fs/btrfs/space-info.c

   1 // SPDX-License-Identifier: GPL-2.0
   2
   3 #include "misc.h"
   4 #include "ctree.h"
   5 #include "space-info.h"
   6 #include "sysfs.h"
   7 #include "volumes.h"
   8 #include "free-space-cache.h"
   9 #include "ordered-data.h"
  10 #include "transaction.h"
  11 #include "block-group.h"
  12
  13 /*
  14  * HOW DOES SPACE RESERVATION WORK
  15  *
  16  * If you want to know about delalloc specifically, there is a separate comment
  17  * for that with the delalloc code.  This comment is about how the whole system
  18  * works generally.
  19  *
  20  * BASIC CONCEPTS
  21  *
  22  *   1) space_info.  This is the ultimate arbiter of how much space we can use.
  23  *   There's a description of the bytes_ fields with the struct declaration,
  24  *   refer to that for specifics on each field.  Suffice it to say that for
  25  *   reservations we care about total_bytes - SUM(space_info->bytes_) when
  26  *   determining if there is space to make an allocation.  There is a space_info
  27  *   for METADATA, SYSTEM, and DATA areas.
  28  *
  29  *   2) block_rsv's.  These are basically buckets for every different type of
  30  *   metadata reservation we have.  You can see the comment in the block_rsv
  31  *   code on the rules for each type, but generally block_rsv->reserved is how
  32  *   much space is accounted for in space_info->bytes_may_use.
  33  *
  34  *   3) btrfs_calc*_size.  These are the worst case calculations we used based
  35  *   on the number of items we will want to modify.  We have one for changing
  36  *   items, and one for inserting new items.  Generally we use these helpers to
  37  *   determine the size of the block reserves, and then use the actual bytes
  38  *   values to adjust the space_info counters.
  39  *
  40  * MAKING RESERVATIONS, THE NORMAL CASE
  41  *
  42  *   We call into either btrfs_reserve_data_bytes() or
  43  *   btrfs_reserve_metadata_bytes(), depending on which we're looking for, with
  44  *   num_bytes we want to reserve.
  45  *
  46  *   ->reserve
  47  *     space_info->bytes_may_reserve += num_bytes
  48  *
  49  *   ->extent allocation
  50  *     Call btrfs_add_reserved_bytes() which does
  51  *     space_info->bytes_may_reserve -= num_bytes
  52  *     space_info->bytes_reserved += extent_bytes
  53  *
  54  *   ->insert reference
  55  *     Call btrfs_update_block_group() which does
  56  *     space_info->bytes_reserved -= extent_bytes
  57  *     space_info->bytes_used += extent_bytes
  58  *
  59  * MAKING RESERVATIONS, FLUSHING NORMALLY (non-priority)
  60  *
  61  *   Assume we are unable to simply make the reservation because we do not have
  62  *   enough space
  63  *
  64  *   -> __reserve_bytes
  65  *     create a reserve_ticket with ->bytes set to our reservation, add it to
  66  *     the tail of space_info->tickets, kick async flush thread
  67  *
  68  *   ->handle_reserve_ticket
  69  *     wait on ticket->wait for ->bytes to be reduced to 0, or ->error to be set
  70  *     on the ticket.
  71  *
  72  *   -> btrfs_async_reclaim_metadata_space/btrfs_async_reclaim_data_space
  73  *     Flushes various things attempting to free up space.
  74  *
  75  *   -> btrfs_try_granting_tickets()
  76  *     This is called by anything that either subtracts space from
  77  *     space_info->bytes_may_use, ->bytes_pinned, etc, or adds to the
  78  *     space_info->total_bytes.  This loops through the ->priority_tickets and
  79  *     then the ->tickets list checking to see if the reservation can be
  80  *     completed.  If it can the space is added to space_info->bytes_may_use and
  81  *     the ticket is woken up.
  82  *
  83  *   -> ticket wakeup
  84  *     Check if ->bytes == 0, if it does we got our reservation and we can carry
  85  *     on, if not return the appropriate error (ENOSPC, but can be EINTR if we
  86  *     were interrupted.)
  87  *
  88  * MAKING RESERVATIONS, FLUSHING HIGH PRIORITY
  89  *
  90  *   Same as the above, except we add ourselves to the
  91  *   space_info->priority_tickets, and we do not use ticket->wait, we simply
  92  *   call flush_space() ourselves for the states that are safe for us to call
  93  *   without deadlocking and hope for the best.
  94  *
  95  * THE FLUSHING STATES
  96  *
  97  *   Generally speaking we will have two cases for each state, a "nice" state
  98  *   and a "ALL THE THINGS" state.  In btrfs we delay a lot of work in order to
  99  *   reduce the locking over head on the various trees, and even to keep from
 100  *   doing any work at all in the case of delayed refs.  Each of these delayed
 101  *   things however hold reservations, and so letting them run allows us to
 102  *   reclaim space so we can make new reservations.
 103  *
 104  *   FLUSH_DELAYED_ITEMS
 105  *     Every inode has a delayed item to update the inode.  Take a simple write
 106  *     for example, we would update the inode item at write time to update the
 107  *     mtime, and then again at finish_ordered_io() time in order to update the
 108  *     isize or bytes.  We keep these delayed items to coalesce these operations
 109  *     into a single operation done on demand.  These are an easy way to reclaim
 110  *     metadata space.
 111  *
 112  *   FLUSH_DELALLOC
 113  *     Look at the delalloc comment to get an idea of how much space is reserved
 114  *     for delayed allocation.  We can reclaim some of this space simply by
 115  *     running delalloc, but usually we need to wait for ordered extents to
 116  *     reclaim the bulk of this space.
 117  *
 118  *   FLUSH_DELAYED_REFS
 119  *     We have a block reserve for the outstanding delayed refs space, and every
 120  *     delayed ref operation holds a reservation.  Running these is a quick way
 121  *     to reclaim space, but we want to hold this until the end because COW can
 122  *     churn a lot and we can avoid making some extent tree modifications if we
 123  *     are able to delay for as long as possible.
 124  *
 125  *   ALLOC_CHUNK
 126  *     We will skip this the first time through space reservation, because of
 127  *     overcommit and we don't want to have a lot of useless metadata space when
 128  *     our worst case reservations will likely never come true.
 129  *
 130  *   RUN_DELAYED_IPUTS
 131  *     If we're freeing inodes we're likely freeing checksums, file extent
 132  *     items, and extent tree items.  Loads of space could be freed up by these
 133  *     operations, however they won't be usable until the transaction commits.
 134  *
 135  *   COMMIT_TRANS
 136  *     may_commit_transaction() is the ultimate arbiter on whether we commit the
 137  *     transaction or not.  In order to avoid constantly churning we do all the
 138  *     above flushing first and then commit the transaction as the last resort.
 139  *     However we need to take into account things like pinned space that would
 140  *     be freed, plus any delayed work we may not have gotten rid of in the case
 141  *     of metadata.
 142  *
 143  * OVERCOMMIT
 144  *
 145  *   Because we hold so many reservations for metadata we will allow you to
 146  *   reserve more space than is currently free in the currently allocate
 147  *   metadata space.  This only happens with metadata, data does not allow
 148  *   overcommitting.
 149  *
 150  *   You can see the current logic for when we allow overcommit in
 151  *   btrfs_can_overcommit(), but it only applies to unallocated space.  If there
 152  *   is no unallocated space to be had, all reservations are kept within the
 153  *   free space in the allocated metadata chunks.
 154  *
 155  *   Because of overcommitting, you generally want to use the
 156  *   btrfs_can_overcommit() logic for metadata allocations, as it does the right
 157  *   thing with or without extra unallocated space.
 158  */
 159
 160 u64 __pure btrfs_space_info_used(struct btrfs_space_info *s_info,
 161                           bool may_use_included)
 162 {
 163         ASSERT(s_info);
 164         return s_info->bytes_used + s_info->bytes_reserved +
 165                 s_info->bytes_pinned + s_info->bytes_readonly +
 166                 (may_use_included ? s_info->bytes_may_use : 0);
 167 }
 168
 169 /*
 170  * after adding space to the filesystem, we need to clear the full flags
 171  * on all the space infos.
 172  */
 173 void btrfs_clear_space_info_full(struct btrfs_fs_info *info)
 174 {
 175         struct list_head *head = &info->space_info;
 176         struct btrfs_space_info *found;
 177
 178         list_for_each_entry(found, head, list)
 179                 found->full = 0;
 180 }
 181
 182 static int create_space_info(struct btrfs_fs_info *info, u64 flags)
 183 {
 184
 185         struct btrfs_space_info *space_info;
 186         int i;
 187         int ret;
 188
 189         space_info = kzalloc(sizeof(*space_info), GFP_NOFS);
 190         if (!space_info)
 191                 return -ENOMEM;
 192
 193         ret = percpu_counter_init(&space_info->total_bytes_pinned, 0,
 194                                  GFP_KERNEL);
 195         if (ret) {
 196                 kfree(space_info);
 197                 return ret;
 198         }
 199
 200         for (i = 0; i < BTRFS_NR_RAID_TYPES; i++)
 201                 INIT_LIST_HEAD(&space_info->block_groups[i]);
 202         init_rwsem(&space_info->groups_sem);
 203         spin_lock_init(&space_info->lock);
 204         space_info->flags = flags & BTRFS_BLOCK_GROUP_TYPE_MASK;
 205         space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
 206         INIT_LIST_HEAD(&space_info->ro_bgs);
 207         INIT_LIST_HEAD(&space_info->tickets);
 208         INIT_LIST_HEAD(&space_info->priority_tickets);
 209
 210         ret = btrfs_sysfs_add_space_info_type(info, space_info);
 211         if (ret)
 212                 return ret;
 213
 214         list_add(&space_info->list, &info->space_info);
 215         if (flags & BTRFS_BLOCK_GROUP_DATA)
 216                 info->data_sinfo = space_info;
 217
 218         return ret;
 219 }
 220
 221 int btrfs_init_space_info(struct btrfs_fs_info *fs_info)
 222 {
 223         struct btrfs_super_block *disk_super;
 224         u64 features;
 225         u64 flags;
 226         int mixed = 0;
 227         int ret;
 228
 229         disk_super = fs_info->super_copy;
 230         if (!btrfs_super_root(disk_super))
 231                 return -EINVAL;
 232
 233         features = btrfs_super_incompat_flags(disk_super);
 234         if (features & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS)
 235                 mixed = 1;
 236
 237         flags = BTRFS_BLOCK_GROUP_SYSTEM;
 238         ret = create_space_info(fs_info, flags);
 239         if (ret)
 240                 goto out;
 241
 242         if (mixed) {
 243                 flags = BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DATA;
 244                 ret = create_space_info(fs_info, flags);
 245         } else {
 246                 flags = BTRFS_BLOCK_GROUP_METADATA;
 247                 ret = create_space_info(fs_info, flags);
 248                 if (ret)
 249                         goto out;
 250
 251                 flags = BTRFS_BLOCK_GROUP_DATA;
 252                 ret = create_space_info(fs_info, flags);
 253         }
 254 out:
 255         return ret;
 256 }
 257
 258 void btrfs_update_space_info(struct btrfs_fs_info *info, u64 flags,
 259                              u64 total_bytes, u64 bytes_used,
 260                              u64 bytes_readonly,
 261                              struct btrfs_space_info **space_info)
 262 {
 263         struct btrfs_space_info *found;
 264         int factor;
 265
 266         factor = btrfs_bg_type_to_factor(flags);
 267
 268         found = btrfs_find_space_info(info, flags);
 269         ASSERT(found);
 270         spin_lock(&found->lock);
 271         found->total_bytes += total_bytes;
 272         found->disk_total += total_bytes * factor;
 273         found->bytes_used += bytes_used;
 274         found->disk_used += bytes_used * factor;
 275         found->bytes_readonly += bytes_readonly;
 276         if (total_bytes > 0)
 277                 found->full = 0;
 278         btrfs_try_granting_tickets(info, found);
 279         spin_unlock(&found->lock);
 280         *space_info = found;
 281 }
 282
 283 struct btrfs_space_info *btrfs_find_space_info(struct btrfs_fs_info *info,
 284                                                u64 flags)
 285 {
 286         struct list_head *head = &info->space_info;
 287         struct btrfs_space_info *found;
 288
 289         flags &= BTRFS_BLOCK_GROUP_TYPE_MASK;
 290
 291         list_for_each_entry(found, head, list) {
 292                 if (found->flags & flags)
 293                         return found;
 294         }
 295         return NULL;
 296 }
 297
 298 static u64 calc_available_free_space(struct btrfs_fs_info *fs_info,
 299                           struct btrfs_space_info *space_info,
 300                           enum btrfs_reserve_flush_enum flush)
 301 {
 302         u64 profile;
 303         u64 avail;
 304         int factor;
 305
 306         if (space_info->flags & BTRFS_BLOCK_GROUP_SYSTEM)
 307                 profile = btrfs_system_alloc_profile(fs_info);
 308         else
 309                 profile = btrfs_metadata_alloc_profile(fs_info);
 310
 311         avail = atomic64_read(&fs_info->free_chunk_space);
 312
 313         /*
 314          * If we have dup, raid1 or raid10 then only half of the free
 315          * space is actually usable.  For raid56, the space info used
 316          * doesn't include the parity drive, so we don't have to
 317          * change the math
 318          */
 319         factor = btrfs_bg_type_to_factor(profile);
 320         avail = div_u64(avail, factor);
 321
 322         /*
 323          * If we aren't flushing all things, let us overcommit up to
 324          * 1/2th of the space. If we can flush, don't let us overcommit
 325          * too much, let it overcommit up to 1/8 of the space.
 326          */
 327         if (flush == BTRFS_RESERVE_FLUSH_ALL)
 328                 avail >>= 3;
 329         else
 330                 avail >>= 1;
 331         return avail;
 332 }
 333
 334 int btrfs_can_overcommit(struct btrfs_fs_info *fs_info,
 335                          struct btrfs_space_info *space_info, u64 bytes,
 336                          enum btrfs_reserve_flush_enum flush)
 337 {
 338         u64 avail;
 339         u64 used;
 340
 341         /* Don't overcommit when in mixed mode */
 342         if (space_info->flags & BTRFS_BLOCK_GROUP_DATA)
 343                 return 0;
 344
 345         used = btrfs_space_info_used(space_info, true);
 346         avail = calc_available_free_space(fs_info, space_info, flush);
 347
 348         if (used + bytes < space_info->total_bytes + avail)
 349                 return 1;
 350         return 0;
 351 }
 352
 353 static void remove_ticket(struct btrfs_space_info *space_info,
 354                           struct reserve_ticket *ticket)
 355 {
 356         if (!list_empty(&ticket->list)) {
 357                 list_del_init(&ticket->list);
 358                 ASSERT(space_info->reclaim_size >= ticket->bytes);
 359                 space_info->reclaim_size -= ticket->bytes;
 360         }
 361 }
 362
 363 /*
 364  * This is for space we already have accounted in space_info->bytes_may_use, so
 365  * basically when we're returning space from block_rsv's.
 366  */
 367 void btrfs_try_granting_tickets(struct btrfs_fs_info *fs_info,
 368                                 struct btrfs_space_info *space_info)
 369 {
 370         struct list_head *head;
 371         enum btrfs_reserve_flush_enum flush = BTRFS_RESERVE_NO_FLUSH;
 372
 373         lockdep_assert_held(&space_info->lock);
 374
 375         head = &space_info->priority_tickets;
 376 again:
 377         while (!list_empty(head)) {
 378                 struct reserve_ticket *ticket;
 379                 u64 used = btrfs_space_info_used(space_info, true);
 380
 381                 ticket = list_first_entry(head, struct reserve_ticket, list);
 382
 383                 /* Check and see if our ticket can be satisified now. */
 384                 if ((used + ticket->bytes <= space_info->total_bytes) ||
 385                     btrfs_can_overcommit(fs_info, space_info, ticket->bytes,
 386                                          flush)) {
 387                         btrfs_space_info_update_bytes_may_use(fs_info,
 388                                                               space_info,
 389                                                               ticket->bytes);
 390                         remove_ticket(space_info, ticket);
 391                         ticket->bytes = 0;
 392                         space_info->tickets_id++;
 393                         wake_up(&ticket->wait);
 394                 } else {
 395                         break;
 396                 }
 397         }
 398
 399         if (head == &space_info->priority_tickets) {
 400                 head = &space_info->tickets;
 401                 flush = BTRFS_RESERVE_FLUSH_ALL;
 402                 goto again;
 403         }
 404 }
 405
 406 #define DUMP_BLOCK_RSV(fs_info, rsv_name)                               \
 407 do {                                                                    \
 408         struct btrfs_block_rsv *__rsv = &(fs_info)->rsv_name;           \
 409         spin_lock(&__rsv->lock);                                        \
 410         btrfs_info(fs_info, #rsv_name ": size %llu reserved %llu",      \
 411                    __rsv->size, __rsv->reserved);                       \
 412         spin_unlock(&__rsv->lock);                                      \
 413 } while (0)
 414
 415 static void __btrfs_dump_space_info(struct btrfs_fs_info *fs_info,
 416                                     struct btrfs_space_info *info)
 417 {
 418         lockdep_assert_held(&info->lock);
 419
 420         /* The free space could be negative in case of overcommit */
 421         btrfs_info(fs_info, "space_info %llu has %lld free, is %sfull",
 422                    info->flags,
 423                    (s64)(info->total_bytes - btrfs_space_info_used(info, true)),
 424                    info->full ? "" : "not ");
 425         btrfs_info(fs_info,
 426                 "space_info total=%llu, used=%llu, pinned=%llu, reserved=%llu, may_use=%llu, readonly=%llu",
 427                 info->total_bytes, info->bytes_used, info->bytes_pinned,
 428                 info->bytes_reserved, info->bytes_may_use,
 429                 info->bytes_readonly);
 430
 431         DUMP_BLOCK_RSV(fs_info, global_block_rsv);
 432         DUMP_BLOCK_RSV(fs_info, trans_block_rsv);
 433         DUMP_BLOCK_RSV(fs_info, chunk_block_rsv);
 434         DUMP_BLOCK_RSV(fs_info, delayed_block_rsv);
 435         DUMP_BLOCK_RSV(fs_info, delayed_refs_rsv);
 436
 437 }
 438
 439 void btrfs_dump_space_info(struct btrfs_fs_info *fs_info,
 440                            struct btrfs_space_info *info, u64 bytes,
 441                            int dump_block_groups)
 442 {
 443         struct btrfs_block_group *cache;
 444         int index = 0;
 445
 446         spin_lock(&info->lock);
 447         __btrfs_dump_space_info(fs_info, info);
 448         spin_unlock(&info->lock);
 449
 450         if (!dump_block_groups)
 451                 return;
 452
 453         down_read(&info->groups_sem);
 454 again:
 455         list_for_each_entry(cache, &info->block_groups[index], list) {
 456                 spin_lock(&cache->lock);
 457                 btrfs_info(fs_info,
 458                         "block group %llu has %llu bytes, %llu used %llu pinned %llu reserved %s",
 459                         cache->start, cache->length, cache->used, cache->pinned,
 460                         cache->reserved, cache->ro ? "[readonly]" : "");
 461                 spin_unlock(&cache->lock);
 462                 btrfs_dump_free_space(cache, bytes);
 463         }
 464         if (++index < BTRFS_NR_RAID_TYPES)
 465                 goto again;
 466         up_read(&info->groups_sem);
 467 }
 468
 469 static inline u64 calc_reclaim_items_nr(struct btrfs_fs_info *fs_info,
 470                                         u64 to_reclaim)
 471 {
 472         u64 bytes;
 473         u64 nr;
 474
 475         bytes = btrfs_calc_insert_metadata_size(fs_info, 1);
 476         nr = div64_u64(to_reclaim, bytes);
 477         if (!nr)
 478                 nr = 1;
 479         return nr;
 480 }
 481
 482 #define EXTENT_SIZE_PER_ITEM    SZ_256K
 483
 484 /*
 485  * shrink metadata reservation for delalloc
 486  */
 487 static void shrink_delalloc(struct btrfs_fs_info *fs_info,
 488                             struct btrfs_space_info *space_info,
 489                             u64 to_reclaim, bool wait_ordered)
 490 {
 491         struct btrfs_trans_handle *trans;
 492         u64 delalloc_bytes;
 493         u64 dio_bytes;
 494         u64 items;
 495         long time_left;
 496         int loops;
 497
 498         /* Calc the number of the pages we need flush for space reservation */
 499         if (to_reclaim == U64_MAX) {
 500                 items = U64_MAX;
 501         } else {
 502                 /*
 503                  * to_reclaim is set to however much metadata we need to
 504                  * reclaim, but reclaiming that much data doesn't really track
 505                  * exactly, so increase the amount to reclaim by 2x in order to
 506                  * make sure we're flushing enough delalloc to hopefully reclaim
 507                  * some metadata reservations.
 508                  */
 509                 items = calc_reclaim_items_nr(fs_info, to_reclaim) * 2;
 510                 to_reclaim = items * EXTENT_SIZE_PER_ITEM;
 511         }
 512
 513         trans = (struct btrfs_trans_handle *)current->journal_info;
 514
 515         delalloc_bytes = percpu_counter_sum_positive(
 516                                                 &fs_info->delalloc_bytes);
 517         dio_bytes = percpu_counter_sum_positive(&fs_info->dio_bytes);
 518         if (delalloc_bytes == 0 && dio_bytes == 0) {
 519                 if (trans)
 520                         return;
 521                 if (wait_ordered)
 522                         btrfs_wait_ordered_roots(fs_info, items, 0, (u64)-1);
 523                 return;
 524         }
 525
 526         /*
 527          * If we are doing more ordered than delalloc we need to just wait on
 528          * ordered extents, otherwise we'll waste time trying to flush delalloc
 529          * that likely won't give us the space back we need.
 530          */
 531         if (dio_bytes > delalloc_bytes)
 532                 wait_ordered = true;
 533
 534         loops = 0;
 535         while ((delalloc_bytes || dio_bytes) && loops < 3) {
 536                 u64 nr_pages = min(delalloc_bytes, to_reclaim) >> PAGE_SHIFT;
 537
 538                 btrfs_start_delalloc_roots(fs_info, nr_pages, true);
 539
 540                 loops++;
 541                 if (wait_ordered && !trans) {
 542                         btrfs_wait_ordered_roots(fs_info, items, 0, (u64)-1);
 543                 } else {
 544                         time_left = schedule_timeout_killable(1);
 545                         if (time_left)
 546                                 break;
 547                 }
 548
 549                 spin_lock(&space_info->lock);
 550                 if (list_empty(&space_info->tickets) &&
 551                     list_empty(&space_info->priority_tickets)) {
 552                         spin_unlock(&space_info->lock);
 553                         break;
 554                 }
 555                 spin_unlock(&space_info->lock);
 556
 557                 delalloc_bytes = percpu_counter_sum_positive(
 558                                                 &fs_info->delalloc_bytes);
 559                 dio_bytes = percpu_counter_sum_positive(&fs_info->dio_bytes);
 560         }
 561 }
 562
 563 /**
 564  * maybe_commit_transaction - possibly commit the transaction if its ok to
 565  * @root - the root we're allocating for
 566  * @bytes - the number of bytes we want to reserve
 567  * @force - force the commit
 568  *
 569  * This will check to make sure that committing the transaction will actually
 570  * get us somewhere and then commit the transaction if it does.  Otherwise it
 571  * will return -ENOSPC.
 572  */
 573 static int may_commit_transaction(struct btrfs_fs_info *fs_info,
 574                                   struct btrfs_space_info *space_info)
 575 {
 576         struct reserve_ticket *ticket = NULL;
 577         struct btrfs_block_rsv *delayed_rsv = &fs_info->delayed_block_rsv;
 578         struct btrfs_block_rsv *delayed_refs_rsv = &fs_info->delayed_refs_rsv;
 579         struct btrfs_block_rsv *trans_rsv = &fs_info->trans_block_rsv;
 580         struct btrfs_trans_handle *trans;
 581         u64 reclaim_bytes = 0;
 582         u64 bytes_needed = 0;
 583         u64 cur_free_bytes = 0;
 584
 585         trans = (struct btrfs_trans_handle *)current->journal_info;
 586         if (trans)
 587                 return -EAGAIN;
 588
 589         spin_lock(&space_info->lock);
 590         cur_free_bytes = btrfs_space_info_used(space_info, true);
 591         if (cur_free_bytes < space_info->total_bytes)
 592                 cur_free_bytes = space_info->total_bytes - cur_free_bytes;
 593         else
 594                 cur_free_bytes = 0;
 595
 596         if (!list_empty(&space_info->priority_tickets))
 597                 ticket = list_first_entry(&space_info->priority_tickets,
 598                                           struct reserve_ticket, list);
 599         else if (!list_empty(&space_info->tickets))
 600                 ticket = list_first_entry(&space_info->tickets,
 601                                           struct reserve_ticket, list);
 602         if (ticket)
 603                 bytes_needed = ticket->bytes;
 604
 605         if (bytes_needed > cur_free_bytes)
 606                 bytes_needed -= cur_free_bytes;
 607         else
 608                 bytes_needed = 0;
 609         spin_unlock(&space_info->lock);
 610
 611         if (!bytes_needed)
 612                 return 0;
 613
 614         trans = btrfs_join_transaction(fs_info->extent_root);
 615         if (IS_ERR(trans))
 616                 return PTR_ERR(trans);
 617
 618         /*
 619          * See if there is enough pinned space to make this reservation, or if
 620          * we have block groups that are going to be freed, allowing us to
 621          * possibly do a chunk allocation the next loop through.
 622          */
 623         if (test_bit(BTRFS_TRANS_HAVE_FREE_BGS, &trans->transaction->flags) ||
 624             __percpu_counter_compare(&space_info->total_bytes_pinned,
 625                                      bytes_needed,
 626                                      BTRFS_TOTAL_BYTES_PINNED_BATCH) >= 0)
 627                 goto commit;
 628
 629         /*
 630          * See if there is some space in the delayed insertion reserve for this
 631          * reservation.  If the space_info's don't match (like for DATA or
 632          * SYSTEM) then just go enospc, reclaiming this space won't recover any
 633          * space to satisfy those reservations.
 634          */
 635         if (space_info != delayed_rsv->space_info)
 636                 goto enospc;
 637
 638         spin_lock(&delayed_rsv->lock);
 639         reclaim_bytes += delayed_rsv->reserved;
 640         spin_unlock(&delayed_rsv->lock);
 641
 642         spin_lock(&delayed_refs_rsv->lock);
 643         reclaim_bytes += delayed_refs_rsv->reserved;
 644         spin_unlock(&delayed_refs_rsv->lock);
 645
 646         spin_lock(&trans_rsv->lock);
 647         reclaim_bytes += trans_rsv->reserved;
 648         spin_unlock(&trans_rsv->lock);
 649
 650         if (reclaim_bytes >= bytes_needed)
 651                 goto commit;
 652         bytes_needed -= reclaim_bytes;
 653
 654         if (__percpu_counter_compare(&space_info->total_bytes_pinned,
 655                                    bytes_needed,
 656                                    BTRFS_TOTAL_BYTES_PINNED_BATCH) < 0)
 657                 goto enospc;
 658
 659 commit:
 660         return btrfs_commit_transaction(trans);
 661 enospc:
 662         btrfs_end_transaction(trans);
 663         return -ENOSPC;
 664 }
 665
 666 /*
 667  * Try to flush some data based on policy set by @state. This is only advisory
 668  * and may fail for various reasons. The caller is supposed to examine the
 669  * state of @space_info to detect the outcome.
 670  */
 671 static void flush_space(struct btrfs_fs_info *fs_info,
 672                        struct btrfs_space_info *space_info, u64 num_bytes,
 673                        int state)
 674 {
 675         struct btrfs_root *root = fs_info->extent_root;
 676         struct btrfs_trans_handle *trans;
 677         int nr;
 678         int ret = 0;
 679
 680         switch (state) {
 681         case FLUSH_DELAYED_ITEMS_NR:
 682         case FLUSH_DELAYED_ITEMS:
 683                 if (state == FLUSH_DELAYED_ITEMS_NR)
 684                         nr = calc_reclaim_items_nr(fs_info, num_bytes) * 2;
 685                 else
 686                         nr = -1;
 687
 688                 trans = btrfs_join_transaction(root);
 689                 if (IS_ERR(trans)) {
 690                         ret = PTR_ERR(trans);
 691                         break;
 692                 }
 693                 ret = btrfs_run_delayed_items_nr(trans, nr);
 694                 btrfs_end_transaction(trans);
 695                 break;
 696         case FLUSH_DELALLOC:
 697         case FLUSH_DELALLOC_WAIT:
 698                 shrink_delalloc(fs_info, space_info, num_bytes,
 699                                 state == FLUSH_DELALLOC_WAIT);
 700                 break;
 701         case FLUSH_DELAYED_REFS_NR:
 702         case FLUSH_DELAYED_REFS:
 703                 trans = btrfs_join_transaction(root);
 704                 if (IS_ERR(trans)) {
 705                         ret = PTR_ERR(trans);
 706                         break;
 707                 }
 708                 if (state == FLUSH_DELAYED_REFS_NR)
 709                         nr = calc_reclaim_items_nr(fs_info, num_bytes);
 710                 else
 711                         nr = 0;
 712                 btrfs_run_delayed_refs(trans, nr);
 713                 btrfs_end_transaction(trans);
 714                 break;
 715         case ALLOC_CHUNK:
 716         case ALLOC_CHUNK_FORCE:
 717                 trans = btrfs_join_transaction(root);
 718                 if (IS_ERR(trans)) {
 719                         ret = PTR_ERR(trans);
 720                         break;
 721                 }
 722                 ret = btrfs_chunk_alloc(trans,
 723                                 btrfs_get_alloc_profile(fs_info, space_info->flags),
 724                                 (state == ALLOC_CHUNK) ? CHUNK_ALLOC_NO_FORCE :
 725                                         CHUNK_ALLOC_FORCE);
 726                 btrfs_end_transaction(trans);
 727                 if (ret > 0 || ret == -ENOSPC)
 728                         ret = 0;
 729                 break;
 730         case RUN_DELAYED_IPUTS:
 731                 /*
 732                  * If we have pending delayed iputs then we could free up a
 733                  * bunch of pinned space, so make sure we run the iputs before
 734                  * we do our pinned bytes check below.
 735                  */
 736                 btrfs_run_delayed_iputs(fs_info);
 737                 btrfs_wait_on_delayed_iputs(fs_info);
 738                 break;
 739         case COMMIT_TRANS:
 740                 ret = may_commit_transaction(fs_info, space_info);
 741                 break;
 742         default:
 743                 ret = -ENOSPC;
 744                 break;
 745         }
 746
 747         trace_btrfs_flush_space(fs_info, space_info->flags, num_bytes, state,
 748                                 ret);
 749         return;
 750 }
 751
 752 static inline u64
 753 btrfs_calc_reclaim_metadata_size(struct btrfs_fs_info *fs_info,
 754                                  struct btrfs_space_info *space_info)
 755 {
 756         u64 used;
 757         u64 avail;
 758         u64 expected;
 759         u64 to_reclaim = space_info->reclaim_size;
 760
 761         lockdep_assert_held(&space_info->lock);
 762
 763         avail = calc_available_free_space(fs_info, space_info,
 764                                           BTRFS_RESERVE_FLUSH_ALL);
 765         used = btrfs_space_info_used(space_info, true);
 766
 767         /*
 768          * We may be flushing because suddenly we have less space than we had
 769          * before, and now we're well over-committed based on our current free
 770          * space.  If that's the case add in our overage so we make sure to put
 771          * appropriate pressure on the flushing state machine.
 772          */
 773         if (space_info->total_bytes + avail < used)
 774                 to_reclaim += used - (space_info->total_bytes + avail);
 775
 776         if (to_reclaim)
 777                 return to_reclaim;
 778
 779         to_reclaim = min_t(u64, num_online_cpus() * SZ_1M, SZ_16M);
 780         if (btrfs_can_overcommit(fs_info, space_info, to_reclaim,
 781                                  BTRFS_RESERVE_FLUSH_ALL))
 782                 return 0;
 783
 784         used = btrfs_space_info_used(space_info, true);
 785
 786         if (btrfs_can_overcommit(fs_info, space_info, SZ_1M,
 787                                  BTRFS_RESERVE_FLUSH_ALL))
 788                 expected = div_factor_fine(space_info->total_bytes, 95);
 789         else
 790                 expected = div_factor_fine(space_info->total_bytes, 90);
 791
 792         if (used > expected)
 793                 to_reclaim = used - expected;
 794         else
 795                 to_reclaim = 0;
 796         to_reclaim = min(to_reclaim, space_info->bytes_may_use +
 797                                      space_info->bytes_reserved);
 798         return to_reclaim;
 799 }
 800
 801 static inline int need_do_async_reclaim(struct btrfs_fs_info *fs_info,
 802                                         struct btrfs_space_info *space_info,
 803                                         u64 used)
 804 {
 805         u64 thresh = div_factor_fine(space_info->total_bytes, 98);
 806
 807         /* If we're just plain full then async reclaim just slows us down. */
 808         if ((space_info->bytes_used + space_info->bytes_reserved) >= thresh)
 809                 return 0;
 810
 811         if (!btrfs_calc_reclaim_metadata_size(fs_info, space_info))
 812                 return 0;
 813
 814         return (used >= thresh && !btrfs_fs_closing(fs_info) &&
 815                 !test_bit(BTRFS_FS_STATE_REMOUNTING, &fs_info->fs_state));
 816 }
 817
 818 static bool steal_from_global_rsv(struct btrfs_fs_info *fs_info,
 819                                   struct btrfs_space_info *space_info,
 820                                   struct reserve_ticket *ticket)
 821 {
 822         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
 823         u64 min_bytes;
 824
 825         if (global_rsv->space_info != space_info)
 826                 return false;
 827
 828         spin_lock(&global_rsv->lock);
 829         min_bytes = div_factor(global_rsv->size, 1);
 830         if (global_rsv->reserved < min_bytes + ticket->bytes) {
 831                 spin_unlock(&global_rsv->lock);
 832                 return false;
 833         }
 834         global_rsv->reserved -= ticket->bytes;
 835         remove_ticket(space_info, ticket);
 836         ticket->bytes = 0;
 837         wake_up(&ticket->wait);
 838         space_info->tickets_id++;
 839         if (global_rsv->reserved < global_rsv->size)
 840                 global_rsv->full = 0;
 841         spin_unlock(&global_rsv->lock);
 842
 843         return true;
 844 }
 845
 846 /*
 847  * maybe_fail_all_tickets - we've exhausted our flushing, start failing tickets
 848  * @fs_info - fs_info for this fs
 849  * @space_info - the space info we were flushing
 850  *
 851  * We call this when we've exhausted our flushing ability and haven't made
 852  * progress in satisfying tickets.  The reservation code handles tickets in
 853  * order, so if there is a large ticket first and then smaller ones we could
 854  * very well satisfy the smaller tickets.  This will attempt to wake up any
 855  * tickets in the list to catch this case.
 856  *
 857  * This function returns true if it was able to make progress by clearing out
 858  * other tickets, or if it stumbles across a ticket that was smaller than the
 859  * first ticket.
 860  */
 861 static bool maybe_fail_all_tickets(struct btrfs_fs_info *fs_info,
 862                                    struct btrfs_space_info *space_info)
 863 {
 864         struct reserve_ticket *ticket;
 865         u64 tickets_id = space_info->tickets_id;
 866         u64 first_ticket_bytes = 0;
 867
 868         if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
 869                 btrfs_info(fs_info, "cannot satisfy tickets, dumping space info");
 870                 __btrfs_dump_space_info(fs_info, space_info);
 871         }
 872
 873         while (!list_empty(&space_info->tickets) &&
 874                tickets_id == space_info->tickets_id) {
 875                 ticket = list_first_entry(&space_info->tickets,
 876                                           struct reserve_ticket, list);
 877
 878                 if (ticket->steal &&
 879                     steal_from_global_rsv(fs_info, space_info, ticket))
 880                         return true;
 881
 882                 /*
 883                  * may_commit_transaction will avoid committing the transaction
 884                  * if it doesn't feel like the space reclaimed by the commit
 885                  * would result in the ticket succeeding.  However if we have a
 886                  * smaller ticket in the queue it may be small enough to be
 887                  * satisified by committing the transaction, so if any
 888                  * subsequent ticket is smaller than the first ticket go ahead
 889                  * and send us back for another loop through the enospc flushing
 890                  * code.
 891                  */
 892                 if (first_ticket_bytes == 0)
 893                         first_ticket_bytes = ticket->bytes;
 894                 else if (first_ticket_bytes > ticket->bytes)
 895                         return true;
 896
 897                 if (btrfs_test_opt(fs_info, ENOSPC_DEBUG))
 898                         btrfs_info(fs_info, "failing ticket with %llu bytes",
 899                                    ticket->bytes);
 900
 901                 remove_ticket(space_info, ticket);
 902                 ticket->error = -ENOSPC;
 903                 wake_up(&ticket->wait);
 904
 905                 /*
 906                  * We're just throwing tickets away, so more flushing may not
 907                  * trip over btrfs_try_granting_tickets, so we need to call it
 908                  * here to see if we can make progress with the next ticket in
 909                  * the list.
 910                  */
 911                 btrfs_try_granting_tickets(fs_info, space_info);
 912         }
 913         return (tickets_id != space_info->tickets_id);
 914 }
 915
 916 /*
 917  * This is for normal flushers, we can wait all goddamned day if we want to.  We
 918  * will loop and continuously try to flush as long as we are making progress.
 919  * We count progress as clearing off tickets each time we have to loop.
 920  */
 921 static void btrfs_async_reclaim_metadata_space(struct work_struct *work)
 922 {
 923         struct btrfs_fs_info *fs_info;
 924         struct btrfs_space_info *space_info;
 925         u64 to_reclaim;
 926         int flush_state;
 927         int commit_cycles = 0;
 928         u64 last_tickets_id;
 929
 930         fs_info = container_of(work, struct btrfs_fs_info, async_reclaim_work);
 931         space_info = btrfs_find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
 932
 933         spin_lock(&space_info->lock);
 934         to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info, space_info);
 935         if (!to_reclaim) {
 936                 space_info->flush = 0;
 937                 spin_unlock(&space_info->lock);
 938                 return;
 939         }
 940         last_tickets_id = space_info->tickets_id;
 941         spin_unlock(&space_info->lock);
 942
 943         flush_state = FLUSH_DELAYED_ITEMS_NR;
 944         do {
 945                 flush_space(fs_info, space_info, to_reclaim, flush_state);
 946                 spin_lock(&space_info->lock);
 947                 if (list_empty(&space_info->tickets)) {
 948                         space_info->flush = 0;
 949                         spin_unlock(&space_info->lock);
 950                         return;
 951                 }
 952                 to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info,
 953                                                               space_info);
 954                 if (last_tickets_id == space_info->tickets_id) {
 955                         flush_state++;
 956                 } else {
 957                         last_tickets_id = space_info->tickets_id;
 958                         flush_state = FLUSH_DELAYED_ITEMS_NR;
 959                         if (commit_cycles)
 960                                 commit_cycles--;
 961                 }
 962
 963                 /*
 964                  * We don't want to force a chunk allocation until we've tried
 965                  * pretty hard to reclaim space.  Think of the case where we
 966                  * freed up a bunch of space and so have a lot of pinned space
 967                  * to reclaim.  We would rather use that than possibly create a
 968                  * underutilized metadata chunk.  So if this is our first run
 969                  * through the flushing state machine skip ALLOC_CHUNK_FORCE and
 970                  * commit the transaction.  If nothing has changed the next go
 971                  * around then we can force a chunk allocation.
 972                  */
 973                 if (flush_state == ALLOC_CHUNK_FORCE && !commit_cycles)
 974                         flush_state++;
 975
 976                 if (flush_state > COMMIT_TRANS) {
 977                         commit_cycles++;
 978                         if (commit_cycles > 2) {
 979                                 if (maybe_fail_all_tickets(fs_info, space_info)) {
 980                                         flush_state = FLUSH_DELAYED_ITEMS_NR;
 981                                         commit_cycles--;
 982                                 } else {
 983                                         space_info->flush = 0;
 984                                 }
 985                         } else {
 986                                 flush_state = FLUSH_DELAYED_ITEMS_NR;
 987                         }
 988                 }
 989                 spin_unlock(&space_info->lock);
 990         } while (flush_state <= COMMIT_TRANS);
 991 }
 992
 993 /*
 994  * FLUSH_DELALLOC_WAIT:
 995  *   Space is freed from flushing delalloc in one of two ways.
 996  *
 997  *   1) compression is on and we allocate less space than we reserved
 998  *   2) we are overwriting existing space
 999  *
1000  *   For #1 that extra space is reclaimed as soon as the delalloc pages are
1001  *   COWed, by way of btrfs_add_reserved_bytes() which adds the actual extent
1002  *   length to ->bytes_reserved, and subtracts the reserved space from
1003  *   ->bytes_may_use.
1004  *
1005  *   For #2 this is trickier.  Once the ordered extent runs we will drop the
1006  *   extent in the range we are overwriting, which creates a delayed ref for
1007  *   that freed extent.  This however is not reclaimed until the transaction
1008  *   commits, thus the next stages.
1009  *
1010  * RUN_DELAYED_IPUTS
1011  *   If we are freeing inodes, we want to make sure all delayed iputs have
1012  *   completed, because they could have been on an inode with i_nlink == 0, and
1013  *   thus have been truncated and freed up space.  But again this space is not
1014  *   immediately re-usable, it comes in the form of a delayed ref, which must be
1015  *   run and then the transaction must be committed.
1016  *
1017  * FLUSH_DELAYED_REFS
1018  *   The above two cases generate delayed refs that will affect
1019  *   ->total_bytes_pinned.  However this counter can be inconsistent with
1020  *   reality if there are outstanding delayed refs.  This is because we adjust
1021  *   the counter based solely on the current set of delayed refs and disregard
1022  *   any on-disk state which might include more refs.  So for example, if we
1023  *   have an extent with 2 references, but we only drop 1, we'll see that there
1024  *   is a negative delayed ref count for the extent and assume that the space
1025  *   will be freed, and thus increase ->total_bytes_pinned.
1026  *
1027  *   Running the delayed refs gives us the actual real view of what will be
1028  *   freed at the transaction commit time.  This stage will not actually free
1029  *   space for us, it just makes sure that may_commit_transaction() has all of
1030  *   the information it needs to make the right decision.
1031  *
1032  * COMMIT_TRANS
1033  *   This is where we reclaim all of the pinned space generated by the previous
1034  *   two stages.  We will not commit the transaction if we don't think we're
1035  *   likely to satisfy our request, which means if our current free space +
1036  *   total_bytes_pinned < reservation we will not commit.  This is why the
1037  *   previous states are actually important, to make sure we know for sure
1038  *   whether committing the transaction will allow us to make progress.
1039  *
1040  * ALLOC_CHUNK_FORCE
1041  *   For data we start with alloc chunk force, however we could have been full
1042  *   before, and then the transaction commit could have freed new block groups,
1043  *   so if we now have space to allocate do the force chunk allocation.
1044  */
1045 static const enum btrfs_flush_state data_flush_states[] = {
1046         FLUSH_DELALLOC_WAIT,
1047         RUN_DELAYED_IPUTS,
1048         FLUSH_DELAYED_REFS,
1049         COMMIT_TRANS,
1050         ALLOC_CHUNK_FORCE,
1051 };
1052
1053 static void btrfs_async_reclaim_data_space(struct work_struct *work)
1054 {
1055         struct btrfs_fs_info *fs_info;
1056         struct btrfs_space_info *space_info;
1057         u64 last_tickets_id;
1058         int flush_state = 0;
1059
1060         fs_info = container_of(work, struct btrfs_fs_info, async_data_reclaim_work);
1061         space_info = fs_info->data_sinfo;
1062
1063         spin_lock(&space_info->lock);
1064         if (list_empty(&space_info->tickets)) {
1065                 space_info->flush = 0;
1066                 spin_unlock(&space_info->lock);
1067                 return;
1068         }
1069         last_tickets_id = space_info->tickets_id;
1070         spin_unlock(&space_info->lock);
1071
1072         while (!space_info->full) {
1073                 flush_space(fs_info, space_info, U64_MAX, ALLOC_CHUNK_FORCE);
1074                 spin_lock(&space_info->lock);
1075                 if (list_empty(&space_info->tickets)) {
1076                         space_info->flush = 0;
1077                         spin_unlock(&space_info->lock);
1078                         return;
1079                 }
1080                 last_tickets_id = space_info->tickets_id;
1081                 spin_unlock(&space_info->lock);
1082         }
1083
1084         while (flush_state < ARRAY_SIZE(data_flush_states)) {
1085                 flush_space(fs_info, space_info, U64_MAX,
1086                             data_flush_states[flush_state]);
1087                 spin_lock(&space_info->lock);
1088                 if (list_empty(&space_info->tickets)) {
1089                         space_info->flush = 0;
1090                         spin_unlock(&space_info->lock);
1091                         return;
1092                 }
1093
1094                 if (last_tickets_id == space_info->tickets_id) {
1095                         flush_state++;
1096                 } else {
1097                         last_tickets_id = space_info->tickets_id;
1098                         flush_state = 0;
1099                 }
1100
1101                 if (flush_state >= ARRAY_SIZE(data_flush_states)) {
1102                         if (space_info->full) {
1103                                 if (maybe_fail_all_tickets(fs_info, space_info))
1104                                         flush_state = 0;
1105                                 else
1106                                         space_info->flush = 0;
1107                         } else {
1108                                 flush_state = 0;
1109                         }
1110                 }
1111                 spin_unlock(&space_info->lock);
1112         }
1113 }
1114
1115 void btrfs_init_async_reclaim_work(struct btrfs_fs_info *fs_info)
1116 {
1117         INIT_WORK(&fs_info->async_reclaim_work, btrfs_async_reclaim_metadata_space);
1118         INIT_WORK(&fs_info->async_data_reclaim_work, btrfs_async_reclaim_data_space);
1119 }
1120
1121 static const enum btrfs_flush_state priority_flush_states[] = {
1122         FLUSH_DELAYED_ITEMS_NR,
1123         FLUSH_DELAYED_ITEMS,
1124         ALLOC_CHUNK,
1125 };
1126
1127 static const enum btrfs_flush_state evict_flush_states[] = {
1128         FLUSH_DELAYED_ITEMS_NR,
1129         FLUSH_DELAYED_ITEMS,
1130         FLUSH_DELAYED_REFS_NR,
1131         FLUSH_DELAYED_REFS,
1132         FLUSH_DELALLOC,
1133         FLUSH_DELALLOC_WAIT,
1134         ALLOC_CHUNK,
1135         COMMIT_TRANS,
1136 };
1137
1138 static void priority_reclaim_metadata_space(struct btrfs_fs_info *fs_info,
1139                                 struct btrfs_space_info *space_info,
1140                                 struct reserve_ticket *ticket,
1141                                 const enum btrfs_flush_state *states,
1142                                 int states_nr)
1143 {
1144         u64 to_reclaim;
1145         int flush_state;
1146
1147         spin_lock(&space_info->lock);
1148         to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info, space_info);
1149         if (!to_reclaim) {
1150                 spin_unlock(&space_info->lock);
1151                 return;
1152         }
1153         spin_unlock(&space_info->lock);
1154
1155         flush_state = 0;
1156         do {
1157                 flush_space(fs_info, space_info, to_reclaim, states[flush_state]);
1158                 flush_state++;
1159                 spin_lock(&space_info->lock);
1160                 if (ticket->bytes == 0) {
1161                         spin_unlock(&space_info->lock);
1162                         return;
1163                 }
1164                 spin_unlock(&space_info->lock);
1165         } while (flush_state < states_nr);
1166 }
1167
1168 static void priority_reclaim_data_space(struct btrfs_fs_info *fs_info,
1169                                         struct btrfs_space_info *space_info,
1170                                         struct reserve_ticket *ticket)
1171 {
1172         while (!space_info->full) {
1173                 flush_space(fs_info, space_info, U64_MAX, ALLOC_CHUNK_FORCE);
1174                 spin_lock(&space_info->lock);
1175                 if (ticket->bytes == 0) {
1176                         spin_unlock(&space_info->lock);
1177                         return;
1178                 }
1179                 spin_unlock(&space_info->lock);
1180         }
1181 }
1182
1183 static void wait_reserve_ticket(struct btrfs_fs_info *fs_info,
1184                                 struct btrfs_space_info *space_info,
1185                                 struct reserve_ticket *ticket)
1186
1187 {
1188         DEFINE_WAIT(wait);
1189         int ret = 0;
1190
1191         spin_lock(&space_info->lock);
1192         while (ticket->bytes > 0 && ticket->error == 0) {
1193                 ret = prepare_to_wait_event(&ticket->wait, &wait, TASK_KILLABLE);
1194                 if (ret) {
1195                         /*
1196                          * Delete us from the list. After we unlock the space
1197                          * info, we don't want the async reclaim job to reserve
1198                          * space for this ticket. If that would happen, then the
1199                          * ticket's task would not known that space was reserved
1200                          * despite getting an error, resulting in a space leak
1201                          * (bytes_may_use counter of our space_info).
1202                          */
1203                         remove_ticket(space_info, ticket);
1204                         ticket->error = -EINTR;
1205                         break;
1206                 }
1207                 spin_unlock(&space_info->lock);
1208
1209                 schedule();
1210
1211                 finish_wait(&ticket->wait, &wait);
1212                 spin_lock(&space_info->lock);
1213         }
1214         spin_unlock(&space_info->lock);
1215 }
1216
1217 /**
1218  * handle_reserve_ticket - do the appropriate flushing and waiting for a ticket
1219  * @fs_info - the fs
1220  * @space_info - the space_info for the reservation
1221  * @ticket - the ticket for the reservation
1222  * @flush - how much we can flush
1223  *
1224  * This does the work of figuring out how to flush for the ticket, waiting for
1225  * the reservation, and returning the appropriate error if there is one.
1226  */
1227 static int handle_reserve_ticket(struct btrfs_fs_info *fs_info,
1228                                  struct btrfs_space_info *space_info,
1229                                  struct reserve_ticket *ticket,
1230                                  enum btrfs_reserve_flush_enum flush)
1231 {
1232         int ret;
1233
1234         switch (flush) {
1235         case BTRFS_RESERVE_FLUSH_DATA:
1236         case BTRFS_RESERVE_FLUSH_ALL:
1237         case BTRFS_RESERVE_FLUSH_ALL_STEAL:
1238                 wait_reserve_ticket(fs_info, space_info, ticket);
1239                 break;
1240         case BTRFS_RESERVE_FLUSH_LIMIT:
1241                 priority_reclaim_metadata_space(fs_info, space_info, ticket,
1242                                                 priority_flush_states,
1243                                                 ARRAY_SIZE(priority_flush_states));
1244                 break;
1245         case BTRFS_RESERVE_FLUSH_EVICT:
1246                 priority_reclaim_metadata_space(fs_info, space_info, ticket,
1247                                                 evict_flush_states,
1248                                                 ARRAY_SIZE(evict_flush_states));
1249                 break;
1250         case BTRFS_RESERVE_FLUSH_FREE_SPACE_INODE:
1251                 priority_reclaim_data_space(fs_info, space_info, ticket);
1252                 break;
1253         default:
1254                 ASSERT(0);
1255                 break;
1256         }
1257
1258         spin_lock(&space_info->lock);
1259         ret = ticket->error;
1260         if (ticket->bytes || ticket->error) {
1261                 /*
1262                  * We were a priority ticket, so we need to delete ourselves
1263                  * from the list.  Because we could have other priority tickets
1264                  * behind us that require less space, run
1265                  * btrfs_try_granting_tickets() to see if their reservations can
1266                  * now be made.
1267                  */
1268                 if (!list_empty(&ticket->list)) {
1269                         remove_ticket(space_info, ticket);
1270                         btrfs_try_granting_tickets(fs_info, space_info);
1271                 }
1272
1273                 if (!ret)
1274                         ret = -ENOSPC;
1275         }
1276         spin_unlock(&space_info->lock);
1277         ASSERT(list_empty(&ticket->list));
1278         /*
1279          * Check that we can't have an error set if the reservation succeeded,
1280          * as that would confuse tasks and lead them to error out without
1281          * releasing reserved space (if an error happens the expectation is that
1282          * space wasn't reserved at all).
1283          */
1284         ASSERT(!(ticket->bytes == 0 && ticket->error));
1285         return ret;
1286 }
1287
1288 /*
1289  * This returns true if this flush state will go through the ordinary flushing
1290  * code.
1291  */
1292 static inline bool is_normal_flushing(enum btrfs_reserve_flush_enum flush)
1293 {
1294         return  (flush == BTRFS_RESERVE_FLUSH_ALL) ||
1295                 (flush == BTRFS_RESERVE_FLUSH_ALL_STEAL);
1296 }
1297
1298 /**
1299  * reserve_metadata_bytes - try to reserve bytes from the block_rsv's space
1300  * @root - the root we're allocating for
1301  * @space_info - the space info we want to allocate from
1302  * @orig_bytes - the number of bytes we want
1303  * @flush - whether or not we can flush to make our reservation
1304  *
1305  * This will reserve orig_bytes number of bytes from the space info associated
1306  * with the block_rsv.  If there is not enough space it will make an attempt to
1307  * flush out space to make room.  It will do this by flushing delalloc if
1308  * possible or committing the transaction.  If flush is 0 then no attempts to
1309  * regain reservations will be made and this will fail if there is not enough
1310  * space already.
1311  */
1312 static int __reserve_bytes(struct btrfs_fs_info *fs_info,
1313                            struct btrfs_space_info *space_info, u64 orig_bytes,
1314                            enum btrfs_reserve_flush_enum flush)
1315 {
1316         struct work_struct *async_work;
1317         struct reserve_ticket ticket;
1318         u64 used;
1319         int ret = 0;
1320         bool pending_tickets;
1321
1322         ASSERT(orig_bytes);
1323         ASSERT(!current->journal_info || flush != BTRFS_RESERVE_FLUSH_ALL);
1324
1325         if (flush == BTRFS_RESERVE_FLUSH_DATA)
1326                 async_work = &fs_info->async_data_reclaim_work;
1327         else
1328                 async_work = &fs_info->async_reclaim_work;
1329
1330         spin_lock(&space_info->lock);
1331         ret = -ENOSPC;
1332         used = btrfs_space_info_used(space_info, true);
1333
1334         /*
1335          * We don't want NO_FLUSH allocations to jump everybody, they can
1336          * generally handle ENOSPC in a different way, so treat them the same as
1337          * normal flushers when it comes to skipping pending tickets.
1338          */
1339         if (is_normal_flushing(flush) || (flush == BTRFS_RESERVE_NO_FLUSH))
1340                 pending_tickets = !list_empty(&space_info->tickets) ||
1341                         !list_empty(&space_info->priority_tickets);
1342         else
1343                 pending_tickets = !list_empty(&space_info->priority_tickets);
1344
1345         /*
1346          * Carry on if we have enough space (short-circuit) OR call
1347          * can_overcommit() to ensure we can overcommit to continue.
1348          */
1349         if (!pending_tickets &&
1350             ((used + orig_bytes <= space_info->total_bytes) ||
1351              btrfs_can_overcommit(fs_info, space_info, orig_bytes, flush))) {
1352                 btrfs_space_info_update_bytes_may_use(fs_info, space_info,
1353                                                       orig_bytes);
1354                 ret = 0;
1355         }
1356
1357         /*
1358          * If we couldn't make a reservation then setup our reservation ticket
1359          * and kick the async worker if it's not already running.
1360          *
1361          * If we are a priority flusher then we just need to add our ticket to
1362          * the list and we will do our own flushing further down.
1363          */
1364         if (ret && flush != BTRFS_RESERVE_NO_FLUSH) {
1365                 ticket.bytes = orig_bytes;
1366                 ticket.error = 0;
1367                 space_info->reclaim_size += ticket.bytes;
1368                 init_waitqueue_head(&ticket.wait);
1369                 ticket.steal = (flush == BTRFS_RESERVE_FLUSH_ALL_STEAL);
1370                 if (flush == BTRFS_RESERVE_FLUSH_ALL ||
1371                     flush == BTRFS_RESERVE_FLUSH_ALL_STEAL ||
1372                     flush == BTRFS_RESERVE_FLUSH_DATA) {
1373                         list_add_tail(&ticket.list, &space_info->tickets);
1374                         if (!space_info->flush) {
1375                                 space_info->flush = 1;
1376                                 trace_btrfs_trigger_flush(fs_info,
1377                                                           space_info->flags,
1378                                                           orig_bytes, flush,
1379                                                           "enospc");
1380                                 queue_work(system_unbound_wq, async_work);
1381                         }
1382                 } else {
1383                         list_add_tail(&ticket.list,
1384                                       &space_info->priority_tickets);
1385                 }
1386         } else if (!ret && space_info->flags & BTRFS_BLOCK_GROUP_METADATA) {
1387                 used += orig_bytes;
1388                 /*
1389                  * We will do the space reservation dance during log replay,
1390                  * which means we won't have fs_info->fs_root set, so don't do
1391                  * the async reclaim as we will panic.
1392                  */
1393                 if (!test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags) &&
1394                     need_do_async_reclaim(fs_info, space_info, used) &&
1395                     !work_busy(&fs_info->async_reclaim_work)) {
1396                         trace_btrfs_trigger_flush(fs_info, space_info->flags,
1397                                                   orig_bytes, flush, "preempt");
1398                         queue_work(system_unbound_wq,
1399                                    &fs_info->async_reclaim_work);
1400                 }
1401         }
1402         spin_unlock(&space_info->lock);
1403         if (!ret || flush == BTRFS_RESERVE_NO_FLUSH)
1404                 return ret;
1405
1406         return handle_reserve_ticket(fs_info, space_info, &ticket, flush);
1407 }
1408
1409 /**
1410  * reserve_metadata_bytes - try to reserve bytes from the block_rsv's space
1411  * @root - the root we're allocating for
1412  * @block_rsv - the block_rsv we're allocating for
1413  * @orig_bytes - the number of bytes we want
1414  * @flush - whether or not we can flush to make our reservation
1415  *
1416  * This will reserve orig_bytes number of bytes from the space info associated
1417  * with the block_rsv.  If there is not enough space it will make an attempt to
1418  * flush out space to make room.  It will do this by flushing delalloc if
1419  * possible or committing the transaction.  If flush is 0 then no attempts to
1420  * regain reservations will be made and this will fail if there is not enough
1421  * space already.
1422  */
1423 int btrfs_reserve_metadata_bytes(struct btrfs_root *root,
1424                                  struct btrfs_block_rsv *block_rsv,
1425                                  u64 orig_bytes,
1426                                  enum btrfs_reserve_flush_enum flush)
1427 {
1428         struct btrfs_fs_info *fs_info = root->fs_info;
1429         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
1430         int ret;
1431
1432         ret = __reserve_bytes(fs_info, block_rsv->space_info, orig_bytes, flush);
1433         if (ret == -ENOSPC &&
1434             unlikely(root->orphan_cleanup_state == ORPHAN_CLEANUP_STARTED)) {
1435                 if (block_rsv != global_rsv &&
1436                     !btrfs_block_rsv_use_bytes(global_rsv, orig_bytes))
1437                         ret = 0;
1438         }
1439         if (ret == -ENOSPC) {
1440                 trace_btrfs_space_reservation(fs_info, "space_info:enospc",
1441                                               block_rsv->space_info->flags,
1442                                               orig_bytes, 1);
1443
1444                 if (btrfs_test_opt(fs_info, ENOSPC_DEBUG))
1445                         btrfs_dump_space_info(fs_info, block_rsv->space_info,
1446                                               orig_bytes, 0);
1447         }
1448         return ret;
1449 }
1450
1451 /**
1452  * btrfs_reserve_data_bytes - try to reserve data bytes for an allocation
1453  * @fs_info - the filesystem
1454  * @bytes - the number of bytes we need
1455  * @flush - how we are allowed to flush
1456  *
1457  * This will reserve bytes from the data space info.  If there is not enough
1458  * space then we will attempt to flush space as specified by flush.
1459  */
1460 int btrfs_reserve_data_bytes(struct btrfs_fs_info *fs_info, u64 bytes,
1461                              enum btrfs_reserve_flush_enum flush)
1462 {
1463         struct btrfs_space_info *data_sinfo = fs_info->data_sinfo;
1464         int ret;
1465
1466         ASSERT(flush == BTRFS_RESERVE_FLUSH_DATA ||
1467                flush == BTRFS_RESERVE_FLUSH_FREE_SPACE_INODE);
1468         ASSERT(!current->journal_info || flush != BTRFS_RESERVE_FLUSH_DATA);
1469
1470         ret = __reserve_bytes(fs_info, data_sinfo, bytes, flush);
1471         if (ret == -ENOSPC) {
1472                 trace_btrfs_space_reservation(fs_info, "space_info:enospc",
1473                                               data_sinfo->flags, bytes, 1);
1474                 if (btrfs_test_opt(fs_info, ENOSPC_DEBUG))
1475                         btrfs_dump_space_info(fs_info, data_sinfo, bytes, 0);
1476         }
1477         return ret;
1478 }