fs/bcachefs/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 "btree_write_buffer.h"
   7 #include "buckets.h"
   8 #include "errcode.h"
   9 #include "error.h"
  10 #include "journal.h"
  11 #include "journal_io.h"
  12 #include "journal_reclaim.h"
  13 #include "replicas.h"
  14 #include "sb-members.h"
  15 #include "trace.h"
  16
  17 #include <linux/kthread.h>
  18 #include <linux/sched/mm.h>
  19
  20 /* Free space calculations: */
  21
  22 static unsigned journal_space_from(struct journal_device *ja,
  23                                    enum journal_space_from from)
  24 {
  25         switch (from) {
  26         case journal_space_discarded:
  27                 return ja->discard_idx;
  28         case journal_space_clean_ondisk:
  29                 return ja->dirty_idx_ondisk;
  30         case journal_space_clean:
  31                 return ja->dirty_idx;
  32         default:
  33                 BUG();
  34         }
  35 }
  36
  37 unsigned bch2_journal_dev_buckets_available(struct journal *j,
  38                                             struct journal_device *ja,
  39                                             enum journal_space_from from)
  40 {
  41         unsigned available = (journal_space_from(ja, from) -
  42                               ja->cur_idx - 1 + ja->nr) % ja->nr;
  43
  44         /*
  45          * Don't use the last bucket unless writing the new last_seq
  46          * will make another bucket available:
  47          */
  48         if (available && ja->dirty_idx_ondisk == ja->dirty_idx)
  49                 --available;
  50
  51         return available;
  52 }
  53
  54 void bch2_journal_set_watermark(struct journal *j)
  55 {
  56         struct bch_fs *c = container_of(j, struct bch_fs, journal);
  57         bool low_on_space = j->space[journal_space_clean].total * 4 <=
  58                 j->space[journal_space_total].total;
  59         bool low_on_pin = fifo_free(&j->pin) < j->pin.size / 4;
  60         bool low_on_wb = bch2_btree_write_buffer_must_wait(c);
  61         unsigned watermark = low_on_space || low_on_pin || low_on_wb
  62                 ? BCH_WATERMARK_reclaim
  63                 : BCH_WATERMARK_stripe;
  64
  65         if (track_event_change(&c->times[BCH_TIME_blocked_journal_low_on_space], low_on_space) ||
  66             track_event_change(&c->times[BCH_TIME_blocked_journal_low_on_pin], low_on_pin) ||
  67             track_event_change(&c->times[BCH_TIME_blocked_write_buffer_full], low_on_wb))
  68                 trace_and_count(c, journal_full, c);
  69
  70         mod_bit(JOURNAL_space_low, &j->flags, low_on_space || low_on_pin);
  71
  72         swap(watermark, j->watermark);
  73         if (watermark > j->watermark)
  74                 journal_wake(j);
  75 }
  76
  77 static struct journal_space
  78 journal_dev_space_available(struct journal *j, struct bch_dev *ca,
  79                             enum journal_space_from from)
  80 {
  81         struct journal_device *ja = &ca->journal;
  82         unsigned sectors, buckets, unwritten;
  83         u64 seq;
  84
  85         if (from == journal_space_total)
  86                 return (struct journal_space) {
  87                         .next_entry     = ca->mi.bucket_size,
  88                         .total          = ca->mi.bucket_size * ja->nr,
  89                 };
  90
  91         buckets = bch2_journal_dev_buckets_available(j, ja, from);
  92         sectors = ja->sectors_free;
  93
  94         /*
  95          * We that we don't allocate the space for a journal entry
  96          * until we write it out - thus, account for it here:
  97          */
  98         for (seq = journal_last_unwritten_seq(j);
  99              seq <= journal_cur_seq(j);
 100              seq++) {
 101                 unwritten = j->buf[seq & JOURNAL_BUF_MASK].sectors;
 102
 103                 if (!unwritten)
 104                         continue;
 105
 106                 /* entry won't fit on this device, skip: */
 107                 if (unwritten > ca->mi.bucket_size)
 108                         continue;
 109
 110                 if (unwritten >= sectors) {
 111                         if (!buckets) {
 112                                 sectors = 0;
 113                                 break;
 114                         }
 115
 116                         buckets--;
 117                         sectors = ca->mi.bucket_size;
 118                 }
 119
 120                 sectors -= unwritten;
 121         }
 122
 123         if (sectors < ca->mi.bucket_size && buckets) {
 124                 buckets--;
 125                 sectors = ca->mi.bucket_size;
 126         }
 127
 128         return (struct journal_space) {
 129                 .next_entry     = sectors,
 130                 .total          = sectors + buckets * ca->mi.bucket_size,
 131         };
 132 }
 133
 134 static struct journal_space __journal_space_available(struct journal *j, unsigned nr_devs_want,
 135                             enum journal_space_from from)
 136 {
 137         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 138         unsigned pos, nr_devs = 0;
 139         struct journal_space space, dev_space[BCH_SB_MEMBERS_MAX];
 140
 141         BUG_ON(nr_devs_want > ARRAY_SIZE(dev_space));
 142
 143         rcu_read_lock();
 144         for_each_member_device_rcu(c, ca, &c->rw_devs[BCH_DATA_journal]) {
 145                 if (!ca->journal.nr)
 146                         continue;
 147
 148                 space = journal_dev_space_available(j, ca, from);
 149                 if (!space.next_entry)
 150                         continue;
 151
 152                 for (pos = 0; pos < nr_devs; pos++)
 153                         if (space.total > dev_space[pos].total)
 154                                 break;
 155
 156                 array_insert_item(dev_space, nr_devs, pos, space);
 157         }
 158         rcu_read_unlock();
 159
 160         if (nr_devs < nr_devs_want)
 161                 return (struct journal_space) { 0, 0 };
 162
 163         /*
 164          * We sorted largest to smallest, and we want the smallest out of the
 165          * @nr_devs_want largest devices:
 166          */
 167         return dev_space[nr_devs_want - 1];
 168 }
 169
 170 void bch2_journal_space_available(struct journal *j)
 171 {
 172         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 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 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(c, ca, &c->rw_devs[BCH_DATA_journal]) {
 184                 struct journal_device *ja = &ca->journal;
 185
 186                 if (!ja->nr)
 187                         continue;
 188
 189                 while (ja->dirty_idx != ja->cur_idx &&
 190                        ja->bucket_seq[ja->dirty_idx] < journal_last_seq(j))
 191                         ja->dirty_idx = (ja->dirty_idx + 1) % ja->nr;
 192
 193                 while (ja->dirty_idx_ondisk != ja->dirty_idx &&
 194                        ja->bucket_seq[ja->dirty_idx_ondisk] < j->last_seq_ondisk)
 195                         ja->dirty_idx_ondisk = (ja->dirty_idx_ondisk + 1) % ja->nr;
 196
 197                 if (ja->discard_idx != ja->dirty_idx_ondisk)
 198                         can_discard = true;
 199
 200                 max_entry_size = min_t(unsigned, max_entry_size, ca->mi.bucket_size);
 201                 nr_online++;
 202         }
 203         rcu_read_unlock();
 204
 205         j->can_discard = can_discard;
 206
 207         if (nr_online < metadata_replicas_required(c)) {
 208                 struct printbuf buf = PRINTBUF;
 209                 buf.atomic++;
 210                 prt_printf(&buf, "insufficient writeable journal devices available: have %u, need %u\n"
 211                            "rw journal devs:", nr_online, metadata_replicas_required(c));
 212
 213                 rcu_read_lock();
 214                 for_each_member_device_rcu(c, ca, &c->rw_devs[BCH_DATA_journal])
 215                         prt_printf(&buf, " %s", ca->name);
 216                 rcu_read_unlock();
 217
 218                 bch_err(c, "%s", buf.buf);
 219                 printbuf_exit(&buf);
 220                 ret = JOURNAL_ERR_insufficient_devices;
 221                 goto out;
 222         }
 223
 224         nr_devs_want = min_t(unsigned, nr_online, c->opts.metadata_replicas);
 225
 226         for (unsigned i = 0; i < journal_space_nr; i++)
 227                 j->space[i] = __journal_space_available(j, nr_devs_want, i);
 228
 229         clean_ondisk    = j->space[journal_space_clean_ondisk].total;
 230         clean           = j->space[journal_space_clean].total;
 231         total           = j->space[journal_space_total].total;
 232
 233         if (!j->space[journal_space_discarded].next_entry)
 234                 ret = JOURNAL_ERR_journal_full;
 235
 236         if ((j->space[journal_space_clean_ondisk].next_entry <
 237              j->space[journal_space_clean_ondisk].total) &&
 238             (clean - clean_ondisk <= total / 8) &&
 239             (clean_ondisk * 2 > clean))
 240                 set_bit(JOURNAL_may_skip_flush, &j->flags);
 241         else
 242                 clear_bit(JOURNAL_may_skip_flush, &j->flags);
 243
 244         bch2_journal_set_watermark(j);
 245 out:
 246         j->cur_entry_sectors    = !ret ? j->space[journal_space_discarded].next_entry : 0;
 247         j->cur_entry_error      = ret;
 248
 249         if (!ret)
 250                 journal_wake(j);
 251 }
 252
 253 /* Discards - last part of journal reclaim: */
 254
 255 static bool should_discard_bucket(struct journal *j, struct journal_device *ja)
 256 {
 257         bool ret;
 258
 259         spin_lock(&j->lock);
 260         ret = ja->discard_idx != ja->dirty_idx_ondisk;
 261         spin_unlock(&j->lock);
 262
 263         return ret;
 264 }
 265
 266 /*
 267  * Advance ja->discard_idx as long as it points to buckets that are no longer
 268  * dirty, issuing discards if necessary:
 269  */
 270 void bch2_journal_do_discards(struct journal *j)
 271 {
 272         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 273
 274         mutex_lock(&j->discard_lock);
 275
 276         for_each_rw_member(c, ca) {
 277                 struct journal_device *ja = &ca->journal;
 278
 279                 while (should_discard_bucket(j, ja)) {
 280                         if (!c->opts.nochanges &&
 281                             ca->mi.discard &&
 282                             bdev_max_discard_sectors(ca->disk_sb.bdev))
 283                                 blkdev_issue_discard(ca->disk_sb.bdev,
 284                                         bucket_to_sector(ca,
 285                                                 ja->buckets[ja->discard_idx]),
 286                                         ca->mi.bucket_size, GFP_NOFS);
 287
 288                         spin_lock(&j->lock);
 289                         ja->discard_idx = (ja->discard_idx + 1) % ja->nr;
 290
 291                         bch2_journal_space_available(j);
 292                         spin_unlock(&j->lock);
 293                 }
 294         }
 295
 296         mutex_unlock(&j->discard_lock);
 297 }
 298
 299 /*
 300  * Journal entry pinning - machinery for holding a reference on a given journal
 301  * entry, holding it open to ensure it gets replayed during recovery:
 302  */
 303
 304 void bch2_journal_reclaim_fast(struct journal *j)
 305 {
 306         bool popped = false;
 307
 308         lockdep_assert_held(&j->lock);
 309
 310         /*
 311          * Unpin journal entries whose reference counts reached zero, meaning
 312          * all btree nodes got written out
 313          */
 314         while (!fifo_empty(&j->pin) &&
 315                j->pin.front <= j->seq_ondisk &&
 316                !atomic_read(&fifo_peek_front(&j->pin).count)) {
 317                 j->pin.front++;
 318                 popped = true;
 319         }
 320
 321         if (popped)
 322                 bch2_journal_space_available(j);
 323 }
 324
 325 bool __bch2_journal_pin_put(struct journal *j, u64 seq)
 326 {
 327         struct journal_entry_pin_list *pin_list = journal_seq_pin(j, seq);
 328
 329         return atomic_dec_and_test(&pin_list->count);
 330 }
 331
 332 void bch2_journal_pin_put(struct journal *j, u64 seq)
 333 {
 334         if (__bch2_journal_pin_put(j, seq)) {
 335                 spin_lock(&j->lock);
 336                 bch2_journal_reclaim_fast(j);
 337                 spin_unlock(&j->lock);
 338         }
 339 }
 340
 341 static inline bool __journal_pin_drop(struct journal *j,
 342                                       struct journal_entry_pin *pin)
 343 {
 344         struct journal_entry_pin_list *pin_list;
 345
 346         if (!journal_pin_active(pin))
 347                 return false;
 348
 349         if (j->flush_in_progress == pin)
 350                 j->flush_in_progress_dropped = true;
 351
 352         pin_list = journal_seq_pin(j, pin->seq);
 353         pin->seq = 0;
 354         list_del_init(&pin->list);
 355
 356         /*
 357          * Unpinning a journal entry may make journal_next_bucket() succeed, if
 358          * writing a new last_seq will now make another bucket available:
 359          */
 360         return atomic_dec_and_test(&pin_list->count) &&
 361                 pin_list == &fifo_peek_front(&j->pin);
 362 }
 363
 364 void bch2_journal_pin_drop(struct journal *j,
 365                            struct journal_entry_pin *pin)
 366 {
 367         spin_lock(&j->lock);
 368         if (__journal_pin_drop(j, pin))
 369                 bch2_journal_reclaim_fast(j);
 370         spin_unlock(&j->lock);
 371 }
 372
 373 static enum journal_pin_type journal_pin_type(journal_pin_flush_fn fn)
 374 {
 375         if (fn == bch2_btree_node_flush0 ||
 376             fn == bch2_btree_node_flush1)
 377                 return JOURNAL_PIN_btree;
 378         else if (fn == bch2_btree_key_cache_journal_flush)
 379                 return JOURNAL_PIN_key_cache;
 380         else
 381                 return JOURNAL_PIN_other;
 382 }
 383
 384 static inline void bch2_journal_pin_set_locked(struct journal *j, u64 seq,
 385                           struct journal_entry_pin *pin,
 386                           journal_pin_flush_fn flush_fn,
 387                           enum journal_pin_type type)
 388 {
 389         struct journal_entry_pin_list *pin_list = journal_seq_pin(j, seq);
 390
 391         /*
 392          * flush_fn is how we identify journal pins in debugfs, so must always
 393          * exist, even if it doesn't do anything:
 394          */
 395         BUG_ON(!flush_fn);
 396
 397         atomic_inc(&pin_list->count);
 398         pin->seq        = seq;
 399         pin->flush      = flush_fn;
 400         list_add(&pin->list, &pin_list->list[type]);
 401 }
 402
 403 void bch2_journal_pin_copy(struct journal *j,
 404                            struct journal_entry_pin *dst,
 405                            struct journal_entry_pin *src,
 406                            journal_pin_flush_fn flush_fn)
 407 {
 408         spin_lock(&j->lock);
 409
 410         u64 seq = READ_ONCE(src->seq);
 411
 412         if (seq < journal_last_seq(j)) {
 413                 /*
 414                  * bch2_journal_pin_copy() raced with bch2_journal_pin_drop() on
 415                  * the src pin - with the pin dropped, the entry to pin might no
 416                  * longer to exist, but that means there's no longer anything to
 417                  * copy and we can bail out here:
 418                  */
 419                 spin_unlock(&j->lock);
 420                 return;
 421         }
 422
 423         bool reclaim = __journal_pin_drop(j, dst);
 424
 425         bch2_journal_pin_set_locked(j, seq, dst, flush_fn, journal_pin_type(flush_fn));
 426
 427         if (reclaim)
 428                 bch2_journal_reclaim_fast(j);
 429
 430         /*
 431          * If the journal is currently full,  we might want to call flush_fn
 432          * immediately:
 433          */
 434         if (seq == journal_last_seq(j))
 435                 journal_wake(j);
 436         spin_unlock(&j->lock);
 437 }
 438
 439 void bch2_journal_pin_set(struct journal *j, u64 seq,
 440                           struct journal_entry_pin *pin,
 441                           journal_pin_flush_fn flush_fn)
 442 {
 443         spin_lock(&j->lock);
 444
 445         BUG_ON(seq < journal_last_seq(j));
 446
 447         bool reclaim = __journal_pin_drop(j, pin);
 448
 449         bch2_journal_pin_set_locked(j, seq, pin, flush_fn, journal_pin_type(flush_fn));
 450
 451         if (reclaim)
 452                 bch2_journal_reclaim_fast(j);
 453         /*
 454          * If the journal is currently full,  we might want to call flush_fn
 455          * immediately:
 456          */
 457         if (seq == journal_last_seq(j))
 458                 journal_wake(j);
 459
 460         spin_unlock(&j->lock);
 461 }
 462
 463 /**
 464  * bch2_journal_pin_flush: ensure journal pin callback is no longer running
 465  * @j:          journal object
 466  * @pin:        pin to flush
 467  */
 468 void bch2_journal_pin_flush(struct journal *j, struct journal_entry_pin *pin)
 469 {
 470         BUG_ON(journal_pin_active(pin));
 471
 472         wait_event(j->pin_flush_wait, j->flush_in_progress != pin);
 473 }
 474
 475 /*
 476  * Journal reclaim: flush references to open journal entries to reclaim space in
 477  * the journal
 478  *
 479  * May be done by the journal code in the background as needed to free up space
 480  * for more journal entries, or as part of doing a clean shutdown, or to migrate
 481  * data off of a specific device:
 482  */
 483
 484 static struct journal_entry_pin *
 485 journal_get_next_pin(struct journal *j,
 486                      u64 seq_to_flush,
 487                      unsigned allowed_below_seq,
 488                      unsigned allowed_above_seq,
 489                      u64 *seq)
 490 {
 491         struct journal_entry_pin_list *pin_list;
 492         struct journal_entry_pin *ret = NULL;
 493         unsigned i;
 494
 495         fifo_for_each_entry_ptr(pin_list, &j->pin, *seq) {
 496                 if (*seq > seq_to_flush && !allowed_above_seq)
 497                         break;
 498
 499                 for (i = 0; i < JOURNAL_PIN_NR; i++)
 500                         if ((((1U << i) & allowed_below_seq) && *seq <= seq_to_flush) ||
 501                             ((1U << i) & allowed_above_seq)) {
 502                                 ret = list_first_entry_or_null(&pin_list->list[i],
 503                                         struct journal_entry_pin, list);
 504                                 if (ret)
 505                                         return ret;
 506                         }
 507         }
 508
 509         return NULL;
 510 }
 511
 512 /* returns true if we did work */
 513 static size_t journal_flush_pins(struct journal *j,
 514                                  u64 seq_to_flush,
 515                                  unsigned allowed_below_seq,
 516                                  unsigned allowed_above_seq,
 517                                  unsigned min_any,
 518                                  unsigned min_key_cache)
 519 {
 520         struct journal_entry_pin *pin;
 521         size_t nr_flushed = 0;
 522         journal_pin_flush_fn flush_fn;
 523         u64 seq;
 524         int err;
 525
 526         lockdep_assert_held(&j->reclaim_lock);
 527
 528         while (1) {
 529                 unsigned allowed_above = allowed_above_seq;
 530                 unsigned allowed_below = allowed_below_seq;
 531
 532                 if (min_any) {
 533                         allowed_above |= ~0;
 534                         allowed_below |= ~0;
 535                 }
 536
 537                 if (min_key_cache) {
 538                         allowed_above |= 1U << JOURNAL_PIN_key_cache;
 539                         allowed_below |= 1U << JOURNAL_PIN_key_cache;
 540                 }
 541
 542                 cond_resched();
 543
 544                 j->last_flushed = jiffies;
 545
 546                 spin_lock(&j->lock);
 547                 pin = journal_get_next_pin(j, seq_to_flush, allowed_below, allowed_above, &seq);
 548                 if (pin) {
 549                         BUG_ON(j->flush_in_progress);
 550                         j->flush_in_progress = pin;
 551                         j->flush_in_progress_dropped = false;
 552                         flush_fn = pin->flush;
 553                 }
 554                 spin_unlock(&j->lock);
 555
 556                 if (!pin)
 557                         break;
 558
 559                 if (min_key_cache && pin->flush == bch2_btree_key_cache_journal_flush)
 560                         min_key_cache--;
 561
 562                 if (min_any)
 563                         min_any--;
 564
 565                 err = flush_fn(j, pin, seq);
 566
 567                 spin_lock(&j->lock);
 568                 /* Pin might have been dropped or rearmed: */
 569                 if (likely(!err && !j->flush_in_progress_dropped))
 570                         list_move(&pin->list, &journal_seq_pin(j, seq)->flushed);
 571                 j->flush_in_progress = NULL;
 572                 j->flush_in_progress_dropped = false;
 573                 spin_unlock(&j->lock);
 574
 575                 wake_up(&j->pin_flush_wait);
 576
 577                 if (err)
 578                         break;
 579
 580                 nr_flushed++;
 581         }
 582
 583         return nr_flushed;
 584 }
 585
 586 static u64 journal_seq_to_flush(struct journal *j)
 587 {
 588         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 589         u64 seq_to_flush = 0;
 590
 591         spin_lock(&j->lock);
 592
 593         for_each_rw_member(c, ca) {
 594                 struct journal_device *ja = &ca->journal;
 595                 unsigned nr_buckets, bucket_to_flush;
 596
 597                 if (!ja->nr)
 598                         continue;
 599
 600                 /* Try to keep the journal at most half full: */
 601                 nr_buckets = ja->nr / 2;
 602
 603                 nr_buckets = min(nr_buckets, ja->nr);
 604
 605                 bucket_to_flush = (ja->cur_idx + nr_buckets) % ja->nr;
 606                 seq_to_flush = max(seq_to_flush,
 607                                    ja->bucket_seq[bucket_to_flush]);
 608         }
 609
 610         /* Also flush if the pin fifo is more than half full */
 611         seq_to_flush = max_t(s64, seq_to_flush,
 612                              (s64) journal_cur_seq(j) -
 613                              (j->pin.size >> 1));
 614         spin_unlock(&j->lock);
 615
 616         return seq_to_flush;
 617 }
 618
 619 /**
 620  * __bch2_journal_reclaim - free up journal buckets
 621  * @j:          journal object
 622  * @direct:     direct or background reclaim?
 623  * @kicked:     requested to run since we last ran?
 624  * Returns:     0 on success, or -EIO if the journal has been shutdown
 625  *
 626  * Background journal reclaim writes out btree nodes. It should be run
 627  * early enough so that we never completely run out of journal buckets.
 628  *
 629  * High watermarks for triggering background reclaim:
 630  * - FIFO has fewer than 512 entries left
 631  * - fewer than 25% journal buckets free
 632  *
 633  * Background reclaim runs until low watermarks are reached:
 634  * - FIFO has more than 1024 entries left
 635  * - more than 50% journal buckets free
 636  *
 637  * As long as a reclaim can complete in the time it takes to fill up
 638  * 512 journal entries or 25% of all journal buckets, then
 639  * journal_next_bucket() should not stall.
 640  */
 641 static int __bch2_journal_reclaim(struct journal *j, bool direct, bool kicked)
 642 {
 643         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 644         struct btree_cache *bc = &c->btree_cache;
 645         bool kthread = (current->flags & PF_KTHREAD) != 0;
 646         u64 seq_to_flush;
 647         size_t min_nr, min_key_cache, nr_flushed;
 648         unsigned flags;
 649         int ret = 0;
 650
 651         /*
 652          * We can't invoke memory reclaim while holding the reclaim_lock -
 653          * journal reclaim is required to make progress for memory reclaim
 654          * (cleaning the caches), so we can't get stuck in memory reclaim while
 655          * we're holding the reclaim lock:
 656          */
 657         lockdep_assert_held(&j->reclaim_lock);
 658         flags = memalloc_noreclaim_save();
 659
 660         do {
 661                 if (kthread && kthread_should_stop())
 662                         break;
 663
 664                 if (bch2_journal_error(j)) {
 665                         ret = -EIO;
 666                         break;
 667                 }
 668
 669                 bch2_journal_do_discards(j);
 670
 671                 seq_to_flush = journal_seq_to_flush(j);
 672                 min_nr = 0;
 673
 674                 /*
 675                  * If it's been longer than j->reclaim_delay_ms since we last flushed,
 676                  * make sure to flush at least one journal pin:
 677                  */
 678                 if (time_after(jiffies, j->last_flushed +
 679                                msecs_to_jiffies(c->opts.journal_reclaim_delay)))
 680                         min_nr = 1;
 681
 682                 if (j->watermark != BCH_WATERMARK_stripe)
 683                         min_nr = 1;
 684
 685                 size_t btree_cache_live = bc->live[0].nr + bc->live[1].nr;
 686                 if (atomic_long_read(&bc->nr_dirty) * 2 > btree_cache_live)
 687                         min_nr = 1;
 688
 689                 min_key_cache = min(bch2_nr_btree_keys_need_flush(c), (size_t) 128);
 690
 691                 trace_and_count(c, journal_reclaim_start, c,
 692                                 direct, kicked,
 693                                 min_nr, min_key_cache,
 694                                 atomic_long_read(&bc->nr_dirty), btree_cache_live,
 695                                 atomic_long_read(&c->btree_key_cache.nr_dirty),
 696                                 atomic_long_read(&c->btree_key_cache.nr_keys));
 697
 698                 nr_flushed = journal_flush_pins(j, seq_to_flush,
 699                                                 ~0, 0,
 700                                                 min_nr, min_key_cache);
 701
 702                 if (direct)
 703                         j->nr_direct_reclaim += nr_flushed;
 704                 else
 705                         j->nr_background_reclaim += nr_flushed;
 706                 trace_and_count(c, journal_reclaim_finish, c, nr_flushed);
 707
 708                 if (nr_flushed)
 709                         wake_up(&j->reclaim_wait);
 710         } while ((min_nr || min_key_cache) && nr_flushed && !direct);
 711
 712         memalloc_noreclaim_restore(flags);
 713
 714         return ret;
 715 }
 716
 717 int bch2_journal_reclaim(struct journal *j)
 718 {
 719         return __bch2_journal_reclaim(j, true, true);
 720 }
 721
 722 static int bch2_journal_reclaim_thread(void *arg)
 723 {
 724         struct journal *j = arg;
 725         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 726         unsigned long delay, now;
 727         bool journal_empty;
 728         int ret = 0;
 729
 730         set_freezable();
 731
 732         j->last_flushed = jiffies;
 733
 734         while (!ret && !kthread_should_stop()) {
 735                 bool kicked = j->reclaim_kicked;
 736
 737                 j->reclaim_kicked = false;
 738
 739                 mutex_lock(&j->reclaim_lock);
 740                 ret = __bch2_journal_reclaim(j, false, kicked);
 741                 mutex_unlock(&j->reclaim_lock);
 742
 743                 now = jiffies;
 744                 delay = msecs_to_jiffies(c->opts.journal_reclaim_delay);
 745                 j->next_reclaim = j->last_flushed + delay;
 746
 747                 if (!time_in_range(j->next_reclaim, now, now + delay))
 748                         j->next_reclaim = now + delay;
 749
 750                 while (1) {
 751                         set_current_state(TASK_INTERRUPTIBLE|TASK_FREEZABLE);
 752                         if (kthread_should_stop())
 753                                 break;
 754                         if (j->reclaim_kicked)
 755                                 break;
 756
 757                         spin_lock(&j->lock);
 758                         journal_empty = fifo_empty(&j->pin);
 759                         spin_unlock(&j->lock);
 760
 761                         if (journal_empty)
 762                                 schedule();
 763                         else if (time_after(j->next_reclaim, jiffies))
 764                                 schedule_timeout(j->next_reclaim - jiffies);
 765                         else
 766                                 break;
 767                 }
 768                 __set_current_state(TASK_RUNNING);
 769         }
 770
 771         return 0;
 772 }
 773
 774 void bch2_journal_reclaim_stop(struct journal *j)
 775 {
 776         struct task_struct *p = j->reclaim_thread;
 777
 778         j->reclaim_thread = NULL;
 779
 780         if (p) {
 781                 kthread_stop(p);
 782                 put_task_struct(p);
 783         }
 784 }
 785
 786 int bch2_journal_reclaim_start(struct journal *j)
 787 {
 788         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 789         struct task_struct *p;
 790         int ret;
 791
 792         if (j->reclaim_thread)
 793                 return 0;
 794
 795         p = kthread_create(bch2_journal_reclaim_thread, j,
 796                            "bch-reclaim/%s", c->name);
 797         ret = PTR_ERR_OR_ZERO(p);
 798         bch_err_msg(c, ret, "creating journal reclaim thread");
 799         if (ret)
 800                 return ret;
 801
 802         get_task_struct(p);
 803         j->reclaim_thread = p;
 804         wake_up_process(p);
 805         return 0;
 806 }
 807
 808 static int journal_flush_done(struct journal *j, u64 seq_to_flush,
 809                               bool *did_work)
 810 {
 811         int ret;
 812
 813         ret = bch2_journal_error(j);
 814         if (ret)
 815                 return ret;
 816
 817         mutex_lock(&j->reclaim_lock);
 818
 819         if (journal_flush_pins(j, seq_to_flush,
 820                                (1U << JOURNAL_PIN_key_cache)|
 821                                (1U << JOURNAL_PIN_other), 0, 0, 0) ||
 822             journal_flush_pins(j, seq_to_flush,
 823                                (1U << JOURNAL_PIN_btree), 0, 0, 0))
 824                 *did_work = true;
 825
 826         if (seq_to_flush > journal_cur_seq(j))
 827                 bch2_journal_entry_close(j);
 828
 829         spin_lock(&j->lock);
 830         /*
 831          * If journal replay hasn't completed, the unreplayed journal entries
 832          * hold refs on their corresponding sequence numbers
 833          */
 834         ret = !test_bit(JOURNAL_replay_done, &j->flags) ||
 835                 journal_last_seq(j) > seq_to_flush ||
 836                 !fifo_used(&j->pin);
 837
 838         spin_unlock(&j->lock);
 839         mutex_unlock(&j->reclaim_lock);
 840
 841         return ret;
 842 }
 843
 844 bool bch2_journal_flush_pins(struct journal *j, u64 seq_to_flush)
 845 {
 846         /* time_stats this */
 847         bool did_work = false;
 848
 849         if (!test_bit(JOURNAL_running, &j->flags))
 850                 return false;
 851
 852         closure_wait_event(&j->async_wait,
 853                 journal_flush_done(j, seq_to_flush, &did_work));
 854
 855         return did_work;
 856 }
 857
 858 int bch2_journal_flush_device_pins(struct journal *j, int dev_idx)
 859 {
 860         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 861         struct journal_entry_pin_list *p;
 862         u64 iter, seq = 0;
 863         int ret = 0;
 864
 865         spin_lock(&j->lock);
 866         fifo_for_each_entry_ptr(p, &j->pin, iter)
 867                 if (dev_idx >= 0
 868                     ? bch2_dev_list_has_dev(p->devs, dev_idx)
 869                     : p->devs.nr < c->opts.metadata_replicas)
 870                         seq = iter;
 871         spin_unlock(&j->lock);
 872
 873         bch2_journal_flush_pins(j, seq);
 874
 875         ret = bch2_journal_error(j);
 876         if (ret)
 877                 return ret;
 878
 879         mutex_lock(&c->replicas_gc_lock);
 880         bch2_replicas_gc_start(c, 1 << BCH_DATA_journal);
 881
 882         /*
 883          * Now that we've populated replicas_gc, write to the journal to mark
 884          * active journal devices. This handles the case where the journal might
 885          * be empty. Otherwise we could clear all journal replicas and
 886          * temporarily put the fs into an unrecoverable state. Journal recovery
 887          * expects to find devices marked for journal data on unclean mount.
 888          */
 889         ret = bch2_journal_meta(&c->journal);
 890         if (ret)
 891                 goto err;
 892
 893         seq = 0;
 894         spin_lock(&j->lock);
 895         while (!ret) {
 896                 struct bch_replicas_padded replicas;
 897
 898                 seq = max(seq, journal_last_seq(j));
 899                 if (seq >= j->pin.back)
 900                         break;
 901                 bch2_devlist_to_replicas(&replicas.e, BCH_DATA_journal,
 902                                          journal_seq_pin(j, seq)->devs);
 903                 seq++;
 904
 905                 if (replicas.e.nr_devs) {
 906                         spin_unlock(&j->lock);
 907                         ret = bch2_mark_replicas(c, &replicas.e);
 908                         spin_lock(&j->lock);
 909                 }
 910         }
 911         spin_unlock(&j->lock);
 912 err:
 913         ret = bch2_replicas_gc_end(c, ret);
 914         mutex_unlock(&c->replicas_gc_lock);
 915
 916         return ret;
 917 }