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 "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 void journal_set_remaining(struct journal *j, unsigned u64s_remaining)
  54 {
  55         union journal_preres_state old, new;
  56         u64 v = atomic64_read(&j->prereserved.counter);
  57
  58         do {
  59                 old.v = new.v = v;
  60                 new.remaining = u64s_remaining;
  61         } while ((v = atomic64_cmpxchg(&j->prereserved.counter,
  62                                        old.v, new.v)) != old.v);
  63 }
  64
  65 static struct journal_space
  66 journal_dev_space_available(struct journal *j, struct bch_dev *ca,
  67                             enum journal_space_from from)
  68 {
  69         struct journal_device *ja = &ca->journal;
  70         unsigned sectors, buckets, unwritten;
  71         u64 seq;
  72
  73         if (from == journal_space_total)
  74                 return (struct journal_space) {
  75                         .next_entry     = ca->mi.bucket_size,
  76                         .total          = ca->mi.bucket_size * ja->nr,
  77                 };
  78
  79         buckets = bch2_journal_dev_buckets_available(j, ja, from);
  80         sectors = ja->sectors_free;
  81
  82         /*
  83          * We that we don't allocate the space for a journal entry
  84          * until we write it out - thus, account for it here:
  85          */
  86         for (seq = journal_last_unwritten_seq(j);
  87              seq <= journal_cur_seq(j);
  88              seq++) {
  89                 unwritten = j->buf[seq & JOURNAL_BUF_MASK].sectors;
  90
  91                 if (!unwritten)
  92                         continue;
  93
  94                 /* entry won't fit on this device, skip: */
  95                 if (unwritten > ca->mi.bucket_size)
  96                         continue;
  97
  98                 if (unwritten >= sectors) {
  99                         if (!buckets) {
 100                                 sectors = 0;
 101                                 break;
 102                         }
 103
 104                         buckets--;
 105                         sectors = ca->mi.bucket_size;
 106                 }
 107
 108                 sectors -= unwritten;
 109         }
 110
 111         if (sectors < ca->mi.bucket_size && buckets) {
 112                 buckets--;
 113                 sectors = ca->mi.bucket_size;
 114         }
 115
 116         return (struct journal_space) {
 117                 .next_entry     = sectors,
 118                 .total          = sectors + buckets * ca->mi.bucket_size,
 119         };
 120 }
 121
 122 static struct journal_space __journal_space_available(struct journal *j, unsigned nr_devs_want,
 123                             enum journal_space_from from)
 124 {
 125         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 126         struct bch_dev *ca;
 127         unsigned i, pos, nr_devs = 0;
 128         struct journal_space space, dev_space[BCH_SB_MEMBERS_MAX];
 129
 130         BUG_ON(nr_devs_want > ARRAY_SIZE(dev_space));
 131
 132         rcu_read_lock();
 133         for_each_member_device_rcu(ca, c, i,
 134                                    &c->rw_devs[BCH_DATA_journal]) {
 135                 if (!ca->journal.nr)
 136                         continue;
 137
 138                 space = journal_dev_space_available(j, ca, from);
 139                 if (!space.next_entry)
 140                         continue;
 141
 142                 for (pos = 0; pos < nr_devs; pos++)
 143                         if (space.total > dev_space[pos].total)
 144                                 break;
 145
 146                 array_insert_item(dev_space, nr_devs, pos, space);
 147         }
 148         rcu_read_unlock();
 149
 150         if (nr_devs < nr_devs_want)
 151                 return (struct journal_space) { 0, 0 };
 152
 153         /*
 154          * We sorted largest to smallest, and we want the smallest out of the
 155          * @nr_devs_want largest devices:
 156          */
 157         return dev_space[nr_devs_want - 1];
 158 }
 159
 160 void bch2_journal_space_available(struct journal *j)
 161 {
 162         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 163         struct bch_dev *ca;
 164         unsigned clean, clean_ondisk, total;
 165         s64 u64s_remaining = 0;
 166         unsigned max_entry_size  = min(j->buf[0].buf_size >> 9,
 167                                        j->buf[1].buf_size >> 9);
 168         unsigned i, nr_online = 0, nr_devs_want;
 169         bool can_discard = false;
 170         int ret = 0;
 171
 172         lockdep_assert_held(&j->lock);
 173
 174         rcu_read_lock();
 175         for_each_member_device_rcu(ca, c, i,
 176                                    &c->rw_devs[BCH_DATA_journal]) {
 177                 struct journal_device *ja = &ca->journal;
 178
 179                 if (!ja->nr)
 180                         continue;
 181
 182                 while (ja->dirty_idx != ja->cur_idx &&
 183                        ja->bucket_seq[ja->dirty_idx] < journal_last_seq(j))
 184                         ja->dirty_idx = (ja->dirty_idx + 1) % ja->nr;
 185
 186                 while (ja->dirty_idx_ondisk != ja->dirty_idx &&
 187                        ja->bucket_seq[ja->dirty_idx_ondisk] < j->last_seq_ondisk)
 188                         ja->dirty_idx_ondisk = (ja->dirty_idx_ondisk + 1) % ja->nr;
 189
 190                 if (ja->discard_idx != ja->dirty_idx_ondisk)
 191                         can_discard = true;
 192
 193                 max_entry_size = min_t(unsigned, max_entry_size, ca->mi.bucket_size);
 194                 nr_online++;
 195         }
 196         rcu_read_unlock();
 197
 198         j->can_discard = can_discard;
 199
 200         if (nr_online < c->opts.metadata_replicas_required) {
 201                 ret = JOURNAL_ERR_insufficient_devices;
 202                 goto out;
 203         }
 204
 205         nr_devs_want = min_t(unsigned, nr_online, c->opts.metadata_replicas);
 206
 207         for (i = 0; i < journal_space_nr; i++)
 208                 j->space[i] = __journal_space_available(j, nr_devs_want, i);
 209
 210         clean_ondisk    = j->space[journal_space_clean_ondisk].total;
 211         clean           = j->space[journal_space_clean].total;
 212         total           = j->space[journal_space_total].total;
 213
 214         if (!j->space[journal_space_discarded].next_entry)
 215                 ret = JOURNAL_ERR_journal_full;
 216
 217         if ((j->space[journal_space_clean_ondisk].next_entry <
 218              j->space[journal_space_clean_ondisk].total) &&
 219             (clean - clean_ondisk <= total / 8) &&
 220             (clean_ondisk * 2 > clean))
 221                 set_bit(JOURNAL_MAY_SKIP_FLUSH, &j->flags);
 222         else
 223                 clear_bit(JOURNAL_MAY_SKIP_FLUSH, &j->flags);
 224
 225         u64s_remaining  = (u64) clean << 6;
 226         u64s_remaining -= (u64) total << 3;
 227         u64s_remaining = max(0LL, u64s_remaining);
 228         u64s_remaining /= 4;
 229         u64s_remaining = min_t(u64, u64s_remaining, U32_MAX);
 230 out:
 231         j->cur_entry_sectors    = !ret ? j->space[journal_space_discarded].next_entry : 0;
 232         j->cur_entry_error      = ret;
 233         journal_set_remaining(j, u64s_remaining);
 234         journal_set_watermark(j);
 235
 236         if (!ret)
 237                 journal_wake(j);
 238 }
 239
 240 /* Discards - last part of journal reclaim: */
 241
 242 static bool should_discard_bucket(struct journal *j, struct journal_device *ja)
 243 {
 244         bool ret;
 245
 246         spin_lock(&j->lock);
 247         ret = ja->discard_idx != ja->dirty_idx_ondisk;
 248         spin_unlock(&j->lock);
 249
 250         return ret;
 251 }
 252
 253 /*
 254  * Advance ja->discard_idx as long as it points to buckets that are no longer
 255  * dirty, issuing discards if necessary:
 256  */
 257 void bch2_journal_do_discards(struct journal *j)
 258 {
 259         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 260         struct bch_dev *ca;
 261         unsigned iter;
 262
 263         mutex_lock(&j->discard_lock);
 264
 265         for_each_rw_member(ca, c, iter) {
 266                 struct journal_device *ja = &ca->journal;
 267
 268                 while (should_discard_bucket(j, ja)) {
 269                         if (!c->opts.nochanges &&
 270                             ca->mi.discard &&
 271                             bdev_max_discard_sectors(ca->disk_sb.bdev))
 272                                 blkdev_issue_discard(ca->disk_sb.bdev,
 273                                         bucket_to_sector(ca,
 274                                                 ja->buckets[ja->discard_idx]),
 275                                         ca->mi.bucket_size, GFP_NOFS);
 276
 277                         spin_lock(&j->lock);
 278                         ja->discard_idx = (ja->discard_idx + 1) % ja->nr;
 279
 280                         bch2_journal_space_available(j);
 281                         spin_unlock(&j->lock);
 282                 }
 283         }
 284
 285         mutex_unlock(&j->discard_lock);
 286 }
 287
 288 /*
 289  * Journal entry pinning - machinery for holding a reference on a given journal
 290  * entry, holding it open to ensure it gets replayed during recovery:
 291  */
 292
 293 void bch2_journal_reclaim_fast(struct journal *j)
 294 {
 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                 j->pin.front++;
 306                 popped = true;
 307         }
 308
 309         if (popped)
 310                 bch2_journal_space_available(j);
 311 }
 312
 313 bool __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         return atomic_dec_and_test(&pin_list->count);
 318 }
 319
 320 void bch2_journal_pin_put(struct journal *j, u64 seq)
 321 {
 322         if (__bch2_journal_pin_put(j, seq)) {
 323                 spin_lock(&j->lock);
 324                 bch2_journal_reclaim_fast(j);
 325                 spin_unlock(&j->lock);
 326         }
 327 }
 328
 329 static inline bool __journal_pin_drop(struct journal *j,
 330                                       struct journal_entry_pin *pin)
 331 {
 332         struct journal_entry_pin_list *pin_list;
 333
 334         if (!journal_pin_active(pin))
 335                 return false;
 336
 337         if (j->flush_in_progress == pin)
 338                 j->flush_in_progress_dropped = true;
 339
 340         pin_list = journal_seq_pin(j, pin->seq);
 341         pin->seq = 0;
 342         list_del_init(&pin->list);
 343
 344         /*
 345          * Unpinning a journal entry may make journal_next_bucket() succeed, if
 346          * writing a new last_seq will now make another bucket available:
 347          */
 348         return atomic_dec_and_test(&pin_list->count) &&
 349                 pin_list == &fifo_peek_front(&j->pin);
 350 }
 351
 352 void bch2_journal_pin_drop(struct journal *j,
 353                            struct journal_entry_pin *pin)
 354 {
 355         spin_lock(&j->lock);
 356         if (__journal_pin_drop(j, pin))
 357                 bch2_journal_reclaim_fast(j);
 358         spin_unlock(&j->lock);
 359 }
 360
 361 static enum journal_pin_type journal_pin_type(journal_pin_flush_fn fn)
 362 {
 363         if (fn == bch2_btree_node_flush0 ||
 364             fn == bch2_btree_node_flush1)
 365                 return JOURNAL_PIN_btree;
 366         else if (fn == bch2_btree_key_cache_journal_flush)
 367                 return JOURNAL_PIN_key_cache;
 368         else
 369                 return JOURNAL_PIN_other;
 370 }
 371
 372 void bch2_journal_pin_set(struct journal *j, u64 seq,
 373                           struct journal_entry_pin *pin,
 374                           journal_pin_flush_fn flush_fn)
 375 {
 376         struct journal_entry_pin_list *pin_list;
 377         bool reclaim;
 378
 379         spin_lock(&j->lock);
 380
 381         if (seq < journal_last_seq(j)) {
 382                 /*
 383                  * bch2_journal_pin_copy() raced with bch2_journal_pin_drop() on
 384                  * the src pin - with the pin dropped, the entry to pin might no
 385                  * longer to exist, but that means there's no longer anything to
 386                  * copy and we can bail out here:
 387                  */
 388                 spin_unlock(&j->lock);
 389                 return;
 390         }
 391
 392         pin_list = journal_seq_pin(j, seq);
 393
 394         reclaim = __journal_pin_drop(j, pin);
 395
 396         atomic_inc(&pin_list->count);
 397         pin->seq        = seq;
 398         pin->flush      = flush_fn;
 399
 400         if (flush_fn)
 401                 list_add(&pin->list, &pin_list->list[journal_pin_type(flush_fn)]);
 402         else
 403                 list_add(&pin->list, &pin_list->flushed);
 404
 405         if (reclaim)
 406                 bch2_journal_reclaim_fast(j);
 407         spin_unlock(&j->lock);
 408
 409         /*
 410          * If the journal is currently full,  we might want to call flush_fn
 411          * immediately:
 412          */
 413         journal_wake(j);
 414 }
 415
 416 /**
 417  * bch2_journal_pin_flush: ensure journal pin callback is no longer running
 418  * @j:          journal object
 419  * @pin:        pin to flush
 420  */
 421 void bch2_journal_pin_flush(struct journal *j, struct journal_entry_pin *pin)
 422 {
 423         BUG_ON(journal_pin_active(pin));
 424
 425         wait_event(j->pin_flush_wait, j->flush_in_progress != pin);
 426 }
 427
 428 /*
 429  * Journal reclaim: flush references to open journal entries to reclaim space in
 430  * the journal
 431  *
 432  * May be done by the journal code in the background as needed to free up space
 433  * for more journal entries, or as part of doing a clean shutdown, or to migrate
 434  * data off of a specific device:
 435  */
 436
 437 static struct journal_entry_pin *
 438 journal_get_next_pin(struct journal *j,
 439                      u64 seq_to_flush,
 440                      unsigned allowed_below_seq,
 441                      unsigned allowed_above_seq,
 442                      u64 *seq)
 443 {
 444         struct journal_entry_pin_list *pin_list;
 445         struct journal_entry_pin *ret = NULL;
 446         unsigned i;
 447
 448         fifo_for_each_entry_ptr(pin_list, &j->pin, *seq) {
 449                 if (*seq > seq_to_flush && !allowed_above_seq)
 450                         break;
 451
 452                 for (i = 0; i < JOURNAL_PIN_NR; i++)
 453                         if ((((1U << i) & allowed_below_seq) && *seq <= seq_to_flush) ||
 454                             ((1U << i) & allowed_above_seq)) {
 455                                 ret = list_first_entry_or_null(&pin_list->list[i],
 456                                         struct journal_entry_pin, list);
 457                                 if (ret)
 458                                         return ret;
 459                         }
 460         }
 461
 462         return NULL;
 463 }
 464
 465 /* returns true if we did work */
 466 static size_t journal_flush_pins(struct journal *j,
 467                                  u64 seq_to_flush,
 468                                  unsigned allowed_below_seq,
 469                                  unsigned allowed_above_seq,
 470                                  unsigned min_any,
 471                                  unsigned min_key_cache)
 472 {
 473         struct journal_entry_pin *pin;
 474         size_t nr_flushed = 0;
 475         journal_pin_flush_fn flush_fn;
 476         u64 seq;
 477         int err;
 478
 479         lockdep_assert_held(&j->reclaim_lock);
 480
 481         while (1) {
 482                 unsigned allowed_above = allowed_above_seq;
 483                 unsigned allowed_below = allowed_below_seq;
 484
 485                 if (min_any) {
 486                         allowed_above |= ~0;
 487                         allowed_below |= ~0;
 488                 }
 489
 490                 if (min_key_cache) {
 491                         allowed_above |= 1U << JOURNAL_PIN_key_cache;
 492                         allowed_below |= 1U << JOURNAL_PIN_key_cache;
 493                 }
 494
 495                 cond_resched();
 496
 497                 j->last_flushed = jiffies;
 498
 499                 spin_lock(&j->lock);
 500                 pin = journal_get_next_pin(j, seq_to_flush, allowed_below, allowed_above, &seq);
 501                 if (pin) {
 502                         BUG_ON(j->flush_in_progress);
 503                         j->flush_in_progress = pin;
 504                         j->flush_in_progress_dropped = false;
 505                         flush_fn = pin->flush;
 506                 }
 507                 spin_unlock(&j->lock);
 508
 509                 if (!pin)
 510                         break;
 511
 512                 if (min_key_cache && pin->flush == bch2_btree_key_cache_journal_flush)
 513                         min_key_cache--;
 514
 515                 if (min_any)
 516                         min_any--;
 517
 518                 err = flush_fn(j, pin, seq);
 519
 520                 spin_lock(&j->lock);
 521                 /* Pin might have been dropped or rearmed: */
 522                 if (likely(!err && !j->flush_in_progress_dropped))
 523                         list_move(&pin->list, &journal_seq_pin(j, seq)->flushed);
 524                 j->flush_in_progress = NULL;
 525                 j->flush_in_progress_dropped = false;
 526                 spin_unlock(&j->lock);
 527
 528                 wake_up(&j->pin_flush_wait);
 529
 530                 if (err)
 531                         break;
 532
 533                 nr_flushed++;
 534         }
 535
 536         return nr_flushed;
 537 }
 538
 539 static u64 journal_seq_to_flush(struct journal *j)
 540 {
 541         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 542         struct bch_dev *ca;
 543         u64 seq_to_flush = 0;
 544         unsigned iter;
 545
 546         spin_lock(&j->lock);
 547
 548         for_each_rw_member(ca, c, iter) {
 549                 struct journal_device *ja = &ca->journal;
 550                 unsigned nr_buckets, bucket_to_flush;
 551
 552                 if (!ja->nr)
 553                         continue;
 554
 555                 /* Try to keep the journal at most half full: */
 556                 nr_buckets = ja->nr / 2;
 557
 558                 /* And include pre-reservations: */
 559                 nr_buckets += DIV_ROUND_UP(j->prereserved.reserved,
 560                                            (ca->mi.bucket_size << 6) -
 561                                            journal_entry_overhead(j));
 562
 563                 nr_buckets = min(nr_buckets, ja->nr);
 564
 565                 bucket_to_flush = (ja->cur_idx + nr_buckets) % ja->nr;
 566                 seq_to_flush = max(seq_to_flush,
 567                                    ja->bucket_seq[bucket_to_flush]);
 568         }
 569
 570         /* Also flush if the pin fifo is more than half full */
 571         seq_to_flush = max_t(s64, seq_to_flush,
 572                              (s64) journal_cur_seq(j) -
 573                              (j->pin.size >> 1));
 574         spin_unlock(&j->lock);
 575
 576         return seq_to_flush;
 577 }
 578
 579 /**
 580  * __bch2_journal_reclaim - free up journal buckets
 581  * @j:          journal object
 582  * @direct:     direct or background reclaim?
 583  * @kicked:     requested to run since we last ran?
 584  * Returns:     0 on success, or -EIO if the journal has been shutdown
 585  *
 586  * Background journal reclaim writes out btree nodes. It should be run
 587  * early enough so that we never completely run out of journal buckets.
 588  *
 589  * High watermarks for triggering background reclaim:
 590  * - FIFO has fewer than 512 entries left
 591  * - fewer than 25% journal buckets free
 592  *
 593  * Background reclaim runs until low watermarks are reached:
 594  * - FIFO has more than 1024 entries left
 595  * - more than 50% journal buckets free
 596  *
 597  * As long as a reclaim can complete in the time it takes to fill up
 598  * 512 journal entries or 25% of all journal buckets, then
 599  * journal_next_bucket() should not stall.
 600  */
 601 static int __bch2_journal_reclaim(struct journal *j, bool direct, bool kicked)
 602 {
 603         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 604         bool kthread = (current->flags & PF_KTHREAD) != 0;
 605         u64 seq_to_flush;
 606         size_t min_nr, min_key_cache, nr_flushed;
 607         unsigned flags;
 608         int ret = 0;
 609
 610         /*
 611          * We can't invoke memory reclaim while holding the reclaim_lock -
 612          * journal reclaim is required to make progress for memory reclaim
 613          * (cleaning the caches), so we can't get stuck in memory reclaim while
 614          * we're holding the reclaim lock:
 615          */
 616         lockdep_assert_held(&j->reclaim_lock);
 617         flags = memalloc_noreclaim_save();
 618
 619         do {
 620                 if (kthread && kthread_should_stop())
 621                         break;
 622
 623                 if (bch2_journal_error(j)) {
 624                         ret = -EIO;
 625                         break;
 626                 }
 627
 628                 bch2_journal_do_discards(j);
 629
 630                 seq_to_flush = journal_seq_to_flush(j);
 631                 min_nr = 0;
 632
 633                 /*
 634                  * If it's been longer than j->reclaim_delay_ms since we last flushed,
 635                  * make sure to flush at least one journal pin:
 636                  */
 637                 if (time_after(jiffies, j->last_flushed +
 638                                msecs_to_jiffies(c->opts.journal_reclaim_delay)))
 639                         min_nr = 1;
 640
 641                 if (j->prereserved.reserved * 4 > j->prereserved.remaining)
 642                         min_nr = 1;
 643
 644                 if (fifo_free(&j->pin) <= 32)
 645                         min_nr = 1;
 646
 647                 if (atomic_read(&c->btree_cache.dirty) * 2 > c->btree_cache.used)
 648                         min_nr = 1;
 649
 650                 min_key_cache = min(bch2_nr_btree_keys_need_flush(c), (size_t) 128);
 651
 652                 trace_and_count(c, journal_reclaim_start, c,
 653                                 direct, kicked,
 654                                 min_nr, min_key_cache,
 655                                 j->prereserved.reserved,
 656                                 j->prereserved.remaining,
 657                                 atomic_read(&c->btree_cache.dirty),
 658                                 c->btree_cache.used,
 659                                 atomic_long_read(&c->btree_key_cache.nr_dirty),
 660                                 atomic_long_read(&c->btree_key_cache.nr_keys));
 661
 662                 nr_flushed = journal_flush_pins(j, seq_to_flush,
 663                                                 ~0, 0,
 664                                                 min_nr, min_key_cache);
 665
 666                 if (direct)
 667                         j->nr_direct_reclaim += nr_flushed;
 668                 else
 669                         j->nr_background_reclaim += nr_flushed;
 670                 trace_and_count(c, journal_reclaim_finish, c, nr_flushed);
 671
 672                 if (nr_flushed)
 673                         wake_up(&j->reclaim_wait);
 674         } while ((min_nr || min_key_cache) && nr_flushed && !direct);
 675
 676         memalloc_noreclaim_restore(flags);
 677
 678         return ret;
 679 }
 680
 681 int bch2_journal_reclaim(struct journal *j)
 682 {
 683         return __bch2_journal_reclaim(j, true, true);
 684 }
 685
 686 static int bch2_journal_reclaim_thread(void *arg)
 687 {
 688         struct journal *j = arg;
 689         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 690         unsigned long delay, now;
 691         bool journal_empty;
 692         int ret = 0;
 693
 694         set_freezable();
 695
 696         j->last_flushed = jiffies;
 697
 698         while (!ret && !kthread_should_stop()) {
 699                 bool kicked = j->reclaim_kicked;
 700
 701                 j->reclaim_kicked = false;
 702
 703                 mutex_lock(&j->reclaim_lock);
 704                 ret = __bch2_journal_reclaim(j, false, kicked);
 705                 mutex_unlock(&j->reclaim_lock);
 706
 707                 now = jiffies;
 708                 delay = msecs_to_jiffies(c->opts.journal_reclaim_delay);
 709                 j->next_reclaim = j->last_flushed + delay;
 710
 711                 if (!time_in_range(j->next_reclaim, now, now + delay))
 712                         j->next_reclaim = now + delay;
 713
 714                 while (1) {
 715                         set_current_state(TASK_INTERRUPTIBLE|TASK_FREEZABLE);
 716                         if (kthread_should_stop())
 717                                 break;
 718                         if (j->reclaim_kicked)
 719                                 break;
 720
 721                         spin_lock(&j->lock);
 722                         journal_empty = fifo_empty(&j->pin);
 723                         spin_unlock(&j->lock);
 724
 725                         if (journal_empty)
 726                                 schedule();
 727                         else if (time_after(j->next_reclaim, jiffies))
 728                                 schedule_timeout(j->next_reclaim - jiffies);
 729                         else
 730                                 break;
 731                 }
 732                 __set_current_state(TASK_RUNNING);
 733         }
 734
 735         return 0;
 736 }
 737
 738 void bch2_journal_reclaim_stop(struct journal *j)
 739 {
 740         struct task_struct *p = j->reclaim_thread;
 741
 742         j->reclaim_thread = NULL;
 743
 744         if (p) {
 745                 kthread_stop(p);
 746                 put_task_struct(p);
 747         }
 748 }
 749
 750 int bch2_journal_reclaim_start(struct journal *j)
 751 {
 752         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 753         struct task_struct *p;
 754         int ret;
 755
 756         if (j->reclaim_thread)
 757                 return 0;
 758
 759         p = kthread_create(bch2_journal_reclaim_thread, j,
 760                            "bch-reclaim/%s", c->name);
 761         ret = PTR_ERR_OR_ZERO(p);
 762         if (ret) {
 763                 bch_err_msg(c, ret, "creating journal reclaim thread");
 764                 return ret;
 765         }
 766
 767         get_task_struct(p);
 768         j->reclaim_thread = p;
 769         wake_up_process(p);
 770         return 0;
 771 }
 772
 773 static int journal_flush_done(struct journal *j, u64 seq_to_flush,
 774                               bool *did_work)
 775 {
 776         int ret;
 777
 778         ret = bch2_journal_error(j);
 779         if (ret)
 780                 return ret;
 781
 782         mutex_lock(&j->reclaim_lock);
 783
 784         if (journal_flush_pins(j, seq_to_flush,
 785                                (1U << JOURNAL_PIN_key_cache)|
 786                                (1U << JOURNAL_PIN_other), 0, 0, 0) ||
 787             journal_flush_pins(j, seq_to_flush,
 788                                (1U << JOURNAL_PIN_btree), 0, 0, 0))
 789                 *did_work = true;
 790
 791         spin_lock(&j->lock);
 792         /*
 793          * If journal replay hasn't completed, the unreplayed journal entries
 794          * hold refs on their corresponding sequence numbers
 795          */
 796         ret = !test_bit(JOURNAL_REPLAY_DONE, &j->flags) ||
 797                 journal_last_seq(j) > seq_to_flush ||
 798                 !fifo_used(&j->pin);
 799
 800         spin_unlock(&j->lock);
 801         mutex_unlock(&j->reclaim_lock);
 802
 803         return ret;
 804 }
 805
 806 bool bch2_journal_flush_pins(struct journal *j, u64 seq_to_flush)
 807 {
 808         bool did_work = false;
 809
 810         if (!test_bit(JOURNAL_STARTED, &j->flags))
 811                 return false;
 812
 813         closure_wait_event(&j->async_wait,
 814                 journal_flush_done(j, seq_to_flush, &did_work));
 815
 816         return did_work;
 817 }
 818
 819 int bch2_journal_flush_device_pins(struct journal *j, int dev_idx)
 820 {
 821         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 822         struct journal_entry_pin_list *p;
 823         u64 iter, seq = 0;
 824         int ret = 0;
 825
 826         spin_lock(&j->lock);
 827         fifo_for_each_entry_ptr(p, &j->pin, iter)
 828                 if (dev_idx >= 0
 829                     ? bch2_dev_list_has_dev(p->devs, dev_idx)
 830                     : p->devs.nr < c->opts.metadata_replicas)
 831                         seq = iter;
 832         spin_unlock(&j->lock);
 833
 834         bch2_journal_flush_pins(j, seq);
 835
 836         ret = bch2_journal_error(j);
 837         if (ret)
 838                 return ret;
 839
 840         mutex_lock(&c->replicas_gc_lock);
 841         bch2_replicas_gc_start(c, 1 << BCH_DATA_journal);
 842
 843         /*
 844          * Now that we've populated replicas_gc, write to the journal to mark
 845          * active journal devices. This handles the case where the journal might
 846          * be empty. Otherwise we could clear all journal replicas and
 847          * temporarily put the fs into an unrecoverable state. Journal recovery
 848          * expects to find devices marked for journal data on unclean mount.
 849          */
 850         ret = bch2_journal_meta(&c->journal);
 851         if (ret)
 852                 goto err;
 853
 854         seq = 0;
 855         spin_lock(&j->lock);
 856         while (!ret) {
 857                 struct bch_replicas_padded replicas;
 858
 859                 seq = max(seq, journal_last_seq(j));
 860                 if (seq >= j->pin.back)
 861                         break;
 862                 bch2_devlist_to_replicas(&replicas.e, BCH_DATA_journal,
 863                                          journal_seq_pin(j, seq)->devs);
 864                 seq++;
 865
 866                 spin_unlock(&j->lock);
 867                 ret = bch2_mark_replicas(c, &replicas.e);
 868                 spin_lock(&j->lock);
 869         }
 870         spin_unlock(&j->lock);
 871 err:
 872         ret = bch2_replicas_gc_end(c, ret);
 873         mutex_unlock(&c->replicas_gc_lock);
 874
 875         return ret;
 876 }