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 inline void journal_set_watermark(struct journal *j)
  54 {
  55         struct bch_fs *c = container_of(j, struct bch_fs, journal);
  56         bool low_on_space = j->space[journal_space_clean].total * 4 <=
  57                 j->space[journal_space_total].total;
  58         bool low_on_pin = fifo_free(&j->pin) < j->pin.size / 4;
  59         unsigned watermark = low_on_space || low_on_pin
  60                 ? BCH_WATERMARK_reclaim
  61                 : BCH_WATERMARK_stripe;
  62
  63         if (track_event_change(&c->times[BCH_TIME_blocked_journal_low_on_space],
  64                                &j->low_on_space_start, low_on_space) ||
  65             track_event_change(&c->times[BCH_TIME_blocked_journal_low_on_pin],
  66                                &j->low_on_pin_start, low_on_pin))
  67                 trace_and_count(c, journal_full, c);
  68
  69         swap(watermark, j->watermark);
  70         if (watermark > j->watermark)
  71                 journal_wake(j);
  72 }
  73
  74 static struct journal_space
  75 journal_dev_space_available(struct journal *j, struct bch_dev *ca,
  76                             enum journal_space_from from)
  77 {
  78         struct journal_device *ja = &ca->journal;
  79         unsigned sectors, buckets, unwritten;
  80         u64 seq;
  81
  82         if (from == journal_space_total)
  83                 return (struct journal_space) {
  84                         .next_entry     = ca->mi.bucket_size,
  85                         .total          = ca->mi.bucket_size * ja->nr,
  86                 };
  87
  88         buckets = bch2_journal_dev_buckets_available(j, ja, from);
  89         sectors = ja->sectors_free;
  90
  91         /*
  92          * We that we don't allocate the space for a journal entry
  93          * until we write it out - thus, account for it here:
  94          */
  95         for (seq = journal_last_unwritten_seq(j);
  96              seq <= journal_cur_seq(j);
  97              seq++) {
  98                 unwritten = j->buf[seq & JOURNAL_BUF_MASK].sectors;
  99
 100                 if (!unwritten)
 101                         continue;
 102
 103                 /* entry won't fit on this device, skip: */
 104                 if (unwritten > ca->mi.bucket_size)
 105                         continue;
 106
 107                 if (unwritten >= sectors) {
 108                         if (!buckets) {
 109                                 sectors = 0;
 110                                 break;
 111                         }
 112
 113                         buckets--;
 114                         sectors = ca->mi.bucket_size;
 115                 }
 116
 117                 sectors -= unwritten;
 118         }
 119
 120         if (sectors < ca->mi.bucket_size && buckets) {
 121                 buckets--;
 122                 sectors = ca->mi.bucket_size;
 123         }
 124
 125         return (struct journal_space) {
 126                 .next_entry     = sectors,
 127                 .total          = sectors + buckets * ca->mi.bucket_size,
 128         };
 129 }
 130
 131 static struct journal_space __journal_space_available(struct journal *j, unsigned nr_devs_want,
 132                             enum journal_space_from from)
 133 {
 134         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 135         struct bch_dev *ca;
 136         unsigned i, pos, nr_devs = 0;
 137         struct journal_space space, dev_space[BCH_SB_MEMBERS_MAX];
 138
 139         BUG_ON(nr_devs_want > ARRAY_SIZE(dev_space));
 140
 141         rcu_read_lock();
 142         for_each_member_device_rcu(ca, c, i,
 143                                    &c->rw_devs[BCH_DATA_journal]) {
 144                 if (!ca->journal.nr)
 145                         continue;
 146
 147                 space = journal_dev_space_available(j, ca, from);
 148                 if (!space.next_entry)
 149                         continue;
 150
 151                 for (pos = 0; pos < nr_devs; pos++)
 152                         if (space.total > dev_space[pos].total)
 153                                 break;
 154
 155                 array_insert_item(dev_space, nr_devs, pos, space);
 156         }
 157         rcu_read_unlock();
 158
 159         if (nr_devs < nr_devs_want)
 160                 return (struct journal_space) { 0, 0 };
 161
 162         /*
 163          * We sorted largest to smallest, and we want the smallest out of the
 164          * @nr_devs_want largest devices:
 165          */
 166         return dev_space[nr_devs_want - 1];
 167 }
 168
 169 void bch2_journal_space_available(struct journal *j)
 170 {
 171         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 172         struct bch_dev *ca;
 173         unsigned clean, clean_ondisk, total;
 174         unsigned max_entry_size  = min(j->buf[0].buf_size >> 9,
 175                                        j->buf[1].buf_size >> 9);
 176         unsigned i, nr_online = 0, nr_devs_want;
 177         bool can_discard = false;
 178         int ret = 0;
 179
 180         lockdep_assert_held(&j->lock);
 181
 182         rcu_read_lock();
 183         for_each_member_device_rcu(ca, c, i,
 184                                    &c->rw_devs[BCH_DATA_journal]) {
 185                 struct journal_device *ja = &ca->journal;
 186
 187                 if (!ja->nr)
 188                         continue;
 189
 190                 while (ja->dirty_idx != ja->cur_idx &&
 191                        ja->bucket_seq[ja->dirty_idx] < journal_last_seq(j))
 192                         ja->dirty_idx = (ja->dirty_idx + 1) % ja->nr;
 193
 194                 while (ja->dirty_idx_ondisk != ja->dirty_idx &&
 195                        ja->bucket_seq[ja->dirty_idx_ondisk] < j->last_seq_ondisk)
 196                         ja->dirty_idx_ondisk = (ja->dirty_idx_ondisk + 1) % ja->nr;
 197
 198                 if (ja->discard_idx != ja->dirty_idx_ondisk)
 199                         can_discard = true;
 200
 201                 max_entry_size = min_t(unsigned, max_entry_size, ca->mi.bucket_size);
 202                 nr_online++;
 203         }
 204         rcu_read_unlock();
 205
 206         j->can_discard = can_discard;
 207
 208         if (nr_online < c->opts.metadata_replicas_required) {
 209                 ret = JOURNAL_ERR_insufficient_devices;
 210                 goto out;
 211         }
 212
 213         nr_devs_want = min_t(unsigned, nr_online, c->opts.metadata_replicas);
 214
 215         for (i = 0; i < journal_space_nr; i++)
 216                 j->space[i] = __journal_space_available(j, nr_devs_want, i);
 217
 218         clean_ondisk    = j->space[journal_space_clean_ondisk].total;
 219         clean           = j->space[journal_space_clean].total;
 220         total           = j->space[journal_space_total].total;
 221
 222         if (!j->space[journal_space_discarded].next_entry)
 223                 ret = JOURNAL_ERR_journal_full;
 224
 225         if ((j->space[journal_space_clean_ondisk].next_entry <
 226              j->space[journal_space_clean_ondisk].total) &&
 227             (clean - clean_ondisk <= total / 8) &&
 228             (clean_ondisk * 2 > clean))
 229                 set_bit(JOURNAL_MAY_SKIP_FLUSH, &j->flags);
 230         else
 231                 clear_bit(JOURNAL_MAY_SKIP_FLUSH, &j->flags);
 232
 233         journal_set_watermark(j);
 234 out:
 235         j->cur_entry_sectors    = !ret ? j->space[journal_space_discarded].next_entry : 0;
 236         j->cur_entry_error      = ret;
 237
 238         if (!ret)
 239                 journal_wake(j);
 240 }
 241
 242 /* Discards - last part of journal reclaim: */
 243
 244 static bool should_discard_bucket(struct journal *j, struct journal_device *ja)
 245 {
 246         bool ret;
 247
 248         spin_lock(&j->lock);
 249         ret = ja->discard_idx != ja->dirty_idx_ondisk;
 250         spin_unlock(&j->lock);
 251
 252         return ret;
 253 }
 254
 255 /*
 256  * Advance ja->discard_idx as long as it points to buckets that are no longer
 257  * dirty, issuing discards if necessary:
 258  */
 259 void bch2_journal_do_discards(struct journal *j)
 260 {
 261         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 262         struct bch_dev *ca;
 263         unsigned iter;
 264
 265         mutex_lock(&j->discard_lock);
 266
 267         for_each_rw_member(ca, c, iter) {
 268                 struct journal_device *ja = &ca->journal;
 269
 270                 while (should_discard_bucket(j, ja)) {
 271                         if (!c->opts.nochanges &&
 272                             ca->mi.discard &&
 273                             bdev_max_discard_sectors(ca->disk_sb.bdev))
 274                                 blkdev_issue_discard(ca->disk_sb.bdev,
 275                                         bucket_to_sector(ca,
 276                                                 ja->buckets[ja->discard_idx]),
 277                                         ca->mi.bucket_size, GFP_NOFS);
 278
 279                         spin_lock(&j->lock);
 280                         ja->discard_idx = (ja->discard_idx + 1) % ja->nr;
 281
 282                         bch2_journal_space_available(j);
 283                         spin_unlock(&j->lock);
 284                 }
 285         }
 286
 287         mutex_unlock(&j->discard_lock);
 288 }
 289
 290 /*
 291  * Journal entry pinning - machinery for holding a reference on a given journal
 292  * entry, holding it open to ensure it gets replayed during recovery:
 293  */
 294
 295 void bch2_journal_reclaim_fast(struct journal *j)
 296 {
 297         bool popped = false;
 298
 299         lockdep_assert_held(&j->lock);
 300
 301         /*
 302          * Unpin journal entries whose reference counts reached zero, meaning
 303          * all btree nodes got written out
 304          */
 305         while (!fifo_empty(&j->pin) &&
 306                !atomic_read(&fifo_peek_front(&j->pin).count)) {
 307                 j->pin.front++;
 308                 popped = true;
 309         }
 310
 311         if (popped)
 312                 bch2_journal_space_available(j);
 313 }
 314
 315 bool __bch2_journal_pin_put(struct journal *j, u64 seq)
 316 {
 317         struct journal_entry_pin_list *pin_list = journal_seq_pin(j, seq);
 318
 319         return atomic_dec_and_test(&pin_list->count);
 320 }
 321
 322 void bch2_journal_pin_put(struct journal *j, u64 seq)
 323 {
 324         if (__bch2_journal_pin_put(j, seq)) {
 325                 spin_lock(&j->lock);
 326                 bch2_journal_reclaim_fast(j);
 327                 spin_unlock(&j->lock);
 328         }
 329 }
 330
 331 static inline bool __journal_pin_drop(struct journal *j,
 332                                       struct journal_entry_pin *pin)
 333 {
 334         struct journal_entry_pin_list *pin_list;
 335
 336         if (!journal_pin_active(pin))
 337                 return false;
 338
 339         if (j->flush_in_progress == pin)
 340                 j->flush_in_progress_dropped = true;
 341
 342         pin_list = journal_seq_pin(j, pin->seq);
 343         pin->seq = 0;
 344         list_del_init(&pin->list);
 345
 346         /*
 347          * Unpinning a journal entry may make journal_next_bucket() succeed, if
 348          * writing a new last_seq will now make another bucket available:
 349          */
 350         return atomic_dec_and_test(&pin_list->count) &&
 351                 pin_list == &fifo_peek_front(&j->pin);
 352 }
 353
 354 void bch2_journal_pin_drop(struct journal *j,
 355                            struct journal_entry_pin *pin)
 356 {
 357         spin_lock(&j->lock);
 358         if (__journal_pin_drop(j, pin))
 359                 bch2_journal_reclaim_fast(j);
 360         spin_unlock(&j->lock);
 361 }
 362
 363 static enum journal_pin_type journal_pin_type(journal_pin_flush_fn fn)
 364 {
 365         if (fn == bch2_btree_node_flush0 ||
 366             fn == bch2_btree_node_flush1)
 367                 return JOURNAL_PIN_btree;
 368         else if (fn == bch2_btree_key_cache_journal_flush)
 369                 return JOURNAL_PIN_key_cache;
 370         else
 371                 return JOURNAL_PIN_other;
 372 }
 373
 374 static inline void bch2_journal_pin_set_locked(struct journal *j, u64 seq,
 375                           struct journal_entry_pin *pin,
 376                           journal_pin_flush_fn flush_fn,
 377                           enum journal_pin_type type)
 378 {
 379         struct journal_entry_pin_list *pin_list = journal_seq_pin(j, seq);
 380
 381         /*
 382          * flush_fn is how we identify journal pins in debugfs, so must always
 383          * exist, even if it doesn't do anything:
 384          */
 385         BUG_ON(!flush_fn);
 386
 387         atomic_inc(&pin_list->count);
 388         pin->seq        = seq;
 389         pin->flush      = flush_fn;
 390         list_add(&pin->list, &pin_list->list[type]);
 391 }
 392
 393 void bch2_journal_pin_copy(struct journal *j,
 394                            struct journal_entry_pin *dst,
 395                            struct journal_entry_pin *src,
 396                            journal_pin_flush_fn flush_fn)
 397 {
 398         bool reclaim;
 399
 400         spin_lock(&j->lock);
 401
 402         u64 seq = READ_ONCE(src->seq);
 403
 404         if (seq < journal_last_seq(j)) {
 405                 /*
 406                  * bch2_journal_pin_copy() raced with bch2_journal_pin_drop() on
 407                  * the src pin - with the pin dropped, the entry to pin might no
 408                  * longer to exist, but that means there's no longer anything to
 409                  * copy and we can bail out here:
 410                  */
 411                 spin_unlock(&j->lock);
 412                 return;
 413         }
 414
 415         reclaim = __journal_pin_drop(j, dst);
 416
 417         bch2_journal_pin_set_locked(j, seq, dst, flush_fn, journal_pin_type(flush_fn));
 418
 419         if (reclaim)
 420                 bch2_journal_reclaim_fast(j);
 421         spin_unlock(&j->lock);
 422
 423         /*
 424          * If the journal is currently full,  we might want to call flush_fn
 425          * immediately:
 426          */
 427         journal_wake(j);
 428 }
 429
 430 void bch2_journal_pin_set(struct journal *j, u64 seq,
 431                           struct journal_entry_pin *pin,
 432                           journal_pin_flush_fn flush_fn)
 433 {
 434         bool reclaim;
 435
 436         spin_lock(&j->lock);
 437
 438         BUG_ON(seq < journal_last_seq(j));
 439
 440         reclaim = __journal_pin_drop(j, pin);
 441
 442         bch2_journal_pin_set_locked(j, seq, pin, flush_fn, journal_pin_type(flush_fn));
 443
 444         if (reclaim)
 445                 bch2_journal_reclaim_fast(j);
 446         spin_unlock(&j->lock);
 447
 448         /*
 449          * If the journal is currently full,  we might want to call flush_fn
 450          * immediately:
 451          */
 452         journal_wake(j);
 453 }
 454
 455 /**
 456  * bch2_journal_pin_flush: ensure journal pin callback is no longer running
 457  * @j:          journal object
 458  * @pin:        pin to flush
 459  */
 460 void bch2_journal_pin_flush(struct journal *j, struct journal_entry_pin *pin)
 461 {
 462         BUG_ON(journal_pin_active(pin));
 463
 464         wait_event(j->pin_flush_wait, j->flush_in_progress != pin);
 465 }
 466
 467 /*
 468  * Journal reclaim: flush references to open journal entries to reclaim space in
 469  * the journal
 470  *
 471  * May be done by the journal code in the background as needed to free up space
 472  * for more journal entries, or as part of doing a clean shutdown, or to migrate
 473  * data off of a specific device:
 474  */
 475
 476 static struct journal_entry_pin *
 477 journal_get_next_pin(struct journal *j,
 478                      u64 seq_to_flush,
 479                      unsigned allowed_below_seq,
 480                      unsigned allowed_above_seq,
 481                      u64 *seq)
 482 {
 483         struct journal_entry_pin_list *pin_list;
 484         struct journal_entry_pin *ret = NULL;
 485         unsigned i;
 486
 487         fifo_for_each_entry_ptr(pin_list, &j->pin, *seq) {
 488                 if (*seq > seq_to_flush && !allowed_above_seq)
 489                         break;
 490
 491                 for (i = 0; i < JOURNAL_PIN_NR; i++)
 492                         if ((((1U << i) & allowed_below_seq) && *seq <= seq_to_flush) ||
 493                             ((1U << i) & allowed_above_seq)) {
 494                                 ret = list_first_entry_or_null(&pin_list->list[i],
 495                                         struct journal_entry_pin, list);
 496                                 if (ret)
 497                                         return ret;
 498                         }
 499         }
 500
 501         return NULL;
 502 }
 503
 504 /* returns true if we did work */
 505 static size_t journal_flush_pins(struct journal *j,
 506                                  u64 seq_to_flush,
 507                                  unsigned allowed_below_seq,
 508                                  unsigned allowed_above_seq,
 509                                  unsigned min_any,
 510                                  unsigned min_key_cache)
 511 {
 512         struct journal_entry_pin *pin;
 513         size_t nr_flushed = 0;
 514         journal_pin_flush_fn flush_fn;
 515         u64 seq;
 516         int err;
 517
 518         lockdep_assert_held(&j->reclaim_lock);
 519
 520         while (1) {
 521                 unsigned allowed_above = allowed_above_seq;
 522                 unsigned allowed_below = allowed_below_seq;
 523
 524                 if (min_any) {
 525                         allowed_above |= ~0;
 526                         allowed_below |= ~0;
 527                 }
 528
 529                 if (min_key_cache) {
 530                         allowed_above |= 1U << JOURNAL_PIN_key_cache;
 531                         allowed_below |= 1U << JOURNAL_PIN_key_cache;
 532                 }
 533
 534                 cond_resched();
 535
 536                 j->last_flushed = jiffies;
 537
 538                 spin_lock(&j->lock);
 539                 pin = journal_get_next_pin(j, seq_to_flush, allowed_below, allowed_above, &seq);
 540                 if (pin) {
 541                         BUG_ON(j->flush_in_progress);
 542                         j->flush_in_progress = pin;
 543                         j->flush_in_progress_dropped = false;
 544                         flush_fn = pin->flush;
 545                 }
 546                 spin_unlock(&j->lock);
 547
 548                 if (!pin)
 549                         break;
 550
 551                 if (min_key_cache && pin->flush == bch2_btree_key_cache_journal_flush)
 552                         min_key_cache--;
 553
 554                 if (min_any)
 555                         min_any--;
 556
 557                 err = flush_fn(j, pin, seq);
 558
 559                 spin_lock(&j->lock);
 560                 /* Pin might have been dropped or rearmed: */
 561                 if (likely(!err && !j->flush_in_progress_dropped))
 562                         list_move(&pin->list, &journal_seq_pin(j, seq)->flushed);
 563                 j->flush_in_progress = NULL;
 564                 j->flush_in_progress_dropped = false;
 565                 spin_unlock(&j->lock);
 566
 567                 wake_up(&j->pin_flush_wait);
 568
 569                 if (err)
 570                         break;
 571
 572                 nr_flushed++;
 573         }
 574
 575         return nr_flushed;
 576 }
 577
 578 static u64 journal_seq_to_flush(struct journal *j)
 579 {
 580         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 581         struct bch_dev *ca;
 582         u64 seq_to_flush = 0;
 583         unsigned iter;
 584
 585         spin_lock(&j->lock);
 586
 587         for_each_rw_member(ca, c, iter) {
 588                 struct journal_device *ja = &ca->journal;
 589                 unsigned nr_buckets, bucket_to_flush;
 590
 591                 if (!ja->nr)
 592                         continue;
 593
 594                 /* Try to keep the journal at most half full: */
 595                 nr_buckets = ja->nr / 2;
 596
 597                 nr_buckets = min(nr_buckets, ja->nr);
 598
 599                 bucket_to_flush = (ja->cur_idx + nr_buckets) % ja->nr;
 600                 seq_to_flush = max(seq_to_flush,
 601                                    ja->bucket_seq[bucket_to_flush]);
 602         }
 603
 604         /* Also flush if the pin fifo is more than half full */
 605         seq_to_flush = max_t(s64, seq_to_flush,
 606                              (s64) journal_cur_seq(j) -
 607                              (j->pin.size >> 1));
 608         spin_unlock(&j->lock);
 609
 610         return seq_to_flush;
 611 }
 612
 613 /**
 614  * __bch2_journal_reclaim - free up journal buckets
 615  * @j:          journal object
 616  * @direct:     direct or background reclaim?
 617  * @kicked:     requested to run since we last ran?
 618  * Returns:     0 on success, or -EIO if the journal has been shutdown
 619  *
 620  * Background journal reclaim writes out btree nodes. It should be run
 621  * early enough so that we never completely run out of journal buckets.
 622  *
 623  * High watermarks for triggering background reclaim:
 624  * - FIFO has fewer than 512 entries left
 625  * - fewer than 25% journal buckets free
 626  *
 627  * Background reclaim runs until low watermarks are reached:
 628  * - FIFO has more than 1024 entries left
 629  * - more than 50% journal buckets free
 630  *
 631  * As long as a reclaim can complete in the time it takes to fill up
 632  * 512 journal entries or 25% of all journal buckets, then
 633  * journal_next_bucket() should not stall.
 634  */
 635 static int __bch2_journal_reclaim(struct journal *j, bool direct, bool kicked)
 636 {
 637         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 638         bool kthread = (current->flags & PF_KTHREAD) != 0;
 639         u64 seq_to_flush;
 640         size_t min_nr, min_key_cache, nr_flushed;
 641         unsigned flags;
 642         int ret = 0;
 643
 644         /*
 645          * We can't invoke memory reclaim while holding the reclaim_lock -
 646          * journal reclaim is required to make progress for memory reclaim
 647          * (cleaning the caches), so we can't get stuck in memory reclaim while
 648          * we're holding the reclaim lock:
 649          */
 650         lockdep_assert_held(&j->reclaim_lock);
 651         flags = memalloc_noreclaim_save();
 652
 653         do {
 654                 if (kthread && kthread_should_stop())
 655                         break;
 656
 657                 if (bch2_journal_error(j)) {
 658                         ret = -EIO;
 659                         break;
 660                 }
 661
 662                 bch2_journal_do_discards(j);
 663
 664                 seq_to_flush = journal_seq_to_flush(j);
 665                 min_nr = 0;
 666
 667                 /*
 668                  * If it's been longer than j->reclaim_delay_ms since we last flushed,
 669                  * make sure to flush at least one journal pin:
 670                  */
 671                 if (time_after(jiffies, j->last_flushed +
 672                                msecs_to_jiffies(c->opts.journal_reclaim_delay)))
 673                         min_nr = 1;
 674
 675                 if (j->watermark != BCH_WATERMARK_stripe)
 676                         min_nr = 1;
 677
 678                 if (atomic_read(&c->btree_cache.dirty) * 2 > c->btree_cache.used)
 679                         min_nr = 1;
 680
 681                 min_key_cache = min(bch2_nr_btree_keys_need_flush(c), (size_t) 128);
 682
 683                 trace_and_count(c, journal_reclaim_start, c,
 684                                 direct, kicked,
 685                                 min_nr, min_key_cache,
 686                                 atomic_read(&c->btree_cache.dirty),
 687                                 c->btree_cache.used,
 688                                 atomic_long_read(&c->btree_key_cache.nr_dirty),
 689                                 atomic_long_read(&c->btree_key_cache.nr_keys));
 690
 691                 nr_flushed = journal_flush_pins(j, seq_to_flush,
 692                                                 ~0, 0,
 693                                                 min_nr, min_key_cache);
 694
 695                 if (direct)
 696                         j->nr_direct_reclaim += nr_flushed;
 697                 else
 698                         j->nr_background_reclaim += nr_flushed;
 699                 trace_and_count(c, journal_reclaim_finish, c, nr_flushed);
 700
 701                 if (nr_flushed)
 702                         wake_up(&j->reclaim_wait);
 703         } while ((min_nr || min_key_cache) && nr_flushed && !direct);
 704
 705         memalloc_noreclaim_restore(flags);
 706
 707         return ret;
 708 }
 709
 710 int bch2_journal_reclaim(struct journal *j)
 711 {
 712         return __bch2_journal_reclaim(j, true, true);
 713 }
 714
 715 static int bch2_journal_reclaim_thread(void *arg)
 716 {
 717         struct journal *j = arg;
 718         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 719         unsigned long delay, now;
 720         bool journal_empty;
 721         int ret = 0;
 722
 723         set_freezable();
 724
 725         j->last_flushed = jiffies;
 726
 727         while (!ret && !kthread_should_stop()) {
 728                 bool kicked = j->reclaim_kicked;
 729
 730                 j->reclaim_kicked = false;
 731
 732                 mutex_lock(&j->reclaim_lock);
 733                 ret = __bch2_journal_reclaim(j, false, kicked);
 734                 mutex_unlock(&j->reclaim_lock);
 735
 736                 now = jiffies;
 737                 delay = msecs_to_jiffies(c->opts.journal_reclaim_delay);
 738                 j->next_reclaim = j->last_flushed + delay;
 739
 740                 if (!time_in_range(j->next_reclaim, now, now + delay))
 741                         j->next_reclaim = now + delay;
 742
 743                 while (1) {
 744                         set_current_state(TASK_INTERRUPTIBLE|TASK_FREEZABLE);
 745                         if (kthread_should_stop())
 746                                 break;
 747                         if (j->reclaim_kicked)
 748                                 break;
 749
 750                         spin_lock(&j->lock);
 751                         journal_empty = fifo_empty(&j->pin);
 752                         spin_unlock(&j->lock);
 753
 754                         if (journal_empty)
 755                                 schedule();
 756                         else if (time_after(j->next_reclaim, jiffies))
 757                                 schedule_timeout(j->next_reclaim - jiffies);
 758                         else
 759                                 break;
 760                 }
 761                 __set_current_state(TASK_RUNNING);
 762         }
 763
 764         return 0;
 765 }
 766
 767 void bch2_journal_reclaim_stop(struct journal *j)
 768 {
 769         struct task_struct *p = j->reclaim_thread;
 770
 771         j->reclaim_thread = NULL;
 772
 773         if (p) {
 774                 kthread_stop(p);
 775                 put_task_struct(p);
 776         }
 777 }
 778
 779 int bch2_journal_reclaim_start(struct journal *j)
 780 {
 781         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 782         struct task_struct *p;
 783         int ret;
 784
 785         if (j->reclaim_thread)
 786                 return 0;
 787
 788         p = kthread_create(bch2_journal_reclaim_thread, j,
 789                            "bch-reclaim/%s", c->name);
 790         ret = PTR_ERR_OR_ZERO(p);
 791         if (ret) {
 792                 bch_err_msg(c, ret, "creating journal reclaim thread");
 793                 return ret;
 794         }
 795
 796         get_task_struct(p);
 797         j->reclaim_thread = p;
 798         wake_up_process(p);
 799         return 0;
 800 }
 801
 802 static int journal_flush_done(struct journal *j, u64 seq_to_flush,
 803                               bool *did_work)
 804 {
 805         int ret;
 806
 807         ret = bch2_journal_error(j);
 808         if (ret)
 809                 return ret;
 810
 811         mutex_lock(&j->reclaim_lock);
 812
 813         if (journal_flush_pins(j, seq_to_flush,
 814                                (1U << JOURNAL_PIN_key_cache)|
 815                                (1U << JOURNAL_PIN_other), 0, 0, 0) ||
 816             journal_flush_pins(j, seq_to_flush,
 817                                (1U << JOURNAL_PIN_btree), 0, 0, 0))
 818                 *did_work = true;
 819
 820         spin_lock(&j->lock);
 821         /*
 822          * If journal replay hasn't completed, the unreplayed journal entries
 823          * hold refs on their corresponding sequence numbers
 824          */
 825         ret = !test_bit(JOURNAL_REPLAY_DONE, &j->flags) ||
 826                 journal_last_seq(j) > seq_to_flush ||
 827                 !fifo_used(&j->pin);
 828
 829         spin_unlock(&j->lock);
 830         mutex_unlock(&j->reclaim_lock);
 831
 832         return ret;
 833 }
 834
 835 bool bch2_journal_flush_pins(struct journal *j, u64 seq_to_flush)
 836 {
 837         /* time_stats this */
 838         bool did_work = false;
 839
 840         if (!test_bit(JOURNAL_STARTED, &j->flags))
 841                 return false;
 842
 843         closure_wait_event(&j->async_wait,
 844                 journal_flush_done(j, seq_to_flush, &did_work));
 845
 846         return did_work;
 847 }
 848
 849 int bch2_journal_flush_device_pins(struct journal *j, int dev_idx)
 850 {
 851         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 852         struct journal_entry_pin_list *p;
 853         u64 iter, seq = 0;
 854         int ret = 0;
 855
 856         spin_lock(&j->lock);
 857         fifo_for_each_entry_ptr(p, &j->pin, iter)
 858                 if (dev_idx >= 0
 859                     ? bch2_dev_list_has_dev(p->devs, dev_idx)
 860                     : p->devs.nr < c->opts.metadata_replicas)
 861                         seq = iter;
 862         spin_unlock(&j->lock);
 863
 864         bch2_journal_flush_pins(j, seq);
 865
 866         ret = bch2_journal_error(j);
 867         if (ret)
 868                 return ret;
 869
 870         mutex_lock(&c->replicas_gc_lock);
 871         bch2_replicas_gc_start(c, 1 << BCH_DATA_journal);
 872
 873         /*
 874          * Now that we've populated replicas_gc, write to the journal to mark
 875          * active journal devices. This handles the case where the journal might
 876          * be empty. Otherwise we could clear all journal replicas and
 877          * temporarily put the fs into an unrecoverable state. Journal recovery
 878          * expects to find devices marked for journal data on unclean mount.
 879          */
 880         ret = bch2_journal_meta(&c->journal);
 881         if (ret)
 882                 goto err;
 883
 884         seq = 0;
 885         spin_lock(&j->lock);
 886         while (!ret) {
 887                 struct bch_replicas_padded replicas;
 888
 889                 seq = max(seq, journal_last_seq(j));
 890                 if (seq >= j->pin.back)
 891                         break;
 892                 bch2_devlist_to_replicas(&replicas.e, BCH_DATA_journal,
 893                                          journal_seq_pin(j, seq)->devs);
 894                 seq++;
 895
 896                 spin_unlock(&j->lock);
 897                 ret = bch2_mark_replicas(c, &replicas.e);
 898                 spin_lock(&j->lock);
 899         }
 900         spin_unlock(&j->lock);
 901 err:
 902         ret = bch2_replicas_gc_end(c, ret);
 903         mutex_unlock(&c->replicas_gc_lock);
 904
 905         return ret;
 906 }