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 "btree_update.h"
   6 #include "errcode.h"
   7 #include "error.h"
   8 #include "journal.h"
   9 #include "journal_io.h"
  10 #include "journal_reclaim.h"
  11 #include "replicas.h"
  12 #include "super.h"
  13 #include "trace.h"
  14
  15 #include <linux/kthread.h>
  16 #include <linux/sched/mm.h>
  17
  18 /* Free space calculations: */
  19
  20 static unsigned journal_space_from(struct journal_device *ja,
  21                                    enum journal_space_from from)
  22 {
  23         switch (from) {
  24         case journal_space_discarded:
  25                 return ja->discard_idx;
  26         case journal_space_clean_ondisk:
  27                 return ja->dirty_idx_ondisk;
  28         case journal_space_clean:
  29                 return ja->dirty_idx;
  30         default:
  31                 BUG();
  32         }
  33 }
  34
  35 unsigned bch2_journal_dev_buckets_available(struct journal *j,
  36                                             struct journal_device *ja,
  37                                             enum journal_space_from from)
  38 {
  39         unsigned available = (journal_space_from(ja, from) -
  40                               ja->cur_idx - 1 + ja->nr) % ja->nr;
  41
  42         /*
  43          * Don't use the last bucket unless writing the new last_seq
  44          * will make another bucket available:
  45          */
  46         if (available && ja->dirty_idx_ondisk == ja->dirty_idx)
  47                 --available;
  48
  49         return available;
  50 }
  51
  52 static void journal_set_remaining(struct journal *j, unsigned u64s_remaining)
  53 {
  54         union journal_preres_state old, new;
  55         u64 v = atomic64_read(&j->prereserved.counter);
  56
  57         do {
  58                 old.v = new.v = v;
  59                 new.remaining = u64s_remaining;
  60         } while ((v = atomic64_cmpxchg(&j->prereserved.counter,
  61                                        old.v, new.v)) != old.v);
  62 }
  63
  64 static struct journal_space
  65 journal_dev_space_available(struct journal *j, struct bch_dev *ca,
  66                             enum journal_space_from from)
  67 {
  68         struct journal_device *ja = &ca->journal;
  69         unsigned sectors, buckets, unwritten;
  70         u64 seq;
  71
  72         if (from == journal_space_total)
  73                 return (struct journal_space) {
  74                         .next_entry     = ca->mi.bucket_size,
  75                         .total          = ca->mi.bucket_size * ja->nr,
  76                 };
  77
  78         buckets = bch2_journal_dev_buckets_available(j, ja, from);
  79         sectors = ja->sectors_free;
  80
  81         /*
  82          * We that we don't allocate the space for a journal entry
  83          * until we write it out - thus, account for it here:
  84          */
  85         for (seq = journal_last_unwritten_seq(j);
  86              seq <= journal_cur_seq(j);
  87              seq++) {
  88                 unwritten = j->buf[seq & JOURNAL_BUF_MASK].sectors;
  89
  90                 if (!unwritten)
  91                         continue;
  92
  93                 /* entry won't fit on this device, skip: */
  94                 if (unwritten > ca->mi.bucket_size)
  95                         continue;
  96
  97                 if (unwritten >= sectors) {
  98                         if (!buckets) {
  99                                 sectors = 0;
 100                                 break;
 101                         }
 102
 103                         buckets--;
 104                         sectors = ca->mi.bucket_size;
 105                 }
 106
 107                 sectors -= unwritten;
 108         }
 109
 110         if (sectors < ca->mi.bucket_size && buckets) {
 111                 buckets--;
 112                 sectors = ca->mi.bucket_size;
 113         }
 114
 115         return (struct journal_space) {
 116                 .next_entry     = sectors,
 117                 .total          = sectors + buckets * ca->mi.bucket_size,
 118         };
 119 }
 120
 121 static struct journal_space __journal_space_available(struct journal *j, unsigned nr_devs_want,
 122                             enum journal_space_from from)
 123 {
 124         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 125         struct bch_dev *ca;
 126         unsigned i, pos, nr_devs = 0;
 127         struct journal_space space, dev_space[BCH_SB_MEMBERS_MAX];
 128
 129         BUG_ON(nr_devs_want > ARRAY_SIZE(dev_space));
 130
 131         rcu_read_lock();
 132         for_each_member_device_rcu(ca, c, i,
 133                                    &c->rw_devs[BCH_DATA_journal]) {
 134                 if (!ca->journal.nr)
 135                         continue;
 136
 137                 space = journal_dev_space_available(j, ca, from);
 138                 if (!space.next_entry)
 139                         continue;
 140
 141                 for (pos = 0; pos < nr_devs; pos++)
 142                         if (space.total > dev_space[pos].total)
 143                                 break;
 144
 145                 array_insert_item(dev_space, nr_devs, pos, space);
 146         }
 147         rcu_read_unlock();
 148
 149         if (nr_devs < nr_devs_want)
 150                 return (struct journal_space) { 0, 0 };
 151
 152         /*
 153          * We sorted largest to smallest, and we want the smallest out of the
 154          * @nr_devs_want largest devices:
 155          */
 156         return dev_space[nr_devs_want - 1];
 157 }
 158
 159 void bch2_journal_space_available(struct journal *j)
 160 {
 161         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 162         struct bch_dev *ca;
 163         unsigned clean, clean_ondisk, total;
 164         s64 u64s_remaining = 0;
 165         unsigned max_entry_size  = min(j->buf[0].buf_size >> 9,
 166                                        j->buf[1].buf_size >> 9);
 167         unsigned i, nr_online = 0, nr_devs_want;
 168         bool can_discard = false;
 169         int ret = 0;
 170
 171         lockdep_assert_held(&j->lock);
 172
 173         rcu_read_lock();
 174         for_each_member_device_rcu(ca, c, i,
 175                                    &c->rw_devs[BCH_DATA_journal]) {
 176                 struct journal_device *ja = &ca->journal;
 177
 178                 if (!ja->nr)
 179                         continue;
 180
 181                 while (ja->dirty_idx != ja->cur_idx &&
 182                        ja->bucket_seq[ja->dirty_idx] < journal_last_seq(j))
 183                         ja->dirty_idx = (ja->dirty_idx + 1) % ja->nr;
 184
 185                 while (ja->dirty_idx_ondisk != ja->dirty_idx &&
 186                        ja->bucket_seq[ja->dirty_idx_ondisk] < j->last_seq_ondisk)
 187                         ja->dirty_idx_ondisk = (ja->dirty_idx_ondisk + 1) % ja->nr;
 188
 189                 if (ja->discard_idx != ja->dirty_idx_ondisk)
 190                         can_discard = true;
 191
 192                 max_entry_size = min_t(unsigned, max_entry_size, ca->mi.bucket_size);
 193                 nr_online++;
 194         }
 195         rcu_read_unlock();
 196
 197         j->can_discard = can_discard;
 198
 199         if (nr_online < c->opts.metadata_replicas_required) {
 200                 ret = JOURNAL_ERR_insufficient_devices;
 201                 goto out;
 202         }
 203
 204         nr_devs_want = min_t(unsigned, nr_online, c->opts.metadata_replicas);
 205
 206         for (i = 0; i < journal_space_nr; i++)
 207                 j->space[i] = __journal_space_available(j, nr_devs_want, i);
 208
 209         clean_ondisk    = j->space[journal_space_clean_ondisk].total;
 210         clean           = j->space[journal_space_clean].total;
 211         total           = j->space[journal_space_total].total;
 212
 213         if (!j->space[journal_space_discarded].next_entry)
 214                 ret = JOURNAL_ERR_journal_full;
 215
 216         if ((j->space[journal_space_clean_ondisk].next_entry <
 217              j->space[journal_space_clean_ondisk].total) &&
 218             (clean - clean_ondisk <= total / 8) &&
 219             (clean_ondisk * 2 > clean))
 220                 set_bit(JOURNAL_MAY_SKIP_FLUSH, &j->flags);
 221         else
 222                 clear_bit(JOURNAL_MAY_SKIP_FLUSH, &j->flags);
 223
 224         u64s_remaining  = (u64) clean << 6;
 225         u64s_remaining -= (u64) total << 3;
 226         u64s_remaining = max(0LL, u64s_remaining);
 227         u64s_remaining /= 4;
 228         u64s_remaining = min_t(u64, u64s_remaining, U32_MAX);
 229 out:
 230         j->cur_entry_sectors    = !ret ? j->space[journal_space_discarded].next_entry : 0;
 231         j->cur_entry_error      = ret;
 232         journal_set_remaining(j, u64s_remaining);
 233         journal_set_watermark(j);
 234
 235         if (!ret)
 236                 journal_wake(j);
 237 }
 238
 239 /* Discards - last part of journal reclaim: */
 240
 241 static bool should_discard_bucket(struct journal *j, struct journal_device *ja)
 242 {
 243         bool ret;
 244
 245         spin_lock(&j->lock);
 246         ret = ja->discard_idx != ja->dirty_idx_ondisk;
 247         spin_unlock(&j->lock);
 248
 249         return ret;
 250 }
 251
 252 /*
 253  * Advance ja->discard_idx as long as it points to buckets that are no longer
 254  * dirty, issuing discards if necessary:
 255  */
 256 void bch2_journal_do_discards(struct journal *j)
 257 {
 258         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 259         struct bch_dev *ca;
 260         unsigned iter;
 261
 262         mutex_lock(&j->discard_lock);
 263
 264         for_each_rw_member(ca, c, iter) {
 265                 struct journal_device *ja = &ca->journal;
 266
 267                 while (should_discard_bucket(j, ja)) {
 268                         if (!c->opts.nochanges &&
 269                             ca->mi.discard &&
 270                             bdev_max_discard_sectors(ca->disk_sb.bdev))
 271                                 blkdev_issue_discard(ca->disk_sb.bdev,
 272                                         bucket_to_sector(ca,
 273                                                 ja->buckets[ja->discard_idx]),
 274                                         ca->mi.bucket_size, GFP_NOIO);
 275
 276                         spin_lock(&j->lock);
 277                         ja->discard_idx = (ja->discard_idx + 1) % ja->nr;
 278
 279                         bch2_journal_space_available(j);
 280                         spin_unlock(&j->lock);
 281                 }
 282         }
 283
 284         mutex_unlock(&j->discard_lock);
 285 }
 286
 287 /*
 288  * Journal entry pinning - machinery for holding a reference on a given journal
 289  * entry, holding it open to ensure it gets replayed during recovery:
 290  */
 291
 292 static void bch2_journal_reclaim_fast(struct journal *j)
 293 {
 294         struct journal_entry_pin_list temp;
 295         bool popped = false;
 296
 297         lockdep_assert_held(&j->lock);
 298
 299         /*
 300          * Unpin journal entries whose reference counts reached zero, meaning
 301          * all btree nodes got written out
 302          */
 303         while (!fifo_empty(&j->pin) &&
 304                !atomic_read(&fifo_peek_front(&j->pin).count)) {
 305                 fifo_pop(&j->pin, temp);
 306                 popped = true;
 307         }
 308
 309         if (popped)
 310                 bch2_journal_space_available(j);
 311 }
 312
 313 void __bch2_journal_pin_put(struct journal *j, u64 seq)
 314 {
 315         struct journal_entry_pin_list *pin_list = journal_seq_pin(j, seq);
 316
 317         if (atomic_dec_and_test(&pin_list->count))
 318                 bch2_journal_reclaim_fast(j);
 319 }
 320
 321 void bch2_journal_pin_put(struct journal *j, u64 seq)
 322 {
 323         struct journal_entry_pin_list *pin_list = journal_seq_pin(j, seq);
 324
 325         if (atomic_dec_and_test(&pin_list->count)) {
 326                 spin_lock(&j->lock);
 327                 bch2_journal_reclaim_fast(j);
 328                 spin_unlock(&j->lock);
 329         }
 330 }
 331
 332 static inline bool __journal_pin_drop(struct journal *j,
 333                                       struct journal_entry_pin *pin)
 334 {
 335         struct journal_entry_pin_list *pin_list;
 336
 337         if (!journal_pin_active(pin))
 338                 return false;
 339
 340         if (j->flush_in_progress == pin)
 341                 j->flush_in_progress_dropped = true;
 342
 343         pin_list = journal_seq_pin(j, pin->seq);
 344         pin->seq = 0;
 345         list_del_init(&pin->list);
 346
 347         /*
 348          * Unpinning a journal entry make make journal_next_bucket() succeed, if
 349          * writing a new last_seq will now make another bucket available:
 350          */
 351         return atomic_dec_and_test(&pin_list->count) &&
 352                 pin_list == &fifo_peek_front(&j->pin);
 353 }
 354
 355 void bch2_journal_pin_drop(struct journal *j,
 356                            struct journal_entry_pin *pin)
 357 {
 358         spin_lock(&j->lock);
 359         if (__journal_pin_drop(j, pin))
 360                 bch2_journal_reclaim_fast(j);
 361         spin_unlock(&j->lock);
 362 }
 363
 364 enum journal_pin_type journal_pin_type(journal_pin_flush_fn fn)
 365 {
 366         if (fn == bch2_btree_node_flush0 ||
 367             fn == bch2_btree_node_flush1)
 368                 return JOURNAL_PIN_btree;
 369         else if (fn == bch2_btree_key_cache_journal_flush)
 370                 return JOURNAL_PIN_key_cache;
 371         else
 372                 return JOURNAL_PIN_other;
 373 }
 374
 375 void bch2_journal_pin_set(struct journal *j, u64 seq,
 376                           struct journal_entry_pin *pin,
 377                           journal_pin_flush_fn flush_fn)
 378 {
 379         struct journal_entry_pin_list *pin_list;
 380         bool reclaim;
 381
 382         spin_lock(&j->lock);
 383
 384         if (seq < journal_last_seq(j)) {
 385                 /*
 386                  * bch2_journal_pin_copy() raced with bch2_journal_pin_drop() on
 387                  * the src pin - with the pin dropped, the entry to pin might no
 388                  * longer to exist, but that means there's no longer anything to
 389                  * copy and we can bail out here:
 390                  */
 391                 spin_unlock(&j->lock);
 392                 return;
 393         }
 394
 395         pin_list = journal_seq_pin(j, seq);
 396
 397         reclaim = __journal_pin_drop(j, pin);
 398
 399         atomic_inc(&pin_list->count);
 400         pin->seq        = seq;
 401         pin->flush      = flush_fn;
 402
 403         if (flush_fn)
 404                 list_add(&pin->list, &pin_list->list[journal_pin_type(flush_fn)]);
 405         else
 406                 list_add(&pin->list, &pin_list->flushed);
 407
 408         if (reclaim)
 409                 bch2_journal_reclaim_fast(j);
 410         spin_unlock(&j->lock);
 411
 412         /*
 413          * If the journal is currently full,  we might want to call flush_fn
 414          * immediately:
 415          */
 416         journal_wake(j);
 417 }
 418
 419 /**
 420  * bch2_journal_pin_flush: ensure journal pin callback is no longer running
 421  */
 422 void bch2_journal_pin_flush(struct journal *j, struct journal_entry_pin *pin)
 423 {
 424         BUG_ON(journal_pin_active(pin));
 425
 426         wait_event(j->pin_flush_wait, j->flush_in_progress != pin);
 427 }
 428
 429 /*
 430  * Journal reclaim: flush references to open journal entries to reclaim space in
 431  * the journal
 432  *
 433  * May be done by the journal code in the background as needed to free up space
 434  * for more journal entries, or as part of doing a clean shutdown, or to migrate
 435  * data off of a specific device:
 436  */
 437
 438 static struct journal_entry_pin *
 439 journal_get_next_pin(struct journal *j,
 440                      u64 seq_to_flush,
 441                      unsigned allowed_below_seq,
 442                      unsigned allowed_above_seq,
 443                      u64 *seq)
 444 {
 445         struct journal_entry_pin_list *pin_list;
 446         struct journal_entry_pin *ret = NULL;
 447         unsigned i;
 448
 449         fifo_for_each_entry_ptr(pin_list, &j->pin, *seq) {
 450                 if (*seq > seq_to_flush && !allowed_above_seq)
 451                         break;
 452
 453                 for (i = 0; i < JOURNAL_PIN_NR; i++)
 454                         if ((((1U << i) & allowed_below_seq) && *seq <= seq_to_flush) ||
 455                             ((1U << i) & allowed_above_seq)) {
 456                                 ret = list_first_entry_or_null(&pin_list->list[i],
 457                                         struct journal_entry_pin, list);
 458                                 if (ret)
 459                                         return ret;
 460                         }
 461         }
 462
 463         return NULL;
 464 }
 465
 466 /* returns true if we did work */
 467 static size_t journal_flush_pins(struct journal *j,
 468                                  u64 seq_to_flush,
 469                                  unsigned allowed_below_seq,
 470                                  unsigned allowed_above_seq,
 471                                  unsigned min_any,
 472                                  unsigned min_key_cache)
 473 {
 474         struct journal_entry_pin *pin;
 475         size_t nr_flushed = 0;
 476         journal_pin_flush_fn flush_fn;
 477         u64 seq;
 478         int err;
 479
 480         lockdep_assert_held(&j->reclaim_lock);
 481
 482         while (1) {
 483                 unsigned allowed_above = allowed_above_seq;
 484                 unsigned allowed_below = allowed_below_seq;
 485
 486                 if (min_any) {
 487                         allowed_above |= ~0;
 488                         allowed_below |= ~0;
 489                 }
 490
 491                 if (min_key_cache) {
 492                         allowed_above |= 1U << JOURNAL_PIN_key_cache;
 493                         allowed_below |= 1U << JOURNAL_PIN_key_cache;
 494                 }
 495
 496                 cond_resched();
 497
 498                 j->last_flushed = jiffies;
 499
 500                 spin_lock(&j->lock);
 501                 pin = journal_get_next_pin(j, seq_to_flush, allowed_below, allowed_above, &seq);
 502                 if (pin) {
 503                         BUG_ON(j->flush_in_progress);
 504                         j->flush_in_progress = pin;
 505                         j->flush_in_progress_dropped = false;
 506                         flush_fn = pin->flush;
 507                 }
 508                 spin_unlock(&j->lock);
 509
 510                 if (!pin)
 511                         break;
 512
 513                 if (min_key_cache && pin->flush == bch2_btree_key_cache_journal_flush)
 514                         min_key_cache--;
 515
 516                 if (min_any)
 517                         min_any--;
 518
 519                 err = flush_fn(j, pin, seq);
 520
 521                 spin_lock(&j->lock);
 522                 /* Pin might have been dropped or rearmed: */
 523                 if (likely(!err && !j->flush_in_progress_dropped))
 524                         list_move(&pin->list, &journal_seq_pin(j, seq)->flushed);
 525                 j->flush_in_progress = NULL;
 526                 j->flush_in_progress_dropped = false;
 527                 spin_unlock(&j->lock);
 528
 529                 wake_up(&j->pin_flush_wait);
 530
 531                 if (err)
 532                         break;
 533
 534                 nr_flushed++;
 535         }
 536
 537         return nr_flushed;
 538 }
 539
 540 static u64 journal_seq_to_flush(struct journal *j)
 541 {
 542         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 543         struct bch_dev *ca;
 544         u64 seq_to_flush = 0;
 545         unsigned iter;
 546
 547         spin_lock(&j->lock);
 548
 549         for_each_rw_member(ca, c, iter) {
 550                 struct journal_device *ja = &ca->journal;
 551                 unsigned nr_buckets, bucket_to_flush;
 552
 553                 if (!ja->nr)
 554                         continue;
 555
 556                 /* Try to keep the journal at most half full: */
 557                 nr_buckets = ja->nr / 2;
 558
 559                 /* And include pre-reservations: */
 560                 nr_buckets += DIV_ROUND_UP(j->prereserved.reserved,
 561                                            (ca->mi.bucket_size << 6) -
 562                                            journal_entry_overhead(j));
 563
 564                 nr_buckets = min(nr_buckets, ja->nr);
 565
 566                 bucket_to_flush = (ja->cur_idx + nr_buckets) % ja->nr;
 567                 seq_to_flush = max(seq_to_flush,
 568                                    ja->bucket_seq[bucket_to_flush]);
 569         }
 570
 571         /* Also flush if the pin fifo is more than half full */
 572         seq_to_flush = max_t(s64, seq_to_flush,
 573                              (s64) journal_cur_seq(j) -
 574                              (j->pin.size >> 1));
 575         spin_unlock(&j->lock);
 576
 577         return seq_to_flush;
 578 }
 579
 580 /**
 581  * bch2_journal_reclaim - free up journal buckets
 582  *
 583  * Background journal reclaim writes out btree nodes. It should be run
 584  * early enough so that we never completely run out of journal buckets.
 585  *
 586  * High watermarks for triggering background reclaim:
 587  * - FIFO has fewer than 512 entries left
 588  * - fewer than 25% journal buckets free
 589  *
 590  * Background reclaim runs until low watermarks are reached:
 591  * - FIFO has more than 1024 entries left
 592  * - more than 50% journal buckets free
 593  *
 594  * As long as a reclaim can complete in the time it takes to fill up
 595  * 512 journal entries or 25% of all journal buckets, then
 596  * journal_next_bucket() should not stall.
 597  */
 598 static int __bch2_journal_reclaim(struct journal *j, bool direct, bool kicked)
 599 {
 600         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 601         bool kthread = (current->flags & PF_KTHREAD) != 0;
 602         u64 seq_to_flush;
 603         size_t min_nr, min_key_cache, nr_flushed;
 604         unsigned flags;
 605         int ret = 0;
 606
 607         /*
 608          * We can't invoke memory reclaim while holding the reclaim_lock -
 609          * journal reclaim is required to make progress for memory reclaim
 610          * (cleaning the caches), so we can't get stuck in memory reclaim while
 611          * we're holding the reclaim lock:
 612          */
 613         lockdep_assert_held(&j->reclaim_lock);
 614         flags = memalloc_noreclaim_save();
 615
 616         do {
 617                 if (kthread && kthread_should_stop())
 618                         break;
 619
 620                 if (bch2_journal_error(j)) {
 621                         ret = -EIO;
 622                         break;
 623                 }
 624
 625                 bch2_journal_do_discards(j);
 626
 627                 seq_to_flush = journal_seq_to_flush(j);
 628                 min_nr = 0;
 629
 630                 /*
 631                  * If it's been longer than j->reclaim_delay_ms since we last flushed,
 632                  * make sure to flush at least one journal pin:
 633                  */
 634                 if (time_after(jiffies, j->last_flushed +
 635                                msecs_to_jiffies(c->opts.journal_reclaim_delay)))
 636                         min_nr = 1;
 637
 638                 if (j->prereserved.reserved * 4 > j->prereserved.remaining)
 639                         min_nr = 1;
 640
 641                 if (fifo_free(&j->pin) <= 32)
 642                         min_nr = 1;
 643
 644                 if (atomic_read(&c->btree_cache.dirty) * 2 > c->btree_cache.used)
 645                         min_nr = 1;
 646
 647                 min_key_cache = min(bch2_nr_btree_keys_need_flush(c), (size_t) 128);
 648
 649                 trace_and_count(c, journal_reclaim_start, c,
 650                                 direct, kicked,
 651                                 min_nr, min_key_cache,
 652                                 j->prereserved.reserved,
 653                                 j->prereserved.remaining,
 654                                 atomic_read(&c->btree_cache.dirty),
 655                                 c->btree_cache.used,
 656                                 atomic_long_read(&c->btree_key_cache.nr_dirty),
 657                                 atomic_long_read(&c->btree_key_cache.nr_keys));
 658
 659                 nr_flushed = journal_flush_pins(j, seq_to_flush,
 660                                                 ~0, 0,
 661                                                 min_nr, min_key_cache);
 662
 663                 if (direct)
 664                         j->nr_direct_reclaim += nr_flushed;
 665                 else
 666                         j->nr_background_reclaim += nr_flushed;
 667                 trace_and_count(c, journal_reclaim_finish, c, nr_flushed);
 668
 669                 if (nr_flushed)
 670                         wake_up(&j->reclaim_wait);
 671         } while ((min_nr || min_key_cache) && nr_flushed && !direct);
 672
 673         memalloc_noreclaim_restore(flags);
 674
 675         return ret;
 676 }
 677
 678 int bch2_journal_reclaim(struct journal *j)
 679 {
 680         return __bch2_journal_reclaim(j, true, true);
 681 }
 682
 683 static int bch2_journal_reclaim_thread(void *arg)
 684 {
 685         struct journal *j = arg;
 686         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 687         unsigned long delay, now;
 688         bool journal_empty;
 689         int ret = 0;
 690
 691         set_freezable();
 692
 693         j->last_flushed = jiffies;
 694
 695         while (!ret && !kthread_should_stop()) {
 696                 bool kicked = j->reclaim_kicked;
 697
 698                 j->reclaim_kicked = false;
 699
 700                 mutex_lock(&j->reclaim_lock);
 701                 ret = __bch2_journal_reclaim(j, false, kicked);
 702                 mutex_unlock(&j->reclaim_lock);
 703
 704                 now = jiffies;
 705                 delay = msecs_to_jiffies(c->opts.journal_reclaim_delay);
 706                 j->next_reclaim = j->last_flushed + delay;
 707
 708                 if (!time_in_range(j->next_reclaim, now, now + delay))
 709                         j->next_reclaim = now + delay;
 710
 711                 while (1) {
 712                         set_current_state(TASK_INTERRUPTIBLE|TASK_FREEZABLE);
 713                         if (kthread_should_stop())
 714                                 break;
 715                         if (j->reclaim_kicked)
 716                                 break;
 717
 718                         spin_lock(&j->lock);
 719                         journal_empty = fifo_empty(&j->pin);
 720                         spin_unlock(&j->lock);
 721
 722                         if (journal_empty)
 723                                 schedule();
 724                         else if (time_after(j->next_reclaim, jiffies))
 725                                 schedule_timeout(j->next_reclaim - jiffies);
 726                         else
 727                                 break;
 728                 }
 729                 __set_current_state(TASK_RUNNING);
 730         }
 731
 732         return 0;
 733 }
 734
 735 void bch2_journal_reclaim_stop(struct journal *j)
 736 {
 737         struct task_struct *p = j->reclaim_thread;
 738
 739         j->reclaim_thread = NULL;
 740
 741         if (p) {
 742                 kthread_stop(p);
 743                 put_task_struct(p);
 744         }
 745 }
 746
 747 int bch2_journal_reclaim_start(struct journal *j)
 748 {
 749         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 750         struct task_struct *p;
 751         int ret;
 752
 753         if (j->reclaim_thread)
 754                 return 0;
 755
 756         p = kthread_create(bch2_journal_reclaim_thread, j,
 757                            "bch-reclaim/%s", c->name);
 758         ret = PTR_ERR_OR_ZERO(p);
 759         if (ret) {
 760                 bch_err(c, "error creating journal reclaim thread: %s", bch2_err_str(ret));
 761                 return ret;
 762         }
 763
 764         get_task_struct(p);
 765         j->reclaim_thread = p;
 766         wake_up_process(p);
 767         return 0;
 768 }
 769
 770 static int journal_flush_done(struct journal *j, u64 seq_to_flush,
 771                               bool *did_work)
 772 {
 773         int ret;
 774
 775         ret = bch2_journal_error(j);
 776         if (ret)
 777                 return ret;
 778
 779         mutex_lock(&j->reclaim_lock);
 780
 781         if (journal_flush_pins(j, seq_to_flush,
 782                                (1U << JOURNAL_PIN_key_cache)|
 783                                (1U << JOURNAL_PIN_other), 0, 0, 0) ||
 784             journal_flush_pins(j, seq_to_flush,
 785                                (1U << JOURNAL_PIN_btree), 0, 0, 0))
 786                 *did_work = true;
 787
 788         spin_lock(&j->lock);
 789         /*
 790          * If journal replay hasn't completed, the unreplayed journal entries
 791          * hold refs on their corresponding sequence numbers
 792          */
 793         ret = !test_bit(JOURNAL_REPLAY_DONE, &j->flags) ||
 794                 journal_last_seq(j) > seq_to_flush ||
 795                 !fifo_used(&j->pin);
 796
 797         spin_unlock(&j->lock);
 798         mutex_unlock(&j->reclaim_lock);
 799
 800         return ret;
 801 }
 802
 803 bool bch2_journal_flush_pins(struct journal *j, u64 seq_to_flush)
 804 {
 805         bool did_work = false;
 806
 807         if (!test_bit(JOURNAL_STARTED, &j->flags))
 808                 return false;
 809
 810         closure_wait_event(&j->async_wait,
 811                 journal_flush_done(j, seq_to_flush, &did_work));
 812
 813         return did_work;
 814 }
 815
 816 int bch2_journal_flush_device_pins(struct journal *j, int dev_idx)
 817 {
 818         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 819         struct journal_entry_pin_list *p;
 820         u64 iter, seq = 0;
 821         int ret = 0;
 822
 823         spin_lock(&j->lock);
 824         fifo_for_each_entry_ptr(p, &j->pin, iter)
 825                 if (dev_idx >= 0
 826                     ? bch2_dev_list_has_dev(p->devs, dev_idx)
 827                     : p->devs.nr < c->opts.metadata_replicas)
 828                         seq = iter;
 829         spin_unlock(&j->lock);
 830
 831         bch2_journal_flush_pins(j, seq);
 832
 833         ret = bch2_journal_error(j);
 834         if (ret)
 835                 return ret;
 836
 837         mutex_lock(&c->replicas_gc_lock);
 838         bch2_replicas_gc_start(c, 1 << BCH_DATA_journal);
 839
 840         seq = 0;
 841
 842         spin_lock(&j->lock);
 843         while (!ret) {
 844                 struct bch_replicas_padded replicas;
 845
 846                 seq = max(seq, journal_last_seq(j));
 847                 if (seq >= j->pin.back)
 848                         break;
 849                 bch2_devlist_to_replicas(&replicas.e, BCH_DATA_journal,
 850                                          journal_seq_pin(j, seq)->devs);
 851                 seq++;
 852
 853                 spin_unlock(&j->lock);
 854                 ret = bch2_mark_replicas(c, &replicas.e);
 855                 spin_lock(&j->lock);
 856         }
 857         spin_unlock(&j->lock);
 858
 859         ret = bch2_replicas_gc_end(c, ret);
 860         mutex_unlock(&c->replicas_gc_lock);
 861
 862         return ret;
 863 }