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