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 "error.h"
   6 #include "journal.h"
   7 #include "journal_io.h"
   8 #include "journal_reclaim.h"
   9 #include "replicas.h"
  10 #include "super.h"
  11
  12 #include <linux/kthread.h>
  13 #include <linux/sched/mm.h>
  14 #include <trace/events/bcachefs.h>
  15
  16 /* Free space calculations: */
  17
  18 static unsigned journal_space_from(struct journal_device *ja,
  19                                    enum journal_space_from from)
  20 {
  21         switch (from) {
  22         case journal_space_discarded:
  23                 return ja->discard_idx;
  24         case journal_space_clean_ondisk:
  25                 return ja->dirty_idx_ondisk;
  26         case journal_space_clean:
  27                 return ja->dirty_idx;
  28         default:
  29                 BUG();
  30         }
  31 }
  32
  33 unsigned bch2_journal_dev_buckets_available(struct journal *j,
  34                                             struct journal_device *ja,
  35                                             enum journal_space_from from)
  36 {
  37         unsigned available = (journal_space_from(ja, from) -
  38                               ja->cur_idx - 1 + ja->nr) % ja->nr;
  39
  40         /*
  41          * Don't use the last bucket unless writing the new last_seq
  42          * will make another bucket available:
  43          */
  44         if (available && ja->dirty_idx_ondisk == ja->dirty_idx)
  45                 --available;
  46
  47         return available;
  48 }
  49
  50 static void journal_set_remaining(struct journal *j, unsigned u64s_remaining)
  51 {
  52         union journal_preres_state old, new;
  53         u64 v = atomic64_read(&j->prereserved.counter);
  54
  55         do {
  56                 old.v = new.v = v;
  57                 new.remaining = u64s_remaining;
  58         } while ((v = atomic64_cmpxchg(&j->prereserved.counter,
  59                                        old.v, new.v)) != old.v);
  60 }
  61
  62 static struct journal_space
  63 journal_dev_space_available(struct journal *j, struct bch_dev *ca,
  64                             enum journal_space_from from)
  65 {
  66         struct journal_device *ja = &ca->journal;
  67         unsigned sectors, buckets, unwritten;
  68         u64 seq;
  69
  70         if (from == journal_space_total)
  71                 return (struct journal_space) {
  72                         .next_entry     = ca->mi.bucket_size,
  73                         .total          = ca->mi.bucket_size * ja->nr,
  74                 };
  75
  76         buckets = bch2_journal_dev_buckets_available(j, ja, from);
  77         sectors = ja->sectors_free;
  78
  79         /*
  80          * We that we don't allocate the space for a journal entry
  81          * until we write it out - thus, account for it here:
  82          */
  83         for (seq = journal_last_unwritten_seq(j);
  84              seq <= journal_cur_seq(j);
  85              seq++) {
  86                 unwritten = j->buf[seq & JOURNAL_BUF_MASK].sectors;
  87
  88                 if (!unwritten)
  89                         continue;
  90
  91                 /* entry won't fit on this device, skip: */
  92                 if (unwritten > ca->mi.bucket_size)
  93                         continue;
  94
  95                 if (unwritten >= sectors) {
  96                         if (!buckets) {
  97                                 sectors = 0;
  98                                 break;
  99                         }
 100
 101                         buckets--;
 102                         sectors = ca->mi.bucket_size;
 103                 }
 104
 105                 sectors -= unwritten;
 106         }
 107
 108         if (sectors < ca->mi.bucket_size && buckets) {
 109                 buckets--;
 110                 sectors = ca->mi.bucket_size;
 111         }
 112
 113         return (struct journal_space) {
 114                 .next_entry     = sectors,
 115                 .total          = sectors + buckets * ca->mi.bucket_size,
 116         };
 117 }
 118
 119 static struct journal_space __journal_space_available(struct journal *j, unsigned nr_devs_want,
 120                             enum journal_space_from from)
 121 {
 122         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 123         struct bch_dev *ca;
 124         unsigned i, pos, nr_devs = 0;
 125         struct journal_space space, dev_space[BCH_SB_MEMBERS_MAX];
 126
 127         BUG_ON(nr_devs_want > ARRAY_SIZE(dev_space));
 128
 129         rcu_read_lock();
 130         for_each_member_device_rcu(ca, c, i,
 131                                    &c->rw_devs[BCH_DATA_journal]) {
 132                 if (!ca->journal.nr)
 133                         continue;
 134
 135                 space = journal_dev_space_available(j, ca, from);
 136                 if (!space.next_entry)
 137                         continue;
 138
 139                 for (pos = 0; pos < nr_devs; pos++)
 140                         if (space.total > dev_space[pos].total)
 141                                 break;
 142
 143                 array_insert_item(dev_space, nr_devs, pos, space);
 144         }
 145         rcu_read_unlock();
 146
 147         if (nr_devs < nr_devs_want)
 148                 return (struct journal_space) { 0, 0 };
 149
 150         /*
 151          * We sorted largest to smallest, and we want the smallest out of the
 152          * @nr_devs_want largest devices:
 153          */
 154         return dev_space[nr_devs_want - 1];
 155 }
 156
 157 void bch2_journal_space_available(struct journal *j)
 158 {
 159         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 160         struct bch_dev *ca;
 161         unsigned clean, clean_ondisk, total;
 162         s64 u64s_remaining = 0;
 163         unsigned max_entry_size  = min(j->buf[0].buf_size >> 9,
 164                                        j->buf[1].buf_size >> 9);
 165         unsigned i, nr_online = 0, nr_devs_want;
 166         bool can_discard = false;
 167         int ret = 0;
 168
 169         lockdep_assert_held(&j->lock);
 170
 171         rcu_read_lock();
 172         for_each_member_device_rcu(ca, c, i,
 173                                    &c->rw_devs[BCH_DATA_journal]) {
 174                 struct journal_device *ja = &ca->journal;
 175
 176                 if (!ja->nr)
 177                         continue;
 178
 179                 while (ja->dirty_idx != ja->cur_idx &&
 180                        ja->bucket_seq[ja->dirty_idx] < journal_last_seq(j))
 181                         ja->dirty_idx = (ja->dirty_idx + 1) % ja->nr;
 182
 183                 while (ja->dirty_idx_ondisk != ja->dirty_idx &&
 184                        ja->bucket_seq[ja->dirty_idx_ondisk] < j->last_seq_ondisk)
 185                         ja->dirty_idx_ondisk = (ja->dirty_idx_ondisk + 1) % ja->nr;
 186
 187                 if (ja->discard_idx != ja->dirty_idx_ondisk)
 188                         can_discard = true;
 189
 190                 max_entry_size = min_t(unsigned, max_entry_size, ca->mi.bucket_size);
 191                 nr_online++;
 192         }
 193         rcu_read_unlock();
 194
 195         j->can_discard = can_discard;
 196
 197         if (nr_online < c->opts.metadata_replicas_required) {
 198                 ret = JOURNAL_ERR_insufficient_devices;
 199                 goto out;
 200         }
 201
 202         nr_devs_want = min_t(unsigned, nr_online, c->opts.metadata_replicas);
 203
 204         for (i = 0; i < journal_space_nr; i++)
 205                 j->space[i] = __journal_space_available(j, nr_devs_want, i);
 206
 207         clean_ondisk    = j->space[journal_space_clean_ondisk].total;
 208         clean           = j->space[journal_space_clean].total;
 209         total           = j->space[journal_space_total].total;
 210
 211         if (!clean_ondisk &&
 212             journal_cur_seq(j) == j->seq_ondisk) {
 213                 struct printbuf buf = PRINTBUF;
 214
 215                 __bch2_journal_debug_to_text(&buf, j);
 216                 bch_err(c, "journal stuck\n%s", buf.buf);
 217                 printbuf_exit(&buf);
 218
 219                 /*
 220                  * Hack: bch2_fatal_error() calls bch2_journal_halt() which
 221                  * takes journal lock:
 222                  */
 223                 spin_unlock(&j->lock);
 224                 bch2_fatal_error(c);
 225                 spin_lock(&j->lock);
 226
 227                 ret = JOURNAL_ERR_journal_stuck;
 228         } else if (!j->space[journal_space_discarded].next_entry)
 229                 ret = JOURNAL_ERR_journal_full;
 230
 231         if ((j->space[journal_space_clean_ondisk].next_entry <
 232              j->space[journal_space_clean_ondisk].total) &&
 233             (clean - clean_ondisk <= total / 8) &&
 234             (clean_ondisk * 2 > clean ))
 235                 set_bit(JOURNAL_MAY_SKIP_FLUSH, &j->flags);
 236         else
 237                 clear_bit(JOURNAL_MAY_SKIP_FLUSH, &j->flags);
 238
 239         u64s_remaining  = (u64) clean << 6;
 240         u64s_remaining -= (u64) total << 3;
 241         u64s_remaining = max(0LL, u64s_remaining);
 242         u64s_remaining /= 4;
 243         u64s_remaining = min_t(u64, u64s_remaining, U32_MAX);
 244 out:
 245         j->cur_entry_sectors    = !ret ? j->space[journal_space_discarded].next_entry : 0;
 246         j->cur_entry_error      = ret;
 247         journal_set_remaining(j, u64s_remaining);
 248         journal_set_watermark(j);
 249
 250         if (!ret)
 251                 journal_wake(j);
 252 }
 253
 254 /* Discards - last part of journal reclaim: */
 255
 256 static bool should_discard_bucket(struct journal *j, struct journal_device *ja)
 257 {
 258         bool ret;
 259
 260         spin_lock(&j->lock);
 261         ret = ja->discard_idx != ja->dirty_idx_ondisk;
 262         spin_unlock(&j->lock);
 263
 264         return ret;
 265 }
 266
 267 /*
 268  * Advance ja->discard_idx as long as it points to buckets that are no longer
 269  * dirty, issuing discards if necessary:
 270  */
 271 void bch2_journal_do_discards(struct journal *j)
 272 {
 273         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 274         struct bch_dev *ca;
 275         unsigned iter;
 276
 277         mutex_lock(&j->discard_lock);
 278
 279         for_each_rw_member(ca, c, iter) {
 280                 struct journal_device *ja = &ca->journal;
 281
 282                 while (should_discard_bucket(j, ja)) {
 283                         if (!c->opts.nochanges &&
 284                             ca->mi.discard &&
 285                             blk_queue_discard(bdev_get_queue(ca->disk_sb.bdev)))
 286                                 blkdev_issue_discard(ca->disk_sb.bdev,
 287                                         bucket_to_sector(ca,
 288                                                 ja->buckets[ja->discard_idx]),
 289                                         ca->mi.bucket_size, GFP_NOIO, 0);
 290
 291                         spin_lock(&j->lock);
 292                         ja->discard_idx = (ja->discard_idx + 1) % ja->nr;
 293
 294                         bch2_journal_space_available(j);
 295                         spin_unlock(&j->lock);
 296                 }
 297         }
 298
 299         mutex_unlock(&j->discard_lock);
 300 }
 301
 302 /*
 303  * Journal entry pinning - machinery for holding a reference on a given journal
 304  * entry, holding it open to ensure it gets replayed during recovery:
 305  */
 306
 307 static void bch2_journal_reclaim_fast(struct journal *j)
 308 {
 309         struct journal_entry_pin_list temp;
 310         bool popped = false;
 311
 312         lockdep_assert_held(&j->lock);
 313
 314         /*
 315          * Unpin journal entries whose reference counts reached zero, meaning
 316          * all btree nodes got written out
 317          */
 318         while (!fifo_empty(&j->pin) &&
 319                !atomic_read(&fifo_peek_front(&j->pin).count)) {
 320                 BUG_ON(!list_empty(&fifo_peek_front(&j->pin).list));
 321                 BUG_ON(!list_empty(&fifo_peek_front(&j->pin).flushed));
 322                 BUG_ON(!fifo_pop(&j->pin, temp));
 323                 popped = true;
 324         }
 325
 326         if (popped)
 327                 bch2_journal_space_available(j);
 328 }
 329
 330 void __bch2_journal_pin_put(struct journal *j, u64 seq)
 331 {
 332         struct journal_entry_pin_list *pin_list = journal_seq_pin(j, seq);
 333
 334         if (atomic_dec_and_test(&pin_list->count))
 335                 bch2_journal_reclaim_fast(j);
 336 }
 337
 338 void bch2_journal_pin_put(struct journal *j, u64 seq)
 339 {
 340         struct journal_entry_pin_list *pin_list = journal_seq_pin(j, seq);
 341
 342         if (atomic_dec_and_test(&pin_list->count)) {
 343                 spin_lock(&j->lock);
 344                 bch2_journal_reclaim_fast(j);
 345                 spin_unlock(&j->lock);
 346         }
 347 }
 348
 349 static inline void __journal_pin_drop(struct journal *j,
 350                                       struct journal_entry_pin *pin)
 351 {
 352         struct journal_entry_pin_list *pin_list;
 353
 354         if (!journal_pin_active(pin))
 355                 return;
 356
 357         if (j->flush_in_progress == pin)
 358                 j->flush_in_progress_dropped = true;
 359
 360         pin_list = journal_seq_pin(j, pin->seq);
 361         pin->seq = 0;
 362         list_del_init(&pin->list);
 363
 364         /*
 365          * Unpinning a journal entry make make journal_next_bucket() succeed, if
 366          * writing a new last_seq will now make another bucket available:
 367          */
 368         if (atomic_dec_and_test(&pin_list->count) &&
 369             pin_list == &fifo_peek_front(&j->pin))
 370                 bch2_journal_reclaim_fast(j);
 371 }
 372
 373 void bch2_journal_pin_drop(struct journal *j,
 374                            struct journal_entry_pin *pin)
 375 {
 376         spin_lock(&j->lock);
 377         __journal_pin_drop(j, pin);
 378         spin_unlock(&j->lock);
 379 }
 380
 381 void bch2_journal_pin_set(struct journal *j, u64 seq,
 382                           struct journal_entry_pin *pin,
 383                           journal_pin_flush_fn flush_fn)
 384 {
 385         struct journal_entry_pin_list *pin_list;
 386
 387         spin_lock(&j->lock);
 388
 389         if (seq < journal_last_seq(j)) {
 390                 /*
 391                  * bch2_journal_pin_copy() raced with bch2_journal_pin_drop() on
 392                  * the src pin - with the pin dropped, the entry to pin might no
 393                  * longer to exist, but that means there's no longer anything to
 394                  * copy and we can bail out here:
 395                  */
 396                 spin_unlock(&j->lock);
 397                 return;
 398         }
 399
 400         pin_list = journal_seq_pin(j, seq);
 401
 402         __journal_pin_drop(j, pin);
 403
 404         atomic_inc(&pin_list->count);
 405         pin->seq        = seq;
 406         pin->flush      = flush_fn;
 407
 408         if (flush_fn == bch2_btree_key_cache_journal_flush)
 409                 list_add(&pin->list, &pin_list->key_cache_list);
 410         else if (flush_fn)
 411                 list_add(&pin->list, &pin_list->list);
 412         else
 413                 list_add(&pin->list, &pin_list->flushed);
 414         spin_unlock(&j->lock);
 415
 416         /*
 417          * If the journal is currently full,  we might want to call flush_fn
 418          * immediately:
 419          */
 420         journal_wake(j);
 421 }
 422
 423 /**
 424  * bch2_journal_pin_flush: ensure journal pin callback is no longer running
 425  */
 426 void bch2_journal_pin_flush(struct journal *j, struct journal_entry_pin *pin)
 427 {
 428         BUG_ON(journal_pin_active(pin));
 429
 430         wait_event(j->pin_flush_wait, j->flush_in_progress != pin);
 431 }
 432
 433 /*
 434  * Journal reclaim: flush references to open journal entries to reclaim space in
 435  * the journal
 436  *
 437  * May be done by the journal code in the background as needed to free up space
 438  * for more journal entries, or as part of doing a clean shutdown, or to migrate
 439  * data off of a specific device:
 440  */
 441
 442 static struct journal_entry_pin *
 443 journal_get_next_pin(struct journal *j,
 444                      bool get_any,
 445                      bool get_key_cache,
 446                      u64 max_seq, u64 *seq)
 447 {
 448         struct journal_entry_pin_list *pin_list;
 449         struct journal_entry_pin *ret = NULL;
 450
 451         fifo_for_each_entry_ptr(pin_list, &j->pin, *seq) {
 452                 if (*seq > max_seq && !get_any && !get_key_cache)
 453                         break;
 454
 455                 if (*seq <= max_seq || get_any) {
 456                         ret = list_first_entry_or_null(&pin_list->list,
 457                                 struct journal_entry_pin, list);
 458                         if (ret)
 459                                 return ret;
 460                 }
 461
 462                 if (*seq <= max_seq || get_any || get_key_cache) {
 463                         ret = list_first_entry_or_null(&pin_list->key_cache_list,
 464                                 struct journal_entry_pin, list);
 465                         if (ret)
 466                                 return ret;
 467                 }
 468         }
 469
 470         return NULL;
 471 }
 472
 473 /* returns true if we did work */
 474 static size_t journal_flush_pins(struct journal *j, u64 seq_to_flush,
 475                                  unsigned min_any,
 476                                  unsigned min_key_cache)
 477 {
 478         struct journal_entry_pin *pin;
 479         size_t nr_flushed = 0;
 480         journal_pin_flush_fn flush_fn;
 481         u64 seq;
 482         int err;
 483
 484         lockdep_assert_held(&j->reclaim_lock);
 485
 486         while (1) {
 487                 cond_resched();
 488
 489                 j->last_flushed = jiffies;
 490
 491                 spin_lock(&j->lock);
 492                 pin = journal_get_next_pin(j,
 493                                            min_any != 0,
 494                                            min_key_cache != 0,
 495                                            seq_to_flush, &seq);
 496                 if (pin) {
 497                         BUG_ON(j->flush_in_progress);
 498                         j->flush_in_progress = pin;
 499                         j->flush_in_progress_dropped = false;
 500                         flush_fn = pin->flush;
 501                 }
 502                 spin_unlock(&j->lock);
 503
 504                 if (!pin)
 505                         break;
 506
 507                 if (min_key_cache && pin->flush == bch2_btree_key_cache_journal_flush)
 508                         min_key_cache--;
 509
 510                 if (min_any)
 511                         min_any--;
 512
 513                 err = flush_fn(j, pin, seq);
 514
 515                 spin_lock(&j->lock);
 516                 /* Pin might have been dropped or rearmed: */
 517                 if (likely(!err && !j->flush_in_progress_dropped))
 518                         list_move(&pin->list, &journal_seq_pin(j, seq)->flushed);
 519                 j->flush_in_progress = NULL;
 520                 j->flush_in_progress_dropped = false;
 521                 spin_unlock(&j->lock);
 522
 523                 wake_up(&j->pin_flush_wait);
 524
 525                 if (err)
 526                         break;
 527
 528                 nr_flushed++;
 529         }
 530
 531         return nr_flushed;
 532 }
 533
 534 static u64 journal_seq_to_flush(struct journal *j)
 535 {
 536         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 537         struct bch_dev *ca;
 538         u64 seq_to_flush = 0;
 539         unsigned iter;
 540
 541         spin_lock(&j->lock);
 542
 543         for_each_rw_member(ca, c, iter) {
 544                 struct journal_device *ja = &ca->journal;
 545                 unsigned nr_buckets, bucket_to_flush;
 546
 547                 if (!ja->nr)
 548                         continue;
 549
 550                 /* Try to keep the journal at most half full: */
 551                 nr_buckets = ja->nr / 2;
 552
 553                 /* And include pre-reservations: */
 554                 nr_buckets += DIV_ROUND_UP(j->prereserved.reserved,
 555                                            (ca->mi.bucket_size << 6) -
 556                                            journal_entry_overhead(j));
 557
 558                 nr_buckets = min(nr_buckets, ja->nr);
 559
 560                 bucket_to_flush = (ja->cur_idx + nr_buckets) % ja->nr;
 561                 seq_to_flush = max(seq_to_flush,
 562                                    ja->bucket_seq[bucket_to_flush]);
 563         }
 564
 565         /* Also flush if the pin fifo is more than half full */
 566         seq_to_flush = max_t(s64, seq_to_flush,
 567                              (s64) journal_cur_seq(j) -
 568                              (j->pin.size >> 1));
 569         spin_unlock(&j->lock);
 570
 571         return seq_to_flush;
 572 }
 573
 574 /**
 575  * bch2_journal_reclaim - free up journal buckets
 576  *
 577  * Background journal reclaim writes out btree nodes. It should be run
 578  * early enough so that we never completely run out of journal buckets.
 579  *
 580  * High watermarks for triggering background reclaim:
 581  * - FIFO has fewer than 512 entries left
 582  * - fewer than 25% journal buckets free
 583  *
 584  * Background reclaim runs until low watermarks are reached:
 585  * - FIFO has more than 1024 entries left
 586  * - more than 50% journal buckets free
 587  *
 588  * As long as a reclaim can complete in the time it takes to fill up
 589  * 512 journal entries or 25% of all journal buckets, then
 590  * journal_next_bucket() should not stall.
 591  */
 592 static int __bch2_journal_reclaim(struct journal *j, bool direct, bool kicked)
 593 {
 594         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 595         bool kthread = (current->flags & PF_KTHREAD) != 0;
 596         u64 seq_to_flush;
 597         size_t min_nr, min_key_cache, nr_flushed;
 598         unsigned flags;
 599         int ret = 0;
 600
 601         /*
 602          * We can't invoke memory reclaim while holding the reclaim_lock -
 603          * journal reclaim is required to make progress for memory reclaim
 604          * (cleaning the caches), so we can't get stuck in memory reclaim while
 605          * we're holding the reclaim lock:
 606          */
 607         lockdep_assert_held(&j->reclaim_lock);
 608         flags = memalloc_noreclaim_save();
 609
 610         do {
 611                 if (kthread && kthread_should_stop())
 612                         break;
 613
 614                 if (bch2_journal_error(j)) {
 615                         ret = -EIO;
 616                         break;
 617                 }
 618
 619                 bch2_journal_do_discards(j);
 620
 621                 seq_to_flush = journal_seq_to_flush(j);
 622                 min_nr = 0;
 623
 624                 /*
 625                  * If it's been longer than j->reclaim_delay_ms since we last flushed,
 626                  * make sure to flush at least one journal pin:
 627                  */
 628                 if (time_after(jiffies, j->last_flushed +
 629                                msecs_to_jiffies(c->opts.journal_reclaim_delay)))
 630                         min_nr = 1;
 631
 632                 if (j->prereserved.reserved * 4 > j->prereserved.remaining)
 633                         min_nr = 1;
 634
 635                 if (fifo_free(&j->pin) <= 32)
 636                         min_nr = 1;
 637
 638                 if (atomic_read(&c->btree_cache.dirty) * 2 > c->btree_cache.used)
 639                         min_nr = 1;
 640
 641                 min_key_cache = min(bch2_nr_btree_keys_need_flush(c), (size_t) 128);
 642
 643                 trace_journal_reclaim_start(c, direct, kicked,
 644                                 min_nr, min_key_cache,
 645                                 j->prereserved.reserved,
 646                                 j->prereserved.remaining,
 647                                 atomic_read(&c->btree_cache.dirty),
 648                                 c->btree_cache.used,
 649                                 atomic_long_read(&c->btree_key_cache.nr_dirty),
 650                                 atomic_long_read(&c->btree_key_cache.nr_keys));
 651
 652                 nr_flushed = journal_flush_pins(j, seq_to_flush,
 653                                                 min_nr, min_key_cache);
 654
 655                 if (direct)
 656                         j->nr_direct_reclaim += nr_flushed;
 657                 else
 658                         j->nr_background_reclaim += nr_flushed;
 659                 trace_journal_reclaim_finish(c, nr_flushed);
 660
 661                 if (nr_flushed)
 662                         wake_up(&j->reclaim_wait);
 663         } while ((min_nr || min_key_cache) && nr_flushed && !direct);
 664
 665         memalloc_noreclaim_restore(flags);
 666
 667         return ret;
 668 }
 669
 670 int bch2_journal_reclaim(struct journal *j)
 671 {
 672         return __bch2_journal_reclaim(j, true, true);
 673 }
 674
 675 static int bch2_journal_reclaim_thread(void *arg)
 676 {
 677         struct journal *j = arg;
 678         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 679         unsigned long delay, now;
 680         bool journal_empty;
 681         int ret = 0;
 682
 683         set_freezable();
 684
 685         j->last_flushed = jiffies;
 686
 687         while (!ret && !kthread_should_stop()) {
 688                 bool kicked = j->reclaim_kicked;
 689
 690                 j->reclaim_kicked = false;
 691
 692                 mutex_lock(&j->reclaim_lock);
 693                 ret = __bch2_journal_reclaim(j, false, kicked);
 694                 mutex_unlock(&j->reclaim_lock);
 695
 696                 now = jiffies;
 697                 delay = msecs_to_jiffies(c->opts.journal_reclaim_delay);
 698                 j->next_reclaim = j->last_flushed + delay;
 699
 700                 if (!time_in_range(j->next_reclaim, now, now + delay))
 701                         j->next_reclaim = now + delay;
 702
 703                 while (1) {
 704                         set_current_state(TASK_INTERRUPTIBLE);
 705                         if (kthread_should_stop())
 706                                 break;
 707                         if (j->reclaim_kicked)
 708                                 break;
 709
 710                         spin_lock(&j->lock);
 711                         journal_empty = fifo_empty(&j->pin);
 712                         spin_unlock(&j->lock);
 713
 714                         if (journal_empty)
 715                                 freezable_schedule();
 716                         else if (time_after(j->next_reclaim, jiffies))
 717                                 freezable_schedule_timeout(j->next_reclaim - jiffies);
 718                         else
 719                                 break;
 720                 }
 721                 __set_current_state(TASK_RUNNING);
 722         }
 723
 724         return 0;
 725 }
 726
 727 void bch2_journal_reclaim_stop(struct journal *j)
 728 {
 729         struct task_struct *p = j->reclaim_thread;
 730
 731         j->reclaim_thread = NULL;
 732
 733         if (p) {
 734                 kthread_stop(p);
 735                 put_task_struct(p);
 736         }
 737 }
 738
 739 int bch2_journal_reclaim_start(struct journal *j)
 740 {
 741         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 742         struct task_struct *p;
 743
 744         if (j->reclaim_thread)
 745                 return 0;
 746
 747         p = kthread_create(bch2_journal_reclaim_thread, j,
 748                            "bch-reclaim/%s", c->name);
 749         if (IS_ERR(p)) {
 750                 bch_err(c, "error creating journal reclaim thread: %li", PTR_ERR(p));
 751                 return PTR_ERR(p);
 752         }
 753
 754         get_task_struct(p);
 755         j->reclaim_thread = p;
 756         wake_up_process(p);
 757         return 0;
 758 }
 759
 760 static int journal_flush_done(struct journal *j, u64 seq_to_flush,
 761                               bool *did_work)
 762 {
 763         int ret;
 764
 765         ret = bch2_journal_error(j);
 766         if (ret)
 767                 return ret;
 768
 769         mutex_lock(&j->reclaim_lock);
 770
 771         if (journal_flush_pins(j, seq_to_flush, 0, 0))
 772                 *did_work = true;
 773
 774         spin_lock(&j->lock);
 775         /*
 776          * If journal replay hasn't completed, the unreplayed journal entries
 777          * hold refs on their corresponding sequence numbers
 778          */
 779         ret = !test_bit(JOURNAL_REPLAY_DONE, &j->flags) ||
 780                 journal_last_seq(j) > seq_to_flush ||
 781                 !fifo_used(&j->pin);
 782
 783         spin_unlock(&j->lock);
 784         mutex_unlock(&j->reclaim_lock);
 785
 786         return ret;
 787 }
 788
 789 bool bch2_journal_flush_pins(struct journal *j, u64 seq_to_flush)
 790 {
 791         bool did_work = false;
 792
 793         if (!test_bit(JOURNAL_STARTED, &j->flags))
 794                 return false;
 795
 796         closure_wait_event(&j->async_wait,
 797                 journal_flush_done(j, seq_to_flush, &did_work));
 798
 799         return did_work;
 800 }
 801
 802 int bch2_journal_flush_device_pins(struct journal *j, int dev_idx)
 803 {
 804         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 805         struct journal_entry_pin_list *p;
 806         u64 iter, seq = 0;
 807         int ret = 0;
 808
 809         spin_lock(&j->lock);
 810         fifo_for_each_entry_ptr(p, &j->pin, iter)
 811                 if (dev_idx >= 0
 812                     ? bch2_dev_list_has_dev(p->devs, dev_idx)
 813                     : p->devs.nr < c->opts.metadata_replicas)
 814                         seq = iter;
 815         spin_unlock(&j->lock);
 816
 817         bch2_journal_flush_pins(j, seq);
 818
 819         ret = bch2_journal_error(j);
 820         if (ret)
 821                 return ret;
 822
 823         mutex_lock(&c->replicas_gc_lock);
 824         bch2_replicas_gc_start(c, 1 << BCH_DATA_journal);
 825
 826         seq = 0;
 827
 828         spin_lock(&j->lock);
 829         while (!ret) {
 830                 struct bch_replicas_padded replicas;
 831
 832                 seq = max(seq, journal_last_seq(j));
 833                 if (seq >= j->pin.back)
 834                         break;
 835                 bch2_devlist_to_replicas(&replicas.e, BCH_DATA_journal,
 836                                          journal_seq_pin(j, seq)->devs);
 837                 seq++;
 838
 839                 spin_unlock(&j->lock);
 840                 ret = bch2_mark_replicas(c, &replicas.e);
 841                 spin_lock(&j->lock);
 842         }
 843         spin_unlock(&j->lock);
 844
 845         ret = bch2_replicas_gc_end(c, ret);
 846         mutex_unlock(&c->replicas_gc_lock);
 847
 848         return ret;
 849 }