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 "errcode.h"
   6 #include "error.h"
   7 #include "journal.h"
   8 #include "journal_io.h"
   9 #include "journal_reclaim.h"
  10 #include "replicas.h"
  11 #include "super.h"
  12
  13 #include <linux/kthread.h>
  14 #include <linux/sched/mm.h>
  15 #include <trace/events/bcachefs.h>
  16
  17 /* Free space calculations: */
  18
  19 static unsigned journal_space_from(struct journal_device *ja,
  20                                    enum journal_space_from from)
  21 {
  22         switch (from) {
  23         case journal_space_discarded:
  24                 return ja->discard_idx;
  25         case journal_space_clean_ondisk:
  26                 return ja->dirty_idx_ondisk;
  27         case journal_space_clean:
  28                 return ja->dirty_idx;
  29         default:
  30                 BUG();
  31         }
  32 }
  33
  34 unsigned bch2_journal_dev_buckets_available(struct journal *j,
  35                                             struct journal_device *ja,
  36                                             enum journal_space_from from)
  37 {
  38         unsigned available = (journal_space_from(ja, from) -
  39                               ja->cur_idx - 1 + ja->nr) % ja->nr;
  40
  41         /*
  42          * Don't use the last bucket unless writing the new last_seq
  43          * will make another bucket available:
  44          */
  45         if (available && ja->dirty_idx_ondisk == ja->dirty_idx)
  46                 --available;
  47
  48         return available;
  49 }
  50
  51 static void journal_set_remaining(struct journal *j, unsigned u64s_remaining)
  52 {
  53         union journal_preres_state old, new;
  54         u64 v = atomic64_read(&j->prereserved.counter);
  55
  56         do {
  57                 old.v = new.v = v;
  58                 new.remaining = u64s_remaining;
  59         } while ((v = atomic64_cmpxchg(&j->prereserved.counter,
  60                                        old.v, new.v)) != old.v);
  61 }
  62
  63 static struct journal_space
  64 journal_dev_space_available(struct journal *j, struct bch_dev *ca,
  65                             enum journal_space_from from)
  66 {
  67         struct journal_device *ja = &ca->journal;
  68         unsigned sectors, buckets, unwritten;
  69         u64 seq;
  70
  71         if (from == journal_space_total)
  72                 return (struct journal_space) {
  73                         .next_entry     = ca->mi.bucket_size,
  74                         .total          = ca->mi.bucket_size * ja->nr,
  75                 };
  76
  77         buckets = bch2_journal_dev_buckets_available(j, ja, from);
  78         sectors = ja->sectors_free;
  79
  80         /*
  81          * We that we don't allocate the space for a journal entry
  82          * until we write it out - thus, account for it here:
  83          */
  84         for (seq = journal_last_unwritten_seq(j);
  85              seq <= journal_cur_seq(j);
  86              seq++) {
  87                 unwritten = j->buf[seq & JOURNAL_BUF_MASK].sectors;
  88
  89                 if (!unwritten)
  90                         continue;
  91
  92                 /* entry won't fit on this device, skip: */
  93                 if (unwritten > ca->mi.bucket_size)
  94                         continue;
  95
  96                 if (unwritten >= sectors) {
  97                         if (!buckets) {
  98                                 sectors = 0;
  99                                 break;
 100                         }
 101
 102                         buckets--;
 103                         sectors = ca->mi.bucket_size;
 104                 }
 105
 106                 sectors -= unwritten;
 107         }
 108
 109         if (sectors < ca->mi.bucket_size && buckets) {
 110                 buckets--;
 111                 sectors = ca->mi.bucket_size;
 112         }
 113
 114         return (struct journal_space) {
 115                 .next_entry     = sectors,
 116                 .total          = sectors + buckets * ca->mi.bucket_size,
 117         };
 118 }
 119
 120 static struct journal_space __journal_space_available(struct journal *j, unsigned nr_devs_want,
 121                             enum journal_space_from from)
 122 {
 123         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 124         struct bch_dev *ca;
 125         unsigned i, pos, nr_devs = 0;
 126         struct journal_space space, dev_space[BCH_SB_MEMBERS_MAX];
 127
 128         BUG_ON(nr_devs_want > ARRAY_SIZE(dev_space));
 129
 130         rcu_read_lock();
 131         for_each_member_device_rcu(ca, c, i,
 132                                    &c->rw_devs[BCH_DATA_journal]) {
 133                 if (!ca->journal.nr)
 134                         continue;
 135
 136                 space = journal_dev_space_available(j, ca, from);
 137                 if (!space.next_entry)
 138                         continue;
 139
 140                 for (pos = 0; pos < nr_devs; pos++)
 141                         if (space.total > dev_space[pos].total)
 142                                 break;
 143
 144                 array_insert_item(dev_space, nr_devs, pos, space);
 145         }
 146         rcu_read_unlock();
 147
 148         if (nr_devs < nr_devs_want)
 149                 return (struct journal_space) { 0, 0 };
 150
 151         /*
 152          * We sorted largest to smallest, and we want the smallest out of the
 153          * @nr_devs_want largest devices:
 154          */
 155         return dev_space[nr_devs_want - 1];
 156 }
 157
 158 void bch2_journal_space_available(struct journal *j)
 159 {
 160         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 161         struct bch_dev *ca;
 162         unsigned clean, clean_ondisk, total;
 163         s64 u64s_remaining = 0;
 164         unsigned max_entry_size  = min(j->buf[0].buf_size >> 9,
 165                                        j->buf[1].buf_size >> 9);
 166         unsigned i, nr_online = 0, nr_devs_want;
 167         bool can_discard = false;
 168         int ret = 0;
 169
 170         lockdep_assert_held(&j->lock);
 171
 172         rcu_read_lock();
 173         for_each_member_device_rcu(ca, c, i,
 174                                    &c->rw_devs[BCH_DATA_journal]) {
 175                 struct journal_device *ja = &ca->journal;
 176
 177                 if (!ja->nr)
 178                         continue;
 179
 180                 while (ja->dirty_idx != ja->cur_idx &&
 181                        ja->bucket_seq[ja->dirty_idx] < journal_last_seq(j))
 182                         ja->dirty_idx = (ja->dirty_idx + 1) % ja->nr;
 183
 184                 while (ja->dirty_idx_ondisk != ja->dirty_idx &&
 185                        ja->bucket_seq[ja->dirty_idx_ondisk] < j->last_seq_ondisk)
 186                         ja->dirty_idx_ondisk = (ja->dirty_idx_ondisk + 1) % ja->nr;
 187
 188                 if (ja->discard_idx != ja->dirty_idx_ondisk)
 189                         can_discard = true;
 190
 191                 max_entry_size = min_t(unsigned, max_entry_size, ca->mi.bucket_size);
 192                 nr_online++;
 193         }
 194         rcu_read_unlock();
 195
 196         j->can_discard = can_discard;
 197
 198         if (nr_online < c->opts.metadata_replicas_required) {
 199                 ret = JOURNAL_ERR_insufficient_devices;
 200                 goto out;
 201         }
 202
 203         nr_devs_want = min_t(unsigned, nr_online, c->opts.metadata_replicas);
 204
 205         for (i = 0; i < journal_space_nr; i++)
 206                 j->space[i] = __journal_space_available(j, nr_devs_want, i);
 207
 208         clean_ondisk    = j->space[journal_space_clean_ondisk].total;
 209         clean           = j->space[journal_space_clean].total;
 210         total           = j->space[journal_space_total].total;
 211
 212         if (!clean_ondisk &&
 213             journal_cur_seq(j) == j->seq_ondisk) {
 214                 struct printbuf buf = PRINTBUF;
 215
 216                 __bch2_journal_debug_to_text(&buf, j);
 217                 bch_err(c, "journal stuck\n%s", buf.buf);
 218                 printbuf_exit(&buf);
 219
 220                 /*
 221                  * Hack: bch2_fatal_error() calls bch2_journal_halt() which
 222                  * takes journal lock:
 223                  */
 224                 spin_unlock(&j->lock);
 225                 bch2_fatal_error(c);
 226                 spin_lock(&j->lock);
 227
 228                 ret = JOURNAL_ERR_journal_stuck;
 229         } else if (!j->space[journal_space_discarded].next_entry)
 230                 ret = JOURNAL_ERR_journal_full;
 231
 232         if ((j->space[journal_space_clean_ondisk].next_entry <
 233              j->space[journal_space_clean_ondisk].total) &&
 234             (clean - clean_ondisk <= total / 8) &&
 235             (clean_ondisk * 2 > clean ))
 236                 set_bit(JOURNAL_MAY_SKIP_FLUSH, &j->flags);
 237         else
 238                 clear_bit(JOURNAL_MAY_SKIP_FLUSH, &j->flags);
 239
 240         u64s_remaining  = (u64) clean << 6;
 241         u64s_remaining -= (u64) total << 3;
 242         u64s_remaining = max(0LL, u64s_remaining);
 243         u64s_remaining /= 4;
 244         u64s_remaining = min_t(u64, u64s_remaining, U32_MAX);
 245 out:
 246         j->cur_entry_sectors    = !ret ? j->space[journal_space_discarded].next_entry : 0;
 247         j->cur_entry_error      = ret;
 248         journal_set_remaining(j, u64s_remaining);
 249         journal_set_watermark(j);
 250
 251         if (!ret)
 252                 journal_wake(j);
 253 }
 254
 255 /* Discards - last part of journal reclaim: */
 256
 257 static bool should_discard_bucket(struct journal *j, struct journal_device *ja)
 258 {
 259         bool ret;
 260
 261         spin_lock(&j->lock);
 262         ret = ja->discard_idx != ja->dirty_idx_ondisk;
 263         spin_unlock(&j->lock);
 264
 265         return ret;
 266 }
 267
 268 /*
 269  * Advance ja->discard_idx as long as it points to buckets that are no longer
 270  * dirty, issuing discards if necessary:
 271  */
 272 void bch2_journal_do_discards(struct journal *j)
 273 {
 274         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 275         struct bch_dev *ca;
 276         unsigned iter;
 277
 278         mutex_lock(&j->discard_lock);
 279
 280         for_each_rw_member(ca, c, iter) {
 281                 struct journal_device *ja = &ca->journal;
 282
 283                 while (should_discard_bucket(j, ja)) {
 284                         if (!c->opts.nochanges &&
 285                             ca->mi.discard &&
 286                             bdev_max_discard_sectors(ca->disk_sb.bdev))
 287                                 blkdev_issue_discard(ca->disk_sb.bdev,
 288                                         bucket_to_sector(ca,
 289                                                 ja->buckets[ja->discard_idx]),
 290                                         ca->mi.bucket_size, GFP_NOIO);
 291
 292                         spin_lock(&j->lock);
 293                         ja->discard_idx = (ja->discard_idx + 1) % ja->nr;
 294
 295                         bch2_journal_space_available(j);
 296                         spin_unlock(&j->lock);
 297                 }
 298         }
 299
 300         mutex_unlock(&j->discard_lock);
 301 }
 302
 303 /*
 304  * Journal entry pinning - machinery for holding a reference on a given journal
 305  * entry, holding it open to ensure it gets replayed during recovery:
 306  */
 307
 308 static void bch2_journal_reclaim_fast(struct journal *j)
 309 {
 310         struct journal_entry_pin_list temp;
 311         bool popped = false;
 312
 313         lockdep_assert_held(&j->lock);
 314
 315         /*
 316          * Unpin journal entries whose reference counts reached zero, meaning
 317          * all btree nodes got written out
 318          */
 319         while (!fifo_empty(&j->pin) &&
 320                !atomic_read(&fifo_peek_front(&j->pin).count)) {
 321                 BUG_ON(!list_empty(&fifo_peek_front(&j->pin).list));
 322                 BUG_ON(!list_empty(&fifo_peek_front(&j->pin).flushed));
 323                 BUG_ON(!fifo_pop(&j->pin, temp));
 324                 popped = true;
 325         }
 326
 327         if (popped)
 328                 bch2_journal_space_available(j);
 329 }
 330
 331 void __bch2_journal_pin_put(struct journal *j, u64 seq)
 332 {
 333         struct journal_entry_pin_list *pin_list = journal_seq_pin(j, seq);
 334
 335         if (atomic_dec_and_test(&pin_list->count))
 336                 bch2_journal_reclaim_fast(j);
 337 }
 338
 339 void bch2_journal_pin_put(struct journal *j, u64 seq)
 340 {
 341         struct journal_entry_pin_list *pin_list = journal_seq_pin(j, seq);
 342
 343         if (atomic_dec_and_test(&pin_list->count)) {
 344                 spin_lock(&j->lock);
 345                 bch2_journal_reclaim_fast(j);
 346                 spin_unlock(&j->lock);
 347         }
 348 }
 349
 350 static inline void __journal_pin_drop(struct journal *j,
 351                                       struct journal_entry_pin *pin)
 352 {
 353         struct journal_entry_pin_list *pin_list;
 354
 355         if (!journal_pin_active(pin))
 356                 return;
 357
 358         if (j->flush_in_progress == pin)
 359                 j->flush_in_progress_dropped = true;
 360
 361         pin_list = journal_seq_pin(j, pin->seq);
 362         pin->seq = 0;
 363         list_del_init(&pin->list);
 364
 365         /*
 366          * Unpinning a journal entry make make journal_next_bucket() succeed, if
 367          * writing a new last_seq will now make another bucket available:
 368          */
 369         if (atomic_dec_and_test(&pin_list->count) &&
 370             pin_list == &fifo_peek_front(&j->pin))
 371                 bch2_journal_reclaim_fast(j);
 372 }
 373
 374 void bch2_journal_pin_drop(struct journal *j,
 375                            struct journal_entry_pin *pin)
 376 {
 377         spin_lock(&j->lock);
 378         __journal_pin_drop(j, pin);
 379         spin_unlock(&j->lock);
 380 }
 381
 382 void bch2_journal_pin_set(struct journal *j, u64 seq,
 383                           struct journal_entry_pin *pin,
 384                           journal_pin_flush_fn flush_fn)
 385 {
 386         struct journal_entry_pin_list *pin_list;
 387
 388         spin_lock(&j->lock);
 389
 390         if (seq < journal_last_seq(j)) {
 391                 /*
 392                  * bch2_journal_pin_copy() raced with bch2_journal_pin_drop() on
 393                  * the src pin - with the pin dropped, the entry to pin might no
 394                  * longer to exist, but that means there's no longer anything to
 395                  * copy and we can bail out here:
 396                  */
 397                 spin_unlock(&j->lock);
 398                 return;
 399         }
 400
 401         pin_list = journal_seq_pin(j, seq);
 402
 403         __journal_pin_drop(j, pin);
 404
 405         atomic_inc(&pin_list->count);
 406         pin->seq        = seq;
 407         pin->flush      = flush_fn;
 408
 409         if (flush_fn == bch2_btree_key_cache_journal_flush)
 410                 list_add(&pin->list, &pin_list->key_cache_list);
 411         else if (flush_fn)
 412                 list_add(&pin->list, &pin_list->list);
 413         else
 414                 list_add(&pin->list, &pin_list->flushed);
 415         spin_unlock(&j->lock);
 416
 417         /*
 418          * If the journal is currently full,  we might want to call flush_fn
 419          * immediately:
 420          */
 421         journal_wake(j);
 422 }
 423
 424 /**
 425  * bch2_journal_pin_flush: ensure journal pin callback is no longer running
 426  */
 427 void bch2_journal_pin_flush(struct journal *j, struct journal_entry_pin *pin)
 428 {
 429         BUG_ON(journal_pin_active(pin));
 430
 431         wait_event(j->pin_flush_wait, j->flush_in_progress != pin);
 432 }
 433
 434 /*
 435  * Journal reclaim: flush references to open journal entries to reclaim space in
 436  * the journal
 437  *
 438  * May be done by the journal code in the background as needed to free up space
 439  * for more journal entries, or as part of doing a clean shutdown, or to migrate
 440  * data off of a specific device:
 441  */
 442
 443 static struct journal_entry_pin *
 444 journal_get_next_pin(struct journal *j,
 445                      bool get_any,
 446                      bool get_key_cache,
 447                      u64 max_seq, u64 *seq)
 448 {
 449         struct journal_entry_pin_list *pin_list;
 450         struct journal_entry_pin *ret = NULL;
 451
 452         fifo_for_each_entry_ptr(pin_list, &j->pin, *seq) {
 453                 if (*seq > max_seq && !get_any && !get_key_cache)
 454                         break;
 455
 456                 if (*seq <= max_seq || get_any) {
 457                         ret = list_first_entry_or_null(&pin_list->list,
 458                                 struct journal_entry_pin, list);
 459                         if (ret)
 460                                 return ret;
 461                 }
 462
 463                 if (*seq <= max_seq || get_any || get_key_cache) {
 464                         ret = list_first_entry_or_null(&pin_list->key_cache_list,
 465                                 struct journal_entry_pin, list);
 466                         if (ret)
 467                                 return ret;
 468                 }
 469         }
 470
 471         return NULL;
 472 }
 473
 474 /* returns true if we did work */
 475 static size_t journal_flush_pins(struct journal *j, u64 seq_to_flush,
 476                                  unsigned min_any,
 477                                  unsigned min_key_cache)
 478 {
 479         struct journal_entry_pin *pin;
 480         size_t nr_flushed = 0;
 481         journal_pin_flush_fn flush_fn;
 482         u64 seq;
 483         int err;
 484
 485         lockdep_assert_held(&j->reclaim_lock);
 486
 487         while (1) {
 488                 cond_resched();
 489
 490                 j->last_flushed = jiffies;
 491
 492                 spin_lock(&j->lock);
 493                 pin = journal_get_next_pin(j,
 494                                            min_any != 0,
 495                                            min_key_cache != 0,
 496                                            seq_to_flush, &seq);
 497                 if (pin) {
 498                         BUG_ON(j->flush_in_progress);
 499                         j->flush_in_progress = pin;
 500                         j->flush_in_progress_dropped = false;
 501                         flush_fn = pin->flush;
 502                 }
 503                 spin_unlock(&j->lock);
 504
 505                 if (!pin)
 506                         break;
 507
 508                 if (min_key_cache && pin->flush == bch2_btree_key_cache_journal_flush)
 509                         min_key_cache--;
 510
 511                 if (min_any)
 512                         min_any--;
 513
 514                 err = flush_fn(j, pin, seq);
 515
 516                 spin_lock(&j->lock);
 517                 /* Pin might have been dropped or rearmed: */
 518                 if (likely(!err && !j->flush_in_progress_dropped))
 519                         list_move(&pin->list, &journal_seq_pin(j, seq)->flushed);
 520                 j->flush_in_progress = NULL;
 521                 j->flush_in_progress_dropped = false;
 522                 spin_unlock(&j->lock);
 523
 524                 wake_up(&j->pin_flush_wait);
 525
 526                 if (err)
 527                         break;
 528
 529                 nr_flushed++;
 530         }
 531
 532         return nr_flushed;
 533 }
 534
 535 static u64 journal_seq_to_flush(struct journal *j)
 536 {
 537         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 538         struct bch_dev *ca;
 539         u64 seq_to_flush = 0;
 540         unsigned iter;
 541
 542         spin_lock(&j->lock);
 543
 544         for_each_rw_member(ca, c, iter) {
 545                 struct journal_device *ja = &ca->journal;
 546                 unsigned nr_buckets, bucket_to_flush;
 547
 548                 if (!ja->nr)
 549                         continue;
 550
 551                 /* Try to keep the journal at most half full: */
 552                 nr_buckets = ja->nr / 2;
 553
 554                 /* And include pre-reservations: */
 555                 nr_buckets += DIV_ROUND_UP(j->prereserved.reserved,
 556                                            (ca->mi.bucket_size << 6) -
 557                                            journal_entry_overhead(j));
 558
 559                 nr_buckets = min(nr_buckets, ja->nr);
 560
 561                 bucket_to_flush = (ja->cur_idx + nr_buckets) % ja->nr;
 562                 seq_to_flush = max(seq_to_flush,
 563                                    ja->bucket_seq[bucket_to_flush]);
 564         }
 565
 566         /* Also flush if the pin fifo is more than half full */
 567         seq_to_flush = max_t(s64, seq_to_flush,
 568                              (s64) journal_cur_seq(j) -
 569                              (j->pin.size >> 1));
 570         spin_unlock(&j->lock);
 571
 572         return seq_to_flush;
 573 }
 574
 575 /**
 576  * bch2_journal_reclaim - free up journal buckets
 577  *
 578  * Background journal reclaim writes out btree nodes. It should be run
 579  * early enough so that we never completely run out of journal buckets.
 580  *
 581  * High watermarks for triggering background reclaim:
 582  * - FIFO has fewer than 512 entries left
 583  * - fewer than 25% journal buckets free
 584  *
 585  * Background reclaim runs until low watermarks are reached:
 586  * - FIFO has more than 1024 entries left
 587  * - more than 50% journal buckets free
 588  *
 589  * As long as a reclaim can complete in the time it takes to fill up
 590  * 512 journal entries or 25% of all journal buckets, then
 591  * journal_next_bucket() should not stall.
 592  */
 593 static int __bch2_journal_reclaim(struct journal *j, bool direct, bool kicked)
 594 {
 595         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 596         bool kthread = (current->flags & PF_KTHREAD) != 0;
 597         u64 seq_to_flush;
 598         size_t min_nr, min_key_cache, nr_flushed;
 599         unsigned flags;
 600         int ret = 0;
 601
 602         /*
 603          * We can't invoke memory reclaim while holding the reclaim_lock -
 604          * journal reclaim is required to make progress for memory reclaim
 605          * (cleaning the caches), so we can't get stuck in memory reclaim while
 606          * we're holding the reclaim lock:
 607          */
 608         lockdep_assert_held(&j->reclaim_lock);
 609         flags = memalloc_noreclaim_save();
 610
 611         do {
 612                 if (kthread && kthread_should_stop())
 613                         break;
 614
 615                 if (bch2_journal_error(j)) {
 616                         ret = -EIO;
 617                         break;
 618                 }
 619
 620                 bch2_journal_do_discards(j);
 621
 622                 seq_to_flush = journal_seq_to_flush(j);
 623                 min_nr = 0;
 624
 625                 /*
 626                  * If it's been longer than j->reclaim_delay_ms since we last flushed,
 627                  * make sure to flush at least one journal pin:
 628                  */
 629                 if (time_after(jiffies, j->last_flushed +
 630                                msecs_to_jiffies(c->opts.journal_reclaim_delay)))
 631                         min_nr = 1;
 632
 633                 if (j->prereserved.reserved * 4 > j->prereserved.remaining)
 634                         min_nr = 1;
 635
 636                 if (fifo_free(&j->pin) <= 32)
 637                         min_nr = 1;
 638
 639                 if (atomic_read(&c->btree_cache.dirty) * 2 > c->btree_cache.used)
 640                         min_nr = 1;
 641
 642                 min_key_cache = min(bch2_nr_btree_keys_need_flush(c), (size_t) 128);
 643
 644                 trace_and_count(c, journal_reclaim_start, c,
 645                                 direct, kicked,
 646                                 min_nr, min_key_cache,
 647                                 j->prereserved.reserved,
 648                                 j->prereserved.remaining,
 649                                 atomic_read(&c->btree_cache.dirty),
 650                                 c->btree_cache.used,
 651                                 atomic_long_read(&c->btree_key_cache.nr_dirty),
 652                                 atomic_long_read(&c->btree_key_cache.nr_keys));
 653
 654                 nr_flushed = journal_flush_pins(j, seq_to_flush,
 655                                                 min_nr, min_key_cache);
 656
 657                 if (direct)
 658                         j->nr_direct_reclaim += nr_flushed;
 659                 else
 660                         j->nr_background_reclaim += nr_flushed;
 661                 trace_and_count(c, journal_reclaim_finish, c, nr_flushed);
 662
 663                 if (nr_flushed)
 664                         wake_up(&j->reclaim_wait);
 665         } while ((min_nr || min_key_cache) && nr_flushed && !direct);
 666
 667         memalloc_noreclaim_restore(flags);
 668
 669         return ret;
 670 }
 671
 672 int bch2_journal_reclaim(struct journal *j)
 673 {
 674         return __bch2_journal_reclaim(j, true, true);
 675 }
 676
 677 static int bch2_journal_reclaim_thread(void *arg)
 678 {
 679         struct journal *j = arg;
 680         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 681         unsigned long delay, now;
 682         bool journal_empty;
 683         int ret = 0;
 684
 685         set_freezable();
 686
 687         j->last_flushed = jiffies;
 688
 689         while (!ret && !kthread_should_stop()) {
 690                 bool kicked = j->reclaim_kicked;
 691
 692                 j->reclaim_kicked = false;
 693
 694                 mutex_lock(&j->reclaim_lock);
 695                 ret = __bch2_journal_reclaim(j, false, kicked);
 696                 mutex_unlock(&j->reclaim_lock);
 697
 698                 now = jiffies;
 699                 delay = msecs_to_jiffies(c->opts.journal_reclaim_delay);
 700                 j->next_reclaim = j->last_flushed + delay;
 701
 702                 if (!time_in_range(j->next_reclaim, now, now + delay))
 703                         j->next_reclaim = now + delay;
 704
 705                 while (1) {
 706                         set_current_state(TASK_INTERRUPTIBLE);
 707                         if (kthread_should_stop())
 708                                 break;
 709                         if (j->reclaim_kicked)
 710                                 break;
 711
 712                         spin_lock(&j->lock);
 713                         journal_empty = fifo_empty(&j->pin);
 714                         spin_unlock(&j->lock);
 715
 716                         if (journal_empty)
 717                                 freezable_schedule();
 718                         else if (time_after(j->next_reclaim, jiffies))
 719                                 freezable_schedule_timeout(j->next_reclaim - jiffies);
 720                         else
 721                                 break;
 722                 }
 723                 __set_current_state(TASK_RUNNING);
 724         }
 725
 726         return 0;
 727 }
 728
 729 void bch2_journal_reclaim_stop(struct journal *j)
 730 {
 731         struct task_struct *p = j->reclaim_thread;
 732
 733         j->reclaim_thread = NULL;
 734
 735         if (p) {
 736                 kthread_stop(p);
 737                 put_task_struct(p);
 738         }
 739 }
 740
 741 int bch2_journal_reclaim_start(struct journal *j)
 742 {
 743         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 744         struct task_struct *p;
 745         int ret;
 746
 747         if (j->reclaim_thread)
 748                 return 0;
 749
 750         p = kthread_create(bch2_journal_reclaim_thread, j,
 751                            "bch-reclaim/%s", c->name);
 752         ret = PTR_ERR_OR_ZERO(p);
 753         if (ret) {
 754                 bch_err(c, "error creating journal reclaim thread: %s", bch2_err_str(ret));
 755                 return ret;
 756         }
 757
 758         get_task_struct(p);
 759         j->reclaim_thread = p;
 760         wake_up_process(p);
 761         return 0;
 762 }
 763
 764 static int journal_flush_done(struct journal *j, u64 seq_to_flush,
 765                               bool *did_work)
 766 {
 767         int ret;
 768
 769         ret = bch2_journal_error(j);
 770         if (ret)
 771                 return ret;
 772
 773         mutex_lock(&j->reclaim_lock);
 774
 775         if (journal_flush_pins(j, seq_to_flush, 0, 0))
 776                 *did_work = true;
 777
 778         spin_lock(&j->lock);
 779         /*
 780          * If journal replay hasn't completed, the unreplayed journal entries
 781          * hold refs on their corresponding sequence numbers
 782          */
 783         ret = !test_bit(JOURNAL_REPLAY_DONE, &j->flags) ||
 784                 journal_last_seq(j) > seq_to_flush ||
 785                 !fifo_used(&j->pin);
 786
 787         spin_unlock(&j->lock);
 788         mutex_unlock(&j->reclaim_lock);
 789
 790         return ret;
 791 }
 792
 793 bool bch2_journal_flush_pins(struct journal *j, u64 seq_to_flush)
 794 {
 795         bool did_work = false;
 796
 797         if (!test_bit(JOURNAL_STARTED, &j->flags))
 798                 return false;
 799
 800         closure_wait_event(&j->async_wait,
 801                 journal_flush_done(j, seq_to_flush, &did_work));
 802
 803         return did_work;
 804 }
 805
 806 int bch2_journal_flush_device_pins(struct journal *j, int dev_idx)
 807 {
 808         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 809         struct journal_entry_pin_list *p;
 810         u64 iter, seq = 0;
 811         int ret = 0;
 812
 813         spin_lock(&j->lock);
 814         fifo_for_each_entry_ptr(p, &j->pin, iter)
 815                 if (dev_idx >= 0
 816                     ? bch2_dev_list_has_dev(p->devs, dev_idx)
 817                     : p->devs.nr < c->opts.metadata_replicas)
 818                         seq = iter;
 819         spin_unlock(&j->lock);
 820
 821         bch2_journal_flush_pins(j, seq);
 822
 823         ret = bch2_journal_error(j);
 824         if (ret)
 825                 return ret;
 826
 827         mutex_lock(&c->replicas_gc_lock);
 828         bch2_replicas_gc_start(c, 1 << BCH_DATA_journal);
 829
 830         seq = 0;
 831
 832         spin_lock(&j->lock);
 833         while (!ret) {
 834                 struct bch_replicas_padded replicas;
 835
 836                 seq = max(seq, journal_last_seq(j));
 837                 if (seq >= j->pin.back)
 838                         break;
 839                 bch2_devlist_to_replicas(&replicas.e, BCH_DATA_journal,
 840                                          journal_seq_pin(j, seq)->devs);
 841                 seq++;
 842
 843                 spin_unlock(&j->lock);
 844                 ret = bch2_mark_replicas(c, &replicas.e);
 845                 spin_lock(&j->lock);
 846         }
 847         spin_unlock(&j->lock);
 848
 849         ret = bch2_replicas_gc_end(c, ret);
 850         mutex_unlock(&c->replicas_gc_lock);
 851
 852         return ret;
 853 }