1 // SPDX-License-Identifier: GPL-2.0
4 #include "btree_key_cache.h"
5 #include "btree_update.h"
6 #include "btree_write_buffer.h"
11 #include "journal_io.h"
12 #include "journal_reclaim.h"
14 #include "sb-members.h"
17 #include <linux/kthread.h>
18 #include <linux/sched/mm.h>
20 /* Free space calculations: */
22 static unsigned journal_space_from(struct journal_device *ja,
23 enum journal_space_from 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:
37 unsigned bch2_journal_dev_buckets_available(struct journal *j,
38 struct journal_device *ja,
39 enum journal_space_from from)
41 unsigned available = (journal_space_from(ja, from) -
42 ja->cur_idx - 1 + ja->nr) % ja->nr;
45 * Don't use the last bucket unless writing the new last_seq
46 * will make another bucket available:
48 if (available && ja->dirty_idx_ondisk == ja->dirty_idx)
54 void bch2_journal_set_watermark(struct journal *j)
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;
65 if (track_event_change(&c->times[BCH_TIME_blocked_journal_low_on_space],
66 &j->low_on_space_start, low_on_space) ||
67 track_event_change(&c->times[BCH_TIME_blocked_journal_low_on_pin],
68 &j->low_on_pin_start, low_on_pin) ||
69 track_event_change(&c->times[BCH_TIME_blocked_write_buffer_full],
70 &j->write_buffer_full_start, low_on_wb))
71 trace_and_count(c, journal_full, c);
73 swap(watermark, j->watermark);
74 if (watermark > j->watermark)
78 static struct journal_space
79 journal_dev_space_available(struct journal *j, struct bch_dev *ca,
80 enum journal_space_from from)
82 struct journal_device *ja = &ca->journal;
83 unsigned sectors, buckets, unwritten;
86 if (from == journal_space_total)
87 return (struct journal_space) {
88 .next_entry = ca->mi.bucket_size,
89 .total = ca->mi.bucket_size * ja->nr,
92 buckets = bch2_journal_dev_buckets_available(j, ja, from);
93 sectors = ja->sectors_free;
96 * We that we don't allocate the space for a journal entry
97 * until we write it out - thus, account for it here:
99 for (seq = journal_last_unwritten_seq(j);
100 seq <= journal_cur_seq(j);
102 unwritten = j->buf[seq & JOURNAL_BUF_MASK].sectors;
107 /* entry won't fit on this device, skip: */
108 if (unwritten > ca->mi.bucket_size)
111 if (unwritten >= sectors) {
118 sectors = ca->mi.bucket_size;
121 sectors -= unwritten;
124 if (sectors < ca->mi.bucket_size && buckets) {
126 sectors = ca->mi.bucket_size;
129 return (struct journal_space) {
130 .next_entry = sectors,
131 .total = sectors + buckets * ca->mi.bucket_size,
135 static struct journal_space __journal_space_available(struct journal *j, unsigned nr_devs_want,
136 enum journal_space_from from)
138 struct bch_fs *c = container_of(j, struct bch_fs, journal);
140 unsigned i, pos, nr_devs = 0;
141 struct journal_space space, dev_space[BCH_SB_MEMBERS_MAX];
143 BUG_ON(nr_devs_want > ARRAY_SIZE(dev_space));
146 for_each_member_device_rcu(ca, c, i,
147 &c->rw_devs[BCH_DATA_journal]) {
151 space = journal_dev_space_available(j, ca, from);
152 if (!space.next_entry)
155 for (pos = 0; pos < nr_devs; pos++)
156 if (space.total > dev_space[pos].total)
159 array_insert_item(dev_space, nr_devs, pos, space);
163 if (nr_devs < nr_devs_want)
164 return (struct journal_space) { 0, 0 };
167 * We sorted largest to smallest, and we want the smallest out of the
168 * @nr_devs_want largest devices:
170 return dev_space[nr_devs_want - 1];
173 void bch2_journal_space_available(struct journal *j)
175 struct bch_fs *c = container_of(j, struct bch_fs, journal);
177 unsigned clean, clean_ondisk, total;
178 unsigned max_entry_size = min(j->buf[0].buf_size >> 9,
179 j->buf[1].buf_size >> 9);
180 unsigned i, nr_online = 0, nr_devs_want;
181 bool can_discard = false;
184 lockdep_assert_held(&j->lock);
187 for_each_member_device_rcu(ca, c, i,
188 &c->rw_devs[BCH_DATA_journal]) {
189 struct journal_device *ja = &ca->journal;
194 while (ja->dirty_idx != ja->cur_idx &&
195 ja->bucket_seq[ja->dirty_idx] < journal_last_seq(j))
196 ja->dirty_idx = (ja->dirty_idx + 1) % ja->nr;
198 while (ja->dirty_idx_ondisk != ja->dirty_idx &&
199 ja->bucket_seq[ja->dirty_idx_ondisk] < j->last_seq_ondisk)
200 ja->dirty_idx_ondisk = (ja->dirty_idx_ondisk + 1) % ja->nr;
202 if (ja->discard_idx != ja->dirty_idx_ondisk)
205 max_entry_size = min_t(unsigned, max_entry_size, ca->mi.bucket_size);
210 j->can_discard = can_discard;
212 if (nr_online < c->opts.metadata_replicas_required) {
213 ret = JOURNAL_ERR_insufficient_devices;
217 nr_devs_want = min_t(unsigned, nr_online, c->opts.metadata_replicas);
219 for (i = 0; i < journal_space_nr; i++)
220 j->space[i] = __journal_space_available(j, nr_devs_want, i);
222 clean_ondisk = j->space[journal_space_clean_ondisk].total;
223 clean = j->space[journal_space_clean].total;
224 total = j->space[journal_space_total].total;
226 if (!j->space[journal_space_discarded].next_entry)
227 ret = JOURNAL_ERR_journal_full;
229 if ((j->space[journal_space_clean_ondisk].next_entry <
230 j->space[journal_space_clean_ondisk].total) &&
231 (clean - clean_ondisk <= total / 8) &&
232 (clean_ondisk * 2 > clean))
233 set_bit(JOURNAL_MAY_SKIP_FLUSH, &j->flags);
235 clear_bit(JOURNAL_MAY_SKIP_FLUSH, &j->flags);
237 bch2_journal_set_watermark(j);
239 j->cur_entry_sectors = !ret ? j->space[journal_space_discarded].next_entry : 0;
240 j->cur_entry_error = ret;
246 /* Discards - last part of journal reclaim: */
248 static bool should_discard_bucket(struct journal *j, struct journal_device *ja)
253 ret = ja->discard_idx != ja->dirty_idx_ondisk;
254 spin_unlock(&j->lock);
260 * Advance ja->discard_idx as long as it points to buckets that are no longer
261 * dirty, issuing discards if necessary:
263 void bch2_journal_do_discards(struct journal *j)
265 struct bch_fs *c = container_of(j, struct bch_fs, journal);
269 mutex_lock(&j->discard_lock);
271 for_each_rw_member(ca, c, iter) {
272 struct journal_device *ja = &ca->journal;
274 while (should_discard_bucket(j, ja)) {
275 if (!c->opts.nochanges &&
277 bdev_max_discard_sectors(ca->disk_sb.bdev))
278 blkdev_issue_discard(ca->disk_sb.bdev,
280 ja->buckets[ja->discard_idx]),
281 ca->mi.bucket_size, GFP_NOFS);
284 ja->discard_idx = (ja->discard_idx + 1) % ja->nr;
286 bch2_journal_space_available(j);
287 spin_unlock(&j->lock);
291 mutex_unlock(&j->discard_lock);
295 * Journal entry pinning - machinery for holding a reference on a given journal
296 * entry, holding it open to ensure it gets replayed during recovery:
299 void bch2_journal_reclaim_fast(struct journal *j)
303 lockdep_assert_held(&j->lock);
306 * Unpin journal entries whose reference counts reached zero, meaning
307 * all btree nodes got written out
309 while (!fifo_empty(&j->pin) &&
310 j->pin.front <= j->seq_ondisk &&
311 !atomic_read(&fifo_peek_front(&j->pin).count)) {
317 bch2_journal_space_available(j);
320 bool __bch2_journal_pin_put(struct journal *j, u64 seq)
322 struct journal_entry_pin_list *pin_list = journal_seq_pin(j, seq);
324 return atomic_dec_and_test(&pin_list->count);
327 void bch2_journal_pin_put(struct journal *j, u64 seq)
329 if (__bch2_journal_pin_put(j, seq)) {
331 bch2_journal_reclaim_fast(j);
332 spin_unlock(&j->lock);
336 static inline bool __journal_pin_drop(struct journal *j,
337 struct journal_entry_pin *pin)
339 struct journal_entry_pin_list *pin_list;
341 if (!journal_pin_active(pin))
344 if (j->flush_in_progress == pin)
345 j->flush_in_progress_dropped = true;
347 pin_list = journal_seq_pin(j, pin->seq);
349 list_del_init(&pin->list);
352 * Unpinning a journal entry may make journal_next_bucket() succeed, if
353 * writing a new last_seq will now make another bucket available:
355 return atomic_dec_and_test(&pin_list->count) &&
356 pin_list == &fifo_peek_front(&j->pin);
359 void bch2_journal_pin_drop(struct journal *j,
360 struct journal_entry_pin *pin)
363 if (__journal_pin_drop(j, pin))
364 bch2_journal_reclaim_fast(j);
365 spin_unlock(&j->lock);
368 static enum journal_pin_type journal_pin_type(journal_pin_flush_fn fn)
370 if (fn == bch2_btree_node_flush0 ||
371 fn == bch2_btree_node_flush1)
372 return JOURNAL_PIN_btree;
373 else if (fn == bch2_btree_key_cache_journal_flush)
374 return JOURNAL_PIN_key_cache;
376 return JOURNAL_PIN_other;
379 static inline void bch2_journal_pin_set_locked(struct journal *j, u64 seq,
380 struct journal_entry_pin *pin,
381 journal_pin_flush_fn flush_fn,
382 enum journal_pin_type type)
384 struct journal_entry_pin_list *pin_list = journal_seq_pin(j, seq);
387 * flush_fn is how we identify journal pins in debugfs, so must always
388 * exist, even if it doesn't do anything:
392 atomic_inc(&pin_list->count);
394 pin->flush = flush_fn;
395 list_add(&pin->list, &pin_list->list[type]);
398 void bch2_journal_pin_copy(struct journal *j,
399 struct journal_entry_pin *dst,
400 struct journal_entry_pin *src,
401 journal_pin_flush_fn flush_fn)
407 u64 seq = READ_ONCE(src->seq);
409 if (seq < journal_last_seq(j)) {
411 * bch2_journal_pin_copy() raced with bch2_journal_pin_drop() on
412 * the src pin - with the pin dropped, the entry to pin might no
413 * longer to exist, but that means there's no longer anything to
414 * copy and we can bail out here:
416 spin_unlock(&j->lock);
420 reclaim = __journal_pin_drop(j, dst);
422 bch2_journal_pin_set_locked(j, seq, dst, flush_fn, journal_pin_type(flush_fn));
425 bch2_journal_reclaim_fast(j);
426 spin_unlock(&j->lock);
429 * If the journal is currently full, we might want to call flush_fn
435 void bch2_journal_pin_set(struct journal *j, u64 seq,
436 struct journal_entry_pin *pin,
437 journal_pin_flush_fn flush_fn)
443 BUG_ON(seq < journal_last_seq(j));
445 reclaim = __journal_pin_drop(j, pin);
447 bch2_journal_pin_set_locked(j, seq, pin, flush_fn, journal_pin_type(flush_fn));
450 bch2_journal_reclaim_fast(j);
451 spin_unlock(&j->lock);
454 * If the journal is currently full, we might want to call flush_fn
461 * bch2_journal_pin_flush: ensure journal pin callback is no longer running
465 void bch2_journal_pin_flush(struct journal *j, struct journal_entry_pin *pin)
467 BUG_ON(journal_pin_active(pin));
469 wait_event(j->pin_flush_wait, j->flush_in_progress != pin);
473 * Journal reclaim: flush references to open journal entries to reclaim space in
476 * May be done by the journal code in the background as needed to free up space
477 * for more journal entries, or as part of doing a clean shutdown, or to migrate
478 * data off of a specific device:
481 static struct journal_entry_pin *
482 journal_get_next_pin(struct journal *j,
484 unsigned allowed_below_seq,
485 unsigned allowed_above_seq,
488 struct journal_entry_pin_list *pin_list;
489 struct journal_entry_pin *ret = NULL;
492 fifo_for_each_entry_ptr(pin_list, &j->pin, *seq) {
493 if (*seq > seq_to_flush && !allowed_above_seq)
496 for (i = 0; i < JOURNAL_PIN_NR; i++)
497 if ((((1U << i) & allowed_below_seq) && *seq <= seq_to_flush) ||
498 ((1U << i) & allowed_above_seq)) {
499 ret = list_first_entry_or_null(&pin_list->list[i],
500 struct journal_entry_pin, list);
509 /* returns true if we did work */
510 static size_t journal_flush_pins(struct journal *j,
512 unsigned allowed_below_seq,
513 unsigned allowed_above_seq,
515 unsigned min_key_cache)
517 struct journal_entry_pin *pin;
518 size_t nr_flushed = 0;
519 journal_pin_flush_fn flush_fn;
523 lockdep_assert_held(&j->reclaim_lock);
526 unsigned allowed_above = allowed_above_seq;
527 unsigned allowed_below = allowed_below_seq;
535 allowed_above |= 1U << JOURNAL_PIN_key_cache;
536 allowed_below |= 1U << JOURNAL_PIN_key_cache;
541 j->last_flushed = jiffies;
544 pin = journal_get_next_pin(j, seq_to_flush, allowed_below, allowed_above, &seq);
546 BUG_ON(j->flush_in_progress);
547 j->flush_in_progress = pin;
548 j->flush_in_progress_dropped = false;
549 flush_fn = pin->flush;
551 spin_unlock(&j->lock);
556 if (min_key_cache && pin->flush == bch2_btree_key_cache_journal_flush)
562 err = flush_fn(j, pin, seq);
565 /* Pin might have been dropped or rearmed: */
566 if (likely(!err && !j->flush_in_progress_dropped))
567 list_move(&pin->list, &journal_seq_pin(j, seq)->flushed);
568 j->flush_in_progress = NULL;
569 j->flush_in_progress_dropped = false;
570 spin_unlock(&j->lock);
572 wake_up(&j->pin_flush_wait);
583 static u64 journal_seq_to_flush(struct journal *j)
585 struct bch_fs *c = container_of(j, struct bch_fs, journal);
587 u64 seq_to_flush = 0;
592 for_each_rw_member(ca, c, iter) {
593 struct journal_device *ja = &ca->journal;
594 unsigned nr_buckets, bucket_to_flush;
599 /* Try to keep the journal at most half full: */
600 nr_buckets = ja->nr / 2;
602 nr_buckets = min(nr_buckets, ja->nr);
604 bucket_to_flush = (ja->cur_idx + nr_buckets) % ja->nr;
605 seq_to_flush = max(seq_to_flush,
606 ja->bucket_seq[bucket_to_flush]);
609 /* Also flush if the pin fifo is more than half full */
610 seq_to_flush = max_t(s64, seq_to_flush,
611 (s64) journal_cur_seq(j) -
613 spin_unlock(&j->lock);
619 * __bch2_journal_reclaim - free up journal buckets
621 * @direct: direct or background reclaim?
622 * @kicked: requested to run since we last ran?
623 * Returns: 0 on success, or -EIO if the journal has been shutdown
625 * Background journal reclaim writes out btree nodes. It should be run
626 * early enough so that we never completely run out of journal buckets.
628 * High watermarks for triggering background reclaim:
629 * - FIFO has fewer than 512 entries left
630 * - fewer than 25% journal buckets free
632 * Background reclaim runs until low watermarks are reached:
633 * - FIFO has more than 1024 entries left
634 * - more than 50% journal buckets free
636 * As long as a reclaim can complete in the time it takes to fill up
637 * 512 journal entries or 25% of all journal buckets, then
638 * journal_next_bucket() should not stall.
640 static int __bch2_journal_reclaim(struct journal *j, bool direct, bool kicked)
642 struct bch_fs *c = container_of(j, struct bch_fs, journal);
643 bool kthread = (current->flags & PF_KTHREAD) != 0;
645 size_t min_nr, min_key_cache, nr_flushed;
650 * We can't invoke memory reclaim while holding the reclaim_lock -
651 * journal reclaim is required to make progress for memory reclaim
652 * (cleaning the caches), so we can't get stuck in memory reclaim while
653 * we're holding the reclaim lock:
655 lockdep_assert_held(&j->reclaim_lock);
656 flags = memalloc_noreclaim_save();
659 if (kthread && kthread_should_stop())
662 if (bch2_journal_error(j)) {
667 bch2_journal_do_discards(j);
669 seq_to_flush = journal_seq_to_flush(j);
673 * If it's been longer than j->reclaim_delay_ms since we last flushed,
674 * make sure to flush at least one journal pin:
676 if (time_after(jiffies, j->last_flushed +
677 msecs_to_jiffies(c->opts.journal_reclaim_delay)))
680 if (j->watermark != BCH_WATERMARK_stripe)
683 if (atomic_read(&c->btree_cache.dirty) * 2 > c->btree_cache.used)
686 min_key_cache = min(bch2_nr_btree_keys_need_flush(c), (size_t) 128);
688 trace_and_count(c, journal_reclaim_start, c,
690 min_nr, min_key_cache,
691 atomic_read(&c->btree_cache.dirty),
693 atomic_long_read(&c->btree_key_cache.nr_dirty),
694 atomic_long_read(&c->btree_key_cache.nr_keys));
696 nr_flushed = journal_flush_pins(j, seq_to_flush,
698 min_nr, min_key_cache);
701 j->nr_direct_reclaim += nr_flushed;
703 j->nr_background_reclaim += nr_flushed;
704 trace_and_count(c, journal_reclaim_finish, c, nr_flushed);
707 wake_up(&j->reclaim_wait);
708 } while ((min_nr || min_key_cache) && nr_flushed && !direct);
710 memalloc_noreclaim_restore(flags);
715 int bch2_journal_reclaim(struct journal *j)
717 return __bch2_journal_reclaim(j, true, true);
720 static int bch2_journal_reclaim_thread(void *arg)
722 struct journal *j = arg;
723 struct bch_fs *c = container_of(j, struct bch_fs, journal);
724 unsigned long delay, now;
730 j->last_flushed = jiffies;
732 while (!ret && !kthread_should_stop()) {
733 bool kicked = j->reclaim_kicked;
735 j->reclaim_kicked = false;
737 mutex_lock(&j->reclaim_lock);
738 ret = __bch2_journal_reclaim(j, false, kicked);
739 mutex_unlock(&j->reclaim_lock);
742 delay = msecs_to_jiffies(c->opts.journal_reclaim_delay);
743 j->next_reclaim = j->last_flushed + delay;
745 if (!time_in_range(j->next_reclaim, now, now + delay))
746 j->next_reclaim = now + delay;
749 set_current_state(TASK_INTERRUPTIBLE|TASK_FREEZABLE);
750 if (kthread_should_stop())
752 if (j->reclaim_kicked)
756 journal_empty = fifo_empty(&j->pin);
757 spin_unlock(&j->lock);
761 else if (time_after(j->next_reclaim, jiffies))
762 schedule_timeout(j->next_reclaim - jiffies);
766 __set_current_state(TASK_RUNNING);
772 void bch2_journal_reclaim_stop(struct journal *j)
774 struct task_struct *p = j->reclaim_thread;
776 j->reclaim_thread = NULL;
784 int bch2_journal_reclaim_start(struct journal *j)
786 struct bch_fs *c = container_of(j, struct bch_fs, journal);
787 struct task_struct *p;
790 if (j->reclaim_thread)
793 p = kthread_create(bch2_journal_reclaim_thread, j,
794 "bch-reclaim/%s", c->name);
795 ret = PTR_ERR_OR_ZERO(p);
797 bch_err_msg(c, ret, "creating journal reclaim thread");
802 j->reclaim_thread = p;
807 static int journal_flush_done(struct journal *j, u64 seq_to_flush,
812 ret = bch2_journal_error(j);
816 mutex_lock(&j->reclaim_lock);
818 if (journal_flush_pins(j, seq_to_flush,
819 (1U << JOURNAL_PIN_key_cache)|
820 (1U << JOURNAL_PIN_other), 0, 0, 0) ||
821 journal_flush_pins(j, seq_to_flush,
822 (1U << JOURNAL_PIN_btree), 0, 0, 0))
825 if (seq_to_flush > journal_cur_seq(j))
826 bch2_journal_entry_close(j);
830 * If journal replay hasn't completed, the unreplayed journal entries
831 * hold refs on their corresponding sequence numbers
833 ret = !test_bit(JOURNAL_REPLAY_DONE, &j->flags) ||
834 journal_last_seq(j) > seq_to_flush ||
837 spin_unlock(&j->lock);
838 mutex_unlock(&j->reclaim_lock);
843 bool bch2_journal_flush_pins(struct journal *j, u64 seq_to_flush)
845 /* time_stats this */
846 bool did_work = false;
848 if (!test_bit(JOURNAL_STARTED, &j->flags))
851 closure_wait_event(&j->async_wait,
852 journal_flush_done(j, seq_to_flush, &did_work));
857 int bch2_journal_flush_device_pins(struct journal *j, int dev_idx)
859 struct bch_fs *c = container_of(j, struct bch_fs, journal);
860 struct journal_entry_pin_list *p;
865 fifo_for_each_entry_ptr(p, &j->pin, iter)
867 ? bch2_dev_list_has_dev(p->devs, dev_idx)
868 : p->devs.nr < c->opts.metadata_replicas)
870 spin_unlock(&j->lock);
872 bch2_journal_flush_pins(j, seq);
874 ret = bch2_journal_error(j);
878 mutex_lock(&c->replicas_gc_lock);
879 bch2_replicas_gc_start(c, 1 << BCH_DATA_journal);
882 * Now that we've populated replicas_gc, write to the journal to mark
883 * active journal devices. This handles the case where the journal might
884 * be empty. Otherwise we could clear all journal replicas and
885 * temporarily put the fs into an unrecoverable state. Journal recovery
886 * expects to find devices marked for journal data on unclean mount.
888 ret = bch2_journal_meta(&c->journal);
895 struct bch_replicas_padded replicas;
897 seq = max(seq, journal_last_seq(j));
898 if (seq >= j->pin.back)
900 bch2_devlist_to_replicas(&replicas.e, BCH_DATA_journal,
901 journal_seq_pin(j, seq)->devs);
904 spin_unlock(&j->lock);
905 ret = bch2_mark_replicas(c, &replicas.e);
908 spin_unlock(&j->lock);
910 ret = bch2_replicas_gc_end(c, ret);
911 mutex_unlock(&c->replicas_gc_lock);