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  *     This will commit the transaction.  Historically we had a lot of logic
 137  *     surrounding whether or not we'd commit the transaction, but this waits born
 138  *     out of a pre-tickets era where we could end up committing the transaction
 139  *     thousands of times in a row without making progress.  Now thanks to our
 140  *     ticketing system we know if we're not making progress and can error
 141  *     everybody out after a few commits rather than burning the disk hoping for
 142  *     a different answer.
 143  *
 144  * OVERCOMMIT
 145  *
 146  *   Because we hold so many reservations for metadata we will allow you to
 147  *   reserve more space than is currently free in the currently allocate
 148  *   metadata space.  This only happens with metadata, data does not allow
 149  *   overcommitting.
 150  *
 151  *   You can see the current logic for when we allow overcommit in
 152  *   btrfs_can_overcommit(), but it only applies to unallocated space.  If there
 153  *   is no unallocated space to be had, all reservations are kept within the
 154  *   free space in the allocated metadata chunks.
 155  *
 156  *   Because of overcommitting, you generally want to use the
 157  *   btrfs_can_overcommit() logic for metadata allocations, as it does the right
 158  *   thing with or without extra unallocated space.
 159  */
 160
 161 u64 __pure btrfs_space_info_used(struct btrfs_space_info *s_info,
 162                           bool may_use_included)
 163 {
 164         ASSERT(s_info);
 165         return s_info->bytes_used + s_info->bytes_reserved +
 166                 s_info->bytes_pinned + s_info->bytes_readonly +
 167                 s_info->bytes_zone_unusable +
 168                 (may_use_included ? s_info->bytes_may_use : 0);
 169 }
 170
 171 /*
 172  * after adding space to the filesystem, we need to clear the full flags
 173  * on all the space infos.
 174  */
 175 void btrfs_clear_space_info_full(struct btrfs_fs_info *info)
 176 {
 177         struct list_head *head = &info->space_info;
 178         struct btrfs_space_info *found;
 179
 180         list_for_each_entry(found, head, list)
 181                 found->full = 0;
 182 }
 183
 184 static int create_space_info(struct btrfs_fs_info *info, u64 flags)
 185 {
 186
 187         struct btrfs_space_info *space_info;
 188         int i;
 189         int ret;
 190
 191         space_info = kzalloc(sizeof(*space_info), GFP_NOFS);
 192         if (!space_info)
 193                 return -ENOMEM;
 194
 195         for (i = 0; i < BTRFS_NR_RAID_TYPES; i++)
 196                 INIT_LIST_HEAD(&space_info->block_groups[i]);
 197         init_rwsem(&space_info->groups_sem);
 198         spin_lock_init(&space_info->lock);
 199         space_info->flags = flags & BTRFS_BLOCK_GROUP_TYPE_MASK;
 200         space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
 201         INIT_LIST_HEAD(&space_info->ro_bgs);
 202         INIT_LIST_HEAD(&space_info->tickets);
 203         INIT_LIST_HEAD(&space_info->priority_tickets);
 204         space_info->clamp = 1;
 205
 206         ret = btrfs_sysfs_add_space_info_type(info, space_info);
 207         if (ret)
 208                 return ret;
 209
 210         list_add(&space_info->list, &info->space_info);
 211         if (flags & BTRFS_BLOCK_GROUP_DATA)
 212                 info->data_sinfo = space_info;
 213
 214         return ret;
 215 }
 216
 217 int btrfs_init_space_info(struct btrfs_fs_info *fs_info)
 218 {
 219         struct btrfs_super_block *disk_super;
 220         u64 features;
 221         u64 flags;
 222         int mixed = 0;
 223         int ret;
 224
 225         disk_super = fs_info->super_copy;
 226         if (!btrfs_super_root(disk_super))
 227                 return -EINVAL;
 228
 229         features = btrfs_super_incompat_flags(disk_super);
 230         if (features & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS)
 231                 mixed = 1;
 232
 233         flags = BTRFS_BLOCK_GROUP_SYSTEM;
 234         ret = create_space_info(fs_info, flags);
 235         if (ret)
 236                 goto out;
 237
 238         if (mixed) {
 239                 flags = BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DATA;
 240                 ret = create_space_info(fs_info, flags);
 241         } else {
 242                 flags = BTRFS_BLOCK_GROUP_METADATA;
 243                 ret = create_space_info(fs_info, flags);
 244                 if (ret)
 245                         goto out;
 246
 247                 flags = BTRFS_BLOCK_GROUP_DATA;
 248                 ret = create_space_info(fs_info, flags);
 249         }
 250 out:
 251         return ret;
 252 }
 253
 254 void btrfs_update_space_info(struct btrfs_fs_info *info, u64 flags,
 255                              u64 total_bytes, u64 bytes_used,
 256                              u64 bytes_readonly, u64 bytes_zone_unusable,
 257                              struct btrfs_space_info **space_info)
 258 {
 259         struct btrfs_space_info *found;
 260         int factor;
 261
 262         factor = btrfs_bg_type_to_factor(flags);
 263
 264         found = btrfs_find_space_info(info, flags);
 265         ASSERT(found);
 266         spin_lock(&found->lock);
 267         found->total_bytes += total_bytes;
 268         found->disk_total += total_bytes * factor;
 269         found->bytes_used += bytes_used;
 270         found->disk_used += bytes_used * factor;
 271         found->bytes_readonly += bytes_readonly;
 272         found->bytes_zone_unusable += bytes_zone_unusable;
 273         if (total_bytes > 0)
 274                 found->full = 0;
 275         btrfs_try_granting_tickets(info, found);
 276         spin_unlock(&found->lock);
 277         *space_info = found;
 278 }
 279
 280 struct btrfs_space_info *btrfs_find_space_info(struct btrfs_fs_info *info,
 281                                                u64 flags)
 282 {
 283         struct list_head *head = &info->space_info;
 284         struct btrfs_space_info *found;
 285
 286         flags &= BTRFS_BLOCK_GROUP_TYPE_MASK;
 287
 288         list_for_each_entry(found, head, list) {
 289                 if (found->flags & flags)
 290                         return found;
 291         }
 292         return NULL;
 293 }
 294
 295 static u64 calc_available_free_space(struct btrfs_fs_info *fs_info,
 296                           struct btrfs_space_info *space_info,
 297                           enum btrfs_reserve_flush_enum flush)
 298 {
 299         u64 profile;
 300         u64 avail;
 301         int factor;
 302
 303         if (space_info->flags & BTRFS_BLOCK_GROUP_SYSTEM)
 304                 profile = btrfs_system_alloc_profile(fs_info);
 305         else
 306                 profile = btrfs_metadata_alloc_profile(fs_info);
 307
 308         avail = atomic64_read(&fs_info->free_chunk_space);
 309
 310         /*
 311          * If we have dup, raid1 or raid10 then only half of the free
 312          * space is actually usable.  For raid56, the space info used
 313          * doesn't include the parity drive, so we don't have to
 314          * change the math
 315          */
 316         factor = btrfs_bg_type_to_factor(profile);
 317         avail = div_u64(avail, factor);
 318
 319         /*
 320          * If we aren't flushing all things, let us overcommit up to
 321          * 1/2th of the space. If we can flush, don't let us overcommit
 322          * too much, let it overcommit up to 1/8 of the space.
 323          */
 324         if (flush == BTRFS_RESERVE_FLUSH_ALL)
 325                 avail >>= 3;
 326         else
 327                 avail >>= 1;
 328         return avail;
 329 }
 330
 331 int btrfs_can_overcommit(struct btrfs_fs_info *fs_info,
 332                          struct btrfs_space_info *space_info, u64 bytes,
 333                          enum btrfs_reserve_flush_enum flush)
 334 {
 335         u64 avail;
 336         u64 used;
 337
 338         /* Don't overcommit when in mixed mode */
 339         if (space_info->flags & BTRFS_BLOCK_GROUP_DATA)
 340                 return 0;
 341
 342         used = btrfs_space_info_used(space_info, true);
 343         avail = calc_available_free_space(fs_info, space_info, flush);
 344
 345         if (used + bytes < space_info->total_bytes + avail)
 346                 return 1;
 347         return 0;
 348 }
 349
 350 static void remove_ticket(struct btrfs_space_info *space_info,
 351                           struct reserve_ticket *ticket)
 352 {
 353         if (!list_empty(&ticket->list)) {
 354                 list_del_init(&ticket->list);
 355                 ASSERT(space_info->reclaim_size >= ticket->bytes);
 356                 space_info->reclaim_size -= ticket->bytes;
 357         }
 358 }
 359
 360 /*
 361  * This is for space we already have accounted in space_info->bytes_may_use, so
 362  * basically when we're returning space from block_rsv's.
 363  */
 364 void btrfs_try_granting_tickets(struct btrfs_fs_info *fs_info,
 365                                 struct btrfs_space_info *space_info)
 366 {
 367         struct list_head *head;
 368         enum btrfs_reserve_flush_enum flush = BTRFS_RESERVE_NO_FLUSH;
 369
 370         lockdep_assert_held(&space_info->lock);
 371
 372         head = &space_info->priority_tickets;
 373 again:
 374         while (!list_empty(head)) {
 375                 struct reserve_ticket *ticket;
 376                 u64 used = btrfs_space_info_used(space_info, true);
 377
 378                 ticket = list_first_entry(head, struct reserve_ticket, list);
 379
 380                 /* Check and see if our ticket can be satisfied now. */
 381                 if ((used + ticket->bytes <= space_info->total_bytes) ||
 382                     btrfs_can_overcommit(fs_info, space_info, ticket->bytes,
 383                                          flush)) {
 384                         btrfs_space_info_update_bytes_may_use(fs_info,
 385                                                               space_info,
 386                                                               ticket->bytes);
 387                         remove_ticket(space_info, ticket);
 388                         ticket->bytes = 0;
 389                         space_info->tickets_id++;
 390                         wake_up(&ticket->wait);
 391                 } else {
 392                         break;
 393                 }
 394         }
 395
 396         if (head == &space_info->priority_tickets) {
 397                 head = &space_info->tickets;
 398                 flush = BTRFS_RESERVE_FLUSH_ALL;
 399                 goto again;
 400         }
 401 }
 402
 403 #define DUMP_BLOCK_RSV(fs_info, rsv_name)                               \
 404 do {                                                                    \
 405         struct btrfs_block_rsv *__rsv = &(fs_info)->rsv_name;           \
 406         spin_lock(&__rsv->lock);                                        \
 407         btrfs_info(fs_info, #rsv_name ": size %llu reserved %llu",      \
 408                    __rsv->size, __rsv->reserved);                       \
 409         spin_unlock(&__rsv->lock);                                      \
 410 } while (0)
 411
 412 static void __btrfs_dump_space_info(struct btrfs_fs_info *fs_info,
 413                                     struct btrfs_space_info *info)
 414 {
 415         lockdep_assert_held(&info->lock);
 416
 417         /* The free space could be negative in case of overcommit */
 418         btrfs_info(fs_info, "space_info %llu has %lld free, is %sfull",
 419                    info->flags,
 420                    (s64)(info->total_bytes - btrfs_space_info_used(info, true)),
 421                    info->full ? "" : "not ");
 422         btrfs_info(fs_info,
 423                 "space_info total=%llu, used=%llu, pinned=%llu, reserved=%llu, may_use=%llu, readonly=%llu zone_unusable=%llu",
 424                 info->total_bytes, info->bytes_used, info->bytes_pinned,
 425                 info->bytes_reserved, info->bytes_may_use,
 426                 info->bytes_readonly, info->bytes_zone_unusable);
 427
 428         DUMP_BLOCK_RSV(fs_info, global_block_rsv);
 429         DUMP_BLOCK_RSV(fs_info, trans_block_rsv);
 430         DUMP_BLOCK_RSV(fs_info, chunk_block_rsv);
 431         DUMP_BLOCK_RSV(fs_info, delayed_block_rsv);
 432         DUMP_BLOCK_RSV(fs_info, delayed_refs_rsv);
 433
 434 }
 435
 436 void btrfs_dump_space_info(struct btrfs_fs_info *fs_info,
 437                            struct btrfs_space_info *info, u64 bytes,
 438                            int dump_block_groups)
 439 {
 440         struct btrfs_block_group *cache;
 441         int index = 0;
 442
 443         spin_lock(&info->lock);
 444         __btrfs_dump_space_info(fs_info, info);
 445         spin_unlock(&info->lock);
 446
 447         if (!dump_block_groups)
 448                 return;
 449
 450         down_read(&info->groups_sem);
 451 again:
 452         list_for_each_entry(cache, &info->block_groups[index], list) {
 453                 spin_lock(&cache->lock);
 454                 btrfs_info(fs_info,
 455                         "block group %llu has %llu bytes, %llu used %llu pinned %llu reserved %llu zone_unusable %s",
 456                         cache->start, cache->length, cache->used, cache->pinned,
 457                         cache->reserved, cache->zone_unusable,
 458                         cache->ro ? "[readonly]" : "");
 459                 spin_unlock(&cache->lock);
 460                 btrfs_dump_free_space(cache, bytes);
 461         }
 462         if (++index < BTRFS_NR_RAID_TYPES)
 463                 goto again;
 464         up_read(&info->groups_sem);
 465 }
 466
 467 static inline u64 calc_reclaim_items_nr(struct btrfs_fs_info *fs_info,
 468                                         u64 to_reclaim)
 469 {
 470         u64 bytes;
 471         u64 nr;
 472
 473         bytes = btrfs_calc_insert_metadata_size(fs_info, 1);
 474         nr = div64_u64(to_reclaim, bytes);
 475         if (!nr)
 476                 nr = 1;
 477         return nr;
 478 }
 479
 480 #define EXTENT_SIZE_PER_ITEM    SZ_256K
 481
 482 /*
 483  * shrink metadata reservation for delalloc
 484  */
 485 static void shrink_delalloc(struct btrfs_fs_info *fs_info,
 486                             struct btrfs_space_info *space_info,
 487                             u64 to_reclaim, bool wait_ordered,
 488                             bool for_preempt)
 489 {
 490         struct btrfs_trans_handle *trans;
 491         u64 delalloc_bytes;
 492         u64 ordered_bytes;
 493         u64 items;
 494         long time_left;
 495         int loops;
 496
 497         delalloc_bytes = percpu_counter_sum_positive(&fs_info->delalloc_bytes);
 498         ordered_bytes = percpu_counter_sum_positive(&fs_info->ordered_bytes);
 499         if (delalloc_bytes == 0 && ordered_bytes == 0)
 500                 return;
 501
 502         /* Calc the number of the pages we need flush for space reservation */
 503         if (to_reclaim == U64_MAX) {
 504                 items = U64_MAX;
 505         } else {
 506                 /*
 507                  * to_reclaim is set to however much metadata we need to
 508                  * reclaim, but reclaiming that much data doesn't really track
 509                  * exactly.  What we really want to do is reclaim full inode's
 510                  * worth of reservations, however that's not available to us
 511                  * here.  We will take a fraction of the delalloc bytes for our
 512                  * flushing loops and hope for the best.  Delalloc will expand
 513                  * the amount we write to cover an entire dirty extent, which
 514                  * will reclaim the metadata reservation for that range.  If
 515                  * it's not enough subsequent flush stages will be more
 516                  * aggressive.
 517                  */
 518                 to_reclaim = max(to_reclaim, delalloc_bytes >> 3);
 519                 items = calc_reclaim_items_nr(fs_info, to_reclaim) * 2;
 520         }
 521
 522         trans = (struct btrfs_trans_handle *)current->journal_info;
 523
 524         /*
 525          * If we are doing more ordered than delalloc we need to just wait on
 526          * ordered extents, otherwise we'll waste time trying to flush delalloc
 527          * that likely won't give us the space back we need.
 528          */
 529         if (ordered_bytes > delalloc_bytes && !for_preempt)
 530                 wait_ordered = true;
 531
 532         loops = 0;
 533         while ((delalloc_bytes || ordered_bytes) && loops < 3) {
 534                 u64 temp = min(delalloc_bytes, to_reclaim) >> PAGE_SHIFT;
 535                 long nr_pages = min_t(u64, temp, LONG_MAX);
 536                 int async_pages;
 537
 538                 btrfs_start_delalloc_roots(fs_info, nr_pages, true);
 539
 540                 /*
 541                  * We need to make sure any outstanding async pages are now
 542                  * processed before we continue.  This is because things like
 543                  * sync_inode() try to be smart and skip writing if the inode is
 544                  * marked clean.  We don't use filemap_fwrite for flushing
 545                  * because we want to control how many pages we write out at a
 546                  * time, thus this is the only safe way to make sure we've
 547                  * waited for outstanding compressed workers to have started
 548                  * their jobs and thus have ordered extents set up properly.
 549                  *
 550                  * This exists because we do not want to wait for each
 551                  * individual inode to finish its async work, we simply want to
 552                  * start the IO on everybody, and then come back here and wait
 553                  * for all of the async work to catch up.  Once we're done with
 554                  * that we know we'll have ordered extents for everything and we
 555                  * can decide if we wait for that or not.
 556                  *
 557                  * If we choose to replace this in the future, make absolutely
 558                  * sure that the proper waiting is being done in the async case,
 559                  * as there have been bugs in that area before.
 560                  */
 561                 async_pages = atomic_read(&fs_info->async_delalloc_pages);
 562                 if (!async_pages)
 563                         goto skip_async;
 564
 565                 /*
 566                  * We don't want to wait forever, if we wrote less pages in this
 567                  * loop than we have outstanding, only wait for that number of
 568                  * pages, otherwise we can wait for all async pages to finish
 569                  * before continuing.
 570                  */
 571                 if (async_pages > nr_pages)
 572                         async_pages -= nr_pages;
 573                 else
 574                         async_pages = 0;
 575                 wait_event(fs_info->async_submit_wait,
 576                            atomic_read(&fs_info->async_delalloc_pages) <=
 577                            async_pages);
 578 skip_async:
 579                 loops++;
 580                 if (wait_ordered && !trans) {
 581                         btrfs_wait_ordered_roots(fs_info, items, 0, (u64)-1);
 582                 } else {
 583                         time_left = schedule_timeout_killable(1);
 584                         if (time_left)
 585                                 break;
 586                 }
 587
 588                 /*
 589                  * If we are for preemption we just want a one-shot of delalloc
 590                  * flushing so we can stop flushing if we decide we don't need
 591                  * to anymore.
 592                  */
 593                 if (for_preempt)
 594                         break;
 595
 596                 spin_lock(&space_info->lock);
 597                 if (list_empty(&space_info->tickets) &&
 598                     list_empty(&space_info->priority_tickets)) {
 599                         spin_unlock(&space_info->lock);
 600                         break;
 601                 }
 602                 spin_unlock(&space_info->lock);
 603
 604                 delalloc_bytes = percpu_counter_sum_positive(
 605                                                 &fs_info->delalloc_bytes);
 606                 ordered_bytes = percpu_counter_sum_positive(
 607                                                 &fs_info->ordered_bytes);
 608         }
 609 }
 610
 611 /*
 612  * Try to flush some data based on policy set by @state. This is only advisory
 613  * and may fail for various reasons. The caller is supposed to examine the
 614  * state of @space_info to detect the outcome.
 615  */
 616 static void flush_space(struct btrfs_fs_info *fs_info,
 617                        struct btrfs_space_info *space_info, u64 num_bytes,
 618                        enum btrfs_flush_state state, bool for_preempt)
 619 {
 620         struct btrfs_root *root = fs_info->tree_root;
 621         struct btrfs_trans_handle *trans;
 622         int nr;
 623         int ret = 0;
 624
 625         switch (state) {
 626         case FLUSH_DELAYED_ITEMS_NR:
 627         case FLUSH_DELAYED_ITEMS:
 628                 if (state == FLUSH_DELAYED_ITEMS_NR)
 629                         nr = calc_reclaim_items_nr(fs_info, num_bytes) * 2;
 630                 else
 631                         nr = -1;
 632
 633                 trans = btrfs_join_transaction(root);
 634                 if (IS_ERR(trans)) {
 635                         ret = PTR_ERR(trans);
 636                         break;
 637                 }
 638                 ret = btrfs_run_delayed_items_nr(trans, nr);
 639                 btrfs_end_transaction(trans);
 640                 break;
 641         case FLUSH_DELALLOC:
 642         case FLUSH_DELALLOC_WAIT:
 643         case FLUSH_DELALLOC_FULL:
 644                 if (state == FLUSH_DELALLOC_FULL)
 645                         num_bytes = U64_MAX;
 646                 shrink_delalloc(fs_info, space_info, num_bytes,
 647                                 state != FLUSH_DELALLOC, for_preempt);
 648                 break;
 649         case FLUSH_DELAYED_REFS_NR:
 650         case FLUSH_DELAYED_REFS:
 651                 trans = btrfs_join_transaction(root);
 652                 if (IS_ERR(trans)) {
 653                         ret = PTR_ERR(trans);
 654                         break;
 655                 }
 656                 if (state == FLUSH_DELAYED_REFS_NR)
 657                         nr = calc_reclaim_items_nr(fs_info, num_bytes);
 658                 else
 659                         nr = 0;
 660                 btrfs_run_delayed_refs(trans, nr);
 661                 btrfs_end_transaction(trans);
 662                 break;
 663         case ALLOC_CHUNK:
 664         case ALLOC_CHUNK_FORCE:
 665                 trans = btrfs_join_transaction(root);
 666                 if (IS_ERR(trans)) {
 667                         ret = PTR_ERR(trans);
 668                         break;
 669                 }
 670                 ret = btrfs_chunk_alloc(trans,
 671                                 btrfs_get_alloc_profile(fs_info, space_info->flags),
 672                                 (state == ALLOC_CHUNK) ? CHUNK_ALLOC_NO_FORCE :
 673                                         CHUNK_ALLOC_FORCE);
 674                 btrfs_end_transaction(trans);
 675                 if (ret > 0 || ret == -ENOSPC)
 676                         ret = 0;
 677                 break;
 678         case RUN_DELAYED_IPUTS:
 679                 /*
 680                  * If we have pending delayed iputs then we could free up a
 681                  * bunch of pinned space, so make sure we run the iputs before
 682                  * we do our pinned bytes check below.
 683                  */
 684                 btrfs_run_delayed_iputs(fs_info);
 685                 btrfs_wait_on_delayed_iputs(fs_info);
 686                 break;
 687         case COMMIT_TRANS:
 688                 ASSERT(current->journal_info == NULL);
 689                 trans = btrfs_join_transaction(root);
 690                 if (IS_ERR(trans)) {
 691                         ret = PTR_ERR(trans);
 692                         break;
 693                 }
 694                 ret = btrfs_commit_transaction(trans);
 695                 break;
 696         default:
 697                 ret = -ENOSPC;
 698                 break;
 699         }
 700
 701         trace_btrfs_flush_space(fs_info, space_info->flags, num_bytes, state,
 702                                 ret, for_preempt);
 703         return;
 704 }
 705
 706 static inline u64
 707 btrfs_calc_reclaim_metadata_size(struct btrfs_fs_info *fs_info,
 708                                  struct btrfs_space_info *space_info)
 709 {
 710         u64 used;
 711         u64 avail;
 712         u64 to_reclaim = space_info->reclaim_size;
 713
 714         lockdep_assert_held(&space_info->lock);
 715
 716         avail = calc_available_free_space(fs_info, space_info,
 717                                           BTRFS_RESERVE_FLUSH_ALL);
 718         used = btrfs_space_info_used(space_info, true);
 719
 720         /*
 721          * We may be flushing because suddenly we have less space than we had
 722          * before, and now we're well over-committed based on our current free
 723          * space.  If that's the case add in our overage so we make sure to put
 724          * appropriate pressure on the flushing state machine.
 725          */
 726         if (space_info->total_bytes + avail < used)
 727                 to_reclaim += used - (space_info->total_bytes + avail);
 728
 729         return to_reclaim;
 730 }
 731
 732 static bool need_preemptive_reclaim(struct btrfs_fs_info *fs_info,
 733                                     struct btrfs_space_info *space_info)
 734 {
 735         u64 global_rsv_size = fs_info->global_block_rsv.reserved;
 736         u64 ordered, delalloc;
 737         u64 thresh = div_factor_fine(space_info->total_bytes, 90);
 738         u64 used;
 739
 740         lockdep_assert_held(&space_info->lock);
 741
 742         /* If we're just plain full then async reclaim just slows us down. */
 743         if ((space_info->bytes_used + space_info->bytes_reserved +
 744              global_rsv_size) >= thresh)
 745                 return false;
 746
 747         used = space_info->bytes_may_use + space_info->bytes_pinned;
 748
 749         /* The total flushable belongs to the global rsv, don't flush. */
 750         if (global_rsv_size >= used)
 751                 return false;
 752
 753         /*
 754          * 128MiB is 1/4 of the maximum global rsv size.  If we have less than
 755          * that devoted to other reservations then there's no sense in flushing,
 756          * we don't have a lot of things that need flushing.
 757          */
 758         if (used - global_rsv_size <= SZ_128M)
 759                 return false;
 760
 761         /*
 762          * We have tickets queued, bail so we don't compete with the async
 763          * flushers.
 764          */
 765         if (space_info->reclaim_size)
 766                 return false;
 767
 768         /*
 769          * If we have over half of the free space occupied by reservations or
 770          * pinned then we want to start flushing.
 771          *
 772          * We do not do the traditional thing here, which is to say
 773          *
 774          *   if (used >= ((total_bytes + avail) / 2))
 775          *     return 1;
 776          *
 777          * because this doesn't quite work how we want.  If we had more than 50%
 778          * of the space_info used by bytes_used and we had 0 available we'd just
 779          * constantly run the background flusher.  Instead we want it to kick in
 780          * if our reclaimable space exceeds our clamped free space.
 781          *
 782          * Our clamping range is 2^1 -> 2^8.  Practically speaking that means
 783          * the following:
 784          *
 785          * Amount of RAM        Minimum threshold       Maximum threshold
 786          *
 787          *        256GiB                     1GiB                  128GiB
 788          *        128GiB                   512MiB                   64GiB
 789          *         64GiB                   256MiB                   32GiB
 790          *         32GiB                   128MiB                   16GiB
 791          *         16GiB                    64MiB                    8GiB
 792          *
 793          * These are the range our thresholds will fall in, corresponding to how
 794          * much delalloc we need for the background flusher to kick in.
 795          */
 796
 797         thresh = calc_available_free_space(fs_info, space_info,
 798                                            BTRFS_RESERVE_FLUSH_ALL);
 799         used = space_info->bytes_used + space_info->bytes_reserved +
 800                space_info->bytes_readonly + global_rsv_size;
 801         if (used < space_info->total_bytes)
 802                 thresh += space_info->total_bytes - used;
 803         thresh >>= space_info->clamp;
 804
 805         used = space_info->bytes_pinned;
 806
 807         /*
 808          * If we have more ordered bytes than delalloc bytes then we're either
 809          * doing a lot of DIO, or we simply don't have a lot of delalloc waiting
 810          * around.  Preemptive flushing is only useful in that it can free up
 811          * space before tickets need to wait for things to finish.  In the case
 812          * of ordered extents, preemptively waiting on ordered extents gets us
 813          * nothing, if our reservations are tied up in ordered extents we'll
 814          * simply have to slow down writers by forcing them to wait on ordered
 815          * extents.
 816          *
 817          * In the case that ordered is larger than delalloc, only include the
 818          * block reserves that we would actually be able to directly reclaim
 819          * from.  In this case if we're heavy on metadata operations this will
 820          * clearly be heavy enough to warrant preemptive flushing.  In the case
 821          * of heavy DIO or ordered reservations, preemptive flushing will just
 822          * waste time and cause us to slow down.
 823          *
 824          * We want to make sure we truly are maxed out on ordered however, so
 825          * cut ordered in half, and if it's still higher than delalloc then we
 826          * can keep flushing.  This is to avoid the case where we start
 827          * flushing, and now delalloc == ordered and we stop preemptively
 828          * flushing when we could still have several gigs of delalloc to flush.
 829          */
 830         ordered = percpu_counter_read_positive(&fs_info->ordered_bytes) >> 1;
 831         delalloc = percpu_counter_read_positive(&fs_info->delalloc_bytes);
 832         if (ordered >= delalloc)
 833                 used += fs_info->delayed_refs_rsv.reserved +
 834                         fs_info->delayed_block_rsv.reserved;
 835         else
 836                 used += space_info->bytes_may_use - global_rsv_size;
 837
 838         return (used >= thresh && !btrfs_fs_closing(fs_info) &&
 839                 !test_bit(BTRFS_FS_STATE_REMOUNTING, &fs_info->fs_state));
 840 }
 841
 842 static bool steal_from_global_rsv(struct btrfs_fs_info *fs_info,
 843                                   struct btrfs_space_info *space_info,
 844                                   struct reserve_ticket *ticket)
 845 {
 846         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
 847         u64 min_bytes;
 848
 849         if (!ticket->steal)
 850                 return false;
 851
 852         if (global_rsv->space_info != space_info)
 853                 return false;
 854
 855         spin_lock(&global_rsv->lock);
 856         min_bytes = div_factor(global_rsv->size, 1);
 857         if (global_rsv->reserved < min_bytes + ticket->bytes) {
 858                 spin_unlock(&global_rsv->lock);
 859                 return false;
 860         }
 861         global_rsv->reserved -= ticket->bytes;
 862         remove_ticket(space_info, ticket);
 863         ticket->bytes = 0;
 864         wake_up(&ticket->wait);
 865         space_info->tickets_id++;
 866         if (global_rsv->reserved < global_rsv->size)
 867                 global_rsv->full = 0;
 868         spin_unlock(&global_rsv->lock);
 869
 870         return true;
 871 }
 872
 873 /*
 874  * maybe_fail_all_tickets - we've exhausted our flushing, start failing tickets
 875  * @fs_info - fs_info for this fs
 876  * @space_info - the space info we were flushing
 877  *
 878  * We call this when we've exhausted our flushing ability and haven't made
 879  * progress in satisfying tickets.  The reservation code handles tickets in
 880  * order, so if there is a large ticket first and then smaller ones we could
 881  * very well satisfy the smaller tickets.  This will attempt to wake up any
 882  * tickets in the list to catch this case.
 883  *
 884  * This function returns true if it was able to make progress by clearing out
 885  * other tickets, or if it stumbles across a ticket that was smaller than the
 886  * first ticket.
 887  */
 888 static bool maybe_fail_all_tickets(struct btrfs_fs_info *fs_info,
 889                                    struct btrfs_space_info *space_info)
 890 {
 891         struct reserve_ticket *ticket;
 892         u64 tickets_id = space_info->tickets_id;
 893         const bool aborted = BTRFS_FS_ERROR(fs_info);
 894
 895         trace_btrfs_fail_all_tickets(fs_info, space_info);
 896
 897         if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
 898                 btrfs_info(fs_info, "cannot satisfy tickets, dumping space info");
 899                 __btrfs_dump_space_info(fs_info, space_info);
 900         }
 901
 902         while (!list_empty(&space_info->tickets) &&
 903                tickets_id == space_info->tickets_id) {
 904                 ticket = list_first_entry(&space_info->tickets,
 905                                           struct reserve_ticket, list);
 906
 907                 if (!aborted && steal_from_global_rsv(fs_info, space_info, ticket))
 908                         return true;
 909
 910                 if (!aborted && btrfs_test_opt(fs_info, ENOSPC_DEBUG))
 911                         btrfs_info(fs_info, "failing ticket with %llu bytes",
 912                                    ticket->bytes);
 913
 914                 remove_ticket(space_info, ticket);
 915                 if (aborted)
 916                         ticket->error = -EIO;
 917                 else
 918                         ticket->error = -ENOSPC;
 919                 wake_up(&ticket->wait);
 920
 921                 /*
 922                  * We're just throwing tickets away, so more flushing may not
 923                  * trip over btrfs_try_granting_tickets, so we need to call it
 924                  * here to see if we can make progress with the next ticket in
 925                  * the list.
 926                  */
 927                 if (!aborted)
 928                         btrfs_try_granting_tickets(fs_info, space_info);
 929         }
 930         return (tickets_id != space_info->tickets_id);
 931 }
 932
 933 /*
 934  * This is for normal flushers, we can wait all goddamned day if we want to.  We
 935  * will loop and continuously try to flush as long as we are making progress.
 936  * We count progress as clearing off tickets each time we have to loop.
 937  */
 938 static void btrfs_async_reclaim_metadata_space(struct work_struct *work)
 939 {
 940         struct btrfs_fs_info *fs_info;
 941         struct btrfs_space_info *space_info;
 942         u64 to_reclaim;
 943         enum btrfs_flush_state flush_state;
 944         int commit_cycles = 0;
 945         u64 last_tickets_id;
 946
 947         fs_info = container_of(work, struct btrfs_fs_info, async_reclaim_work);
 948         space_info = btrfs_find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
 949
 950         spin_lock(&space_info->lock);
 951         to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info, space_info);
 952         if (!to_reclaim) {
 953                 space_info->flush = 0;
 954                 spin_unlock(&space_info->lock);
 955                 return;
 956         }
 957         last_tickets_id = space_info->tickets_id;
 958         spin_unlock(&space_info->lock);
 959
 960         flush_state = FLUSH_DELAYED_ITEMS_NR;
 961         do {
 962                 flush_space(fs_info, space_info, to_reclaim, flush_state, false);
 963                 spin_lock(&space_info->lock);
 964                 if (list_empty(&space_info->tickets)) {
 965                         space_info->flush = 0;
 966                         spin_unlock(&space_info->lock);
 967                         return;
 968                 }
 969                 to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info,
 970                                                               space_info);
 971                 if (last_tickets_id == space_info->tickets_id) {
 972                         flush_state++;
 973                 } else {
 974                         last_tickets_id = space_info->tickets_id;
 975                         flush_state = FLUSH_DELAYED_ITEMS_NR;
 976                         if (commit_cycles)
 977                                 commit_cycles--;
 978                 }
 979
 980                 /*
 981                  * We do not want to empty the system of delalloc unless we're
 982                  * under heavy pressure, so allow one trip through the flushing
 983                  * logic before we start doing a FLUSH_DELALLOC_FULL.
 984                  */
 985                 if (flush_state == FLUSH_DELALLOC_FULL && !commit_cycles)
 986                         flush_state++;
 987
 988                 /*
 989                  * We don't want to force a chunk allocation until we've tried
 990                  * pretty hard to reclaim space.  Think of the case where we
 991                  * freed up a bunch of space and so have a lot of pinned space
 992                  * to reclaim.  We would rather use that than possibly create a
 993                  * underutilized metadata chunk.  So if this is our first run
 994                  * through the flushing state machine skip ALLOC_CHUNK_FORCE and
 995                  * commit the transaction.  If nothing has changed the next go
 996                  * around then we can force a chunk allocation.
 997                  */
 998                 if (flush_state == ALLOC_CHUNK_FORCE && !commit_cycles)
 999                         flush_state++;
1000
1001                 if (flush_state > COMMIT_TRANS) {
1002                         commit_cycles++;
1003                         if (commit_cycles > 2) {
1004                                 if (maybe_fail_all_tickets(fs_info, space_info)) {
1005                                         flush_state = FLUSH_DELAYED_ITEMS_NR;
1006                                         commit_cycles--;
1007                                 } else {
1008                                         space_info->flush = 0;
1009                                 }
1010                         } else {
1011                                 flush_state = FLUSH_DELAYED_ITEMS_NR;
1012                         }
1013                 }
1014                 spin_unlock(&space_info->lock);
1015         } while (flush_state <= COMMIT_TRANS);
1016 }
1017
1018 /*
1019  * This handles pre-flushing of metadata space before we get to the point that
1020  * we need to start blocking threads on tickets.  The logic here is different
1021  * from the other flush paths because it doesn't rely on tickets to tell us how
1022  * much we need to flush, instead it attempts to keep us below the 80% full
1023  * watermark of space by flushing whichever reservation pool is currently the
1024  * largest.
1025  */
1026 static void btrfs_preempt_reclaim_metadata_space(struct work_struct *work)
1027 {
1028         struct btrfs_fs_info *fs_info;
1029         struct btrfs_space_info *space_info;
1030         struct btrfs_block_rsv *delayed_block_rsv;
1031         struct btrfs_block_rsv *delayed_refs_rsv;
1032         struct btrfs_block_rsv *global_rsv;
1033         struct btrfs_block_rsv *trans_rsv;
1034         int loops = 0;
1035
1036         fs_info = container_of(work, struct btrfs_fs_info,
1037                                preempt_reclaim_work);
1038         space_info = btrfs_find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
1039         delayed_block_rsv = &fs_info->delayed_block_rsv;
1040         delayed_refs_rsv = &fs_info->delayed_refs_rsv;
1041         global_rsv = &fs_info->global_block_rsv;
1042         trans_rsv = &fs_info->trans_block_rsv;
1043
1044         spin_lock(&space_info->lock);
1045         while (need_preemptive_reclaim(fs_info, space_info)) {
1046                 enum btrfs_flush_state flush;
1047                 u64 delalloc_size = 0;
1048                 u64 to_reclaim, block_rsv_size;
1049                 u64 global_rsv_size = global_rsv->reserved;
1050
1051                 loops++;
1052
1053                 /*
1054                  * We don't have a precise counter for the metadata being
1055                  * reserved for delalloc, so we'll approximate it by subtracting
1056                  * out the block rsv's space from the bytes_may_use.  If that
1057                  * amount is higher than the individual reserves, then we can
1058                  * assume it's tied up in delalloc reservations.
1059                  */
1060                 block_rsv_size = global_rsv_size +
1061                         delayed_block_rsv->reserved +
1062                         delayed_refs_rsv->reserved +
1063                         trans_rsv->reserved;
1064                 if (block_rsv_size < space_info->bytes_may_use)
1065                         delalloc_size = space_info->bytes_may_use - block_rsv_size;
1066
1067                 /*
1068                  * We don't want to include the global_rsv in our calculation,
1069                  * because that's space we can't touch.  Subtract it from the
1070                  * block_rsv_size for the next checks.
1071                  */
1072                 block_rsv_size -= global_rsv_size;
1073
1074                 /*
1075                  * We really want to avoid flushing delalloc too much, as it
1076                  * could result in poor allocation patterns, so only flush it if
1077                  * it's larger than the rest of the pools combined.
1078                  */
1079                 if (delalloc_size > block_rsv_size) {
1080                         to_reclaim = delalloc_size;
1081                         flush = FLUSH_DELALLOC;
1082                 } else if (space_info->bytes_pinned >
1083                            (delayed_block_rsv->reserved +
1084                             delayed_refs_rsv->reserved)) {
1085                         to_reclaim = space_info->bytes_pinned;
1086                         flush = COMMIT_TRANS;
1087                 } else if (delayed_block_rsv->reserved >
1088                            delayed_refs_rsv->reserved) {
1089                         to_reclaim = delayed_block_rsv->reserved;
1090                         flush = FLUSH_DELAYED_ITEMS_NR;
1091                 } else {
1092                         to_reclaim = delayed_refs_rsv->reserved;
1093                         flush = FLUSH_DELAYED_REFS_NR;
1094                 }
1095
1096                 spin_unlock(&space_info->lock);
1097
1098                 /*
1099                  * We don't want to reclaim everything, just a portion, so scale
1100                  * down the to_reclaim by 1/4.  If it takes us down to 0,
1101                  * reclaim 1 items worth.
1102                  */
1103                 to_reclaim >>= 2;
1104                 if (!to_reclaim)
1105                         to_reclaim = btrfs_calc_insert_metadata_size(fs_info, 1);
1106                 flush_space(fs_info, space_info, to_reclaim, flush, true);
1107                 cond_resched();
1108                 spin_lock(&space_info->lock);
1109         }
1110
1111         /* We only went through once, back off our clamping. */
1112         if (loops == 1 && !space_info->reclaim_size)
1113                 space_info->clamp = max(1, space_info->clamp - 1);
1114         trace_btrfs_done_preemptive_reclaim(fs_info, space_info);
1115         spin_unlock(&space_info->lock);
1116 }
1117
1118 /*
1119  * FLUSH_DELALLOC_WAIT:
1120  *   Space is freed from flushing delalloc in one of two ways.
1121  *
1122  *   1) compression is on and we allocate less space than we reserved
1123  *   2) we are overwriting existing space
1124  *
1125  *   For #1 that extra space is reclaimed as soon as the delalloc pages are
1126  *   COWed, by way of btrfs_add_reserved_bytes() which adds the actual extent
1127  *   length to ->bytes_reserved, and subtracts the reserved space from
1128  *   ->bytes_may_use.
1129  *
1130  *   For #2 this is trickier.  Once the ordered extent runs we will drop the
1131  *   extent in the range we are overwriting, which creates a delayed ref for
1132  *   that freed extent.  This however is not reclaimed until the transaction
1133  *   commits, thus the next stages.
1134  *
1135  * RUN_DELAYED_IPUTS
1136  *   If we are freeing inodes, we want to make sure all delayed iputs have
1137  *   completed, because they could have been on an inode with i_nlink == 0, and
1138  *   thus have been truncated and freed up space.  But again this space is not
1139  *   immediately re-usable, it comes in the form of a delayed ref, which must be
1140  *   run and then the transaction must be committed.
1141  *
1142  * COMMIT_TRANS
1143  *   This is where we reclaim all of the pinned space generated by running the
1144  *   iputs
1145  *
1146  * ALLOC_CHUNK_FORCE
1147  *   For data we start with alloc chunk force, however we could have been full
1148  *   before, and then the transaction commit could have freed new block groups,
1149  *   so if we now have space to allocate do the force chunk allocation.
1150  */
1151 static const enum btrfs_flush_state data_flush_states[] = {
1152         FLUSH_DELALLOC_FULL,
1153         RUN_DELAYED_IPUTS,
1154         COMMIT_TRANS,
1155         ALLOC_CHUNK_FORCE,
1156 };
1157
1158 static void btrfs_async_reclaim_data_space(struct work_struct *work)
1159 {
1160         struct btrfs_fs_info *fs_info;
1161         struct btrfs_space_info *space_info;
1162         u64 last_tickets_id;
1163         enum btrfs_flush_state flush_state = 0;
1164
1165         fs_info = container_of(work, struct btrfs_fs_info, async_data_reclaim_work);
1166         space_info = fs_info->data_sinfo;
1167
1168         spin_lock(&space_info->lock);
1169         if (list_empty(&space_info->tickets)) {
1170                 space_info->flush = 0;
1171                 spin_unlock(&space_info->lock);
1172                 return;
1173         }
1174         last_tickets_id = space_info->tickets_id;
1175         spin_unlock(&space_info->lock);
1176
1177         while (!space_info->full) {
1178                 flush_space(fs_info, space_info, U64_MAX, ALLOC_CHUNK_FORCE, false);
1179                 spin_lock(&space_info->lock);
1180                 if (list_empty(&space_info->tickets)) {
1181                         space_info->flush = 0;
1182                         spin_unlock(&space_info->lock);
1183                         return;
1184                 }
1185
1186                 /* Something happened, fail everything and bail. */
1187                 if (BTRFS_FS_ERROR(fs_info))
1188                         goto aborted_fs;
1189                 last_tickets_id = space_info->tickets_id;
1190                 spin_unlock(&space_info->lock);
1191         }
1192
1193         while (flush_state < ARRAY_SIZE(data_flush_states)) {
1194                 flush_space(fs_info, space_info, U64_MAX,
1195                             data_flush_states[flush_state], false);
1196                 spin_lock(&space_info->lock);
1197                 if (list_empty(&space_info->tickets)) {
1198                         space_info->flush = 0;
1199                         spin_unlock(&space_info->lock);
1200                         return;
1201                 }
1202
1203                 if (last_tickets_id == space_info->tickets_id) {
1204                         flush_state++;
1205                 } else {
1206                         last_tickets_id = space_info->tickets_id;
1207                         flush_state = 0;
1208                 }
1209
1210                 if (flush_state >= ARRAY_SIZE(data_flush_states)) {
1211                         if (space_info->full) {
1212                                 if (maybe_fail_all_tickets(fs_info, space_info))
1213                                         flush_state = 0;
1214                                 else
1215                                         space_info->flush = 0;
1216                         } else {
1217                                 flush_state = 0;
1218                         }
1219
1220                         /* Something happened, fail everything and bail. */
1221                         if (BTRFS_FS_ERROR(fs_info))
1222                                 goto aborted_fs;
1223
1224                 }
1225                 spin_unlock(&space_info->lock);
1226         }
1227         return;
1228
1229 aborted_fs:
1230         maybe_fail_all_tickets(fs_info, space_info);
1231         space_info->flush = 0;
1232         spin_unlock(&space_info->lock);
1233 }
1234
1235 void btrfs_init_async_reclaim_work(struct btrfs_fs_info *fs_info)
1236 {
1237         INIT_WORK(&fs_info->async_reclaim_work, btrfs_async_reclaim_metadata_space);
1238         INIT_WORK(&fs_info->async_data_reclaim_work, btrfs_async_reclaim_data_space);
1239         INIT_WORK(&fs_info->preempt_reclaim_work,
1240                   btrfs_preempt_reclaim_metadata_space);
1241 }
1242
1243 static const enum btrfs_flush_state priority_flush_states[] = {
1244         FLUSH_DELAYED_ITEMS_NR,
1245         FLUSH_DELAYED_ITEMS,
1246         ALLOC_CHUNK,
1247 };
1248
1249 static const enum btrfs_flush_state evict_flush_states[] = {
1250         FLUSH_DELAYED_ITEMS_NR,
1251         FLUSH_DELAYED_ITEMS,
1252         FLUSH_DELAYED_REFS_NR,
1253         FLUSH_DELAYED_REFS,
1254         FLUSH_DELALLOC,
1255         FLUSH_DELALLOC_WAIT,
1256         FLUSH_DELALLOC_FULL,
1257         ALLOC_CHUNK,
1258         COMMIT_TRANS,
1259 };
1260
1261 static void priority_reclaim_metadata_space(struct btrfs_fs_info *fs_info,
1262                                 struct btrfs_space_info *space_info,
1263                                 struct reserve_ticket *ticket,
1264                                 const enum btrfs_flush_state *states,
1265                                 int states_nr)
1266 {
1267         u64 to_reclaim;
1268         int flush_state = 0;
1269
1270         spin_lock(&space_info->lock);
1271         to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info, space_info);
1272         /*
1273          * This is the priority reclaim path, so to_reclaim could be >0 still
1274          * because we may have only satisified the priority tickets and still
1275          * left non priority tickets on the list.  We would then have
1276          * to_reclaim but ->bytes == 0.
1277          */
1278         if (ticket->bytes == 0) {
1279                 spin_unlock(&space_info->lock);
1280                 return;
1281         }
1282
1283         while (flush_state < states_nr) {
1284                 spin_unlock(&space_info->lock);
1285                 flush_space(fs_info, space_info, to_reclaim, states[flush_state],
1286                             false);
1287                 flush_state++;
1288                 spin_lock(&space_info->lock);
1289                 if (ticket->bytes == 0) {
1290                         spin_unlock(&space_info->lock);
1291                         return;
1292                 }
1293         }
1294
1295         /* Attempt to steal from the global rsv if we can. */
1296         if (!steal_from_global_rsv(fs_info, space_info, ticket)) {
1297                 ticket->error = -ENOSPC;
1298                 remove_ticket(space_info, ticket);
1299         }
1300
1301         /*
1302          * We must run try_granting_tickets here because we could be a large
1303          * ticket in front of a smaller ticket that can now be satisfied with
1304          * the available space.
1305          */
1306         btrfs_try_granting_tickets(fs_info, space_info);
1307         spin_unlock(&space_info->lock);
1308 }
1309
1310 static void priority_reclaim_data_space(struct btrfs_fs_info *fs_info,
1311                                         struct btrfs_space_info *space_info,
1312                                         struct reserve_ticket *ticket)
1313 {
1314         spin_lock(&space_info->lock);
1315
1316         /* We could have been granted before we got here. */
1317         if (ticket->bytes == 0) {
1318                 spin_unlock(&space_info->lock);
1319                 return;
1320         }
1321
1322         while (!space_info->full) {
1323                 spin_unlock(&space_info->lock);
1324                 flush_space(fs_info, space_info, U64_MAX, ALLOC_CHUNK_FORCE, false);
1325                 spin_lock(&space_info->lock);
1326                 if (ticket->bytes == 0) {
1327                         spin_unlock(&space_info->lock);
1328                         return;
1329                 }
1330         }
1331
1332         ticket->error = -ENOSPC;
1333         remove_ticket(space_info, ticket);
1334         btrfs_try_granting_tickets(fs_info, space_info);
1335         spin_unlock(&space_info->lock);
1336 }
1337
1338 static void wait_reserve_ticket(struct btrfs_fs_info *fs_info,
1339                                 struct btrfs_space_info *space_info,
1340                                 struct reserve_ticket *ticket)
1341
1342 {
1343         DEFINE_WAIT(wait);
1344         int ret = 0;
1345
1346         spin_lock(&space_info->lock);
1347         while (ticket->bytes > 0 && ticket->error == 0) {
1348                 ret = prepare_to_wait_event(&ticket->wait, &wait, TASK_KILLABLE);
1349                 if (ret) {
1350                         /*
1351                          * Delete us from the list. After we unlock the space
1352                          * info, we don't want the async reclaim job to reserve
1353                          * space for this ticket. If that would happen, then the
1354                          * ticket's task would not known that space was reserved
1355                          * despite getting an error, resulting in a space leak
1356                          * (bytes_may_use counter of our space_info).
1357                          */
1358                         remove_ticket(space_info, ticket);
1359                         ticket->error = -EINTR;
1360                         break;
1361                 }
1362                 spin_unlock(&space_info->lock);
1363
1364                 schedule();
1365
1366                 finish_wait(&ticket->wait, &wait);
1367                 spin_lock(&space_info->lock);
1368         }
1369         spin_unlock(&space_info->lock);
1370 }
1371
1372 /**
1373  * Do the appropriate flushing and waiting for a ticket
1374  *
1375  * @fs_info:    the filesystem
1376  * @space_info: space info for the reservation
1377  * @ticket:     ticket for the reservation
1378  * @start_ns:   timestamp when the reservation started
1379  * @orig_bytes: amount of bytes originally reserved
1380  * @flush:      how much we can flush
1381  *
1382  * This does the work of figuring out how to flush for the ticket, waiting for
1383  * the reservation, and returning the appropriate error if there is one.
1384  */
1385 static int handle_reserve_ticket(struct btrfs_fs_info *fs_info,
1386                                  struct btrfs_space_info *space_info,
1387                                  struct reserve_ticket *ticket,
1388                                  u64 start_ns, u64 orig_bytes,
1389                                  enum btrfs_reserve_flush_enum flush)
1390 {
1391         int ret;
1392
1393         switch (flush) {
1394         case BTRFS_RESERVE_FLUSH_DATA:
1395         case BTRFS_RESERVE_FLUSH_ALL:
1396         case BTRFS_RESERVE_FLUSH_ALL_STEAL:
1397                 wait_reserve_ticket(fs_info, space_info, ticket);
1398                 break;
1399         case BTRFS_RESERVE_FLUSH_LIMIT:
1400                 priority_reclaim_metadata_space(fs_info, space_info, ticket,
1401                                                 priority_flush_states,
1402                                                 ARRAY_SIZE(priority_flush_states));
1403                 break;
1404         case BTRFS_RESERVE_FLUSH_EVICT:
1405                 priority_reclaim_metadata_space(fs_info, space_info, ticket,
1406                                                 evict_flush_states,
1407                                                 ARRAY_SIZE(evict_flush_states));
1408                 break;
1409         case BTRFS_RESERVE_FLUSH_FREE_SPACE_INODE:
1410                 priority_reclaim_data_space(fs_info, space_info, ticket);
1411                 break;
1412         default:
1413                 ASSERT(0);
1414                 break;
1415         }
1416
1417         ret = ticket->error;
1418         ASSERT(list_empty(&ticket->list));
1419         /*
1420          * Check that we can't have an error set if the reservation succeeded,
1421          * as that would confuse tasks and lead them to error out without
1422          * releasing reserved space (if an error happens the expectation is that
1423          * space wasn't reserved at all).
1424          */
1425         ASSERT(!(ticket->bytes == 0 && ticket->error));
1426         trace_btrfs_reserve_ticket(fs_info, space_info->flags, orig_bytes,
1427                                    start_ns, flush, ticket->error);
1428         return ret;
1429 }
1430
1431 /*
1432  * This returns true if this flush state will go through the ordinary flushing
1433  * code.
1434  */
1435 static inline bool is_normal_flushing(enum btrfs_reserve_flush_enum flush)
1436 {
1437         return  (flush == BTRFS_RESERVE_FLUSH_ALL) ||
1438                 (flush == BTRFS_RESERVE_FLUSH_ALL_STEAL);
1439 }
1440
1441 static inline void maybe_clamp_preempt(struct btrfs_fs_info *fs_info,
1442                                        struct btrfs_space_info *space_info)
1443 {
1444         u64 ordered = percpu_counter_sum_positive(&fs_info->ordered_bytes);
1445         u64 delalloc = percpu_counter_sum_positive(&fs_info->delalloc_bytes);
1446
1447         /*
1448          * If we're heavy on ordered operations then clamping won't help us.  We
1449          * need to clamp specifically to keep up with dirty'ing buffered
1450          * writers, because there's not a 1:1 correlation of writing delalloc
1451          * and freeing space, like there is with flushing delayed refs or
1452          * delayed nodes.  If we're already more ordered than delalloc then
1453          * we're keeping up, otherwise we aren't and should probably clamp.
1454          */
1455         if (ordered < delalloc)
1456                 space_info->clamp = min(space_info->clamp + 1, 8);
1457 }
1458
1459 static inline bool can_steal(enum btrfs_reserve_flush_enum flush)
1460 {
1461         return (flush == BTRFS_RESERVE_FLUSH_ALL_STEAL ||
1462                 flush == BTRFS_RESERVE_FLUSH_EVICT);
1463 }
1464
1465 /**
1466  * Try to reserve bytes from the block_rsv's space
1467  *
1468  * @fs_info:    the filesystem
1469  * @space_info: space info we want to allocate from
1470  * @orig_bytes: number of bytes we want
1471  * @flush:      whether or not we can flush to make our reservation
1472  *
1473  * This will reserve orig_bytes number of bytes from the space info associated
1474  * with the block_rsv.  If there is not enough space it will make an attempt to
1475  * flush out space to make room.  It will do this by flushing delalloc if
1476  * possible or committing the transaction.  If flush is 0 then no attempts to
1477  * regain reservations will be made and this will fail if there is not enough
1478  * space already.
1479  */
1480 static int __reserve_bytes(struct btrfs_fs_info *fs_info,
1481                            struct btrfs_space_info *space_info, u64 orig_bytes,
1482                            enum btrfs_reserve_flush_enum flush)
1483 {
1484         struct work_struct *async_work;
1485         struct reserve_ticket ticket;
1486         u64 start_ns = 0;
1487         u64 used;
1488         int ret = 0;
1489         bool pending_tickets;
1490
1491         ASSERT(orig_bytes);
1492         ASSERT(!current->journal_info || flush != BTRFS_RESERVE_FLUSH_ALL);
1493
1494         if (flush == BTRFS_RESERVE_FLUSH_DATA)
1495                 async_work = &fs_info->async_data_reclaim_work;
1496         else
1497                 async_work = &fs_info->async_reclaim_work;
1498
1499         spin_lock(&space_info->lock);
1500         ret = -ENOSPC;
1501         used = btrfs_space_info_used(space_info, true);
1502
1503         /*
1504          * We don't want NO_FLUSH allocations to jump everybody, they can
1505          * generally handle ENOSPC in a different way, so treat them the same as
1506          * normal flushers when it comes to skipping pending tickets.
1507          */
1508         if (is_normal_flushing(flush) || (flush == BTRFS_RESERVE_NO_FLUSH))
1509                 pending_tickets = !list_empty(&space_info->tickets) ||
1510                         !list_empty(&space_info->priority_tickets);
1511         else
1512                 pending_tickets = !list_empty(&space_info->priority_tickets);
1513
1514         /*
1515          * Carry on if we have enough space (short-circuit) OR call
1516          * can_overcommit() to ensure we can overcommit to continue.
1517          */
1518         if (!pending_tickets &&
1519             ((used + orig_bytes <= space_info->total_bytes) ||
1520              btrfs_can_overcommit(fs_info, space_info, orig_bytes, flush))) {
1521                 btrfs_space_info_update_bytes_may_use(fs_info, space_info,
1522                                                       orig_bytes);
1523                 ret = 0;
1524         }
1525
1526         /*
1527          * If we couldn't make a reservation then setup our reservation ticket
1528          * and kick the async worker if it's not already running.
1529          *
1530          * If we are a priority flusher then we just need to add our ticket to
1531          * the list and we will do our own flushing further down.
1532          */
1533         if (ret && flush != BTRFS_RESERVE_NO_FLUSH) {
1534                 ticket.bytes = orig_bytes;
1535                 ticket.error = 0;
1536                 space_info->reclaim_size += ticket.bytes;
1537                 init_waitqueue_head(&ticket.wait);
1538                 ticket.steal = can_steal(flush);
1539                 if (trace_btrfs_reserve_ticket_enabled())
1540                         start_ns = ktime_get_ns();
1541
1542                 if (flush == BTRFS_RESERVE_FLUSH_ALL ||
1543                     flush == BTRFS_RESERVE_FLUSH_ALL_STEAL ||
1544                     flush == BTRFS_RESERVE_FLUSH_DATA) {
1545                         list_add_tail(&ticket.list, &space_info->tickets);
1546                         if (!space_info->flush) {
1547                                 /*
1548                                  * We were forced to add a reserve ticket, so
1549                                  * our preemptive flushing is unable to keep
1550                                  * up.  Clamp down on the threshold for the
1551                                  * preemptive flushing in order to keep up with
1552                                  * the workload.
1553                                  */
1554                                 maybe_clamp_preempt(fs_info, space_info);
1555
1556                                 space_info->flush = 1;
1557                                 trace_btrfs_trigger_flush(fs_info,
1558                                                           space_info->flags,
1559                                                           orig_bytes, flush,
1560                                                           "enospc");
1561                                 queue_work(system_unbound_wq, async_work);
1562                         }
1563                 } else {
1564                         list_add_tail(&ticket.list,
1565                                       &space_info->priority_tickets);
1566                 }
1567         } else if (!ret && space_info->flags & BTRFS_BLOCK_GROUP_METADATA) {
1568                 used += orig_bytes;
1569                 /*
1570                  * We will do the space reservation dance during log replay,
1571                  * which means we won't have fs_info->fs_root set, so don't do
1572                  * the async reclaim as we will panic.
1573                  */
1574                 if (!test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags) &&
1575                     !work_busy(&fs_info->preempt_reclaim_work) &&
1576                     need_preemptive_reclaim(fs_info, space_info)) {
1577                         trace_btrfs_trigger_flush(fs_info, space_info->flags,
1578                                                   orig_bytes, flush, "preempt");
1579                         queue_work(system_unbound_wq,
1580                                    &fs_info->preempt_reclaim_work);
1581                 }
1582         }
1583         spin_unlock(&space_info->lock);
1584         if (!ret || flush == BTRFS_RESERVE_NO_FLUSH)
1585                 return ret;
1586
1587         return handle_reserve_ticket(fs_info, space_info, &ticket, start_ns,
1588                                      orig_bytes, flush);
1589 }
1590
1591 /**
1592  * Trye to reserve metadata bytes from the block_rsv's space
1593  *
1594  * @fs_info:    the filesystem
1595  * @block_rsv:  block_rsv we're allocating for
1596  * @orig_bytes: number of bytes we want
1597  * @flush:      whether or not we can flush to make our reservation
1598  *
1599  * This will reserve orig_bytes number of bytes from the space info associated
1600  * with the block_rsv.  If there is not enough space it will make an attempt to
1601  * flush out space to make room.  It will do this by flushing delalloc if
1602  * possible or committing the transaction.  If flush is 0 then no attempts to
1603  * regain reservations will be made and this will fail if there is not enough
1604  * space already.
1605  */
1606 int btrfs_reserve_metadata_bytes(struct btrfs_fs_info *fs_info,
1607                                  struct btrfs_block_rsv *block_rsv,
1608                                  u64 orig_bytes,
1609                                  enum btrfs_reserve_flush_enum flush)
1610 {
1611         int ret;
1612
1613         ret = __reserve_bytes(fs_info, block_rsv->space_info, orig_bytes, flush);
1614         if (ret == -ENOSPC) {
1615                 trace_btrfs_space_reservation(fs_info, "space_info:enospc",
1616                                               block_rsv->space_info->flags,
1617                                               orig_bytes, 1);
1618
1619                 if (btrfs_test_opt(fs_info, ENOSPC_DEBUG))
1620                         btrfs_dump_space_info(fs_info, block_rsv->space_info,
1621                                               orig_bytes, 0);
1622         }
1623         return ret;
1624 }
1625
1626 /**
1627  * Try to reserve data bytes for an allocation
1628  *
1629  * @fs_info: the filesystem
1630  * @bytes:   number of bytes we need
1631  * @flush:   how we are allowed to flush
1632  *
1633  * This will reserve bytes from the data space info.  If there is not enough
1634  * space then we will attempt to flush space as specified by flush.
1635  */
1636 int btrfs_reserve_data_bytes(struct btrfs_fs_info *fs_info, u64 bytes,
1637                              enum btrfs_reserve_flush_enum flush)
1638 {
1639         struct btrfs_space_info *data_sinfo = fs_info->data_sinfo;
1640         int ret;
1641
1642         ASSERT(flush == BTRFS_RESERVE_FLUSH_DATA ||
1643                flush == BTRFS_RESERVE_FLUSH_FREE_SPACE_INODE);
1644         ASSERT(!current->journal_info || flush != BTRFS_RESERVE_FLUSH_DATA);
1645
1646         ret = __reserve_bytes(fs_info, data_sinfo, bytes, flush);
1647         if (ret == -ENOSPC) {
1648                 trace_btrfs_space_reservation(fs_info, "space_info:enospc",
1649                                               data_sinfo->flags, bytes, 1);
1650                 if (btrfs_test_opt(fs_info, ENOSPC_DEBUG))
1651                         btrfs_dump_space_info(fs_info, data_sinfo, bytes, 0);
1652         }
1653         return ret;
1654 }