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  *   FORCE_COMMIT_TRANS
 144  *     For use by the preemptive flusher.  We use this to bypass the ticketing
 145  *     checks in may_commit_transaction, as we have more information about the
 146  *     overall state of the system and may want to commit the transaction ahead
 147  *     of actual ENOSPC conditions.
 148  *
 149  * OVERCOMMIT
 150  *
 151  *   Because we hold so many reservations for metadata we will allow you to
 152  *   reserve more space than is currently free in the currently allocate
 153  *   metadata space.  This only happens with metadata, data does not allow
 154  *   overcommitting.
 155  *
 156  *   You can see the current logic for when we allow overcommit in
 157  *   btrfs_can_overcommit(), but it only applies to unallocated space.  If there
 158  *   is no unallocated space to be had, all reservations are kept within the
 159  *   free space in the allocated metadata chunks.
 160  *
 161  *   Because of overcommitting, you generally want to use the
 162  *   btrfs_can_overcommit() logic for metadata allocations, as it does the right
 163  *   thing with or without extra unallocated space.
 164  */
 165
 166 u64 __pure btrfs_space_info_used(struct btrfs_space_info *s_info,
 167                           bool may_use_included)
 168 {
 169         ASSERT(s_info);
 170         return s_info->bytes_used + s_info->bytes_reserved +
 171                 s_info->bytes_pinned + s_info->bytes_readonly +
 172                 s_info->bytes_zone_unusable +
 173                 (may_use_included ? s_info->bytes_may_use : 0);
 174 }
 175
 176 /*
 177  * after adding space to the filesystem, we need to clear the full flags
 178  * on all the space infos.
 179  */
 180 void btrfs_clear_space_info_full(struct btrfs_fs_info *info)
 181 {
 182         struct list_head *head = &info->space_info;
 183         struct btrfs_space_info *found;
 184
 185         list_for_each_entry(found, head, list)
 186                 found->full = 0;
 187 }
 188
 189 static int create_space_info(struct btrfs_fs_info *info, u64 flags)
 190 {
 191
 192         struct btrfs_space_info *space_info;
 193         int i;
 194         int ret;
 195
 196         space_info = kzalloc(sizeof(*space_info), GFP_NOFS);
 197         if (!space_info)
 198                 return -ENOMEM;
 199
 200         ret = percpu_counter_init(&space_info->total_bytes_pinned, 0,
 201                                  GFP_KERNEL);
 202         if (ret) {
 203                 kfree(space_info);
 204                 return ret;
 205         }
 206
 207         for (i = 0; i < BTRFS_NR_RAID_TYPES; i++)
 208                 INIT_LIST_HEAD(&space_info->block_groups[i]);
 209         init_rwsem(&space_info->groups_sem);
 210         spin_lock_init(&space_info->lock);
 211         space_info->flags = flags & BTRFS_BLOCK_GROUP_TYPE_MASK;
 212         space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
 213         INIT_LIST_HEAD(&space_info->ro_bgs);
 214         INIT_LIST_HEAD(&space_info->tickets);
 215         INIT_LIST_HEAD(&space_info->priority_tickets);
 216         space_info->clamp = 1;
 217
 218         ret = btrfs_sysfs_add_space_info_type(info, space_info);
 219         if (ret)
 220                 return ret;
 221
 222         list_add(&space_info->list, &info->space_info);
 223         if (flags & BTRFS_BLOCK_GROUP_DATA)
 224                 info->data_sinfo = space_info;
 225
 226         return ret;
 227 }
 228
 229 int btrfs_init_space_info(struct btrfs_fs_info *fs_info)
 230 {
 231         struct btrfs_super_block *disk_super;
 232         u64 features;
 233         u64 flags;
 234         int mixed = 0;
 235         int ret;
 236
 237         disk_super = fs_info->super_copy;
 238         if (!btrfs_super_root(disk_super))
 239                 return -EINVAL;
 240
 241         features = btrfs_super_incompat_flags(disk_super);
 242         if (features & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS)
 243                 mixed = 1;
 244
 245         flags = BTRFS_BLOCK_GROUP_SYSTEM;
 246         ret = create_space_info(fs_info, flags);
 247         if (ret)
 248                 goto out;
 249
 250         if (mixed) {
 251                 flags = BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DATA;
 252                 ret = create_space_info(fs_info, flags);
 253         } else {
 254                 flags = BTRFS_BLOCK_GROUP_METADATA;
 255                 ret = create_space_info(fs_info, flags);
 256                 if (ret)
 257                         goto out;
 258
 259                 flags = BTRFS_BLOCK_GROUP_DATA;
 260                 ret = create_space_info(fs_info, flags);
 261         }
 262 out:
 263         return ret;
 264 }
 265
 266 void btrfs_update_space_info(struct btrfs_fs_info *info, u64 flags,
 267                              u64 total_bytes, u64 bytes_used,
 268                              u64 bytes_readonly, u64 bytes_zone_unusable,
 269                              struct btrfs_space_info **space_info)
 270 {
 271         struct btrfs_space_info *found;
 272         int factor;
 273
 274         factor = btrfs_bg_type_to_factor(flags);
 275
 276         found = btrfs_find_space_info(info, flags);
 277         ASSERT(found);
 278         spin_lock(&found->lock);
 279         found->total_bytes += total_bytes;
 280         found->disk_total += total_bytes * factor;
 281         found->bytes_used += bytes_used;
 282         found->disk_used += bytes_used * factor;
 283         found->bytes_readonly += bytes_readonly;
 284         found->bytes_zone_unusable += bytes_zone_unusable;
 285         if (total_bytes > 0)
 286                 found->full = 0;
 287         btrfs_try_granting_tickets(info, found);
 288         spin_unlock(&found->lock);
 289         *space_info = found;
 290 }
 291
 292 struct btrfs_space_info *btrfs_find_space_info(struct btrfs_fs_info *info,
 293                                                u64 flags)
 294 {
 295         struct list_head *head = &info->space_info;
 296         struct btrfs_space_info *found;
 297
 298         flags &= BTRFS_BLOCK_GROUP_TYPE_MASK;
 299
 300         list_for_each_entry(found, head, list) {
 301                 if (found->flags & flags)
 302                         return found;
 303         }
 304         return NULL;
 305 }
 306
 307 static u64 calc_available_free_space(struct btrfs_fs_info *fs_info,
 308                           struct btrfs_space_info *space_info,
 309                           enum btrfs_reserve_flush_enum flush)
 310 {
 311         u64 profile;
 312         u64 avail;
 313         int factor;
 314
 315         if (space_info->flags & BTRFS_BLOCK_GROUP_SYSTEM)
 316                 profile = btrfs_system_alloc_profile(fs_info);
 317         else
 318                 profile = btrfs_metadata_alloc_profile(fs_info);
 319
 320         avail = atomic64_read(&fs_info->free_chunk_space);
 321
 322         /*
 323          * If we have dup, raid1 or raid10 then only half of the free
 324          * space is actually usable.  For raid56, the space info used
 325          * doesn't include the parity drive, so we don't have to
 326          * change the math
 327          */
 328         factor = btrfs_bg_type_to_factor(profile);
 329         avail = div_u64(avail, factor);
 330
 331         /*
 332          * If we aren't flushing all things, let us overcommit up to
 333          * 1/2th of the space. If we can flush, don't let us overcommit
 334          * too much, let it overcommit up to 1/8 of the space.
 335          */
 336         if (flush == BTRFS_RESERVE_FLUSH_ALL)
 337                 avail >>= 3;
 338         else
 339                 avail >>= 1;
 340         return avail;
 341 }
 342
 343 int btrfs_can_overcommit(struct btrfs_fs_info *fs_info,
 344                          struct btrfs_space_info *space_info, u64 bytes,
 345                          enum btrfs_reserve_flush_enum flush)
 346 {
 347         u64 avail;
 348         u64 used;
 349
 350         /* Don't overcommit when in mixed mode */
 351         if (space_info->flags & BTRFS_BLOCK_GROUP_DATA)
 352                 return 0;
 353
 354         used = btrfs_space_info_used(space_info, true);
 355         avail = calc_available_free_space(fs_info, space_info, flush);
 356
 357         if (used + bytes < space_info->total_bytes + avail)
 358                 return 1;
 359         return 0;
 360 }
 361
 362 static void remove_ticket(struct btrfs_space_info *space_info,
 363                           struct reserve_ticket *ticket)
 364 {
 365         if (!list_empty(&ticket->list)) {
 366                 list_del_init(&ticket->list);
 367                 ASSERT(space_info->reclaim_size >= ticket->bytes);
 368                 space_info->reclaim_size -= ticket->bytes;
 369         }
 370 }
 371
 372 /*
 373  * This is for space we already have accounted in space_info->bytes_may_use, so
 374  * basically when we're returning space from block_rsv's.
 375  */
 376 void btrfs_try_granting_tickets(struct btrfs_fs_info *fs_info,
 377                                 struct btrfs_space_info *space_info)
 378 {
 379         struct list_head *head;
 380         enum btrfs_reserve_flush_enum flush = BTRFS_RESERVE_NO_FLUSH;
 381
 382         lockdep_assert_held(&space_info->lock);
 383
 384         head = &space_info->priority_tickets;
 385 again:
 386         while (!list_empty(head)) {
 387                 struct reserve_ticket *ticket;
 388                 u64 used = btrfs_space_info_used(space_info, true);
 389
 390                 ticket = list_first_entry(head, struct reserve_ticket, list);
 391
 392                 /* Check and see if our ticket can be satisified now. */
 393                 if ((used + ticket->bytes <= space_info->total_bytes) ||
 394                     btrfs_can_overcommit(fs_info, space_info, ticket->bytes,
 395                                          flush)) {
 396                         btrfs_space_info_update_bytes_may_use(fs_info,
 397                                                               space_info,
 398                                                               ticket->bytes);
 399                         remove_ticket(space_info, ticket);
 400                         ticket->bytes = 0;
 401                         space_info->tickets_id++;
 402                         wake_up(&ticket->wait);
 403                 } else {
 404                         break;
 405                 }
 406         }
 407
 408         if (head == &space_info->priority_tickets) {
 409                 head = &space_info->tickets;
 410                 flush = BTRFS_RESERVE_FLUSH_ALL;
 411                 goto again;
 412         }
 413 }
 414
 415 #define DUMP_BLOCK_RSV(fs_info, rsv_name)                               \
 416 do {                                                                    \
 417         struct btrfs_block_rsv *__rsv = &(fs_info)->rsv_name;           \
 418         spin_lock(&__rsv->lock);                                        \
 419         btrfs_info(fs_info, #rsv_name ": size %llu reserved %llu",      \
 420                    __rsv->size, __rsv->reserved);                       \
 421         spin_unlock(&__rsv->lock);                                      \
 422 } while (0)
 423
 424 static void __btrfs_dump_space_info(struct btrfs_fs_info *fs_info,
 425                                     struct btrfs_space_info *info)
 426 {
 427         lockdep_assert_held(&info->lock);
 428
 429         btrfs_info(fs_info, "space_info %llu has %llu free, is %sfull",
 430                    info->flags,
 431                    info->total_bytes - btrfs_space_info_used(info, true),
 432                    info->full ? "" : "not ");
 433         btrfs_info(fs_info,
 434                 "space_info total=%llu, used=%llu, pinned=%llu, reserved=%llu, may_use=%llu, readonly=%llu zone_unusable=%llu",
 435                 info->total_bytes, info->bytes_used, info->bytes_pinned,
 436                 info->bytes_reserved, info->bytes_may_use,
 437                 info->bytes_readonly, info->bytes_zone_unusable);
 438
 439         DUMP_BLOCK_RSV(fs_info, global_block_rsv);
 440         DUMP_BLOCK_RSV(fs_info, trans_block_rsv);
 441         DUMP_BLOCK_RSV(fs_info, chunk_block_rsv);
 442         DUMP_BLOCK_RSV(fs_info, delayed_block_rsv);
 443         DUMP_BLOCK_RSV(fs_info, delayed_refs_rsv);
 444
 445 }
 446
 447 void btrfs_dump_space_info(struct btrfs_fs_info *fs_info,
 448                            struct btrfs_space_info *info, u64 bytes,
 449                            int dump_block_groups)
 450 {
 451         struct btrfs_block_group *cache;
 452         int index = 0;
 453
 454         spin_lock(&info->lock);
 455         __btrfs_dump_space_info(fs_info, info);
 456         spin_unlock(&info->lock);
 457
 458         if (!dump_block_groups)
 459                 return;
 460
 461         down_read(&info->groups_sem);
 462 again:
 463         list_for_each_entry(cache, &info->block_groups[index], list) {
 464                 spin_lock(&cache->lock);
 465                 btrfs_info(fs_info,
 466                         "block group %llu has %llu bytes, %llu used %llu pinned %llu reserved %llu zone_unusable %s",
 467                         cache->start, cache->length, cache->used, cache->pinned,
 468                         cache->reserved, cache->zone_unusable,
 469                         cache->ro ? "[readonly]" : "");
 470                 spin_unlock(&cache->lock);
 471                 btrfs_dump_free_space(cache, bytes);
 472         }
 473         if (++index < BTRFS_NR_RAID_TYPES)
 474                 goto again;
 475         up_read(&info->groups_sem);
 476 }
 477
 478 static inline u64 calc_reclaim_items_nr(struct btrfs_fs_info *fs_info,
 479                                         u64 to_reclaim)
 480 {
 481         u64 bytes;
 482         u64 nr;
 483
 484         bytes = btrfs_calc_insert_metadata_size(fs_info, 1);
 485         nr = div64_u64(to_reclaim, bytes);
 486         if (!nr)
 487                 nr = 1;
 488         return nr;
 489 }
 490
 491 #define EXTENT_SIZE_PER_ITEM    SZ_256K
 492
 493 /*
 494  * shrink metadata reservation for delalloc
 495  */
 496 static void shrink_delalloc(struct btrfs_fs_info *fs_info,
 497                             struct btrfs_space_info *space_info,
 498                             u64 to_reclaim, bool wait_ordered)
 499 {
 500         struct btrfs_trans_handle *trans;
 501         u64 delalloc_bytes;
 502         u64 ordered_bytes;
 503         u64 items;
 504         long time_left;
 505         int loops;
 506
 507         /* Calc the number of the pages we need flush for space reservation */
 508         if (to_reclaim == U64_MAX) {
 509                 items = U64_MAX;
 510         } else {
 511                 /*
 512                  * to_reclaim is set to however much metadata we need to
 513                  * reclaim, but reclaiming that much data doesn't really track
 514                  * exactly, so increase the amount to reclaim by 2x in order to
 515                  * make sure we're flushing enough delalloc to hopefully reclaim
 516                  * some metadata reservations.
 517                  */
 518                 items = calc_reclaim_items_nr(fs_info, to_reclaim) * 2;
 519                 to_reclaim = items * EXTENT_SIZE_PER_ITEM;
 520         }
 521
 522         trans = (struct btrfs_trans_handle *)current->journal_info;
 523
 524         delalloc_bytes = percpu_counter_sum_positive(
 525                                                 &fs_info->delalloc_bytes);
 526         ordered_bytes = percpu_counter_sum_positive(&fs_info->ordered_bytes);
 527         if (delalloc_bytes == 0 && ordered_bytes == 0)
 528                 return;
 529
 530         /*
 531          * If we are doing more ordered than delalloc we need to just wait on
 532          * ordered extents, otherwise we'll waste time trying to flush delalloc
 533          * that likely won't give us the space back we need.
 534          */
 535         if (ordered_bytes > delalloc_bytes)
 536                 wait_ordered = true;
 537
 538         loops = 0;
 539         while ((delalloc_bytes || ordered_bytes) && loops < 3) {
 540                 u64 temp = min(delalloc_bytes, to_reclaim) >> PAGE_SHIFT;
 541                 long nr_pages = min_t(u64, temp, LONG_MAX);
 542
 543                 btrfs_start_delalloc_roots(fs_info, nr_pages, true);
 544
 545                 loops++;
 546                 if (wait_ordered && !trans) {
 547                         btrfs_wait_ordered_roots(fs_info, items, 0, (u64)-1);
 548                 } else {
 549                         time_left = schedule_timeout_killable(1);
 550                         if (time_left)
 551                                 break;
 552                 }
 553
 554                 spin_lock(&space_info->lock);
 555                 if (list_empty(&space_info->tickets) &&
 556                     list_empty(&space_info->priority_tickets)) {
 557                         spin_unlock(&space_info->lock);
 558                         break;
 559                 }
 560                 spin_unlock(&space_info->lock);
 561
 562                 delalloc_bytes = percpu_counter_sum_positive(
 563                                                 &fs_info->delalloc_bytes);
 564                 ordered_bytes = percpu_counter_sum_positive(
 565                                                 &fs_info->ordered_bytes);
 566         }
 567 }
 568
 569 /**
 570  * Possibly commit the transaction if its ok to
 571  *
 572  * @fs_info:    the filesystem
 573  * @space_info: space_info we are checking for commit, either data or metadata
 574  *
 575  * This will check to make sure that committing the transaction will actually
 576  * get us somewhere and then commit the transaction if it does.  Otherwise it
 577  * will return -ENOSPC.
 578  */
 579 static int may_commit_transaction(struct btrfs_fs_info *fs_info,
 580                                   struct btrfs_space_info *space_info)
 581 {
 582         struct reserve_ticket *ticket = NULL;
 583         struct btrfs_block_rsv *delayed_rsv = &fs_info->delayed_block_rsv;
 584         struct btrfs_block_rsv *delayed_refs_rsv = &fs_info->delayed_refs_rsv;
 585         struct btrfs_block_rsv *trans_rsv = &fs_info->trans_block_rsv;
 586         struct btrfs_trans_handle *trans;
 587         u64 reclaim_bytes = 0;
 588         u64 bytes_needed = 0;
 589         u64 cur_free_bytes = 0;
 590
 591         trans = (struct btrfs_trans_handle *)current->journal_info;
 592         if (trans)
 593                 return -EAGAIN;
 594
 595         spin_lock(&space_info->lock);
 596         cur_free_bytes = btrfs_space_info_used(space_info, true);
 597         if (cur_free_bytes < space_info->total_bytes)
 598                 cur_free_bytes = space_info->total_bytes - cur_free_bytes;
 599         else
 600                 cur_free_bytes = 0;
 601
 602         if (!list_empty(&space_info->priority_tickets))
 603                 ticket = list_first_entry(&space_info->priority_tickets,
 604                                           struct reserve_ticket, list);
 605         else if (!list_empty(&space_info->tickets))
 606                 ticket = list_first_entry(&space_info->tickets,
 607                                           struct reserve_ticket, list);
 608         if (ticket)
 609                 bytes_needed = ticket->bytes;
 610
 611         if (bytes_needed > cur_free_bytes)
 612                 bytes_needed -= cur_free_bytes;
 613         else
 614                 bytes_needed = 0;
 615         spin_unlock(&space_info->lock);
 616
 617         if (!bytes_needed)
 618                 return 0;
 619
 620         trans = btrfs_join_transaction(fs_info->extent_root);
 621         if (IS_ERR(trans))
 622                 return PTR_ERR(trans);
 623
 624         /*
 625          * See if there is enough pinned space to make this reservation, or if
 626          * we have block groups that are going to be freed, allowing us to
 627          * possibly do a chunk allocation the next loop through.
 628          */
 629         if (test_bit(BTRFS_TRANS_HAVE_FREE_BGS, &trans->transaction->flags) ||
 630             __percpu_counter_compare(&space_info->total_bytes_pinned,
 631                                      bytes_needed,
 632                                      BTRFS_TOTAL_BYTES_PINNED_BATCH) >= 0)
 633                 goto commit;
 634
 635         /*
 636          * See if there is some space in the delayed insertion reserve for this
 637          * reservation.  If the space_info's don't match (like for DATA or
 638          * SYSTEM) then just go enospc, reclaiming this space won't recover any
 639          * space to satisfy those reservations.
 640          */
 641         if (space_info != delayed_rsv->space_info)
 642                 goto enospc;
 643
 644         spin_lock(&delayed_rsv->lock);
 645         reclaim_bytes += delayed_rsv->reserved;
 646         spin_unlock(&delayed_rsv->lock);
 647
 648         spin_lock(&delayed_refs_rsv->lock);
 649         reclaim_bytes += delayed_refs_rsv->reserved;
 650         spin_unlock(&delayed_refs_rsv->lock);
 651
 652         spin_lock(&trans_rsv->lock);
 653         reclaim_bytes += trans_rsv->reserved;
 654         spin_unlock(&trans_rsv->lock);
 655
 656         if (reclaim_bytes >= bytes_needed)
 657                 goto commit;
 658         bytes_needed -= reclaim_bytes;
 659
 660         if (__percpu_counter_compare(&space_info->total_bytes_pinned,
 661                                    bytes_needed,
 662                                    BTRFS_TOTAL_BYTES_PINNED_BATCH) < 0)
 663                 goto enospc;
 664
 665 commit:
 666         return btrfs_commit_transaction(trans);
 667 enospc:
 668         btrfs_end_transaction(trans);
 669         return -ENOSPC;
 670 }
 671
 672 /*
 673  * Try to flush some data based on policy set by @state. This is only advisory
 674  * and may fail for various reasons. The caller is supposed to examine the
 675  * state of @space_info to detect the outcome.
 676  */
 677 static void flush_space(struct btrfs_fs_info *fs_info,
 678                        struct btrfs_space_info *space_info, u64 num_bytes,
 679                        enum btrfs_flush_state state, bool for_preempt)
 680 {
 681         struct btrfs_root *root = fs_info->extent_root;
 682         struct btrfs_trans_handle *trans;
 683         int nr;
 684         int ret = 0;
 685
 686         switch (state) {
 687         case FLUSH_DELAYED_ITEMS_NR:
 688         case FLUSH_DELAYED_ITEMS:
 689                 if (state == FLUSH_DELAYED_ITEMS_NR)
 690                         nr = calc_reclaim_items_nr(fs_info, num_bytes) * 2;
 691                 else
 692                         nr = -1;
 693
 694                 trans = btrfs_join_transaction(root);
 695                 if (IS_ERR(trans)) {
 696                         ret = PTR_ERR(trans);
 697                         break;
 698                 }
 699                 ret = btrfs_run_delayed_items_nr(trans, nr);
 700                 btrfs_end_transaction(trans);
 701                 break;
 702         case FLUSH_DELALLOC:
 703         case FLUSH_DELALLOC_WAIT:
 704                 shrink_delalloc(fs_info, space_info, num_bytes,
 705                                 state == FLUSH_DELALLOC_WAIT);
 706                 break;
 707         case FLUSH_DELAYED_REFS_NR:
 708         case FLUSH_DELAYED_REFS:
 709                 trans = btrfs_join_transaction(root);
 710                 if (IS_ERR(trans)) {
 711                         ret = PTR_ERR(trans);
 712                         break;
 713                 }
 714                 if (state == FLUSH_DELAYED_REFS_NR)
 715                         nr = calc_reclaim_items_nr(fs_info, num_bytes);
 716                 else
 717                         nr = 0;
 718                 btrfs_run_delayed_refs(trans, nr);
 719                 btrfs_end_transaction(trans);
 720                 break;
 721         case ALLOC_CHUNK:
 722         case ALLOC_CHUNK_FORCE:
 723                 trans = btrfs_join_transaction(root);
 724                 if (IS_ERR(trans)) {
 725                         ret = PTR_ERR(trans);
 726                         break;
 727                 }
 728                 ret = btrfs_chunk_alloc(trans,
 729                                 btrfs_get_alloc_profile(fs_info, space_info->flags),
 730                                 (state == ALLOC_CHUNK) ? CHUNK_ALLOC_NO_FORCE :
 731                                         CHUNK_ALLOC_FORCE);
 732                 btrfs_end_transaction(trans);
 733                 if (ret > 0 || ret == -ENOSPC)
 734                         ret = 0;
 735                 break;
 736         case RUN_DELAYED_IPUTS:
 737                 /*
 738                  * If we have pending delayed iputs then we could free up a
 739                  * bunch of pinned space, so make sure we run the iputs before
 740                  * we do our pinned bytes check below.
 741                  */
 742                 btrfs_run_delayed_iputs(fs_info);
 743                 btrfs_wait_on_delayed_iputs(fs_info);
 744                 break;
 745         case COMMIT_TRANS:
 746                 ret = may_commit_transaction(fs_info, space_info);
 747                 break;
 748         case FORCE_COMMIT_TRANS:
 749                 trans = btrfs_join_transaction(root);
 750                 if (IS_ERR(trans)) {
 751                         ret = PTR_ERR(trans);
 752                         break;
 753                 }
 754                 ret = btrfs_commit_transaction(trans);
 755                 break;
 756         default:
 757                 ret = -ENOSPC;
 758                 break;
 759         }
 760
 761         trace_btrfs_flush_space(fs_info, space_info->flags, num_bytes, state,
 762                                 ret, for_preempt);
 763         return;
 764 }
 765
 766 static inline u64
 767 btrfs_calc_reclaim_metadata_size(struct btrfs_fs_info *fs_info,
 768                                  struct btrfs_space_info *space_info)
 769 {
 770         u64 used;
 771         u64 avail;
 772         u64 to_reclaim = space_info->reclaim_size;
 773
 774         lockdep_assert_held(&space_info->lock);
 775
 776         avail = calc_available_free_space(fs_info, space_info,
 777                                           BTRFS_RESERVE_FLUSH_ALL);
 778         used = btrfs_space_info_used(space_info, true);
 779
 780         /*
 781          * We may be flushing because suddenly we have less space than we had
 782          * before, and now we're well over-committed based on our current free
 783          * space.  If that's the case add in our overage so we make sure to put
 784          * appropriate pressure on the flushing state machine.
 785          */
 786         if (space_info->total_bytes + avail < used)
 787                 to_reclaim += used - (space_info->total_bytes + avail);
 788
 789         return to_reclaim;
 790 }
 791
 792 static bool need_preemptive_reclaim(struct btrfs_fs_info *fs_info,
 793                                     struct btrfs_space_info *space_info)
 794 {
 795         u64 ordered, delalloc;
 796         u64 thresh = div_factor_fine(space_info->total_bytes, 98);
 797         u64 used;
 798
 799         /* If we're just plain full then async reclaim just slows us down. */
 800         if ((space_info->bytes_used + space_info->bytes_reserved) >= thresh)
 801                 return false;
 802
 803         /*
 804          * We have tickets queued, bail so we don't compete with the async
 805          * flushers.
 806          */
 807         if (space_info->reclaim_size)
 808                 return false;
 809
 810         /*
 811          * If we have over half of the free space occupied by reservations or
 812          * pinned then we want to start flushing.
 813          *
 814          * We do not do the traditional thing here, which is to say
 815          *
 816          *   if (used >= ((total_bytes + avail) / 2))
 817          *     return 1;
 818          *
 819          * because this doesn't quite work how we want.  If we had more than 50%
 820          * of the space_info used by bytes_used and we had 0 available we'd just
 821          * constantly run the background flusher.  Instead we want it to kick in
 822          * if our reclaimable space exceeds our clamped free space.
 823          *
 824          * Our clamping range is 2^1 -> 2^8.  Practically speaking that means
 825          * the following:
 826          *
 827          * Amount of RAM        Minimum threshold       Maximum threshold
 828          *
 829          *        256GiB                     1GiB                  128GiB
 830          *        128GiB                   512MiB                   64GiB
 831          *         64GiB                   256MiB                   32GiB
 832          *         32GiB                   128MiB                   16GiB
 833          *         16GiB                    64MiB                    8GiB
 834          *
 835          * These are the range our thresholds will fall in, corresponding to how
 836          * much delalloc we need for the background flusher to kick in.
 837          */
 838
 839         thresh = calc_available_free_space(fs_info, space_info,
 840                                            BTRFS_RESERVE_FLUSH_ALL);
 841         thresh += (space_info->total_bytes - space_info->bytes_used -
 842                    space_info->bytes_reserved - space_info->bytes_readonly);
 843         thresh >>= space_info->clamp;
 844
 845         used = space_info->bytes_pinned;
 846
 847         /*
 848          * If we have more ordered bytes than delalloc bytes then we're either
 849          * doing a lot of DIO, or we simply don't have a lot of delalloc waiting
 850          * around.  Preemptive flushing is only useful in that it can free up
 851          * space before tickets need to wait for things to finish.  In the case
 852          * of ordered extents, preemptively waiting on ordered extents gets us
 853          * nothing, if our reservations are tied up in ordered extents we'll
 854          * simply have to slow down writers by forcing them to wait on ordered
 855          * extents.
 856          *
 857          * In the case that ordered is larger than delalloc, only include the
 858          * block reserves that we would actually be able to directly reclaim
 859          * from.  In this case if we're heavy on metadata operations this will
 860          * clearly be heavy enough to warrant preemptive flushing.  In the case
 861          * of heavy DIO or ordered reservations, preemptive flushing will just
 862          * waste time and cause us to slow down.
 863          */
 864         ordered = percpu_counter_sum_positive(&fs_info->ordered_bytes);
 865         delalloc = percpu_counter_sum_positive(&fs_info->delalloc_bytes);
 866         if (ordered >= delalloc)
 867                 used += fs_info->delayed_refs_rsv.reserved +
 868                         fs_info->delayed_block_rsv.reserved;
 869         else
 870                 used += space_info->bytes_may_use;
 871
 872         return (used >= thresh && !btrfs_fs_closing(fs_info) &&
 873                 !test_bit(BTRFS_FS_STATE_REMOUNTING, &fs_info->fs_state));
 874 }
 875
 876 static bool steal_from_global_rsv(struct btrfs_fs_info *fs_info,
 877                                   struct btrfs_space_info *space_info,
 878                                   struct reserve_ticket *ticket)
 879 {
 880         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
 881         u64 min_bytes;
 882
 883         if (global_rsv->space_info != space_info)
 884                 return false;
 885
 886         spin_lock(&global_rsv->lock);
 887         min_bytes = div_factor(global_rsv->size, 1);
 888         if (global_rsv->reserved < min_bytes + ticket->bytes) {
 889                 spin_unlock(&global_rsv->lock);
 890                 return false;
 891         }
 892         global_rsv->reserved -= ticket->bytes;
 893         remove_ticket(space_info, ticket);
 894         ticket->bytes = 0;
 895         wake_up(&ticket->wait);
 896         space_info->tickets_id++;
 897         if (global_rsv->reserved < global_rsv->size)
 898                 global_rsv->full = 0;
 899         spin_unlock(&global_rsv->lock);
 900
 901         return true;
 902 }
 903
 904 /*
 905  * maybe_fail_all_tickets - we've exhausted our flushing, start failing tickets
 906  * @fs_info - fs_info for this fs
 907  * @space_info - the space info we were flushing
 908  *
 909  * We call this when we've exhausted our flushing ability and haven't made
 910  * progress in satisfying tickets.  The reservation code handles tickets in
 911  * order, so if there is a large ticket first and then smaller ones we could
 912  * very well satisfy the smaller tickets.  This will attempt to wake up any
 913  * tickets in the list to catch this case.
 914  *
 915  * This function returns true if it was able to make progress by clearing out
 916  * other tickets, or if it stumbles across a ticket that was smaller than the
 917  * first ticket.
 918  */
 919 static bool maybe_fail_all_tickets(struct btrfs_fs_info *fs_info,
 920                                    struct btrfs_space_info *space_info)
 921 {
 922         struct reserve_ticket *ticket;
 923         u64 tickets_id = space_info->tickets_id;
 924         u64 first_ticket_bytes = 0;
 925
 926         if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
 927                 btrfs_info(fs_info, "cannot satisfy tickets, dumping space info");
 928                 __btrfs_dump_space_info(fs_info, space_info);
 929         }
 930
 931         while (!list_empty(&space_info->tickets) &&
 932                tickets_id == space_info->tickets_id) {
 933                 ticket = list_first_entry(&space_info->tickets,
 934                                           struct reserve_ticket, list);
 935
 936                 if (ticket->steal &&
 937                     steal_from_global_rsv(fs_info, space_info, ticket))
 938                         return true;
 939
 940                 /*
 941                  * may_commit_transaction will avoid committing the transaction
 942                  * if it doesn't feel like the space reclaimed by the commit
 943                  * would result in the ticket succeeding.  However if we have a
 944                  * smaller ticket in the queue it may be small enough to be
 945                  * satisified by committing the transaction, so if any
 946                  * subsequent ticket is smaller than the first ticket go ahead
 947                  * and send us back for another loop through the enospc flushing
 948                  * code.
 949                  */
 950                 if (first_ticket_bytes == 0)
 951                         first_ticket_bytes = ticket->bytes;
 952                 else if (first_ticket_bytes > ticket->bytes)
 953                         return true;
 954
 955                 if (btrfs_test_opt(fs_info, ENOSPC_DEBUG))
 956                         btrfs_info(fs_info, "failing ticket with %llu bytes",
 957                                    ticket->bytes);
 958
 959                 remove_ticket(space_info, ticket);
 960                 ticket->error = -ENOSPC;
 961                 wake_up(&ticket->wait);
 962
 963                 /*
 964                  * We're just throwing tickets away, so more flushing may not
 965                  * trip over btrfs_try_granting_tickets, so we need to call it
 966                  * here to see if we can make progress with the next ticket in
 967                  * the list.
 968                  */
 969                 btrfs_try_granting_tickets(fs_info, space_info);
 970         }
 971         return (tickets_id != space_info->tickets_id);
 972 }
 973
 974 /*
 975  * This is for normal flushers, we can wait all goddamned day if we want to.  We
 976  * will loop and continuously try to flush as long as we are making progress.
 977  * We count progress as clearing off tickets each time we have to loop.
 978  */
 979 static void btrfs_async_reclaim_metadata_space(struct work_struct *work)
 980 {
 981         struct btrfs_fs_info *fs_info;
 982         struct btrfs_space_info *space_info;
 983         u64 to_reclaim;
 984         enum btrfs_flush_state flush_state;
 985         int commit_cycles = 0;
 986         u64 last_tickets_id;
 987
 988         fs_info = container_of(work, struct btrfs_fs_info, async_reclaim_work);
 989         space_info = btrfs_find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
 990
 991         spin_lock(&space_info->lock);
 992         to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info, space_info);
 993         if (!to_reclaim) {
 994                 space_info->flush = 0;
 995                 spin_unlock(&space_info->lock);
 996                 return;
 997         }
 998         last_tickets_id = space_info->tickets_id;
 999         spin_unlock(&space_info->lock);
1000
1001         flush_state = FLUSH_DELAYED_ITEMS_NR;
1002         do {
1003                 flush_space(fs_info, space_info, to_reclaim, flush_state, false);
1004                 spin_lock(&space_info->lock);
1005                 if (list_empty(&space_info->tickets)) {
1006                         space_info->flush = 0;
1007                         spin_unlock(&space_info->lock);
1008                         return;
1009                 }
1010                 to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info,
1011                                                               space_info);
1012                 if (last_tickets_id == space_info->tickets_id) {
1013                         flush_state++;
1014                 } else {
1015                         last_tickets_id = space_info->tickets_id;
1016                         flush_state = FLUSH_DELAYED_ITEMS_NR;
1017                         if (commit_cycles)
1018                                 commit_cycles--;
1019                 }
1020
1021                 /*
1022                  * We don't want to force a chunk allocation until we've tried
1023                  * pretty hard to reclaim space.  Think of the case where we
1024                  * freed up a bunch of space and so have a lot of pinned space
1025                  * to reclaim.  We would rather use that than possibly create a
1026                  * underutilized metadata chunk.  So if this is our first run
1027                  * through the flushing state machine skip ALLOC_CHUNK_FORCE and
1028                  * commit the transaction.  If nothing has changed the next go
1029                  * around then we can force a chunk allocation.
1030                  */
1031                 if (flush_state == ALLOC_CHUNK_FORCE && !commit_cycles)
1032                         flush_state++;
1033
1034                 if (flush_state > COMMIT_TRANS) {
1035                         commit_cycles++;
1036                         if (commit_cycles > 2) {
1037                                 if (maybe_fail_all_tickets(fs_info, space_info)) {
1038                                         flush_state = FLUSH_DELAYED_ITEMS_NR;
1039                                         commit_cycles--;
1040                                 } else {
1041                                         space_info->flush = 0;
1042                                 }
1043                         } else {
1044                                 flush_state = FLUSH_DELAYED_ITEMS_NR;
1045                         }
1046                 }
1047                 spin_unlock(&space_info->lock);
1048         } while (flush_state <= COMMIT_TRANS);
1049 }
1050
1051 /*
1052  * This handles pre-flushing of metadata space before we get to the point that
1053  * we need to start blocking threads on tickets.  The logic here is different
1054  * from the other flush paths because it doesn't rely on tickets to tell us how
1055  * much we need to flush, instead it attempts to keep us below the 80% full
1056  * watermark of space by flushing whichever reservation pool is currently the
1057  * largest.
1058  */
1059 static void btrfs_preempt_reclaim_metadata_space(struct work_struct *work)
1060 {
1061         struct btrfs_fs_info *fs_info;
1062         struct btrfs_space_info *space_info;
1063         struct btrfs_block_rsv *delayed_block_rsv;
1064         struct btrfs_block_rsv *delayed_refs_rsv;
1065         struct btrfs_block_rsv *global_rsv;
1066         struct btrfs_block_rsv *trans_rsv;
1067         int loops = 0;
1068
1069         fs_info = container_of(work, struct btrfs_fs_info,
1070                                preempt_reclaim_work);
1071         space_info = btrfs_find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
1072         delayed_block_rsv = &fs_info->delayed_block_rsv;
1073         delayed_refs_rsv = &fs_info->delayed_refs_rsv;
1074         global_rsv = &fs_info->global_block_rsv;
1075         trans_rsv = &fs_info->trans_block_rsv;
1076
1077         spin_lock(&space_info->lock);
1078         while (need_preemptive_reclaim(fs_info, space_info)) {
1079                 enum btrfs_flush_state flush;
1080                 u64 delalloc_size = 0;
1081                 u64 to_reclaim, block_rsv_size;
1082                 u64 global_rsv_size = global_rsv->reserved;
1083
1084                 loops++;
1085
1086                 /*
1087                  * We don't have a precise counter for the metadata being
1088                  * reserved for delalloc, so we'll approximate it by subtracting
1089                  * out the block rsv's space from the bytes_may_use.  If that
1090                  * amount is higher than the individual reserves, then we can
1091                  * assume it's tied up in delalloc reservations.
1092                  */
1093                 block_rsv_size = global_rsv_size +
1094                         delayed_block_rsv->reserved +
1095                         delayed_refs_rsv->reserved +
1096                         trans_rsv->reserved;
1097                 if (block_rsv_size < space_info->bytes_may_use)
1098                         delalloc_size = space_info->bytes_may_use - block_rsv_size;
1099                 spin_unlock(&space_info->lock);
1100
1101                 /*
1102                  * We don't want to include the global_rsv in our calculation,
1103                  * because that's space we can't touch.  Subtract it from the
1104                  * block_rsv_size for the next checks.
1105                  */
1106                 block_rsv_size -= global_rsv_size;
1107
1108                 /*
1109                  * We really want to avoid flushing delalloc too much, as it
1110                  * could result in poor allocation patterns, so only flush it if
1111                  * it's larger than the rest of the pools combined.
1112                  */
1113                 if (delalloc_size > block_rsv_size) {
1114                         to_reclaim = delalloc_size;
1115                         flush = FLUSH_DELALLOC;
1116                 } else if (space_info->bytes_pinned >
1117                            (delayed_block_rsv->reserved +
1118                             delayed_refs_rsv->reserved)) {
1119                         to_reclaim = space_info->bytes_pinned;
1120                         flush = FORCE_COMMIT_TRANS;
1121                 } else if (delayed_block_rsv->reserved >
1122                            delayed_refs_rsv->reserved) {
1123                         to_reclaim = delayed_block_rsv->reserved;
1124                         flush = FLUSH_DELAYED_ITEMS_NR;
1125                 } else {
1126                         to_reclaim = delayed_refs_rsv->reserved;
1127                         flush = FLUSH_DELAYED_REFS_NR;
1128                 }
1129
1130                 /*
1131                  * We don't want to reclaim everything, just a portion, so scale
1132                  * down the to_reclaim by 1/4.  If it takes us down to 0,
1133                  * reclaim 1 items worth.
1134                  */
1135                 to_reclaim >>= 2;
1136                 if (!to_reclaim)
1137                         to_reclaim = btrfs_calc_insert_metadata_size(fs_info, 1);
1138                 flush_space(fs_info, space_info, to_reclaim, flush, true);
1139                 cond_resched();
1140                 spin_lock(&space_info->lock);
1141         }
1142
1143         /* We only went through once, back off our clamping. */
1144         if (loops == 1 && !space_info->reclaim_size)
1145                 space_info->clamp = max(1, space_info->clamp - 1);
1146         trace_btrfs_done_preemptive_reclaim(fs_info, space_info);
1147         spin_unlock(&space_info->lock);
1148 }
1149
1150 /*
1151  * FLUSH_DELALLOC_WAIT:
1152  *   Space is freed from flushing delalloc in one of two ways.
1153  *
1154  *   1) compression is on and we allocate less space than we reserved
1155  *   2) we are overwriting existing space
1156  *
1157  *   For #1 that extra space is reclaimed as soon as the delalloc pages are
1158  *   COWed, by way of btrfs_add_reserved_bytes() which adds the actual extent
1159  *   length to ->bytes_reserved, and subtracts the reserved space from
1160  *   ->bytes_may_use.
1161  *
1162  *   For #2 this is trickier.  Once the ordered extent runs we will drop the
1163  *   extent in the range we are overwriting, which creates a delayed ref for
1164  *   that freed extent.  This however is not reclaimed until the transaction
1165  *   commits, thus the next stages.
1166  *
1167  * RUN_DELAYED_IPUTS
1168  *   If we are freeing inodes, we want to make sure all delayed iputs have
1169  *   completed, because they could have been on an inode with i_nlink == 0, and
1170  *   thus have been truncated and freed up space.  But again this space is not
1171  *   immediately re-usable, it comes in the form of a delayed ref, which must be
1172  *   run and then the transaction must be committed.
1173  *
1174  * FLUSH_DELAYED_REFS
1175  *   The above two cases generate delayed refs that will affect
1176  *   ->total_bytes_pinned.  However this counter can be inconsistent with
1177  *   reality if there are outstanding delayed refs.  This is because we adjust
1178  *   the counter based solely on the current set of delayed refs and disregard
1179  *   any on-disk state which might include more refs.  So for example, if we
1180  *   have an extent with 2 references, but we only drop 1, we'll see that there
1181  *   is a negative delayed ref count for the extent and assume that the space
1182  *   will be freed, and thus increase ->total_bytes_pinned.
1183  *
1184  *   Running the delayed refs gives us the actual real view of what will be
1185  *   freed at the transaction commit time.  This stage will not actually free
1186  *   space for us, it just makes sure that may_commit_transaction() has all of
1187  *   the information it needs to make the right decision.
1188  *
1189  * COMMIT_TRANS
1190  *   This is where we reclaim all of the pinned space generated by the previous
1191  *   two stages.  We will not commit the transaction if we don't think we're
1192  *   likely to satisfy our request, which means if our current free space +
1193  *   total_bytes_pinned < reservation we will not commit.  This is why the
1194  *   previous states are actually important, to make sure we know for sure
1195  *   whether committing the transaction will allow us to make progress.
1196  *
1197  * ALLOC_CHUNK_FORCE
1198  *   For data we start with alloc chunk force, however we could have been full
1199  *   before, and then the transaction commit could have freed new block groups,
1200  *   so if we now have space to allocate do the force chunk allocation.
1201  */
1202 static const enum btrfs_flush_state data_flush_states[] = {
1203         FLUSH_DELALLOC_WAIT,
1204         RUN_DELAYED_IPUTS,
1205         FLUSH_DELAYED_REFS,
1206         COMMIT_TRANS,
1207         ALLOC_CHUNK_FORCE,
1208 };
1209
1210 static void btrfs_async_reclaim_data_space(struct work_struct *work)
1211 {
1212         struct btrfs_fs_info *fs_info;
1213         struct btrfs_space_info *space_info;
1214         u64 last_tickets_id;
1215         enum btrfs_flush_state flush_state = 0;
1216
1217         fs_info = container_of(work, struct btrfs_fs_info, async_data_reclaim_work);
1218         space_info = fs_info->data_sinfo;
1219
1220         spin_lock(&space_info->lock);
1221         if (list_empty(&space_info->tickets)) {
1222                 space_info->flush = 0;
1223                 spin_unlock(&space_info->lock);
1224                 return;
1225         }
1226         last_tickets_id = space_info->tickets_id;
1227         spin_unlock(&space_info->lock);
1228
1229         while (!space_info->full) {
1230                 flush_space(fs_info, space_info, U64_MAX, ALLOC_CHUNK_FORCE, false);
1231                 spin_lock(&space_info->lock);
1232                 if (list_empty(&space_info->tickets)) {
1233                         space_info->flush = 0;
1234                         spin_unlock(&space_info->lock);
1235                         return;
1236                 }
1237                 last_tickets_id = space_info->tickets_id;
1238                 spin_unlock(&space_info->lock);
1239         }
1240
1241         while (flush_state < ARRAY_SIZE(data_flush_states)) {
1242                 flush_space(fs_info, space_info, U64_MAX,
1243                             data_flush_states[flush_state], false);
1244                 spin_lock(&space_info->lock);
1245                 if (list_empty(&space_info->tickets)) {
1246                         space_info->flush = 0;
1247                         spin_unlock(&space_info->lock);
1248                         return;
1249                 }
1250
1251                 if (last_tickets_id == space_info->tickets_id) {
1252                         flush_state++;
1253                 } else {
1254                         last_tickets_id = space_info->tickets_id;
1255                         flush_state = 0;
1256                 }
1257
1258                 if (flush_state >= ARRAY_SIZE(data_flush_states)) {
1259                         if (space_info->full) {
1260                                 if (maybe_fail_all_tickets(fs_info, space_info))
1261                                         flush_state = 0;
1262                                 else
1263                                         space_info->flush = 0;
1264                         } else {
1265                                 flush_state = 0;
1266                         }
1267                 }
1268                 spin_unlock(&space_info->lock);
1269         }
1270 }
1271
1272 void btrfs_init_async_reclaim_work(struct btrfs_fs_info *fs_info)
1273 {
1274         INIT_WORK(&fs_info->async_reclaim_work, btrfs_async_reclaim_metadata_space);
1275         INIT_WORK(&fs_info->async_data_reclaim_work, btrfs_async_reclaim_data_space);
1276         INIT_WORK(&fs_info->preempt_reclaim_work,
1277                   btrfs_preempt_reclaim_metadata_space);
1278 }
1279
1280 static const enum btrfs_flush_state priority_flush_states[] = {
1281         FLUSH_DELAYED_ITEMS_NR,
1282         FLUSH_DELAYED_ITEMS,
1283         ALLOC_CHUNK,
1284 };
1285
1286 static const enum btrfs_flush_state evict_flush_states[] = {
1287         FLUSH_DELAYED_ITEMS_NR,
1288         FLUSH_DELAYED_ITEMS,
1289         FLUSH_DELAYED_REFS_NR,
1290         FLUSH_DELAYED_REFS,
1291         FLUSH_DELALLOC,
1292         FLUSH_DELALLOC_WAIT,
1293         ALLOC_CHUNK,
1294         COMMIT_TRANS,
1295 };
1296
1297 static void priority_reclaim_metadata_space(struct btrfs_fs_info *fs_info,
1298                                 struct btrfs_space_info *space_info,
1299                                 struct reserve_ticket *ticket,
1300                                 const enum btrfs_flush_state *states,
1301                                 int states_nr)
1302 {
1303         u64 to_reclaim;
1304         int flush_state;
1305
1306         spin_lock(&space_info->lock);
1307         to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info, space_info);
1308         if (!to_reclaim) {
1309                 spin_unlock(&space_info->lock);
1310                 return;
1311         }
1312         spin_unlock(&space_info->lock);
1313
1314         flush_state = 0;
1315         do {
1316                 flush_space(fs_info, space_info, to_reclaim, states[flush_state],
1317                             false);
1318                 flush_state++;
1319                 spin_lock(&space_info->lock);
1320                 if (ticket->bytes == 0) {
1321                         spin_unlock(&space_info->lock);
1322                         return;
1323                 }
1324                 spin_unlock(&space_info->lock);
1325         } while (flush_state < states_nr);
1326 }
1327
1328 static void priority_reclaim_data_space(struct btrfs_fs_info *fs_info,
1329                                         struct btrfs_space_info *space_info,
1330                                         struct reserve_ticket *ticket)
1331 {
1332         while (!space_info->full) {
1333                 flush_space(fs_info, space_info, U64_MAX, ALLOC_CHUNK_FORCE, false);
1334                 spin_lock(&space_info->lock);
1335                 if (ticket->bytes == 0) {
1336                         spin_unlock(&space_info->lock);
1337                         return;
1338                 }
1339                 spin_unlock(&space_info->lock);
1340         }
1341 }
1342
1343 static void wait_reserve_ticket(struct btrfs_fs_info *fs_info,
1344                                 struct btrfs_space_info *space_info,
1345                                 struct reserve_ticket *ticket)
1346
1347 {
1348         DEFINE_WAIT(wait);
1349         int ret = 0;
1350
1351         spin_lock(&space_info->lock);
1352         while (ticket->bytes > 0 && ticket->error == 0) {
1353                 ret = prepare_to_wait_event(&ticket->wait, &wait, TASK_KILLABLE);
1354                 if (ret) {
1355                         /*
1356                          * Delete us from the list. After we unlock the space
1357                          * info, we don't want the async reclaim job to reserve
1358                          * space for this ticket. If that would happen, then the
1359                          * ticket's task would not known that space was reserved
1360                          * despite getting an error, resulting in a space leak
1361                          * (bytes_may_use counter of our space_info).
1362                          */
1363                         remove_ticket(space_info, ticket);
1364                         ticket->error = -EINTR;
1365                         break;
1366                 }
1367                 spin_unlock(&space_info->lock);
1368
1369                 schedule();
1370
1371                 finish_wait(&ticket->wait, &wait);
1372                 spin_lock(&space_info->lock);
1373         }
1374         spin_unlock(&space_info->lock);
1375 }
1376
1377 /**
1378  * Do the appropriate flushing and waiting for a ticket
1379  *
1380  * @fs_info:    the filesystem
1381  * @space_info: space info for the reservation
1382  * @ticket:     ticket for the reservation
1383  * @start_ns:   timestamp when the reservation started
1384  * @orig_bytes: amount of bytes originally reserved
1385  * @flush:      how much we can flush
1386  *
1387  * This does the work of figuring out how to flush for the ticket, waiting for
1388  * the reservation, and returning the appropriate error if there is one.
1389  */
1390 static int handle_reserve_ticket(struct btrfs_fs_info *fs_info,
1391                                  struct btrfs_space_info *space_info,
1392                                  struct reserve_ticket *ticket,
1393                                  u64 start_ns, u64 orig_bytes,
1394                                  enum btrfs_reserve_flush_enum flush)
1395 {
1396         int ret;
1397
1398         switch (flush) {
1399         case BTRFS_RESERVE_FLUSH_DATA:
1400         case BTRFS_RESERVE_FLUSH_ALL:
1401         case BTRFS_RESERVE_FLUSH_ALL_STEAL:
1402                 wait_reserve_ticket(fs_info, space_info, ticket);
1403                 break;
1404         case BTRFS_RESERVE_FLUSH_LIMIT:
1405                 priority_reclaim_metadata_space(fs_info, space_info, ticket,
1406                                                 priority_flush_states,
1407                                                 ARRAY_SIZE(priority_flush_states));
1408                 break;
1409         case BTRFS_RESERVE_FLUSH_EVICT:
1410                 priority_reclaim_metadata_space(fs_info, space_info, ticket,
1411                                                 evict_flush_states,
1412                                                 ARRAY_SIZE(evict_flush_states));
1413                 break;
1414         case BTRFS_RESERVE_FLUSH_FREE_SPACE_INODE:
1415                 priority_reclaim_data_space(fs_info, space_info, ticket);
1416                 break;
1417         default:
1418                 ASSERT(0);
1419                 break;
1420         }
1421
1422         spin_lock(&space_info->lock);
1423         ret = ticket->error;
1424         if (ticket->bytes || ticket->error) {
1425                 /*
1426                  * We were a priority ticket, so we need to delete ourselves
1427                  * from the list.  Because we could have other priority tickets
1428                  * behind us that require less space, run
1429                  * btrfs_try_granting_tickets() to see if their reservations can
1430                  * now be made.
1431                  */
1432                 if (!list_empty(&ticket->list)) {
1433                         remove_ticket(space_info, ticket);
1434                         btrfs_try_granting_tickets(fs_info, space_info);
1435                 }
1436
1437                 if (!ret)
1438                         ret = -ENOSPC;
1439         }
1440         spin_unlock(&space_info->lock);
1441         ASSERT(list_empty(&ticket->list));
1442         /*
1443          * Check that we can't have an error set if the reservation succeeded,
1444          * as that would confuse tasks and lead them to error out without
1445          * releasing reserved space (if an error happens the expectation is that
1446          * space wasn't reserved at all).
1447          */
1448         ASSERT(!(ticket->bytes == 0 && ticket->error));
1449         trace_btrfs_reserve_ticket(fs_info, space_info->flags, orig_bytes,
1450                                    start_ns, flush, ticket->error);
1451         return ret;
1452 }
1453
1454 /*
1455  * This returns true if this flush state will go through the ordinary flushing
1456  * code.
1457  */
1458 static inline bool is_normal_flushing(enum btrfs_reserve_flush_enum flush)
1459 {
1460         return  (flush == BTRFS_RESERVE_FLUSH_ALL) ||
1461                 (flush == BTRFS_RESERVE_FLUSH_ALL_STEAL);
1462 }
1463
1464 static inline void maybe_clamp_preempt(struct btrfs_fs_info *fs_info,
1465                                        struct btrfs_space_info *space_info)
1466 {
1467         u64 ordered = percpu_counter_sum_positive(&fs_info->ordered_bytes);
1468         u64 delalloc = percpu_counter_sum_positive(&fs_info->delalloc_bytes);
1469
1470         /*
1471          * If we're heavy on ordered operations then clamping won't help us.  We
1472          * need to clamp specifically to keep up with dirty'ing buffered
1473          * writers, because there's not a 1:1 correlation of writing delalloc
1474          * and freeing space, like there is with flushing delayed refs or
1475          * delayed nodes.  If we're already more ordered than delalloc then
1476          * we're keeping up, otherwise we aren't and should probably clamp.
1477          */
1478         if (ordered < delalloc)
1479                 space_info->clamp = min(space_info->clamp + 1, 8);
1480 }
1481
1482 /**
1483  * Try to reserve bytes from the block_rsv's space
1484  *
1485  * @fs_info:    the filesystem
1486  * @space_info: space info we want to allocate from
1487  * @orig_bytes: number of bytes we want
1488  * @flush:      whether or not we can flush to make our reservation
1489  *
1490  * This will reserve orig_bytes number of bytes from the space info associated
1491  * with the block_rsv.  If there is not enough space it will make an attempt to
1492  * flush out space to make room.  It will do this by flushing delalloc if
1493  * possible or committing the transaction.  If flush is 0 then no attempts to
1494  * regain reservations will be made and this will fail if there is not enough
1495  * space already.
1496  */
1497 static int __reserve_bytes(struct btrfs_fs_info *fs_info,
1498                            struct btrfs_space_info *space_info, u64 orig_bytes,
1499                            enum btrfs_reserve_flush_enum flush)
1500 {
1501         struct work_struct *async_work;
1502         struct reserve_ticket ticket;
1503         u64 start_ns = 0;
1504         u64 used;
1505         int ret = 0;
1506         bool pending_tickets;
1507
1508         ASSERT(orig_bytes);
1509         ASSERT(!current->journal_info || flush != BTRFS_RESERVE_FLUSH_ALL);
1510
1511         if (flush == BTRFS_RESERVE_FLUSH_DATA)
1512                 async_work = &fs_info->async_data_reclaim_work;
1513         else
1514                 async_work = &fs_info->async_reclaim_work;
1515
1516         spin_lock(&space_info->lock);
1517         ret = -ENOSPC;
1518         used = btrfs_space_info_used(space_info, true);
1519
1520         /*
1521          * We don't want NO_FLUSH allocations to jump everybody, they can
1522          * generally handle ENOSPC in a different way, so treat them the same as
1523          * normal flushers when it comes to skipping pending tickets.
1524          */
1525         if (is_normal_flushing(flush) || (flush == BTRFS_RESERVE_NO_FLUSH))
1526                 pending_tickets = !list_empty(&space_info->tickets) ||
1527                         !list_empty(&space_info->priority_tickets);
1528         else
1529                 pending_tickets = !list_empty(&space_info->priority_tickets);
1530
1531         /*
1532          * Carry on if we have enough space (short-circuit) OR call
1533          * can_overcommit() to ensure we can overcommit to continue.
1534          */
1535         if (!pending_tickets &&
1536             ((used + orig_bytes <= space_info->total_bytes) ||
1537              btrfs_can_overcommit(fs_info, space_info, orig_bytes, flush))) {
1538                 btrfs_space_info_update_bytes_may_use(fs_info, space_info,
1539                                                       orig_bytes);
1540                 ret = 0;
1541         }
1542
1543         /*
1544          * If we couldn't make a reservation then setup our reservation ticket
1545          * and kick the async worker if it's not already running.
1546          *
1547          * If we are a priority flusher then we just need to add our ticket to
1548          * the list and we will do our own flushing further down.
1549          */
1550         if (ret && flush != BTRFS_RESERVE_NO_FLUSH) {
1551                 ticket.bytes = orig_bytes;
1552                 ticket.error = 0;
1553                 space_info->reclaim_size += ticket.bytes;
1554                 init_waitqueue_head(&ticket.wait);
1555                 ticket.steal = (flush == BTRFS_RESERVE_FLUSH_ALL_STEAL);
1556                 if (trace_btrfs_reserve_ticket_enabled())
1557                         start_ns = ktime_get_ns();
1558
1559                 if (flush == BTRFS_RESERVE_FLUSH_ALL ||
1560                     flush == BTRFS_RESERVE_FLUSH_ALL_STEAL ||
1561                     flush == BTRFS_RESERVE_FLUSH_DATA) {
1562                         list_add_tail(&ticket.list, &space_info->tickets);
1563                         if (!space_info->flush) {
1564                                 space_info->flush = 1;
1565                                 trace_btrfs_trigger_flush(fs_info,
1566                                                           space_info->flags,
1567                                                           orig_bytes, flush,
1568                                                           "enospc");
1569                                 queue_work(system_unbound_wq, async_work);
1570                         }
1571                 } else {
1572                         list_add_tail(&ticket.list,
1573                                       &space_info->priority_tickets);
1574                 }
1575
1576                 /*
1577                  * We were forced to add a reserve ticket, so our preemptive
1578                  * flushing is unable to keep up.  Clamp down on the threshold
1579                  * for the preemptive flushing in order to keep up with the
1580                  * workload.
1581                  */
1582                 maybe_clamp_preempt(fs_info, space_info);
1583         } else if (!ret && space_info->flags & BTRFS_BLOCK_GROUP_METADATA) {
1584                 used += orig_bytes;
1585                 /*
1586                  * We will do the space reservation dance during log replay,
1587                  * which means we won't have fs_info->fs_root set, so don't do
1588                  * the async reclaim as we will panic.
1589                  */
1590                 if (!test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags) &&
1591                     need_preemptive_reclaim(fs_info, space_info) &&
1592                     !work_busy(&fs_info->preempt_reclaim_work)) {
1593                         trace_btrfs_trigger_flush(fs_info, space_info->flags,
1594                                                   orig_bytes, flush, "preempt");
1595                         queue_work(system_unbound_wq,
1596                                    &fs_info->preempt_reclaim_work);
1597                 }
1598         }
1599         spin_unlock(&space_info->lock);
1600         if (!ret || flush == BTRFS_RESERVE_NO_FLUSH)
1601                 return ret;
1602
1603         return handle_reserve_ticket(fs_info, space_info, &ticket, start_ns,
1604                                      orig_bytes, flush);
1605 }
1606
1607 /**
1608  * Trye to reserve metadata bytes from the block_rsv's space
1609  *
1610  * @root:       the root we're allocating for
1611  * @block_rsv:  block_rsv we're allocating for
1612  * @orig_bytes: number of bytes we want
1613  * @flush:      whether or not we can flush to make our reservation
1614  *
1615  * This will reserve orig_bytes number of bytes from the space info associated
1616  * with the block_rsv.  If there is not enough space it will make an attempt to
1617  * flush out space to make room.  It will do this by flushing delalloc if
1618  * possible or committing the transaction.  If flush is 0 then no attempts to
1619  * regain reservations will be made and this will fail if there is not enough
1620  * space already.
1621  */
1622 int btrfs_reserve_metadata_bytes(struct btrfs_root *root,
1623                                  struct btrfs_block_rsv *block_rsv,
1624                                  u64 orig_bytes,
1625                                  enum btrfs_reserve_flush_enum flush)
1626 {
1627         struct btrfs_fs_info *fs_info = root->fs_info;
1628         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
1629         int ret;
1630
1631         ret = __reserve_bytes(fs_info, block_rsv->space_info, orig_bytes, flush);
1632         if (ret == -ENOSPC &&
1633             unlikely(root->orphan_cleanup_state == ORPHAN_CLEANUP_STARTED)) {
1634                 if (block_rsv != global_rsv &&
1635                     !btrfs_block_rsv_use_bytes(global_rsv, orig_bytes))
1636                         ret = 0;
1637         }
1638         if (ret == -ENOSPC) {
1639                 trace_btrfs_space_reservation(fs_info, "space_info:enospc",
1640                                               block_rsv->space_info->flags,
1641                                               orig_bytes, 1);
1642
1643                 if (btrfs_test_opt(fs_info, ENOSPC_DEBUG))
1644                         btrfs_dump_space_info(fs_info, block_rsv->space_info,
1645                                               orig_bytes, 0);
1646         }
1647         return ret;
1648 }
1649
1650 /**
1651  * Try to reserve data bytes for an allocation
1652  *
1653  * @fs_info: the filesystem
1654  * @bytes:   number of bytes we need
1655  * @flush:   how we are allowed to flush
1656  *
1657  * This will reserve bytes from the data space info.  If there is not enough
1658  * space then we will attempt to flush space as specified by flush.
1659  */
1660 int btrfs_reserve_data_bytes(struct btrfs_fs_info *fs_info, u64 bytes,
1661                              enum btrfs_reserve_flush_enum flush)
1662 {
1663         struct btrfs_space_info *data_sinfo = fs_info->data_sinfo;
1664         int ret;
1665
1666         ASSERT(flush == BTRFS_RESERVE_FLUSH_DATA ||
1667                flush == BTRFS_RESERVE_FLUSH_FREE_SPACE_INODE);
1668         ASSERT(!current->journal_info || flush != BTRFS_RESERVE_FLUSH_DATA);
1669
1670         ret = __reserve_bytes(fs_info, data_sinfo, bytes, flush);
1671         if (ret == -ENOSPC) {
1672                 trace_btrfs_space_reservation(fs_info, "space_info:enospc",
1673                                               data_sinfo->flags, bytes, 1);
1674                 if (btrfs_test_opt(fs_info, ENOSPC_DEBUG))
1675                         btrfs_dump_space_info(fs_info, data_sinfo, bytes, 0);
1676         }
1677         return ret;
1678 }