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 static void bch2_journal_reclaim_fast(struct journal *j)
 294 {
 295         struct journal_entry_pin_list temp;
 296         bool popped = false;
 297
 298         lockdep_assert_held(&j->lock);
 299
 300         /*
 301          * Unpin journal entries whose reference counts reached zero, meaning
 302          * all btree nodes got written out
 303          */
 304         while (!fifo_empty(&j->pin) &&
 305                !atomic_read(&fifo_peek_front(&j->pin).count)) {
 306                 fifo_pop(&j->pin, temp);
 307                 popped = true;
 308         }
 309
 310         if (popped)
 311                 bch2_journal_space_available(j);
 312 }
 313
 314 void __bch2_journal_pin_put(struct journal *j, u64 seq)
 315 {
 316         struct journal_entry_pin_list *pin_list = journal_seq_pin(j, seq);
 317
 318         if (atomic_dec_and_test(&pin_list->count))
 319                 bch2_journal_reclaim_fast(j);
 320 }
 321
 322 void bch2_journal_pin_put(struct journal *j, u64 seq)
 323 {
 324         struct journal_entry_pin_list *pin_list = journal_seq_pin(j, seq);
 325
 326         if (atomic_dec_and_test(&pin_list->count)) {
 327                 spin_lock(&j->lock);
 328                 bch2_journal_reclaim_fast(j);
 329                 spin_unlock(&j->lock);
 330         }
 331 }
 332
 333 static inline bool __journal_pin_drop(struct journal *j,
 334                                       struct journal_entry_pin *pin)
 335 {
 336         struct journal_entry_pin_list *pin_list;
 337
 338         if (!journal_pin_active(pin))
 339                 return false;
 340
 341         if (j->flush_in_progress == pin)
 342                 j->flush_in_progress_dropped = true;
 343
 344         pin_list = journal_seq_pin(j, pin->seq);
 345         pin->seq = 0;
 346         list_del_init(&pin->list);
 347
 348         /*
 349          * Unpinning a journal entry may make journal_next_bucket() succeed, if
 350          * writing a new last_seq will now make another bucket available:
 351          */
 352         return atomic_dec_and_test(&pin_list->count) &&
 353                 pin_list == &fifo_peek_front(&j->pin);
 354 }
 355
 356 void bch2_journal_pin_drop(struct journal *j,
 357                            struct journal_entry_pin *pin)
 358 {
 359         spin_lock(&j->lock);
 360         if (__journal_pin_drop(j, pin))
 361                 bch2_journal_reclaim_fast(j);
 362         spin_unlock(&j->lock);
 363 }
 364
 365 static enum journal_pin_type journal_pin_type(journal_pin_flush_fn fn)
 366 {
 367         if (fn == bch2_btree_node_flush0 ||
 368             fn == bch2_btree_node_flush1)
 369                 return JOURNAL_PIN_btree;
 370         else if (fn == bch2_btree_key_cache_journal_flush)
 371                 return JOURNAL_PIN_key_cache;
 372         else
 373                 return JOURNAL_PIN_other;
 374 }
 375
 376 void bch2_journal_pin_set(struct journal *j, u64 seq,
 377                           struct journal_entry_pin *pin,
 378                           journal_pin_flush_fn flush_fn)
 379 {
 380         struct journal_entry_pin_list *pin_list;
 381         bool reclaim;
 382
 383         spin_lock(&j->lock);
 384
 385         if (seq < journal_last_seq(j)) {
 386                 /*
 387                  * bch2_journal_pin_copy() raced with bch2_journal_pin_drop() on
 388                  * the src pin - with the pin dropped, the entry to pin might no
 389                  * longer to exist, but that means there's no longer anything to
 390                  * copy and we can bail out here:
 391                  */
 392                 spin_unlock(&j->lock);
 393                 return;
 394         }
 395
 396         pin_list = journal_seq_pin(j, seq);
 397
 398         reclaim = __journal_pin_drop(j, pin);
 399
 400         atomic_inc(&pin_list->count);
 401         pin->seq        = seq;
 402         pin->flush      = flush_fn;
 403
 404         if (flush_fn)
 405                 list_add(&pin->list, &pin_list->list[journal_pin_type(flush_fn)]);
 406         else
 407                 list_add(&pin->list, &pin_list->flushed);
 408
 409         if (reclaim)
 410                 bch2_journal_reclaim_fast(j);
 411         spin_unlock(&j->lock);
 412
 413         /*
 414          * If the journal is currently full,  we might want to call flush_fn
 415          * immediately:
 416          */
 417         journal_wake(j);
 418 }
 419
 420 /**
 421  * bch2_journal_pin_flush: ensure journal pin callback is no longer running
 422  */
 423 void bch2_journal_pin_flush(struct journal *j, struct journal_entry_pin *pin)
 424 {
 425         BUG_ON(journal_pin_active(pin));
 426
 427         wait_event(j->pin_flush_wait, j->flush_in_progress != pin);
 428 }
 429
 430 /*
 431  * Journal reclaim: flush references to open journal entries to reclaim space in
 432  * the journal
 433  *
 434  * May be done by the journal code in the background as needed to free up space
 435  * for more journal entries, or as part of doing a clean shutdown, or to migrate
 436  * data off of a specific device:
 437  */
 438
 439 static struct journal_entry_pin *
 440 journal_get_next_pin(struct journal *j,
 441                      u64 seq_to_flush,
 442                      unsigned allowed_below_seq,
 443                      unsigned allowed_above_seq,
 444                      u64 *seq)
 445 {
 446         struct journal_entry_pin_list *pin_list;
 447         struct journal_entry_pin *ret = NULL;
 448         unsigned i;
 449
 450         fifo_for_each_entry_ptr(pin_list, &j->pin, *seq) {
 451                 if (*seq > seq_to_flush && !allowed_above_seq)
 452                         break;
 453
 454                 for (i = 0; i < JOURNAL_PIN_NR; i++)
 455                         if ((((1U << i) & allowed_below_seq) && *seq <= seq_to_flush) ||
 456                             ((1U << i) & allowed_above_seq)) {
 457                                 ret = list_first_entry_or_null(&pin_list->list[i],
 458                                         struct journal_entry_pin, list);
 459                                 if (ret)
 460                                         return ret;
 461                         }
 462         }
 463
 464         return NULL;
 465 }
 466
 467 /* returns true if we did work */
 468 static size_t journal_flush_pins(struct journal *j,
 469                                  u64 seq_to_flush,
 470                                  unsigned allowed_below_seq,
 471                                  unsigned allowed_above_seq,
 472                                  unsigned min_any,
 473                                  unsigned min_key_cache)
 474 {
 475         struct journal_entry_pin *pin;
 476         size_t nr_flushed = 0;
 477         journal_pin_flush_fn flush_fn;
 478         u64 seq;
 479         int err;
 480
 481         lockdep_assert_held(&j->reclaim_lock);
 482
 483         while (1) {
 484                 unsigned allowed_above = allowed_above_seq;
 485                 unsigned allowed_below = allowed_below_seq;
 486
 487                 if (min_any) {
 488                         allowed_above |= ~0;
 489                         allowed_below |= ~0;
 490                 }
 491
 492                 if (min_key_cache) {
 493                         allowed_above |= 1U << JOURNAL_PIN_key_cache;
 494                         allowed_below |= 1U << JOURNAL_PIN_key_cache;
 495                 }
 496
 497                 cond_resched();
 498
 499                 j->last_flushed = jiffies;
 500
 501                 spin_lock(&j->lock);
 502                 pin = journal_get_next_pin(j, seq_to_flush, allowed_below, allowed_above, &seq);
 503                 if (pin) {
 504                         BUG_ON(j->flush_in_progress);
 505                         j->flush_in_progress = pin;
 506                         j->flush_in_progress_dropped = false;
 507                         flush_fn = pin->flush;
 508                 }
 509                 spin_unlock(&j->lock);
 510
 511                 if (!pin)
 512                         break;
 513
 514                 if (min_key_cache && pin->flush == bch2_btree_key_cache_journal_flush)
 515                         min_key_cache--;
 516
 517                 if (min_any)
 518                         min_any--;
 519
 520                 err = flush_fn(j, pin, seq);
 521
 522                 spin_lock(&j->lock);
 523                 /* Pin might have been dropped or rearmed: */
 524                 if (likely(!err && !j->flush_in_progress_dropped))
 525                         list_move(&pin->list, &journal_seq_pin(j, seq)->flushed);
 526                 j->flush_in_progress = NULL;
 527                 j->flush_in_progress_dropped = false;
 528                 spin_unlock(&j->lock);
 529
 530                 wake_up(&j->pin_flush_wait);
 531
 532                 if (err)
 533                         break;
 534
 535                 nr_flushed++;
 536         }
 537
 538         return nr_flushed;
 539 }
 540
 541 static u64 journal_seq_to_flush(struct journal *j)
 542 {
 543         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 544         struct bch_dev *ca;
 545         u64 seq_to_flush = 0;
 546         unsigned iter;
 547
 548         spin_lock(&j->lock);
 549
 550         for_each_rw_member(ca, c, iter) {
 551                 struct journal_device *ja = &ca->journal;
 552                 unsigned nr_buckets, bucket_to_flush;
 553
 554                 if (!ja->nr)
 555                         continue;
 556
 557                 /* Try to keep the journal at most half full: */
 558                 nr_buckets = ja->nr / 2;
 559
 560                 /* And include pre-reservations: */
 561                 nr_buckets += DIV_ROUND_UP(j->prereserved.reserved,
 562                                            (ca->mi.bucket_size << 6) -
 563                                            journal_entry_overhead(j));
 564
 565                 nr_buckets = min(nr_buckets, ja->nr);
 566
 567                 bucket_to_flush = (ja->cur_idx + nr_buckets) % ja->nr;
 568                 seq_to_flush = max(seq_to_flush,
 569                                    ja->bucket_seq[bucket_to_flush]);
 570         }
 571
 572         /* Also flush if the pin fifo is more than half full */
 573         seq_to_flush = max_t(s64, seq_to_flush,
 574                              (s64) journal_cur_seq(j) -
 575                              (j->pin.size >> 1));
 576         spin_unlock(&j->lock);
 577
 578         return seq_to_flush;
 579 }
 580
 581 /**
 582  * bch2_journal_reclaim - free up journal buckets
 583  *
 584  * Background journal reclaim writes out btree nodes. It should be run
 585  * early enough so that we never completely run out of journal buckets.
 586  *
 587  * High watermarks for triggering background reclaim:
 588  * - FIFO has fewer than 512 entries left
 589  * - fewer than 25% journal buckets free
 590  *
 591  * Background reclaim runs until low watermarks are reached:
 592  * - FIFO has more than 1024 entries left
 593  * - more than 50% journal buckets free
 594  *
 595  * As long as a reclaim can complete in the time it takes to fill up
 596  * 512 journal entries or 25% of all journal buckets, then
 597  * journal_next_bucket() should not stall.
 598  */
 599 static int __bch2_journal_reclaim(struct journal *j, bool direct, bool kicked)
 600 {
 601         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 602         bool kthread = (current->flags & PF_KTHREAD) != 0;
 603         u64 seq_to_flush;
 604         size_t min_nr, min_key_cache, nr_flushed;
 605         unsigned flags;
 606         int ret = 0;
 607
 608         /*
 609          * We can't invoke memory reclaim while holding the reclaim_lock -
 610          * journal reclaim is required to make progress for memory reclaim
 611          * (cleaning the caches), so we can't get stuck in memory reclaim while
 612          * we're holding the reclaim lock:
 613          */
 614         lockdep_assert_held(&j->reclaim_lock);
 615         flags = memalloc_noreclaim_save();
 616
 617         do {
 618                 if (kthread && kthread_should_stop())
 619                         break;
 620
 621                 if (bch2_journal_error(j)) {
 622                         ret = -EIO;
 623                         break;
 624                 }
 625
 626                 bch2_journal_do_discards(j);
 627
 628                 seq_to_flush = journal_seq_to_flush(j);
 629                 min_nr = 0;
 630
 631                 /*
 632                  * If it's been longer than j->reclaim_delay_ms since we last flushed,
 633                  * make sure to flush at least one journal pin:
 634                  */
 635                 if (time_after(jiffies, j->last_flushed +
 636                                msecs_to_jiffies(c->opts.journal_reclaim_delay)))
 637                         min_nr = 1;
 638
 639                 if (j->prereserved.reserved * 4 > j->prereserved.remaining)
 640                         min_nr = 1;
 641
 642                 if (fifo_free(&j->pin) <= 32)
 643                         min_nr = 1;
 644
 645                 if (atomic_read(&c->btree_cache.dirty) * 2 > c->btree_cache.used)
 646                         min_nr = 1;
 647
 648                 min_key_cache = min(bch2_nr_btree_keys_need_flush(c), (size_t) 128);
 649
 650                 trace_and_count(c, journal_reclaim_start, c,
 651                                 direct, kicked,
 652                                 min_nr, min_key_cache,
 653                                 j->prereserved.reserved,
 654                                 j->prereserved.remaining,
 655                                 atomic_read(&c->btree_cache.dirty),
 656                                 c->btree_cache.used,
 657                                 atomic_long_read(&c->btree_key_cache.nr_dirty),
 658                                 atomic_long_read(&c->btree_key_cache.nr_keys));
 659
 660                 nr_flushed = journal_flush_pins(j, seq_to_flush,
 661                                                 ~0, 0,
 662                                                 min_nr, min_key_cache);
 663
 664                 if (direct)
 665                         j->nr_direct_reclaim += nr_flushed;
 666                 else
 667                         j->nr_background_reclaim += nr_flushed;
 668                 trace_and_count(c, journal_reclaim_finish, c, nr_flushed);
 669
 670                 if (nr_flushed)
 671                         wake_up(&j->reclaim_wait);
 672         } while ((min_nr || min_key_cache) && nr_flushed && !direct);
 673
 674         memalloc_noreclaim_restore(flags);
 675
 676         return ret;
 677 }
 678
 679 int bch2_journal_reclaim(struct journal *j)
 680 {
 681         return __bch2_journal_reclaim(j, true, true);
 682 }
 683
 684 static int bch2_journal_reclaim_thread(void *arg)
 685 {
 686         struct journal *j = arg;
 687         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 688         unsigned long delay, now;
 689         bool journal_empty;
 690         int ret = 0;
 691
 692         set_freezable();
 693
 694         j->last_flushed = jiffies;
 695
 696         while (!ret && !kthread_should_stop()) {
 697                 bool kicked = j->reclaim_kicked;
 698
 699                 j->reclaim_kicked = false;
 700
 701                 mutex_lock(&j->reclaim_lock);
 702                 ret = __bch2_journal_reclaim(j, false, kicked);
 703                 mutex_unlock(&j->reclaim_lock);
 704
 705                 now = jiffies;
 706                 delay = msecs_to_jiffies(c->opts.journal_reclaim_delay);
 707                 j->next_reclaim = j->last_flushed + delay;
 708
 709                 if (!time_in_range(j->next_reclaim, now, now + delay))
 710                         j->next_reclaim = now + delay;
 711
 712                 while (1) {
 713                         set_current_state(TASK_INTERRUPTIBLE|TASK_FREEZABLE);
 714                         if (kthread_should_stop())
 715                                 break;
 716                         if (j->reclaim_kicked)
 717                                 break;
 718
 719                         spin_lock(&j->lock);
 720                         journal_empty = fifo_empty(&j->pin);
 721                         spin_unlock(&j->lock);
 722
 723                         if (journal_empty)
 724                                 schedule();
 725                         else if (time_after(j->next_reclaim, jiffies))
 726                                 schedule_timeout(j->next_reclaim - jiffies);
 727                         else
 728                                 break;
 729                 }
 730                 __set_current_state(TASK_RUNNING);
 731         }
 732
 733         return 0;
 734 }
 735
 736 void bch2_journal_reclaim_stop(struct journal *j)
 737 {
 738         struct task_struct *p = j->reclaim_thread;
 739
 740         j->reclaim_thread = NULL;
 741
 742         if (p) {
 743                 kthread_stop(p);
 744                 put_task_struct(p);
 745         }
 746 }
 747
 748 int bch2_journal_reclaim_start(struct journal *j)
 749 {
 750         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 751         struct task_struct *p;
 752         int ret;
 753
 754         if (j->reclaim_thread)
 755                 return 0;
 756
 757         p = kthread_create(bch2_journal_reclaim_thread, j,
 758                            "bch-reclaim/%s", c->name);
 759         ret = PTR_ERR_OR_ZERO(p);
 760         if (ret) {
 761                 bch_err(c, "error creating journal reclaim thread: %s", bch2_err_str(ret));
 762                 return ret;
 763         }
 764
 765         get_task_struct(p);
 766         j->reclaim_thread = p;
 767         wake_up_process(p);
 768         return 0;
 769 }
 770
 771 static int journal_flush_done(struct journal *j, u64 seq_to_flush,
 772                               bool *did_work)
 773 {
 774         int ret;
 775
 776         ret = bch2_journal_error(j);
 777         if (ret)
 778                 return ret;
 779
 780         mutex_lock(&j->reclaim_lock);
 781
 782         if (journal_flush_pins(j, seq_to_flush,
 783                                (1U << JOURNAL_PIN_key_cache)|
 784                                (1U << JOURNAL_PIN_other), 0, 0, 0) ||
 785             journal_flush_pins(j, seq_to_flush,
 786                                (1U << JOURNAL_PIN_btree), 0, 0, 0))
 787                 *did_work = true;
 788
 789         spin_lock(&j->lock);
 790         /*
 791          * If journal replay hasn't completed, the unreplayed journal entries
 792          * hold refs on their corresponding sequence numbers
 793          */
 794         ret = !test_bit(JOURNAL_REPLAY_DONE, &j->flags) ||
 795                 journal_last_seq(j) > seq_to_flush ||
 796                 !fifo_used(&j->pin);
 797
 798         spin_unlock(&j->lock);
 799         mutex_unlock(&j->reclaim_lock);
 800
 801         return ret;
 802 }
 803
 804 bool bch2_journal_flush_pins(struct journal *j, u64 seq_to_flush)
 805 {
 806         bool did_work = false;
 807
 808         if (!test_bit(JOURNAL_STARTED, &j->flags))
 809                 return false;
 810
 811         closure_wait_event(&j->async_wait,
 812                 journal_flush_done(j, seq_to_flush, &did_work));
 813
 814         return did_work;
 815 }
 816
 817 int bch2_journal_flush_device_pins(struct journal *j, int dev_idx)
 818 {
 819         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 820         struct journal_entry_pin_list *p;
 821         u64 iter, seq = 0;
 822         int ret = 0;
 823
 824         spin_lock(&j->lock);
 825         fifo_for_each_entry_ptr(p, &j->pin, iter)
 826                 if (dev_idx >= 0
 827                     ? bch2_dev_list_has_dev(p->devs, dev_idx)
 828                     : p->devs.nr < c->opts.metadata_replicas)
 829                         seq = iter;
 830         spin_unlock(&j->lock);
 831
 832         bch2_journal_flush_pins(j, seq);
 833
 834         ret = bch2_journal_error(j);
 835         if (ret)
 836                 return ret;
 837
 838         mutex_lock(&c->replicas_gc_lock);
 839         bch2_replicas_gc_start(c, 1 << BCH_DATA_journal);
 840
 841         /*
 842          * Now that we've populated replicas_gc, write to the journal to mark
 843          * active journal devices. This handles the case where the journal might
 844          * be empty. Otherwise we could clear all journal replicas and
 845          * temporarily put the fs into an unrecoverable state. Journal recovery
 846          * expects to find devices marked for journal data on unclean mount.
 847          */
 848         ret = bch2_journal_meta(&c->journal);
 849         if (ret)
 850                 goto err;
 851
 852         seq = 0;
 853         spin_lock(&j->lock);
 854         while (!ret) {
 855                 struct bch_replicas_padded replicas;
 856
 857                 seq = max(seq, journal_last_seq(j));
 858                 if (seq >= j->pin.back)
 859                         break;
 860                 bch2_devlist_to_replicas(&replicas.e, BCH_DATA_journal,
 861                                          journal_seq_pin(j, seq)->devs);
 862                 seq++;
 863
 864                 spin_unlock(&j->lock);
 865                 ret = bch2_mark_replicas(c, &replicas.e);
 866                 spin_lock(&j->lock);
 867         }
 868         spin_unlock(&j->lock);
 869 err:
 870         ret = bch2_replicas_gc_end(c, ret);
 871         mutex_unlock(&c->replicas_gc_lock);
 872
 873         return ret;
 874 }