git.sesse.net Git - bcachefs-tools-debian/blob - libbcachefs/journal_reclaim.c

   1 // SPDX-License-Identifier: GPL-2.0
   2
   3 #include "bcachefs.h"
   4 #include "btree_key_cache.h"
   5 #include "error.h"
   6 #include "journal.h"
   7 #include "journal_io.h"
   8 #include "journal_reclaim.h"
   9 #include "replicas.h"
  10 #include "super.h"
  11
  12 #include <linux/kthread.h>
  13 #include <linux/sched/mm.h>
  14 #include <trace/events/bcachefs.h>
  15
  16 /* Free space calculations: */
  17
  18 static unsigned journal_space_from(struct journal_device *ja,
  19                                    enum journal_space_from from)
  20 {
  21         switch (from) {
  22         case journal_space_discarded:
  23                 return ja->discard_idx;
  24         case journal_space_clean_ondisk:
  25                 return ja->dirty_idx_ondisk;
  26         case journal_space_clean:
  27                 return ja->dirty_idx;
  28         default:
  29                 BUG();
  30         }
  31 }
  32
  33 unsigned bch2_journal_dev_buckets_available(struct journal *j,
  34                                             struct journal_device *ja,
  35                                             enum journal_space_from from)
  36 {
  37         unsigned available = !test_bit(JOURNAL_NOCHANGES, &j->flags)
  38                 ? ((journal_space_from(ja, from) -
  39                     ja->cur_idx - 1 + ja->nr) % ja->nr)
  40                 : ja->nr;
  41
  42         /*
  43          * Don't use the last bucket unless writing the new last_seq
  44          * will make another bucket available:
  45          */
  46         if (available && ja->dirty_idx_ondisk == ja->dirty_idx)
  47                 --available;
  48
  49         return available;
  50 }
  51
  52 static void journal_set_remaining(struct journal *j, unsigned u64s_remaining)
  53 {
  54         union journal_preres_state old, new;
  55         u64 v = atomic64_read(&j->prereserved.counter);
  56
  57         do {
  58                 old.v = new.v = v;
  59                 new.remaining = u64s_remaining;
  60         } while ((v = atomic64_cmpxchg(&j->prereserved.counter,
  61                                        old.v, new.v)) != old.v);
  62 }
  63
  64 static inline unsigned get_unwritten_sectors(struct journal *j, unsigned *idx)
  65 {
  66         unsigned sectors = 0;
  67
  68         while (!sectors && *idx != j->reservations.idx) {
  69                 sectors = j->buf[*idx].sectors;
  70
  71                 *idx = (*idx + 1) & JOURNAL_BUF_MASK;
  72         }
  73
  74         return sectors;
  75 }
  76
  77 static struct journal_space
  78 journal_dev_space_available(struct journal *j, struct bch_dev *ca,
  79                             enum journal_space_from from)
  80 {
  81         struct journal_device *ja = &ca->journal;
  82         unsigned sectors, buckets, unwritten, idx = j->reservations.unwritten_idx;
  83
  84         if (from == journal_space_total)
  85                 return (struct journal_space) {
  86                         .next_entry     = ca->mi.bucket_size,
  87                         .total          = ca->mi.bucket_size * ja->nr,
  88                 };
  89
  90         buckets = bch2_journal_dev_buckets_available(j, ja, from);
  91         sectors = ja->sectors_free;
  92
  93         /*
  94          * We that we don't allocate the space for a journal entry
  95          * until we write it out - thus, account for it here:
  96          */
  97         while ((unwritten = get_unwritten_sectors(j, &idx))) {
  98                 /* entry won't fit on this device, skip: */
  99                 if (unwritten > ca->mi.bucket_size)
 100                         continue;
 101
 102                 if (unwritten >= sectors) {
 103                         if (!buckets) {
 104                                 sectors = 0;
 105                                 break;
 106                         }
 107
 108                         buckets--;
 109                         sectors = ca->mi.bucket_size;
 110                 }
 111
 112                 sectors -= unwritten;
 113         }
 114
 115         if (sectors < ca->mi.bucket_size && buckets) {
 116                 buckets--;
 117                 sectors = ca->mi.bucket_size;
 118         }
 119
 120         return (struct journal_space) {
 121                 .next_entry     = sectors,
 122                 .total          = sectors + buckets * ca->mi.bucket_size,
 123         };
 124 }
 125
 126 static struct journal_space __journal_space_available(struct journal *j, unsigned nr_devs_want,
 127                             enum journal_space_from from)
 128 {
 129         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 130         struct bch_dev *ca;
 131         unsigned i, pos, nr_devs = 0;
 132         struct journal_space space, dev_space[BCH_SB_MEMBERS_MAX];
 133
 134         BUG_ON(nr_devs_want > ARRAY_SIZE(dev_space));
 135
 136         rcu_read_lock();
 137         for_each_member_device_rcu(ca, c, i,
 138                                    &c->rw_devs[BCH_DATA_journal]) {
 139                 if (!ca->journal.nr)
 140                         continue;
 141
 142                 space = journal_dev_space_available(j, ca, from);
 143                 if (!space.next_entry)
 144                         continue;
 145
 146                 for (pos = 0; pos < nr_devs; pos++)
 147                         if (space.total > dev_space[pos].total)
 148                                 break;
 149
 150                 array_insert_item(dev_space, nr_devs, pos, space);
 151         }
 152         rcu_read_unlock();
 153
 154         if (nr_devs < nr_devs_want)
 155                 return (struct journal_space) { 0, 0 };
 156
 157         /*
 158          * We sorted largest to smallest, and we want the smallest out of the
 159          * @nr_devs_want largest devices:
 160          */
 161         return dev_space[nr_devs_want - 1];
 162 }
 163
 164 void bch2_journal_space_available(struct journal *j)
 165 {
 166         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 167         struct bch_dev *ca;
 168         unsigned clean, clean_ondisk, total;
 169         s64 u64s_remaining = 0;
 170         unsigned max_entry_size  = min(j->buf[0].buf_size >> 9,
 171                                        j->buf[1].buf_size >> 9);
 172         unsigned i, nr_online = 0, nr_devs_want;
 173         bool can_discard = false;
 174         int ret = 0;
 175
 176         lockdep_assert_held(&j->lock);
 177
 178         rcu_read_lock();
 179         for_each_member_device_rcu(ca, c, i,
 180                                    &c->rw_devs[BCH_DATA_journal]) {
 181                 struct journal_device *ja = &ca->journal;
 182
 183                 if (!ja->nr)
 184                         continue;
 185
 186                 while (ja->dirty_idx != ja->cur_idx &&
 187                        ja->bucket_seq[ja->dirty_idx] < journal_last_seq(j))
 188                         ja->dirty_idx = (ja->dirty_idx + 1) % ja->nr;
 189
 190                 while (ja->dirty_idx_ondisk != ja->dirty_idx &&
 191                        ja->bucket_seq[ja->dirty_idx_ondisk] < j->last_seq_ondisk)
 192                         ja->dirty_idx_ondisk = (ja->dirty_idx_ondisk + 1) % ja->nr;
 193
 194                 if (ja->discard_idx != ja->dirty_idx_ondisk)
 195                         can_discard = true;
 196
 197                 max_entry_size = min_t(unsigned, max_entry_size, ca->mi.bucket_size);
 198                 nr_online++;
 199         }
 200         rcu_read_unlock();
 201
 202         j->can_discard = can_discard;
 203
 204         if (nr_online < c->opts.metadata_replicas_required) {
 205                 ret = cur_entry_insufficient_devices;
 206                 goto out;
 207         }
 208
 209         nr_devs_want = min_t(unsigned, nr_online, c->opts.metadata_replicas);
 210
 211         for (i = 0; i < journal_space_nr; i++)
 212                 j->space[i] = __journal_space_available(j, nr_devs_want, i);
 213
 214         clean_ondisk    = j->space[journal_space_clean_ondisk].total;
 215         clean           = j->space[journal_space_clean].total;
 216         total           = j->space[journal_space_total].total;
 217
 218         if (!clean_ondisk &&
 219             j->reservations.idx ==
 220             j->reservations.unwritten_idx) {
 221                 char *buf = kmalloc(4096, GFP_ATOMIC);
 222
 223                 bch_err(c, "journal stuck");
 224                 if (buf) {
 225                         __bch2_journal_debug_to_text(&_PBUF(buf, 4096), j);
 226                         pr_err("\n%s", buf);
 227                         kfree(buf);
 228                 }
 229
 230                 bch2_fatal_error(c);
 231                 ret = cur_entry_journal_stuck;
 232         } else if (!j->space[journal_space_discarded].next_entry)
 233                 ret = cur_entry_journal_full;
 234         else if (!fifo_free(&j->pin))
 235                 ret = cur_entry_journal_pin_full;
 236
 237         if ((j->space[journal_space_clean_ondisk].next_entry <
 238              j->space[journal_space_clean_ondisk].total) &&
 239             (clean - clean_ondisk <= total / 8) &&
 240             (clean_ondisk * 2 > clean ))
 241                 set_bit(JOURNAL_MAY_SKIP_FLUSH, &j->flags);
 242         else
 243                 clear_bit(JOURNAL_MAY_SKIP_FLUSH, &j->flags);
 244
 245         u64s_remaining  = (u64) clean << 6;
 246         u64s_remaining -= (u64) total << 3;
 247         u64s_remaining = max(0LL, u64s_remaining);
 248         u64s_remaining /= 4;
 249         u64s_remaining = min_t(u64, u64s_remaining, U32_MAX);
 250 out:
 251         j->cur_entry_sectors    = !ret ? j->space[journal_space_discarded].next_entry : 0;
 252         j->cur_entry_error      = ret;
 253         journal_set_remaining(j, u64s_remaining);
 254         journal_check_may_get_unreserved(j);
 255
 256         if (!ret)
 257                 journal_wake(j);
 258 }
 259
 260 /* Discards - last part of journal reclaim: */
 261
 262 static bool should_discard_bucket(struct journal *j, struct journal_device *ja)
 263 {
 264         bool ret;
 265
 266         spin_lock(&j->lock);
 267         ret = ja->discard_idx != ja->dirty_idx_ondisk;
 268         spin_unlock(&j->lock);
 269
 270         return ret;
 271 }
 272
 273 /*
 274  * Advance ja->discard_idx as long as it points to buckets that are no longer
 275  * dirty, issuing discards if necessary:
 276  */
 277 void bch2_journal_do_discards(struct journal *j)
 278 {
 279         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 280         struct bch_dev *ca;
 281         unsigned iter;
 282
 283         mutex_lock(&j->discard_lock);
 284
 285         for_each_rw_member(ca, c, iter) {
 286                 struct journal_device *ja = &ca->journal;
 287
 288                 while (should_discard_bucket(j, ja)) {
 289                         if (ca->mi.discard &&
 290                             blk_queue_discard(bdev_get_queue(ca->disk_sb.bdev)))
 291                                 blkdev_issue_discard(ca->disk_sb.bdev,
 292                                         bucket_to_sector(ca,
 293                                                 ja->buckets[ja->discard_idx]),
 294                                         ca->mi.bucket_size, GFP_NOIO, 0);
 295
 296                         spin_lock(&j->lock);
 297                         ja->discard_idx = (ja->discard_idx + 1) % ja->nr;
 298
 299                         bch2_journal_space_available(j);
 300                         spin_unlock(&j->lock);
 301                 }
 302         }
 303
 304         mutex_unlock(&j->discard_lock);
 305 }
 306
 307 /*
 308  * Journal entry pinning - machinery for holding a reference on a given journal
 309  * entry, holding it open to ensure it gets replayed during recovery:
 310  */
 311
 312 static void bch2_journal_reclaim_fast(struct journal *j)
 313 {
 314         struct journal_entry_pin_list temp;
 315         bool popped = false;
 316
 317         lockdep_assert_held(&j->lock);
 318
 319         /*
 320          * Unpin journal entries whose reference counts reached zero, meaning
 321          * all btree nodes got written out
 322          */
 323         while (!fifo_empty(&j->pin) &&
 324                !atomic_read(&fifo_peek_front(&j->pin).count)) {
 325                 BUG_ON(!list_empty(&fifo_peek_front(&j->pin).list));
 326                 BUG_ON(!list_empty(&fifo_peek_front(&j->pin).flushed));
 327                 BUG_ON(!fifo_pop(&j->pin, temp));
 328                 popped = true;
 329         }
 330
 331         if (popped)
 332                 bch2_journal_space_available(j);
 333 }
 334
 335 void __bch2_journal_pin_put(struct journal *j, u64 seq)
 336 {
 337         struct journal_entry_pin_list *pin_list = journal_seq_pin(j, seq);
 338
 339         if (atomic_dec_and_test(&pin_list->count))
 340                 bch2_journal_reclaim_fast(j);
 341 }
 342
 343 void bch2_journal_pin_put(struct journal *j, u64 seq)
 344 {
 345         struct journal_entry_pin_list *pin_list = journal_seq_pin(j, seq);
 346
 347         if (atomic_dec_and_test(&pin_list->count)) {
 348                 spin_lock(&j->lock);
 349                 bch2_journal_reclaim_fast(j);
 350                 spin_unlock(&j->lock);
 351         }
 352 }
 353
 354 static inline void __journal_pin_drop(struct journal *j,
 355                                       struct journal_entry_pin *pin)
 356 {
 357         struct journal_entry_pin_list *pin_list;
 358
 359         if (!journal_pin_active(pin))
 360                 return;
 361
 362         if (j->flush_in_progress == pin)
 363                 j->flush_in_progress_dropped = true;
 364
 365         pin_list = journal_seq_pin(j, pin->seq);
 366         pin->seq = 0;
 367         list_del_init(&pin->list);
 368
 369         /*
 370          * Unpinning a journal entry make make journal_next_bucket() succeed, if
 371          * writing a new last_seq will now make another bucket available:
 372          */
 373         if (atomic_dec_and_test(&pin_list->count) &&
 374             pin_list == &fifo_peek_front(&j->pin))
 375                 bch2_journal_reclaim_fast(j);
 376         else if (fifo_used(&j->pin) == 1 &&
 377                  atomic_read(&pin_list->count) == 1)
 378                 journal_wake(j);
 379 }
 380
 381 void bch2_journal_pin_drop(struct journal *j,
 382                            struct journal_entry_pin *pin)
 383 {
 384         spin_lock(&j->lock);
 385         __journal_pin_drop(j, pin);
 386         spin_unlock(&j->lock);
 387 }
 388
 389 void bch2_journal_pin_set(struct journal *j, u64 seq,
 390                           struct journal_entry_pin *pin,
 391                           journal_pin_flush_fn flush_fn)
 392 {
 393         struct journal_entry_pin_list *pin_list;
 394
 395         spin_lock(&j->lock);
 396
 397         if (seq < journal_last_seq(j)) {
 398                 /*
 399                  * bch2_journal_pin_copy() raced with bch2_journal_pin_drop() on
 400                  * the src pin - with the pin dropped, the entry to pin might no
 401                  * longer to exist, but that means there's no longer anything to
 402                  * copy and we can bail out here:
 403                  */
 404                 spin_unlock(&j->lock);
 405                 return;
 406         }
 407
 408         pin_list = journal_seq_pin(j, seq);
 409
 410         __journal_pin_drop(j, pin);
 411
 412         atomic_inc(&pin_list->count);
 413         pin->seq        = seq;
 414         pin->flush      = flush_fn;
 415
 416         if (flush_fn == bch2_btree_key_cache_journal_flush)
 417                 list_add(&pin->list, &pin_list->key_cache_list);
 418         else if (flush_fn)
 419                 list_add(&pin->list, &pin_list->list);
 420         else
 421                 list_add(&pin->list, &pin_list->flushed);
 422         spin_unlock(&j->lock);
 423
 424         /*
 425          * If the journal is currently full,  we might want to call flush_fn
 426          * immediately:
 427          */
 428         journal_wake(j);
 429 }
 430
 431 /**
 432  * bch2_journal_pin_flush: ensure journal pin callback is no longer running
 433  */
 434 void bch2_journal_pin_flush(struct journal *j, struct journal_entry_pin *pin)
 435 {
 436         BUG_ON(journal_pin_active(pin));
 437
 438         wait_event(j->pin_flush_wait, j->flush_in_progress != pin);
 439 }
 440
 441 /*
 442  * Journal reclaim: flush references to open journal entries to reclaim space in
 443  * the journal
 444  *
 445  * May be done by the journal code in the background as needed to free up space
 446  * for more journal entries, or as part of doing a clean shutdown, or to migrate
 447  * data off of a specific device:
 448  */
 449
 450 static struct journal_entry_pin *
 451 journal_get_next_pin(struct journal *j,
 452                      bool get_any,
 453                      bool get_key_cache,
 454                      u64 max_seq, u64 *seq)
 455 {
 456         struct journal_entry_pin_list *pin_list;
 457         struct journal_entry_pin *ret = NULL;
 458
 459         fifo_for_each_entry_ptr(pin_list, &j->pin, *seq) {
 460                 if (*seq > max_seq && !get_any && !get_key_cache)
 461                         break;
 462
 463                 if (*seq <= max_seq || get_any) {
 464                         ret = list_first_entry_or_null(&pin_list->list,
 465                                 struct journal_entry_pin, list);
 466                         if (ret)
 467                                 return ret;
 468                 }
 469
 470                 if (*seq <= max_seq || get_any || get_key_cache) {
 471                         ret = list_first_entry_or_null(&pin_list->key_cache_list,
 472                                 struct journal_entry_pin, list);
 473                         if (ret)
 474                                 return ret;
 475                 }
 476         }
 477
 478         return NULL;
 479 }
 480
 481 /* returns true if we did work */
 482 static size_t journal_flush_pins(struct journal *j, u64 seq_to_flush,
 483                                  unsigned min_any,
 484                                  unsigned min_key_cache)
 485 {
 486         struct journal_entry_pin *pin;
 487         size_t nr_flushed = 0;
 488         journal_pin_flush_fn flush_fn;
 489         u64 seq;
 490         int err;
 491
 492         if (!test_bit(JOURNAL_RECLAIM_STARTED, &j->flags))
 493                 return 0;
 494
 495         lockdep_assert_held(&j->reclaim_lock);
 496
 497         while (1) {
 498                 cond_resched();
 499
 500                 j->last_flushed = jiffies;
 501
 502                 spin_lock(&j->lock);
 503                 pin = journal_get_next_pin(j,
 504                                            min_any != 0,
 505                                            min_key_cache != 0,
 506                                            seq_to_flush, &seq);
 507                 if (pin) {
 508                         BUG_ON(j->flush_in_progress);
 509                         j->flush_in_progress = pin;
 510                         j->flush_in_progress_dropped = false;
 511                         flush_fn = pin->flush;
 512                 }
 513                 spin_unlock(&j->lock);
 514
 515                 if (!pin)
 516                         break;
 517
 518                 if (min_key_cache && pin->flush == bch2_btree_key_cache_journal_flush)
 519                         min_key_cache--;
 520
 521                 if (min_any)
 522                         min_any--;
 523
 524                 err = flush_fn(j, pin, seq);
 525
 526                 spin_lock(&j->lock);
 527                 /* Pin might have been dropped or rearmed: */
 528                 if (likely(!err && !j->flush_in_progress_dropped))
 529                         list_move(&pin->list, &journal_seq_pin(j, seq)->flushed);
 530                 j->flush_in_progress = NULL;
 531                 j->flush_in_progress_dropped = false;
 532                 spin_unlock(&j->lock);
 533
 534                 wake_up(&j->pin_flush_wait);
 535
 536                 if (err)
 537                         break;
 538
 539                 nr_flushed++;
 540         }
 541
 542         return nr_flushed;
 543 }
 544
 545 static u64 journal_seq_to_flush(struct journal *j)
 546 {
 547         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 548         struct bch_dev *ca;
 549         u64 seq_to_flush = 0;
 550         unsigned iter;
 551
 552         spin_lock(&j->lock);
 553
 554         for_each_rw_member(ca, c, iter) {
 555                 struct journal_device *ja = &ca->journal;
 556                 unsigned nr_buckets, bucket_to_flush;
 557
 558                 if (!ja->nr)
 559                         continue;
 560
 561                 /* Try to keep the journal at most half full: */
 562                 nr_buckets = ja->nr / 2;
 563
 564                 /* And include pre-reservations: */
 565                 nr_buckets += DIV_ROUND_UP(j->prereserved.reserved,
 566                                            (ca->mi.bucket_size << 6) -
 567                                            journal_entry_overhead(j));
 568
 569                 nr_buckets = min(nr_buckets, ja->nr);
 570
 571                 bucket_to_flush = (ja->cur_idx + nr_buckets) % ja->nr;
 572                 seq_to_flush = max(seq_to_flush,
 573                                    ja->bucket_seq[bucket_to_flush]);
 574         }
 575
 576         /* Also flush if the pin fifo is more than half full */
 577         seq_to_flush = max_t(s64, seq_to_flush,
 578                              (s64) journal_cur_seq(j) -
 579                              (j->pin.size >> 1));
 580         spin_unlock(&j->lock);
 581
 582         return seq_to_flush;
 583 }
 584
 585 /**
 586  * bch2_journal_reclaim - free up journal buckets
 587  *
 588  * Background journal reclaim writes out btree nodes. It should be run
 589  * early enough so that we never completely run out of journal buckets.
 590  *
 591  * High watermarks for triggering background reclaim:
 592  * - FIFO has fewer than 512 entries left
 593  * - fewer than 25% journal buckets free
 594  *
 595  * Background reclaim runs until low watermarks are reached:
 596  * - FIFO has more than 1024 entries left
 597  * - more than 50% journal buckets free
 598  *
 599  * As long as a reclaim can complete in the time it takes to fill up
 600  * 512 journal entries or 25% of all journal buckets, then
 601  * journal_next_bucket() should not stall.
 602  */
 603 static int __bch2_journal_reclaim(struct journal *j, bool direct)
 604 {
 605         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 606         bool kthread = (current->flags & PF_KTHREAD) != 0;
 607         u64 seq_to_flush;
 608         size_t min_nr, min_key_cache, nr_flushed;
 609         unsigned flags;
 610         int ret = 0;
 611
 612         /*
 613          * We can't invoke memory reclaim while holding the reclaim_lock -
 614          * journal reclaim is required to make progress for memory reclaim
 615          * (cleaning the caches), so we can't get stuck in memory reclaim while
 616          * we're holding the reclaim lock:
 617          */
 618         lockdep_assert_held(&j->reclaim_lock);
 619         flags = memalloc_noreclaim_save();
 620
 621         do {
 622                 if (kthread && kthread_should_stop())
 623                         break;
 624
 625                 if (bch2_journal_error(j)) {
 626                         ret = -EIO;
 627                         break;
 628                 }
 629
 630                 bch2_journal_do_discards(j);
 631
 632                 seq_to_flush = journal_seq_to_flush(j);
 633                 min_nr = 0;
 634
 635                 /*
 636                  * If it's been longer than j->reclaim_delay_ms since we last flushed,
 637                  * make sure to flush at least one journal pin:
 638                  */
 639                 if (time_after(jiffies, j->last_flushed +
 640                                msecs_to_jiffies(j->reclaim_delay_ms)))
 641                         min_nr = 1;
 642
 643                 if (j->prereserved.reserved * 4 > j->prereserved.remaining)
 644                         min_nr = 1;
 645
 646                 if (fifo_free(&j->pin) <= 32)
 647                         min_nr = 1;
 648
 649                 trace_journal_reclaim_start(c,
 650                                 min_nr,
 651                                 j->prereserved.reserved,
 652                                 j->prereserved.remaining,
 653                                 atomic_read(&c->btree_cache.dirty),
 654                                 c->btree_cache.used,
 655                                 atomic_long_read(&c->btree_key_cache.nr_dirty),
 656                                 atomic_long_read(&c->btree_key_cache.nr_keys));
 657
 658                 min_key_cache = min(bch2_nr_btree_keys_need_flush(c), (size_t) 128);
 659
 660                 nr_flushed = journal_flush_pins(j, seq_to_flush,
 661                                                 min_nr, min_key_cache);
 662
 663                 if (direct)
 664                         j->nr_direct_reclaim += nr_flushed;
 665                 else
 666                         j->nr_background_reclaim += nr_flushed;
 667                 trace_journal_reclaim_finish(c, nr_flushed);
 668
 669                 if (nr_flushed)
 670                         wake_up(&j->reclaim_wait);
 671         } while ((min_nr || min_key_cache) && !direct);
 672
 673         memalloc_noreclaim_restore(flags);
 674
 675         return ret;
 676 }
 677
 678 int bch2_journal_reclaim(struct journal *j)
 679 {
 680         return __bch2_journal_reclaim(j, true);
 681 }
 682
 683 static int bch2_journal_reclaim_thread(void *arg)
 684 {
 685         struct journal *j = arg;
 686         unsigned long delay, now;
 687         int ret = 0;
 688
 689         set_freezable();
 690
 691         kthread_wait_freezable(test_bit(JOURNAL_RECLAIM_STARTED, &j->flags));
 692
 693         j->last_flushed = jiffies;
 694
 695         while (!ret && !kthread_should_stop()) {
 696                 j->reclaim_kicked = false;
 697
 698                 mutex_lock(&j->reclaim_lock);
 699                 ret = __bch2_journal_reclaim(j, false);
 700                 mutex_unlock(&j->reclaim_lock);
 701
 702                 now = jiffies;
 703                 delay = msecs_to_jiffies(j->reclaim_delay_ms);
 704                 j->next_reclaim = j->last_flushed + delay;
 705
 706                 if (!time_in_range(j->next_reclaim, now, now + delay))
 707                         j->next_reclaim = now + delay;
 708
 709                 while (1) {
 710                         set_current_state(TASK_INTERRUPTIBLE);
 711                         if (kthread_should_stop())
 712                                 break;
 713                         if (j->reclaim_kicked)
 714                                 break;
 715                         if (time_after_eq(jiffies, j->next_reclaim))
 716                                 break;
 717                         freezable_schedule_timeout(j->next_reclaim - jiffies);
 718
 719                 }
 720                 __set_current_state(TASK_RUNNING);
 721         }
 722
 723         return 0;
 724 }
 725
 726 void bch2_journal_reclaim_stop(struct journal *j)
 727 {
 728         struct task_struct *p = j->reclaim_thread;
 729
 730         j->reclaim_thread = NULL;
 731
 732         if (p) {
 733                 kthread_stop(p);
 734                 put_task_struct(p);
 735         }
 736 }
 737
 738 int bch2_journal_reclaim_start(struct journal *j)
 739 {
 740         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 741         struct task_struct *p;
 742
 743         if (j->reclaim_thread)
 744                 return 0;
 745
 746         p = kthread_create(bch2_journal_reclaim_thread, j,
 747                            "bch-reclaim/%s", c->name);
 748         if (IS_ERR(p)) {
 749                 bch_err(c, "error creating journal reclaim thread: %li", PTR_ERR(p));
 750                 return PTR_ERR(p);
 751         }
 752
 753         get_task_struct(p);
 754         j->reclaim_thread = p;
 755         wake_up_process(p);
 756         return 0;
 757 }
 758
 759 static int journal_flush_done(struct journal *j, u64 seq_to_flush,
 760                               bool *did_work)
 761 {
 762         int ret;
 763
 764         ret = bch2_journal_error(j);
 765         if (ret)
 766                 return ret;
 767
 768         mutex_lock(&j->reclaim_lock);
 769
 770         *did_work = journal_flush_pins(j, seq_to_flush, 0, 0) != 0;
 771
 772         spin_lock(&j->lock);
 773         /*
 774          * If journal replay hasn't completed, the unreplayed journal entries
 775          * hold refs on their corresponding sequence numbers
 776          */
 777         ret = !test_bit(JOURNAL_REPLAY_DONE, &j->flags) ||
 778                 journal_last_seq(j) > seq_to_flush ||
 779                 (fifo_used(&j->pin) == 1 &&
 780                  atomic_read(&fifo_peek_front(&j->pin).count) == 1);
 781
 782         spin_unlock(&j->lock);
 783         mutex_unlock(&j->reclaim_lock);
 784
 785         return ret;
 786 }
 787
 788 bool bch2_journal_flush_pins(struct journal *j, u64 seq_to_flush)
 789 {
 790         bool did_work = false;
 791
 792         if (!test_bit(JOURNAL_STARTED, &j->flags))
 793                 return false;
 794
 795         closure_wait_event(&j->async_wait,
 796                 journal_flush_done(j, seq_to_flush, &did_work));
 797
 798         return did_work;
 799 }
 800
 801 int bch2_journal_flush_device_pins(struct journal *j, int dev_idx)
 802 {
 803         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 804         struct journal_entry_pin_list *p;
 805         u64 iter, seq = 0;
 806         int ret = 0;
 807
 808         spin_lock(&j->lock);
 809         fifo_for_each_entry_ptr(p, &j->pin, iter)
 810                 if (dev_idx >= 0
 811                     ? bch2_dev_list_has_dev(p->devs, dev_idx)
 812                     : p->devs.nr < c->opts.metadata_replicas)
 813                         seq = iter;
 814         spin_unlock(&j->lock);
 815
 816         bch2_journal_flush_pins(j, seq);
 817
 818         ret = bch2_journal_error(j);
 819         if (ret)
 820                 return ret;
 821
 822         mutex_lock(&c->replicas_gc_lock);
 823         bch2_replicas_gc_start(c, 1 << BCH_DATA_journal);
 824
 825         seq = 0;
 826
 827         spin_lock(&j->lock);
 828         while (!ret && seq < j->pin.back) {
 829                 struct bch_replicas_padded replicas;
 830
 831                 seq = max(seq, journal_last_seq(j));
 832                 bch2_devlist_to_replicas(&replicas.e, BCH_DATA_journal,
 833                                          journal_seq_pin(j, seq)->devs);
 834                 seq++;
 835
 836                 spin_unlock(&j->lock);
 837                 ret = bch2_mark_replicas(c, &replicas.e);
 838                 spin_lock(&j->lock);
 839         }
 840         spin_unlock(&j->lock);
 841
 842         ret = bch2_replicas_gc_end(c, ret);
 843         mutex_unlock(&c->replicas_gc_lock);
 844
 845         return ret;
 846 }