fs/btrfs/ordered-data.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Copyright (C) 2007 Oracle.  All rights reserved.
   4  */
   5
   6 #include <linux/slab.h>
   7 #include <linux/blkdev.h>
   8 #include <linux/writeback.h>
   9 #include <linux/sched/mm.h>
  10 #include "messages.h"
  11 #include "misc.h"
  12 #include "ctree.h"
  13 #include "transaction.h"
  14 #include "btrfs_inode.h"
  15 #include "extent_io.h"
  16 #include "disk-io.h"
  17 #include "compression.h"
  18 #include "delalloc-space.h"
  19 #include "qgroup.h"
  20 #include "subpage.h"
  21 #include "file.h"
  22 #include "super.h"
  23
  24 static struct kmem_cache *btrfs_ordered_extent_cache;
  25
  26 static u64 entry_end(struct btrfs_ordered_extent *entry)
  27 {
  28         if (entry->file_offset + entry->num_bytes < entry->file_offset)
  29                 return (u64)-1;
  30         return entry->file_offset + entry->num_bytes;
  31 }
  32
  33 /* returns NULL if the insertion worked, or it returns the node it did find
  34  * in the tree
  35  */
  36 static struct rb_node *tree_insert(struct rb_root *root, u64 file_offset,
  37                                    struct rb_node *node)
  38 {
  39         struct rb_node **p = &root->rb_node;
  40         struct rb_node *parent = NULL;
  41         struct btrfs_ordered_extent *entry;
  42
  43         while (*p) {
  44                 parent = *p;
  45                 entry = rb_entry(parent, struct btrfs_ordered_extent, rb_node);
  46
  47                 if (file_offset < entry->file_offset)
  48                         p = &(*p)->rb_left;
  49                 else if (file_offset >= entry_end(entry))
  50                         p = &(*p)->rb_right;
  51                 else
  52                         return parent;
  53         }
  54
  55         rb_link_node(node, parent, p);
  56         rb_insert_color(node, root);
  57         return NULL;
  58 }
  59
  60 /*
  61  * look for a given offset in the tree, and if it can't be found return the
  62  * first lesser offset
  63  */
  64 static struct rb_node *__tree_search(struct rb_root *root, u64 file_offset,
  65                                      struct rb_node **prev_ret)
  66 {
  67         struct rb_node *n = root->rb_node;
  68         struct rb_node *prev = NULL;
  69         struct rb_node *test;
  70         struct btrfs_ordered_extent *entry;
  71         struct btrfs_ordered_extent *prev_entry = NULL;
  72
  73         while (n) {
  74                 entry = rb_entry(n, struct btrfs_ordered_extent, rb_node);
  75                 prev = n;
  76                 prev_entry = entry;
  77
  78                 if (file_offset < entry->file_offset)
  79                         n = n->rb_left;
  80                 else if (file_offset >= entry_end(entry))
  81                         n = n->rb_right;
  82                 else
  83                         return n;
  84         }
  85         if (!prev_ret)
  86                 return NULL;
  87
  88         while (prev && file_offset >= entry_end(prev_entry)) {
  89                 test = rb_next(prev);
  90                 if (!test)
  91                         break;
  92                 prev_entry = rb_entry(test, struct btrfs_ordered_extent,
  93                                       rb_node);
  94                 if (file_offset < entry_end(prev_entry))
  95                         break;
  96
  97                 prev = test;
  98         }
  99         if (prev)
 100                 prev_entry = rb_entry(prev, struct btrfs_ordered_extent,
 101                                       rb_node);
 102         while (prev && file_offset < entry_end(prev_entry)) {
 103                 test = rb_prev(prev);
 104                 if (!test)
 105                         break;
 106                 prev_entry = rb_entry(test, struct btrfs_ordered_extent,
 107                                       rb_node);
 108                 prev = test;
 109         }
 110         *prev_ret = prev;
 111         return NULL;
 112 }
 113
 114 static int range_overlaps(struct btrfs_ordered_extent *entry, u64 file_offset,
 115                           u64 len)
 116 {
 117         if (file_offset + len <= entry->file_offset ||
 118             entry->file_offset + entry->num_bytes <= file_offset)
 119                 return 0;
 120         return 1;
 121 }
 122
 123 /*
 124  * look find the first ordered struct that has this offset, otherwise
 125  * the first one less than this offset
 126  */
 127 static inline struct rb_node *tree_search(struct btrfs_ordered_inode_tree *tree,
 128                                           u64 file_offset)
 129 {
 130         struct rb_root *root = &tree->tree;
 131         struct rb_node *prev = NULL;
 132         struct rb_node *ret;
 133         struct btrfs_ordered_extent *entry;
 134
 135         if (tree->last) {
 136                 entry = rb_entry(tree->last, struct btrfs_ordered_extent,
 137                                  rb_node);
 138                 if (in_range(file_offset, entry->file_offset, entry->num_bytes))
 139                         return tree->last;
 140         }
 141         ret = __tree_search(root, file_offset, &prev);
 142         if (!ret)
 143                 ret = prev;
 144         if (ret)
 145                 tree->last = ret;
 146         return ret;
 147 }
 148
 149 /*
 150  * Add an ordered extent to the per-inode tree.
 151  *
 152  * @inode:           Inode that this extent is for.
 153  * @file_offset:     Logical offset in file where the extent starts.
 154  * @num_bytes:       Logical length of extent in file.
 155  * @ram_bytes:       Full length of unencoded data.
 156  * @disk_bytenr:     Offset of extent on disk.
 157  * @disk_num_bytes:  Size of extent on disk.
 158  * @offset:          Offset into unencoded data where file data starts.
 159  * @flags:           Flags specifying type of extent (1 << BTRFS_ORDERED_*).
 160  * @compress_type:   Compression algorithm used for data.
 161  *
 162  * Most of these parameters correspond to &struct btrfs_file_extent_item. The
 163  * tree is given a single reference on the ordered extent that was inserted.
 164  *
 165  * Return: 0 or -ENOMEM.
 166  */
 167 int btrfs_add_ordered_extent(struct btrfs_inode *inode, u64 file_offset,
 168                              u64 num_bytes, u64 ram_bytes, u64 disk_bytenr,
 169                              u64 disk_num_bytes, u64 offset, unsigned flags,
 170                              int compress_type)
 171 {
 172         struct btrfs_root *root = inode->root;
 173         struct btrfs_fs_info *fs_info = root->fs_info;
 174         struct btrfs_ordered_inode_tree *tree = &inode->ordered_tree;
 175         struct rb_node *node;
 176         struct btrfs_ordered_extent *entry;
 177         int ret;
 178
 179         if (flags &
 180             ((1 << BTRFS_ORDERED_NOCOW) | (1 << BTRFS_ORDERED_PREALLOC))) {
 181                 /* For nocow write, we can release the qgroup rsv right now */
 182                 ret = btrfs_qgroup_free_data(inode, NULL, file_offset, num_bytes);
 183                 if (ret < 0)
 184                         return ret;
 185                 ret = 0;
 186         } else {
 187                 /*
 188                  * The ordered extent has reserved qgroup space, release now
 189                  * and pass the reserved number for qgroup_record to free.
 190                  */
 191                 ret = btrfs_qgroup_release_data(inode, file_offset, num_bytes);
 192                 if (ret < 0)
 193                         return ret;
 194         }
 195         entry = kmem_cache_zalloc(btrfs_ordered_extent_cache, GFP_NOFS);
 196         if (!entry)
 197                 return -ENOMEM;
 198
 199         entry->file_offset = file_offset;
 200         entry->num_bytes = num_bytes;
 201         entry->ram_bytes = ram_bytes;
 202         entry->disk_bytenr = disk_bytenr;
 203         entry->disk_num_bytes = disk_num_bytes;
 204         entry->offset = offset;
 205         entry->bytes_left = num_bytes;
 206         entry->inode = igrab(&inode->vfs_inode);
 207         entry->compress_type = compress_type;
 208         entry->truncated_len = (u64)-1;
 209         entry->qgroup_rsv = ret;
 210         entry->physical = (u64)-1;
 211
 212         ASSERT((flags & ~BTRFS_ORDERED_TYPE_FLAGS) == 0);
 213         entry->flags = flags;
 214
 215         percpu_counter_add_batch(&fs_info->ordered_bytes, num_bytes,
 216                                  fs_info->delalloc_batch);
 217
 218         /* one ref for the tree */
 219         refcount_set(&entry->refs, 1);
 220         init_waitqueue_head(&entry->wait);
 221         INIT_LIST_HEAD(&entry->list);
 222         INIT_LIST_HEAD(&entry->log_list);
 223         INIT_LIST_HEAD(&entry->root_extent_list);
 224         INIT_LIST_HEAD(&entry->work_list);
 225         init_completion(&entry->completion);
 226
 227         trace_btrfs_ordered_extent_add(inode, entry);
 228
 229         spin_lock_irq(&tree->lock);
 230         node = tree_insert(&tree->tree, file_offset,
 231                            &entry->rb_node);
 232         if (node)
 233                 btrfs_panic(fs_info, -EEXIST,
 234                                 "inconsistency in ordered tree at offset %llu",
 235                                 file_offset);
 236         spin_unlock_irq(&tree->lock);
 237
 238         spin_lock(&root->ordered_extent_lock);
 239         list_add_tail(&entry->root_extent_list,
 240                       &root->ordered_extents);
 241         root->nr_ordered_extents++;
 242         if (root->nr_ordered_extents == 1) {
 243                 spin_lock(&fs_info->ordered_root_lock);
 244                 BUG_ON(!list_empty(&root->ordered_root));
 245                 list_add_tail(&root->ordered_root, &fs_info->ordered_roots);
 246                 spin_unlock(&fs_info->ordered_root_lock);
 247         }
 248         spin_unlock(&root->ordered_extent_lock);
 249
 250         /*
 251          * We don't need the count_max_extents here, we can assume that all of
 252          * that work has been done at higher layers, so this is truly the
 253          * smallest the extent is going to get.
 254          */
 255         spin_lock(&inode->lock);
 256         btrfs_mod_outstanding_extents(inode, 1);
 257         spin_unlock(&inode->lock);
 258
 259         return 0;
 260 }
 261
 262 /*
 263  * Add a struct btrfs_ordered_sum into the list of checksums to be inserted
 264  * when an ordered extent is finished.  If the list covers more than one
 265  * ordered extent, it is split across multiples.
 266  */
 267 void btrfs_add_ordered_sum(struct btrfs_ordered_extent *entry,
 268                            struct btrfs_ordered_sum *sum)
 269 {
 270         struct btrfs_ordered_inode_tree *tree;
 271
 272         tree = &BTRFS_I(entry->inode)->ordered_tree;
 273         spin_lock_irq(&tree->lock);
 274         list_add_tail(&sum->list, &entry->list);
 275         spin_unlock_irq(&tree->lock);
 276 }
 277
 278 static void finish_ordered_fn(struct btrfs_work *work)
 279 {
 280         struct btrfs_ordered_extent *ordered_extent;
 281
 282         ordered_extent = container_of(work, struct btrfs_ordered_extent, work);
 283         btrfs_finish_ordered_io(ordered_extent);
 284 }
 285
 286 /*
 287  * Mark all ordered extents io inside the specified range finished.
 288  *
 289  * @page:        The involved page for the operation.
 290  *               For uncompressed buffered IO, the page status also needs to be
 291  *               updated to indicate whether the pending ordered io is finished.
 292  *               Can be NULL for direct IO and compressed write.
 293  *               For these cases, callers are ensured they won't execute the
 294  *               endio function twice.
 295  *
 296  * This function is called for endio, thus the range must have ordered
 297  * extent(s) covering it.
 298  */
 299 void btrfs_mark_ordered_io_finished(struct btrfs_inode *inode,
 300                                     struct page *page, u64 file_offset,
 301                                     u64 num_bytes, bool uptodate)
 302 {
 303         struct btrfs_ordered_inode_tree *tree = &inode->ordered_tree;
 304         struct btrfs_fs_info *fs_info = inode->root->fs_info;
 305         struct btrfs_workqueue *wq;
 306         struct rb_node *node;
 307         struct btrfs_ordered_extent *entry = NULL;
 308         unsigned long flags;
 309         u64 cur = file_offset;
 310
 311         if (btrfs_is_free_space_inode(inode))
 312                 wq = fs_info->endio_freespace_worker;
 313         else
 314                 wq = fs_info->endio_write_workers;
 315
 316         if (page)
 317                 ASSERT(page->mapping && page_offset(page) <= file_offset &&
 318                        file_offset + num_bytes <= page_offset(page) + PAGE_SIZE);
 319
 320         spin_lock_irqsave(&tree->lock, flags);
 321         while (cur < file_offset + num_bytes) {
 322                 u64 entry_end;
 323                 u64 end;
 324                 u32 len;
 325
 326                 node = tree_search(tree, cur);
 327                 /* No ordered extents at all */
 328                 if (!node)
 329                         break;
 330
 331                 entry = rb_entry(node, struct btrfs_ordered_extent, rb_node);
 332                 entry_end = entry->file_offset + entry->num_bytes;
 333                 /*
 334                  * |<-- OE --->|  |
 335                  *                cur
 336                  * Go to next OE.
 337                  */
 338                 if (cur >= entry_end) {
 339                         node = rb_next(node);
 340                         /* No more ordered extents, exit */
 341                         if (!node)
 342                                 break;
 343                         entry = rb_entry(node, struct btrfs_ordered_extent,
 344                                          rb_node);
 345
 346                         /* Go to next ordered extent and continue */
 347                         cur = entry->file_offset;
 348                         continue;
 349                 }
 350                 /*
 351                  * |    |<--- OE --->|
 352                  * cur
 353                  * Go to the start of OE.
 354                  */
 355                 if (cur < entry->file_offset) {
 356                         cur = entry->file_offset;
 357                         continue;
 358                 }
 359
 360                 /*
 361                  * Now we are definitely inside one ordered extent.
 362                  *
 363                  * |<--- OE --->|
 364                  *      |
 365                  *      cur
 366                  */
 367                 end = min(entry->file_offset + entry->num_bytes,
 368                           file_offset + num_bytes) - 1;
 369                 ASSERT(end + 1 - cur < U32_MAX);
 370                 len = end + 1 - cur;
 371
 372                 if (page) {
 373                         /*
 374                          * Ordered (Private2) bit indicates whether we still
 375                          * have pending io unfinished for the ordered extent.
 376                          *
 377                          * If there's no such bit, we need to skip to next range.
 378                          */
 379                         if (!btrfs_page_test_ordered(fs_info, page, cur, len)) {
 380                                 cur += len;
 381                                 continue;
 382                         }
 383                         btrfs_page_clear_ordered(fs_info, page, cur, len);
 384                 }
 385
 386                 /* Now we're fine to update the accounting */
 387                 if (unlikely(len > entry->bytes_left)) {
 388                         WARN_ON(1);
 389                         btrfs_crit(fs_info,
 390 "bad ordered extent accounting, root=%llu ino=%llu OE offset=%llu OE len=%llu to_dec=%u left=%llu",
 391                                    inode->root->root_key.objectid,
 392                                    btrfs_ino(inode),
 393                                    entry->file_offset,
 394                                    entry->num_bytes,
 395                                    len, entry->bytes_left);
 396                         entry->bytes_left = 0;
 397                 } else {
 398                         entry->bytes_left -= len;
 399                 }
 400
 401                 if (!uptodate)
 402                         set_bit(BTRFS_ORDERED_IOERR, &entry->flags);
 403
 404                 /*
 405                  * All the IO of the ordered extent is finished, we need to queue
 406                  * the finish_func to be executed.
 407                  */
 408                 if (entry->bytes_left == 0) {
 409                         set_bit(BTRFS_ORDERED_IO_DONE, &entry->flags);
 410                         cond_wake_up(&entry->wait);
 411                         refcount_inc(&entry->refs);
 412                         trace_btrfs_ordered_extent_mark_finished(inode, entry);
 413                         spin_unlock_irqrestore(&tree->lock, flags);
 414                         btrfs_init_work(&entry->work, finish_ordered_fn, NULL, NULL);
 415                         btrfs_queue_work(wq, &entry->work);
 416                         spin_lock_irqsave(&tree->lock, flags);
 417                 }
 418                 cur += len;
 419         }
 420         spin_unlock_irqrestore(&tree->lock, flags);
 421 }
 422
 423 /*
 424  * Finish IO for one ordered extent across a given range.  The range can only
 425  * contain one ordered extent.
 426  *
 427  * @cached:      The cached ordered extent. If not NULL, we can skip the tree
 428  *               search and use the ordered extent directly.
 429  *               Will be also used to store the finished ordered extent.
 430  * @file_offset: File offset for the finished IO
 431  * @io_size:     Length of the finish IO range
 432  *
 433  * Return true if the ordered extent is finished in the range, and update
 434  * @cached.
 435  * Return false otherwise.
 436  *
 437  * NOTE: The range can NOT cross multiple ordered extents.
 438  * Thus caller should ensure the range doesn't cross ordered extents.
 439  */
 440 bool btrfs_dec_test_ordered_pending(struct btrfs_inode *inode,
 441                                     struct btrfs_ordered_extent **cached,
 442                                     u64 file_offset, u64 io_size)
 443 {
 444         struct btrfs_ordered_inode_tree *tree = &inode->ordered_tree;
 445         struct rb_node *node;
 446         struct btrfs_ordered_extent *entry = NULL;
 447         unsigned long flags;
 448         bool finished = false;
 449
 450         spin_lock_irqsave(&tree->lock, flags);
 451         if (cached && *cached) {
 452                 entry = *cached;
 453                 goto have_entry;
 454         }
 455
 456         node = tree_search(tree, file_offset);
 457         if (!node)
 458                 goto out;
 459
 460         entry = rb_entry(node, struct btrfs_ordered_extent, rb_node);
 461 have_entry:
 462         if (!in_range(file_offset, entry->file_offset, entry->num_bytes))
 463                 goto out;
 464
 465         if (io_size > entry->bytes_left)
 466                 btrfs_crit(inode->root->fs_info,
 467                            "bad ordered accounting left %llu size %llu",
 468                        entry->bytes_left, io_size);
 469
 470         entry->bytes_left -= io_size;
 471
 472         if (entry->bytes_left == 0) {
 473                 /*
 474                  * Ensure only one caller can set the flag and finished_ret
 475                  * accordingly
 476                  */
 477                 finished = !test_and_set_bit(BTRFS_ORDERED_IO_DONE, &entry->flags);
 478                 /* test_and_set_bit implies a barrier */
 479                 cond_wake_up_nomb(&entry->wait);
 480         }
 481 out:
 482         if (finished && cached && entry) {
 483                 *cached = entry;
 484                 refcount_inc(&entry->refs);
 485                 trace_btrfs_ordered_extent_dec_test_pending(inode, entry);
 486         }
 487         spin_unlock_irqrestore(&tree->lock, flags);
 488         return finished;
 489 }
 490
 491 /*
 492  * used to drop a reference on an ordered extent.  This will free
 493  * the extent if the last reference is dropped
 494  */
 495 void btrfs_put_ordered_extent(struct btrfs_ordered_extent *entry)
 496 {
 497         struct list_head *cur;
 498         struct btrfs_ordered_sum *sum;
 499
 500         trace_btrfs_ordered_extent_put(BTRFS_I(entry->inode), entry);
 501
 502         if (refcount_dec_and_test(&entry->refs)) {
 503                 ASSERT(list_empty(&entry->root_extent_list));
 504                 ASSERT(list_empty(&entry->log_list));
 505                 ASSERT(RB_EMPTY_NODE(&entry->rb_node));
 506                 if (entry->inode)
 507                         btrfs_add_delayed_iput(BTRFS_I(entry->inode));
 508                 while (!list_empty(&entry->list)) {
 509                         cur = entry->list.next;
 510                         sum = list_entry(cur, struct btrfs_ordered_sum, list);
 511                         list_del(&sum->list);
 512                         kvfree(sum);
 513                 }
 514                 kmem_cache_free(btrfs_ordered_extent_cache, entry);
 515         }
 516 }
 517
 518 /*
 519  * remove an ordered extent from the tree.  No references are dropped
 520  * and waiters are woken up.
 521  */
 522 void btrfs_remove_ordered_extent(struct btrfs_inode *btrfs_inode,
 523                                  struct btrfs_ordered_extent *entry)
 524 {
 525         struct btrfs_ordered_inode_tree *tree;
 526         struct btrfs_root *root = btrfs_inode->root;
 527         struct btrfs_fs_info *fs_info = root->fs_info;
 528         struct rb_node *node;
 529         bool pending;
 530         bool freespace_inode;
 531
 532         /*
 533          * If this is a free space inode the thread has not acquired the ordered
 534          * extents lockdep map.
 535          */
 536         freespace_inode = btrfs_is_free_space_inode(btrfs_inode);
 537
 538         btrfs_lockdep_acquire(fs_info, btrfs_trans_pending_ordered);
 539         /* This is paired with btrfs_add_ordered_extent. */
 540         spin_lock(&btrfs_inode->lock);
 541         btrfs_mod_outstanding_extents(btrfs_inode, -1);
 542         spin_unlock(&btrfs_inode->lock);
 543         if (root != fs_info->tree_root) {
 544                 u64 release;
 545
 546                 if (test_bit(BTRFS_ORDERED_ENCODED, &entry->flags))
 547                         release = entry->disk_num_bytes;
 548                 else
 549                         release = entry->num_bytes;
 550                 btrfs_delalloc_release_metadata(btrfs_inode, release, false);
 551         }
 552
 553         percpu_counter_add_batch(&fs_info->ordered_bytes, -entry->num_bytes,
 554                                  fs_info->delalloc_batch);
 555
 556         tree = &btrfs_inode->ordered_tree;
 557         spin_lock_irq(&tree->lock);
 558         node = &entry->rb_node;
 559         rb_erase(node, &tree->tree);
 560         RB_CLEAR_NODE(node);
 561         if (tree->last == node)
 562                 tree->last = NULL;
 563         set_bit(BTRFS_ORDERED_COMPLETE, &entry->flags);
 564         pending = test_and_clear_bit(BTRFS_ORDERED_PENDING, &entry->flags);
 565         spin_unlock_irq(&tree->lock);
 566
 567         /*
 568          * The current running transaction is waiting on us, we need to let it
 569          * know that we're complete and wake it up.
 570          */
 571         if (pending) {
 572                 struct btrfs_transaction *trans;
 573
 574                 /*
 575                  * The checks for trans are just a formality, it should be set,
 576                  * but if it isn't we don't want to deref/assert under the spin
 577                  * lock, so be nice and check if trans is set, but ASSERT() so
 578                  * if it isn't set a developer will notice.
 579                  */
 580                 spin_lock(&fs_info->trans_lock);
 581                 trans = fs_info->running_transaction;
 582                 if (trans)
 583                         refcount_inc(&trans->use_count);
 584                 spin_unlock(&fs_info->trans_lock);
 585
 586                 ASSERT(trans);
 587                 if (trans) {
 588                         if (atomic_dec_and_test(&trans->pending_ordered))
 589                                 wake_up(&trans->pending_wait);
 590                         btrfs_put_transaction(trans);
 591                 }
 592         }
 593
 594         btrfs_lockdep_release(fs_info, btrfs_trans_pending_ordered);
 595
 596         spin_lock(&root->ordered_extent_lock);
 597         list_del_init(&entry->root_extent_list);
 598         root->nr_ordered_extents--;
 599
 600         trace_btrfs_ordered_extent_remove(btrfs_inode, entry);
 601
 602         if (!root->nr_ordered_extents) {
 603                 spin_lock(&fs_info->ordered_root_lock);
 604                 BUG_ON(list_empty(&root->ordered_root));
 605                 list_del_init(&root->ordered_root);
 606                 spin_unlock(&fs_info->ordered_root_lock);
 607         }
 608         spin_unlock(&root->ordered_extent_lock);
 609         wake_up(&entry->wait);
 610         if (!freespace_inode)
 611                 btrfs_lockdep_release(fs_info, btrfs_ordered_extent);
 612 }
 613
 614 static void btrfs_run_ordered_extent_work(struct btrfs_work *work)
 615 {
 616         struct btrfs_ordered_extent *ordered;
 617
 618         ordered = container_of(work, struct btrfs_ordered_extent, flush_work);
 619         btrfs_start_ordered_extent(ordered);
 620         complete(&ordered->completion);
 621 }
 622
 623 /*
 624  * wait for all the ordered extents in a root.  This is done when balancing
 625  * space between drives.
 626  */
 627 u64 btrfs_wait_ordered_extents(struct btrfs_root *root, u64 nr,
 628                                const u64 range_start, const u64 range_len)
 629 {
 630         struct btrfs_fs_info *fs_info = root->fs_info;
 631         LIST_HEAD(splice);
 632         LIST_HEAD(skipped);
 633         LIST_HEAD(works);
 634         struct btrfs_ordered_extent *ordered, *next;
 635         u64 count = 0;
 636         const u64 range_end = range_start + range_len;
 637
 638         mutex_lock(&root->ordered_extent_mutex);
 639         spin_lock(&root->ordered_extent_lock);
 640         list_splice_init(&root->ordered_extents, &splice);
 641         while (!list_empty(&splice) && nr) {
 642                 ordered = list_first_entry(&splice, struct btrfs_ordered_extent,
 643                                            root_extent_list);
 644
 645                 if (range_end <= ordered->disk_bytenr ||
 646                     ordered->disk_bytenr + ordered->disk_num_bytes <= range_start) {
 647                         list_move_tail(&ordered->root_extent_list, &skipped);
 648                         cond_resched_lock(&root->ordered_extent_lock);
 649                         continue;
 650                 }
 651
 652                 list_move_tail(&ordered->root_extent_list,
 653                                &root->ordered_extents);
 654                 refcount_inc(&ordered->refs);
 655                 spin_unlock(&root->ordered_extent_lock);
 656
 657                 btrfs_init_work(&ordered->flush_work,
 658                                 btrfs_run_ordered_extent_work, NULL, NULL);
 659                 list_add_tail(&ordered->work_list, &works);
 660                 btrfs_queue_work(fs_info->flush_workers, &ordered->flush_work);
 661
 662                 cond_resched();
 663                 spin_lock(&root->ordered_extent_lock);
 664                 if (nr != U64_MAX)
 665                         nr--;
 666                 count++;
 667         }
 668         list_splice_tail(&skipped, &root->ordered_extents);
 669         list_splice_tail(&splice, &root->ordered_extents);
 670         spin_unlock(&root->ordered_extent_lock);
 671
 672         list_for_each_entry_safe(ordered, next, &works, work_list) {
 673                 list_del_init(&ordered->work_list);
 674                 wait_for_completion(&ordered->completion);
 675                 btrfs_put_ordered_extent(ordered);
 676                 cond_resched();
 677         }
 678         mutex_unlock(&root->ordered_extent_mutex);
 679
 680         return count;
 681 }
 682
 683 void btrfs_wait_ordered_roots(struct btrfs_fs_info *fs_info, u64 nr,
 684                              const u64 range_start, const u64 range_len)
 685 {
 686         struct btrfs_root *root;
 687         struct list_head splice;
 688         u64 done;
 689
 690         INIT_LIST_HEAD(&splice);
 691
 692         mutex_lock(&fs_info->ordered_operations_mutex);
 693         spin_lock(&fs_info->ordered_root_lock);
 694         list_splice_init(&fs_info->ordered_roots, &splice);
 695         while (!list_empty(&splice) && nr) {
 696                 root = list_first_entry(&splice, struct btrfs_root,
 697                                         ordered_root);
 698                 root = btrfs_grab_root(root);
 699                 BUG_ON(!root);
 700                 list_move_tail(&root->ordered_root,
 701                                &fs_info->ordered_roots);
 702                 spin_unlock(&fs_info->ordered_root_lock);
 703
 704                 done = btrfs_wait_ordered_extents(root, nr,
 705                                                   range_start, range_len);
 706                 btrfs_put_root(root);
 707
 708                 spin_lock(&fs_info->ordered_root_lock);
 709                 if (nr != U64_MAX) {
 710                         nr -= done;
 711                 }
 712         }
 713         list_splice_tail(&splice, &fs_info->ordered_roots);
 714         spin_unlock(&fs_info->ordered_root_lock);
 715         mutex_unlock(&fs_info->ordered_operations_mutex);
 716 }
 717
 718 /*
 719  * Start IO and wait for a given ordered extent to finish.
 720  *
 721  * Wait on page writeback for all the pages in the extent and the IO completion
 722  * code to insert metadata into the btree corresponding to the extent.
 723  */
 724 void btrfs_start_ordered_extent(struct btrfs_ordered_extent *entry)
 725 {
 726         u64 start = entry->file_offset;
 727         u64 end = start + entry->num_bytes - 1;
 728         struct btrfs_inode *inode = BTRFS_I(entry->inode);
 729         bool freespace_inode;
 730
 731         trace_btrfs_ordered_extent_start(inode, entry);
 732
 733         /*
 734          * If this is a free space inode do not take the ordered extents lockdep
 735          * map.
 736          */
 737         freespace_inode = btrfs_is_free_space_inode(inode);
 738
 739         /*
 740          * pages in the range can be dirty, clean or writeback.  We
 741          * start IO on any dirty ones so the wait doesn't stall waiting
 742          * for the flusher thread to find them
 743          */
 744         if (!test_bit(BTRFS_ORDERED_DIRECT, &entry->flags))
 745                 filemap_fdatawrite_range(inode->vfs_inode.i_mapping, start, end);
 746
 747         if (!freespace_inode)
 748                 btrfs_might_wait_for_event(inode->root->fs_info, btrfs_ordered_extent);
 749         wait_event(entry->wait, test_bit(BTRFS_ORDERED_COMPLETE, &entry->flags));
 750 }
 751
 752 /*
 753  * Used to wait on ordered extents across a large range of bytes.
 754  */
 755 int btrfs_wait_ordered_range(struct inode *inode, u64 start, u64 len)
 756 {
 757         int ret = 0;
 758         int ret_wb = 0;
 759         u64 end;
 760         u64 orig_end;
 761         struct btrfs_ordered_extent *ordered;
 762
 763         if (start + len < start) {
 764                 orig_end = OFFSET_MAX;
 765         } else {
 766                 orig_end = start + len - 1;
 767                 if (orig_end > OFFSET_MAX)
 768                         orig_end = OFFSET_MAX;
 769         }
 770
 771         /* start IO across the range first to instantiate any delalloc
 772          * extents
 773          */
 774         ret = btrfs_fdatawrite_range(inode, start, orig_end);
 775         if (ret)
 776                 return ret;
 777
 778         /*
 779          * If we have a writeback error don't return immediately. Wait first
 780          * for any ordered extents that haven't completed yet. This is to make
 781          * sure no one can dirty the same page ranges and call writepages()
 782          * before the ordered extents complete - to avoid failures (-EEXIST)
 783          * when adding the new ordered extents to the ordered tree.
 784          */
 785         ret_wb = filemap_fdatawait_range(inode->i_mapping, start, orig_end);
 786
 787         end = orig_end;
 788         while (1) {
 789                 ordered = btrfs_lookup_first_ordered_extent(BTRFS_I(inode), end);
 790                 if (!ordered)
 791                         break;
 792                 if (ordered->file_offset > orig_end) {
 793                         btrfs_put_ordered_extent(ordered);
 794                         break;
 795                 }
 796                 if (ordered->file_offset + ordered->num_bytes <= start) {
 797                         btrfs_put_ordered_extent(ordered);
 798                         break;
 799                 }
 800                 btrfs_start_ordered_extent(ordered);
 801                 end = ordered->file_offset;
 802                 /*
 803                  * If the ordered extent had an error save the error but don't
 804                  * exit without waiting first for all other ordered extents in
 805                  * the range to complete.
 806                  */
 807                 if (test_bit(BTRFS_ORDERED_IOERR, &ordered->flags))
 808                         ret = -EIO;
 809                 btrfs_put_ordered_extent(ordered);
 810                 if (end == 0 || end == start)
 811                         break;
 812                 end--;
 813         }
 814         return ret_wb ? ret_wb : ret;
 815 }
 816
 817 /*
 818  * find an ordered extent corresponding to file_offset.  return NULL if
 819  * nothing is found, otherwise take a reference on the extent and return it
 820  */
 821 struct btrfs_ordered_extent *btrfs_lookup_ordered_extent(struct btrfs_inode *inode,
 822                                                          u64 file_offset)
 823 {
 824         struct btrfs_ordered_inode_tree *tree;
 825         struct rb_node *node;
 826         struct btrfs_ordered_extent *entry = NULL;
 827         unsigned long flags;
 828
 829         tree = &inode->ordered_tree;
 830         spin_lock_irqsave(&tree->lock, flags);
 831         node = tree_search(tree, file_offset);
 832         if (!node)
 833                 goto out;
 834
 835         entry = rb_entry(node, struct btrfs_ordered_extent, rb_node);
 836         if (!in_range(file_offset, entry->file_offset, entry->num_bytes))
 837                 entry = NULL;
 838         if (entry) {
 839                 refcount_inc(&entry->refs);
 840                 trace_btrfs_ordered_extent_lookup(inode, entry);
 841         }
 842 out:
 843         spin_unlock_irqrestore(&tree->lock, flags);
 844         return entry;
 845 }
 846
 847 /* Since the DIO code tries to lock a wide area we need to look for any ordered
 848  * extents that exist in the range, rather than just the start of the range.
 849  */
 850 struct btrfs_ordered_extent *btrfs_lookup_ordered_range(
 851                 struct btrfs_inode *inode, u64 file_offset, u64 len)
 852 {
 853         struct btrfs_ordered_inode_tree *tree;
 854         struct rb_node *node;
 855         struct btrfs_ordered_extent *entry = NULL;
 856
 857         tree = &inode->ordered_tree;
 858         spin_lock_irq(&tree->lock);
 859         node = tree_search(tree, file_offset);
 860         if (!node) {
 861                 node = tree_search(tree, file_offset + len);
 862                 if (!node)
 863                         goto out;
 864         }
 865
 866         while (1) {
 867                 entry = rb_entry(node, struct btrfs_ordered_extent, rb_node);
 868                 if (range_overlaps(entry, file_offset, len))
 869                         break;
 870
 871                 if (entry->file_offset >= file_offset + len) {
 872                         entry = NULL;
 873                         break;
 874                 }
 875                 entry = NULL;
 876                 node = rb_next(node);
 877                 if (!node)
 878                         break;
 879         }
 880 out:
 881         if (entry) {
 882                 refcount_inc(&entry->refs);
 883                 trace_btrfs_ordered_extent_lookup_range(inode, entry);
 884         }
 885         spin_unlock_irq(&tree->lock);
 886         return entry;
 887 }
 888
 889 /*
 890  * Adds all ordered extents to the given list. The list ends up sorted by the
 891  * file_offset of the ordered extents.
 892  */
 893 void btrfs_get_ordered_extents_for_logging(struct btrfs_inode *inode,
 894                                            struct list_head *list)
 895 {
 896         struct btrfs_ordered_inode_tree *tree = &inode->ordered_tree;
 897         struct rb_node *n;
 898
 899         ASSERT(inode_is_locked(&inode->vfs_inode));
 900
 901         spin_lock_irq(&tree->lock);
 902         for (n = rb_first(&tree->tree); n; n = rb_next(n)) {
 903                 struct btrfs_ordered_extent *ordered;
 904
 905                 ordered = rb_entry(n, struct btrfs_ordered_extent, rb_node);
 906
 907                 if (test_bit(BTRFS_ORDERED_LOGGED, &ordered->flags))
 908                         continue;
 909
 910                 ASSERT(list_empty(&ordered->log_list));
 911                 list_add_tail(&ordered->log_list, list);
 912                 refcount_inc(&ordered->refs);
 913                 trace_btrfs_ordered_extent_lookup_for_logging(inode, ordered);
 914         }
 915         spin_unlock_irq(&tree->lock);
 916 }
 917
 918 /*
 919  * lookup and return any extent before 'file_offset'.  NULL is returned
 920  * if none is found
 921  */
 922 struct btrfs_ordered_extent *
 923 btrfs_lookup_first_ordered_extent(struct btrfs_inode *inode, u64 file_offset)
 924 {
 925         struct btrfs_ordered_inode_tree *tree;
 926         struct rb_node *node;
 927         struct btrfs_ordered_extent *entry = NULL;
 928
 929         tree = &inode->ordered_tree;
 930         spin_lock_irq(&tree->lock);
 931         node = tree_search(tree, file_offset);
 932         if (!node)
 933                 goto out;
 934
 935         entry = rb_entry(node, struct btrfs_ordered_extent, rb_node);
 936         refcount_inc(&entry->refs);
 937         trace_btrfs_ordered_extent_lookup_first(inode, entry);
 938 out:
 939         spin_unlock_irq(&tree->lock);
 940         return entry;
 941 }
 942
 943 /*
 944  * Lookup the first ordered extent that overlaps the range
 945  * [@file_offset, @file_offset + @len).
 946  *
 947  * The difference between this and btrfs_lookup_first_ordered_extent() is
 948  * that this one won't return any ordered extent that does not overlap the range.
 949  * And the difference against btrfs_lookup_ordered_extent() is, this function
 950  * ensures the first ordered extent gets returned.
 951  */
 952 struct btrfs_ordered_extent *btrfs_lookup_first_ordered_range(
 953                         struct btrfs_inode *inode, u64 file_offset, u64 len)
 954 {
 955         struct btrfs_ordered_inode_tree *tree = &inode->ordered_tree;
 956         struct rb_node *node;
 957         struct rb_node *cur;
 958         struct rb_node *prev;
 959         struct rb_node *next;
 960         struct btrfs_ordered_extent *entry = NULL;
 961
 962         spin_lock_irq(&tree->lock);
 963         node = tree->tree.rb_node;
 964         /*
 965          * Here we don't want to use tree_search() which will use tree->last
 966          * and screw up the search order.
 967          * And __tree_search() can't return the adjacent ordered extents
 968          * either, thus here we do our own search.
 969          */
 970         while (node) {
 971                 entry = rb_entry(node, struct btrfs_ordered_extent, rb_node);
 972
 973                 if (file_offset < entry->file_offset) {
 974                         node = node->rb_left;
 975                 } else if (file_offset >= entry_end(entry)) {
 976                         node = node->rb_right;
 977                 } else {
 978                         /*
 979                          * Direct hit, got an ordered extent that starts at
 980                          * @file_offset
 981                          */
 982                         goto out;
 983                 }
 984         }
 985         if (!entry) {
 986                 /* Empty tree */
 987                 goto out;
 988         }
 989
 990         cur = &entry->rb_node;
 991         /* We got an entry around @file_offset, check adjacent entries */
 992         if (entry->file_offset < file_offset) {
 993                 prev = cur;
 994                 next = rb_next(cur);
 995         } else {
 996                 prev = rb_prev(cur);
 997                 next = cur;
 998         }
 999         if (prev) {
1000                 entry = rb_entry(prev, struct btrfs_ordered_extent, rb_node);
1001                 if (range_overlaps(entry, file_offset, len))
1002                         goto out;
1003         }
1004         if (next) {
1005                 entry = rb_entry(next, struct btrfs_ordered_extent, rb_node);
1006                 if (range_overlaps(entry, file_offset, len))
1007                         goto out;
1008         }
1009         /* No ordered extent in the range */
1010         entry = NULL;
1011 out:
1012         if (entry) {
1013                 refcount_inc(&entry->refs);
1014                 trace_btrfs_ordered_extent_lookup_first_range(inode, entry);
1015         }
1016
1017         spin_unlock_irq(&tree->lock);
1018         return entry;
1019 }
1020
1021 /*
1022  * Lock the passed range and ensures all pending ordered extents in it are run
1023  * to completion.
1024  *
1025  * @inode:        Inode whose ordered tree is to be searched
1026  * @start:        Beginning of range to flush
1027  * @end:          Last byte of range to lock
1028  * @cached_state: If passed, will return the extent state responsible for the
1029  *                locked range. It's the caller's responsibility to free the
1030  *                cached state.
1031  *
1032  * Always return with the given range locked, ensuring after it's called no
1033  * order extent can be pending.
1034  */
1035 void btrfs_lock_and_flush_ordered_range(struct btrfs_inode *inode, u64 start,
1036                                         u64 end,
1037                                         struct extent_state **cached_state)
1038 {
1039         struct btrfs_ordered_extent *ordered;
1040         struct extent_state *cache = NULL;
1041         struct extent_state **cachedp = &cache;
1042
1043         if (cached_state)
1044                 cachedp = cached_state;
1045
1046         while (1) {
1047                 lock_extent(&inode->io_tree, start, end, cachedp);
1048                 ordered = btrfs_lookup_ordered_range(inode, start,
1049                                                      end - start + 1);
1050                 if (!ordered) {
1051                         /*
1052                          * If no external cached_state has been passed then
1053                          * decrement the extra ref taken for cachedp since we
1054                          * aren't exposing it outside of this function
1055                          */
1056                         if (!cached_state)
1057                                 refcount_dec(&cache->refs);
1058                         break;
1059                 }
1060                 unlock_extent(&inode->io_tree, start, end, cachedp);
1061                 btrfs_start_ordered_extent(ordered);
1062                 btrfs_put_ordered_extent(ordered);
1063         }
1064 }
1065
1066 /*
1067  * Lock the passed range and ensure all pending ordered extents in it are run
1068  * to completion in nowait mode.
1069  *
1070  * Return true if btrfs_lock_ordered_range does not return any extents,
1071  * otherwise false.
1072  */
1073 bool btrfs_try_lock_ordered_range(struct btrfs_inode *inode, u64 start, u64 end,
1074                                   struct extent_state **cached_state)
1075 {
1076         struct btrfs_ordered_extent *ordered;
1077
1078         if (!try_lock_extent(&inode->io_tree, start, end, cached_state))
1079                 return false;
1080
1081         ordered = btrfs_lookup_ordered_range(inode, start, end - start + 1);
1082         if (!ordered)
1083                 return true;
1084
1085         btrfs_put_ordered_extent(ordered);
1086         unlock_extent(&inode->io_tree, start, end, cached_state);
1087
1088         return false;
1089 }
1090
1091
1092 static int clone_ordered_extent(struct btrfs_ordered_extent *ordered, u64 pos,
1093                                 u64 len)
1094 {
1095         struct inode *inode = ordered->inode;
1096         struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
1097         u64 file_offset = ordered->file_offset + pos;
1098         u64 disk_bytenr = ordered->disk_bytenr + pos;
1099         unsigned long flags = ordered->flags & BTRFS_ORDERED_TYPE_FLAGS;
1100
1101         /*
1102          * The splitting extent is already counted and will be added again in
1103          * btrfs_add_ordered_extent_*(). Subtract len to avoid double counting.
1104          */
1105         percpu_counter_add_batch(&fs_info->ordered_bytes, -len,
1106                                  fs_info->delalloc_batch);
1107         WARN_ON_ONCE(flags & (1 << BTRFS_ORDERED_COMPRESSED));
1108         return btrfs_add_ordered_extent(BTRFS_I(inode), file_offset, len, len,
1109                                         disk_bytenr, len, 0, flags,
1110                                         ordered->compress_type);
1111 }
1112
1113 int btrfs_split_ordered_extent(struct btrfs_ordered_extent *ordered, u64 pre,
1114                                 u64 post)
1115 {
1116         struct inode *inode = ordered->inode;
1117         struct btrfs_ordered_inode_tree *tree = &BTRFS_I(inode)->ordered_tree;
1118         struct rb_node *node;
1119         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1120         int ret = 0;
1121
1122         trace_btrfs_ordered_extent_split(BTRFS_I(inode), ordered);
1123
1124         spin_lock_irq(&tree->lock);
1125         /* Remove from tree once */
1126         node = &ordered->rb_node;
1127         rb_erase(node, &tree->tree);
1128         RB_CLEAR_NODE(node);
1129         if (tree->last == node)
1130                 tree->last = NULL;
1131
1132         ordered->file_offset += pre;
1133         ordered->disk_bytenr += pre;
1134         ordered->num_bytes -= (pre + post);
1135         ordered->disk_num_bytes -= (pre + post);
1136         ordered->bytes_left -= (pre + post);
1137
1138         /* Re-insert the node */
1139         node = tree_insert(&tree->tree, ordered->file_offset, &ordered->rb_node);
1140         if (node)
1141                 btrfs_panic(fs_info, -EEXIST,
1142                         "zoned: inconsistency in ordered tree at offset %llu",
1143                             ordered->file_offset);
1144
1145         spin_unlock_irq(&tree->lock);
1146
1147         if (pre)
1148                 ret = clone_ordered_extent(ordered, 0, pre);
1149         if (ret == 0 && post)
1150                 ret = clone_ordered_extent(ordered, pre + ordered->disk_num_bytes,
1151                                            post);
1152
1153         return ret;
1154 }
1155
1156 int __init ordered_data_init(void)
1157 {
1158         btrfs_ordered_extent_cache = kmem_cache_create("btrfs_ordered_extent",
1159                                      sizeof(struct btrfs_ordered_extent), 0,
1160                                      SLAB_MEM_SPREAD,
1161                                      NULL);
1162         if (!btrfs_ordered_extent_cache)
1163                 return -ENOMEM;
1164
1165         return 0;
1166 }
1167
1168 void __cold ordered_data_exit(void)
1169 {
1170         kmem_cache_destroy(btrfs_ordered_extent_cache);
1171 }