1 // SPDX-License-Identifier: GPL-2.0 2 3 #include "misc.h" 4 #include "ctree.h" 5 #include "space-info.h" 6 #include "sysfs.h" 7 #include "volumes.h" 8 #include "free-space-cache.h" 9 #include "ordered-data.h" 10 #include "transaction.h" 11 #include "block-group.h" 12 13 /* 14 * HOW DOES SPACE RESERVATION WORK 15 * 16 * If you want to know about delalloc specifically, there is a separate comment 17 * for that with the delalloc code. This comment is about how the whole system 18 * works generally. 19 * 20 * BASIC CONCEPTS 21 * 22 * 1) space_info. This is the ultimate arbiter of how much space we can use. 23 * There's a description of the bytes_ fields with the struct declaration, 24 * refer to that for specifics on each field. Suffice it to say that for 25 * reservations we care about total_bytes - SUM(space_info->bytes_) when 26 * determining if there is space to make an allocation. There is a space_info 27 * for METADATA, SYSTEM, and DATA areas. 28 * 29 * 2) block_rsv's. These are basically buckets for every different type of 30 * metadata reservation we have. You can see the comment in the block_rsv 31 * code on the rules for each type, but generally block_rsv->reserved is how 32 * much space is accounted for in space_info->bytes_may_use. 33 * 34 * 3) btrfs_calc*_size. These are the worst case calculations we used based 35 * on the number of items we will want to modify. We have one for changing 36 * items, and one for inserting new items. Generally we use these helpers to 37 * determine the size of the block reserves, and then use the actual bytes 38 * values to adjust the space_info counters. 39 * 40 * MAKING RESERVATIONS, THE NORMAL CASE 41 * 42 * We call into either btrfs_reserve_data_bytes() or 43 * btrfs_reserve_metadata_bytes(), depending on which we're looking for, with 44 * num_bytes we want to reserve. 45 * 46 * ->reserve 47 * space_info->bytes_may_reserve += num_bytes 48 * 49 * ->extent allocation 50 * Call btrfs_add_reserved_bytes() which does 51 * space_info->bytes_may_reserve -= num_bytes 52 * space_info->bytes_reserved += extent_bytes 53 * 54 * ->insert reference 55 * Call btrfs_update_block_group() which does 56 * space_info->bytes_reserved -= extent_bytes 57 * space_info->bytes_used += extent_bytes 58 * 59 * MAKING RESERVATIONS, FLUSHING NORMALLY (non-priority) 60 * 61 * Assume we are unable to simply make the reservation because we do not have 62 * enough space 63 * 64 * -> __reserve_bytes 65 * create a reserve_ticket with ->bytes set to our reservation, add it to 66 * the tail of space_info->tickets, kick async flush thread 67 * 68 * ->handle_reserve_ticket 69 * wait on ticket->wait for ->bytes to be reduced to 0, or ->error to be set 70 * on the ticket. 71 * 72 * -> btrfs_async_reclaim_metadata_space/btrfs_async_reclaim_data_space 73 * Flushes various things attempting to free up space. 74 * 75 * -> btrfs_try_granting_tickets() 76 * This is called by anything that either subtracts space from 77 * space_info->bytes_may_use, ->bytes_pinned, etc, or adds to the 78 * space_info->total_bytes. This loops through the ->priority_tickets and 79 * then the ->tickets list checking to see if the reservation can be 80 * completed. If it can the space is added to space_info->bytes_may_use and 81 * the ticket is woken up. 82 * 83 * -> ticket wakeup 84 * Check if ->bytes == 0, if it does we got our reservation and we can carry 85 * on, if not return the appropriate error (ENOSPC, but can be EINTR if we 86 * were interrupted.) 87 * 88 * MAKING RESERVATIONS, FLUSHING HIGH PRIORITY 89 * 90 * Same as the above, except we add ourselves to the 91 * space_info->priority_tickets, and we do not use ticket->wait, we simply 92 * call flush_space() ourselves for the states that are safe for us to call 93 * without deadlocking and hope for the best. 94 * 95 * THE FLUSHING STATES 96 * 97 * Generally speaking we will have two cases for each state, a "nice" state 98 * and a "ALL THE THINGS" state. In btrfs we delay a lot of work in order to 99 * reduce the locking over head on the various trees, and even to keep from 100 * doing any work at all in the case of delayed refs. Each of these delayed 101 * things however hold reservations, and so letting them run allows us to 102 * reclaim space so we can make new reservations. 103 * 104 * FLUSH_DELAYED_ITEMS 105 * Every inode has a delayed item to update the inode. Take a simple write 106 * for example, we would update the inode item at write time to update the 107 * mtime, and then again at finish_ordered_io() time in order to update the 108 * isize or bytes. We keep these delayed items to coalesce these operations 109 * into a single operation done on demand. These are an easy way to reclaim 110 * metadata space. 111 * 112 * FLUSH_DELALLOC 113 * Look at the delalloc comment to get an idea of how much space is reserved 114 * for delayed allocation. We can reclaim some of this space simply by 115 * running delalloc, but usually we need to wait for ordered extents to 116 * reclaim the bulk of this space. 117 * 118 * FLUSH_DELAYED_REFS 119 * We have a block reserve for the outstanding delayed refs space, and every 120 * delayed ref operation holds a reservation. Running these is a quick way 121 * to reclaim space, but we want to hold this until the end because COW can 122 * churn a lot and we can avoid making some extent tree modifications if we 123 * are able to delay for as long as possible. 124 * 125 * ALLOC_CHUNK 126 * We will skip this the first time through space reservation, because of 127 * overcommit and we don't want to have a lot of useless metadata space when 128 * our worst case reservations will likely never come true. 129 * 130 * RUN_DELAYED_IPUTS 131 * If we're freeing inodes we're likely freeing checksums, file extent 132 * items, and extent tree items. Loads of space could be freed up by these 133 * operations, however they won't be usable until the transaction commits. 134 * 135 * COMMIT_TRANS 136 * may_commit_transaction() is the ultimate arbiter on whether we commit the 137 * transaction or not. In order to avoid constantly churning we do all the 138 * above flushing first and then commit the transaction as the last resort. 139 * However we need to take into account things like pinned space that would 140 * be freed, plus any delayed work we may not have gotten rid of in the case 141 * of metadata. 142 * 143 * FORCE_COMMIT_TRANS 144 * For use by the preemptive flusher. We use this to bypass the ticketing 145 * checks in may_commit_transaction, as we have more information about the 146 * overall state of the system and may want to commit the transaction ahead 147 * of actual ENOSPC conditions. 148 * 149 * OVERCOMMIT 150 * 151 * Because we hold so many reservations for metadata we will allow you to 152 * reserve more space than is currently free in the currently allocate 153 * metadata space. This only happens with metadata, data does not allow 154 * overcommitting. 155 * 156 * You can see the current logic for when we allow overcommit in 157 * btrfs_can_overcommit(), but it only applies to unallocated space. If there 158 * is no unallocated space to be had, all reservations are kept within the 159 * free space in the allocated metadata chunks. 160 * 161 * Because of overcommitting, you generally want to use the 162 * btrfs_can_overcommit() logic for metadata allocations, as it does the right 163 * thing with or without extra unallocated space. 164 */ 165 166 u64 __pure btrfs_space_info_used(struct btrfs_space_info *s_info, 167 bool may_use_included) 168 { 169 ASSERT(s_info); 170 return s_info->bytes_used + s_info->bytes_reserved + 171 s_info->bytes_pinned + s_info->bytes_readonly + 172 (may_use_included ? s_info->bytes_may_use : 0); 173 } 174 175 /* 176 * after adding space to the filesystem, we need to clear the full flags 177 * on all the space infos. 178 */ 179 void btrfs_clear_space_info_full(struct btrfs_fs_info *info) 180 { 181 struct list_head *head = &info->space_info; 182 struct btrfs_space_info *found; 183 184 list_for_each_entry(found, head, list) 185 found->full = 0; 186 } 187 188 static int create_space_info(struct btrfs_fs_info *info, u64 flags) 189 { 190 191 struct btrfs_space_info *space_info; 192 int i; 193 int ret; 194 195 space_info = kzalloc(sizeof(*space_info), GFP_NOFS); 196 if (!space_info) 197 return -ENOMEM; 198 199 ret = percpu_counter_init(&space_info->total_bytes_pinned, 0, 200 GFP_KERNEL); 201 if (ret) { 202 kfree(space_info); 203 return ret; 204 } 205 206 for (i = 0; i < BTRFS_NR_RAID_TYPES; i++) 207 INIT_LIST_HEAD(&space_info->block_groups[i]); 208 init_rwsem(&space_info->groups_sem); 209 spin_lock_init(&space_info->lock); 210 space_info->flags = flags & BTRFS_BLOCK_GROUP_TYPE_MASK; 211 space_info->force_alloc = CHUNK_ALLOC_NO_FORCE; 212 INIT_LIST_HEAD(&space_info->ro_bgs); 213 INIT_LIST_HEAD(&space_info->tickets); 214 INIT_LIST_HEAD(&space_info->priority_tickets); 215 216 ret = btrfs_sysfs_add_space_info_type(info, space_info); 217 if (ret) 218 return ret; 219 220 list_add(&space_info->list, &info->space_info); 221 if (flags & BTRFS_BLOCK_GROUP_DATA) 222 info->data_sinfo = space_info; 223 224 return ret; 225 } 226 227 int btrfs_init_space_info(struct btrfs_fs_info *fs_info) 228 { 229 struct btrfs_super_block *disk_super; 230 u64 features; 231 u64 flags; 232 int mixed = 0; 233 int ret; 234 235 disk_super = fs_info->super_copy; 236 if (!btrfs_super_root(disk_super)) 237 return -EINVAL; 238 239 features = btrfs_super_incompat_flags(disk_super); 240 if (features & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS) 241 mixed = 1; 242 243 flags = BTRFS_BLOCK_GROUP_SYSTEM; 244 ret = create_space_info(fs_info, flags); 245 if (ret) 246 goto out; 247 248 if (mixed) { 249 flags = BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DATA; 250 ret = create_space_info(fs_info, flags); 251 } else { 252 flags = BTRFS_BLOCK_GROUP_METADATA; 253 ret = create_space_info(fs_info, flags); 254 if (ret) 255 goto out; 256 257 flags = BTRFS_BLOCK_GROUP_DATA; 258 ret = create_space_info(fs_info, flags); 259 } 260 out: 261 return ret; 262 } 263 264 void btrfs_update_space_info(struct btrfs_fs_info *info, u64 flags, 265 u64 total_bytes, u64 bytes_used, 266 u64 bytes_readonly, 267 struct btrfs_space_info **space_info) 268 { 269 struct btrfs_space_info *found; 270 int factor; 271 272 factor = btrfs_bg_type_to_factor(flags); 273 274 found = btrfs_find_space_info(info, flags); 275 ASSERT(found); 276 spin_lock(&found->lock); 277 found->total_bytes += total_bytes; 278 found->disk_total += total_bytes * factor; 279 found->bytes_used += bytes_used; 280 found->disk_used += bytes_used * factor; 281 found->bytes_readonly += bytes_readonly; 282 if (total_bytes > 0) 283 found->full = 0; 284 btrfs_try_granting_tickets(info, found); 285 spin_unlock(&found->lock); 286 *space_info = found; 287 } 288 289 struct btrfs_space_info *btrfs_find_space_info(struct btrfs_fs_info *info, 290 u64 flags) 291 { 292 struct list_head *head = &info->space_info; 293 struct btrfs_space_info *found; 294 295 flags &= BTRFS_BLOCK_GROUP_TYPE_MASK; 296 297 list_for_each_entry(found, head, list) { 298 if (found->flags & flags) 299 return found; 300 } 301 return NULL; 302 } 303 304 static u64 calc_available_free_space(struct btrfs_fs_info *fs_info, 305 struct btrfs_space_info *space_info, 306 enum btrfs_reserve_flush_enum flush) 307 { 308 u64 profile; 309 u64 avail; 310 int factor; 311 312 if (space_info->flags & BTRFS_BLOCK_GROUP_SYSTEM) 313 profile = btrfs_system_alloc_profile(fs_info); 314 else 315 profile = btrfs_metadata_alloc_profile(fs_info); 316 317 avail = atomic64_read(&fs_info->free_chunk_space); 318 319 /* 320 * If we have dup, raid1 or raid10 then only half of the free 321 * space is actually usable. For raid56, the space info used 322 * doesn't include the parity drive, so we don't have to 323 * change the math 324 */ 325 factor = btrfs_bg_type_to_factor(profile); 326 avail = div_u64(avail, factor); 327 328 /* 329 * If we aren't flushing all things, let us overcommit up to 330 * 1/2th of the space. If we can flush, don't let us overcommit 331 * too much, let it overcommit up to 1/8 of the space. 332 */ 333 if (flush == BTRFS_RESERVE_FLUSH_ALL) 334 avail >>= 3; 335 else 336 avail >>= 1; 337 return avail; 338 } 339 340 int btrfs_can_overcommit(struct btrfs_fs_info *fs_info, 341 struct btrfs_space_info *space_info, u64 bytes, 342 enum btrfs_reserve_flush_enum flush) 343 { 344 u64 avail; 345 u64 used; 346 347 /* Don't overcommit when in mixed mode */ 348 if (space_info->flags & BTRFS_BLOCK_GROUP_DATA) 349 return 0; 350 351 used = btrfs_space_info_used(space_info, true); 352 avail = calc_available_free_space(fs_info, space_info, flush); 353 354 if (used + bytes < space_info->total_bytes + avail) 355 return 1; 356 return 0; 357 } 358 359 static void remove_ticket(struct btrfs_space_info *space_info, 360 struct reserve_ticket *ticket) 361 { 362 if (!list_empty(&ticket->list)) { 363 list_del_init(&ticket->list); 364 ASSERT(space_info->reclaim_size >= ticket->bytes); 365 space_info->reclaim_size -= ticket->bytes; 366 } 367 } 368 369 /* 370 * This is for space we already have accounted in space_info->bytes_may_use, so 371 * basically when we're returning space from block_rsv's. 372 */ 373 void btrfs_try_granting_tickets(struct btrfs_fs_info *fs_info, 374 struct btrfs_space_info *space_info) 375 { 376 struct list_head *head; 377 enum btrfs_reserve_flush_enum flush = BTRFS_RESERVE_NO_FLUSH; 378 379 lockdep_assert_held(&space_info->lock); 380 381 head = &space_info->priority_tickets; 382 again: 383 while (!list_empty(head)) { 384 struct reserve_ticket *ticket; 385 u64 used = btrfs_space_info_used(space_info, true); 386 387 ticket = list_first_entry(head, struct reserve_ticket, list); 388 389 /* Check and see if our ticket can be satisified now. */ 390 if ((used + ticket->bytes <= space_info->total_bytes) || 391 btrfs_can_overcommit(fs_info, space_info, ticket->bytes, 392 flush)) { 393 btrfs_space_info_update_bytes_may_use(fs_info, 394 space_info, 395 ticket->bytes); 396 remove_ticket(space_info, ticket); 397 ticket->bytes = 0; 398 space_info->tickets_id++; 399 wake_up(&ticket->wait); 400 } else { 401 break; 402 } 403 } 404 405 if (head == &space_info->priority_tickets) { 406 head = &space_info->tickets; 407 flush = BTRFS_RESERVE_FLUSH_ALL; 408 goto again; 409 } 410 } 411 412 #define DUMP_BLOCK_RSV(fs_info, rsv_name) \ 413 do { \ 414 struct btrfs_block_rsv *__rsv = &(fs_info)->rsv_name; \ 415 spin_lock(&__rsv->lock); \ 416 btrfs_info(fs_info, #rsv_name ": size %llu reserved %llu", \ 417 __rsv->size, __rsv->reserved); \ 418 spin_unlock(&__rsv->lock); \ 419 } while (0) 420 421 static void __btrfs_dump_space_info(struct btrfs_fs_info *fs_info, 422 struct btrfs_space_info *info) 423 { 424 lockdep_assert_held(&info->lock); 425 426 btrfs_info(fs_info, "space_info %llu has %llu free, is %sfull", 427 info->flags, 428 info->total_bytes - btrfs_space_info_used(info, true), 429 info->full ? "" : "not "); 430 btrfs_info(fs_info, 431 "space_info total=%llu, used=%llu, pinned=%llu, reserved=%llu, may_use=%llu, readonly=%llu", 432 info->total_bytes, info->bytes_used, info->bytes_pinned, 433 info->bytes_reserved, info->bytes_may_use, 434 info->bytes_readonly); 435 436 DUMP_BLOCK_RSV(fs_info, global_block_rsv); 437 DUMP_BLOCK_RSV(fs_info, trans_block_rsv); 438 DUMP_BLOCK_RSV(fs_info, chunk_block_rsv); 439 DUMP_BLOCK_RSV(fs_info, delayed_block_rsv); 440 DUMP_BLOCK_RSV(fs_info, delayed_refs_rsv); 441 442 } 443 444 void btrfs_dump_space_info(struct btrfs_fs_info *fs_info, 445 struct btrfs_space_info *info, u64 bytes, 446 int dump_block_groups) 447 { 448 struct btrfs_block_group *cache; 449 int index = 0; 450 451 spin_lock(&info->lock); 452 __btrfs_dump_space_info(fs_info, info); 453 spin_unlock(&info->lock); 454 455 if (!dump_block_groups) 456 return; 457 458 down_read(&info->groups_sem); 459 again: 460 list_for_each_entry(cache, &info->block_groups[index], list) { 461 spin_lock(&cache->lock); 462 btrfs_info(fs_info, 463 "block group %llu has %llu bytes, %llu used %llu pinned %llu reserved %s", 464 cache->start, cache->length, cache->used, cache->pinned, 465 cache->reserved, cache->ro ? "[readonly]" : ""); 466 spin_unlock(&cache->lock); 467 btrfs_dump_free_space(cache, bytes); 468 } 469 if (++index < BTRFS_NR_RAID_TYPES) 470 goto again; 471 up_read(&info->groups_sem); 472 } 473 474 static inline u64 calc_reclaim_items_nr(struct btrfs_fs_info *fs_info, 475 u64 to_reclaim) 476 { 477 u64 bytes; 478 u64 nr; 479 480 bytes = btrfs_calc_insert_metadata_size(fs_info, 1); 481 nr = div64_u64(to_reclaim, bytes); 482 if (!nr) 483 nr = 1; 484 return nr; 485 } 486 487 #define EXTENT_SIZE_PER_ITEM SZ_256K 488 489 /* 490 * shrink metadata reservation for delalloc 491 */ 492 static void shrink_delalloc(struct btrfs_fs_info *fs_info, 493 struct btrfs_space_info *space_info, 494 u64 to_reclaim, bool wait_ordered) 495 { 496 struct btrfs_trans_handle *trans; 497 u64 delalloc_bytes; 498 u64 ordered_bytes; 499 u64 items; 500 long time_left; 501 int loops; 502 503 /* Calc the number of the pages we need flush for space reservation */ 504 if (to_reclaim == U64_MAX) { 505 items = U64_MAX; 506 } else { 507 /* 508 * to_reclaim is set to however much metadata we need to 509 * reclaim, but reclaiming that much data doesn't really track 510 * exactly, so increase the amount to reclaim by 2x in order to 511 * make sure we're flushing enough delalloc to hopefully reclaim 512 * some metadata reservations. 513 */ 514 items = calc_reclaim_items_nr(fs_info, to_reclaim) * 2; 515 to_reclaim = items * EXTENT_SIZE_PER_ITEM; 516 } 517 518 trans = (struct btrfs_trans_handle *)current->journal_info; 519 520 delalloc_bytes = percpu_counter_sum_positive( 521 &fs_info->delalloc_bytes); 522 ordered_bytes = percpu_counter_sum_positive(&fs_info->ordered_bytes); 523 if (delalloc_bytes == 0 && ordered_bytes == 0) 524 return; 525 526 /* 527 * If we are doing more ordered than delalloc we need to just wait on 528 * ordered extents, otherwise we'll waste time trying to flush delalloc 529 * that likely won't give us the space back we need. 530 */ 531 if (ordered_bytes > delalloc_bytes) 532 wait_ordered = true; 533 534 loops = 0; 535 while ((delalloc_bytes || ordered_bytes) && loops < 3) { 536 u64 temp = min(delalloc_bytes, to_reclaim) >> PAGE_SHIFT; 537 long nr_pages = min_t(u64, temp, LONG_MAX); 538 539 btrfs_start_delalloc_roots(fs_info, nr_pages, true); 540 541 loops++; 542 if (wait_ordered && !trans) { 543 btrfs_wait_ordered_roots(fs_info, items, 0, (u64)-1); 544 } else { 545 time_left = schedule_timeout_killable(1); 546 if (time_left) 547 break; 548 } 549 550 spin_lock(&space_info->lock); 551 if (list_empty(&space_info->tickets) && 552 list_empty(&space_info->priority_tickets)) { 553 spin_unlock(&space_info->lock); 554 break; 555 } 556 spin_unlock(&space_info->lock); 557 558 delalloc_bytes = percpu_counter_sum_positive( 559 &fs_info->delalloc_bytes); 560 ordered_bytes = percpu_counter_sum_positive( 561 &fs_info->ordered_bytes); 562 } 563 } 564 565 /** 566 * Possibly commit the transaction if its ok to 567 * 568 * @fs_info: the filesystem 569 * @space_info: space_info we are checking for commit, either data or metadata 570 * 571 * This will check to make sure that committing the transaction will actually 572 * get us somewhere and then commit the transaction if it does. Otherwise it 573 * will return -ENOSPC. 574 */ 575 static int may_commit_transaction(struct btrfs_fs_info *fs_info, 576 struct btrfs_space_info *space_info) 577 { 578 struct reserve_ticket *ticket = NULL; 579 struct btrfs_block_rsv *delayed_rsv = &fs_info->delayed_block_rsv; 580 struct btrfs_block_rsv *delayed_refs_rsv = &fs_info->delayed_refs_rsv; 581 struct btrfs_block_rsv *trans_rsv = &fs_info->trans_block_rsv; 582 struct btrfs_trans_handle *trans; 583 u64 reclaim_bytes = 0; 584 u64 bytes_needed = 0; 585 u64 cur_free_bytes = 0; 586 587 trans = (struct btrfs_trans_handle *)current->journal_info; 588 if (trans) 589 return -EAGAIN; 590 591 spin_lock(&space_info->lock); 592 cur_free_bytes = btrfs_space_info_used(space_info, true); 593 if (cur_free_bytes < space_info->total_bytes) 594 cur_free_bytes = space_info->total_bytes - cur_free_bytes; 595 else 596 cur_free_bytes = 0; 597 598 if (!list_empty(&space_info->priority_tickets)) 599 ticket = list_first_entry(&space_info->priority_tickets, 600 struct reserve_ticket, list); 601 else if (!list_empty(&space_info->tickets)) 602 ticket = list_first_entry(&space_info->tickets, 603 struct reserve_ticket, list); 604 if (ticket) 605 bytes_needed = ticket->bytes; 606 607 if (bytes_needed > cur_free_bytes) 608 bytes_needed -= cur_free_bytes; 609 else 610 bytes_needed = 0; 611 spin_unlock(&space_info->lock); 612 613 if (!bytes_needed) 614 return 0; 615 616 trans = btrfs_join_transaction(fs_info->extent_root); 617 if (IS_ERR(trans)) 618 return PTR_ERR(trans); 619 620 /* 621 * See if there is enough pinned space to make this reservation, or if 622 * we have block groups that are going to be freed, allowing us to 623 * possibly do a chunk allocation the next loop through. 624 */ 625 if (test_bit(BTRFS_TRANS_HAVE_FREE_BGS, &trans->transaction->flags) || 626 __percpu_counter_compare(&space_info->total_bytes_pinned, 627 bytes_needed, 628 BTRFS_TOTAL_BYTES_PINNED_BATCH) >= 0) 629 goto commit; 630 631 /* 632 * See if there is some space in the delayed insertion reserve for this 633 * reservation. If the space_info's don't match (like for DATA or 634 * SYSTEM) then just go enospc, reclaiming this space won't recover any 635 * space to satisfy those reservations. 636 */ 637 if (space_info != delayed_rsv->space_info) 638 goto enospc; 639 640 spin_lock(&delayed_rsv->lock); 641 reclaim_bytes += delayed_rsv->reserved; 642 spin_unlock(&delayed_rsv->lock); 643 644 spin_lock(&delayed_refs_rsv->lock); 645 reclaim_bytes += delayed_refs_rsv->reserved; 646 spin_unlock(&delayed_refs_rsv->lock); 647 648 spin_lock(&trans_rsv->lock); 649 reclaim_bytes += trans_rsv->reserved; 650 spin_unlock(&trans_rsv->lock); 651 652 if (reclaim_bytes >= bytes_needed) 653 goto commit; 654 bytes_needed -= reclaim_bytes; 655 656 if (__percpu_counter_compare(&space_info->total_bytes_pinned, 657 bytes_needed, 658 BTRFS_TOTAL_BYTES_PINNED_BATCH) < 0) 659 goto enospc; 660 661 commit: 662 return btrfs_commit_transaction(trans); 663 enospc: 664 btrfs_end_transaction(trans); 665 return -ENOSPC; 666 } 667 668 /* 669 * Try to flush some data based on policy set by @state. This is only advisory 670 * and may fail for various reasons. The caller is supposed to examine the 671 * state of @space_info to detect the outcome. 672 */ 673 static void flush_space(struct btrfs_fs_info *fs_info, 674 struct btrfs_space_info *space_info, u64 num_bytes, 675 enum btrfs_flush_state state) 676 { 677 struct btrfs_root *root = fs_info->extent_root; 678 struct btrfs_trans_handle *trans; 679 int nr; 680 int ret = 0; 681 682 switch (state) { 683 case FLUSH_DELAYED_ITEMS_NR: 684 case FLUSH_DELAYED_ITEMS: 685 if (state == FLUSH_DELAYED_ITEMS_NR) 686 nr = calc_reclaim_items_nr(fs_info, num_bytes) * 2; 687 else 688 nr = -1; 689 690 trans = btrfs_join_transaction(root); 691 if (IS_ERR(trans)) { 692 ret = PTR_ERR(trans); 693 break; 694 } 695 ret = btrfs_run_delayed_items_nr(trans, nr); 696 btrfs_end_transaction(trans); 697 break; 698 case FLUSH_DELALLOC: 699 case FLUSH_DELALLOC_WAIT: 700 shrink_delalloc(fs_info, space_info, num_bytes, 701 state == FLUSH_DELALLOC_WAIT); 702 break; 703 case FLUSH_DELAYED_REFS_NR: 704 case FLUSH_DELAYED_REFS: 705 trans = btrfs_join_transaction(root); 706 if (IS_ERR(trans)) { 707 ret = PTR_ERR(trans); 708 break; 709 } 710 if (state == FLUSH_DELAYED_REFS_NR) 711 nr = calc_reclaim_items_nr(fs_info, num_bytes); 712 else 713 nr = 0; 714 btrfs_run_delayed_refs(trans, nr); 715 btrfs_end_transaction(trans); 716 break; 717 case ALLOC_CHUNK: 718 case ALLOC_CHUNK_FORCE: 719 trans = btrfs_join_transaction(root); 720 if (IS_ERR(trans)) { 721 ret = PTR_ERR(trans); 722 break; 723 } 724 ret = btrfs_chunk_alloc(trans, 725 btrfs_get_alloc_profile(fs_info, space_info->flags), 726 (state == ALLOC_CHUNK) ? CHUNK_ALLOC_NO_FORCE : 727 CHUNK_ALLOC_FORCE); 728 btrfs_end_transaction(trans); 729 if (ret > 0 || ret == -ENOSPC) 730 ret = 0; 731 break; 732 case RUN_DELAYED_IPUTS: 733 /* 734 * If we have pending delayed iputs then we could free up a 735 * bunch of pinned space, so make sure we run the iputs before 736 * we do our pinned bytes check below. 737 */ 738 btrfs_run_delayed_iputs(fs_info); 739 btrfs_wait_on_delayed_iputs(fs_info); 740 break; 741 case COMMIT_TRANS: 742 ret = may_commit_transaction(fs_info, space_info); 743 break; 744 case FORCE_COMMIT_TRANS: 745 trans = btrfs_join_transaction(root); 746 if (IS_ERR(trans)) { 747 ret = PTR_ERR(trans); 748 break; 749 } 750 ret = btrfs_commit_transaction(trans); 751 break; 752 default: 753 ret = -ENOSPC; 754 break; 755 } 756 757 trace_btrfs_flush_space(fs_info, space_info->flags, num_bytes, state, 758 ret); 759 return; 760 } 761 762 static inline u64 763 btrfs_calc_reclaim_metadata_size(struct btrfs_fs_info *fs_info, 764 struct btrfs_space_info *space_info) 765 { 766 u64 used; 767 u64 avail; 768 u64 to_reclaim = space_info->reclaim_size; 769 770 lockdep_assert_held(&space_info->lock); 771 772 avail = calc_available_free_space(fs_info, space_info, 773 BTRFS_RESERVE_FLUSH_ALL); 774 used = btrfs_space_info_used(space_info, true); 775 776 /* 777 * We may be flushing because suddenly we have less space than we had 778 * before, and now we're well over-committed based on our current free 779 * space. If that's the case add in our overage so we make sure to put 780 * appropriate pressure on the flushing state machine. 781 */ 782 if (space_info->total_bytes + avail < used) 783 to_reclaim += used - (space_info->total_bytes + avail); 784 785 return to_reclaim; 786 } 787 788 static bool need_preemptive_reclaim(struct btrfs_fs_info *fs_info, 789 struct btrfs_space_info *space_info) 790 { 791 u64 ordered, delalloc; 792 u64 thresh = div_factor_fine(space_info->total_bytes, 98); 793 u64 used; 794 795 /* If we're just plain full then async reclaim just slows us down. */ 796 if ((space_info->bytes_used + space_info->bytes_reserved) >= thresh) 797 return false; 798 799 /* 800 * We have tickets queued, bail so we don't compete with the async 801 * flushers. 802 */ 803 if (space_info->reclaim_size) 804 return false; 805 806 /* 807 * If we have over half of the free space occupied by reservations or 808 * pinned then we want to start flushing. 809 * 810 * We do not do the traditional thing here, which is to say 811 * 812 * if (used >= ((total_bytes + avail) / 2)) 813 * return 1; 814 * 815 * because this doesn't quite work how we want. If we had more than 50% 816 * of the space_info used by bytes_used and we had 0 available we'd just 817 * constantly run the background flusher. Instead we want it to kick in 818 * if our reclaimable space exceeds 50% of our available free space. 819 */ 820 thresh = calc_available_free_space(fs_info, space_info, 821 BTRFS_RESERVE_FLUSH_ALL); 822 thresh += (space_info->total_bytes - space_info->bytes_used - 823 space_info->bytes_reserved - space_info->bytes_readonly); 824 thresh >>= 1; 825 826 used = space_info->bytes_pinned; 827 828 /* 829 * If we have more ordered bytes than delalloc bytes then we're either 830 * doing a lot of DIO, or we simply don't have a lot of delalloc waiting 831 * around. Preemptive flushing is only useful in that it can free up 832 * space before tickets need to wait for things to finish. In the case 833 * of ordered extents, preemptively waiting on ordered extents gets us 834 * nothing, if our reservations are tied up in ordered extents we'll 835 * simply have to slow down writers by forcing them to wait on ordered 836 * extents. 837 * 838 * In the case that ordered is larger than delalloc, only include the 839 * block reserves that we would actually be able to directly reclaim 840 * from. In this case if we're heavy on metadata operations this will 841 * clearly be heavy enough to warrant preemptive flushing. In the case 842 * of heavy DIO or ordered reservations, preemptive flushing will just 843 * waste time and cause us to slow down. 844 */ 845 ordered = percpu_counter_sum_positive(&fs_info->ordered_bytes); 846 delalloc = percpu_counter_sum_positive(&fs_info->delalloc_bytes); 847 if (ordered >= delalloc) 848 used += fs_info->delayed_refs_rsv.reserved + 849 fs_info->delayed_block_rsv.reserved; 850 else 851 used += space_info->bytes_may_use; 852 853 return (used >= thresh && !btrfs_fs_closing(fs_info) && 854 !test_bit(BTRFS_FS_STATE_REMOUNTING, &fs_info->fs_state)); 855 } 856 857 static bool steal_from_global_rsv(struct btrfs_fs_info *fs_info, 858 struct btrfs_space_info *space_info, 859 struct reserve_ticket *ticket) 860 { 861 struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv; 862 u64 min_bytes; 863 864 if (global_rsv->space_info != space_info) 865 return false; 866 867 spin_lock(&global_rsv->lock); 868 min_bytes = div_factor(global_rsv->size, 1); 869 if (global_rsv->reserved < min_bytes + ticket->bytes) { 870 spin_unlock(&global_rsv->lock); 871 return false; 872 } 873 global_rsv->reserved -= ticket->bytes; 874 remove_ticket(space_info, ticket); 875 ticket->bytes = 0; 876 wake_up(&ticket->wait); 877 space_info->tickets_id++; 878 if (global_rsv->reserved < global_rsv->size) 879 global_rsv->full = 0; 880 spin_unlock(&global_rsv->lock); 881 882 return true; 883 } 884 885 /* 886 * maybe_fail_all_tickets - we've exhausted our flushing, start failing tickets 887 * @fs_info - fs_info for this fs 888 * @space_info - the space info we were flushing 889 * 890 * We call this when we've exhausted our flushing ability and haven't made 891 * progress in satisfying tickets. The reservation code handles tickets in 892 * order, so if there is a large ticket first and then smaller ones we could 893 * very well satisfy the smaller tickets. This will attempt to wake up any 894 * tickets in the list to catch this case. 895 * 896 * This function returns true if it was able to make progress by clearing out 897 * other tickets, or if it stumbles across a ticket that was smaller than the 898 * first ticket. 899 */ 900 static bool maybe_fail_all_tickets(struct btrfs_fs_info *fs_info, 901 struct btrfs_space_info *space_info) 902 { 903 struct reserve_ticket *ticket; 904 u64 tickets_id = space_info->tickets_id; 905 u64 first_ticket_bytes = 0; 906 907 if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) { 908 btrfs_info(fs_info, "cannot satisfy tickets, dumping space info"); 909 __btrfs_dump_space_info(fs_info, space_info); 910 } 911 912 while (!list_empty(&space_info->tickets) && 913 tickets_id == space_info->tickets_id) { 914 ticket = list_first_entry(&space_info->tickets, 915 struct reserve_ticket, list); 916 917 if (ticket->steal && 918 steal_from_global_rsv(fs_info, space_info, ticket)) 919 return true; 920 921 /* 922 * may_commit_transaction will avoid committing the transaction 923 * if it doesn't feel like the space reclaimed by the commit 924 * would result in the ticket succeeding. However if we have a 925 * smaller ticket in the queue it may be small enough to be 926 * satisified by committing the transaction, so if any 927 * subsequent ticket is smaller than the first ticket go ahead 928 * and send us back for another loop through the enospc flushing 929 * code. 930 */ 931 if (first_ticket_bytes == 0) 932 first_ticket_bytes = ticket->bytes; 933 else if (first_ticket_bytes > ticket->bytes) 934 return true; 935 936 if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) 937 btrfs_info(fs_info, "failing ticket with %llu bytes", 938 ticket->bytes); 939 940 remove_ticket(space_info, ticket); 941 ticket->error = -ENOSPC; 942 wake_up(&ticket->wait); 943 944 /* 945 * We're just throwing tickets away, so more flushing may not 946 * trip over btrfs_try_granting_tickets, so we need to call it 947 * here to see if we can make progress with the next ticket in 948 * the list. 949 */ 950 btrfs_try_granting_tickets(fs_info, space_info); 951 } 952 return (tickets_id != space_info->tickets_id); 953 } 954 955 /* 956 * This is for normal flushers, we can wait all goddamned day if we want to. We 957 * will loop and continuously try to flush as long as we are making progress. 958 * We count progress as clearing off tickets each time we have to loop. 959 */ 960 static void btrfs_async_reclaim_metadata_space(struct work_struct *work) 961 { 962 struct btrfs_fs_info *fs_info; 963 struct btrfs_space_info *space_info; 964 u64 to_reclaim; 965 enum btrfs_flush_state flush_state; 966 int commit_cycles = 0; 967 u64 last_tickets_id; 968 969 fs_info = container_of(work, struct btrfs_fs_info, async_reclaim_work); 970 space_info = btrfs_find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA); 971 972 spin_lock(&space_info->lock); 973 to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info, space_info); 974 if (!to_reclaim) { 975 space_info->flush = 0; 976 spin_unlock(&space_info->lock); 977 return; 978 } 979 last_tickets_id = space_info->tickets_id; 980 spin_unlock(&space_info->lock); 981 982 flush_state = FLUSH_DELAYED_ITEMS_NR; 983 do { 984 flush_space(fs_info, space_info, to_reclaim, flush_state); 985 spin_lock(&space_info->lock); 986 if (list_empty(&space_info->tickets)) { 987 space_info->flush = 0; 988 spin_unlock(&space_info->lock); 989 return; 990 } 991 to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info, 992 space_info); 993 if (last_tickets_id == space_info->tickets_id) { 994 flush_state++; 995 } else { 996 last_tickets_id = space_info->tickets_id; 997 flush_state = FLUSH_DELAYED_ITEMS_NR; 998 if (commit_cycles) 999 commit_cycles--; 1000 } 1001 1002 /* 1003 * We don't want to force a chunk allocation until we've tried 1004 * pretty hard to reclaim space. Think of the case where we 1005 * freed up a bunch of space and so have a lot of pinned space 1006 * to reclaim. We would rather use that than possibly create a 1007 * underutilized metadata chunk. So if this is our first run 1008 * through the flushing state machine skip ALLOC_CHUNK_FORCE and 1009 * commit the transaction. If nothing has changed the next go 1010 * around then we can force a chunk allocation. 1011 */ 1012 if (flush_state == ALLOC_CHUNK_FORCE && !commit_cycles) 1013 flush_state++; 1014 1015 if (flush_state > COMMIT_TRANS) { 1016 commit_cycles++; 1017 if (commit_cycles > 2) { 1018 if (maybe_fail_all_tickets(fs_info, space_info)) { 1019 flush_state = FLUSH_DELAYED_ITEMS_NR; 1020 commit_cycles--; 1021 } else { 1022 space_info->flush = 0; 1023 } 1024 } else { 1025 flush_state = FLUSH_DELAYED_ITEMS_NR; 1026 } 1027 } 1028 spin_unlock(&space_info->lock); 1029 } while (flush_state <= COMMIT_TRANS); 1030 } 1031 1032 /* 1033 * This handles pre-flushing of metadata space before we get to the point that 1034 * we need to start blocking threads on tickets. The logic here is different 1035 * from the other flush paths because it doesn't rely on tickets to tell us how 1036 * much we need to flush, instead it attempts to keep us below the 80% full 1037 * watermark of space by flushing whichever reservation pool is currently the 1038 * largest. 1039 */ 1040 static void btrfs_preempt_reclaim_metadata_space(struct work_struct *work) 1041 { 1042 struct btrfs_fs_info *fs_info; 1043 struct btrfs_space_info *space_info; 1044 struct btrfs_block_rsv *delayed_block_rsv; 1045 struct btrfs_block_rsv *delayed_refs_rsv; 1046 struct btrfs_block_rsv *global_rsv; 1047 struct btrfs_block_rsv *trans_rsv; 1048 1049 fs_info = container_of(work, struct btrfs_fs_info, 1050 preempt_reclaim_work); 1051 space_info = btrfs_find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA); 1052 delayed_block_rsv = &fs_info->delayed_block_rsv; 1053 delayed_refs_rsv = &fs_info->delayed_refs_rsv; 1054 global_rsv = &fs_info->global_block_rsv; 1055 trans_rsv = &fs_info->trans_block_rsv; 1056 1057 spin_lock(&space_info->lock); 1058 while (need_preemptive_reclaim(fs_info, space_info)) { 1059 enum btrfs_flush_state flush; 1060 u64 delalloc_size = 0; 1061 u64 to_reclaim, block_rsv_size; 1062 u64 global_rsv_size = global_rsv->reserved; 1063 1064 /* 1065 * We don't have a precise counter for the metadata being 1066 * reserved for delalloc, so we'll approximate it by subtracting 1067 * out the block rsv's space from the bytes_may_use. If that 1068 * amount is higher than the individual reserves, then we can 1069 * assume it's tied up in delalloc reservations. 1070 */ 1071 block_rsv_size = global_rsv_size + 1072 delayed_block_rsv->reserved + 1073 delayed_refs_rsv->reserved + 1074 trans_rsv->reserved; 1075 if (block_rsv_size < space_info->bytes_may_use) 1076 delalloc_size = space_info->bytes_may_use - block_rsv_size; 1077 spin_unlock(&space_info->lock); 1078 1079 /* 1080 * We don't want to include the global_rsv in our calculation, 1081 * because that's space we can't touch. Subtract it from the 1082 * block_rsv_size for the next checks. 1083 */ 1084 block_rsv_size -= global_rsv_size; 1085 1086 /* 1087 * We really want to avoid flushing delalloc too much, as it 1088 * could result in poor allocation patterns, so only flush it if 1089 * it's larger than the rest of the pools combined. 1090 */ 1091 if (delalloc_size > block_rsv_size) { 1092 to_reclaim = delalloc_size; 1093 flush = FLUSH_DELALLOC; 1094 } else if (space_info->bytes_pinned > 1095 (delayed_block_rsv->reserved + 1096 delayed_refs_rsv->reserved)) { 1097 to_reclaim = space_info->bytes_pinned; 1098 flush = FORCE_COMMIT_TRANS; 1099 } else if (delayed_block_rsv->reserved > 1100 delayed_refs_rsv->reserved) { 1101 to_reclaim = delayed_block_rsv->reserved; 1102 flush = FLUSH_DELAYED_ITEMS_NR; 1103 } else { 1104 to_reclaim = delayed_refs_rsv->reserved; 1105 flush = FLUSH_DELAYED_REFS_NR; 1106 } 1107 1108 /* 1109 * We don't want to reclaim everything, just a portion, so scale 1110 * down the to_reclaim by 1/4. If it takes us down to 0, 1111 * reclaim 1 items worth. 1112 */ 1113 to_reclaim >>= 2; 1114 if (!to_reclaim) 1115 to_reclaim = btrfs_calc_insert_metadata_size(fs_info, 1); 1116 flush_space(fs_info, space_info, to_reclaim, flush); 1117 cond_resched(); 1118 spin_lock(&space_info->lock); 1119 } 1120 spin_unlock(&space_info->lock); 1121 } 1122 1123 /* 1124 * FLUSH_DELALLOC_WAIT: 1125 * Space is freed from flushing delalloc in one of two ways. 1126 * 1127 * 1) compression is on and we allocate less space than we reserved 1128 * 2) we are overwriting existing space 1129 * 1130 * For #1 that extra space is reclaimed as soon as the delalloc pages are 1131 * COWed, by way of btrfs_add_reserved_bytes() which adds the actual extent 1132 * length to ->bytes_reserved, and subtracts the reserved space from 1133 * ->bytes_may_use. 1134 * 1135 * For #2 this is trickier. Once the ordered extent runs we will drop the 1136 * extent in the range we are overwriting, which creates a delayed ref for 1137 * that freed extent. This however is not reclaimed until the transaction 1138 * commits, thus the next stages. 1139 * 1140 * RUN_DELAYED_IPUTS 1141 * If we are freeing inodes, we want to make sure all delayed iputs have 1142 * completed, because they could have been on an inode with i_nlink == 0, and 1143 * thus have been truncated and freed up space. But again this space is not 1144 * immediately re-usable, it comes in the form of a delayed ref, which must be 1145 * run and then the transaction must be committed. 1146 * 1147 * FLUSH_DELAYED_REFS 1148 * The above two cases generate delayed refs that will affect 1149 * ->total_bytes_pinned. However this counter can be inconsistent with 1150 * reality if there are outstanding delayed refs. This is because we adjust 1151 * the counter based solely on the current set of delayed refs and disregard 1152 * any on-disk state which might include more refs. So for example, if we 1153 * have an extent with 2 references, but we only drop 1, we'll see that there 1154 * is a negative delayed ref count for the extent and assume that the space 1155 * will be freed, and thus increase ->total_bytes_pinned. 1156 * 1157 * Running the delayed refs gives us the actual real view of what will be 1158 * freed at the transaction commit time. This stage will not actually free 1159 * space for us, it just makes sure that may_commit_transaction() has all of 1160 * the information it needs to make the right decision. 1161 * 1162 * COMMIT_TRANS 1163 * This is where we reclaim all of the pinned space generated by the previous 1164 * two stages. We will not commit the transaction if we don't think we're 1165 * likely to satisfy our request, which means if our current free space + 1166 * total_bytes_pinned < reservation we will not commit. This is why the 1167 * previous states are actually important, to make sure we know for sure 1168 * whether committing the transaction will allow us to make progress. 1169 * 1170 * ALLOC_CHUNK_FORCE 1171 * For data we start with alloc chunk force, however we could have been full 1172 * before, and then the transaction commit could have freed new block groups, 1173 * so if we now have space to allocate do the force chunk allocation. 1174 */ 1175 static const enum btrfs_flush_state data_flush_states[] = { 1176 FLUSH_DELALLOC_WAIT, 1177 RUN_DELAYED_IPUTS, 1178 FLUSH_DELAYED_REFS, 1179 COMMIT_TRANS, 1180 ALLOC_CHUNK_FORCE, 1181 }; 1182 1183 static void btrfs_async_reclaim_data_space(struct work_struct *work) 1184 { 1185 struct btrfs_fs_info *fs_info; 1186 struct btrfs_space_info *space_info; 1187 u64 last_tickets_id; 1188 enum btrfs_flush_state flush_state = 0; 1189 1190 fs_info = container_of(work, struct btrfs_fs_info, async_data_reclaim_work); 1191 space_info = fs_info->data_sinfo; 1192 1193 spin_lock(&space_info->lock); 1194 if (list_empty(&space_info->tickets)) { 1195 space_info->flush = 0; 1196 spin_unlock(&space_info->lock); 1197 return; 1198 } 1199 last_tickets_id = space_info->tickets_id; 1200 spin_unlock(&space_info->lock); 1201 1202 while (!space_info->full) { 1203 flush_space(fs_info, space_info, U64_MAX, ALLOC_CHUNK_FORCE); 1204 spin_lock(&space_info->lock); 1205 if (list_empty(&space_info->tickets)) { 1206 space_info->flush = 0; 1207 spin_unlock(&space_info->lock); 1208 return; 1209 } 1210 last_tickets_id = space_info->tickets_id; 1211 spin_unlock(&space_info->lock); 1212 } 1213 1214 while (flush_state < ARRAY_SIZE(data_flush_states)) { 1215 flush_space(fs_info, space_info, U64_MAX, 1216 data_flush_states[flush_state]); 1217 spin_lock(&space_info->lock); 1218 if (list_empty(&space_info->tickets)) { 1219 space_info->flush = 0; 1220 spin_unlock(&space_info->lock); 1221 return; 1222 } 1223 1224 if (last_tickets_id == space_info->tickets_id) { 1225 flush_state++; 1226 } else { 1227 last_tickets_id = space_info->tickets_id; 1228 flush_state = 0; 1229 } 1230 1231 if (flush_state >= ARRAY_SIZE(data_flush_states)) { 1232 if (space_info->full) { 1233 if (maybe_fail_all_tickets(fs_info, space_info)) 1234 flush_state = 0; 1235 else 1236 space_info->flush = 0; 1237 } else { 1238 flush_state = 0; 1239 } 1240 } 1241 spin_unlock(&space_info->lock); 1242 } 1243 } 1244 1245 void btrfs_init_async_reclaim_work(struct btrfs_fs_info *fs_info) 1246 { 1247 INIT_WORK(&fs_info->async_reclaim_work, btrfs_async_reclaim_metadata_space); 1248 INIT_WORK(&fs_info->async_data_reclaim_work, btrfs_async_reclaim_data_space); 1249 INIT_WORK(&fs_info->preempt_reclaim_work, 1250 btrfs_preempt_reclaim_metadata_space); 1251 } 1252 1253 static const enum btrfs_flush_state priority_flush_states[] = { 1254 FLUSH_DELAYED_ITEMS_NR, 1255 FLUSH_DELAYED_ITEMS, 1256 ALLOC_CHUNK, 1257 }; 1258 1259 static const enum btrfs_flush_state evict_flush_states[] = { 1260 FLUSH_DELAYED_ITEMS_NR, 1261 FLUSH_DELAYED_ITEMS, 1262 FLUSH_DELAYED_REFS_NR, 1263 FLUSH_DELAYED_REFS, 1264 FLUSH_DELALLOC, 1265 FLUSH_DELALLOC_WAIT, 1266 ALLOC_CHUNK, 1267 COMMIT_TRANS, 1268 }; 1269 1270 static void priority_reclaim_metadata_space(struct btrfs_fs_info *fs_info, 1271 struct btrfs_space_info *space_info, 1272 struct reserve_ticket *ticket, 1273 const enum btrfs_flush_state *states, 1274 int states_nr) 1275 { 1276 u64 to_reclaim; 1277 int flush_state; 1278 1279 spin_lock(&space_info->lock); 1280 to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info, space_info); 1281 if (!to_reclaim) { 1282 spin_unlock(&space_info->lock); 1283 return; 1284 } 1285 spin_unlock(&space_info->lock); 1286 1287 flush_state = 0; 1288 do { 1289 flush_space(fs_info, space_info, to_reclaim, states[flush_state]); 1290 flush_state++; 1291 spin_lock(&space_info->lock); 1292 if (ticket->bytes == 0) { 1293 spin_unlock(&space_info->lock); 1294 return; 1295 } 1296 spin_unlock(&space_info->lock); 1297 } while (flush_state < states_nr); 1298 } 1299 1300 static void priority_reclaim_data_space(struct btrfs_fs_info *fs_info, 1301 struct btrfs_space_info *space_info, 1302 struct reserve_ticket *ticket) 1303 { 1304 while (!space_info->full) { 1305 flush_space(fs_info, space_info, U64_MAX, ALLOC_CHUNK_FORCE); 1306 spin_lock(&space_info->lock); 1307 if (ticket->bytes == 0) { 1308 spin_unlock(&space_info->lock); 1309 return; 1310 } 1311 spin_unlock(&space_info->lock); 1312 } 1313 } 1314 1315 static void wait_reserve_ticket(struct btrfs_fs_info *fs_info, 1316 struct btrfs_space_info *space_info, 1317 struct reserve_ticket *ticket) 1318 1319 { 1320 DEFINE_WAIT(wait); 1321 int ret = 0; 1322 1323 spin_lock(&space_info->lock); 1324 while (ticket->bytes > 0 && ticket->error == 0) { 1325 ret = prepare_to_wait_event(&ticket->wait, &wait, TASK_KILLABLE); 1326 if (ret) { 1327 /* 1328 * Delete us from the list. After we unlock the space 1329 * info, we don't want the async reclaim job to reserve 1330 * space for this ticket. If that would happen, then the 1331 * ticket's task would not known that space was reserved 1332 * despite getting an error, resulting in a space leak 1333 * (bytes_may_use counter of our space_info). 1334 */ 1335 remove_ticket(space_info, ticket); 1336 ticket->error = -EINTR; 1337 break; 1338 } 1339 spin_unlock(&space_info->lock); 1340 1341 schedule(); 1342 1343 finish_wait(&ticket->wait, &wait); 1344 spin_lock(&space_info->lock); 1345 } 1346 spin_unlock(&space_info->lock); 1347 } 1348 1349 /** 1350 * Do the appropriate flushing and waiting for a ticket 1351 * 1352 * @fs_info: the filesystem 1353 * @space_info: space info for the reservation 1354 * @ticket: ticket for the reservation 1355 * @start_ns: timestamp when the reservation started 1356 * @orig_bytes: amount of bytes originally reserved 1357 * @flush: how much we can flush 1358 * 1359 * This does the work of figuring out how to flush for the ticket, waiting for 1360 * the reservation, and returning the appropriate error if there is one. 1361 */ 1362 static int handle_reserve_ticket(struct btrfs_fs_info *fs_info, 1363 struct btrfs_space_info *space_info, 1364 struct reserve_ticket *ticket, 1365 u64 start_ns, u64 orig_bytes, 1366 enum btrfs_reserve_flush_enum flush) 1367 { 1368 int ret; 1369 1370 switch (flush) { 1371 case BTRFS_RESERVE_FLUSH_DATA: 1372 case BTRFS_RESERVE_FLUSH_ALL: 1373 case BTRFS_RESERVE_FLUSH_ALL_STEAL: 1374 wait_reserve_ticket(fs_info, space_info, ticket); 1375 break; 1376 case BTRFS_RESERVE_FLUSH_LIMIT: 1377 priority_reclaim_metadata_space(fs_info, space_info, ticket, 1378 priority_flush_states, 1379 ARRAY_SIZE(priority_flush_states)); 1380 break; 1381 case BTRFS_RESERVE_FLUSH_EVICT: 1382 priority_reclaim_metadata_space(fs_info, space_info, ticket, 1383 evict_flush_states, 1384 ARRAY_SIZE(evict_flush_states)); 1385 break; 1386 case BTRFS_RESERVE_FLUSH_FREE_SPACE_INODE: 1387 priority_reclaim_data_space(fs_info, space_info, ticket); 1388 break; 1389 default: 1390 ASSERT(0); 1391 break; 1392 } 1393 1394 spin_lock(&space_info->lock); 1395 ret = ticket->error; 1396 if (ticket->bytes || ticket->error) { 1397 /* 1398 * We were a priority ticket, so we need to delete ourselves 1399 * from the list. Because we could have other priority tickets 1400 * behind us that require less space, run 1401 * btrfs_try_granting_tickets() to see if their reservations can 1402 * now be made. 1403 */ 1404 if (!list_empty(&ticket->list)) { 1405 remove_ticket(space_info, ticket); 1406 btrfs_try_granting_tickets(fs_info, space_info); 1407 } 1408 1409 if (!ret) 1410 ret = -ENOSPC; 1411 } 1412 spin_unlock(&space_info->lock); 1413 ASSERT(list_empty(&ticket->list)); 1414 /* 1415 * Check that we can't have an error set if the reservation succeeded, 1416 * as that would confuse tasks and lead them to error out without 1417 * releasing reserved space (if an error happens the expectation is that 1418 * space wasn't reserved at all). 1419 */ 1420 ASSERT(!(ticket->bytes == 0 && ticket->error)); 1421 trace_btrfs_reserve_ticket(fs_info, space_info->flags, orig_bytes, 1422 start_ns, flush, ticket->error); 1423 return ret; 1424 } 1425 1426 /* 1427 * This returns true if this flush state will go through the ordinary flushing 1428 * code. 1429 */ 1430 static inline bool is_normal_flushing(enum btrfs_reserve_flush_enum flush) 1431 { 1432 return (flush == BTRFS_RESERVE_FLUSH_ALL) || 1433 (flush == BTRFS_RESERVE_FLUSH_ALL_STEAL); 1434 } 1435 1436 /** 1437 * Try to reserve bytes from the block_rsv's space 1438 * 1439 * @fs_info: the filesystem 1440 * @space_info: space info we want to allocate from 1441 * @orig_bytes: number of bytes we want 1442 * @flush: whether or not we can flush to make our reservation 1443 * 1444 * This will reserve orig_bytes number of bytes from the space info associated 1445 * with the block_rsv. If there is not enough space it will make an attempt to 1446 * flush out space to make room. It will do this by flushing delalloc if 1447 * possible or committing the transaction. If flush is 0 then no attempts to 1448 * regain reservations will be made and this will fail if there is not enough 1449 * space already. 1450 */ 1451 static int __reserve_bytes(struct btrfs_fs_info *fs_info, 1452 struct btrfs_space_info *space_info, u64 orig_bytes, 1453 enum btrfs_reserve_flush_enum flush) 1454 { 1455 struct work_struct *async_work; 1456 struct reserve_ticket ticket; 1457 u64 start_ns = 0; 1458 u64 used; 1459 int ret = 0; 1460 bool pending_tickets; 1461 1462 ASSERT(orig_bytes); 1463 ASSERT(!current->journal_info || flush != BTRFS_RESERVE_FLUSH_ALL); 1464 1465 if (flush == BTRFS_RESERVE_FLUSH_DATA) 1466 async_work = &fs_info->async_data_reclaim_work; 1467 else 1468 async_work = &fs_info->async_reclaim_work; 1469 1470 spin_lock(&space_info->lock); 1471 ret = -ENOSPC; 1472 used = btrfs_space_info_used(space_info, true); 1473 1474 /* 1475 * We don't want NO_FLUSH allocations to jump everybody, they can 1476 * generally handle ENOSPC in a different way, so treat them the same as 1477 * normal flushers when it comes to skipping pending tickets. 1478 */ 1479 if (is_normal_flushing(flush) || (flush == BTRFS_RESERVE_NO_FLUSH)) 1480 pending_tickets = !list_empty(&space_info->tickets) || 1481 !list_empty(&space_info->priority_tickets); 1482 else 1483 pending_tickets = !list_empty(&space_info->priority_tickets); 1484 1485 /* 1486 * Carry on if we have enough space (short-circuit) OR call 1487 * can_overcommit() to ensure we can overcommit to continue. 1488 */ 1489 if (!pending_tickets && 1490 ((used + orig_bytes <= space_info->total_bytes) || 1491 btrfs_can_overcommit(fs_info, space_info, orig_bytes, flush))) { 1492 btrfs_space_info_update_bytes_may_use(fs_info, space_info, 1493 orig_bytes); 1494 ret = 0; 1495 } 1496 1497 /* 1498 * If we couldn't make a reservation then setup our reservation ticket 1499 * and kick the async worker if it's not already running. 1500 * 1501 * If we are a priority flusher then we just need to add our ticket to 1502 * the list and we will do our own flushing further down. 1503 */ 1504 if (ret && flush != BTRFS_RESERVE_NO_FLUSH) { 1505 ticket.bytes = orig_bytes; 1506 ticket.error = 0; 1507 space_info->reclaim_size += ticket.bytes; 1508 init_waitqueue_head(&ticket.wait); 1509 ticket.steal = (flush == BTRFS_RESERVE_FLUSH_ALL_STEAL); 1510 if (trace_btrfs_reserve_ticket_enabled()) 1511 start_ns = ktime_get_ns(); 1512 1513 if (flush == BTRFS_RESERVE_FLUSH_ALL || 1514 flush == BTRFS_RESERVE_FLUSH_ALL_STEAL || 1515 flush == BTRFS_RESERVE_FLUSH_DATA) { 1516 list_add_tail(&ticket.list, &space_info->tickets); 1517 if (!space_info->flush) { 1518 space_info->flush = 1; 1519 trace_btrfs_trigger_flush(fs_info, 1520 space_info->flags, 1521 orig_bytes, flush, 1522 "enospc"); 1523 queue_work(system_unbound_wq, async_work); 1524 } 1525 } else { 1526 list_add_tail(&ticket.list, 1527 &space_info->priority_tickets); 1528 } 1529 } else if (!ret && space_info->flags & BTRFS_BLOCK_GROUP_METADATA) { 1530 used += orig_bytes; 1531 /* 1532 * We will do the space reservation dance during log replay, 1533 * which means we won't have fs_info->fs_root set, so don't do 1534 * the async reclaim as we will panic. 1535 */ 1536 if (!test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags) && 1537 need_preemptive_reclaim(fs_info, space_info) && 1538 !work_busy(&fs_info->preempt_reclaim_work)) { 1539 trace_btrfs_trigger_flush(fs_info, space_info->flags, 1540 orig_bytes, flush, "preempt"); 1541 queue_work(system_unbound_wq, 1542 &fs_info->preempt_reclaim_work); 1543 } 1544 } 1545 spin_unlock(&space_info->lock); 1546 if (!ret || flush == BTRFS_RESERVE_NO_FLUSH) 1547 return ret; 1548 1549 return handle_reserve_ticket(fs_info, space_info, &ticket, start_ns, 1550 orig_bytes, flush); 1551 } 1552 1553 /** 1554 * Trye to reserve metadata bytes from the block_rsv's space 1555 * 1556 * @root: the root we're allocating for 1557 * @block_rsv: block_rsv we're allocating for 1558 * @orig_bytes: number of bytes we want 1559 * @flush: whether or not we can flush to make our reservation 1560 * 1561 * This will reserve orig_bytes number of bytes from the space info associated 1562 * with the block_rsv. If there is not enough space it will make an attempt to 1563 * flush out space to make room. It will do this by flushing delalloc if 1564 * possible or committing the transaction. If flush is 0 then no attempts to 1565 * regain reservations will be made and this will fail if there is not enough 1566 * space already. 1567 */ 1568 int btrfs_reserve_metadata_bytes(struct btrfs_root *root, 1569 struct btrfs_block_rsv *block_rsv, 1570 u64 orig_bytes, 1571 enum btrfs_reserve_flush_enum flush) 1572 { 1573 struct btrfs_fs_info *fs_info = root->fs_info; 1574 struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv; 1575 int ret; 1576 1577 ret = __reserve_bytes(fs_info, block_rsv->space_info, orig_bytes, flush); 1578 if (ret == -ENOSPC && 1579 unlikely(root->orphan_cleanup_state == ORPHAN_CLEANUP_STARTED)) { 1580 if (block_rsv != global_rsv && 1581 !btrfs_block_rsv_use_bytes(global_rsv, orig_bytes)) 1582 ret = 0; 1583 } 1584 if (ret == -ENOSPC) { 1585 trace_btrfs_space_reservation(fs_info, "space_info:enospc", 1586 block_rsv->space_info->flags, 1587 orig_bytes, 1); 1588 1589 if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) 1590 btrfs_dump_space_info(fs_info, block_rsv->space_info, 1591 orig_bytes, 0); 1592 } 1593 return ret; 1594 } 1595 1596 /** 1597 * Try to reserve data bytes for an allocation 1598 * 1599 * @fs_info: the filesystem 1600 * @bytes: number of bytes we need 1601 * @flush: how we are allowed to flush 1602 * 1603 * This will reserve bytes from the data space info. If there is not enough 1604 * space then we will attempt to flush space as specified by flush. 1605 */ 1606 int btrfs_reserve_data_bytes(struct btrfs_fs_info *fs_info, u64 bytes, 1607 enum btrfs_reserve_flush_enum flush) 1608 { 1609 struct btrfs_space_info *data_sinfo = fs_info->data_sinfo; 1610 int ret; 1611 1612 ASSERT(flush == BTRFS_RESERVE_FLUSH_DATA || 1613 flush == BTRFS_RESERVE_FLUSH_FREE_SPACE_INODE); 1614 ASSERT(!current->journal_info || flush != BTRFS_RESERVE_FLUSH_DATA); 1615 1616 ret = __reserve_bytes(fs_info, data_sinfo, bytes, flush); 1617 if (ret == -ENOSPC) { 1618 trace_btrfs_space_reservation(fs_info, "space_info:enospc", 1619 data_sinfo->flags, bytes, 1); 1620 if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) 1621 btrfs_dump_space_info(fs_info, data_sinfo, bytes, 0); 1622 } 1623 return ret; 1624 } 1625