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 <linux/fsverity.h>
  20 #include "ctree.h"
  21 #include "disk-io.h"
  22 #include "transaction.h"
  23 #include "btrfs_inode.h"
  24 #include "print-tree.h"
  25 #include "tree-log.h"
  26 #include "locking.h"
  27 #include "volumes.h"
  28 #include "qgroup.h"
  29 #include "compression.h"
  30 #include "delalloc-space.h"
  31 #include "reflink.h"
  32 #include "subpage.h"
  33
  34 static struct kmem_cache *btrfs_inode_defrag_cachep;
  35 /*
  36  * when auto defrag is enabled we
  37  * queue up these defrag structs to remember which
  38  * inodes need defragging passes
  39  */
  40 struct inode_defrag {
  41         struct rb_node rb_node;
  42         /* objectid */
  43         u64 ino;
  44         /*
  45          * transid where the defrag was added, we search for
  46          * extents newer than this
  47          */
  48         u64 transid;
  49
  50         /* root objectid */
  51         u64 root;
  52
  53         /* last offset we were able to defrag */
  54         u64 last_offset;
  55
  56         /* if we've wrapped around back to zero once already */
  57         int cycled;
  58 };
  59
  60 static int __compare_inode_defrag(struct inode_defrag *defrag1,
  61                                   struct inode_defrag *defrag2)
  62 {
  63         if (defrag1->root > defrag2->root)
  64                 return 1;
  65         else if (defrag1->root < defrag2->root)
  66                 return -1;
  67         else if (defrag1->ino > defrag2->ino)
  68                 return 1;
  69         else if (defrag1->ino < defrag2->ino)
  70                 return -1;
  71         else
  72                 return 0;
  73 }
  74
  75 /* pop a record for an inode into the defrag tree.  The lock
  76  * must be held already
  77  *
  78  * If you're inserting a record for an older transid than an
  79  * existing record, the transid already in the tree is lowered
  80  *
  81  * If an existing record is found the defrag item you
  82  * pass in is freed
  83  */
  84 static int __btrfs_add_inode_defrag(struct btrfs_inode *inode,
  85                                     struct inode_defrag *defrag)
  86 {
  87         struct btrfs_fs_info *fs_info = inode->root->fs_info;
  88         struct inode_defrag *entry;
  89         struct rb_node **p;
  90         struct rb_node *parent = NULL;
  91         int ret;
  92
  93         p = &fs_info->defrag_inodes.rb_node;
  94         while (*p) {
  95                 parent = *p;
  96                 entry = rb_entry(parent, struct inode_defrag, rb_node);
  97
  98                 ret = __compare_inode_defrag(defrag, entry);
  99                 if (ret < 0)
 100                         p = &parent->rb_left;
 101                 else if (ret > 0)
 102                         p = &parent->rb_right;
 103                 else {
 104                         /* if we're reinserting an entry for
 105                          * an old defrag run, make sure to
 106                          * lower the transid of our existing record
 107                          */
 108                         if (defrag->transid < entry->transid)
 109                                 entry->transid = defrag->transid;
 110                         if (defrag->last_offset > entry->last_offset)
 111                                 entry->last_offset = defrag->last_offset;
 112                         return -EEXIST;
 113                 }
 114         }
 115         set_bit(BTRFS_INODE_IN_DEFRAG, &inode->runtime_flags);
 116         rb_link_node(&defrag->rb_node, parent, p);
 117         rb_insert_color(&defrag->rb_node, &fs_info->defrag_inodes);
 118         return 0;
 119 }
 120
 121 static inline int __need_auto_defrag(struct btrfs_fs_info *fs_info)
 122 {
 123         if (!btrfs_test_opt(fs_info, AUTO_DEFRAG))
 124                 return 0;
 125
 126         if (btrfs_fs_closing(fs_info))
 127                 return 0;
 128
 129         return 1;
 130 }
 131
 132 /*
 133  * insert a defrag record for this inode if auto defrag is
 134  * enabled
 135  */
 136 int btrfs_add_inode_defrag(struct btrfs_trans_handle *trans,
 137                            struct btrfs_inode *inode)
 138 {
 139         struct btrfs_root *root = inode->root;
 140         struct btrfs_fs_info *fs_info = root->fs_info;
 141         struct inode_defrag *defrag;
 142         u64 transid;
 143         int ret;
 144
 145         if (!__need_auto_defrag(fs_info))
 146                 return 0;
 147
 148         if (test_bit(BTRFS_INODE_IN_DEFRAG, &inode->runtime_flags))
 149                 return 0;
 150
 151         if (trans)
 152                 transid = trans->transid;
 153         else
 154                 transid = inode->root->last_trans;
 155
 156         defrag = kmem_cache_zalloc(btrfs_inode_defrag_cachep, GFP_NOFS);
 157         if (!defrag)
 158                 return -ENOMEM;
 159
 160         defrag->ino = btrfs_ino(inode);
 161         defrag->transid = transid;
 162         defrag->root = root->root_key.objectid;
 163
 164         spin_lock(&fs_info->defrag_inodes_lock);
 165         if (!test_bit(BTRFS_INODE_IN_DEFRAG, &inode->runtime_flags)) {
 166                 /*
 167                  * If we set IN_DEFRAG flag and evict the inode from memory,
 168                  * and then re-read this inode, this new inode doesn't have
 169                  * IN_DEFRAG flag. At the case, we may find the existed defrag.
 170                  */
 171                 ret = __btrfs_add_inode_defrag(inode, defrag);
 172                 if (ret)
 173                         kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
 174         } else {
 175                 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
 176         }
 177         spin_unlock(&fs_info->defrag_inodes_lock);
 178         return 0;
 179 }
 180
 181 /*
 182  * Requeue the defrag object. If there is a defrag object that points to
 183  * the same inode in the tree, we will merge them together (by
 184  * __btrfs_add_inode_defrag()) and free the one that we want to requeue.
 185  */
 186 static void btrfs_requeue_inode_defrag(struct btrfs_inode *inode,
 187                                        struct inode_defrag *defrag)
 188 {
 189         struct btrfs_fs_info *fs_info = inode->root->fs_info;
 190         int ret;
 191
 192         if (!__need_auto_defrag(fs_info))
 193                 goto out;
 194
 195         /*
 196          * Here we don't check the IN_DEFRAG flag, because we need merge
 197          * them together.
 198          */
 199         spin_lock(&fs_info->defrag_inodes_lock);
 200         ret = __btrfs_add_inode_defrag(inode, defrag);
 201         spin_unlock(&fs_info->defrag_inodes_lock);
 202         if (ret)
 203                 goto out;
 204         return;
 205 out:
 206         kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
 207 }
 208
 209 /*
 210  * pick the defragable inode that we want, if it doesn't exist, we will get
 211  * the next one.
 212  */
 213 static struct inode_defrag *
 214 btrfs_pick_defrag_inode(struct btrfs_fs_info *fs_info, u64 root, u64 ino)
 215 {
 216         struct inode_defrag *entry = NULL;
 217         struct inode_defrag tmp;
 218         struct rb_node *p;
 219         struct rb_node *parent = NULL;
 220         int ret;
 221
 222         tmp.ino = ino;
 223         tmp.root = root;
 224
 225         spin_lock(&fs_info->defrag_inodes_lock);
 226         p = fs_info->defrag_inodes.rb_node;
 227         while (p) {
 228                 parent = p;
 229                 entry = rb_entry(parent, struct inode_defrag, rb_node);
 230
 231                 ret = __compare_inode_defrag(&tmp, entry);
 232                 if (ret < 0)
 233                         p = parent->rb_left;
 234                 else if (ret > 0)
 235                         p = parent->rb_right;
 236                 else
 237                         goto out;
 238         }
 239
 240         if (parent && __compare_inode_defrag(&tmp, entry) > 0) {
 241                 parent = rb_next(parent);
 242                 if (parent)
 243                         entry = rb_entry(parent, struct inode_defrag, rb_node);
 244                 else
 245                         entry = NULL;
 246         }
 247 out:
 248         if (entry)
 249                 rb_erase(parent, &fs_info->defrag_inodes);
 250         spin_unlock(&fs_info->defrag_inodes_lock);
 251         return entry;
 252 }
 253
 254 void btrfs_cleanup_defrag_inodes(struct btrfs_fs_info *fs_info)
 255 {
 256         struct inode_defrag *defrag;
 257         struct rb_node *node;
 258
 259         spin_lock(&fs_info->defrag_inodes_lock);
 260         node = rb_first(&fs_info->defrag_inodes);
 261         while (node) {
 262                 rb_erase(node, &fs_info->defrag_inodes);
 263                 defrag = rb_entry(node, struct inode_defrag, rb_node);
 264                 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
 265
 266                 cond_resched_lock(&fs_info->defrag_inodes_lock);
 267
 268                 node = rb_first(&fs_info->defrag_inodes);
 269         }
 270         spin_unlock(&fs_info->defrag_inodes_lock);
 271 }
 272
 273 #define BTRFS_DEFRAG_BATCH      1024
 274
 275 static int __btrfs_run_defrag_inode(struct btrfs_fs_info *fs_info,
 276                                     struct inode_defrag *defrag)
 277 {
 278         struct btrfs_root *inode_root;
 279         struct inode *inode;
 280         struct btrfs_ioctl_defrag_range_args range;
 281         int num_defrag;
 282         int ret;
 283
 284         /* get the inode */
 285         inode_root = btrfs_get_fs_root(fs_info, defrag->root, true);
 286         if (IS_ERR(inode_root)) {
 287                 ret = PTR_ERR(inode_root);
 288                 goto cleanup;
 289         }
 290
 291         inode = btrfs_iget(fs_info->sb, defrag->ino, inode_root);
 292         btrfs_put_root(inode_root);
 293         if (IS_ERR(inode)) {
 294                 ret = PTR_ERR(inode);
 295                 goto cleanup;
 296         }
 297
 298         /* do a chunk of defrag */
 299         clear_bit(BTRFS_INODE_IN_DEFRAG, &BTRFS_I(inode)->runtime_flags);
 300         memset(&range, 0, sizeof(range));
 301         range.len = (u64)-1;
 302         range.start = defrag->last_offset;
 303
 304         sb_start_write(fs_info->sb);
 305         num_defrag = btrfs_defrag_file(inode, NULL, &range, defrag->transid,
 306                                        BTRFS_DEFRAG_BATCH);
 307         sb_end_write(fs_info->sb);
 308         /*
 309          * if we filled the whole defrag batch, there
 310          * must be more work to do.  Queue this defrag
 311          * again
 312          */
 313         if (num_defrag == BTRFS_DEFRAG_BATCH) {
 314                 defrag->last_offset = range.start;
 315                 btrfs_requeue_inode_defrag(BTRFS_I(inode), defrag);
 316         } else if (defrag->last_offset && !defrag->cycled) {
 317                 /*
 318                  * we didn't fill our defrag batch, but
 319                  * we didn't start at zero.  Make sure we loop
 320                  * around to the start of the file.
 321                  */
 322                 defrag->last_offset = 0;
 323                 defrag->cycled = 1;
 324                 btrfs_requeue_inode_defrag(BTRFS_I(inode), defrag);
 325         } else {
 326                 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
 327         }
 328
 329         iput(inode);
 330         return 0;
 331 cleanup:
 332         kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
 333         return ret;
 334 }
 335
 336 /*
 337  * run through the list of inodes in the FS that need
 338  * defragging
 339  */
 340 int btrfs_run_defrag_inodes(struct btrfs_fs_info *fs_info)
 341 {
 342         struct inode_defrag *defrag;
 343         u64 first_ino = 0;
 344         u64 root_objectid = 0;
 345
 346         atomic_inc(&fs_info->defrag_running);
 347         while (1) {
 348                 /* Pause the auto defragger. */
 349                 if (test_bit(BTRFS_FS_STATE_REMOUNTING,
 350                              &fs_info->fs_state))
 351                         break;
 352
 353                 if (!__need_auto_defrag(fs_info))
 354                         break;
 355
 356                 /* find an inode to defrag */
 357                 defrag = btrfs_pick_defrag_inode(fs_info, root_objectid,
 358                                                  first_ino);
 359                 if (!defrag) {
 360                         if (root_objectid || first_ino) {
 361                                 root_objectid = 0;
 362                                 first_ino = 0;
 363                                 continue;
 364                         } else {
 365                                 break;
 366                         }
 367                 }
 368
 369                 first_ino = defrag->ino + 1;
 370                 root_objectid = defrag->root;
 371
 372                 __btrfs_run_defrag_inode(fs_info, defrag);
 373         }
 374         atomic_dec(&fs_info->defrag_running);
 375
 376         /*
 377          * during unmount, we use the transaction_wait queue to
 378          * wait for the defragger to stop
 379          */
 380         wake_up(&fs_info->transaction_wait);
 381         return 0;
 382 }
 383
 384 /* simple helper to fault in pages and copy.  This should go away
 385  * and be replaced with calls into generic code.
 386  */
 387 static noinline int btrfs_copy_from_user(loff_t pos, size_t write_bytes,
 388                                          struct page **prepared_pages,
 389                                          struct iov_iter *i)
 390 {
 391         size_t copied = 0;
 392         size_t total_copied = 0;
 393         int pg = 0;
 394         int offset = offset_in_page(pos);
 395
 396         while (write_bytes > 0) {
 397                 size_t count = min_t(size_t,
 398                                      PAGE_SIZE - offset, write_bytes);
 399                 struct page *page = prepared_pages[pg];
 400                 /*
 401                  * Copy data from userspace to the current page
 402                  */
 403                 copied = copy_page_from_iter_atomic(page, offset, count, i);
 404
 405                 /* Flush processor's dcache for this page */
 406                 flush_dcache_page(page);
 407
 408                 /*
 409                  * if we get a partial write, we can end up with
 410                  * partially up to date pages.  These add
 411                  * a lot of complexity, so make sure they don't
 412                  * happen by forcing this copy to be retried.
 413                  *
 414                  * The rest of the btrfs_file_write code will fall
 415                  * back to page at a time copies after we return 0.
 416                  */
 417                 if (unlikely(copied < count)) {
 418                         if (!PageUptodate(page)) {
 419                                 iov_iter_revert(i, copied);
 420                                 copied = 0;
 421                         }
 422                         if (!copied)
 423                                 break;
 424                 }
 425
 426                 write_bytes -= copied;
 427                 total_copied += copied;
 428                 offset += copied;
 429                 if (offset == PAGE_SIZE) {
 430                         pg++;
 431                         offset = 0;
 432                 }
 433         }
 434         return total_copied;
 435 }
 436
 437 /*
 438  * unlocks pages after btrfs_file_write is done with them
 439  */
 440 static void btrfs_drop_pages(struct page **pages, size_t num_pages)
 441 {
 442         size_t i;
 443         for (i = 0; i < num_pages; i++) {
 444                 /* page checked is some magic around finding pages that
 445                  * have been modified without going through btrfs_set_page_dirty
 446                  * clear it here. There should be no need to mark the pages
 447                  * accessed as prepare_pages should have marked them accessed
 448                  * in prepare_pages via find_or_create_page()
 449                  */
 450                 ClearPageChecked(pages[i]);
 451                 unlock_page(pages[i]);
 452                 put_page(pages[i]);
 453         }
 454 }
 455
 456 /*
 457  * After btrfs_copy_from_user(), update the following things for delalloc:
 458  * - Mark newly dirtied pages as DELALLOC in the io tree.
 459  *   Used to advise which range is to be written back.
 460  * - Mark modified pages as Uptodate/Dirty and not needing COW fixup
 461  * - Update inode size for past EOF write
 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         ASSERT(num_bytes <= U32_MAX);
 487
 488         end_of_last_block = start_pos + num_bytes - 1;
 489
 490         /*
 491          * The pages may have already been dirty, clear out old accounting so
 492          * we can set things up properly
 493          */
 494         clear_extent_bit(&inode->io_tree, start_pos, end_of_last_block,
 495                          EXTENT_DELALLOC | EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG,
 496                          0, 0, cached);
 497
 498         err = btrfs_set_extent_delalloc(inode, start_pos, end_of_last_block,
 499                                         extra_bits, cached);
 500         if (err)
 501                 return err;
 502
 503         for (i = 0; i < num_pages; i++) {
 504                 struct page *p = pages[i];
 505
 506                 btrfs_page_clamp_set_uptodate(fs_info, p, start_pos, num_bytes);
 507                 ClearPageChecked(p);
 508                 btrfs_page_clamp_set_dirty(fs_info, p, start_pos, num_bytes);
 509         }
 510
 511         /*
 512          * we've only changed i_size in ram, and we haven't updated
 513          * the disk i_size.  There is no need to log the inode
 514          * at this time.
 515          */
 516         if (end_pos > isize)
 517                 i_size_write(&inode->vfs_inode, end_pos);
 518         return 0;
 519 }
 520
 521 /*
 522  * this drops all the extents in the cache that intersect the range
 523  * [start, end].  Existing extents are split as required.
 524  */
 525 void btrfs_drop_extent_cache(struct btrfs_inode *inode, u64 start, u64 end,
 526                              int skip_pinned)
 527 {
 528         struct extent_map *em;
 529         struct extent_map *split = NULL;
 530         struct extent_map *split2 = NULL;
 531         struct extent_map_tree *em_tree = &inode->extent_tree;
 532         u64 len = end - start + 1;
 533         u64 gen;
 534         int ret;
 535         int testend = 1;
 536         unsigned long flags;
 537         int compressed = 0;
 538         bool modified;
 539
 540         WARN_ON(end < start);
 541         if (end == (u64)-1) {
 542                 len = (u64)-1;
 543                 testend = 0;
 544         }
 545         while (1) {
 546                 int no_splits = 0;
 547
 548                 modified = false;
 549                 if (!split)
 550                         split = alloc_extent_map();
 551                 if (!split2)
 552                         split2 = alloc_extent_map();
 553                 if (!split || !split2)
 554                         no_splits = 1;
 555
 556                 write_lock(&em_tree->lock);
 557                 em = lookup_extent_mapping(em_tree, start, len);
 558                 if (!em) {
 559                         write_unlock(&em_tree->lock);
 560                         break;
 561                 }
 562                 flags = em->flags;
 563                 gen = em->generation;
 564                 if (skip_pinned && test_bit(EXTENT_FLAG_PINNED, &em->flags)) {
 565                         if (testend && em->start + em->len >= start + len) {
 566                                 free_extent_map(em);
 567                                 write_unlock(&em_tree->lock);
 568                                 break;
 569                         }
 570                         start = em->start + em->len;
 571                         if (testend)
 572                                 len = start + len - (em->start + em->len);
 573                         free_extent_map(em);
 574                         write_unlock(&em_tree->lock);
 575                         continue;
 576                 }
 577                 compressed = test_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
 578                 clear_bit(EXTENT_FLAG_PINNED, &em->flags);
 579                 clear_bit(EXTENT_FLAG_LOGGING, &flags);
 580                 modified = !list_empty(&em->list);
 581                 if (no_splits)
 582                         goto next;
 583
 584                 if (em->start < start) {
 585                         split->start = em->start;
 586                         split->len = start - em->start;
 587
 588                         if (em->block_start < EXTENT_MAP_LAST_BYTE) {
 589                                 split->orig_start = em->orig_start;
 590                                 split->block_start = em->block_start;
 591
 592                                 if (compressed)
 593                                         split->block_len = em->block_len;
 594                                 else
 595                                         split->block_len = split->len;
 596                                 split->orig_block_len = max(split->block_len,
 597                                                 em->orig_block_len);
 598                                 split->ram_bytes = em->ram_bytes;
 599                         } else {
 600                                 split->orig_start = split->start;
 601                                 split->block_len = 0;
 602                                 split->block_start = em->block_start;
 603                                 split->orig_block_len = 0;
 604                                 split->ram_bytes = split->len;
 605                         }
 606
 607                         split->generation = gen;
 608                         split->flags = flags;
 609                         split->compress_type = em->compress_type;
 610                         replace_extent_mapping(em_tree, em, split, modified);
 611                         free_extent_map(split);
 612                         split = split2;
 613                         split2 = NULL;
 614                 }
 615                 if (testend && em->start + em->len > start + len) {
 616                         u64 diff = start + len - em->start;
 617
 618                         split->start = start + len;
 619                         split->len = em->start + em->len - (start + len);
 620                         split->flags = flags;
 621                         split->compress_type = em->compress_type;
 622                         split->generation = gen;
 623
 624                         if (em->block_start < EXTENT_MAP_LAST_BYTE) {
 625                                 split->orig_block_len = max(em->block_len,
 626                                                     em->orig_block_len);
 627
 628                                 split->ram_bytes = em->ram_bytes;
 629                                 if (compressed) {
 630                                         split->block_len = em->block_len;
 631                                         split->block_start = em->block_start;
 632                                         split->orig_start = em->orig_start;
 633                                 } else {
 634                                         split->block_len = split->len;
 635                                         split->block_start = em->block_start
 636                                                 + diff;
 637                                         split->orig_start = em->orig_start;
 638                                 }
 639                         } else {
 640                                 split->ram_bytes = split->len;
 641                                 split->orig_start = split->start;
 642                                 split->block_len = 0;
 643                                 split->block_start = em->block_start;
 644                                 split->orig_block_len = 0;
 645                         }
 646
 647                         if (extent_map_in_tree(em)) {
 648                                 replace_extent_mapping(em_tree, em, split,
 649                                                        modified);
 650                         } else {
 651                                 ret = add_extent_mapping(em_tree, split,
 652                                                          modified);
 653                                 ASSERT(ret == 0); /* Logic error */
 654                         }
 655                         free_extent_map(split);
 656                         split = NULL;
 657                 }
 658 next:
 659                 if (extent_map_in_tree(em))
 660                         remove_extent_mapping(em_tree, em);
 661                 write_unlock(&em_tree->lock);
 662
 663                 /* once for us */
 664                 free_extent_map(em);
 665                 /* once for the tree*/
 666                 free_extent_map(em);
 667         }
 668         if (split)
 669                 free_extent_map(split);
 670         if (split2)
 671                 free_extent_map(split2);
 672 }
 673
 674 /*
 675  * this is very complex, but the basic idea is to drop all extents
 676  * in the range start - end.  hint_block is filled in with a block number
 677  * that would be a good hint to the block allocator for this file.
 678  *
 679  * If an extent intersects the range but is not entirely inside the range
 680  * it is either truncated or split.  Anything entirely inside the range
 681  * is deleted from the tree.
 682  *
 683  * Note: the VFS' inode number of bytes is not updated, it's up to the caller
 684  * to deal with that. We set the field 'bytes_found' of the arguments structure
 685  * with the number of allocated bytes found in the target range, so that the
 686  * caller can update the inode's number of bytes in an atomic way when
 687  * replacing extents in a range to avoid races with stat(2).
 688  */
 689 int btrfs_drop_extents(struct btrfs_trans_handle *trans,
 690                        struct btrfs_root *root, struct btrfs_inode *inode,
 691                        struct btrfs_drop_extents_args *args)
 692 {
 693         struct btrfs_fs_info *fs_info = root->fs_info;
 694         struct extent_buffer *leaf;
 695         struct btrfs_file_extent_item *fi;
 696         struct btrfs_ref ref = { 0 };
 697         struct btrfs_key key;
 698         struct btrfs_key new_key;
 699         u64 ino = btrfs_ino(inode);
 700         u64 search_start = args->start;
 701         u64 disk_bytenr = 0;
 702         u64 num_bytes = 0;
 703         u64 extent_offset = 0;
 704         u64 extent_end = 0;
 705         u64 last_end = args->start;
 706         int del_nr = 0;
 707         int del_slot = 0;
 708         int extent_type;
 709         int recow;
 710         int ret;
 711         int modify_tree = -1;
 712         int update_refs;
 713         int found = 0;
 714         int leafs_visited = 0;
 715         struct btrfs_path *path = args->path;
 716
 717         args->bytes_found = 0;
 718         args->extent_inserted = false;
 719
 720         /* Must always have a path if ->replace_extent is true */
 721         ASSERT(!(args->replace_extent && !args->path));
 722
 723         if (!path) {
 724                 path = btrfs_alloc_path();
 725                 if (!path) {
 726                         ret = -ENOMEM;
 727                         goto out;
 728                 }
 729         }
 730
 731         if (args->drop_cache)
 732                 btrfs_drop_extent_cache(inode, args->start, args->end - 1, 0);
 733
 734         if (args->start >= inode->disk_i_size && !args->replace_extent)
 735                 modify_tree = 0;
 736
 737         update_refs = (test_bit(BTRFS_ROOT_SHAREABLE, &root->state) ||
 738                        root == fs_info->tree_root);
 739         while (1) {
 740                 recow = 0;
 741                 ret = btrfs_lookup_file_extent(trans, root, path, ino,
 742                                                search_start, modify_tree);
 743                 if (ret < 0)
 744                         break;
 745                 if (ret > 0 && path->slots[0] > 0 && search_start == args->start) {
 746                         leaf = path->nodes[0];
 747                         btrfs_item_key_to_cpu(leaf, &key, path->slots[0] - 1);
 748                         if (key.objectid == ino &&
 749                             key.type == BTRFS_EXTENT_DATA_KEY)
 750                                 path->slots[0]--;
 751                 }
 752                 ret = 0;
 753                 leafs_visited++;
 754 next_slot:
 755                 leaf = path->nodes[0];
 756                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
 757                         BUG_ON(del_nr > 0);
 758                         ret = btrfs_next_leaf(root, path);
 759                         if (ret < 0)
 760                                 break;
 761                         if (ret > 0) {
 762                                 ret = 0;
 763                                 break;
 764                         }
 765                         leafs_visited++;
 766                         leaf = path->nodes[0];
 767                         recow = 1;
 768                 }
 769
 770                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 771
 772                 if (key.objectid > ino)
 773                         break;
 774                 if (WARN_ON_ONCE(key.objectid < ino) ||
 775                     key.type < BTRFS_EXTENT_DATA_KEY) {
 776                         ASSERT(del_nr == 0);
 777                         path->slots[0]++;
 778                         goto next_slot;
 779                 }
 780                 if (key.type > BTRFS_EXTENT_DATA_KEY || key.offset >= args->end)
 781                         break;
 782
 783                 fi = btrfs_item_ptr(leaf, path->slots[0],
 784                                     struct btrfs_file_extent_item);
 785                 extent_type = btrfs_file_extent_type(leaf, fi);
 786
 787                 if (extent_type == BTRFS_FILE_EXTENT_REG ||
 788                     extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
 789                         disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
 790                         num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
 791                         extent_offset = btrfs_file_extent_offset(leaf, fi);
 792                         extent_end = key.offset +
 793                                 btrfs_file_extent_num_bytes(leaf, fi);
 794                 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
 795                         extent_end = key.offset +
 796                                 btrfs_file_extent_ram_bytes(leaf, fi);
 797                 } else {
 798                         /* can't happen */
 799                         BUG();
 800                 }
 801
 802                 /*
 803                  * Don't skip extent items representing 0 byte lengths. They
 804                  * used to be created (bug) if while punching holes we hit
 805                  * -ENOSPC condition. So if we find one here, just ensure we
 806                  * delete it, otherwise we would insert a new file extent item
 807                  * with the same key (offset) as that 0 bytes length file
 808                  * extent item in the call to setup_items_for_insert() later
 809                  * in this function.
 810                  */
 811                 if (extent_end == key.offset && extent_end >= search_start) {
 812                         last_end = extent_end;
 813                         goto delete_extent_item;
 814                 }
 815
 816                 if (extent_end <= search_start) {
 817                         path->slots[0]++;
 818                         goto next_slot;
 819                 }
 820
 821                 found = 1;
 822                 search_start = max(key.offset, args->start);
 823                 if (recow || !modify_tree) {
 824                         modify_tree = -1;
 825                         btrfs_release_path(path);
 826                         continue;
 827                 }
 828
 829                 /*
 830                  *     | - range to drop - |
 831                  *  | -------- extent -------- |
 832                  */
 833                 if (args->start > key.offset && args->end < extent_end) {
 834                         BUG_ON(del_nr > 0);
 835                         if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
 836                                 ret = -EOPNOTSUPP;
 837                                 break;
 838                         }
 839
 840                         memcpy(&new_key, &key, sizeof(new_key));
 841                         new_key.offset = args->start;
 842                         ret = btrfs_duplicate_item(trans, root, path,
 843                                                    &new_key);
 844                         if (ret == -EAGAIN) {
 845                                 btrfs_release_path(path);
 846                                 continue;
 847                         }
 848                         if (ret < 0)
 849                                 break;
 850
 851                         leaf = path->nodes[0];
 852                         fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
 853                                             struct btrfs_file_extent_item);
 854                         btrfs_set_file_extent_num_bytes(leaf, fi,
 855                                                         args->start - key.offset);
 856
 857                         fi = btrfs_item_ptr(leaf, path->slots[0],
 858                                             struct btrfs_file_extent_item);
 859
 860                         extent_offset += args->start - key.offset;
 861                         btrfs_set_file_extent_offset(leaf, fi, extent_offset);
 862                         btrfs_set_file_extent_num_bytes(leaf, fi,
 863                                                         extent_end - args->start);
 864                         btrfs_mark_buffer_dirty(leaf);
 865
 866                         if (update_refs && disk_bytenr > 0) {
 867                                 btrfs_init_generic_ref(&ref,
 868                                                 BTRFS_ADD_DELAYED_REF,
 869                                                 disk_bytenr, num_bytes, 0);
 870                                 btrfs_init_data_ref(&ref,
 871                                                 root->root_key.objectid,
 872                                                 new_key.objectid,
 873                                                 args->start - extent_offset);
 874                                 ret = btrfs_inc_extent_ref(trans, &ref);
 875                                 BUG_ON(ret); /* -ENOMEM */
 876                         }
 877                         key.offset = args->start;
 878                 }
 879                 /*
 880                  * From here on out we will have actually dropped something, so
 881                  * last_end can be updated.
 882                  */
 883                 last_end = extent_end;
 884
 885                 /*
 886                  *  | ---- range to drop ----- |
 887                  *      | -------- extent -------- |
 888                  */
 889                 if (args->start <= key.offset && args->end < extent_end) {
 890                         if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
 891                                 ret = -EOPNOTSUPP;
 892                                 break;
 893                         }
 894
 895                         memcpy(&new_key, &key, sizeof(new_key));
 896                         new_key.offset = args->end;
 897                         btrfs_set_item_key_safe(fs_info, path, &new_key);
 898
 899                         extent_offset += args->end - key.offset;
 900                         btrfs_set_file_extent_offset(leaf, fi, extent_offset);
 901                         btrfs_set_file_extent_num_bytes(leaf, fi,
 902                                                         extent_end - args->end);
 903                         btrfs_mark_buffer_dirty(leaf);
 904                         if (update_refs && disk_bytenr > 0)
 905                                 args->bytes_found += args->end - key.offset;
 906                         break;
 907                 }
 908
 909                 search_start = extent_end;
 910                 /*
 911                  *       | ---- range to drop ----- |
 912                  *  | -------- extent -------- |
 913                  */
 914                 if (args->start > key.offset && args->end >= extent_end) {
 915                         BUG_ON(del_nr > 0);
 916                         if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
 917                                 ret = -EOPNOTSUPP;
 918                                 break;
 919                         }
 920
 921                         btrfs_set_file_extent_num_bytes(leaf, fi,
 922                                                         args->start - key.offset);
 923                         btrfs_mark_buffer_dirty(leaf);
 924                         if (update_refs && disk_bytenr > 0)
 925                                 args->bytes_found += extent_end - args->start;
 926                         if (args->end == extent_end)
 927                                 break;
 928
 929                         path->slots[0]++;
 930                         goto next_slot;
 931                 }
 932
 933                 /*
 934                  *  | ---- range to drop ----- |
 935                  *    | ------ extent ------ |
 936                  */
 937                 if (args->start <= key.offset && args->end >= extent_end) {
 938 delete_extent_item:
 939                         if (del_nr == 0) {
 940                                 del_slot = path->slots[0];
 941                                 del_nr = 1;
 942                         } else {
 943                                 BUG_ON(del_slot + del_nr != path->slots[0]);
 944                                 del_nr++;
 945                         }
 946
 947                         if (update_refs &&
 948                             extent_type == BTRFS_FILE_EXTENT_INLINE) {
 949                                 args->bytes_found += extent_end - key.offset;
 950                                 extent_end = ALIGN(extent_end,
 951                                                    fs_info->sectorsize);
 952                         } else if (update_refs && disk_bytenr > 0) {
 953                                 btrfs_init_generic_ref(&ref,
 954                                                 BTRFS_DROP_DELAYED_REF,
 955                                                 disk_bytenr, num_bytes, 0);
 956                                 btrfs_init_data_ref(&ref,
 957                                                 root->root_key.objectid,
 958                                                 key.objectid,
 959                                                 key.offset - extent_offset);
 960                                 ret = btrfs_free_extent(trans, &ref);
 961                                 BUG_ON(ret); /* -ENOMEM */
 962                                 args->bytes_found += extent_end - key.offset;
 963                         }
 964
 965                         if (args->end == extent_end)
 966                                 break;
 967
 968                         if (path->slots[0] + 1 < btrfs_header_nritems(leaf)) {
 969                                 path->slots[0]++;
 970                                 goto next_slot;
 971                         }
 972
 973                         ret = btrfs_del_items(trans, root, path, del_slot,
 974                                               del_nr);
 975                         if (ret) {
 976                                 btrfs_abort_transaction(trans, ret);
 977                                 break;
 978                         }
 979
 980                         del_nr = 0;
 981                         del_slot = 0;
 982
 983                         btrfs_release_path(path);
 984                         continue;
 985                 }
 986
 987                 BUG();
 988         }
 989
 990         if (!ret && del_nr > 0) {
 991                 /*
 992                  * Set path->slots[0] to first slot, so that after the delete
 993                  * if items are move off from our leaf to its immediate left or
 994                  * right neighbor leafs, we end up with a correct and adjusted
 995                  * path->slots[0] for our insertion (if args->replace_extent).
 996                  */
 997                 path->slots[0] = del_slot;
 998                 ret = btrfs_del_items(trans, root, path, del_slot, del_nr);
 999                 if (ret)
1000                         btrfs_abort_transaction(trans, ret);
1001         }
1002
1003         leaf = path->nodes[0];
1004         /*
1005          * If btrfs_del_items() was called, it might have deleted a leaf, in
1006          * which case it unlocked our path, so check path->locks[0] matches a
1007          * write lock.
1008          */
1009         if (!ret && args->replace_extent && leafs_visited == 1 &&
1010             path->locks[0] == BTRFS_WRITE_LOCK &&
1011             btrfs_leaf_free_space(leaf) >=
1012             sizeof(struct btrfs_item) + args->extent_item_size) {
1013
1014                 key.objectid = ino;
1015                 key.type = BTRFS_EXTENT_DATA_KEY;
1016                 key.offset = args->start;
1017                 if (!del_nr && path->slots[0] < btrfs_header_nritems(leaf)) {
1018                         struct btrfs_key slot_key;
1019
1020                         btrfs_item_key_to_cpu(leaf, &slot_key, path->slots[0]);
1021                         if (btrfs_comp_cpu_keys(&key, &slot_key) > 0)
1022                                 path->slots[0]++;
1023                 }
1024                 setup_items_for_insert(root, path, &key,
1025                                        &args->extent_item_size, 1);
1026                 args->extent_inserted = true;
1027         }
1028
1029         if (!args->path)
1030                 btrfs_free_path(path);
1031         else if (!args->extent_inserted)
1032                 btrfs_release_path(path);
1033 out:
1034         args->drop_end = found ? min(args->end, last_end) : args->end;
1035
1036         return ret;
1037 }
1038
1039 static int extent_mergeable(struct extent_buffer *leaf, int slot,
1040                             u64 objectid, u64 bytenr, u64 orig_offset,
1041                             u64 *start, u64 *end)
1042 {
1043         struct btrfs_file_extent_item *fi;
1044         struct btrfs_key key;
1045         u64 extent_end;
1046
1047         if (slot < 0 || slot >= btrfs_header_nritems(leaf))
1048                 return 0;
1049
1050         btrfs_item_key_to_cpu(leaf, &key, slot);
1051         if (key.objectid != objectid || key.type != BTRFS_EXTENT_DATA_KEY)
1052                 return 0;
1053
1054         fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
1055         if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG ||
1056             btrfs_file_extent_disk_bytenr(leaf, fi) != bytenr ||
1057             btrfs_file_extent_offset(leaf, fi) != key.offset - orig_offset ||
1058             btrfs_file_extent_compression(leaf, fi) ||
1059             btrfs_file_extent_encryption(leaf, fi) ||
1060             btrfs_file_extent_other_encoding(leaf, fi))
1061                 return 0;
1062
1063         extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
1064         if ((*start && *start != key.offset) || (*end && *end != extent_end))
1065                 return 0;
1066
1067         *start = key.offset;
1068         *end = extent_end;
1069         return 1;
1070 }
1071
1072 /*
1073  * Mark extent in the range start - end as written.
1074  *
1075  * This changes extent type from 'pre-allocated' to 'regular'. If only
1076  * part of extent is marked as written, the extent will be split into
1077  * two or three.
1078  */
1079 int btrfs_mark_extent_written(struct btrfs_trans_handle *trans,
1080                               struct btrfs_inode *inode, u64 start, u64 end)
1081 {
1082         struct btrfs_fs_info *fs_info = trans->fs_info;
1083         struct btrfs_root *root = inode->root;
1084         struct extent_buffer *leaf;
1085         struct btrfs_path *path;
1086         struct btrfs_file_extent_item *fi;
1087         struct btrfs_ref ref = { 0 };
1088         struct btrfs_key key;
1089         struct btrfs_key new_key;
1090         u64 bytenr;
1091         u64 num_bytes;
1092         u64 extent_end;
1093         u64 orig_offset;
1094         u64 other_start;
1095         u64 other_end;
1096         u64 split;
1097         int del_nr = 0;
1098         int del_slot = 0;
1099         int recow;
1100         int ret = 0;
1101         u64 ino = btrfs_ino(inode);
1102
1103         path = btrfs_alloc_path();
1104         if (!path)
1105                 return -ENOMEM;
1106 again:
1107         recow = 0;
1108         split = start;
1109         key.objectid = ino;
1110         key.type = BTRFS_EXTENT_DATA_KEY;
1111         key.offset = split;
1112
1113         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1114         if (ret < 0)
1115                 goto out;
1116         if (ret > 0 && path->slots[0] > 0)
1117                 path->slots[0]--;
1118
1119         leaf = path->nodes[0];
1120         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1121         if (key.objectid != ino ||
1122             key.type != BTRFS_EXTENT_DATA_KEY) {
1123                 ret = -EINVAL;
1124                 btrfs_abort_transaction(trans, ret);
1125                 goto out;
1126         }
1127         fi = btrfs_item_ptr(leaf, path->slots[0],
1128                             struct btrfs_file_extent_item);
1129         if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_PREALLOC) {
1130                 ret = -EINVAL;
1131                 btrfs_abort_transaction(trans, ret);
1132                 goto out;
1133         }
1134         extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
1135         if (key.offset > start || extent_end < end) {
1136                 ret = -EINVAL;
1137                 btrfs_abort_transaction(trans, ret);
1138                 goto out;
1139         }
1140
1141         bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1142         num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
1143         orig_offset = key.offset - btrfs_file_extent_offset(leaf, fi);
1144         memcpy(&new_key, &key, sizeof(new_key));
1145
1146         if (start == key.offset && end < extent_end) {
1147                 other_start = 0;
1148                 other_end = start;
1149                 if (extent_mergeable(leaf, path->slots[0] - 1,
1150                                      ino, bytenr, orig_offset,
1151                                      &other_start, &other_end)) {
1152                         new_key.offset = end;
1153                         btrfs_set_item_key_safe(fs_info, path, &new_key);
1154                         fi = btrfs_item_ptr(leaf, path->slots[0],
1155                                             struct btrfs_file_extent_item);
1156                         btrfs_set_file_extent_generation(leaf, fi,
1157                                                          trans->transid);
1158                         btrfs_set_file_extent_num_bytes(leaf, fi,
1159                                                         extent_end - end);
1160                         btrfs_set_file_extent_offset(leaf, fi,
1161                                                      end - orig_offset);
1162                         fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
1163                                             struct btrfs_file_extent_item);
1164                         btrfs_set_file_extent_generation(leaf, fi,
1165                                                          trans->transid);
1166                         btrfs_set_file_extent_num_bytes(leaf, fi,
1167                                                         end - other_start);
1168                         btrfs_mark_buffer_dirty(leaf);
1169                         goto out;
1170                 }
1171         }
1172
1173         if (start > key.offset && end == extent_end) {
1174                 other_start = end;
1175                 other_end = 0;
1176                 if (extent_mergeable(leaf, path->slots[0] + 1,
1177                                      ino, bytenr, orig_offset,
1178                                      &other_start, &other_end)) {
1179                         fi = btrfs_item_ptr(leaf, path->slots[0],
1180                                             struct btrfs_file_extent_item);
1181                         btrfs_set_file_extent_num_bytes(leaf, fi,
1182                                                         start - key.offset);
1183                         btrfs_set_file_extent_generation(leaf, fi,
1184                                                          trans->transid);
1185                         path->slots[0]++;
1186                         new_key.offset = start;
1187                         btrfs_set_item_key_safe(fs_info, path, &new_key);
1188
1189                         fi = btrfs_item_ptr(leaf, path->slots[0],
1190                                             struct btrfs_file_extent_item);
1191                         btrfs_set_file_extent_generation(leaf, fi,
1192                                                          trans->transid);
1193                         btrfs_set_file_extent_num_bytes(leaf, fi,
1194                                                         other_end - start);
1195                         btrfs_set_file_extent_offset(leaf, fi,
1196                                                      start - orig_offset);
1197                         btrfs_mark_buffer_dirty(leaf);
1198                         goto out;
1199                 }
1200         }
1201
1202         while (start > key.offset || end < extent_end) {
1203                 if (key.offset == start)
1204                         split = end;
1205
1206                 new_key.offset = split;
1207                 ret = btrfs_duplicate_item(trans, root, path, &new_key);
1208                 if (ret == -EAGAIN) {
1209                         btrfs_release_path(path);
1210                         goto again;
1211                 }
1212                 if (ret < 0) {
1213                         btrfs_abort_transaction(trans, ret);
1214                         goto out;
1215                 }
1216
1217                 leaf = path->nodes[0];
1218                 fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
1219                                     struct btrfs_file_extent_item);
1220                 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1221                 btrfs_set_file_extent_num_bytes(leaf, fi,
1222                                                 split - key.offset);
1223
1224                 fi = btrfs_item_ptr(leaf, path->slots[0],
1225                                     struct btrfs_file_extent_item);
1226
1227                 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1228                 btrfs_set_file_extent_offset(leaf, fi, split - orig_offset);
1229                 btrfs_set_file_extent_num_bytes(leaf, fi,
1230                                                 extent_end - split);
1231                 btrfs_mark_buffer_dirty(leaf);
1232
1233                 btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF, bytenr,
1234                                        num_bytes, 0);
1235                 btrfs_init_data_ref(&ref, root->root_key.objectid, ino,
1236                                     orig_offset);
1237                 ret = btrfs_inc_extent_ref(trans, &ref);
1238                 if (ret) {
1239                         btrfs_abort_transaction(trans, ret);
1240                         goto out;
1241                 }
1242
1243                 if (split == start) {
1244                         key.offset = start;
1245                 } else {
1246                         if (start != key.offset) {
1247                                 ret = -EINVAL;
1248                                 btrfs_abort_transaction(trans, ret);
1249                                 goto out;
1250                         }
1251                         path->slots[0]--;
1252                         extent_end = end;
1253                 }
1254                 recow = 1;
1255         }
1256
1257         other_start = end;
1258         other_end = 0;
1259         btrfs_init_generic_ref(&ref, BTRFS_DROP_DELAYED_REF, bytenr,
1260                                num_bytes, 0);
1261         btrfs_init_data_ref(&ref, root->root_key.objectid, ino, orig_offset);
1262         if (extent_mergeable(leaf, path->slots[0] + 1,
1263                              ino, bytenr, orig_offset,
1264                              &other_start, &other_end)) {
1265                 if (recow) {
1266                         btrfs_release_path(path);
1267                         goto again;
1268                 }
1269                 extent_end = other_end;
1270                 del_slot = path->slots[0] + 1;
1271                 del_nr++;
1272                 ret = btrfs_free_extent(trans, &ref);
1273                 if (ret) {
1274                         btrfs_abort_transaction(trans, ret);
1275                         goto out;
1276                 }
1277         }
1278         other_start = 0;
1279         other_end = start;
1280         if (extent_mergeable(leaf, path->slots[0] - 1,
1281                              ino, bytenr, orig_offset,
1282                              &other_start, &other_end)) {
1283                 if (recow) {
1284                         btrfs_release_path(path);
1285                         goto again;
1286                 }
1287                 key.offset = other_start;
1288                 del_slot = path->slots[0];
1289                 del_nr++;
1290                 ret = btrfs_free_extent(trans, &ref);
1291                 if (ret) {
1292                         btrfs_abort_transaction(trans, ret);
1293                         goto out;
1294                 }
1295         }
1296         if (del_nr == 0) {
1297                 fi = btrfs_item_ptr(leaf, path->slots[0],
1298                            struct btrfs_file_extent_item);
1299                 btrfs_set_file_extent_type(leaf, fi,
1300                                            BTRFS_FILE_EXTENT_REG);
1301                 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1302                 btrfs_mark_buffer_dirty(leaf);
1303         } else {
1304                 fi = btrfs_item_ptr(leaf, del_slot - 1,
1305                            struct btrfs_file_extent_item);
1306                 btrfs_set_file_extent_type(leaf, fi,
1307                                            BTRFS_FILE_EXTENT_REG);
1308                 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1309                 btrfs_set_file_extent_num_bytes(leaf, fi,
1310                                                 extent_end - key.offset);
1311                 btrfs_mark_buffer_dirty(leaf);
1312
1313                 ret = btrfs_del_items(trans, root, path, del_slot, del_nr);
1314                 if (ret < 0) {
1315                         btrfs_abort_transaction(trans, ret);
1316                         goto out;
1317                 }
1318         }
1319 out:
1320         btrfs_free_path(path);
1321         return ret;
1322 }
1323
1324 /*
1325  * on error we return an unlocked page and the error value
1326  * on success we return a locked page and 0
1327  */
1328 static int prepare_uptodate_page(struct inode *inode,
1329                                  struct page *page, u64 pos,
1330                                  bool force_uptodate)
1331 {
1332         int ret = 0;
1333
1334         if (((pos & (PAGE_SIZE - 1)) || force_uptodate) &&
1335             !PageUptodate(page)) {
1336                 ret = btrfs_readpage(NULL, page);
1337                 if (ret)
1338                         return ret;
1339                 lock_page(page);
1340                 if (!PageUptodate(page)) {
1341                         unlock_page(page);
1342                         return -EIO;
1343                 }
1344
1345                 /*
1346                  * Since btrfs_readpage() will unlock the page before it
1347                  * returns, there is a window where btrfs_releasepage() can be
1348                  * called to release the page.  Here we check both inode
1349                  * mapping and PagePrivate() to make sure the page was not
1350                  * released.
1351                  *
1352                  * The private flag check is essential for subpage as we need
1353                  * to store extra bitmap using page->private.
1354                  */
1355                 if (page->mapping != inode->i_mapping || !PagePrivate(page)) {
1356                         unlock_page(page);
1357                         return -EAGAIN;
1358                 }
1359         }
1360         return 0;
1361 }
1362
1363 /*
1364  * this just gets pages into the page cache and locks them down.
1365  */
1366 static noinline int prepare_pages(struct inode *inode, struct page **pages,
1367                                   size_t num_pages, loff_t pos,
1368                                   size_t write_bytes, bool force_uptodate)
1369 {
1370         int i;
1371         unsigned long index = pos >> PAGE_SHIFT;
1372         gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping);
1373         int err = 0;
1374         int faili;
1375
1376         for (i = 0; i < num_pages; i++) {
1377 again:
1378                 pages[i] = find_or_create_page(inode->i_mapping, index + i,
1379                                                mask | __GFP_WRITE);
1380                 if (!pages[i]) {
1381                         faili = i - 1;
1382                         err = -ENOMEM;
1383                         goto fail;
1384                 }
1385
1386                 err = set_page_extent_mapped(pages[i]);
1387                 if (err < 0) {
1388                         faili = i;
1389                         goto fail;
1390                 }
1391
1392                 if (i == 0)
1393                         err = prepare_uptodate_page(inode, pages[i], pos,
1394                                                     force_uptodate);
1395                 if (!err && i == num_pages - 1)
1396                         err = prepare_uptodate_page(inode, pages[i],
1397                                                     pos + write_bytes, false);
1398                 if (err) {
1399                         put_page(pages[i]);
1400                         if (err == -EAGAIN) {
1401                                 err = 0;
1402                                 goto again;
1403                         }
1404                         faili = i - 1;
1405                         goto fail;
1406                 }
1407                 wait_on_page_writeback(pages[i]);
1408         }
1409
1410         return 0;
1411 fail:
1412         while (faili >= 0) {
1413                 unlock_page(pages[faili]);
1414                 put_page(pages[faili]);
1415                 faili--;
1416         }
1417         return err;
1418
1419 }
1420
1421 /*
1422  * This function locks the extent and properly waits for data=ordered extents
1423  * to finish before allowing the pages to be modified if need.
1424  *
1425  * The return value:
1426  * 1 - the extent is locked
1427  * 0 - the extent is not locked, and everything is OK
1428  * -EAGAIN - need re-prepare the pages
1429  * the other < 0 number - Something wrong happens
1430  */
1431 static noinline int
1432 lock_and_cleanup_extent_if_need(struct btrfs_inode *inode, struct page **pages,
1433                                 size_t num_pages, loff_t pos,
1434                                 size_t write_bytes,
1435                                 u64 *lockstart, u64 *lockend,
1436                                 struct extent_state **cached_state)
1437 {
1438         struct btrfs_fs_info *fs_info = inode->root->fs_info;
1439         u64 start_pos;
1440         u64 last_pos;
1441         int i;
1442         int ret = 0;
1443
1444         start_pos = round_down(pos, fs_info->sectorsize);
1445         last_pos = round_up(pos + write_bytes, fs_info->sectorsize) - 1;
1446
1447         if (start_pos < inode->vfs_inode.i_size) {
1448                 struct btrfs_ordered_extent *ordered;
1449
1450                 lock_extent_bits(&inode->io_tree, start_pos, last_pos,
1451                                 cached_state);
1452                 ordered = btrfs_lookup_ordered_range(inode, start_pos,
1453                                                      last_pos - start_pos + 1);
1454                 if (ordered &&
1455                     ordered->file_offset + ordered->num_bytes > start_pos &&
1456                     ordered->file_offset <= last_pos) {
1457                         unlock_extent_cached(&inode->io_tree, start_pos,
1458                                         last_pos, cached_state);
1459                         for (i = 0; i < num_pages; i++) {
1460                                 unlock_page(pages[i]);
1461                                 put_page(pages[i]);
1462                         }
1463                         btrfs_start_ordered_extent(ordered, 1);
1464                         btrfs_put_ordered_extent(ordered);
1465                         return -EAGAIN;
1466                 }
1467                 if (ordered)
1468                         btrfs_put_ordered_extent(ordered);
1469
1470                 *lockstart = start_pos;
1471                 *lockend = last_pos;
1472                 ret = 1;
1473         }
1474
1475         /*
1476          * We should be called after prepare_pages() which should have locked
1477          * all pages in the range.
1478          */
1479         for (i = 0; i < num_pages; i++)
1480                 WARN_ON(!PageLocked(pages[i]));
1481
1482         return ret;
1483 }
1484
1485 static int check_can_nocow(struct btrfs_inode *inode, loff_t pos,
1486                            size_t *write_bytes, bool nowait)
1487 {
1488         struct btrfs_fs_info *fs_info = inode->root->fs_info;
1489         struct btrfs_root *root = inode->root;
1490         u64 lockstart, lockend;
1491         u64 num_bytes;
1492         int ret;
1493
1494         if (!(inode->flags & (BTRFS_INODE_NODATACOW | BTRFS_INODE_PREALLOC)))
1495                 return 0;
1496
1497         if (!nowait && !btrfs_drew_try_write_lock(&root->snapshot_lock))
1498                 return -EAGAIN;
1499
1500         lockstart = round_down(pos, fs_info->sectorsize);
1501         lockend = round_up(pos + *write_bytes,
1502                            fs_info->sectorsize) - 1;
1503         num_bytes = lockend - lockstart + 1;
1504
1505         if (nowait) {
1506                 struct btrfs_ordered_extent *ordered;
1507
1508                 if (!try_lock_extent(&inode->io_tree, lockstart, lockend))
1509                         return -EAGAIN;
1510
1511                 ordered = btrfs_lookup_ordered_range(inode, lockstart,
1512                                                      num_bytes);
1513                 if (ordered) {
1514                         btrfs_put_ordered_extent(ordered);
1515                         ret = -EAGAIN;
1516                         goto out_unlock;
1517                 }
1518         } else {
1519                 btrfs_lock_and_flush_ordered_range(inode, lockstart,
1520                                                    lockend, NULL);
1521         }
1522
1523         ret = can_nocow_extent(&inode->vfs_inode, lockstart, &num_bytes,
1524                         NULL, NULL, NULL, false);
1525         if (ret <= 0) {
1526                 ret = 0;
1527                 if (!nowait)
1528                         btrfs_drew_write_unlock(&root->snapshot_lock);
1529         } else {
1530                 *write_bytes = min_t(size_t, *write_bytes ,
1531                                      num_bytes - pos + lockstart);
1532         }
1533 out_unlock:
1534         unlock_extent(&inode->io_tree, lockstart, lockend);
1535
1536         return ret;
1537 }
1538
1539 static int check_nocow_nolock(struct btrfs_inode *inode, loff_t pos,
1540                               size_t *write_bytes)
1541 {
1542         return check_can_nocow(inode, pos, write_bytes, true);
1543 }
1544
1545 /*
1546  * Check if we can do nocow write into the range [@pos, @pos + @write_bytes)
1547  *
1548  * @pos:         File offset
1549  * @write_bytes: The length to write, will be updated to the nocow writeable
1550  *               range
1551  *
1552  * This function will flush ordered extents in the range to ensure proper
1553  * nocow checks.
1554  *
1555  * Return:
1556  * >0           and update @write_bytes if we can do nocow write
1557  *  0           if we can't do nocow write
1558  * -EAGAIN      if we can't get the needed lock or there are ordered extents
1559  *              for * (nowait == true) case
1560  * <0           if other error happened
1561  *
1562  * NOTE: Callers need to release the lock by btrfs_check_nocow_unlock().
1563  */
1564 int btrfs_check_nocow_lock(struct btrfs_inode *inode, loff_t pos,
1565                            size_t *write_bytes)
1566 {
1567         return check_can_nocow(inode, pos, write_bytes, false);
1568 }
1569
1570 void btrfs_check_nocow_unlock(struct btrfs_inode *inode)
1571 {
1572         btrfs_drew_write_unlock(&inode->root->snapshot_lock);
1573 }
1574
1575 static void update_time_for_write(struct inode *inode)
1576 {
1577         struct timespec64 now;
1578
1579         if (IS_NOCMTIME(inode))
1580                 return;
1581
1582         now = current_time(inode);
1583         if (!timespec64_equal(&inode->i_mtime, &now))
1584                 inode->i_mtime = now;
1585
1586         if (!timespec64_equal(&inode->i_ctime, &now))
1587                 inode->i_ctime = now;
1588
1589         if (IS_I_VERSION(inode))
1590                 inode_inc_iversion(inode);
1591 }
1592
1593 static int btrfs_write_check(struct kiocb *iocb, struct iov_iter *from,
1594                              size_t count)
1595 {
1596         struct file *file = iocb->ki_filp;
1597         struct inode *inode = file_inode(file);
1598         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1599         loff_t pos = iocb->ki_pos;
1600         int ret;
1601         loff_t oldsize;
1602         loff_t start_pos;
1603
1604         if (iocb->ki_flags & IOCB_NOWAIT) {
1605                 size_t nocow_bytes = count;
1606
1607                 /* We will allocate space in case nodatacow is not set, so bail */
1608                 if (check_nocow_nolock(BTRFS_I(inode), pos, &nocow_bytes) <= 0)
1609                         return -EAGAIN;
1610                 /*
1611                  * There are holes in the range or parts of the range that must
1612                  * be COWed (shared extents, RO block groups, etc), so just bail
1613                  * out.
1614                  */
1615                 if (nocow_bytes < count)
1616                         return -EAGAIN;
1617         }
1618
1619         current->backing_dev_info = inode_to_bdi(inode);
1620         ret = file_remove_privs(file);
1621         if (ret)
1622                 return ret;
1623
1624         /*
1625          * We reserve space for updating the inode when we reserve space for the
1626          * extent we are going to write, so we will enospc out there.  We don't
1627          * need to start yet another transaction to update the inode as we will
1628          * update the inode when we finish writing whatever data we write.
1629          */
1630         update_time_for_write(inode);
1631
1632         start_pos = round_down(pos, fs_info->sectorsize);
1633         oldsize = i_size_read(inode);
1634         if (start_pos > oldsize) {
1635                 /* Expand hole size to cover write data, preventing empty gap */
1636                 loff_t end_pos = round_up(pos + count, fs_info->sectorsize);
1637
1638                 ret = btrfs_cont_expand(BTRFS_I(inode), oldsize, end_pos);
1639                 if (ret) {
1640                         current->backing_dev_info = NULL;
1641                         return ret;
1642                 }
1643         }
1644
1645         return 0;
1646 }
1647
1648 static noinline ssize_t btrfs_buffered_write(struct kiocb *iocb,
1649                                                struct iov_iter *i)
1650 {
1651         struct file *file = iocb->ki_filp;
1652         loff_t pos;
1653         struct inode *inode = file_inode(file);
1654         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1655         struct page **pages = NULL;
1656         struct extent_changeset *data_reserved = NULL;
1657         u64 release_bytes = 0;
1658         u64 lockstart;
1659         u64 lockend;
1660         size_t num_written = 0;
1661         int nrptrs;
1662         ssize_t ret;
1663         bool only_release_metadata = false;
1664         bool force_page_uptodate = false;
1665         loff_t old_isize = i_size_read(inode);
1666         unsigned int ilock_flags = 0;
1667
1668         if (iocb->ki_flags & IOCB_NOWAIT)
1669                 ilock_flags |= BTRFS_ILOCK_TRY;
1670
1671         ret = btrfs_inode_lock(inode, ilock_flags);
1672         if (ret < 0)
1673                 return ret;
1674
1675         ret = generic_write_checks(iocb, i);
1676         if (ret <= 0)
1677                 goto out;
1678
1679         ret = btrfs_write_check(iocb, i, ret);
1680         if (ret < 0)
1681                 goto out;
1682
1683         pos = iocb->ki_pos;
1684         nrptrs = min(DIV_ROUND_UP(iov_iter_count(i), PAGE_SIZE),
1685                         PAGE_SIZE / (sizeof(struct page *)));
1686         nrptrs = min(nrptrs, current->nr_dirtied_pause - current->nr_dirtied);
1687         nrptrs = max(nrptrs, 8);
1688         pages = kmalloc_array(nrptrs, sizeof(struct page *), GFP_KERNEL);
1689         if (!pages) {
1690                 ret = -ENOMEM;
1691                 goto out;
1692         }
1693
1694         while (iov_iter_count(i) > 0) {
1695                 struct extent_state *cached_state = NULL;
1696                 size_t offset = offset_in_page(pos);
1697                 size_t sector_offset;
1698                 size_t write_bytes = min(iov_iter_count(i),
1699                                          nrptrs * (size_t)PAGE_SIZE -
1700                                          offset);
1701                 size_t num_pages;
1702                 size_t reserve_bytes;
1703                 size_t dirty_pages;
1704                 size_t copied;
1705                 size_t dirty_sectors;
1706                 size_t num_sectors;
1707                 int extents_locked;
1708
1709                 /*
1710                  * Fault pages before locking them in prepare_pages
1711                  * to avoid recursive lock
1712                  */
1713                 if (unlikely(iov_iter_fault_in_readable(i, write_bytes))) {
1714                         ret = -EFAULT;
1715                         break;
1716                 }
1717
1718                 only_release_metadata = false;
1719                 sector_offset = pos & (fs_info->sectorsize - 1);
1720
1721                 extent_changeset_release(data_reserved);
1722                 ret = btrfs_check_data_free_space(BTRFS_I(inode),
1723                                                   &data_reserved, pos,
1724                                                   write_bytes);
1725                 if (ret < 0) {
1726                         /*
1727                          * If we don't have to COW at the offset, reserve
1728                          * metadata only. write_bytes may get smaller than
1729                          * requested here.
1730                          */
1731                         if (btrfs_check_nocow_lock(BTRFS_I(inode), pos,
1732                                                    &write_bytes) > 0)
1733                                 only_release_metadata = true;
1734                         else
1735                                 break;
1736                 }
1737
1738                 num_pages = DIV_ROUND_UP(write_bytes + offset, PAGE_SIZE);
1739                 WARN_ON(num_pages > nrptrs);
1740                 reserve_bytes = round_up(write_bytes + sector_offset,
1741                                          fs_info->sectorsize);
1742                 WARN_ON(reserve_bytes == 0);
1743                 ret = btrfs_delalloc_reserve_metadata(BTRFS_I(inode),
1744                                 reserve_bytes);
1745                 if (ret) {
1746                         if (!only_release_metadata)
1747                                 btrfs_free_reserved_data_space(BTRFS_I(inode),
1748                                                 data_reserved, pos,
1749                                                 write_bytes);
1750                         else
1751                                 btrfs_check_nocow_unlock(BTRFS_I(inode));
1752                         break;
1753                 }
1754
1755                 release_bytes = reserve_bytes;
1756 again:
1757                 /*
1758                  * This is going to setup the pages array with the number of
1759                  * pages we want, so we don't really need to worry about the
1760                  * contents of pages from loop to loop
1761                  */
1762                 ret = prepare_pages(inode, pages, num_pages,
1763                                     pos, write_bytes,
1764                                     force_page_uptodate);
1765                 if (ret) {
1766                         btrfs_delalloc_release_extents(BTRFS_I(inode),
1767                                                        reserve_bytes);
1768                         break;
1769                 }
1770
1771                 extents_locked = lock_and_cleanup_extent_if_need(
1772                                 BTRFS_I(inode), pages,
1773                                 num_pages, pos, write_bytes, &lockstart,
1774                                 &lockend, &cached_state);
1775                 if (extents_locked < 0) {
1776                         if (extents_locked == -EAGAIN)
1777                                 goto again;
1778                         btrfs_delalloc_release_extents(BTRFS_I(inode),
1779                                                        reserve_bytes);
1780                         ret = extents_locked;
1781                         break;
1782                 }
1783
1784                 copied = btrfs_copy_from_user(pos, write_bytes, pages, i);
1785
1786                 num_sectors = BTRFS_BYTES_TO_BLKS(fs_info, reserve_bytes);
1787                 dirty_sectors = round_up(copied + sector_offset,
1788                                         fs_info->sectorsize);
1789                 dirty_sectors = BTRFS_BYTES_TO_BLKS(fs_info, dirty_sectors);
1790
1791                 /*
1792                  * if we have trouble faulting in the pages, fall
1793                  * back to one page at a time
1794                  */
1795                 if (copied < write_bytes)
1796                         nrptrs = 1;
1797
1798                 if (copied == 0) {
1799                         force_page_uptodate = true;
1800                         dirty_sectors = 0;
1801                         dirty_pages = 0;
1802                 } else {
1803                         force_page_uptodate = false;
1804                         dirty_pages = DIV_ROUND_UP(copied + offset,
1805                                                    PAGE_SIZE);
1806                 }
1807
1808                 if (num_sectors > dirty_sectors) {
1809                         /* release everything except the sectors we dirtied */
1810                         release_bytes -= dirty_sectors << fs_info->sectorsize_bits;
1811                         if (only_release_metadata) {
1812                                 btrfs_delalloc_release_metadata(BTRFS_I(inode),
1813                                                         release_bytes, true);
1814                         } else {
1815                                 u64 __pos;
1816
1817                                 __pos = round_down(pos,
1818                                                    fs_info->sectorsize) +
1819                                         (dirty_pages << PAGE_SHIFT);
1820                                 btrfs_delalloc_release_space(BTRFS_I(inode),
1821                                                 data_reserved, __pos,
1822                                                 release_bytes, true);
1823                         }
1824                 }
1825
1826                 release_bytes = round_up(copied + sector_offset,
1827                                         fs_info->sectorsize);
1828
1829                 ret = btrfs_dirty_pages(BTRFS_I(inode), pages,
1830                                         dirty_pages, pos, copied,
1831                                         &cached_state, only_release_metadata);
1832
1833                 /*
1834                  * If we have not locked the extent range, because the range's
1835                  * start offset is >= i_size, we might still have a non-NULL
1836                  * cached extent state, acquired while marking the extent range
1837                  * as delalloc through btrfs_dirty_pages(). Therefore free any
1838                  * possible cached extent state to avoid a memory leak.
1839                  */
1840                 if (extents_locked)
1841                         unlock_extent_cached(&BTRFS_I(inode)->io_tree,
1842                                              lockstart, lockend, &cached_state);
1843                 else
1844                         free_extent_state(cached_state);
1845
1846                 btrfs_delalloc_release_extents(BTRFS_I(inode), reserve_bytes);
1847                 if (ret) {
1848                         btrfs_drop_pages(pages, num_pages);
1849                         break;
1850                 }
1851
1852                 release_bytes = 0;
1853                 if (only_release_metadata)
1854                         btrfs_check_nocow_unlock(BTRFS_I(inode));
1855
1856                 btrfs_drop_pages(pages, num_pages);
1857
1858                 cond_resched();
1859
1860                 balance_dirty_pages_ratelimited(inode->i_mapping);
1861
1862                 pos += copied;
1863                 num_written += copied;
1864         }
1865
1866         kfree(pages);
1867
1868         if (release_bytes) {
1869                 if (only_release_metadata) {
1870                         btrfs_check_nocow_unlock(BTRFS_I(inode));
1871                         btrfs_delalloc_release_metadata(BTRFS_I(inode),
1872                                         release_bytes, true);
1873                 } else {
1874                         btrfs_delalloc_release_space(BTRFS_I(inode),
1875                                         data_reserved,
1876                                         round_down(pos, fs_info->sectorsize),
1877                                         release_bytes, true);
1878                 }
1879         }
1880
1881         extent_changeset_free(data_reserved);
1882         if (num_written > 0) {
1883                 pagecache_isize_extended(inode, old_isize, iocb->ki_pos);
1884                 iocb->ki_pos += num_written;
1885         }
1886 out:
1887         btrfs_inode_unlock(inode, ilock_flags);
1888         return num_written ? num_written : ret;
1889 }
1890
1891 static ssize_t check_direct_IO(struct btrfs_fs_info *fs_info,
1892                                const struct iov_iter *iter, loff_t offset)
1893 {
1894         const u32 blocksize_mask = fs_info->sectorsize - 1;
1895
1896         if (offset & blocksize_mask)
1897                 return -EINVAL;
1898
1899         if (iov_iter_alignment(iter) & blocksize_mask)
1900                 return -EINVAL;
1901
1902         return 0;
1903 }
1904
1905 static ssize_t btrfs_direct_write(struct kiocb *iocb, struct iov_iter *from)
1906 {
1907         struct file *file = iocb->ki_filp;
1908         struct inode *inode = file_inode(file);
1909         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1910         loff_t pos;
1911         ssize_t written = 0;
1912         ssize_t written_buffered;
1913         loff_t endbyte;
1914         ssize_t err;
1915         unsigned int ilock_flags = 0;
1916         struct iomap_dio *dio = NULL;
1917
1918         if (iocb->ki_flags & IOCB_NOWAIT)
1919                 ilock_flags |= BTRFS_ILOCK_TRY;
1920
1921         /* If the write DIO is within EOF, use a shared lock */
1922         if (iocb->ki_pos + iov_iter_count(from) <= i_size_read(inode))
1923                 ilock_flags |= BTRFS_ILOCK_SHARED;
1924
1925 relock:
1926         err = btrfs_inode_lock(inode, ilock_flags);
1927         if (err < 0)
1928                 return err;
1929
1930         err = generic_write_checks(iocb, from);
1931         if (err <= 0) {
1932                 btrfs_inode_unlock(inode, ilock_flags);
1933                 return err;
1934         }
1935
1936         err = btrfs_write_check(iocb, from, err);
1937         if (err < 0) {
1938                 btrfs_inode_unlock(inode, ilock_flags);
1939                 goto out;
1940         }
1941
1942         pos = iocb->ki_pos;
1943         /*
1944          * Re-check since file size may have changed just before taking the
1945          * lock or pos may have changed because of O_APPEND in generic_write_check()
1946          */
1947         if ((ilock_flags & BTRFS_ILOCK_SHARED) &&
1948             pos + iov_iter_count(from) > i_size_read(inode)) {
1949                 btrfs_inode_unlock(inode, ilock_flags);
1950                 ilock_flags &= ~BTRFS_ILOCK_SHARED;
1951                 goto relock;
1952         }
1953
1954         if (check_direct_IO(fs_info, from, pos)) {
1955                 btrfs_inode_unlock(inode, ilock_flags);
1956                 goto buffered;
1957         }
1958
1959         dio = __iomap_dio_rw(iocb, from, &btrfs_dio_iomap_ops, &btrfs_dio_ops,
1960                              0);
1961
1962         btrfs_inode_unlock(inode, ilock_flags);
1963
1964         if (IS_ERR_OR_NULL(dio)) {
1965                 err = PTR_ERR_OR_ZERO(dio);
1966                 if (err < 0 && err != -ENOTBLK)
1967                         goto out;
1968         } else {
1969                 written = iomap_dio_complete(dio);
1970         }
1971
1972         if (written < 0 || !iov_iter_count(from)) {
1973                 err = written;
1974                 goto out;
1975         }
1976
1977 buffered:
1978         pos = iocb->ki_pos;
1979         written_buffered = btrfs_buffered_write(iocb, from);
1980         if (written_buffered < 0) {
1981                 err = written_buffered;
1982                 goto out;
1983         }
1984         /*
1985          * Ensure all data is persisted. We want the next direct IO read to be
1986          * able to read what was just written.
1987          */
1988         endbyte = pos + written_buffered - 1;
1989         err = btrfs_fdatawrite_range(inode, pos, endbyte);
1990         if (err)
1991                 goto out;
1992         err = filemap_fdatawait_range(inode->i_mapping, pos, endbyte);
1993         if (err)
1994                 goto out;
1995         written += written_buffered;
1996         iocb->ki_pos = pos + written_buffered;
1997         invalidate_mapping_pages(file->f_mapping, pos >> PAGE_SHIFT,
1998                                  endbyte >> PAGE_SHIFT);
1999 out:
2000         return written ? written : err;
2001 }
2002
2003 static ssize_t btrfs_file_write_iter(struct kiocb *iocb,
2004                                     struct iov_iter *from)
2005 {
2006         struct file *file = iocb->ki_filp;
2007         struct btrfs_inode *inode = BTRFS_I(file_inode(file));
2008         ssize_t num_written = 0;
2009         const bool sync = iocb->ki_flags & IOCB_DSYNC;
2010
2011         /*
2012          * If the fs flips readonly due to some impossible error, although we
2013          * have opened a file as writable, we have to stop this write operation
2014          * to ensure consistency.
2015          */
2016         if (test_bit(BTRFS_FS_STATE_ERROR, &inode->root->fs_info->fs_state))
2017                 return -EROFS;
2018
2019         if (!(iocb->ki_flags & IOCB_DIRECT) &&
2020             (iocb->ki_flags & IOCB_NOWAIT))
2021                 return -EOPNOTSUPP;
2022
2023         if (sync)
2024                 atomic_inc(&inode->sync_writers);
2025
2026         if (iocb->ki_flags & IOCB_DIRECT)
2027                 num_written = btrfs_direct_write(iocb, from);
2028         else
2029                 num_written = btrfs_buffered_write(iocb, from);
2030
2031         btrfs_set_inode_last_sub_trans(inode);
2032
2033         if (num_written > 0)
2034                 num_written = generic_write_sync(iocb, num_written);
2035
2036         if (sync)
2037                 atomic_dec(&inode->sync_writers);
2038
2039         current->backing_dev_info = NULL;
2040         return num_written;
2041 }
2042
2043 int btrfs_release_file(struct inode *inode, struct file *filp)
2044 {
2045         struct btrfs_file_private *private = filp->private_data;
2046
2047         if (private && private->filldir_buf)
2048                 kfree(private->filldir_buf);
2049         kfree(private);
2050         filp->private_data = NULL;
2051
2052         /*
2053          * Set by setattr when we are about to truncate a file from a non-zero
2054          * size to a zero size.  This tries to flush down new bytes that may
2055          * have been written if the application were using truncate to replace
2056          * a file in place.
2057          */
2058         if (test_and_clear_bit(BTRFS_INODE_FLUSH_ON_CLOSE,
2059                                &BTRFS_I(inode)->runtime_flags))
2060                         filemap_flush(inode->i_mapping);
2061         return 0;
2062 }
2063
2064 static int start_ordered_ops(struct inode *inode, loff_t start, loff_t end)
2065 {
2066         int ret;
2067         struct blk_plug plug;
2068
2069         /*
2070          * This is only called in fsync, which would do synchronous writes, so
2071          * a plug can merge adjacent IOs as much as possible.  Esp. in case of
2072          * multiple disks using raid profile, a large IO can be split to
2073          * several segments of stripe length (currently 64K).
2074          */
2075         blk_start_plug(&plug);
2076         atomic_inc(&BTRFS_I(inode)->sync_writers);
2077         ret = btrfs_fdatawrite_range(inode, start, end);
2078         atomic_dec(&BTRFS_I(inode)->sync_writers);
2079         blk_finish_plug(&plug);
2080
2081         return ret;
2082 }
2083
2084 static inline bool skip_inode_logging(const struct btrfs_log_ctx *ctx)
2085 {
2086         struct btrfs_inode *inode = BTRFS_I(ctx->inode);
2087         struct btrfs_fs_info *fs_info = inode->root->fs_info;
2088
2089         if (btrfs_inode_in_log(inode, fs_info->generation) &&
2090             list_empty(&ctx->ordered_extents))
2091                 return true;
2092
2093         /*
2094          * If we are doing a fast fsync we can not bail out if the inode's
2095          * last_trans is <= then the last committed transaction, because we only
2096          * update the last_trans of the inode during ordered extent completion,
2097          * and for a fast fsync we don't wait for that, we only wait for the
2098          * writeback to complete.
2099          */
2100         if (inode->last_trans <= fs_info->last_trans_committed &&
2101             (test_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &inode->runtime_flags) ||
2102              list_empty(&ctx->ordered_extents)))
2103                 return true;
2104
2105         return false;
2106 }
2107
2108 /*
2109  * fsync call for both files and directories.  This logs the inode into
2110  * the tree log instead of forcing full commits whenever possible.
2111  *
2112  * It needs to call filemap_fdatawait so that all ordered extent updates are
2113  * in the metadata btree are up to date for copying to the log.
2114  *
2115  * It drops the inode mutex before doing the tree log commit.  This is an
2116  * important optimization for directories because holding the mutex prevents
2117  * new operations on the dir while we write to disk.
2118  */
2119 int btrfs_sync_file(struct file *file, loff_t start, loff_t end, int datasync)
2120 {
2121         struct dentry *dentry = file_dentry(file);
2122         struct inode *inode = d_inode(dentry);
2123         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2124         struct btrfs_root *root = BTRFS_I(inode)->root;
2125         struct btrfs_trans_handle *trans;
2126         struct btrfs_log_ctx ctx;
2127         int ret = 0, err;
2128         u64 len;
2129         bool full_sync;
2130
2131         trace_btrfs_sync_file(file, datasync);
2132
2133         btrfs_init_log_ctx(&ctx, inode);
2134
2135         /*
2136          * Always set the range to a full range, otherwise we can get into
2137          * several problems, from missing file extent items to represent holes
2138          * when not using the NO_HOLES feature, to log tree corruption due to
2139          * races between hole detection during logging and completion of ordered
2140          * extents outside the range, to missing checksums due to ordered extents
2141          * for which we flushed only a subset of their pages.
2142          */
2143         start = 0;
2144         end = LLONG_MAX;
2145         len = (u64)LLONG_MAX + 1;
2146
2147         /*
2148          * We write the dirty pages in the range and wait until they complete
2149          * out of the ->i_mutex. If so, we can flush the dirty pages by
2150          * multi-task, and make the performance up.  See
2151          * btrfs_wait_ordered_range for an explanation of the ASYNC check.
2152          */
2153         ret = start_ordered_ops(inode, start, end);
2154         if (ret)
2155                 goto out;
2156
2157         btrfs_inode_lock(inode, BTRFS_ILOCK_MMAP);
2158
2159         atomic_inc(&root->log_batch);
2160
2161         /*
2162          * Always check for the full sync flag while holding the inode's lock,
2163          * to avoid races with other tasks. The flag must be either set all the
2164          * time during logging or always off all the time while logging.
2165          */
2166         full_sync = test_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2167                              &BTRFS_I(inode)->runtime_flags);
2168
2169         /*
2170          * Before we acquired the inode's lock and the mmap lock, someone may
2171          * have dirtied more pages in the target range. We need to make sure
2172          * that writeback for any such pages does not start while we are logging
2173          * the inode, because if it does, any of the following might happen when
2174          * we are not doing a full inode sync:
2175          *
2176          * 1) We log an extent after its writeback finishes but before its
2177          *    checksums are added to the csum tree, leading to -EIO errors
2178          *    when attempting to read the extent after a log replay.
2179          *
2180          * 2) We can end up logging an extent before its writeback finishes.
2181          *    Therefore after the log replay we will have a file extent item
2182          *    pointing to an unwritten extent (and no data checksums as well).
2183          *
2184          * So trigger writeback for any eventual new dirty pages and then we
2185          * wait for all ordered extents to complete below.
2186          */
2187         ret = start_ordered_ops(inode, start, end);
2188         if (ret) {
2189                 btrfs_inode_unlock(inode, BTRFS_ILOCK_MMAP);
2190                 goto out;
2191         }
2192
2193         /*
2194          * We have to do this here to avoid the priority inversion of waiting on
2195          * IO of a lower priority task while holding a transaction open.
2196          *
2197          * For a full fsync we wait for the ordered extents to complete while
2198          * for a fast fsync we wait just for writeback to complete, and then
2199          * attach the ordered extents to the transaction so that a transaction
2200          * commit waits for their completion, to avoid data loss if we fsync,
2201          * the current transaction commits before the ordered extents complete
2202          * and a power failure happens right after that.
2203          *
2204          * For zoned filesystem, if a write IO uses a ZONE_APPEND command, the
2205          * logical address recorded in the ordered extent may change. We need
2206          * to wait for the IO to stabilize the logical address.
2207          */
2208         if (full_sync || btrfs_is_zoned(fs_info)) {
2209                 ret = btrfs_wait_ordered_range(inode, start, len);
2210         } else {
2211                 /*
2212                  * Get our ordered extents as soon as possible to avoid doing
2213                  * checksum lookups in the csum tree, and use instead the
2214                  * checksums attached to the ordered extents.
2215                  */
2216                 btrfs_get_ordered_extents_for_logging(BTRFS_I(inode),
2217                                                       &ctx.ordered_extents);
2218                 ret = filemap_fdatawait_range(inode->i_mapping, start, end);
2219         }
2220
2221         if (ret)
2222                 goto out_release_extents;
2223
2224         atomic_inc(&root->log_batch);
2225
2226         smp_mb();
2227         if (skip_inode_logging(&ctx)) {
2228                 /*
2229                  * We've had everything committed since the last time we were
2230                  * modified so clear this flag in case it was set for whatever
2231                  * reason, it's no longer relevant.
2232                  */
2233                 clear_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2234                           &BTRFS_I(inode)->runtime_flags);
2235                 /*
2236                  * An ordered extent might have started before and completed
2237                  * already with io errors, in which case the inode was not
2238                  * updated and we end up here. So check the inode's mapping
2239                  * for any errors that might have happened since we last
2240                  * checked called fsync.
2241                  */
2242                 ret = filemap_check_wb_err(inode->i_mapping, file->f_wb_err);
2243                 goto out_release_extents;
2244         }
2245
2246         /*
2247          * We use start here because we will need to wait on the IO to complete
2248          * in btrfs_sync_log, which could require joining a transaction (for
2249          * example checking cross references in the nocow path).  If we use join
2250          * here we could get into a situation where we're waiting on IO to
2251          * happen that is blocked on a transaction trying to commit.  With start
2252          * we inc the extwriter counter, so we wait for all extwriters to exit
2253          * before we start blocking joiners.  This comment is to keep somebody
2254          * from thinking they are super smart and changing this to
2255          * btrfs_join_transaction *cough*Josef*cough*.
2256          */
2257         trans = btrfs_start_transaction(root, 0);
2258         if (IS_ERR(trans)) {
2259                 ret = PTR_ERR(trans);
2260                 goto out_release_extents;
2261         }
2262         trans->in_fsync = true;
2263
2264         ret = btrfs_log_dentry_safe(trans, dentry, &ctx);
2265         btrfs_release_log_ctx_extents(&ctx);
2266         if (ret < 0) {
2267                 /* Fallthrough and commit/free transaction. */
2268                 ret = 1;
2269         }
2270
2271         /* we've logged all the items and now have a consistent
2272          * version of the file in the log.  It is possible that
2273          * someone will come in and modify the file, but that's
2274          * fine because the log is consistent on disk, and we
2275          * have references to all of the file's extents
2276          *
2277          * It is possible that someone will come in and log the
2278          * file again, but that will end up using the synchronization
2279          * inside btrfs_sync_log to keep things safe.
2280          */
2281         btrfs_inode_unlock(inode, BTRFS_ILOCK_MMAP);
2282
2283         if (ret != BTRFS_NO_LOG_SYNC) {
2284                 if (!ret) {
2285                         ret = btrfs_sync_log(trans, root, &ctx);
2286                         if (!ret) {
2287                                 ret = btrfs_end_transaction(trans);
2288                                 goto out;
2289                         }
2290                 }
2291                 if (!full_sync) {
2292                         ret = btrfs_wait_ordered_range(inode, start, len);
2293                         if (ret) {
2294                                 btrfs_end_transaction(trans);
2295                                 goto out;
2296                         }
2297                 }
2298                 ret = btrfs_commit_transaction(trans);
2299         } else {
2300                 ret = btrfs_end_transaction(trans);
2301         }
2302 out:
2303         ASSERT(list_empty(&ctx.list));
2304         err = file_check_and_advance_wb_err(file);
2305         if (!ret)
2306                 ret = err;
2307         return ret > 0 ? -EIO : ret;
2308
2309 out_release_extents:
2310         btrfs_release_log_ctx_extents(&ctx);
2311         btrfs_inode_unlock(inode, BTRFS_ILOCK_MMAP);
2312         goto out;
2313 }
2314
2315 static const struct vm_operations_struct btrfs_file_vm_ops = {
2316         .fault          = filemap_fault,
2317         .map_pages      = filemap_map_pages,
2318         .page_mkwrite   = btrfs_page_mkwrite,
2319 };
2320
2321 static int btrfs_file_mmap(struct file  *filp, struct vm_area_struct *vma)
2322 {
2323         struct address_space *mapping = filp->f_mapping;
2324
2325         if (!mapping->a_ops->readpage)
2326                 return -ENOEXEC;
2327
2328         file_accessed(filp);
2329         vma->vm_ops = &btrfs_file_vm_ops;
2330
2331         return 0;
2332 }
2333
2334 static int hole_mergeable(struct btrfs_inode *inode, struct extent_buffer *leaf,
2335                           int slot, u64 start, u64 end)
2336 {
2337         struct btrfs_file_extent_item *fi;
2338         struct btrfs_key key;
2339
2340         if (slot < 0 || slot >= btrfs_header_nritems(leaf))
2341                 return 0;
2342
2343         btrfs_item_key_to_cpu(leaf, &key, slot);
2344         if (key.objectid != btrfs_ino(inode) ||
2345             key.type != BTRFS_EXTENT_DATA_KEY)
2346                 return 0;
2347
2348         fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
2349
2350         if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG)
2351                 return 0;
2352
2353         if (btrfs_file_extent_disk_bytenr(leaf, fi))
2354                 return 0;
2355
2356         if (key.offset == end)
2357                 return 1;
2358         if (key.offset + btrfs_file_extent_num_bytes(leaf, fi) == start)
2359                 return 1;
2360         return 0;
2361 }
2362
2363 static int fill_holes(struct btrfs_trans_handle *trans,
2364                 struct btrfs_inode *inode,
2365                 struct btrfs_path *path, u64 offset, u64 end)
2366 {
2367         struct btrfs_fs_info *fs_info = trans->fs_info;
2368         struct btrfs_root *root = inode->root;
2369         struct extent_buffer *leaf;
2370         struct btrfs_file_extent_item *fi;
2371         struct extent_map *hole_em;
2372         struct extent_map_tree *em_tree = &inode->extent_tree;
2373         struct btrfs_key key;
2374         int ret;
2375
2376         if (btrfs_fs_incompat(fs_info, NO_HOLES))
2377                 goto out;
2378
2379         key.objectid = btrfs_ino(inode);
2380         key.type = BTRFS_EXTENT_DATA_KEY;
2381         key.offset = offset;
2382
2383         ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
2384         if (ret <= 0) {
2385                 /*
2386                  * We should have dropped this offset, so if we find it then
2387                  * something has gone horribly wrong.
2388                  */
2389                 if (ret == 0)
2390                         ret = -EINVAL;
2391                 return ret;
2392         }
2393
2394         leaf = path->nodes[0];
2395         if (hole_mergeable(inode, leaf, path->slots[0] - 1, offset, end)) {
2396                 u64 num_bytes;
2397
2398                 path->slots[0]--;
2399                 fi = btrfs_item_ptr(leaf, path->slots[0],
2400                                     struct btrfs_file_extent_item);
2401                 num_bytes = btrfs_file_extent_num_bytes(leaf, fi) +
2402                         end - offset;
2403                 btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2404                 btrfs_set_file_extent_ram_bytes(leaf, fi, num_bytes);
2405                 btrfs_set_file_extent_offset(leaf, fi, 0);
2406                 btrfs_mark_buffer_dirty(leaf);
2407                 goto out;
2408         }
2409
2410         if (hole_mergeable(inode, leaf, path->slots[0], offset, end)) {
2411                 u64 num_bytes;
2412
2413                 key.offset = offset;
2414                 btrfs_set_item_key_safe(fs_info, path, &key);
2415                 fi = btrfs_item_ptr(leaf, path->slots[0],
2416                                     struct btrfs_file_extent_item);
2417                 num_bytes = btrfs_file_extent_num_bytes(leaf, fi) + end -
2418                         offset;
2419                 btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2420                 btrfs_set_file_extent_ram_bytes(leaf, fi, num_bytes);
2421                 btrfs_set_file_extent_offset(leaf, fi, 0);
2422                 btrfs_mark_buffer_dirty(leaf);
2423                 goto out;
2424         }
2425         btrfs_release_path(path);
2426
2427         ret = btrfs_insert_file_extent(trans, root, btrfs_ino(inode),
2428                         offset, 0, 0, end - offset, 0, end - offset, 0, 0, 0);
2429         if (ret)
2430                 return ret;
2431
2432 out:
2433         btrfs_release_path(path);
2434
2435         hole_em = alloc_extent_map();
2436         if (!hole_em) {
2437                 btrfs_drop_extent_cache(inode, offset, end - 1, 0);
2438                 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &inode->runtime_flags);
2439         } else {
2440                 hole_em->start = offset;
2441                 hole_em->len = end - offset;
2442                 hole_em->ram_bytes = hole_em->len;
2443                 hole_em->orig_start = offset;
2444
2445                 hole_em->block_start = EXTENT_MAP_HOLE;
2446                 hole_em->block_len = 0;
2447                 hole_em->orig_block_len = 0;
2448                 hole_em->compress_type = BTRFS_COMPRESS_NONE;
2449                 hole_em->generation = trans->transid;
2450
2451                 do {
2452                         btrfs_drop_extent_cache(inode, offset, end - 1, 0);
2453                         write_lock(&em_tree->lock);
2454                         ret = add_extent_mapping(em_tree, hole_em, 1);
2455                         write_unlock(&em_tree->lock);
2456                 } while (ret == -EEXIST);
2457                 free_extent_map(hole_em);
2458                 if (ret)
2459                         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2460                                         &inode->runtime_flags);
2461         }
2462
2463         return 0;
2464 }
2465
2466 /*
2467  * Find a hole extent on given inode and change start/len to the end of hole
2468  * extent.(hole/vacuum extent whose em->start <= start &&
2469  *         em->start + em->len > start)
2470  * When a hole extent is found, return 1 and modify start/len.
2471  */
2472 static int find_first_non_hole(struct btrfs_inode *inode, u64 *start, u64 *len)
2473 {
2474         struct btrfs_fs_info *fs_info = inode->root->fs_info;
2475         struct extent_map *em;
2476         int ret = 0;
2477
2478         em = btrfs_get_extent(inode, NULL, 0,
2479                               round_down(*start, fs_info->sectorsize),
2480                               round_up(*len, fs_info->sectorsize));
2481         if (IS_ERR(em))
2482                 return PTR_ERR(em);
2483
2484         /* Hole or vacuum extent(only exists in no-hole mode) */
2485         if (em->block_start == EXTENT_MAP_HOLE) {
2486                 ret = 1;
2487                 *len = em->start + em->len > *start + *len ?
2488                        0 : *start + *len - em->start - em->len;
2489                 *start = em->start + em->len;
2490         }
2491         free_extent_map(em);
2492         return ret;
2493 }
2494
2495 static int btrfs_punch_hole_lock_range(struct inode *inode,
2496                                        const u64 lockstart,
2497                                        const u64 lockend,
2498                                        struct extent_state **cached_state)
2499 {
2500         /*
2501          * For subpage case, if the range is not at page boundary, we could
2502          * have pages at the leading/tailing part of the range.
2503          * This could lead to dead loop since filemap_range_has_page()
2504          * will always return true.
2505          * So here we need to do extra page alignment for
2506          * filemap_range_has_page().
2507          */
2508         const u64 page_lockstart = round_up(lockstart, PAGE_SIZE);
2509         const u64 page_lockend = round_down(lockend + 1, PAGE_SIZE) - 1;
2510
2511         while (1) {
2512                 struct btrfs_ordered_extent *ordered;
2513                 int ret;
2514
2515                 truncate_pagecache_range(inode, lockstart, lockend);
2516
2517                 lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2518                                  cached_state);
2519                 ordered = btrfs_lookup_first_ordered_extent(BTRFS_I(inode),
2520                                                             lockend);
2521
2522                 /*
2523                  * We need to make sure we have no ordered extents in this range
2524                  * and nobody raced in and read a page in this range, if we did
2525                  * we need to try again.
2526                  */
2527                 if ((!ordered ||
2528                     (ordered->file_offset + ordered->num_bytes <= lockstart ||
2529                      ordered->file_offset > lockend)) &&
2530                      !filemap_range_has_page(inode->i_mapping,
2531                                              page_lockstart, page_lockend)) {
2532                         if (ordered)
2533                                 btrfs_put_ordered_extent(ordered);
2534                         break;
2535                 }
2536                 if (ordered)
2537                         btrfs_put_ordered_extent(ordered);
2538                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart,
2539                                      lockend, cached_state);
2540                 ret = btrfs_wait_ordered_range(inode, lockstart,
2541                                                lockend - lockstart + 1);
2542                 if (ret)
2543                         return ret;
2544         }
2545         return 0;
2546 }
2547
2548 static int btrfs_insert_replace_extent(struct btrfs_trans_handle *trans,
2549                                      struct btrfs_inode *inode,
2550                                      struct btrfs_path *path,
2551                                      struct btrfs_replace_extent_info *extent_info,
2552                                      const u64 replace_len,
2553                                      const u64 bytes_to_drop)
2554 {
2555         struct btrfs_fs_info *fs_info = trans->fs_info;
2556         struct btrfs_root *root = inode->root;
2557         struct btrfs_file_extent_item *extent;
2558         struct extent_buffer *leaf;
2559         struct btrfs_key key;
2560         int slot;
2561         struct btrfs_ref ref = { 0 };
2562         int ret;
2563
2564         if (replace_len == 0)
2565                 return 0;
2566
2567         if (extent_info->disk_offset == 0 &&
2568             btrfs_fs_incompat(fs_info, NO_HOLES)) {
2569                 btrfs_update_inode_bytes(inode, 0, bytes_to_drop);
2570                 return 0;
2571         }
2572
2573         key.objectid = btrfs_ino(inode);
2574         key.type = BTRFS_EXTENT_DATA_KEY;
2575         key.offset = extent_info->file_offset;
2576         ret = btrfs_insert_empty_item(trans, root, path, &key,
2577                                       sizeof(struct btrfs_file_extent_item));
2578         if (ret)
2579                 return ret;
2580         leaf = path->nodes[0];
2581         slot = path->slots[0];
2582         write_extent_buffer(leaf, extent_info->extent_buf,
2583                             btrfs_item_ptr_offset(leaf, slot),
2584                             sizeof(struct btrfs_file_extent_item));
2585         extent = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
2586         ASSERT(btrfs_file_extent_type(leaf, extent) != BTRFS_FILE_EXTENT_INLINE);
2587         btrfs_set_file_extent_offset(leaf, extent, extent_info->data_offset);
2588         btrfs_set_file_extent_num_bytes(leaf, extent, replace_len);
2589         if (extent_info->is_new_extent)
2590                 btrfs_set_file_extent_generation(leaf, extent, trans->transid);
2591         btrfs_mark_buffer_dirty(leaf);
2592         btrfs_release_path(path);
2593
2594         ret = btrfs_inode_set_file_extent_range(inode, extent_info->file_offset,
2595                                                 replace_len);
2596         if (ret)
2597                 return ret;
2598
2599         /* If it's a hole, nothing more needs to be done. */
2600         if (extent_info->disk_offset == 0) {
2601                 btrfs_update_inode_bytes(inode, 0, bytes_to_drop);
2602                 return 0;
2603         }
2604
2605         btrfs_update_inode_bytes(inode, replace_len, bytes_to_drop);
2606
2607         if (extent_info->is_new_extent && extent_info->insertions == 0) {
2608                 key.objectid = extent_info->disk_offset;
2609                 key.type = BTRFS_EXTENT_ITEM_KEY;
2610                 key.offset = extent_info->disk_len;
2611                 ret = btrfs_alloc_reserved_file_extent(trans, root,
2612                                                        btrfs_ino(inode),
2613                                                        extent_info->file_offset,
2614                                                        extent_info->qgroup_reserved,
2615                                                        &key);
2616         } else {
2617                 u64 ref_offset;
2618
2619                 btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF,
2620                                        extent_info->disk_offset,
2621                                        extent_info->disk_len, 0);
2622                 ref_offset = extent_info->file_offset - extent_info->data_offset;
2623                 btrfs_init_data_ref(&ref, root->root_key.objectid,
2624                                     btrfs_ino(inode), ref_offset);
2625                 ret = btrfs_inc_extent_ref(trans, &ref);
2626         }
2627
2628         extent_info->insertions++;
2629
2630         return ret;
2631 }
2632
2633 /*
2634  * The respective range must have been previously locked, as well as the inode.
2635  * The end offset is inclusive (last byte of the range).
2636  * @extent_info is NULL for fallocate's hole punching and non-NULL when replacing
2637  * the file range with an extent.
2638  * When not punching a hole, we don't want to end up in a state where we dropped
2639  * extents without inserting a new one, so we must abort the transaction to avoid
2640  * a corruption.
2641  */
2642 int btrfs_replace_file_extents(struct btrfs_inode *inode,
2643                                struct btrfs_path *path, const u64 start,
2644                                const u64 end,
2645                                struct btrfs_replace_extent_info *extent_info,
2646                                struct btrfs_trans_handle **trans_out)
2647 {
2648         struct btrfs_drop_extents_args drop_args = { 0 };
2649         struct btrfs_root *root = inode->root;
2650         struct btrfs_fs_info *fs_info = root->fs_info;
2651         u64 min_size = btrfs_calc_insert_metadata_size(fs_info, 1);
2652         u64 ino_size = round_up(inode->vfs_inode.i_size, fs_info->sectorsize);
2653         struct btrfs_trans_handle *trans = NULL;
2654         struct btrfs_block_rsv *rsv;
2655         unsigned int rsv_count;
2656         u64 cur_offset;
2657         u64 len = end - start;
2658         int ret = 0;
2659
2660         if (end <= start)
2661                 return -EINVAL;
2662
2663         rsv = btrfs_alloc_block_rsv(fs_info, BTRFS_BLOCK_RSV_TEMP);
2664         if (!rsv) {
2665                 ret = -ENOMEM;
2666                 goto out;
2667         }
2668         rsv->size = btrfs_calc_insert_metadata_size(fs_info, 1);
2669         rsv->failfast = 1;
2670
2671         /*
2672          * 1 - update the inode
2673          * 1 - removing the extents in the range
2674          * 1 - adding the hole extent if no_holes isn't set or if we are
2675          *     replacing the range with a new extent
2676          */
2677         if (!btrfs_fs_incompat(fs_info, NO_HOLES) || extent_info)
2678                 rsv_count = 3;
2679         else
2680                 rsv_count = 2;
2681
2682         trans = btrfs_start_transaction(root, rsv_count);
2683         if (IS_ERR(trans)) {
2684                 ret = PTR_ERR(trans);
2685                 trans = NULL;
2686                 goto out_free;
2687         }
2688
2689         ret = btrfs_block_rsv_migrate(&fs_info->trans_block_rsv, rsv,
2690                                       min_size, false);
2691         BUG_ON(ret);
2692         trans->block_rsv = rsv;
2693
2694         cur_offset = start;
2695         drop_args.path = path;
2696         drop_args.end = end + 1;
2697         drop_args.drop_cache = true;
2698         while (cur_offset < end) {
2699                 drop_args.start = cur_offset;
2700                 ret = btrfs_drop_extents(trans, root, inode, &drop_args);
2701                 /* If we are punching a hole decrement the inode's byte count */
2702                 if (!extent_info)
2703                         btrfs_update_inode_bytes(inode, 0,
2704                                                  drop_args.bytes_found);
2705                 if (ret != -ENOSPC) {
2706                         /*
2707                          * When cloning we want to avoid transaction aborts when
2708                          * nothing was done and we are attempting to clone parts
2709                          * of inline extents, in such cases -EOPNOTSUPP is
2710                          * returned by __btrfs_drop_extents() without having
2711                          * changed anything in the file.
2712                          */
2713                         if (extent_info && !extent_info->is_new_extent &&
2714                             ret && ret != -EOPNOTSUPP)
2715                                 btrfs_abort_transaction(trans, ret);
2716                         break;
2717                 }
2718
2719                 trans->block_rsv = &fs_info->trans_block_rsv;
2720
2721                 if (!extent_info && cur_offset < drop_args.drop_end &&
2722                     cur_offset < ino_size) {
2723                         ret = fill_holes(trans, inode, path, cur_offset,
2724                                          drop_args.drop_end);
2725                         if (ret) {
2726                                 /*
2727                                  * If we failed then we didn't insert our hole
2728                                  * entries for the area we dropped, so now the
2729                                  * fs is corrupted, so we must abort the
2730                                  * transaction.
2731                                  */
2732                                 btrfs_abort_transaction(trans, ret);
2733                                 break;
2734                         }
2735                 } else if (!extent_info && cur_offset < drop_args.drop_end) {
2736                         /*
2737                          * We are past the i_size here, but since we didn't
2738                          * insert holes we need to clear the mapped area so we
2739                          * know to not set disk_i_size in this area until a new
2740                          * file extent is inserted here.
2741                          */
2742                         ret = btrfs_inode_clear_file_extent_range(inode,
2743                                         cur_offset,
2744                                         drop_args.drop_end - cur_offset);
2745                         if (ret) {
2746                                 /*
2747                                  * We couldn't clear our area, so we could
2748                                  * presumably adjust up and corrupt the fs, so
2749                                  * we need to abort.
2750                                  */
2751                                 btrfs_abort_transaction(trans, ret);
2752                                 break;
2753                         }
2754                 }
2755
2756                 if (extent_info &&
2757                     drop_args.drop_end > extent_info->file_offset) {
2758                         u64 replace_len = drop_args.drop_end -
2759                                           extent_info->file_offset;
2760
2761                         ret = btrfs_insert_replace_extent(trans, inode, path,
2762                                         extent_info, replace_len,
2763                                         drop_args.bytes_found);
2764                         if (ret) {
2765                                 btrfs_abort_transaction(trans, ret);
2766                                 break;
2767                         }
2768                         extent_info->data_len -= replace_len;
2769                         extent_info->data_offset += replace_len;
2770                         extent_info->file_offset += replace_len;
2771                 }
2772
2773                 ret = btrfs_update_inode(trans, root, inode);
2774                 if (ret)
2775                         break;
2776
2777                 btrfs_end_transaction(trans);
2778                 btrfs_btree_balance_dirty(fs_info);
2779
2780                 trans = btrfs_start_transaction(root, rsv_count);
2781                 if (IS_ERR(trans)) {
2782                         ret = PTR_ERR(trans);
2783                         trans = NULL;
2784                         break;
2785                 }
2786
2787                 ret = btrfs_block_rsv_migrate(&fs_info->trans_block_rsv,
2788                                               rsv, min_size, false);
2789                 BUG_ON(ret);    /* shouldn't happen */
2790                 trans->block_rsv = rsv;
2791
2792                 cur_offset = drop_args.drop_end;
2793                 len = end - cur_offset;
2794                 if (!extent_info && len) {
2795                         ret = find_first_non_hole(inode, &cur_offset, &len);
2796                         if (unlikely(ret < 0))
2797                                 break;
2798                         if (ret && !len) {
2799                                 ret = 0;
2800                                 break;
2801                         }
2802                 }
2803         }
2804
2805         /*
2806          * If we were cloning, force the next fsync to be a full one since we
2807          * we replaced (or just dropped in the case of cloning holes when
2808          * NO_HOLES is enabled) file extent items and did not setup new extent
2809          * maps for the replacement extents (or holes).
2810          */
2811         if (extent_info && !extent_info->is_new_extent)
2812                 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &inode->runtime_flags);
2813
2814         if (ret)
2815                 goto out_trans;
2816
2817         trans->block_rsv = &fs_info->trans_block_rsv;
2818         /*
2819          * If we are using the NO_HOLES feature we might have had already an
2820          * hole that overlaps a part of the region [lockstart, lockend] and
2821          * ends at (or beyond) lockend. Since we have no file extent items to
2822          * represent holes, drop_end can be less than lockend and so we must
2823          * make sure we have an extent map representing the existing hole (the
2824          * call to __btrfs_drop_extents() might have dropped the existing extent
2825          * map representing the existing hole), otherwise the fast fsync path
2826          * will not record the existence of the hole region
2827          * [existing_hole_start, lockend].
2828          */
2829         if (drop_args.drop_end <= end)
2830                 drop_args.drop_end = end + 1;
2831         /*
2832          * Don't insert file hole extent item if it's for a range beyond eof
2833          * (because it's useless) or if it represents a 0 bytes range (when
2834          * cur_offset == drop_end).
2835          */
2836         if (!extent_info && cur_offset < ino_size &&
2837             cur_offset < drop_args.drop_end) {
2838                 ret = fill_holes(trans, inode, path, cur_offset,
2839                                  drop_args.drop_end);
2840                 if (ret) {
2841                         /* Same comment as above. */
2842                         btrfs_abort_transaction(trans, ret);
2843                         goto out_trans;
2844                 }
2845         } else if (!extent_info && cur_offset < drop_args.drop_end) {
2846                 /* See the comment in the loop above for the reasoning here. */
2847                 ret = btrfs_inode_clear_file_extent_range(inode, cur_offset,
2848                                         drop_args.drop_end - cur_offset);
2849                 if (ret) {
2850                         btrfs_abort_transaction(trans, ret);
2851                         goto out_trans;
2852                 }
2853
2854         }
2855         if (extent_info) {
2856                 ret = btrfs_insert_replace_extent(trans, inode, path,
2857                                 extent_info, extent_info->data_len,
2858                                 drop_args.bytes_found);
2859                 if (ret) {
2860                         btrfs_abort_transaction(trans, ret);
2861                         goto out_trans;
2862                 }
2863         }
2864
2865 out_trans:
2866         if (!trans)
2867                 goto out_free;
2868
2869         trans->block_rsv = &fs_info->trans_block_rsv;
2870         if (ret)
2871                 btrfs_end_transaction(trans);
2872         else
2873                 *trans_out = trans;
2874 out_free:
2875         btrfs_free_block_rsv(fs_info, rsv);
2876 out:
2877         return ret;
2878 }
2879
2880 static int btrfs_punch_hole(struct inode *inode, loff_t offset, loff_t len)
2881 {
2882         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2883         struct btrfs_root *root = BTRFS_I(inode)->root;
2884         struct extent_state *cached_state = NULL;
2885         struct btrfs_path *path;
2886         struct btrfs_trans_handle *trans = NULL;
2887         u64 lockstart;
2888         u64 lockend;
2889         u64 tail_start;
2890         u64 tail_len;
2891         u64 orig_start = offset;
2892         int ret = 0;
2893         bool same_block;
2894         u64 ino_size;
2895         bool truncated_block = false;
2896         bool updated_inode = false;
2897
2898         ret = btrfs_wait_ordered_range(inode, offset, len);
2899         if (ret)
2900                 return ret;
2901
2902         btrfs_inode_lock(inode, BTRFS_ILOCK_MMAP);
2903         ino_size = round_up(inode->i_size, fs_info->sectorsize);
2904         ret = find_first_non_hole(BTRFS_I(inode), &offset, &len);
2905         if (ret < 0)
2906                 goto out_only_mutex;
2907         if (ret && !len) {
2908                 /* Already in a large hole */
2909                 ret = 0;
2910                 goto out_only_mutex;
2911         }
2912
2913         lockstart = round_up(offset, btrfs_inode_sectorsize(BTRFS_I(inode)));
2914         lockend = round_down(offset + len,
2915                              btrfs_inode_sectorsize(BTRFS_I(inode))) - 1;
2916         same_block = (BTRFS_BYTES_TO_BLKS(fs_info, offset))
2917                 == (BTRFS_BYTES_TO_BLKS(fs_info, offset + len - 1));
2918         /*
2919          * We needn't truncate any block which is beyond the end of the file
2920          * because we are sure there is no data there.
2921          */
2922         /*
2923          * Only do this if we are in the same block and we aren't doing the
2924          * entire block.
2925          */
2926         if (same_block && len < fs_info->sectorsize) {
2927                 if (offset < ino_size) {
2928                         truncated_block = true;
2929                         ret = btrfs_truncate_block(BTRFS_I(inode), offset, len,
2930                                                    0);
2931                 } else {
2932                         ret = 0;
2933                 }
2934                 goto out_only_mutex;
2935         }
2936
2937         /* zero back part of the first block */
2938         if (offset < ino_size) {
2939                 truncated_block = true;
2940                 ret = btrfs_truncate_block(BTRFS_I(inode), offset, 0, 0);
2941                 if (ret) {
2942                         btrfs_inode_unlock(inode, BTRFS_ILOCK_MMAP);
2943                         return ret;
2944                 }
2945         }
2946
2947         /* Check the aligned pages after the first unaligned page,
2948          * if offset != orig_start, which means the first unaligned page
2949          * including several following pages are already in holes,
2950          * the extra check can be skipped */
2951         if (offset == orig_start) {
2952                 /* after truncate page, check hole again */
2953                 len = offset + len - lockstart;
2954                 offset = lockstart;
2955                 ret = find_first_non_hole(BTRFS_I(inode), &offset, &len);
2956                 if (ret < 0)
2957                         goto out_only_mutex;
2958                 if (ret && !len) {
2959                         ret = 0;
2960                         goto out_only_mutex;
2961                 }
2962                 lockstart = offset;
2963         }
2964
2965         /* Check the tail unaligned part is in a hole */
2966         tail_start = lockend + 1;
2967         tail_len = offset + len - tail_start;
2968         if (tail_len) {
2969                 ret = find_first_non_hole(BTRFS_I(inode), &tail_start, &tail_len);
2970                 if (unlikely(ret < 0))
2971                         goto out_only_mutex;
2972                 if (!ret) {
2973                         /* zero the front end of the last page */
2974                         if (tail_start + tail_len < ino_size) {
2975                                 truncated_block = true;
2976                                 ret = btrfs_truncate_block(BTRFS_I(inode),
2977                                                         tail_start + tail_len,
2978                                                         0, 1);
2979                                 if (ret)
2980                                         goto out_only_mutex;
2981                         }
2982                 }
2983         }
2984
2985         if (lockend < lockstart) {
2986                 ret = 0;
2987                 goto out_only_mutex;
2988         }
2989
2990         ret = btrfs_punch_hole_lock_range(inode, lockstart, lockend,
2991                                           &cached_state);
2992         if (ret)
2993                 goto out_only_mutex;
2994
2995         path = btrfs_alloc_path();
2996         if (!path) {
2997                 ret = -ENOMEM;
2998                 goto out;
2999         }
3000
3001         ret = btrfs_replace_file_extents(BTRFS_I(inode), path, lockstart,
3002                                          lockend, NULL, &trans);
3003         btrfs_free_path(path);
3004         if (ret)
3005                 goto out;
3006
3007         ASSERT(trans != NULL);
3008         inode_inc_iversion(inode);
3009         inode->i_mtime = inode->i_ctime = current_time(inode);
3010         ret = btrfs_update_inode(trans, root, BTRFS_I(inode));
3011         updated_inode = true;
3012         btrfs_end_transaction(trans);
3013         btrfs_btree_balance_dirty(fs_info);
3014 out:
3015         unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
3016                              &cached_state);
3017 out_only_mutex:
3018         if (!updated_inode && truncated_block && !ret) {
3019                 /*
3020                  * If we only end up zeroing part of a page, we still need to
3021                  * update the inode item, so that all the time fields are
3022                  * updated as well as the necessary btrfs inode in memory fields
3023                  * for detecting, at fsync time, if the inode isn't yet in the
3024                  * log tree or it's there but not up to date.
3025                  */
3026                 struct timespec64 now = current_time(inode);
3027
3028                 inode_inc_iversion(inode);
3029                 inode->i_mtime = now;
3030                 inode->i_ctime = now;
3031                 trans = btrfs_start_transaction(root, 1);
3032                 if (IS_ERR(trans)) {
3033                         ret = PTR_ERR(trans);
3034                 } else {
3035                         int ret2;
3036
3037                         ret = btrfs_update_inode(trans, root, BTRFS_I(inode));
3038                         ret2 = btrfs_end_transaction(trans);
3039                         if (!ret)
3040                                 ret = ret2;
3041                 }
3042         }
3043         btrfs_inode_unlock(inode, BTRFS_ILOCK_MMAP);
3044         return ret;
3045 }
3046
3047 /* Helper structure to record which range is already reserved */
3048 struct falloc_range {
3049         struct list_head list;
3050         u64 start;
3051         u64 len;
3052 };
3053
3054 /*
3055  * Helper function to add falloc range
3056  *
3057  * Caller should have locked the larger range of extent containing
3058  * [start, len)
3059  */
3060 static int add_falloc_range(struct list_head *head, u64 start, u64 len)
3061 {
3062         struct falloc_range *range = NULL;
3063
3064         if (!list_empty(head)) {
3065                 /*
3066                  * As fallocate iterates by bytenr order, we only need to check
3067                  * the last range.
3068                  */
3069                 range = list_last_entry(head, struct falloc_range, list);
3070                 if (range->start + range->len == start) {
3071                         range->len += len;
3072                         return 0;
3073                 }
3074         }
3075
3076         range = kmalloc(sizeof(*range), GFP_KERNEL);
3077         if (!range)
3078                 return -ENOMEM;
3079         range->start = start;
3080         range->len = len;
3081         list_add_tail(&range->list, head);
3082         return 0;
3083 }
3084
3085 static int btrfs_fallocate_update_isize(struct inode *inode,
3086                                         const u64 end,
3087                                         const int mode)
3088 {
3089         struct btrfs_trans_handle *trans;
3090         struct btrfs_root *root = BTRFS_I(inode)->root;
3091         int ret;
3092         int ret2;
3093
3094         if (mode & FALLOC_FL_KEEP_SIZE || end <= i_size_read(inode))
3095                 return 0;
3096
3097         trans = btrfs_start_transaction(root, 1);
3098         if (IS_ERR(trans))
3099                 return PTR_ERR(trans);
3100
3101         inode->i_ctime = current_time(inode);
3102         i_size_write(inode, end);
3103         btrfs_inode_safe_disk_i_size_write(BTRFS_I(inode), 0);
3104         ret = btrfs_update_inode(trans, root, BTRFS_I(inode));
3105         ret2 = btrfs_end_transaction(trans);
3106
3107         return ret ? ret : ret2;
3108 }
3109
3110 enum {
3111         RANGE_BOUNDARY_WRITTEN_EXTENT,
3112         RANGE_BOUNDARY_PREALLOC_EXTENT,
3113         RANGE_BOUNDARY_HOLE,
3114 };
3115
3116 static int btrfs_zero_range_check_range_boundary(struct btrfs_inode *inode,
3117                                                  u64 offset)
3118 {
3119         const u64 sectorsize = btrfs_inode_sectorsize(inode);
3120         struct extent_map *em;
3121         int ret;
3122
3123         offset = round_down(offset, sectorsize);
3124         em = btrfs_get_extent(inode, NULL, 0, offset, sectorsize);
3125         if (IS_ERR(em))
3126                 return PTR_ERR(em);
3127
3128         if (em->block_start == EXTENT_MAP_HOLE)
3129                 ret = RANGE_BOUNDARY_HOLE;
3130         else if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
3131                 ret = RANGE_BOUNDARY_PREALLOC_EXTENT;
3132         else
3133                 ret = RANGE_BOUNDARY_WRITTEN_EXTENT;
3134
3135         free_extent_map(em);
3136         return ret;
3137 }
3138
3139 static int btrfs_zero_range(struct inode *inode,
3140                             loff_t offset,
3141                             loff_t len,
3142                             const int mode)
3143 {
3144         struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
3145         struct extent_map *em;
3146         struct extent_changeset *data_reserved = NULL;
3147         int ret;
3148         u64 alloc_hint = 0;
3149         const u64 sectorsize = btrfs_inode_sectorsize(BTRFS_I(inode));
3150         u64 alloc_start = round_down(offset, sectorsize);
3151         u64 alloc_end = round_up(offset + len, sectorsize);
3152         u64 bytes_to_reserve = 0;
3153         bool space_reserved = false;
3154
3155         inode_dio_wait(inode);
3156
3157         em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, alloc_start,
3158                               alloc_end - alloc_start);
3159         if (IS_ERR(em)) {
3160                 ret = PTR_ERR(em);
3161                 goto out;
3162         }
3163
3164         /*
3165          * Avoid hole punching and extent allocation for some cases. More cases
3166          * could be considered, but these are unlikely common and we keep things
3167          * as simple as possible for now. Also, intentionally, if the target
3168          * range contains one or more prealloc extents together with regular
3169          * extents and holes, we drop all the existing extents and allocate a
3170          * new prealloc extent, so that we get a larger contiguous disk extent.
3171          */
3172         if (em->start <= alloc_start &&
3173             test_bit(EXTENT_FLAG_PREALLOC, &em->flags)) {
3174                 const u64 em_end = em->start + em->len;
3175
3176                 if (em_end >= offset + len) {
3177                         /*
3178                          * The whole range is already a prealloc extent,
3179                          * do nothing except updating the inode's i_size if
3180                          * needed.
3181                          */
3182                         free_extent_map(em);
3183                         ret = btrfs_fallocate_update_isize(inode, offset + len,
3184                                                            mode);
3185                         goto out;
3186                 }
3187                 /*
3188                  * Part of the range is already a prealloc extent, so operate
3189                  * only on the remaining part of the range.
3190                  */
3191                 alloc_start = em_end;
3192                 ASSERT(IS_ALIGNED(alloc_start, sectorsize));
3193                 len = offset + len - alloc_start;
3194                 offset = alloc_start;
3195                 alloc_hint = em->block_start + em->len;
3196         }
3197         free_extent_map(em);
3198
3199         if (BTRFS_BYTES_TO_BLKS(fs_info, offset) ==
3200             BTRFS_BYTES_TO_BLKS(fs_info, offset + len - 1)) {
3201                 em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, alloc_start,
3202                                       sectorsize);
3203                 if (IS_ERR(em)) {
3204                         ret = PTR_ERR(em);
3205                         goto out;
3206                 }
3207
3208                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags)) {
3209                         free_extent_map(em);
3210                         ret = btrfs_fallocate_update_isize(inode, offset + len,
3211                                                            mode);
3212                         goto out;
3213                 }
3214                 if (len < sectorsize && em->block_start != EXTENT_MAP_HOLE) {
3215                         free_extent_map(em);
3216                         ret = btrfs_truncate_block(BTRFS_I(inode), offset, len,
3217                                                    0);
3218                         if (!ret)
3219                                 ret = btrfs_fallocate_update_isize(inode,
3220                                                                    offset + len,
3221                                                                    mode);
3222                         return ret;
3223                 }
3224                 free_extent_map(em);
3225                 alloc_start = round_down(offset, sectorsize);
3226                 alloc_end = alloc_start + sectorsize;
3227                 goto reserve_space;
3228         }
3229
3230         alloc_start = round_up(offset, sectorsize);
3231         alloc_end = round_down(offset + len, sectorsize);
3232
3233         /*
3234          * For unaligned ranges, check the pages at the boundaries, they might
3235          * map to an extent, in which case we need to partially zero them, or
3236          * they might map to a hole, in which case we need our allocation range
3237          * to cover them.
3238          */
3239         if (!IS_ALIGNED(offset, sectorsize)) {
3240                 ret = btrfs_zero_range_check_range_boundary(BTRFS_I(inode),
3241                                                             offset);
3242                 if (ret < 0)
3243                         goto out;
3244                 if (ret == RANGE_BOUNDARY_HOLE) {
3245                         alloc_start = round_down(offset, sectorsize);
3246                         ret = 0;
3247                 } else if (ret == RANGE_BOUNDARY_WRITTEN_EXTENT) {
3248                         ret = btrfs_truncate_block(BTRFS_I(inode), offset, 0, 0);
3249                         if (ret)
3250                                 goto out;
3251                 } else {
3252                         ret = 0;
3253                 }
3254         }
3255
3256         if (!IS_ALIGNED(offset + len, sectorsize)) {
3257                 ret = btrfs_zero_range_check_range_boundary(BTRFS_I(inode),
3258                                                             offset + len);
3259                 if (ret < 0)
3260                         goto out;
3261                 if (ret == RANGE_BOUNDARY_HOLE) {
3262                         alloc_end = round_up(offset + len, sectorsize);
3263                         ret = 0;
3264                 } else if (ret == RANGE_BOUNDARY_WRITTEN_EXTENT) {
3265                         ret = btrfs_truncate_block(BTRFS_I(inode), offset + len,
3266                                                    0, 1);
3267                         if (ret)
3268                                 goto out;
3269                 } else {
3270                         ret = 0;
3271                 }
3272         }
3273
3274 reserve_space:
3275         if (alloc_start < alloc_end) {
3276                 struct extent_state *cached_state = NULL;
3277                 const u64 lockstart = alloc_start;
3278                 const u64 lockend = alloc_end - 1;
3279
3280                 bytes_to_reserve = alloc_end - alloc_start;
3281                 ret = btrfs_alloc_data_chunk_ondemand(BTRFS_I(inode),
3282                                                       bytes_to_reserve);
3283                 if (ret < 0)
3284                         goto out;
3285                 space_reserved = true;
3286                 ret = btrfs_punch_hole_lock_range(inode, lockstart, lockend,
3287                                                   &cached_state);
3288                 if (ret)
3289                         goto out;
3290                 ret = btrfs_qgroup_reserve_data(BTRFS_I(inode), &data_reserved,
3291                                                 alloc_start, bytes_to_reserve);
3292                 if (ret) {
3293                         unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart,
3294                                              lockend, &cached_state);
3295                         goto out;
3296                 }
3297                 ret = btrfs_prealloc_file_range(inode, mode, alloc_start,
3298                                                 alloc_end - alloc_start,
3299                                                 i_blocksize(inode),
3300                                                 offset + len, &alloc_hint);
3301                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart,
3302                                      lockend, &cached_state);
3303                 /* btrfs_prealloc_file_range releases reserved space on error */
3304                 if (ret) {
3305                         space_reserved = false;
3306                         goto out;
3307                 }
3308         }
3309         ret = btrfs_fallocate_update_isize(inode, offset + len, mode);
3310  out:
3311         if (ret && space_reserved)
3312                 btrfs_free_reserved_data_space(BTRFS_I(inode), data_reserved,
3313                                                alloc_start, bytes_to_reserve);
3314         extent_changeset_free(data_reserved);
3315
3316         return ret;
3317 }
3318
3319 static long btrfs_fallocate(struct file *file, int mode,
3320                             loff_t offset, loff_t len)
3321 {
3322         struct inode *inode = file_inode(file);
3323         struct extent_state *cached_state = NULL;
3324         struct extent_changeset *data_reserved = NULL;
3325         struct falloc_range *range;
3326         struct falloc_range *tmp;
3327         struct list_head reserve_list;
3328         u64 cur_offset;
3329         u64 last_byte;
3330         u64 alloc_start;
3331         u64 alloc_end;
3332         u64 alloc_hint = 0;
3333         u64 locked_end;
3334         u64 actual_end = 0;
3335         struct extent_map *em;
3336         int blocksize = btrfs_inode_sectorsize(BTRFS_I(inode));
3337         int ret;
3338
3339         /* Do not allow fallocate in ZONED mode */
3340         if (btrfs_is_zoned(btrfs_sb(inode->i_sb)))
3341                 return -EOPNOTSUPP;
3342
3343         alloc_start = round_down(offset, blocksize);
3344         alloc_end = round_up(offset + len, blocksize);
3345         cur_offset = alloc_start;
3346
3347         /* Make sure we aren't being give some crap mode */
3348         if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE |
3349                      FALLOC_FL_ZERO_RANGE))
3350                 return -EOPNOTSUPP;
3351
3352         if (mode & FALLOC_FL_PUNCH_HOLE)
3353                 return btrfs_punch_hole(inode, offset, len);
3354
3355         /*
3356          * Only trigger disk allocation, don't trigger qgroup reserve
3357          *
3358          * For qgroup space, it will be checked later.
3359          */
3360         if (!(mode & FALLOC_FL_ZERO_RANGE)) {
3361                 ret = btrfs_alloc_data_chunk_ondemand(BTRFS_I(inode),
3362                                                       alloc_end - alloc_start);
3363                 if (ret < 0)
3364                         return ret;
3365         }
3366
3367         btrfs_inode_lock(inode, BTRFS_ILOCK_MMAP);
3368
3369         if (!(mode & FALLOC_FL_KEEP_SIZE) && offset + len > inode->i_size) {
3370                 ret = inode_newsize_ok(inode, offset + len);
3371                 if (ret)
3372                         goto out;
3373         }
3374
3375         /*
3376          * TODO: Move these two operations after we have checked
3377          * accurate reserved space, or fallocate can still fail but
3378          * with page truncated or size expanded.
3379          *
3380          * But that's a minor problem and won't do much harm BTW.
3381          */
3382         if (alloc_start > inode->i_size) {
3383                 ret = btrfs_cont_expand(BTRFS_I(inode), i_size_read(inode),
3384                                         alloc_start);
3385                 if (ret)
3386                         goto out;
3387         } else if (offset + len > inode->i_size) {
3388                 /*
3389                  * If we are fallocating from the end of the file onward we
3390                  * need to zero out the end of the block if i_size lands in the
3391                  * middle of a block.
3392                  */
3393                 ret = btrfs_truncate_block(BTRFS_I(inode), inode->i_size, 0, 0);
3394                 if (ret)
3395                         goto out;
3396         }
3397
3398         /*
3399          * wait for ordered IO before we have any locks.  We'll loop again
3400          * below with the locks held.
3401          */
3402         ret = btrfs_wait_ordered_range(inode, alloc_start,
3403                                        alloc_end - alloc_start);
3404         if (ret)
3405                 goto out;
3406
3407         if (mode & FALLOC_FL_ZERO_RANGE) {
3408                 ret = btrfs_zero_range(inode, offset, len, mode);
3409                 btrfs_inode_unlock(inode, BTRFS_ILOCK_MMAP);
3410                 return ret;
3411         }
3412
3413         locked_end = alloc_end - 1;
3414         while (1) {
3415                 struct btrfs_ordered_extent *ordered;
3416
3417                 /* the extent lock is ordered inside the running
3418                  * transaction
3419                  */
3420                 lock_extent_bits(&BTRFS_I(inode)->io_tree, alloc_start,
3421                                  locked_end, &cached_state);
3422                 ordered = btrfs_lookup_first_ordered_extent(BTRFS_I(inode),
3423                                                             locked_end);
3424
3425                 if (ordered &&
3426                     ordered->file_offset + ordered->num_bytes > alloc_start &&
3427                     ordered->file_offset < alloc_end) {
3428                         btrfs_put_ordered_extent(ordered);
3429                         unlock_extent_cached(&BTRFS_I(inode)->io_tree,
3430                                              alloc_start, locked_end,
3431                                              &cached_state);
3432                         /*
3433                          * we can't wait on the range with the transaction
3434                          * running or with the extent lock held
3435                          */
3436                         ret = btrfs_wait_ordered_range(inode, alloc_start,
3437                                                        alloc_end - alloc_start);
3438                         if (ret)
3439                                 goto out;
3440                 } else {
3441                         if (ordered)
3442                                 btrfs_put_ordered_extent(ordered);
3443                         break;
3444                 }
3445         }
3446
3447         /* First, check if we exceed the qgroup limit */
3448         INIT_LIST_HEAD(&reserve_list);
3449         while (cur_offset < alloc_end) {
3450                 em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, cur_offset,
3451                                       alloc_end - cur_offset);
3452                 if (IS_ERR(em)) {
3453                         ret = PTR_ERR(em);
3454                         break;
3455                 }
3456                 last_byte = min(extent_map_end(em), alloc_end);
3457                 actual_end = min_t(u64, extent_map_end(em), offset + len);
3458                 last_byte = ALIGN(last_byte, blocksize);
3459                 if (em->block_start == EXTENT_MAP_HOLE ||
3460                     (cur_offset >= inode->i_size &&
3461                      !test_bit(EXTENT_FLAG_PREALLOC, &em->flags))) {
3462                         ret = add_falloc_range(&reserve_list, cur_offset,
3463                                                last_byte - cur_offset);
3464                         if (ret < 0) {
3465                                 free_extent_map(em);
3466                                 break;
3467                         }
3468                         ret = btrfs_qgroup_reserve_data(BTRFS_I(inode),
3469                                         &data_reserved, cur_offset,
3470                                         last_byte - cur_offset);
3471                         if (ret < 0) {
3472                                 cur_offset = last_byte;
3473                                 free_extent_map(em);
3474                                 break;
3475                         }
3476                 } else {
3477                         /*
3478                          * Do not need to reserve unwritten extent for this
3479                          * range, free reserved data space first, otherwise
3480                          * it'll result in false ENOSPC error.
3481                          */
3482                         btrfs_free_reserved_data_space(BTRFS_I(inode),
3483                                 data_reserved, cur_offset,
3484                                 last_byte - cur_offset);
3485                 }
3486                 free_extent_map(em);
3487                 cur_offset = last_byte;
3488         }
3489
3490         /*
3491          * If ret is still 0, means we're OK to fallocate.
3492          * Or just cleanup the list and exit.
3493          */
3494         list_for_each_entry_safe(range, tmp, &reserve_list, list) {
3495                 if (!ret)
3496                         ret = btrfs_prealloc_file_range(inode, mode,
3497                                         range->start,
3498                                         range->len, i_blocksize(inode),
3499                                         offset + len, &alloc_hint);
3500                 else
3501                         btrfs_free_reserved_data_space(BTRFS_I(inode),
3502                                         data_reserved, range->start,
3503                                         range->len);
3504                 list_del(&range->list);
3505                 kfree(range);
3506         }
3507         if (ret < 0)
3508                 goto out_unlock;
3509
3510         /*
3511          * We didn't need to allocate any more space, but we still extended the
3512          * size of the file so we need to update i_size and the inode item.
3513          */
3514         ret = btrfs_fallocate_update_isize(inode, actual_end, mode);
3515 out_unlock:
3516         unlock_extent_cached(&BTRFS_I(inode)->io_tree, alloc_start, locked_end,
3517                              &cached_state);
3518 out:
3519         btrfs_inode_unlock(inode, BTRFS_ILOCK_MMAP);
3520         /* Let go of our reservation. */
3521         if (ret != 0 && !(mode & FALLOC_FL_ZERO_RANGE))
3522                 btrfs_free_reserved_data_space(BTRFS_I(inode), data_reserved,
3523                                 cur_offset, alloc_end - cur_offset);
3524         extent_changeset_free(data_reserved);
3525         return ret;
3526 }
3527
3528 static loff_t find_desired_extent(struct btrfs_inode *inode, loff_t offset,
3529                                   int whence)
3530 {
3531         struct btrfs_fs_info *fs_info = inode->root->fs_info;
3532         struct extent_map *em = NULL;
3533         struct extent_state *cached_state = NULL;
3534         loff_t i_size = inode->vfs_inode.i_size;
3535         u64 lockstart;
3536         u64 lockend;
3537         u64 start;
3538         u64 len;
3539         int ret = 0;
3540
3541         if (i_size == 0 || offset >= i_size)
3542                 return -ENXIO;
3543
3544         /*
3545          * offset can be negative, in this case we start finding DATA/HOLE from
3546          * the very start of the file.
3547          */
3548         start = max_t(loff_t, 0, offset);
3549
3550         lockstart = round_down(start, fs_info->sectorsize);
3551         lockend = round_up(i_size, fs_info->sectorsize);
3552         if (lockend <= lockstart)
3553                 lockend = lockstart + fs_info->sectorsize;
3554         lockend--;
3555         len = lockend - lockstart + 1;
3556
3557         lock_extent_bits(&inode->io_tree, lockstart, lockend, &cached_state);
3558
3559         while (start < i_size) {
3560                 em = btrfs_get_extent_fiemap(inode, start, len);
3561                 if (IS_ERR(em)) {
3562                         ret = PTR_ERR(em);
3563                         em = NULL;
3564                         break;
3565                 }
3566
3567                 if (whence == SEEK_HOLE &&
3568                     (em->block_start == EXTENT_MAP_HOLE ||
3569                      test_bit(EXTENT_FLAG_PREALLOC, &em->flags)))
3570                         break;
3571                 else if (whence == SEEK_DATA &&
3572                            (em->block_start != EXTENT_MAP_HOLE &&
3573                             !test_bit(EXTENT_FLAG_PREALLOC, &em->flags)))
3574                         break;
3575
3576                 start = em->start + em->len;
3577                 free_extent_map(em);
3578                 em = NULL;
3579                 cond_resched();
3580         }
3581         free_extent_map(em);
3582         unlock_extent_cached(&inode->io_tree, lockstart, lockend,
3583                              &cached_state);
3584         if (ret) {
3585                 offset = ret;
3586         } else {
3587                 if (whence == SEEK_DATA && start >= i_size)
3588                         offset = -ENXIO;
3589                 else
3590                         offset = min_t(loff_t, start, i_size);
3591         }
3592
3593         return offset;
3594 }
3595
3596 static loff_t btrfs_file_llseek(struct file *file, loff_t offset, int whence)
3597 {
3598         struct inode *inode = file->f_mapping->host;
3599
3600         switch (whence) {
3601         default:
3602                 return generic_file_llseek(file, offset, whence);
3603         case SEEK_DATA:
3604         case SEEK_HOLE:
3605                 btrfs_inode_lock(inode, BTRFS_ILOCK_SHARED);
3606                 offset = find_desired_extent(BTRFS_I(inode), offset, whence);
3607                 btrfs_inode_unlock(inode, BTRFS_ILOCK_SHARED);
3608                 break;
3609         }
3610
3611         if (offset < 0)
3612                 return offset;
3613
3614         return vfs_setpos(file, offset, inode->i_sb->s_maxbytes);
3615 }
3616
3617 static int btrfs_file_open(struct inode *inode, struct file *filp)
3618 {
3619         int ret;
3620
3621         filp->f_mode |= FMODE_NOWAIT | FMODE_BUF_RASYNC;
3622
3623         ret = fsverity_file_open(inode, filp);
3624         if (ret)
3625                 return ret;
3626         return generic_file_open(inode, filp);
3627 }
3628
3629 static int check_direct_read(struct btrfs_fs_info *fs_info,
3630                              const struct iov_iter *iter, loff_t offset)
3631 {
3632         int ret;
3633         int i, seg;
3634
3635         ret = check_direct_IO(fs_info, iter, offset);
3636         if (ret < 0)
3637                 return ret;
3638
3639         if (!iter_is_iovec(iter))
3640                 return 0;
3641
3642         for (seg = 0; seg < iter->nr_segs; seg++)
3643                 for (i = seg + 1; i < iter->nr_segs; i++)
3644                         if (iter->iov[seg].iov_base == iter->iov[i].iov_base)
3645                                 return -EINVAL;
3646         return 0;
3647 }
3648
3649 static ssize_t btrfs_direct_read(struct kiocb *iocb, struct iov_iter *to)
3650 {
3651         struct inode *inode = file_inode(iocb->ki_filp);
3652         ssize_t ret;
3653
3654         if (fsverity_active(inode))
3655                 return 0;
3656
3657         if (check_direct_read(btrfs_sb(inode->i_sb), to, iocb->ki_pos))
3658                 return 0;
3659
3660         btrfs_inode_lock(inode, BTRFS_ILOCK_SHARED);
3661         ret = iomap_dio_rw(iocb, to, &btrfs_dio_iomap_ops, &btrfs_dio_ops, 0);
3662         btrfs_inode_unlock(inode, BTRFS_ILOCK_SHARED);
3663         return ret;
3664 }
3665
3666 static ssize_t btrfs_file_read_iter(struct kiocb *iocb, struct iov_iter *to)
3667 {
3668         ssize_t ret = 0;
3669
3670         if (iocb->ki_flags & IOCB_DIRECT) {
3671                 ret = btrfs_direct_read(iocb, to);
3672                 if (ret < 0 || !iov_iter_count(to) ||
3673                     iocb->ki_pos >= i_size_read(file_inode(iocb->ki_filp)))
3674                         return ret;
3675         }
3676
3677         return filemap_read(iocb, to, ret);
3678 }
3679
3680 const struct file_operations btrfs_file_operations = {
3681         .llseek         = btrfs_file_llseek,
3682         .read_iter      = btrfs_file_read_iter,
3683         .splice_read    = generic_file_splice_read,
3684         .write_iter     = btrfs_file_write_iter,
3685         .splice_write   = iter_file_splice_write,
3686         .mmap           = btrfs_file_mmap,
3687         .open           = btrfs_file_open,
3688         .release        = btrfs_release_file,
3689         .fsync          = btrfs_sync_file,
3690         .fallocate      = btrfs_fallocate,
3691         .unlocked_ioctl = btrfs_ioctl,
3692 #ifdef CONFIG_COMPAT
3693         .compat_ioctl   = btrfs_compat_ioctl,
3694 #endif
3695         .remap_file_range = btrfs_remap_file_range,
3696 };
3697
3698 void __cold btrfs_auto_defrag_exit(void)
3699 {
3700         kmem_cache_destroy(btrfs_inode_defrag_cachep);
3701 }
3702
3703 int __init btrfs_auto_defrag_init(void)
3704 {
3705         btrfs_inode_defrag_cachep = kmem_cache_create("btrfs_inode_defrag",
3706                                         sizeof(struct inode_defrag), 0,
3707                                         SLAB_MEM_SPREAD,
3708                                         NULL);
3709         if (!btrfs_inode_defrag_cachep)
3710                 return -ENOMEM;
3711
3712         return 0;
3713 }
3714
3715 int btrfs_fdatawrite_range(struct inode *inode, loff_t start, loff_t end)
3716 {
3717         int ret;
3718
3719         /*
3720          * So with compression we will find and lock a dirty page and clear the
3721          * first one as dirty, setup an async extent, and immediately return
3722          * with the entire range locked but with nobody actually marked with
3723          * writeback.  So we can't just filemap_write_and_wait_range() and
3724          * expect it to work since it will just kick off a thread to do the
3725          * actual work.  So we need to call filemap_fdatawrite_range _again_
3726          * since it will wait on the page lock, which won't be unlocked until
3727          * after the pages have been marked as writeback and so we're good to go
3728          * from there.  We have to do this otherwise we'll miss the ordered
3729          * extents and that results in badness.  Please Josef, do not think you
3730          * know better and pull this out at some point in the future, it is
3731          * right and you are wrong.
3732          */
3733         ret = filemap_fdatawrite_range(inode->i_mapping, start, end);
3734         if (!ret && test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
3735                              &BTRFS_I(inode)->runtime_flags))
3736                 ret = filemap_fdatawrite_range(inode->i_mapping, start, end);
3737
3738         return ret;
3739 }