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 "journal.h"
   6 #include "journal_io.h"
   7 #include "journal_reclaim.h"
   8 #include "replicas.h"
   9 #include "super.h"
  10
  11 #include <linux/kthread.h>
  12 #include <linux/sched/mm.h>
  13 #include <trace/events/bcachefs.h>
  14
  15 /* Free space calculations: */
  16
  17 static unsigned journal_space_from(struct journal_device *ja,
  18                                    enum journal_space_from from)
  19 {
  20         switch (from) {
  21         case journal_space_discarded:
  22                 return ja->discard_idx;
  23         case journal_space_clean_ondisk:
  24                 return ja->dirty_idx_ondisk;
  25         case journal_space_clean:
  26                 return ja->dirty_idx;
  27         default:
  28                 BUG();
  29         }
  30 }
  31
  32 unsigned bch2_journal_dev_buckets_available(struct journal *j,
  33                                             struct journal_device *ja,
  34                                             enum journal_space_from from)
  35 {
  36         unsigned available = (journal_space_from(ja, from) -
  37                               ja->cur_idx - 1 + ja->nr) % ja->nr;
  38
  39         /*
  40          * Don't use the last bucket unless writing the new last_seq
  41          * will make another bucket available:
  42          */
  43         if (available && ja->dirty_idx_ondisk == ja->dirty_idx)
  44                 --available;
  45
  46         return available;
  47 }
  48
  49 static void journal_set_remaining(struct journal *j, unsigned u64s_remaining)
  50 {
  51         union journal_preres_state old, new;
  52         u64 v = atomic64_read(&j->prereserved.counter);
  53
  54         do {
  55                 old.v = new.v = v;
  56                 new.remaining = u64s_remaining;
  57         } while ((v = atomic64_cmpxchg(&j->prereserved.counter,
  58                                        old.v, new.v)) != old.v);
  59 }
  60
  61 static struct journal_space {
  62         unsigned        next_entry;
  63         unsigned        remaining;
  64 } __journal_space_available(struct journal *j, unsigned nr_devs_want,
  65                             enum journal_space_from from)
  66 {
  67         struct bch_fs *c = container_of(j, struct bch_fs, journal);
  68         struct bch_dev *ca;
  69         unsigned sectors_next_entry     = UINT_MAX;
  70         unsigned sectors_total          = UINT_MAX;
  71         unsigned i, nr_devs = 0;
  72         unsigned unwritten_sectors = j->reservations.prev_buf_unwritten
  73                 ? journal_prev_buf(j)->sectors
  74                 : 0;
  75
  76         rcu_read_lock();
  77         for_each_member_device_rcu(ca, c, i,
  78                                    &c->rw_devs[BCH_DATA_journal]) {
  79                 struct journal_device *ja = &ca->journal;
  80                 unsigned buckets_this_device, sectors_this_device;
  81
  82                 if (!ja->nr)
  83                         continue;
  84
  85                 buckets_this_device = bch2_journal_dev_buckets_available(j, ja, from);
  86                 sectors_this_device = ja->sectors_free;
  87
  88                 /*
  89                  * We that we don't allocate the space for a journal entry
  90                  * until we write it out - thus, account for it here:
  91                  */
  92                 if (unwritten_sectors >= sectors_this_device) {
  93                         if (!buckets_this_device)
  94                                 continue;
  95
  96                         buckets_this_device--;
  97                         sectors_this_device = ca->mi.bucket_size;
  98                 }
  99
 100                 sectors_this_device -= unwritten_sectors;
 101
 102                 if (sectors_this_device < ca->mi.bucket_size &&
 103                     buckets_this_device) {
 104                         buckets_this_device--;
 105                         sectors_this_device = ca->mi.bucket_size;
 106                 }
 107
 108                 if (!sectors_this_device)
 109                         continue;
 110
 111                 sectors_next_entry = min(sectors_next_entry,
 112                                          sectors_this_device);
 113
 114                 sectors_total = min(sectors_total,
 115                         buckets_this_device * ca->mi.bucket_size +
 116                         sectors_this_device);
 117
 118                 nr_devs++;
 119         }
 120         rcu_read_unlock();
 121
 122         if (nr_devs < nr_devs_want)
 123                 return (struct journal_space) { 0, 0 };
 124
 125         return (struct journal_space) {
 126                 .next_entry     = sectors_next_entry,
 127                 .remaining      = max_t(int, 0, sectors_total - sectors_next_entry),
 128         };
 129 }
 130
 131 void bch2_journal_space_available(struct journal *j)
 132 {
 133         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 134         struct bch_dev *ca;
 135         struct journal_space discarded, clean_ondisk, clean;
 136         unsigned overhead, u64s_remaining = 0;
 137         unsigned max_entry_size  = min(j->buf[0].buf_size >> 9,
 138                                        j->buf[1].buf_size >> 9);
 139         unsigned i, nr_online = 0, nr_devs_want;
 140         bool can_discard = false;
 141         int ret = 0;
 142
 143         lockdep_assert_held(&j->lock);
 144
 145         rcu_read_lock();
 146         for_each_member_device_rcu(ca, c, i,
 147                                    &c->rw_devs[BCH_DATA_journal]) {
 148                 struct journal_device *ja = &ca->journal;
 149
 150                 if (!ja->nr)
 151                         continue;
 152
 153                 while (ja->dirty_idx != ja->cur_idx &&
 154                        ja->bucket_seq[ja->dirty_idx] < journal_last_seq(j))
 155                         ja->dirty_idx = (ja->dirty_idx + 1) % ja->nr;
 156
 157                 while (ja->dirty_idx_ondisk != ja->dirty_idx &&
 158                        ja->bucket_seq[ja->dirty_idx_ondisk] < j->last_seq_ondisk)
 159                         ja->dirty_idx_ondisk = (ja->dirty_idx_ondisk + 1) % ja->nr;
 160
 161                 if (ja->discard_idx != ja->dirty_idx_ondisk)
 162                         can_discard = true;
 163
 164                 max_entry_size = min_t(unsigned, max_entry_size, ca->mi.bucket_size);
 165                 nr_online++;
 166         }
 167         rcu_read_unlock();
 168
 169         j->can_discard = can_discard;
 170
 171         if (nr_online < c->opts.metadata_replicas_required) {
 172                 ret = cur_entry_insufficient_devices;
 173                 goto out;
 174         }
 175
 176         if (!fifo_free(&j->pin)) {
 177                 ret = cur_entry_journal_pin_full;
 178                 goto out;
 179         }
 180
 181         nr_devs_want = min_t(unsigned, nr_online, c->opts.metadata_replicas);
 182
 183         discarded       = __journal_space_available(j, nr_devs_want, journal_space_discarded);
 184         clean_ondisk    = __journal_space_available(j, nr_devs_want, journal_space_clean_ondisk);
 185         clean           = __journal_space_available(j, nr_devs_want, journal_space_clean);
 186
 187         if (!discarded.next_entry)
 188                 ret = cur_entry_journal_full;
 189
 190         overhead = DIV_ROUND_UP(clean.remaining, max_entry_size) *
 191                 journal_entry_overhead(j);
 192         u64s_remaining = clean.remaining << 6;
 193         u64s_remaining = max_t(int, 0, u64s_remaining - overhead);
 194         u64s_remaining /= 4;
 195 out:
 196         j->cur_entry_sectors    = !ret ? discarded.next_entry : 0;
 197         j->cur_entry_error      = ret;
 198         journal_set_remaining(j, u64s_remaining);
 199         journal_check_may_get_unreserved(j);
 200
 201         if (!ret)
 202                 journal_wake(j);
 203 }
 204
 205 /* Discards - last part of journal reclaim: */
 206
 207 static bool should_discard_bucket(struct journal *j, struct journal_device *ja)
 208 {
 209         bool ret;
 210
 211         spin_lock(&j->lock);
 212         ret = ja->discard_idx != ja->dirty_idx_ondisk;
 213         spin_unlock(&j->lock);
 214
 215         return ret;
 216 }
 217
 218 /*
 219  * Advance ja->discard_idx as long as it points to buckets that are no longer
 220  * dirty, issuing discards if necessary:
 221  */
 222 void bch2_journal_do_discards(struct journal *j)
 223 {
 224         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 225         struct bch_dev *ca;
 226         unsigned iter;
 227
 228         mutex_lock(&j->discard_lock);
 229
 230         for_each_rw_member(ca, c, iter) {
 231                 struct journal_device *ja = &ca->journal;
 232
 233                 while (should_discard_bucket(j, ja)) {
 234                         if (ca->mi.discard &&
 235                             blk_queue_discard(bdev_get_queue(ca->disk_sb.bdev)))
 236                                 blkdev_issue_discard(ca->disk_sb.bdev,
 237                                         bucket_to_sector(ca,
 238                                                 ja->buckets[ja->discard_idx]),
 239                                         ca->mi.bucket_size, GFP_NOIO, 0);
 240
 241                         spin_lock(&j->lock);
 242                         ja->discard_idx = (ja->discard_idx + 1) % ja->nr;
 243
 244                         bch2_journal_space_available(j);
 245                         spin_unlock(&j->lock);
 246                 }
 247         }
 248
 249         mutex_unlock(&j->discard_lock);
 250 }
 251
 252 /*
 253  * Journal entry pinning - machinery for holding a reference on a given journal
 254  * entry, holding it open to ensure it gets replayed during recovery:
 255  */
 256
 257 static void bch2_journal_reclaim_fast(struct journal *j)
 258 {
 259         struct journal_entry_pin_list temp;
 260         bool popped = false;
 261
 262         lockdep_assert_held(&j->lock);
 263
 264         /*
 265          * Unpin journal entries whose reference counts reached zero, meaning
 266          * all btree nodes got written out
 267          */
 268         while (!fifo_empty(&j->pin) &&
 269                !atomic_read(&fifo_peek_front(&j->pin).count)) {
 270                 BUG_ON(!list_empty(&fifo_peek_front(&j->pin).list));
 271                 BUG_ON(!list_empty(&fifo_peek_front(&j->pin).flushed));
 272                 BUG_ON(!fifo_pop(&j->pin, temp));
 273                 popped = true;
 274         }
 275
 276         if (popped)
 277                 bch2_journal_space_available(j);
 278 }
 279
 280 void bch2_journal_pin_put(struct journal *j, u64 seq)
 281 {
 282         struct journal_entry_pin_list *pin_list = journal_seq_pin(j, seq);
 283
 284         if (atomic_dec_and_test(&pin_list->count)) {
 285                 spin_lock(&j->lock);
 286                 bch2_journal_reclaim_fast(j);
 287                 spin_unlock(&j->lock);
 288         }
 289 }
 290
 291 static inline void __journal_pin_drop(struct journal *j,
 292                                       struct journal_entry_pin *pin)
 293 {
 294         struct journal_entry_pin_list *pin_list;
 295
 296         if (!journal_pin_active(pin))
 297                 return;
 298
 299         pin_list = journal_seq_pin(j, pin->seq);
 300         pin->seq = 0;
 301         list_del_init(&pin->list);
 302
 303         /*
 304          * Unpinning a journal entry make make journal_next_bucket() succeed, if
 305          * writing a new last_seq will now make another bucket available:
 306          */
 307         if (atomic_dec_and_test(&pin_list->count) &&
 308             pin_list == &fifo_peek_front(&j->pin))
 309                 bch2_journal_reclaim_fast(j);
 310         else if (fifo_used(&j->pin) == 1 &&
 311                  atomic_read(&pin_list->count) == 1)
 312                 journal_wake(j);
 313 }
 314
 315 void bch2_journal_pin_drop(struct journal *j,
 316                            struct journal_entry_pin *pin)
 317 {
 318         spin_lock(&j->lock);
 319         __journal_pin_drop(j, pin);
 320         spin_unlock(&j->lock);
 321 }
 322
 323 static void bch2_journal_pin_add_locked(struct journal *j, u64 seq,
 324                             struct journal_entry_pin *pin,
 325                             journal_pin_flush_fn flush_fn)
 326 {
 327         struct journal_entry_pin_list *pin_list = journal_seq_pin(j, seq);
 328
 329         __journal_pin_drop(j, pin);
 330
 331         BUG_ON(!atomic_read(&pin_list->count) && seq == journal_last_seq(j));
 332
 333         atomic_inc(&pin_list->count);
 334         pin->seq        = seq;
 335         pin->flush      = flush_fn;
 336
 337         list_add(&pin->list, flush_fn ? &pin_list->list : &pin_list->flushed);
 338 }
 339
 340 void __bch2_journal_pin_add(struct journal *j, u64 seq,
 341                             struct journal_entry_pin *pin,
 342                             journal_pin_flush_fn flush_fn)
 343 {
 344         spin_lock(&j->lock);
 345         bch2_journal_pin_add_locked(j, seq, pin, flush_fn);
 346         spin_unlock(&j->lock);
 347
 348         /*
 349          * If the journal is currently full,  we might want to call flush_fn
 350          * immediately:
 351          */
 352         journal_wake(j);
 353 }
 354
 355 void bch2_journal_pin_update(struct journal *j, u64 seq,
 356                              struct journal_entry_pin *pin,
 357                              journal_pin_flush_fn flush_fn)
 358 {
 359         if (journal_pin_active(pin) && pin->seq < seq)
 360                 return;
 361
 362         spin_lock(&j->lock);
 363
 364         if (pin->seq != seq) {
 365                 bch2_journal_pin_add_locked(j, seq, pin, flush_fn);
 366         } else {
 367                 struct journal_entry_pin_list *pin_list =
 368                         journal_seq_pin(j, seq);
 369
 370                 /*
 371                  * If the pin is already pinning the right sequence number, it
 372                  * still might've already been flushed:
 373                  */
 374                 list_move(&pin->list, &pin_list->list);
 375         }
 376
 377         spin_unlock(&j->lock);
 378
 379         /*
 380          * If the journal is currently full,  we might want to call flush_fn
 381          * immediately:
 382          */
 383         journal_wake(j);
 384 }
 385
 386 void bch2_journal_pin_copy(struct journal *j,
 387                            struct journal_entry_pin *dst,
 388                            struct journal_entry_pin *src,
 389                            journal_pin_flush_fn flush_fn)
 390 {
 391         spin_lock(&j->lock);
 392
 393         if (journal_pin_active(src) &&
 394             (!journal_pin_active(dst) || src->seq < dst->seq))
 395                 bch2_journal_pin_add_locked(j, src->seq, dst, flush_fn);
 396
 397         spin_unlock(&j->lock);
 398 }
 399
 400 /**
 401  * bch2_journal_pin_flush: ensure journal pin callback is no longer running
 402  */
 403 void bch2_journal_pin_flush(struct journal *j, struct journal_entry_pin *pin)
 404 {
 405         BUG_ON(journal_pin_active(pin));
 406
 407         wait_event(j->pin_flush_wait, j->flush_in_progress != pin);
 408 }
 409
 410 /*
 411  * Journal reclaim: flush references to open journal entries to reclaim space in
 412  * the journal
 413  *
 414  * May be done by the journal code in the background as needed to free up space
 415  * for more journal entries, or as part of doing a clean shutdown, or to migrate
 416  * data off of a specific device:
 417  */
 418
 419 static struct journal_entry_pin *
 420 journal_get_next_pin(struct journal *j, u64 max_seq, u64 *seq)
 421 {
 422         struct journal_entry_pin_list *pin_list;
 423         struct journal_entry_pin *ret = NULL;
 424
 425         if (!test_bit(JOURNAL_RECLAIM_STARTED, &j->flags))
 426                 return NULL;
 427
 428         spin_lock(&j->lock);
 429
 430         fifo_for_each_entry_ptr(pin_list, &j->pin, *seq)
 431                 if (*seq > max_seq ||
 432                     (ret = list_first_entry_or_null(&pin_list->list,
 433                                 struct journal_entry_pin, list)))
 434                         break;
 435
 436         if (ret) {
 437                 list_move(&ret->list, &pin_list->flushed);
 438                 BUG_ON(j->flush_in_progress);
 439                 j->flush_in_progress = ret;
 440         }
 441
 442         spin_unlock(&j->lock);
 443
 444         return ret;
 445 }
 446
 447 /* returns true if we did work */
 448 static u64 journal_flush_pins(struct journal *j, u64 seq_to_flush,
 449                               unsigned min_nr)
 450 {
 451         struct journal_entry_pin *pin;
 452         u64 seq, ret = 0;
 453
 454         lockdep_assert_held(&j->reclaim_lock);
 455
 456         while (1) {
 457                 cond_resched();
 458
 459                 j->last_flushed = jiffies;
 460
 461                 pin = journal_get_next_pin(j, min_nr
 462                                 ? U64_MAX : seq_to_flush, &seq);
 463                 if (!pin)
 464                         break;
 465
 466                 if (min_nr)
 467                         min_nr--;
 468
 469                 pin->flush(j, pin, seq);
 470
 471                 BUG_ON(j->flush_in_progress != pin);
 472                 j->flush_in_progress = NULL;
 473                 wake_up(&j->pin_flush_wait);
 474                 ret++;
 475         }
 476
 477         return ret;
 478 }
 479
 480 static u64 journal_seq_to_flush(struct journal *j)
 481 {
 482         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 483         struct bch_dev *ca;
 484         u64 seq_to_flush = 0;
 485         unsigned iter;
 486
 487         spin_lock(&j->lock);
 488
 489         for_each_rw_member(ca, c, iter) {
 490                 struct journal_device *ja = &ca->journal;
 491                 unsigned nr_buckets, bucket_to_flush;
 492
 493                 if (!ja->nr)
 494                         continue;
 495
 496                 /* Try to keep the journal at most half full: */
 497                 nr_buckets = ja->nr / 2;
 498
 499                 /* And include pre-reservations: */
 500                 nr_buckets += DIV_ROUND_UP(j->prereserved.reserved,
 501                                            (ca->mi.bucket_size << 6) -
 502                                            journal_entry_overhead(j));
 503
 504                 nr_buckets = min(nr_buckets, ja->nr);
 505
 506                 bucket_to_flush = (ja->cur_idx + nr_buckets) % ja->nr;
 507                 seq_to_flush = max(seq_to_flush,
 508                                    ja->bucket_seq[bucket_to_flush]);
 509         }
 510
 511         /* Also flush if the pin fifo is more than half full */
 512         seq_to_flush = max_t(s64, seq_to_flush,
 513                              (s64) journal_cur_seq(j) -
 514                              (j->pin.size >> 1));
 515         spin_unlock(&j->lock);
 516
 517         return seq_to_flush;
 518 }
 519
 520 /**
 521  * bch2_journal_reclaim - free up journal buckets
 522  *
 523  * Background journal reclaim writes out btree nodes. It should be run
 524  * early enough so that we never completely run out of journal buckets.
 525  *
 526  * High watermarks for triggering background reclaim:
 527  * - FIFO has fewer than 512 entries left
 528  * - fewer than 25% journal buckets free
 529  *
 530  * Background reclaim runs until low watermarks are reached:
 531  * - FIFO has more than 1024 entries left
 532  * - more than 50% journal buckets free
 533  *
 534  * As long as a reclaim can complete in the time it takes to fill up
 535  * 512 journal entries or 25% of all journal buckets, then
 536  * journal_next_bucket() should not stall.
 537  */
 538 static void __bch2_journal_reclaim(struct journal *j, bool direct)
 539 {
 540         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 541         bool kthread = (current->flags & PF_KTHREAD) != 0;
 542         u64 seq_to_flush, nr_flushed = 0;
 543         size_t min_nr;
 544         unsigned flags;
 545
 546         /*
 547          * We can't invoke memory reclaim while holding the reclaim_lock -
 548          * journal reclaim is required to make progress for memory reclaim
 549          * (cleaning the caches), so we can't get stuck in memory reclaim while
 550          * we're holding the reclaim lock:
 551          */
 552         lockdep_assert_held(&j->reclaim_lock);
 553         flags = memalloc_noreclaim_save();
 554
 555         do {
 556                 if (kthread && kthread_should_stop())
 557                         break;
 558
 559                 bch2_journal_do_discards(j);
 560
 561                 seq_to_flush = journal_seq_to_flush(j);
 562                 min_nr = 0;
 563
 564                 /*
 565                  * If it's been longer than j->reclaim_delay_ms since we last flushed,
 566                  * make sure to flush at least one journal pin:
 567                  */
 568                 if (time_after(jiffies, j->last_flushed +
 569                                msecs_to_jiffies(j->reclaim_delay_ms)))
 570                         min_nr = 1;
 571
 572                 if (j->prereserved.reserved * 2 > j->prereserved.remaining)
 573                         min_nr = 1;
 574
 575                 if (atomic_read(&c->btree_cache.dirty) * 4 >
 576                     c->btree_cache.used  * 3)
 577                         min_nr = 1;
 578
 579                 min_nr = max(min_nr, bch2_nr_btree_keys_need_flush(c));
 580
 581                 trace_journal_reclaim_start(c,
 582                                 min_nr,
 583                                 j->prereserved.reserved,
 584                                 j->prereserved.remaining,
 585                                 atomic_read(&c->btree_cache.dirty),
 586                                 c->btree_cache.used,
 587                                 c->btree_key_cache.nr_dirty,
 588                                 c->btree_key_cache.nr_keys);
 589
 590                 nr_flushed = journal_flush_pins(j, seq_to_flush, min_nr);
 591
 592                 if (direct)
 593                         j->nr_direct_reclaim += nr_flushed;
 594                 else
 595                         j->nr_background_reclaim += nr_flushed;
 596                 trace_journal_reclaim_finish(c, nr_flushed);
 597         } while (min_nr);
 598
 599         memalloc_noreclaim_restore(flags);
 600 }
 601
 602 void bch2_journal_reclaim(struct journal *j)
 603 {
 604         __bch2_journal_reclaim(j, true);
 605 }
 606
 607 static int bch2_journal_reclaim_thread(void *arg)
 608 {
 609         struct journal *j = arg;
 610         unsigned long next;
 611
 612         set_freezable();
 613
 614         kthread_wait_freezable(test_bit(JOURNAL_RECLAIM_STARTED, &j->flags));
 615
 616         while (!kthread_should_stop()) {
 617                 j->reclaim_kicked = false;
 618
 619                 mutex_lock(&j->reclaim_lock);
 620                 __bch2_journal_reclaim(j, false);
 621                 mutex_unlock(&j->reclaim_lock);
 622
 623                 next = j->last_flushed + msecs_to_jiffies(j->reclaim_delay_ms);
 624
 625                 while (1) {
 626                         set_current_state(TASK_INTERRUPTIBLE);
 627                         if (kthread_should_stop())
 628                                 break;
 629                         if (j->reclaim_kicked)
 630                                 break;
 631                         if (time_after_eq(jiffies, next))
 632                                 break;
 633                         schedule_timeout(next - jiffies);
 634                         try_to_freeze();
 635
 636                 }
 637                 __set_current_state(TASK_RUNNING);
 638         }
 639
 640         return 0;
 641 }
 642
 643 void bch2_journal_reclaim_stop(struct journal *j)
 644 {
 645         struct task_struct *p = j->reclaim_thread;
 646
 647         j->reclaim_thread = NULL;
 648
 649         if (p) {
 650                 kthread_stop(p);
 651                 put_task_struct(p);
 652         }
 653 }
 654
 655 int bch2_journal_reclaim_start(struct journal *j)
 656 {
 657         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 658         struct task_struct *p;
 659
 660         if (j->reclaim_thread)
 661                 return 0;
 662
 663         p = kthread_create(bch2_journal_reclaim_thread, j,
 664                            "bch-reclaim/%s", c->name);
 665         if (IS_ERR(p))
 666                 return PTR_ERR(p);
 667
 668         get_task_struct(p);
 669         j->reclaim_thread = p;
 670         wake_up_process(p);
 671         return 0;
 672 }
 673
 674 static int journal_flush_done(struct journal *j, u64 seq_to_flush,
 675                               bool *did_work)
 676 {
 677         int ret;
 678
 679         ret = bch2_journal_error(j);
 680         if (ret)
 681                 return ret;
 682
 683         mutex_lock(&j->reclaim_lock);
 684
 685         *did_work = journal_flush_pins(j, seq_to_flush, 0) != 0;
 686
 687         spin_lock(&j->lock);
 688         /*
 689          * If journal replay hasn't completed, the unreplayed journal entries
 690          * hold refs on their corresponding sequence numbers
 691          */
 692         ret = !test_bit(JOURNAL_REPLAY_DONE, &j->flags) ||
 693                 journal_last_seq(j) > seq_to_flush ||
 694                 (fifo_used(&j->pin) == 1 &&
 695                  atomic_read(&fifo_peek_front(&j->pin).count) == 1);
 696
 697         spin_unlock(&j->lock);
 698         mutex_unlock(&j->reclaim_lock);
 699
 700         return ret;
 701 }
 702
 703 bool bch2_journal_flush_pins(struct journal *j, u64 seq_to_flush)
 704 {
 705         bool did_work = false;
 706
 707         if (!test_bit(JOURNAL_STARTED, &j->flags))
 708                 return false;
 709
 710         closure_wait_event(&j->async_wait,
 711                 journal_flush_done(j, seq_to_flush, &did_work));
 712
 713         return did_work;
 714 }
 715
 716 int bch2_journal_flush_device_pins(struct journal *j, int dev_idx)
 717 {
 718         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 719         struct journal_entry_pin_list *p;
 720         u64 iter, seq = 0;
 721         int ret = 0;
 722
 723         spin_lock(&j->lock);
 724         fifo_for_each_entry_ptr(p, &j->pin, iter)
 725                 if (dev_idx >= 0
 726                     ? bch2_dev_list_has_dev(p->devs, dev_idx)
 727                     : p->devs.nr < c->opts.metadata_replicas)
 728                         seq = iter;
 729         spin_unlock(&j->lock);
 730
 731         bch2_journal_flush_pins(j, seq);
 732
 733         ret = bch2_journal_error(j);
 734         if (ret)
 735                 return ret;
 736
 737         mutex_lock(&c->replicas_gc_lock);
 738         bch2_replicas_gc_start(c, 1 << BCH_DATA_journal);
 739
 740         seq = 0;
 741
 742         spin_lock(&j->lock);
 743         while (!ret && seq < j->pin.back) {
 744                 struct bch_replicas_padded replicas;
 745
 746                 seq = max(seq, journal_last_seq(j));
 747                 bch2_devlist_to_replicas(&replicas.e, BCH_DATA_journal,
 748                                          journal_seq_pin(j, seq)->devs);
 749                 seq++;
 750
 751                 spin_unlock(&j->lock);
 752                 ret = bch2_mark_replicas(c, &replicas.e);
 753                 spin_lock(&j->lock);
 754         }
 755         spin_unlock(&j->lock);
 756
 757         ret = bch2_replicas_gc_end(c, ret);
 758         mutex_unlock(&c->replicas_gc_lock);
 759
 760         return ret;
 761 }