:
1365                 pages[i] = find_or_create_page(inode->i_mapping, index + i,
1366                                                mask | __GFP_WRITE);
1367                 if (!pages[i]) {
1368                         faili = i - 1;
1369                         err = -ENOMEM;
1370                         goto fail;
1371                 }
1372
1373                 if (i == 0)
1374                         err = prepare_uptodate_page(inode, pages[i], pos,
1375                                                     force_uptodate);
1376                 if (!err && i == num_pages - 1)
1377                         err = prepare_uptodate_page(inode, pages[i],
1378                                                     pos + write_bytes, false);
1379                 if (err) {
1380                         put_page(pages[i]);
1381                         if (err == -EAGAIN) {
1382                                 err = 0;
1383                                 goto again;
1384                         }
1385                         faili = i - 1;
1386                         goto fail;
1387                 }
1388                 wait_on_page_writeback(pages[i]);
1389         }
1390
1391         return 0;
1392 fail:
1393         while (faili >= 0) {
1394                 unlock_page(pages[faili]);
1395                 put_page(pages[faili]);
1396                 faili--;
1397         }
1398         return err;
1399
1400 }
1401
1402 /*
1403  * This function locks the extent and properly waits for data=ordered extents
1404  * to finish before allowing the pages to be modified if need.
1405  *
1406  * The return value:
1407  * 1 - the extent is locked
1408  * 0 - the extent is not locked, and everything is OK
1409  * -EAGAIN - need re-prepare the pages
1410  * the other < 0 number - Something wrong happens
1411  */
1412 static noinline int
1413 lock_and_cleanup_extent_if_need(struct btrfs_inode *inode, struct page **pages,
1414                                 size_t num_pages, loff_t pos,
1415                                 size_t write_bytes,
1416                                 u64 *lockstart, u64 *lockend,
1417                                 struct extent_state **cached_state)
1418 {
1419         struct btrfs_fs_info *fs_info = inode->root->fs_info;
1420         u64 start_pos;
1421         u64 last_pos;
1422         int i;
1423         int ret = 0;
1424
1425         start_pos = round_down(pos, fs_info->sectorsize);
1426         last_pos = round_up(pos + write_bytes, fs_info->sectorsize) - 1;
1427
1428         if (start_pos < inode->vfs_inode.i_size) {
1429                 struct btrfs_ordered_extent *ordered;
1430
1431                 lock_extent_bits(&inode->io_tree, start_pos, last_pos,
1432                                 cached_state);
1433                 ordered = btrfs_lookup_ordered_range(inode, start_pos,
1434                                                      last_pos - start_pos + 1);
1435                 if (ordered &&
1436                     ordered->file_offset + ordered->num_bytes > start_pos &&
1437                     ordered->file_offset <= last_pos) {
1438                         unlock_extent_cached(&inode->io_tree, start_pos,
1439                                         last_pos, cached_state);
1440                         for (i = 0; i < num_pages; i++) {
1441                                 unlock_page(pages[i]);
1442                                 put_page(pages[i]);
1443                         }
1444                         btrfs_start_ordered_extent(ordered, 1);
1445                         btrfs_put_ordered_extent(ordered);
1446                         return -EAGAIN;
1447                 }
1448                 if (ordered)
1449                         btrfs_put_ordered_extent(ordered);
1450
1451                 *lockstart = start_pos;
1452                 *lockend = last_pos;
1453                 ret = 1;
1454         }
1455
1456         /*
1457          * It's possible the pages are dirty right now, but we don't want
1458          * to clean them yet because copy_from_user may catch a page fault
1459          * and we might have to fall back to one page at a time.  If that
1460          * happens, we'll unlock these pages and we'd have a window where
1461          * reclaim could sneak in and drop the once-dirty page on the floor
1462          * without writing it.
1463          *
1464          * We have the pages locked and the extent range locked, so there's
1465          * no way someone can start IO on any dirty pages in this range.
1466          *
1467          * We'll call btrfs_dirty_pages() later on, and that will flip around
1468          * delalloc bits and dirty the pages as required.
1469          */
1470         for (i = 0; i < num_pages; i++) {
1471                 set_page_extent_mapped(pages[i]);
1472                 WARN_ON(!PageLocked(pages[i]));
1473         }
1474
1475         return ret;
1476 }
1477
1478 static int check_can_nocow(struct btrfs_inode *inode, loff_t pos,
1479                            size_t *write_bytes, bool nowait)
1480 {
1481         struct btrfs_fs_info *fs_info = inode->root->fs_info;
1482         struct btrfs_root *root = inode->root;
1483         u64 lockstart, lockend;
1484         u64 num_bytes;
1485         int ret;
1486
1487         if (!(inode->flags & (BTRFS_INODE_NODATACOW | BTRFS_INODE_PREALLOC)))
1488                 return 0;
1489
1490         if (!nowait && !btrfs_drew_try_write_lock(&root->snapshot_lock))
1491                 return -EAGAIN;
1492
1493         lockstart = round_down(pos, fs_info->sectorsize);
1494         lockend = round_up(pos + *write_bytes,
1495                            fs_info->sectorsize) - 1;
1496         num_bytes = lockend - lockstart + 1;
1497
1498         if (nowait) {
1499                 struct btrfs_ordered_extent *ordered;
1500
1501                 if (!try_lock_extent(&inode->io_tree, lockstart, lockend))
1502                         return -EAGAIN;
1503
1504                 ordered = btrfs_lookup_ordered_range(inode, lockstart,
1505                                                      num_bytes);
1506                 if (ordered) {
1507                         btrfs_put_ordered_extent(ordered);
1508                         ret = -EAGAIN;
1509                         goto out_unlock;
1510                 }
1511         } else {
1512                 btrfs_lock_and_flush_ordered_range(inode, lockstart,
1513                                                    lockend, NULL);
1514         }
1515
1516         ret = can_nocow_extent(&inode->vfs_inode, lockstart, &num_bytes,
1517                         NULL, NULL, NULL, false);
1518         if (ret <= 0) {
1519                 ret = 0;
1520                 if (!nowait)
1521                         btrfs_drew_write_unlock(&root->snapshot_lock);
1522         } else {
1523                 *write_bytes = min_t(size_t, *write_bytes ,
1524                                      num_bytes - pos + lockstart);
1525         }
1526 out_unlock:
1527         unlock_extent(&inode->io_tree, lockstart, lockend);
1528
1529         return ret;
1530 }
1531
1532 static int check_nocow_nolock(struct btrfs_inode *inode, loff_t pos,
1533                               size_t *write_bytes)
1534 {
1535         return check_can_nocow(inode, pos, write_bytes, true);
1536 }
1537
1538 /*
1539  * Check if we can do nocow write into the range [@pos, @pos + @write_bytes)
1540  *
1541  * @pos:         File offset
1542  * @write_bytes: The length to write, will be updated to the nocow writeable
1543  *               range
1544  *
1545  * This function will flush ordered extents in the range to ensure proper
1546  * nocow checks.
1547  *
1548  * Return:
1549  * >0           and update @write_bytes if we can do nocow write
1550  *  0           if we can't do nocow write
1551  * -EAGAIN      if we can't get the needed lock or there are ordered extents
1552  *              for * (nowait == true) case
1553  * <0           if other error happened
1554  *
1555  * NOTE: Callers need to release the lock by btrfs_check_nocow_unlock().
1556  */
1557 int btrfs_check_nocow_lock(struct btrfs_inode *inode, loff_t pos,
1558                            size_t *write_bytes)
1559 {
1560         return check_can_nocow(inode, pos, write_bytes, false);
1561 }
1562
1563 void btrfs_check_nocow_unlock(struct btrfs_inode *inode)
1564 {
1565         btrfs_drew_write_unlock(&inode->root->snapshot_lock);
1566 }
1567
1568 static void update_time_for_write(struct inode *inode)
1569 {
1570         struct timespec64 now;
1571
1572         if (IS_NOCMTIME(inode))
1573                 return;
1574
1575         now = current_time(inode);
1576         if (!timespec64_equal(&inode->i_mtime, &now))
1577                 inode->i_mtime = now;
1578
1579         if (!timespec64_equal(&inode->i_ctime, &now))
1580                 inode->i_ctime = now;
1581
1582         if (IS_I_VERSION(inode))
1583                 inode_inc_iversion(inode);
1584 }
1585
1586 static int btrfs_write_check(struct kiocb *iocb, struct iov_iter *from,
1587                              size_t count)
1588 {
1589         struct file *file = iocb->ki_filp;
1590         struct inode *inode = file_inode(file);
1591         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1592         loff_t pos = iocb->ki_pos;
1593         int ret;
1594         loff_t oldsize;
1595         loff_t start_pos;
1596
1597         if (iocb->ki_flags & IOCB_NOWAIT) {
1598                 size_t nocow_bytes = count;
1599
1600                 /* We will allocate space in case nodatacow is not set, so bail */
1601                 if (check_nocow_nolock(BTRFS_I(inode), pos, &nocow_bytes) <= 0)
1602                         return -EAGAIN;
1603                 /*
1604                  * There are holes in the range or parts of the range that must
1605                  * be COWed (shared extents, RO block groups, etc), so just bail
1606                  * out.
1607                  */
1608                 if (nocow_bytes < count)
1609                         return -EAGAIN;
1610         }
1611
1612         current->backing_dev_info = inode_to_bdi(inode);
1613         ret = file_remove_privs(file);
1614         if (ret)
1615                 return ret;
1616
1617         /*
1618          * We reserve space for updating the inode when we reserve space for the
1619          * extent we are going to write, so we will enospc out there.  We don't
1620          * need to start yet another transaction to update the inode as we will
1621          * update the inode when we finish writing whatever data we write.
1622          */
1623         update_time_for_write(inode);
1624
1625         start_pos = round_down(pos, fs_info->sectorsize);
1626         oldsize = i_size_read(inode);
1627         if (start_pos > oldsize) {
1628                 /* Expand hole size to cover write data, preventing empty gap */
1629                 loff_t end_pos = round_up(pos + count, fs_info->sectorsize);
1630
1631                 ret = btrfs_cont_expand(BTRFS_I(inode), oldsize, end_pos);
1632                 if (ret) {
1633                         current->backing_dev_info = NULL;
1634                         return ret;
1635                 }
1636         }
1637
1638         return 0;
1639 }
1640
1641 static noinline ssize_t btrfs_buffered_write(struct kiocb *iocb,
1642                                                struct iov_iter *i)
1643 {
1644         struct file *file = iocb->ki_filp;
1645         loff_t pos;
1646         struct inode *inode = file_inode(file);
1647         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1648         struct page **pages = NULL;
1649         struct extent_changeset *data_reserved = NULL;
1650         u64 release_bytes = 0;
1651         u64 lockstart;
1652         u64 lockend;
1653         size_t num_written = 0;
1654         int nrptrs;
1655         ssize_t ret;
1656         bool only_release_metadata = false;
1657         bool force_page_uptodate = false;
1658         loff_t old_isize = i_size_read(inode);
1659         unsigned int ilock_flags = 0;
1660
1661         if (iocb->ki_flags & IOCB_NOWAIT)
1662                 ilock_flags |= BTRFS_ILOCK_TRY;
1663
1664         ret = btrfs_inode_lock(inode, ilock_flags);
1665         if (ret < 0)
1666                 return ret;
1667
1668         ret = generic_write_checks(iocb, i);
1669         if (ret <= 0)
1670                 goto out;
1671
1672         ret = btrfs_write_check(iocb, i, ret);
1673         if (ret < 0)
1674                 goto out;
1675
1676         pos = iocb->ki_pos;
1677         nrptrs = min(DIV_ROUND_UP(iov_iter_count(i), PAGE_SIZE),
1678                         PAGE_SIZE / (sizeof(struct page *)));
1679         nrptrs = min(nrptrs, current->nr_dirtied_pause - current->nr_dirtied);
1680         nrptrs = max(nrptrs, 8);
1681         pages = kmalloc_array(nrptrs, sizeof(struct page *), GFP_KERNEL);
1682         if (!pages) {
1683                 ret = -ENOMEM;
1684                 goto out;
1685         }
1686
1687         while (iov_iter_count(i) > 0) {
1688                 struct extent_state *cached_state = NULL;
1689                 size_t offset = offset_in_page(pos);
1690                 size_t sector_offset;
1691                 size_t write_bytes = min(iov_iter_count(i),
1692                                          nrptrs * (size_t)PAGE_SIZE -
1693                                          offset);
1694                 size_t num_pages;
1695                 size_t reserve_bytes;
1696                 size_t dirty_pages;
1697                 size_t copied;
1698                 size_t dirty_sectors;
1699                 size_t num_sectors;
1700                 int extents_locked;
1701
1702                 /*
1703                  * Fault pages before locking them in prepare_pages
1704                  * to avoid recursive lock
1705                  */
1706                 if (unlikely(iov_iter_fault_in_readable(i, write_bytes))) {
1707                         ret = -EFAULT;
1708                         break;
1709                 }
1710
1711                 only_release_metadata = false;
1712                 sector_offset = pos & (fs_info->sectorsize - 1);
1713
1714                 extent_changeset_release(data_reserved);
1715                 ret = btrfs_check_data_free_space(BTRFS_I(inode),
1716                                                   &data_reserved, pos,
1717                                                   write_bytes);
1718                 if (ret < 0) {
1719                         /*
1720                          * If we don't have to COW at the offset, reserve
1721                          * metadata only. write_bytes may get smaller than
1722                          * requested here.
1723                          */
1724                         if (btrfs_check_nocow_lock(BTRFS_I(inode), pos,
1725                                                    &write_bytes) > 0)
1726                                 only_release_metadata = true;
1727                         else
1728                                 break;
1729                 }
1730
1731                 num_pages = DIV_ROUND_UP(write_bytes + offset, PAGE_SIZE);
1732                 WARN_ON(num_pages > nrptrs);
1733                 reserve_bytes = round_up(write_bytes + sector_offset,
1734                                          fs_info->sectorsize);
1735                 WARN_ON(reserve_bytes == 0);
1736                 ret = btrfs_delalloc_reserve_metadata(BTRFS_I(inode),
1737                                 reserve_bytes);
1738                 if (ret) {
1739                         if (!only_release_metadata)
1740                                 btrfs_free_reserved_data_space(BTRFS_I(inode),
1741                                                 data_reserved, pos,
1742                                                 write_bytes);
1743                         else
1744                                 btrfs_check_nocow_unlock(BTRFS_I(inode));
1745                         break;
1746                 }
1747
1748                 release_bytes = reserve_bytes;
1749 again:
1750                 /*
1751                  * This is going to setup the pages array with the number of
1752                  * pages we want, so we don't really need to worry about the
1753                  * contents of pages from loop to loop
1754                  */
1755                 ret = prepare_pages(inode, pages, num_pages,
1756                                     pos, write_bytes,
1757                                     force_page_uptodate);
1758                 if (ret) {
1759                         btrfs_delalloc_release_extents(BTRFS_I(inode),
1760                                                        reserve_bytes);
1761                         break;
1762                 }
1763
1764                 extents_locked = lock_and_cleanup_extent_if_need(
1765                                 BTRFS_I(inode), pages,
1766                                 num_pages, pos, write_bytes, &lockstart,
1767                                 &lockend, &cached_state);
1768                 if (extents_locked < 0) {
1769                         if (extents_locked == -EAGAIN)
1770                                 goto again;
1771                         btrfs_delalloc_release_extents(BTRFS_I(inode),
1772                                                        reserve_bytes);
1773                         ret = extents_locked;
1774                         break;
1775                 }
1776
1777                 copied = btrfs_copy_from_user(pos, write_bytes, pages, i);
1778
1779                 num_sectors = BTRFS_BYTES_TO_BLKS(fs_info, reserve_bytes);
1780                 dirty_sectors = round_up(copied + sector_offset,
1781                                         fs_info->sectorsize);
1782                 dirty_sectors = BTRFS_BYTES_TO_BLKS(fs_info, dirty_sectors);
1783
1784                 /*
1785                  * if we have trouble faulting in the pages, fall
1786                  * back to one page at a time
1787                  */
1788                 if (copied < write_bytes)
1789                         nrptrs = 1;
1790
1791                 if (copied == 0) {
1792                         force_page_uptodate = true;
1793                         dirty_sectors = 0;
1794                         dirty_pages = 0;
1795                 } else {
1796                         force_page_uptodate = false;
1797                         dirty_pages = DIV_ROUND_UP(copied + offset,
1798                                                    PAGE_SIZE);
1799                 }
1800
1801                 if (num_sectors > dirty_sectors) {
1802                         /* release everything except the sectors we dirtied */
1803                         release_bytes -= dirty_sectors << fs_info->sectorsize_bits;
1804                         if (only_release_metadata) {
1805                                 btrfs_delalloc_release_metadata(BTRFS_I(inode),
1806                                                         release_bytes, true);
1807                         } else {
1808                                 u64 __pos;
1809
1810                                 __pos = round_down(pos,
1811                                                    fs_info->sectorsize) +
1812                                         (dirty_pages << PAGE_SHIFT);
1813                                 btrfs_delalloc_release_space(BTRFS_I(inode),
1814                                                 data_reserved, __pos,
1815                                                 release_bytes, true);
1816                         }
1817                 }
1818
1819                 release_bytes = round_up(copied + sector_offset,
1820                                         fs_info->sectorsize);
1821
1822                 ret = btrfs_dirty_pages(BTRFS_I(inode), pages,
1823                                         dirty_pages, pos, copied,
1824                                         &cached_state, only_release_metadata);
1825
1826                 /*
1827                  * If we have not locked the extent range, because the range's
1828                  * start offset is >= i_size, we might still have a non-NULL
1829                  * cached extent state, acquired while marking the extent range
1830                  * as delalloc through btrfs_dirty_pages(). Therefore free any
1831                  * possible cached extent state to avoid a memory leak.
1832                  */
1833                 if (extents_locked)
1834                         unlock_extent_cached(&BTRFS_I(inode)->io_tree,
1835                                              lockstart, lockend, &cached_state);
1836                 else
1837                         free_extent_state(cached_state);
1838
1839                 btrfs_delalloc_release_extents(BTRFS_I(inode), reserve_bytes);
1840                 if (ret) {
1841                         btrfs_drop_pages(pages, num_pages);
1842                         break;
1843                 }
1844
1845                 release_bytes = 0;
1846                 if (only_release_metadata)
1847                         btrfs_check_nocow_unlock(BTRFS_I(inode));
1848
1849                 btrfs_drop_pages(pages, num_pages);
1850
1851                 cond_resched();
1852
1853                 balance_dirty_pages_ratelimited(inode->i_mapping);
1854
1855                 pos += copied;
1856                 num_written += copied;
1857         }
1858
1859         kfree(pages);
1860
1861         if (release_bytes) {
1862                 if (only_release_metadata) {
1863                         btrfs_check_nocow_unlock(BTRFS_I(inode));
1864                         btrfs_delalloc_release_metadata(BTRFS_I(inode),
1865                                         release_bytes, true);
1866                 } else {
1867                         btrfs_delalloc_release_space(BTRFS_I(inode),
1868                                         data_reserved,
1869                                         round_down(pos, fs_info->sectorsize),
1870                                         release_bytes, true);
1871                 }
1872         }
1873
1874         extent_changeset_free(data_reserved);
1875         if (num_written > 0) {
1876                 pagecache_isize_extended(inode, old_isize, iocb->ki_pos);
1877                 iocb->ki_pos += num_written;
1878         }
1879 out:
1880         btrfs_inode_unlock(inode, ilock_flags);
1881         return num_written ? num_written : ret;
1882 }
1883
1884 static ssize_t check_direct_IO(struct btrfs_fs_info *fs_info,
1885                                const struct iov_iter *iter, loff_t offset)
1886 {
1887         const u32 blocksize_mask = fs_info->sectorsize - 1;
1888
1889         if (offset & blocksize_mask)
1890                 return -EINVAL;
1891
1892         if (iov_iter_alignment(iter) & blocksize_mask)
1893                 return -EINVAL;
1894
1895         return 0;
1896 }
1897
1898 static ssize_t btrfs_direct_write(struct kiocb *iocb, struct iov_iter *from)
1899 {
1900         struct file *file = iocb->ki_filp;
1901         struct inode *inode = file_inode(file);
1902         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1903         loff_t pos;
1904         ssize_t written = 0;
1905         ssize_t written_buffered;
1906         loff_t endbyte;
1907         ssize_t err;
1908         unsigned int ilock_flags = 0;
1909         struct iomap_dio *dio = NULL;
1910
1911         if (iocb->ki_flags & IOCB_NOWAIT)
1912                 ilock_flags |= BTRFS_ILOCK_TRY;
1913
1914         /* If the write DIO is within EOF, use a shared lock */
1915         if (iocb->ki_pos + iov_iter_count(from) <= i_size_read(inode))
1916                 ilock_flags |= BTRFS_ILOCK_SHARED;
1917
1918 relock:
1919         err = btrfs_inode_lock(inode, ilock_flags);
1920         if (err < 0)
1921                 return err;
1922
1923         err = generic_write_checks(iocb, from);
1924         if (err <= 0) {
1925                 btrfs_inode_unlock(inode, ilock_flags);
1926                 return err;
1927         }
1928
1929         err = btrfs_write_check(iocb, from, err);
1930         if (err < 0) {
1931                 btrfs_inode_unlock(inode, ilock_flags);
1932                 goto out;
1933         }
1934
1935         pos = iocb->ki_pos;
1936         /*
1937          * Re-check since file size may have changed just before taking the
1938          * lock or pos may have changed because of O_APPEND in generic_write_check()
1939          */
1940         if ((ilock_flags & BTRFS_ILOCK_SHARED) &&
1941             pos + iov_iter_count(from) > i_size_read(inode)) {
1942                 btrfs_inode_unlock(inode, ilock_flags);
1943                 ilock_flags &= ~BTRFS_ILOCK_SHARED;
1944                 goto relock;
1945         }
1946
1947         if (check_direct_IO(fs_info, from, pos)) {
1948                 btrfs_inode_unlock(inode, ilock_flags);
1949                 goto buffered;
1950         }
1951
1952         dio = __iomap_dio_rw(iocb, from, &btrfs_dio_iomap_ops,
1953                              &btrfs_dio_ops, is_sync_kiocb(iocb));
1954
1955         btrfs_inode_unlock(inode, ilock_flags);
1956
1957         if (IS_ERR_OR_NULL(dio)) {
1958                 err = PTR_ERR_OR_ZERO(dio);
1959                 if (err < 0 && err != -ENOTBLK)
1960                         goto out;
1961         } else {
1962                 written = iomap_dio_complete(dio);
1963         }
1964
1965         if (written < 0 || !iov_iter_count(from)) {
1966                 err = written;
1967                 goto out;
1968         }
1969
1970 buffered:
1971         pos = iocb->ki_pos;
1972         written_buffered = btrfs_buffered_write(iocb, from);
1973         if (written_buffered < 0) {
1974                 err = written_buffered;
1975                 goto out;
1976         }
1977         /*
1978          * Ensure all data is persisted. We want the next direct IO read to be
1979          * able to read what was just written.
1980          */
1981         endbyte = pos + written_buffered - 1;
1982         err = btrfs_fdatawrite_range(inode, pos, endbyte);
1983         if (err)
1984                 goto out;
1985         err = filemap_fdatawait_range(inode->i_mapping, pos, endbyte);
1986         if (err)
1987                 goto out;
1988         written += written_buffered;
1989         iocb->ki_pos = pos + written_buffered;
1990         invalidate_mapping_pages(file->f_mapping, pos >> PAGE_SHIFT,
1991                                  endbyte >> PAGE_SHIFT);
1992 out:
1993         return written ? written : err;
1994 }
1995
1996 static ssize_t btrfs_file_write_iter(struct kiocb *iocb,
1997                                     struct iov_iter *from)
1998 {
1999         struct file *file = iocb->ki_filp;
2000         struct inode *inode = file_inode(file);
2001         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2002         struct btrfs_root *root = BTRFS_I(inode)->root;
2003         ssize_t num_written = 0;
2004         const bool sync = iocb->ki_flags & IOCB_DSYNC;
2005
2006         /*
2007          * If the fs flips readonly due to some impossible error, although we
2008          * have opened a file as writable, we have to stop this write operation
2009          * to ensure consistency.
2010          */
2011         if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state))
2012                 return -EROFS;
2013
2014         if (!(iocb->ki_flags & IOCB_DIRECT) &&
2015             (iocb->ki_flags & IOCB_NOWAIT))
2016                 return -EOPNOTSUPP;
2017
2018         if (sync)
2019                 atomic_inc(&BTRFS_I(inode)->sync_writers);
2020
2021         if (iocb->ki_flags & IOCB_DIRECT)
2022                 num_written = btrfs_direct_write(iocb, from);
2023         else
2024                 num_written = btrfs_buffered_write(iocb, from);
2025
2026         /*
2027          * We also have to set last_sub_trans to the current log transid,
2028          * otherwise subsequent syncs to a file that's been synced in this
2029          * transaction will appear to have already occurred.
2030          */
2031         spin_lock(&BTRFS_I(inode)->lock);
2032         BTRFS_I(inode)->last_sub_trans = root->log_transid;
2033         spin_unlock(&BTRFS_I(inode)->lock);
2034         if (num_written > 0)
2035                 num_written = generic_write_sync(iocb, num_written);
2036
2037         if (sync)
2038                 atomic_dec(&BTRFS_I(inode)->sync_writers);
2039
2040         current->backing_dev_info = NULL;
2041         return num_written;
2042 }
2043
2044 int btrfs_release_file(struct inode *inode, struct file *filp)
2045 {
2046         struct btrfs_file_private *private = filp->private_data;
2047
2048         if (private && private->filldir_buf)
2049                 kfree(private->filldir_buf);
2050         kfree(private);
2051         filp->private_data = NULL;
2052
2053         /*
2054          * Set by setattr when we are about to truncate a file from a non-zero
2055          * size to a zero size.  This tries to flush down new bytes that may
2056          * have been written if the application were using truncate to replace
2057          * a file in place.
2058          */
2059         if (test_and_clear_bit(BTRFS_INODE_FLUSH_ON_CLOSE,
2060                                &BTRFS_I(inode)->runtime_flags))
2061                         filemap_flush(inode->i_mapping);
2062         return 0;
2063 }
2064
2065 static int start_ordered_ops(struct inode *inode, loff_t start, loff_t end)
2066 {
2067         int ret;
2068         struct blk_plug plug;
2069
2070         /*
2071          * This is only called in fsync, which would do synchronous writes, so
2072          * a plug can merge adjacent IOs as much as possible.  Esp. in case of
2073          * multiple disks using raid profile, a large IO can be split to
2074          * several segments of stripe length (currently 64K).
2075          */
2076         blk_start_plug(&plug);
2077         atomic_inc(&BTRFS_I(inode)->sync_writers);
2078         ret = btrfs_fdatawrite_range(inode, start, end);
2079         atomic_dec(&BTRFS_I(inode)->sync_writers);
2080         blk_finish_plug(&plug);
2081
2082         return ret;
2083 }
2084
2085 /*
2086  * fsync call for both files and directories.  This logs the inode into
2087  * the tree log instead of forcing full commits whenever possible.
2088  *
2089  * It needs to call filemap_fdatawait so that all ordered extent updates are
2090  * in the metadata btree are up to date for copying to the log.
2091  *
2092  * It drops the inode mutex before doing the tree log commit.  This is an
2093  * important optimization for directories because holding the mutex prevents
2094  * new operations on the dir while we write to disk.
2095  */
2096 int btrfs_sync_file(struct file *file, loff_t start, loff_t end, int datasync)
2097 {
2098         struct dentry *dentry = file_dentry(file);
2099         struct inode *inode = d_inode(dentry);
2100         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2101         struct btrfs_root *root = BTRFS_I(inode)->root;
2102         struct btrfs_trans_handle *trans;
2103         struct btrfs_log_ctx ctx;
2104         int ret = 0, err;
2105         u64 len;
2106         bool full_sync;
2107
2108         trace_btrfs_sync_file(file, datasync);
2109
2110         btrfs_init_log_ctx(&ctx, inode);
2111
2112         /*
2113          * Always set the range to a full range, otherwise we can get into
2114          * several problems, from missing file extent items to represent holes
2115          * when not using the NO_HOLES feature, to log tree corruption due to
2116          * races between hole detection during logging and completion of ordered
2117          * extents outside the range, to missing checksums due to ordered extents
2118          * for which we flushed only a subset of their pages.
2119          */
2120         start = 0;
2121         end = LLONG_MAX;
2122         len = (u64)LLONG_MAX + 1;
2123
2124         /*
2125          * We write the dirty pages in the range and wait until they complete
2126          * out of the ->i_mutex. If so, we can flush the dirty pages by
2127          * multi-task, and make the performance up.  See
2128          * btrfs_wait_ordered_range for an explanation of the ASYNC check.
2129          */
2130         ret = start_ordered_ops(inode, start, end);
2131         if (ret)
2132                 goto out;
2133
2134         inode_lock(inode);
2135
2136         atomic_inc(&root->log_batch);
2137
2138         /*
2139          * Always check for the full sync flag while holding the inode's lock,
2140          * to avoid races with other tasks. The flag must be either set all the
2141          * time during logging or always off all the time while logging.
2142          */
2143         full_sync = test_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2144                              &BTRFS_I(inode)->runtime_flags);
2145
2146         /*
2147          * Before we acquired the inode's lock, someone may have dirtied more
2148          * pages in the target range. We need to make sure that writeback for
2149          * any such pages does not start while we are logging the inode, because
2150          * if it does, any of the following might happen when we are not doing a
2151          * full inode sync:
2152          *
2153          * 1) We log an extent after its writeback finishes but before its
2154          *    checksums are added to the csum tree, leading to -EIO errors
2155          *    when attempting to read the extent after a log replay.
2156          *
2157          * 2) We can end up logging an extent before its writeback finishes.
2158          *    Therefore after the log replay we will have a file extent item
2159          *    pointing to an unwritten extent (and no data checksums as well).
2160          *
2161          * So trigger writeback for any eventual new dirty pages and then we
2162          * wait for all ordered extents to complete below.
2163          */
2164         ret = start_ordered_ops(inode, start, end);
2165         if (ret) {
2166                 inode_unlock(inode);
2167                 goto out;
2168         }
2169
2170         /*
2171          * We have to do this here to avoid the priority inversion of waiting on
2172          * IO of a lower priority task while holding a transaction open.
2173          *
2174          * For a full fsync we wait for the ordered extents to complete while
2175          * for a fast fsync we wait just for writeback to complete, and then
2176          * attach the ordered extents to the transaction so that a transaction
2177          * commit waits for their completion, to avoid data loss if we fsync,
2178          * the current transaction commits before the ordered extents complete
2179          * and a power failure happens right after that.
2180          */
2181         if (full_sync) {
2182                 ret = btrfs_wait_ordered_range(inode, start, len);
2183         } else {
2184                 /*
2185                  * Get our ordered extents as soon as possible to avoid doing
2186                  * checksum lookups in the csum tree, and use instead the
2187                  * checksums attached to the ordered extents.
2188                  */
2189                 btrfs_get_ordered_extents_for_logging(BTRFS_I(inode),
2190                                                       &ctx.ordered_extents);
2191                 ret = filemap_fdatawait_range(inode->i_mapping, start, end);
2192         }
2193
2194         if (ret)
2195                 goto out_release_extents;
2196
2197         atomic_inc(&root->log_batch);
2198
2199         /*
2200          * If we are doing a fast fsync we can not bail out if the inode's
2201          * last_trans is <= then the last committed transaction, because we only
2202          * update the last_trans of the inode during ordered extent completion,
2203          * and for a fast fsync we don't wait for that, we only wait for the
2204          * writeback to complete.
2205          */
2206         smp_mb();
2207         if (btrfs_inode_in_log(BTRFS_I(inode), fs_info->generation) ||
2208             (BTRFS_I(inode)->last_trans <= fs_info->last_trans_committed &&
2209              (full_sync || list_empty(&ctx.ordered_extents)))) {
2210                 /*
2211                  * We've had everything committed since the last time we were
2212                  * modified so clear this flag in case it was set for whatever
2213                  * reason, it's no longer relevant.
2214                  */
2215                 clear_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2216                           &BTRFS_I(inode)->runtime_flags);
2217                 /*
2218                  * An ordered extent might have started before and completed
2219                  * already with io errors, in which case the inode was not
2220                  * updated and we end up here. So check the inode's mapping
2221                  * for any errors that might have happened since we last
2222                  * checked called fsync.
2223                  */
2224                 ret = filemap_check_wb_err(inode->i_mapping, file->f_wb_err);
2225                 goto out_release_extents;
2226         }
2227
2228         /*
2229          * We use start here because we will need to wait on the IO to complete
2230          * in btrfs_sync_log, which could require joining a transaction (for
2231          * example checking cross references in the nocow path).  If we use join
2232          * here we could get into a situation where we're waiting on IO to
2233          * happen that is blocked on a transaction trying to commit.  With start
2234          * we inc the extwriter counter, so we wait for all extwriters to exit
2235          * before we start blocking joiners.  This comment is to keep somebody
2236          * from thinking they are super smart and changing this to
2237          * btrfs_join_transaction *cough*Josef*cough*.
2238          */
2239         trans = btrfs_start_transaction(root, 0);
2240         if (IS_ERR(trans)) {
2241                 ret = PTR_ERR(trans);
2242                 goto out_release_extents;
2243         }
2244
2245         ret = btrfs_log_dentry_safe(trans, dentry, &ctx);
2246         btrfs_release_log_ctx_extents(&ctx);
2247         if (ret < 0) {
2248                 /* Fallthrough and commit/free transaction. */
2249                 ret = 1;
2250         }
2251
2252         /* we've logged all the items and now have a consistent
2253          * version of the file in the log.  It is possible that
2254          * someone will come in and modify the file, but that's
2255          * fine because the log is consistent on disk, and we
2256          * have references to all of the file's extents
2257          *
2258          * It is possible that someone will come in and log the
2259          * file again, but that will end up using the synchronization
2260          * inside btrfs_sync_log to keep things safe.
2261          */
2262         inode_unlock(inode);
2263
2264         if (ret != BTRFS_NO_LOG_SYNC) {
2265                 if (!ret) {
2266                         ret = btrfs_sync_log(trans, root, &ctx);
2267                         if (!ret) {
2268                                 ret = btrfs_end_transaction(trans);
2269                                 goto out;
2270                         }
2271                 }
2272                 if (!full_sync) {
2273                         ret = btrfs_wait_ordered_range(inode, start, len);
2274                         if (ret) {
2275                                 btrfs_end_transaction(trans);
2276                                 goto out;
2277                         }
2278                 }
2279                 ret = btrfs_commit_transaction(trans);
2280         } else {
2281                 ret = btrfs_end_transaction(trans);
2282         }
2283 out:
2284         ASSERT(list_empty(&ctx.list));
2285         err = file_check_and_advance_wb_err(file);
2286         if (!ret)
2287                 ret = err;
2288         return ret > 0 ? -EIO : ret;
2289
2290 out_release_extents:
2291         btrfs_release_log_ctx_extents(&ctx);
2292         inode_unlock(inode);
2293         goto out;
2294 }
2295
2296 static const struct vm_operations_struct btrfs_file_vm_ops = {
2297         .fault          = filemap_fault,
2298         .map_pages      = filemap_map_pages,
2299         .page_mkwrite   = btrfs_page_mkwrite,
2300 };
2301
2302 static int btrfs_file_mmap(struct file  *filp, struct vm_area_struct *vma)
2303 {
2304         struct address_space *mapping = filp->f_mapping;
2305
2306         if (!mapping->a_ops->readpage)
2307                 return -ENOEXEC;
2308
2309         file_accessed(filp);
2310         vma->vm_ops = &btrfs_file_vm_ops;
2311
2312         return 0;
2313 }
2314
2315 static int hole_mergeable(struct btrfs_inode *inode, struct extent_buffer *leaf,
2316                           int slot, u64 start, u64 end)
2317 {
2318         struct btrfs_file_extent_item *fi;
2319         struct btrfs_key key;
2320
2321         if (slot < 0 || slot >= btrfs_header_nritems(leaf))
2322                 return 0;
2323
2324         btrfs_item_key_to_cpu(leaf, &key, slot);
2325         if (key.objectid != btrfs_ino(inode) ||
2326             key.type != BTRFS_EXTENT_DATA_KEY)
2327                 return 0;
2328
2329         fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
2330
2331         if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG)
2332                 return 0;
2333
2334         if (btrfs_file_extent_disk_bytenr(leaf, fi))
2335                 return 0;
2336
2337         if (key.offset == end)
2338                 return 1;
2339         if (key.offset + btrfs_file_extent_num_bytes(leaf, fi) == start)
2340                 return 1;
2341         return 0;
2342 }
2343
2344 static int fill_holes(struct btrfs_trans_handle *trans,
2345                 struct btrfs_inode *inode,
2346                 struct btrfs_path *path, u64 offset, u64 end)
2347 {
2348         struct btrfs_fs_info *fs_info = trans->fs_info;
2349         struct btrfs_root *root = inode->root;
2350         struct extent_buffer *leaf;
2351         struct btrfs_file_extent_item *fi;
2352         struct extent_map *hole_em;
2353         struct extent_map_tree *em_tree = &inode->extent_tree;
2354         struct btrfs_key key;
2355         int ret;
2356
2357         if (btrfs_fs_incompat(fs_info, NO_HOLES))
2358                 goto out;
2359
2360         key.objectid = btrfs_ino(inode);
2361         key.type = BTRFS_EXTENT_DATA_KEY;
2362         key.offset = offset;
2363
2364         ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
2365         if (ret <= 0) {
2366                 /*
2367                  * We should have dropped this offset, so if we find it then
2368                  * something has gone horribly wrong.
2369                  */
2370                 if (ret == 0)
2371                         ret = -EINVAL;
2372                 return ret;
2373         }
2374
2375         leaf = path->nodes[0];
2376         if (hole_mergeable(inode, leaf, path->slots[0] - 1, offset, end)) {
2377                 u64 num_bytes;
2378
2379                 path->slots[0]--;
2380                 fi = btrfs_item_ptr(leaf, path->slots[0],
2381                                     struct btrfs_file_extent_item);
2382                 num_bytes = btrfs_file_extent_num_bytes(leaf, fi) +
2383                         end - offset;
2384                 btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2385                 btrfs_set_file_extent_ram_bytes(leaf, fi, num_bytes);
2386                 btrfs_set_file_extent_offset(leaf, fi, 0);
2387                 btrfs_mark_buffer_dirty(leaf);
2388                 goto out;
2389         }
2390
2391         if (hole_mergeable(inode, leaf, path->slots[0], offset, end)) {
2392                 u64 num_bytes;
2393
2394                 key.offset = offset;
2395                 btrfs_set_item_key_safe(fs_info, path, &key);
2396                 fi = btrfs_item_ptr(leaf, path->slots[0],
2397                                     struct btrfs_file_extent_item);
2398                 num_bytes = btrfs_file_extent_num_bytes(leaf, fi) + end -
2399                         offset;
2400                 btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2401                 btrfs_set_file_extent_ram_bytes(leaf, fi, num_bytes);
2402                 btrfs_set_file_extent_offset(leaf, fi, 0);
2403                 btrfs_mark_buffer_dirty(leaf);
2404                 goto out;
2405         }
2406         btrfs_release_path(path);
2407
2408         ret = btrfs_insert_file_extent(trans, root, btrfs_ino(inode),
2409                         offset, 0, 0, end - offset, 0, end - offset, 0, 0, 0);
2410         if (ret)
2411                 return ret;
2412
2413 out:
2414         btrfs_release_path(path);
2415
2416         hole_em = alloc_extent_map();
2417         if (!hole_em) {
2418                 btrfs_drop_extent_cache(inode, offset, end - 1, 0);
2419                 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &inode->runtime_flags);
2420         } else {
2421                 hole_em->start = offset;
2422                 hole_em->len = end - offset;
2423                 hole_em->ram_bytes = hole_em->len;
2424                 hole_em->orig_start = offset;
2425
2426                 hole_em->block_start = EXTENT_MAP_HOLE;
2427                 hole_em->block_len = 0;
2428                 hole_em->orig_block_len = 0;
2429                 hole_em->compress_type = BTRFS_COMPRESS_NONE;
2430                 hole_em->generation = trans->transid;
2431
2432                 do {
2433                         btrfs_drop_extent_cache(inode, offset, end - 1, 0);
2434                         write_lock(&em_tree->lock);
2435                         ret = add_extent_mapping(em_tree, hole_em, 1);
2436                         write_unlock(&em_tree->lock);
2437                 } while (ret == -EEXIST);
2438                 free_extent_map(hole_em);
2439                 if (ret)
2440                         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2441                                         &inode->runtime_flags);
2442         }
2443
2444         return 0;
2445 }
2446
2447 /*
2448  * Find a hole extent on given inode and change start/len to the end of hole
2449  * extent.(hole/vacuum extent whose em->start <= start &&
2450  *         em->start + em->len > start)
2451  * When a hole extent is found, return 1 and modify start/len.
2452  */
2453 static int find_first_non_hole(struct btrfs_inode *inode, u64 *start, u64 *len)
2454 {
2455         struct btrfs_fs_info *fs_info = inode->root->fs_info;
2456         struct extent_map *em;
2457         int ret = 0;
2458
2459         em = btrfs_get_extent(inode, NULL, 0,
2460                               round_down(*start, fs_info->sectorsize),
2461                               round_up(*len, fs_info->sectorsize));
2462         if (IS_ERR(em))
2463                 return PTR_ERR(em);
2464
2465         /* Hole or vacuum extent(only exists in no-hole mode) */
2466         if (em->block_start == EXTENT_MAP_HOLE) {
2467                 ret = 1;
2468                 *len = em->start + em->len > *start + *len ?
2469                        0 : *start + *len - em->start - em->len;
2470                 *start = em->start + em->len;
2471         }
2472         free_extent_map(em);
2473         return ret;
2474 }
2475
2476 static int btrfs_punch_hole_lock_range(struct inode *inode,
2477                                        const u64 lockstart,
2478                                        const u64 lockend,
2479                                        struct extent_state **cached_state)
2480 {
2481         while (1) {
2482                 struct btrfs_ordered_extent *ordered;
2483                 int ret;
2484
2485                 truncate_pagecache_range(inode, lockstart, lockend);
2486
2487                 lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2488                                  cached_state);
2489                 ordered = btrfs_lookup_first_ordered_extent(BTRFS_I(inode),
2490                                                             lockend);
2491
2492                 /*
2493                  * We need to make sure we have no ordered extents in this range
2494                  * and nobody raced in and read a page in this range, if we did
2495                  * we need to try again.
2496                  */
2497                 if ((!ordered ||
2498                     (ordered->file_offset + ordered->num_bytes <= lockstart ||
2499                      ordered->file_offset > lockend)) &&
2500                      !filemap_range_has_page(inode->i_mapping,
2501                                              lockstart, lockend)) {
2502                         if (ordered)
2503                                 btrfs_put_ordered_extent(ordered);
2504                         break;
2505                 }
2506                 if (ordered)
2507                         btrfs_put_ordered_extent(ordered);
2508                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart,
2509                                      lockend, cached_state);
2510                 ret = btrfs_wait_ordered_range(inode, lockstart,
2511                                                lockend - lockstart + 1);
2512                 if (ret)
2513                         return ret;
2514         }
2515         return 0;
2516 }
2517
2518 static int btrfs_insert_replace_extent(struct btrfs_trans_handle *trans,
2519                                      struct btrfs_inode *inode,
2520                                      struct btrfs_path *path,
2521                                      struct btrfs_replace_extent_info *extent_info,
2522                                      const u64 replace_len,
2523                                      const u64 bytes_to_drop)
2524 {
2525         struct btrfs_fs_info *fs_info = trans->fs_info;
2526         struct btrfs_root *root = inode->root;
2527         struct btrfs_file_extent_item *extent;
2528         struct extent_buffer *leaf;
2529         struct btrfs_key key;
2530         int slot;
2531         struct btrfs_ref ref = { 0 };
2532         int ret;
2533
2534         if (replace_len == 0)
2535                 return 0;
2536
2537         if (extent_info->disk_offset == 0 &&
2538             btrfs_fs_incompat(fs_info, NO_HOLES)) {
2539                 btrfs_update_inode_bytes(inode, 0, bytes_to_drop);
2540                 return 0;
2541         }
2542
2543         key.objectid = btrfs_ino(inode);
2544         key.type = BTRFS_EXTENT_DATA_KEY;
2545         key.offset = extent_info->file_offset;
2546         ret = btrfs_insert_empty_item(trans, root, path, &key,
2547                                       sizeof(struct btrfs_file_extent_item));
2548         if (ret)
2549                 return ret;
2550         leaf = path->nodes[0];
2551         slot = path->slots[0];
2552         write_extent_buffer(leaf, extent_info->extent_buf,
2553                             btrfs_item_ptr_offset(leaf, slot),
2554                             sizeof(struct btrfs_file_extent_item));
2555         extent = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
2556         ASSERT(btrfs_file_extent_type(leaf, extent) != BTRFS_FILE_EXTENT_INLINE);
2557         btrfs_set_file_extent_offset(leaf, extent, extent_info->data_offset);
2558         btrfs_set_file_extent_num_bytes(leaf, extent, replace_len);
2559         if (extent_info->is_new_extent)
2560                 btrfs_set_file_extent_generation(leaf, extent, trans->transid);
2561         btrfs_mark_buffer_dirty(leaf);
2562         btrfs_release_path(path);
2563
2564         ret = btrfs_inode_set_file_extent_range(inode, extent_info->file_offset,
2565                                                 replace_len);
2566         if (ret)
2567                 return ret;
2568
2569         /* If it's a hole, nothing more needs to be done. */
2570         if (extent_info->disk_offset == 0) {
2571                 btrfs_update_inode_bytes(inode, 0, bytes_to_drop);
2572                 return 0;
2573         }
2574
2575         btrfs_update_inode_bytes(inode, replace_len, bytes_to_drop);
2576
2577         if (extent_info->is_new_extent && extent_info->insertions == 0) {
2578                 key.objectid = extent_info->disk_offset;
2579                 key.type = BTRFS_EXTENT_ITEM_KEY;
2580                 key.offset = extent_info->disk_len;
2581                 ret = btrfs_alloc_reserved_file_extent(trans, root,
2582                                                        btrfs_ino(inode),
2583                                                        extent_info->file_offset,
2584                                                        extent_info->qgroup_reserved,
2585                                                        &key);
2586         } else {
2587                 u64 ref_offset;
2588
2589                 btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF,
2590                                        extent_info->disk_offset,
2591                                        extent_info->disk_len, 0);
2592                 ref_offset = extent_info->file_offset - extent_info->data_offset;
2593                 btrfs_init_data_ref(&ref, root->root_key.objectid,
2594                                     btrfs_ino(inode), ref_offset);
2595                 ret = btrfs_inc_extent_ref(trans, &ref);
2596         }
2597
2598         extent_info->insertions++;
2599
2600         return ret;
2601 }
2602
2603 /*
2604  * The respective range must have been previously locked, as well as the inode.
2605  * The end offset is inclusive (last byte of the range).
2606  * @extent_info is NULL for fallocate's hole punching and non-NULL when replacing
2607  * the file range with an extent.
2608  * When not punching a hole, we don't want to end up in a state where we dropped
2609  * extents without inserting a new one, so we must abort the transaction to avoid
2610  * a corruption.
2611  */
2612 int btrfs_replace_file_extents(struct inode *inode, struct btrfs_path *path,
2613                            const u64 start, const u64 end,
2614                            struct btrfs_replace_extent_info *extent_info,
2615                            struct btrfs_trans_handle **trans_out)
2616 {
2617         struct btrfs_drop_extents_args drop_args = { 0 };
2618         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2619         u64 min_size = btrfs_calc_insert_metadata_size(fs_info, 1);
2620         u64 ino_size = round_up(inode->i_size, fs_info->sectorsize);
2621         struct btrfs_root *root = BTRFS_I(inode)->root;
2622         struct btrfs_trans_handle *trans = NULL;
2623         struct btrfs_block_rsv *rsv;
2624         unsigned int rsv_count;
2625         u64 cur_offset;
2626         u64 len = end - start;
2627         int ret = 0;
2628
2629         if (end <= start)
2630                 return -EINVAL;
2631
2632         rsv = btrfs_alloc_block_rsv(fs_info, BTRFS_BLOCK_RSV_TEMP);
2633         if (!rsv) {
2634                 ret = -ENOMEM;
2635                 goto out;
2636         }
2637         rsv->size = btrfs_calc_insert_metadata_size(fs_info, 1);
2638         rsv->failfast = 1;
2639
2640         /*
2641          * 1 - update the inode
2642          * 1 - removing the extents in the range
2643          * 1 - adding the hole extent if no_holes isn't set or if we are
2644          *     replacing the range with a new extent
2645          */
2646         if (!btrfs_fs_incompat(fs_info, NO_HOLES) || extent_info)
2647                 rsv_count = 3;
2648         else
2649                 rsv_count = 2;
2650
2651         trans = btrfs_start_transaction(root, rsv_count);
2652         if (IS_ERR(trans)) {
2653                 ret = PTR_ERR(trans);
2654                 trans = NULL;
2655                 goto out_free;
2656         }
2657
2658         ret = btrfs_block_rsv_migrate(&fs_info->trans_block_rsv, rsv,
2659                                       min_size, false);
2660         BUG_ON(ret);
2661         trans->block_rsv = rsv;
2662
2663         cur_offset = start;
2664         drop_args.path = path;
2665         drop_args.end = end + 1;
2666         drop_args.drop_cache = true;
2667         while (cur_offset < end) {
2668                 drop_args.start = cur_offset;
2669                 ret = btrfs_drop_extents(trans, root, BTRFS_I(inode), &drop_args);
2670                 /* If we are punching a hole decrement the inode's byte count */
2671                 if (!extent_info)
2672                         btrfs_update_inode_bytes(BTRFS_I(inode), 0,
2673                                                  drop_args.bytes_found);
2674                 if (ret != -ENOSPC) {
2675                         /*
2676                          * When cloning we want to avoid transaction aborts when
2677                          * nothing was done and we are attempting to clone parts
2678                          * of inline extents, in such cases -EOPNOTSUPP is
2679                          * returned by __btrfs_drop_extents() without having
2680                          * changed anything in the file.
2681                          */
2682                         if (extent_info && !extent_info->is_new_extent &&
2683                             ret && ret != -EOPNOTSUPP)
2684                                 btrfs_abort_transaction(trans, ret);
2685                         break;
2686                 }
2687
2688                 trans->block_rsv = &fs_info->trans_block_rsv;
2689
2690                 if (!extent_info && cur_offset < drop_args.drop_end &&
2691                     cur_offset < ino_size) {
2692                         ret = fill_holes(trans, BTRFS_I(inode), path,
2693                                          cur_offset, drop_args.drop_end);
2694                         if (ret) {
2695                                 /*
2696                                  * If we failed then we didn't insert our hole
2697                                  * entries for the area we dropped, so now the
2698                                  * fs is corrupted, so we must abort the
2699                                  * transaction.
2700                                  */
2701                                 btrfs_abort_transaction(trans, ret);
2702                                 break;
2703                         }
2704                 } else if (!extent_info && cur_offset < drop_args.drop_end) {
2705                         /*
2706                          * We are past the i_size here, but since we didn't
2707                          * insert holes we need to clear the mapped area so we
2708                          * know to not set disk_i_size in this area until a new
2709                          * file extent is inserted here.
2710                          */
2711                         ret = btrfs_inode_clear_file_extent_range(BTRFS_I(inode),
2712                                         cur_offset,
2713                                         drop_args.drop_end - cur_offset);
2714                         if (ret) {
2715                                 /*
2716                                  * We couldn't clear our area, so we could
2717                                  * presumably adjust up and corrupt the fs, so
2718                                  * we need to abort.
2719                                  */
2720                                 btrfs_abort_transaction(trans, ret);
2721                                 break;
2722                         }
2723                 }
2724
2725                 if (extent_info &&
2726                     drop_args.drop_end > extent_info->file_offset) {
2727                         u64 replace_len = drop_args.drop_end -
2728                                           extent_info->file_offset;
2729
2730                         ret = btrfs_insert_replace_extent(trans, BTRFS_I(inode),
2731                                         path, extent_info, replace_len,
2732                                         drop_args.bytes_found);
2733                         if (ret) {
2734                                 btrfs_abort_transaction(trans, ret);
2735                                 break;
2736                         }
2737                         extent_info->data_len -= replace_len;
2738                         extent_info->data_offset += replace_len;
2739                         extent_info->file_offset += replace_len;
2740                 }
2741
2742                 cur_offset = drop_args.drop_end;
2743
2744                 ret = btrfs_update_inode(trans, root, BTRFS_I(inode));
2745                 if (ret)
2746                         break;
2747
2748                 btrfs_end_transaction(trans);
2749                 btrfs_btree_balance_dirty(fs_info);
2750
2751                 trans = btrfs_start_transaction(root, rsv_count);
2752                 if (IS_ERR(trans)) {
2753                         ret = PTR_ERR(trans);
2754                         trans = NULL;
2755                         break;
2756                 }
2757
2758                 ret = btrfs_block_rsv_migrate(&fs_info->trans_block_rsv,
2759                                               rsv, min_size, false);
2760                 BUG_ON(ret);    /* shouldn't happen */
2761                 trans->block_rsv = rsv;
2762
2763                 if (!extent_info) {
2764                         ret = find_first_non_hole(BTRFS_I(inode), &cur_offset,
2765                                                   &len);
2766                         if (unlikely(ret < 0))
2767                                 break;
2768                         if (ret && !len) {
2769                                 ret = 0;
2770                                 break;
2771                         }
2772                 }
2773         }
2774
2775         /*
2776          * If we were cloning, force the next fsync to be a full one since we
2777          * we replaced (or just dropped in the case of cloning holes when
2778          * NO_HOLES is enabled) extents and extent maps.
2779          * This is for the sake of simplicity, and cloning into files larger
2780          * than 16Mb would force the full fsync any way (when
2781          * try_release_extent_mapping() is invoked during page cache truncation.
2782          */
2783         if (extent_info && !extent_info->is_new_extent)
2784                 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2785                         &BTRFS_I(inode)->runtime_flags);
2786
2787         if (ret)
2788                 goto out_trans;
2789
2790         trans->block_rsv = &fs_info->trans_block_rsv;
2791         /*
2792          * If we are using the NO_HOLES feature we might have had already an
2793          * hole that overlaps a part of the region [lockstart, lockend] and
2794          * ends at (or beyond) lockend. Since we have no file extent items to
2795          * represent holes, drop_end can be less than lockend and so we must
2796          * make sure we have an extent map representing the existing hole (the
2797          * call to __btrfs_drop_extents() might have dropped the existing extent
2798          * map representing the existing hole), otherwise the fast fsync path
2799          * will not record the existence of the hole region
2800          * [existing_hole_start, lockend].
2801          */
2802         if (drop_args.drop_end <= end)
2803                 drop_args.drop_end = end + 1;
2804         /*
2805          * Don't insert file hole extent item if it's for a range beyond eof
2806          * (because it's useless) or if it represents a 0 bytes range (when
2807          * cur_offset == drop_end).
2808          */
2809         if (!extent_info && cur_offset < ino_size &&
2810             cur_offset < drop_args.drop_end) {
2811                 ret = fill_holes(trans, BTRFS_I(inode), path,
2812                                  cur_offset, drop_args.drop_end);
2813                 if (ret) {
2814                         /* Same comment as above. */
2815                         btrfs_abort_transaction(trans, ret);
2816                         goto out_trans;
2817                 }
2818         } else if (!extent_info && cur_offset < drop_args.drop_end) {
2819                 /* See the comment in the loop above for the reasoning here. */
2820                 ret = btrfs_inode_clear_file_extent_range(BTRFS_I(inode),
2821                                 cur_offset, drop_args.drop_end - cur_offset);
2822                 if (ret) {
2823                         btrfs_abort_transaction(trans, ret);
2824                         goto out_trans;
2825                 }
2826
2827         }
2828         if (extent_info) {
2829                 ret = btrfs_insert_replace_extent(trans, BTRFS_I(inode), path,
2830                                 extent_info, extent_info->data_len,
2831                                 drop_args.bytes_found);
2832                 if (ret) {
2833                         btrfs_abort_transaction(trans, ret);
2834                         goto out_trans;
2835                 }
2836         }
2837
2838 out_trans:
2839         if (!trans)
2840                 goto out_free;
2841
2842         trans->block_rsv = &fs_info->trans_block_rsv;
2843         if (ret)
2844                 btrfs_end_transaction(trans);
2845         else
2846                 *trans_out = trans;
2847 out_free:
2848         btrfs_free_block_rsv(fs_info, rsv);
2849 out:
2850         return ret;
2851 }
2852
2853 static int btrfs_punch_hole(struct inode *inode, loff_t offset, loff_t len)
2854 {
2855         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2856         struct btrfs_root *root = BTRFS_I(inode)->root;
2857         struct extent_state *cached_state = NULL;
2858         struct btrfs_path *path;
2859         struct btrfs_trans_handle *trans = NULL;
2860         u64 lockstart;
2861         u64 lockend;
2862         u64 tail_start;
2863         u64 tail_len;
2864         u64 orig_start = offset;
2865         int ret = 0;
2866         bool same_block;
2867         u64 ino_size;
2868         bool truncated_block = false;
2869         bool updated_inode = false;
2870
2871         ret = btrfs_wait_ordered_range(inode, offset, len);
2872         if (ret)
2873                 return ret;
2874
2875         inode_lock(inode);
2876         ino_size = round_up(inode->i_size, fs_info->sectorsize);
2877         ret = find_first_non_hole(BTRFS_I(inode), &offset, &len);
2878         if (ret < 0)
2879                 goto out_only_mutex;
2880         if (ret && !len) {
2881                 /* Already in a large hole */
2882                 ret = 0;
2883                 goto out_only_mutex;
2884         }
2885
2886         lockstart = round_up(offset, btrfs_inode_sectorsize(BTRFS_I(inode)));
2887         lockend = round_down(offset + len,
2888                              btrfs_inode_sectorsize(BTRFS_I(inode))) - 1;
2889         same_block = (BTRFS_BYTES_TO_BLKS(fs_info, offset))
2890                 == (BTRFS_BYTES_TO_BLKS(fs_info, offset + len - 1));
2891         /*
2892          * We needn't truncate any block which is beyond the end of the file
2893          * because we are sure there is no data there.
2894          */
2895         /*
2896          * Only do this if we are in the same block and we aren't doing the
2897          * entire block.
2898          */
2899         if (same_block && len < fs_info->sectorsize) {
2900                 if (offset < ino_size) {
2901                         truncated_block = true;
2902                         ret = btrfs_truncate_block(BTRFS_I(inode), offset, len,
2903                                                    0);
2904                 } else {
2905                         ret = 0;
2906                 }
2907                 goto out_only_mutex;
2908         }
2909
2910         /* zero back part of the first block */
2911         if (offset < ino_size) {
2912                 truncated_block = true;
2913                 ret = btrfs_truncate_block(BTRFS_I(inode), offset, 0, 0);
2914                 if (ret) {
2915                         inode_unlock(inode);
2916                         return ret;
2917                 }
2918         }
2919
2920         /* Check the aligned pages after the first unaligned page,
2921          * if offset != orig_start, which means the first unaligned page
2922          * including several following pages are already in holes,
2923          * the extra check can be skipped */
2924         if (offset == orig_start) {
2925                 /* after truncate page, check hole again */
2926                 len = offset + len - lockstart;
2927                 offset = lockstart;
2928                 ret = find_first_non_hole(BTRFS_I(inode), &offset, &len);
2929                 if (ret < 0)
2930                         goto out_only_mutex;
2931                 if (ret && !len) {
2932                         ret = 0;
2933                         goto out_only_mutex;
2934                 }
2935                 lockstart = offset;
2936         }
2937
2938         /* Check the tail unaligned part is in a hole */
2939         tail_start = lockend + 1;
2940         tail_len = offset + len - tail_start;
2941         if (tail_len) {
2942                 ret = find_first_non_hole(BTRFS_I(inode), &tail_start, &tail_len);
2943                 if (unlikely(ret < 0))
2944                         goto out_only_mutex;
2945                 if (!ret) {
2946                         /* zero the front end of the last page */
2947                         if (tail_start + tail_len < ino_size) {
2948                                 truncated_block = true;
2949                                 ret = btrfs_truncate_block(BTRFS_I(inode),
2950                                                         tail_start + tail_len,
2951                                                         0, 1);
2952                                 if (ret)
2953                                         goto out_only_mutex;
2954                         }
2955                 }
2956         }
2957
2958         if (lockend < lockstart) {
2959                 ret = 0;
2960                 goto out_only_mutex;
2961         }
2962
2963         ret = btrfs_punch_hole_lock_range(inode, lockstart, lockend,
2964                                           &cached_state);
2965         if (ret)
2966                 goto out_only_mutex;
2967
2968         path = btrfs_alloc_path();
2969         if (!path) {
2970                 ret = -ENOMEM;
2971                 goto out;
2972         }
2973
2974         ret = btrfs_replace_file_extents(inode, path, lockstart, lockend, NULL,
2975                                      &trans);
2976         btrfs_free_path(path);
2977         if (ret)
2978                 goto out;
2979
2980         ASSERT(trans != NULL);
2981         inode_inc_iversion(inode);
2982         inode->i_mtime = inode->i_ctime = current_time(inode);
2983         ret = btrfs_update_inode(trans, root, BTRFS_I(inode));
2984         updated_inode = true;
2985         btrfs_end_transaction(trans);
2986         btrfs_btree_balance_dirty(fs_info);
2987 out:
2988         unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2989                              &cached_state);
2990 out_only_mutex:
2991         if (!updated_inode && truncated_block && !ret) {
2992                 /*
2993                  * If we only end up zeroing part of a page, we still need to
2994                  * update the inode item, so that all the time fields are
2995                  * updated as well as the necessary btrfs inode in memory fields
2996                  * for detecting, at fsync time, if the inode isn't yet in the
2997                  * log tree or it's there but not up to date.
2998                  */
2999                 struct timespec64 now = current_time(inode);
3000
3001                 inode_inc_iversion(inode);
3002                 inode->i_mtime = now;
3003                 inode->i_ctime = now;
3004                 trans = btrfs_start_transaction(root, 1);
3005                 if (IS_ERR(trans)) {
3006                         ret = PTR_ERR(trans);
3007                 } else {
3008                         int ret2;
3009
3010                         ret = btrfs_update_inode(trans, root, BTRFS_I(inode));
3011                         ret2 = btrfs_end_transaction(trans);
3012                         if (!ret)
3013                                 ret = ret2;
3014                 }
3015         }
3016         inode_unlock(inode);
3017         return ret;
3018 }
3019
3020 /* Helper structure to record which range is already reserved */
3021 struct falloc_range {
3022         struct list_head list;
3023         u64 start;
3024         u64 len;
3025 };
3026
3027 /*
3028  * Helper function to add falloc range
3029  *
3030  * Caller should have locked the larger range of extent containing
3031  * [start, len)
3032  */
3033 static int add_falloc_range(struct list_head *head, u64 start, u64 len)
3034 {
3035         struct falloc_range *prev = NULL;
3036         struct falloc_range *range = NULL;
3037
3038         if (list_empty(head))
3039                 goto insert;
3040
3041         /*
3042          * As fallocate iterate by bytenr order, we only need to check
3043          * the last range.
3044          */
3045         prev = list_entry(head->prev, struct falloc_range, list);
3046         if (prev->start + prev->len == start) {
3047                 prev->len += len;
3048                 return 0;
3049         }
3050 insert:
3051         range = kmalloc(sizeof(*range), GFP_KERNEL);
3052         if (!range)
3053                 return -ENOMEM;
3054         range->start = start;
3055         range->len = len;
3056         list_add_tail(&range->list, head);
3057         return 0;
3058 }
3059
3060 static int btrfs_fallocate_update_isize(struct inode *inode,
3061                                         const u64 end,
3062                                         const int mode)
3063 {
3064         struct btrfs_trans_handle *trans;
3065         struct btrfs_root *root = BTRFS_I(inode)->root;
3066         int ret;
3067         int ret2;
3068
3069         if (mode & FALLOC_FL_KEEP_SIZE || end <= i_size_read(inode))
3070                 return 0;
3071
3072         trans = btrfs_start_transaction(root, 1);
3073         if (IS_ERR(trans))
3074                 return PTR_ERR(trans);
3075
3076         inode->i_ctime = current_time(inode);
3077         i_size_write(inode, end);
3078         btrfs_inode_safe_disk_i_size_write(BTRFS_I(inode), 0);
3079         ret = btrfs_update_inode(trans, root, BTRFS_I(inode));
3080         ret2 = btrfs_end_transaction(trans);
3081
3082         return ret ? ret : ret2;
3083 }
3084
3085 enum {
3086         RANGE_BOUNDARY_WRITTEN_EXTENT,
3087         RANGE_BOUNDARY_PREALLOC_EXTENT,
3088         RANGE_BOUNDARY_HOLE,
3089 };
3090
3091 static int btrfs_zero_range_check_range_boundary(struct btrfs_inode *inode,
3092                                                  u64 offset)
3093 {
3094         const u64 sectorsize = btrfs_inode_sectorsize(inode);
3095         struct extent_map *em;
3096         int ret;
3097
3098         offset = round_down(offset, sectorsize);
3099         em = btrfs_get_extent(inode, NULL, 0, offset, sectorsize);
3100         if (IS_ERR(em))
3101                 return PTR_ERR(em);
3102
3103         if (em->block_start == EXTENT_MAP_HOLE)
3104                 ret = RANGE_BOUNDARY_HOLE;
3105         else if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
3106                 ret = RANGE_BOUNDARY_PREALLOC_EXTENT;
3107         else
3108                 ret = RANGE_BOUNDARY_WRITTEN_EXTENT;
3109
3110         free_extent_map(em);
3111         return ret;
3112 }
3113
3114 static int btrfs_zero_range(struct inode *inode,
3115                             loff_t offset,
3116                             loff_t len,
3117                             const int mode)
3118 {
3119         struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
3120         struct extent_map *em;
3121         struct extent_changeset *data_reserved = NULL;
3122         int ret;
3123         u64 alloc_hint = 0;
3124         const u64 sectorsize = btrfs_inode_sectorsize(BTRFS_I(inode));
3125         u64 alloc_start = round_down(offset, sectorsize);
3126         u64 alloc_end = round_up(offset + len, sectorsize);
3127         u64 bytes_to_reserve = 0;
3128         bool space_reserved = false;
3129
3130         inode_dio_wait(inode);
3131
3132         em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, alloc_start,
3133                               alloc_end - alloc_start);
3134         if (IS_ERR(em)) {
3135                 ret = PTR_ERR(em);
3136                 goto out;
3137         }
3138
3139         /*
3140          * Avoid hole punching and extent allocation for some cases. More cases
3141          * could be considered, but these are unlikely common and we keep things
3142          * as simple as possible for now. Also, intentionally, if the target
3143          * range contains one or more prealloc extents together with regular
3144          * extents and holes, we drop all the existing extents and allocate a
3145          * new prealloc extent, so that we get a larger contiguous disk extent.
3146          */
3147         if (em->start <= alloc_start &&
3148             test_bit(EXTENT_FLAG_PREALLOC, &em->flags)) {
3149                 const u64 em_end = em->start + em->len;
3150
3151                 if (em_end >= offset + len) {
3152                         /*
3153                          * The whole range is already a prealloc extent,
3154                          * do nothing except updating the inode's i_size if
3155                          * needed.
3156                          */
3157                         free_extent_map(em);
3158                         ret = btrfs_fallocate_update_isize(inode, offset + len,
3159                                                            mode);
3160                         goto out;
3161                 }
3162                 /*
3163                  * Part of the range is already a prealloc extent, so operate
3164                  * only on the remaining part of the range.
3165                  */
3166                 alloc_start = em_end;
3167                 ASSERT(IS_ALIGNED(alloc_start, sectorsize));
3168                 len = offset + len - alloc_start;
3169                 offset = alloc_start;
3170                 alloc_hint = em->block_start + em->len;
3171         }
3172         free_extent_map(em);
3173
3174         if (BTRFS_BYTES_TO_BLKS(fs_info, offset) ==
3175             BTRFS_BYTES_TO_BLKS(fs_info, offset + len - 1)) {
3176                 em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, alloc_start,
3177                                       sectorsize);
3178                 if (IS_ERR(em)) {
3179                         ret = PTR_ERR(em);
3180                         goto out;
3181                 }
3182
3183                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags)) {
3184                         free_extent_map(em);
3185                         ret = btrfs_fallocate_update_isize(inode, offset + len,
3186                                                            mode);
3187                         goto out;
3188                 }
3189                 if (len < sectorsize && em->block_start != EXTENT_MAP_HOLE) {
3190                         free_extent_map(em);
3191                         ret = btrfs_truncate_block(BTRFS_I(inode), offset, len,
3192                                                    0);
3193                         if (!ret)
3194                                 ret = btrfs_fallocate_update_isize(inode,
3195                                                                    offset + len,
3196                                                                    mode);
3197                         return ret;
3198                 }
3199                 free_extent_map(em);
3200                 alloc_start = round_down(offset, sectorsize);
3201                 alloc_end = alloc_start + sectorsize;
3202                 goto reserve_space;
3203         }
3204
3205         alloc_start = round_up(offset, sectorsize);
3206         alloc_end = round_down(offset + len, sectorsize);
3207
3208         /*
3209          * For unaligned ranges, check the pages at the boundaries, they might
3210          * map to an extent, in which case we need to partially zero them, or
3211          * they might map to a hole, in which case we need our allocation range
3212          * to cover them.
3213          */
3214         if (!IS_ALIGNED(offset, sectorsize)) {
3215                 ret = btrfs_zero_range_check_range_boundary(BTRFS_I(inode),
3216                                                             offset);
3217                 if (ret < 0)
3218                         goto out;
3219                 if (ret == RANGE_BOUNDARY_HOLE) {
3220                         alloc_start = round_down(offset, sectorsize);
3221                         ret = 0;
3222                 } else if (ret == RANGE_BOUNDARY_WRITTEN_EXTENT) {
3223                         ret = btrfs_truncate_block(BTRFS_I(inode), offset, 0, 0);
3224                         if (ret)
3225                                 goto out;
3226                 } else {
3227                         ret = 0;
3228                 }
3229         }
3230
3231         if (!IS_ALIGNED(offset + len, sectorsize)) {
3232                 ret = btrfs_zero_range_check_range_boundary(BTRFS_I(inode),
3233                                                             offset + len);
3234                 if (ret < 0)
3235                         goto out;
3236                 if (ret == RANGE_BOUNDARY_HOLE) {
3237                         alloc_end = round_up(offset + len, sectorsize);
3238                         ret = 0;
3239                 } else if (ret == RANGE_BOUNDARY_WRITTEN_EXTENT) {
3240                         ret = btrfs_truncate_block(BTRFS_I(inode), offset + len,
3241                                                    0, 1);
3242                         if (ret)
3243                                 goto out;
3244                 } else {
3245                         ret = 0;
3246                 }
3247         }
3248
3249 reserve_space:
3250         if (alloc_start < alloc_end) {
3251                 struct extent_state *cached_state = NULL;
3252                 const u64 lockstart = alloc_start;
3253                 const u64 lockend = alloc_end - 1;
3254
3255                 bytes_to_reserve = alloc_end - alloc_start;
3256                 ret = btrfs_alloc_data_chunk_ondemand(BTRFS_I(inode),
3257                                                       bytes_to_reserve);
3258                 if (ret < 0)
3259                         goto out;
3260                 space_reserved = true;
3261                 ret = btrfs_punch_hole_lock_range(inode, lockstart, lockend,
3262                                                   &cached_state);
3263                 if (ret)
3264                         goto out;
3265                 ret = btrfs_qgroup_reserve_data(BTRFS_I(inode), &data_reserved,
3266                                                 alloc_start, bytes_to_reserve);
3267                 if (ret)
3268                         goto out;
3269                 ret = btrfs_prealloc_file_range(inode, mode, alloc_start,
3270                                                 alloc_end - alloc_start,
3271                                                 i_blocksize(inode),
3272                                                 offset + len, &alloc_hint);
3273                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart,
3274                                      lockend, &cached_state);
3275                 /* btrfs_prealloc_file_range releases reserved space on error */
3276                 if (ret) {
3277                         space_reserved = false;
3278                         goto out;
3279                 }
3280         }
3281         ret = btrfs_fallocate_update_isize(inode, offset + len, mode);
3282  out:
3283         if (ret && space_reserved)
3284                 btrfs_free_reserved_data_space(BTRFS_I(inode), data_reserved,
3285                                                alloc_start, bytes_to_reserve);
3286         extent_changeset_free(data_reserved);
3287
3288         return ret;
3289 }
3290
3291 static long btrfs_fallocate(struct file *file, int mode,
3292                             loff_t offset, loff_t len)
3293 {
3294         struct inode *inode = file_inode(file);
3295         struct extent_state *cached_state = NULL;
3296         struct extent_changeset *data_reserved = NULL;
3297         struct falloc_range *range;
3298         struct falloc_range *tmp;
3299         struct list_head reserve_list;
3300         u64 cur_offset;
3301         u64 last_byte;
3302         u64 alloc_start;
3303         u64 alloc_end;
3304         u64 alloc_hint = 0;
3305         u64 locked_end;
3306         u64 actual_end = 0;
3307         struct extent_map *em;
3308         int blocksize = btrfs_inode_sectorsize(BTRFS_I(inode));
3309         int ret;
3310
3311         /* Do not allow fallocate in ZONED mode */
3312         if (btrfs_is_zoned(btrfs_sb(inode->i_sb)))
3313                 return -EOPNOTSUPP;
3314
3315         alloc_start = round_down(offset, blocksize);
3316         alloc_end = round_up(offset + len, blocksize);
3317         cur_offset = alloc_start;
3318
3319         /* Make sure we aren't being give some crap mode */
3320         if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE |
3321                      FALLOC_FL_ZERO_RANGE))
3322                 return -EOPNOTSUPP;
3323
3324         if (mode & FALLOC_FL_PUNCH_HOLE)
3325                 return btrfs_punch_hole(inode, offset, len);
3326
3327         /*
3328          * Only trigger disk allocation, don't trigger qgroup reserve
3329          *
3330          * For qgroup space, it will be checked later.
3331          */
3332         if (!(mode & FALLOC_FL_ZERO_RANGE)) {
3333                 ret = btrfs_alloc_data_chunk_ondemand(BTRFS_I(inode),
3334                                                       alloc_end - alloc_start);
3335                 if (ret < 0)
3336                         return ret;
3337         }
3338
3339         btrfs_inode_lock(inode, 0);
3340
3341         if (!(mode & FALLOC_FL_KEEP_SIZE) && offset + len > inode->i_size) {
3342                 ret = inode_newsize_ok(inode, offset + len);
3343                 if (ret)
3344                         goto out;
3345         }
3346
3347         /*
3348          * TODO: Move these two operations after we have checked
3349          * accurate reserved space, or fallocate can still fail but
3350          * with page truncated or size expanded.
3351          *
3352          * But that's a minor problem and won't do much harm BTW.
3353          */
3354         if (alloc_start > inode->i_size) {
3355                 ret = btrfs_cont_expand(BTRFS_I(inode), i_size_read(inode),
3356                                         alloc_start);
3357                 if (ret)
3358                         goto out;
3359         } else if (offset + len > inode->i_size) {
3360                 /*
3361                  * If we are fallocating from the end of the file onward we
3362                  * need to zero out the end of the block if i_size lands in the
3363                  * middle of a block.
3364                  */
3365                 ret = btrfs_truncate_block(BTRFS_I(inode), inode->i_size, 0, 0);
3366                 if (ret)
3367                         goto out;
3368         }
3369
3370         /*
3371          * wait for ordered IO before we have any locks.  We'll loop again
3372          * below with the locks held.
3373          */
3374         ret = btrfs_wait_ordered_range(inode, alloc_start,
3375                                        alloc_end - alloc_start);
3376         if (ret)
3377                 goto out;
3378
3379         if (mode & FALLOC_FL_ZERO_RANGE) {
3380                 ret = btrfs_zero_range(inode, offset, len, mode);
3381                 inode_unlock(inode);
3382                 return ret;
3383         }
3384
3385         locked_end = alloc_end - 1;
3386         while (1) {
3387                 struct btrfs_ordered_extent *ordered;
3388
3389                 /* the extent lock is ordered inside the running
3390                  * transaction
3391                  */
3392                 lock_extent_bits(&BTRFS_I(inode)->io_tree, alloc_start,
3393                                  locked_end, &cached_state);
3394                 ordered = btrfs_lookup_first_ordered_extent(BTRFS_I(inode),
3395                                                             locked_end);
3396
3397                 if (ordered &&
3398                     ordered->file_offset + ordered->num_bytes > alloc_start &&
3399                     ordered->file_offset < alloc_end) {
3400                         btrfs_put_ordered_extent(ordered);
3401                         unlock_extent_cached(&BTRFS_I(inode)->io_tree,
3402                                              alloc_start, locked_end,
3403                                              &cached_state);
3404                         /*
3405                          * we can't wait on the range with the transaction
3406                          * running or with the extent lock held
3407                          */
3408                         ret = btrfs_wait_ordered_range(inode, alloc_start,
3409                                                        alloc_end - alloc_start);
3410                         if (ret)
3411                                 goto out;
3412                 } else {
3413                         if (ordered)
3414                                 btrfs_put_ordered_extent(ordered);
3415                         break;
3416                 }
3417         }
3418
3419         /* First, check if we exceed the qgroup limit */
3420         INIT_LIST_HEAD(&reserve_list);
3421         while (cur_offset < alloc_end) {
3422                 em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, cur_offset,
3423                                       alloc_end - cur_offset);
3424                 if (IS_ERR(em)) {
3425                         ret = PTR_ERR(em);
3426                         break;
3427                 }
3428                 last_byte = min(extent_map_end(em), alloc_end);
3429                 actual_end = min_t(u64, extent_map_end(em), offset + len);
3430                 last_byte = ALIGN(last_byte, blocksize);
3431                 if (em->block_start == EXTENT_MAP_HOLE ||
3432                     (cur_offset >= inode->i_size &&
3433                      !test_bit(EXTENT_FLAG_PREALLOC, &em->flags))) {
3434                         ret = add_falloc_range(&reserve_list, cur_offset,
3435                                                last_byte - cur_offset);
3436                         if (ret < 0) {
3437                                 free_extent_map(em);
3438                                 break;
3439                         }
3440                         ret = btrfs_qgroup_reserve_data(BTRFS_I(inode),
3441                                         &data_reserved, cur_offset,
3442                                         last_byte - cur_offset);
3443                         if (ret < 0) {
3444                                 cur_offset = last_byte;
3445                                 free_extent_map(em);
3446                                 break;
3447                         }
3448                 } else {
3449                         /*
3450                          * Do not need to reserve unwritten extent for this
3451                          * range, free reserved data space first, otherwise
3452                          * it'll result in false ENOSPC error.
3453                          */
3454                         btrfs_free_reserved_data_space(BTRFS_I(inode),
3455                                 data_reserved, cur_offset,
3456                                 last_byte - cur_offset);
3457                 }
3458                 free_extent_map(em);
3459                 cur_offset = last_byte;
3460         }
3461
3462         /*
3463          * If ret is still 0, means we're OK to fallocate.
3464          * Or just cleanup the list and exit.
3465          */
3466         list_for_each_entry_safe(range, tmp, &reserve_list, list) {
3467                 if (!ret)
3468                         ret = btrfs_prealloc_file_range(inode, mode,
3469                                         range->start,
3470                                         range->len, i_blocksize(inode),
3471                                         offset + len, &alloc_hint);
3472                 else
3473                         btrfs_free_reserved_data_space(BTRFS_I(inode),
3474                                         data_reserved, range->start,
3475                                         range->len);
3476                 list_del(&range->list);
3477                 kfree(range);
3478         }
3479         if (ret < 0)
3480                 goto out_unlock;
3481
3482         /*
3483          * We didn't need to allocate any more space, but we still extended the
3484          * size of the file so we need to update i_size and the inode item.
3485          */
3486         ret = btrfs_fallocate_update_isize(inode, actual_end, mode);
3487 out_unlock:
3488         unlock_extent_cached(&BTRFS_I(inode)->io_tree, alloc_start, locked_end,
3489                              &cached_state);
3490 out:
3491         inode_unlock(inode);
3492         /* Let go of our reservation. */
3493         if (ret != 0 && !(mode & FALLOC_FL_ZERO_RANGE))
3494                 btrfs_free_reserved_data_space(BTRFS_I(inode), data_reserved,
3495                                 cur_offset, alloc_end - cur_offset);
3496         extent_changeset_free(data_reserved);
3497         return ret;
3498 }
3499
3500 static loff_t find_desired_extent(struct inode *inode, loff_t offset,
3501                                   int whence)
3502 {
3503         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
3504         struct extent_map *em = NULL;
3505         struct extent_state *cached_state = NULL;
3506         loff_t i_size = inode->i_size;
3507         u64 lockstart;
3508         u64 lockend;
3509         u64 start;
3510         u64 len;
3511         int ret = 0;
3512
3513         if (i_size == 0 || offset >= i_size)
3514                 return -ENXIO;
3515
3516         /*
3517          * offset can be negative, in this case we start finding DATA/HOLE from
3518          * the very start of the file.
3519          */
3520         start = max_t(loff_t, 0, offset);
3521
3522         lockstart = round_down(start, fs_info->sectorsize);
3523         lockend = round_up(i_size, fs_info->sectorsize);
3524         if (lockend <= lockstart)
3525                 lockend = lockstart + fs_info->sectorsize;
3526         lockend--;
3527         len = lockend - lockstart + 1;
3528
3529         lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend,
3530                          &cached_state);
3531
3532         while (start < i_size) {
3533                 em = btrfs_get_extent_fiemap(BTRFS_I(inode), start, len);
3534                 if (IS_ERR(em)) {
3535                         ret = PTR_ERR(em);
3536                         em = NULL;
3537                         break;
3538                 }
3539
3540                 if (whence == SEEK_HOLE &&
3541                     (em->block_start == EXTENT_MAP_HOLE ||
3542                      test_bit(EXTENT_FLAG_PREALLOC, &em->flags)))
3543                         break;
3544                 else if (whence == SEEK_DATA &&
3545                            (em->block_start != EXTENT_MAP_HOLE &&
3546                             !test_bit(EXTENT_FLAG_PREALLOC, &em->flags)))
3547                         break;
3548
3549                 start = em->start + em->len;
3550                 free_extent_map(em);
3551                 em = NULL;
3552                 cond_resched();
3553         }
3554         free_extent_map(em);
3555         unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
3556                              &cached_state);
3557         if (ret) {
3558                 offset = ret;
3559         } else {
3560                 if (whence == SEEK_DATA && start >= i_size)
3561                         offset = -ENXIO;
3562                 else
3563                         offset = min_t(loff_t, start, i_size);
3564         }
3565
3566         return offset;
3567 }
3568
3569 static loff_t btrfs_file_llseek(struct file *file, loff_t offset, int whence)
3570 {
3571         struct inode *inode = file->f_mapping->host;
3572
3573         switch (whence) {
3574         default:
3575                 return generic_file_llseek(file, offset, whence);
3576         case SEEK_DATA:
3577         case SEEK_HOLE:
3578                 btrfs_inode_lock(inode, BTRFS_ILOCK_SHARED);
3579                 offset = find_desired_extent(inode, offset, whence);
3580                 btrfs_inode_unlock(inode, BTRFS_ILOCK_SHARED);
3581                 break;
3582         }
3583
3584         if (offset < 0)
3585                 return offset;
3586
3587         return vfs_setpos(file, offset, inode->i_sb->s_maxbytes);
3588 }
3589
3590 static int btrfs_file_open(struct inode *inode, struct file *filp)
3591 {
3592         filp->f_mode |= FMODE_NOWAIT | FMODE_BUF_RASYNC;
3593         return generic_file_open(inode, filp);
3594 }
3595
3596 static int check_direct_read(struct btrfs_fs_info *fs_info,
3597                              const struct iov_iter *iter, loff_t offset)
3598 {
3599         int ret;
3600         int i, seg;
3601
3602         ret = check_direct_IO(fs_info, iter, offset);
3603         if (ret < 0)
3604                 return ret;
3605
3606         if (!iter_is_iovec(iter))
3607                 return 0;
3608
3609         for (seg = 0; seg < iter->nr_segs; seg++)
3610                 for (i = seg + 1; i < iter->nr_segs; i++)
3611                         if (iter->iov[seg].iov_base == iter->iov[i].iov_base)
3612                                 return -EINVAL;
3613         return 0;
3614 }
3615
3616 static ssize_t btrfs_direct_read(struct kiocb *iocb, struct iov_iter *to)
3617 {
3618         struct inode *inode = file_inode(iocb->ki_filp);
3619         ssize_t ret;
3620
3621         if (check_direct_read(btrfs_sb(inode->i_sb), to, iocb->ki_pos))
3622                 return 0;
3623
3624         btrfs_inode_lock(inode, BTRFS_ILOCK_SHARED);
3625         ret = iomap_dio_rw(iocb, to, &btrfs_dio_iomap_ops, &btrfs_dio_ops,
3626                            is_sync_kiocb(iocb));
3627         btrfs_inode_unlock(inode, BTRFS_ILOCK_SHARED);
3628         return ret;
3629 }
3630
3631 static ssize_t btrfs_file_read_iter(struct kiocb *iocb, struct iov_iter *to)
3632 {
3633         ssize_t ret = 0;
3634
3635         if (iocb->ki_flags & IOCB_DIRECT) {
3636                 ret = btrfs_direct_read(iocb, to);
3637                 if (ret < 0 || !iov_iter_count(to) ||
3638                     iocb->ki_pos >= i_size_read(file_inode(iocb->ki_filp)))
3639                         return ret;
3640         }
3641
3642         return generic_file_buffered_read(iocb, to, ret);
3643 }
3644
3645 const struct file_operations btrfs_file_operations = {
3646         .llseek         = btrfs_file_llseek,
3647         .read_iter      = btrfs_file_read_iter,
3648         .splice_read    = generic_file_splice_read,
3649         .write_iter     = btrfs_file_write_iter,
3650         .splice_write   = iter_file_splice_write,
3651         .mmap           = btrfs_file_mmap,
3652         .open           = btrfs_file_open,
3653         .release        = btrfs_release_file,
3654         .fsync          = btrfs_sync_file,
3655         .fallocate      = btrfs_fallocate,
3656         .unlocked_ioctl = btrfs_ioctl,
3657 #ifdef CONFIG_COMPAT
3658         .compat_ioctl   = btrfs_compat_ioctl,
3659 #endif
3660         .remap_file_range = btrfs_remap_file_range,
3661 };
3662
3663 void __cold btrfs_auto_defrag_exit(void)
3664 {
3665         kmem_cache_destroy(btrfs_inode_defrag_cachep);
3666 }
3667
3668 int __init btrfs_auto_defrag_init(void)
3669 {
3670         btrfs_inode_defrag_cachep = kmem_cache_create("btrfs_inode_defrag",
3671                                         sizeof(struct inode_defrag), 0,
3672                                         SLAB_MEM_SPREAD,
3673                                         NULL);
3674         if (!btrfs_inode_defrag_cachep)
3675                 return -ENOMEM;
3676
3677         return 0;
3678 }
3679
3680 int btrfs_fdatawrite_range(struct inode *inode, loff_t start, loff_t end)
3681 {
3682         int ret;
3683
3684         /*
3685          * So with compression we will find and lock a dirty page and clear the
3686          * first one as dirty, setup an async extent, and immediately return
3687          * with the entire range locked but with nobody actually marked with
3688          * writeback.  So we can't just filemap_write_and_wait_range() and
3689          * expect it to work since it will just kick off a thread to do the
3690          * actual work.  So we need to call filemap_fdatawrite_range _again_
3691          * since it will wait on the page lock, which won't be unlocked until
3692          * after the pages have been marked as writeback and so we're good to go
3693          * from there.  We have to do this otherwise we'll miss the ordered
3694          * extents and that results in badness.  Please Josef, do not think you
3695          * know better and pull this out at some point in the future, it is
3696          * right and you are wrong.
3697          */
3698         ret = filemap_fdatawrite_range(inode->i_mapping, start, end);
3699         if (!ret && test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
3700                              &BTRFS_I(inode)->runtime_flags))
3701                 ret = filemap_fdatawrite_range(inode->i_mapping, start, end);
3702
3703         return ret;
3704 }