fs/bcachefs/journal_reclaim.c

   1 // SPDX-License-Identifier: GPL-2.0
   2
   3 #include "bcachefs.h"
   4 #include "btree_key_cache.h"
   5 #include "btree_update.h"
   6 #include "buckets.h"
   7 #include "errcode.h"
   8 #include "error.h"
   9 #include "journal.h"
  10 #include "journal_io.h"
  11 #include "journal_reclaim.h"
  12 #include "replicas.h"
  13 #include "sb-members.h"
  14 #include "trace.h"
  15
  16 #include <linux/kthread.h>
  17 #include <linux/sched/mm.h>
  18
  19 /* Free space calculations: */
  20
  21 static unsigned journal_space_from(struct journal_device *ja,
  22                                    enum journal_space_from from)
  23 {
  24         switch (from) {
  25         case journal_space_discarded:
  26                 return ja->discard_idx;
  27         case journal_space_clean_ondisk:
  28                 return ja->dirty_idx_ondisk;
  29         case journal_space_clean:
  30                 return ja->dirty_idx;
  31         default:
  32                 BUG();
  33         }
  34 }
  35
  36 unsigned bch2_journal_dev_buckets_available(struct journal *j,
  37                                             struct journal_device *ja,
  38                                             enum journal_space_from from)
  39 {
  40         unsigned available = (journal_space_from(ja, from) -
  41                               ja->cur_idx - 1 + ja->nr) % ja->nr;
  42
  43         /*
  44          * Don't use the last bucket unless writing the new last_seq
  45          * will make another bucket available:
  46          */
  47         if (available && ja->dirty_idx_ondisk == ja->dirty_idx)
  48                 --available;
  49
  50         return available;
  51 }
  52
  53 static inline void journal_set_watermark(struct journal *j, bool low_on_space)
  54 {
  55         unsigned watermark = BCH_WATERMARK_stripe;
  56
  57         if (low_on_space)
  58                 watermark = max_t(unsigned, watermark, BCH_WATERMARK_reclaim);
  59         if (fifo_free(&j->pin) < j->pin.size / 4)
  60                 watermark = max_t(unsigned, watermark, BCH_WATERMARK_reclaim);
  61
  62         if (watermark == j->watermark)
  63                 return;
  64
  65         swap(watermark, j->watermark);
  66         if (watermark > j->watermark)
  67                 journal_wake(j);
  68 }
  69
  70 static struct journal_space
  71 journal_dev_space_available(struct journal *j, struct bch_dev *ca,
  72                             enum journal_space_from from)
  73 {
  74         struct journal_device *ja = &ca->journal;
  75         unsigned sectors, buckets, unwritten;
  76         u64 seq;
  77
  78         if (from == journal_space_total)
  79                 return (struct journal_space) {
  80                         .next_entry     = ca->mi.bucket_size,
  81                         .total          = ca->mi.bucket_size * ja->nr,
  82                 };
  83
  84         buckets = bch2_journal_dev_buckets_available(j, ja, from);
  85         sectors = ja->sectors_free;
  86
  87         /*
  88          * We that we don't allocate the space for a journal entry
  89          * until we write it out - thus, account for it here:
  90          */
  91         for (seq = journal_last_unwritten_seq(j);
  92              seq <= journal_cur_seq(j);
  93              seq++) {
  94                 unwritten = j->buf[seq & JOURNAL_BUF_MASK].sectors;
  95
  96                 if (!unwritten)
  97                         continue;
  98
  99                 /* entry won't fit on this device, skip: */
 100                 if (unwritten > ca->mi.bucket_size)
 101                         continue;
 102
 103                 if (unwritten >= sectors) {
 104                         if (!buckets) {
 105                                 sectors = 0;
 106                                 break;
 107                         }
 108
 109                         buckets--;
 110                         sectors = ca->mi.bucket_size;
 111                 }
 112
 113                 sectors -= unwritten;
 114         }
 115
 116         if (sectors < ca->mi.bucket_size && buckets) {
 117                 buckets--;
 118                 sectors = ca->mi.bucket_size;
 119         }
 120
 121         return (struct journal_space) {
 122                 .next_entry     = sectors,
 123                 .total          = sectors + buckets * ca->mi.bucket_size,
 124         };
 125 }
 126
 127 static struct journal_space __journal_space_available(struct journal *j, unsigned nr_devs_want,
 128                             enum journal_space_from from)
 129 {
 130         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 131         struct bch_dev *ca;
 132         unsigned i, pos, nr_devs = 0;
 133         struct journal_space space, dev_space[BCH_SB_MEMBERS_MAX];
 134
 135         BUG_ON(nr_devs_want > ARRAY_SIZE(dev_space));
 136
 137         rcu_read_lock();
 138         for_each_member_device_rcu(ca, c, i,
 139                                    &c->rw_devs[BCH_DATA_journal]) {
 140                 if (!ca->journal.nr)
 141                         continue;
 142
 143                 space = journal_dev_space_available(j, ca, from);
 144                 if (!space.next_entry)
 145                         continue;
 146
 147                 for (pos = 0; pos < nr_devs; pos++)
 148                         if (space.total > dev_space[pos].total)
 149                                 break;
 150
 151                 array_insert_item(dev_space, nr_devs, pos, space);
 152         }
 153         rcu_read_unlock();
 154
 155         if (nr_devs < nr_devs_want)
 156                 return (struct journal_space) { 0, 0 };
 157
 158         /*
 159          * We sorted largest to smallest, and we want the smallest out of the
 160          * @nr_devs_want largest devices:
 161          */
 162         return dev_space[nr_devs_want - 1];
 163 }
 164
 165 void bch2_journal_space_available(struct journal *j)
 166 {
 167         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 168         struct bch_dev *ca;
 169         unsigned clean, clean_ondisk, total;
 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 = JOURNAL_ERR_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 (!j->space[journal_space_discarded].next_entry)
 219                 ret = JOURNAL_ERR_journal_full;
 220
 221         if ((j->space[journal_space_clean_ondisk].next_entry <
 222              j->space[journal_space_clean_ondisk].total) &&
 223             (clean - clean_ondisk <= total / 8) &&
 224             (clean_ondisk * 2 > clean))
 225                 set_bit(JOURNAL_MAY_SKIP_FLUSH, &j->flags);
 226         else
 227                 clear_bit(JOURNAL_MAY_SKIP_FLUSH, &j->flags);
 228
 229         journal_set_watermark(j, clean * 4 <= total);
 230 out:
 231         j->cur_entry_sectors    = !ret ? j->space[journal_space_discarded].next_entry : 0;
 232         j->cur_entry_error      = ret;
 233
 234         if (!ret)
 235                 journal_wake(j);
 236 }
 237
 238 /* Discards - last part of journal reclaim: */
 239
 240 static bool should_discard_bucket(struct journal *j, struct journal_device *ja)
 241 {
 242         bool ret;
 243
 244         spin_lock(&j->lock);
 245         ret = ja->discard_idx != ja->dirty_idx_ondisk;
 246         spin_unlock(&j->lock);
 247
 248         return ret;
 249 }
 250
 251 /*
 252  * Advance ja->discard_idx as long as it points to buckets that are no longer
 253  * dirty, issuing discards if necessary:
 254  */
 255 void bch2_journal_do_discards(struct journal *j)
 256 {
 257         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 258         struct bch_dev *ca;
 259         unsigned iter;
 260
 261         mutex_lock(&j->discard_lock);
 262
 263         for_each_rw_member(ca, c, iter) {
 264                 struct journal_device *ja = &ca->journal;
 265
 266                 while (should_discard_bucket(j, ja)) {
 267                         if (!c->opts.nochanges &&
 268                             ca->mi.discard &&
 269                             bdev_max_discard_sectors(ca->disk_sb.bdev))
 270                                 blkdev_issue_discard(ca->disk_sb.bdev,
 271                                         bucket_to_sector(ca,
 272                                                 ja->buckets[ja->discard_idx]),
 273                                         ca->mi.bucket_size, GFP_NOFS);
 274
 275                         spin_lock(&j->lock);
 276                         ja->discard_idx = (ja->discard_idx + 1) % ja->nr;
 277
 278                         bch2_journal_space_available(j);
 279                         spin_unlock(&j->lock);
 280                 }
 281         }
 282
 283         mutex_unlock(&j->discard_lock);
 284 }
 285
 286 /*
 287  * Journal entry pinning - machinery for holding a reference on a given journal
 288  * entry, holding it open to ensure it gets replayed during recovery:
 289  */
 290
 291 void bch2_journal_reclaim_fast(struct journal *j)
 292 {
 293         bool popped = false;
 294
 295         lockdep_assert_held(&j->lock);
 296
 297         /*
 298          * Unpin journal entries whose reference counts reached zero, meaning
 299          * all btree nodes got written out
 300          */
 301         while (!fifo_empty(&j->pin) &&
 302                !atomic_read(&fifo_peek_front(&j->pin).count)) {
 303                 j->pin.front++;
 304                 popped = true;
 305         }
 306
 307         if (popped)
 308                 bch2_journal_space_available(j);
 309 }
 310
 311 bool __bch2_journal_pin_put(struct journal *j, u64 seq)
 312 {
 313         struct journal_entry_pin_list *pin_list = journal_seq_pin(j, seq);
 314
 315         return atomic_dec_and_test(&pin_list->count);
 316 }
 317
 318 void bch2_journal_pin_put(struct journal *j, u64 seq)
 319 {
 320         if (__bch2_journal_pin_put(j, seq)) {
 321                 spin_lock(&j->lock);
 322                 bch2_journal_reclaim_fast(j);
 323                 spin_unlock(&j->lock);
 324         }
 325 }
 326
 327 static inline bool __journal_pin_drop(struct journal *j,
 328                                       struct journal_entry_pin *pin)
 329 {
 330         struct journal_entry_pin_list *pin_list;
 331
 332         if (!journal_pin_active(pin))
 333                 return false;
 334
 335         if (j->flush_in_progress == pin)
 336                 j->flush_in_progress_dropped = true;
 337
 338         pin_list = journal_seq_pin(j, pin->seq);
 339         pin->seq = 0;
 340         list_del_init(&pin->list);
 341
 342         /*
 343          * Unpinning a journal entry may make journal_next_bucket() succeed, if
 344          * writing a new last_seq will now make another bucket available:
 345          */
 346         return atomic_dec_and_test(&pin_list->count) &&
 347                 pin_list == &fifo_peek_front(&j->pin);
 348 }
 349
 350 void bch2_journal_pin_drop(struct journal *j,
 351                            struct journal_entry_pin *pin)
 352 {
 353         spin_lock(&j->lock);
 354         if (__journal_pin_drop(j, pin))
 355                 bch2_journal_reclaim_fast(j);
 356         spin_unlock(&j->lock);
 357 }
 358
 359 static enum journal_pin_type journal_pin_type(journal_pin_flush_fn fn)
 360 {
 361         if (fn == bch2_btree_node_flush0 ||
 362             fn == bch2_btree_node_flush1)
 363                 return JOURNAL_PIN_btree;
 364         else if (fn == bch2_btree_key_cache_journal_flush)
 365                 return JOURNAL_PIN_key_cache;
 366         else
 367                 return JOURNAL_PIN_other;
 368 }
 369
 370 void bch2_journal_pin_set(struct journal *j, u64 seq,
 371                           struct journal_entry_pin *pin,
 372                           journal_pin_flush_fn flush_fn)
 373 {
 374         struct journal_entry_pin_list *pin_list;
 375         bool reclaim;
 376
 377         spin_lock(&j->lock);
 378
 379         if (seq < journal_last_seq(j)) {
 380                 /*
 381                  * bch2_journal_pin_copy() raced with bch2_journal_pin_drop() on
 382                  * the src pin - with the pin dropped, the entry to pin might no
 383                  * longer to exist, but that means there's no longer anything to
 384                  * copy and we can bail out here:
 385                  */
 386                 spin_unlock(&j->lock);
 387                 return;
 388         }
 389
 390         pin_list = journal_seq_pin(j, seq);
 391
 392         reclaim = __journal_pin_drop(j, pin);
 393
 394         atomic_inc(&pin_list->count);
 395         pin->seq        = seq;
 396         pin->flush      = flush_fn;
 397
 398         if (flush_fn)
 399                 list_add(&pin->list, &pin_list->list[journal_pin_type(flush_fn)]);
 400         else
 401                 list_add(&pin->list, &pin_list->flushed);
 402
 403         if (reclaim)
 404                 bch2_journal_reclaim_fast(j);
 405         spin_unlock(&j->lock);
 406
 407         /*
 408          * If the journal is currently full,  we might want to call flush_fn
 409          * immediately:
 410          */
 411         journal_wake(j);
 412 }
 413
 414 /**
 415  * bch2_journal_pin_flush: ensure journal pin callback is no longer running
 416  * @j:          journal object
 417  * @pin:        pin to flush
 418  */
 419 void bch2_journal_pin_flush(struct journal *j, struct journal_entry_pin *pin)
 420 {
 421         BUG_ON(journal_pin_active(pin));
 422
 423         wait_event(j->pin_flush_wait, j->flush_in_progress != pin);
 424 }
 425
 426 /*
 427  * Journal reclaim: flush references to open journal entries to reclaim space in
 428  * the journal
 429  *
 430  * May be done by the journal code in the background as needed to free up space
 431  * for more journal entries, or as part of doing a clean shutdown, or to migrate
 432  * data off of a specific device:
 433  */
 434
 435 static struct journal_entry_pin *
 436 journal_get_next_pin(struct journal *j,
 437                      u64 seq_to_flush,
 438                      unsigned allowed_below_seq,
 439                      unsigned allowed_above_seq,
 440                      u64 *seq)
 441 {
 442         struct journal_entry_pin_list *pin_list;
 443         struct journal_entry_pin *ret = NULL;
 444         unsigned i;
 445
 446         fifo_for_each_entry_ptr(pin_list, &j->pin, *seq) {
 447                 if (*seq > seq_to_flush && !allowed_above_seq)
 448                         break;
 449
 450                 for (i = 0; i < JOURNAL_PIN_NR; i++)
 451                         if ((((1U << i) & allowed_below_seq) && *seq <= seq_to_flush) ||
 452                             ((1U << i) & allowed_above_seq)) {
 453                                 ret = list_first_entry_or_null(&pin_list->list[i],
 454                                         struct journal_entry_pin, list);
 455                                 if (ret)
 456                                         return ret;
 457                         }
 458         }
 459
 460         return NULL;
 461 }
 462
 463 /* returns true if we did work */
 464 static size_t journal_flush_pins(struct journal *j,
 465                                  u64 seq_to_flush,
 466                                  unsigned allowed_below_seq,
 467                                  unsigned allowed_above_seq,
 468                                  unsigned min_any,
 469                                  unsigned min_key_cache)
 470 {
 471         struct journal_entry_pin *pin;
 472         size_t nr_flushed = 0;
 473         journal_pin_flush_fn flush_fn;
 474         u64 seq;
 475         int err;
 476
 477         lockdep_assert_held(&j->reclaim_lock);
 478
 479         while (1) {
 480                 unsigned allowed_above = allowed_above_seq;
 481                 unsigned allowed_below = allowed_below_seq;
 482
 483                 if (min_any) {
 484                         allowed_above |= ~0;
 485                         allowed_below |= ~0;
 486                 }
 487
 488                 if (min_key_cache) {
 489                         allowed_above |= 1U << JOURNAL_PIN_key_cache;
 490                         allowed_below |= 1U << JOURNAL_PIN_key_cache;
 491                 }
 492
 493                 cond_resched();
 494
 495                 j->last_flushed = jiffies;
 496
 497                 spin_lock(&j->lock);
 498                 pin = journal_get_next_pin(j, seq_to_flush, allowed_below, allowed_above, &seq);
 499                 if (pin) {
 500                         BUG_ON(j->flush_in_progress);
 501                         j->flush_in_progress = pin;
 502                         j->flush_in_progress_dropped = false;
 503                         flush_fn = pin->flush;
 504                 }
 505                 spin_unlock(&j->lock);
 506
 507                 if (!pin)
 508                         break;
 509
 510                 if (min_key_cache && pin->flush == bch2_btree_key_cache_journal_flush)
 511                         min_key_cache--;
 512
 513                 if (min_any)
 514                         min_any--;
 515
 516                 err = flush_fn(j, pin, seq);
 517
 518                 spin_lock(&j->lock);
 519                 /* Pin might have been dropped or rearmed: */
 520                 if (likely(!err && !j->flush_in_progress_dropped))
 521                         list_move(&pin->list, &journal_seq_pin(j, seq)->flushed);
 522                 j->flush_in_progress = NULL;
 523                 j->flush_in_progress_dropped = false;
 524                 spin_unlock(&j->lock);
 525
 526                 wake_up(&j->pin_flush_wait);
 527
 528                 if (err)
 529                         break;
 530
 531                 nr_flushed++;
 532         }
 533
 534         return nr_flushed;
 535 }
 536
 537 static u64 journal_seq_to_flush(struct journal *j)
 538 {
 539         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 540         struct bch_dev *ca;
 541         u64 seq_to_flush = 0;
 542         unsigned iter;
 543
 544         spin_lock(&j->lock);
 545
 546         for_each_rw_member(ca, c, iter) {
 547                 struct journal_device *ja = &ca->journal;
 548                 unsigned nr_buckets, bucket_to_flush;
 549
 550                 if (!ja->nr)
 551                         continue;
 552
 553                 /* Try to keep the journal at most half full: */
 554                 nr_buckets = ja->nr / 2;
 555
 556                 nr_buckets = min(nr_buckets, ja->nr);
 557
 558                 bucket_to_flush = (ja->cur_idx + nr_buckets) % ja->nr;
 559                 seq_to_flush = max(seq_to_flush,
 560                                    ja->bucket_seq[bucket_to_flush]);
 561         }
 562
 563         /* Also flush if the pin fifo is more than half full */
 564         seq_to_flush = max_t(s64, seq_to_flush,
 565                              (s64) journal_cur_seq(j) -
 566                              (j->pin.size >> 1));
 567         spin_unlock(&j->lock);
 568
 569         return seq_to_flush;
 570 }
 571
 572 /**
 573  * __bch2_journal_reclaim - free up journal buckets
 574  * @j:          journal object
 575  * @direct:     direct or background reclaim?
 576  * @kicked:     requested to run since we last ran?
 577  * Returns:     0 on success, or -EIO if the journal has been shutdown
 578  *
 579  * Background journal reclaim writes out btree nodes. It should be run
 580  * early enough so that we never completely run out of journal buckets.
 581  *
 582  * High watermarks for triggering background reclaim:
 583  * - FIFO has fewer than 512 entries left
 584  * - fewer than 25% journal buckets free
 585  *
 586  * Background reclaim runs until low watermarks are reached:
 587  * - FIFO has more than 1024 entries left
 588  * - more than 50% journal buckets free
 589  *
 590  * As long as a reclaim can complete in the time it takes to fill up
 591  * 512 journal entries or 25% of all journal buckets, then
 592  * journal_next_bucket() should not stall.
 593  */
 594 static int __bch2_journal_reclaim(struct journal *j, bool direct, bool kicked)
 595 {
 596         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 597         bool kthread = (current->flags & PF_KTHREAD) != 0;
 598         u64 seq_to_flush;
 599         size_t min_nr, min_key_cache, nr_flushed;
 600         unsigned flags;
 601         int ret = 0;
 602
 603         /*
 604          * We can't invoke memory reclaim while holding the reclaim_lock -
 605          * journal reclaim is required to make progress for memory reclaim
 606          * (cleaning the caches), so we can't get stuck in memory reclaim while
 607          * we're holding the reclaim lock:
 608          */
 609         lockdep_assert_held(&j->reclaim_lock);
 610         flags = memalloc_noreclaim_save();
 611
 612         do {
 613                 if (kthread && kthread_should_stop())
 614                         break;
 615
 616                 if (bch2_journal_error(j)) {
 617                         ret = -EIO;
 618                         break;
 619                 }
 620
 621                 bch2_journal_do_discards(j);
 622
 623                 seq_to_flush = journal_seq_to_flush(j);
 624                 min_nr = 0;
 625
 626                 /*
 627                  * If it's been longer than j->reclaim_delay_ms since we last flushed,
 628                  * make sure to flush at least one journal pin:
 629                  */
 630                 if (time_after(jiffies, j->last_flushed +
 631                                msecs_to_jiffies(c->opts.journal_reclaim_delay)))
 632                         min_nr = 1;
 633
 634                 if (j->watermark != BCH_WATERMARK_stripe)
 635                         min_nr = 1;
 636
 637                 if (atomic_read(&c->btree_cache.dirty) * 2 > c->btree_cache.used)
 638                         min_nr = 1;
 639
 640                 min_key_cache = min(bch2_nr_btree_keys_need_flush(c), (size_t) 128);
 641
 642                 trace_and_count(c, journal_reclaim_start, c,
 643                                 direct, kicked,
 644                                 min_nr, min_key_cache,
 645                                 atomic_read(&c->btree_cache.dirty),
 646                                 c->btree_cache.used,
 647                                 atomic_long_read(&c->btree_key_cache.nr_dirty),
 648                                 atomic_long_read(&c->btree_key_cache.nr_keys));
 649
 650                 nr_flushed = journal_flush_pins(j, seq_to_flush,
 651                                                 ~0, 0,
 652                                                 min_nr, min_key_cache);
 653
 654                 if (direct)
 655                         j->nr_direct_reclaim += nr_flushed;
 656                 else
 657                         j->nr_background_reclaim += nr_flushed;
 658                 trace_and_count(c, journal_reclaim_finish, c, nr_flushed);
 659
 660                 if (nr_flushed)
 661                         wake_up(&j->reclaim_wait);
 662         } while ((min_nr || min_key_cache) && nr_flushed && !direct);
 663
 664         memalloc_noreclaim_restore(flags);
 665
 666         return ret;
 667 }
 668
 669 int bch2_journal_reclaim(struct journal *j)
 670 {
 671         return __bch2_journal_reclaim(j, true, true);
 672 }
 673
 674 static int bch2_journal_reclaim_thread(void *arg)
 675 {
 676         struct journal *j = arg;
 677         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 678         unsigned long delay, now;
 679         bool journal_empty;
 680         int ret = 0;
 681
 682         set_freezable();
 683
 684         j->last_flushed = jiffies;
 685
 686         while (!ret && !kthread_should_stop()) {
 687                 bool kicked = j->reclaim_kicked;
 688
 689                 j->reclaim_kicked = false;
 690
 691                 mutex_lock(&j->reclaim_lock);
 692                 ret = __bch2_journal_reclaim(j, false, kicked);
 693                 mutex_unlock(&j->reclaim_lock);
 694
 695                 now = jiffies;
 696                 delay = msecs_to_jiffies(c->opts.journal_reclaim_delay);
 697                 j->next_reclaim = j->last_flushed + delay;
 698
 699                 if (!time_in_range(j->next_reclaim, now, now + delay))
 700                         j->next_reclaim = now + delay;
 701
 702                 while (1) {
 703                         set_current_state(TASK_INTERRUPTIBLE|TASK_FREEZABLE);
 704                         if (kthread_should_stop())
 705                                 break;
 706                         if (j->reclaim_kicked)
 707                                 break;
 708
 709                         spin_lock(&j->lock);
 710                         journal_empty = fifo_empty(&j->pin);
 711                         spin_unlock(&j->lock);
 712
 713                         if (journal_empty)
 714                                 schedule();
 715                         else if (time_after(j->next_reclaim, jiffies))
 716                                 schedule_timeout(j->next_reclaim - jiffies);
 717                         else
 718                                 break;
 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         int ret;
 743
 744         if (j->reclaim_thread)
 745                 return 0;
 746
 747         p = kthread_create(bch2_journal_reclaim_thread, j,
 748                            "bch-reclaim/%s", c->name);
 749         ret = PTR_ERR_OR_ZERO(p);
 750         if (ret) {
 751                 bch_err_msg(c, ret, "creating journal reclaim thread");
 752                 return ret;
 753         }
 754
 755         get_task_struct(p);
 756         j->reclaim_thread = p;
 757         wake_up_process(p);
 758         return 0;
 759 }
 760
 761 static int journal_flush_done(struct journal *j, u64 seq_to_flush,
 762                               bool *did_work)
 763 {
 764         int ret;
 765
 766         ret = bch2_journal_error(j);
 767         if (ret)
 768                 return ret;
 769
 770         mutex_lock(&j->reclaim_lock);
 771
 772         if (journal_flush_pins(j, seq_to_flush,
 773                                (1U << JOURNAL_PIN_key_cache)|
 774                                (1U << JOURNAL_PIN_other), 0, 0, 0) ||
 775             journal_flush_pins(j, seq_to_flush,
 776                                (1U << JOURNAL_PIN_btree), 0, 0, 0))
 777                 *did_work = true;
 778
 779         spin_lock(&j->lock);
 780         /*
 781          * If journal replay hasn't completed, the unreplayed journal entries
 782          * hold refs on their corresponding sequence numbers
 783          */
 784         ret = !test_bit(JOURNAL_REPLAY_DONE, &j->flags) ||
 785                 journal_last_seq(j) > seq_to_flush ||
 786                 !fifo_used(&j->pin);
 787
 788         spin_unlock(&j->lock);
 789         mutex_unlock(&j->reclaim_lock);
 790
 791         return ret;
 792 }
 793
 794 bool bch2_journal_flush_pins(struct journal *j, u64 seq_to_flush)
 795 {
 796         bool did_work = false;
 797
 798         if (!test_bit(JOURNAL_STARTED, &j->flags))
 799                 return false;
 800
 801         closure_wait_event(&j->async_wait,
 802                 journal_flush_done(j, seq_to_flush, &did_work));
 803
 804         return did_work;
 805 }
 806
 807 int bch2_journal_flush_device_pins(struct journal *j, int dev_idx)
 808 {
 809         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 810         struct journal_entry_pin_list *p;
 811         u64 iter, seq = 0;
 812         int ret = 0;
 813
 814         spin_lock(&j->lock);
 815         fifo_for_each_entry_ptr(p, &j->pin, iter)
 816                 if (dev_idx >= 0
 817                     ? bch2_dev_list_has_dev(p->devs, dev_idx)
 818                     : p->devs.nr < c->opts.metadata_replicas)
 819                         seq = iter;
 820         spin_unlock(&j->lock);
 821
 822         bch2_journal_flush_pins(j, seq);
 823
 824         ret = bch2_journal_error(j);
 825         if (ret)
 826                 return ret;
 827
 828         mutex_lock(&c->replicas_gc_lock);
 829         bch2_replicas_gc_start(c, 1 << BCH_DATA_journal);
 830
 831         /*
 832          * Now that we've populated replicas_gc, write to the journal to mark
 833          * active journal devices. This handles the case where the journal might
 834          * be empty. Otherwise we could clear all journal replicas and
 835          * temporarily put the fs into an unrecoverable state. Journal recovery
 836          * expects to find devices marked for journal data on unclean mount.
 837          */
 838         ret = bch2_journal_meta(&c->journal);
 839         if (ret)
 840                 goto err;
 841
 842         seq = 0;
 843         spin_lock(&j->lock);
 844         while (!ret) {
 845                 struct bch_replicas_padded replicas;
 846
 847                 seq = max(seq, journal_last_seq(j));
 848                 if (seq >= j->pin.back)
 849                         break;
 850                 bch2_devlist_to_replicas(&replicas.e, BCH_DATA_journal,
 851                                          journal_seq_pin(j, seq)->devs);
 852                 seq++;
 853
 854                 spin_unlock(&j->lock);
 855                 ret = bch2_mark_replicas(c, &replicas.e);
 856                 spin_lock(&j->lock);
 857         }
 858         spin_unlock(&j->lock);
 859 err:
 860         ret = bch2_replicas_gc_end(c, ret);
 861         mutex_unlock(&c->replicas_gc_lock);
 862
 863         return ret;
 864 }