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);
139 unsigned pos, nr_devs = 0;
140 struct journal_space space, dev_space[BCH_SB_MEMBERS_MAX];
142 BUG_ON(nr_devs_want > ARRAY_SIZE(dev_space));
145 for_each_member_device_rcu(c, ca, &c->rw_devs[BCH_DATA_journal]) {
149 space = journal_dev_space_available(j, ca, from);
150 if (!space.next_entry)
153 for (pos = 0; pos < nr_devs; pos++)
154 if (space.total > dev_space[pos].total)
157 array_insert_item(dev_space, nr_devs, pos, space);
161 if (nr_devs < nr_devs_want)
162 return (struct journal_space) { 0, 0 };
165 * We sorted largest to smallest, and we want the smallest out of the
166 * @nr_devs_want largest devices:
168 return dev_space[nr_devs_want - 1];
171 void bch2_journal_space_available(struct journal *j)
173 struct bch_fs *c = container_of(j, struct bch_fs, journal);
174 unsigned clean, clean_ondisk, total;
175 unsigned max_entry_size = min(j->buf[0].buf_size >> 9,
176 j->buf[1].buf_size >> 9);
177 unsigned nr_online = 0, nr_devs_want;
178 bool can_discard = false;
181 lockdep_assert_held(&j->lock);
184 for_each_member_device_rcu(c, ca, &c->rw_devs[BCH_DATA_journal]) {
185 struct journal_device *ja = &ca->journal;
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;
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;
198 if (ja->discard_idx != ja->dirty_idx_ondisk)
201 max_entry_size = min_t(unsigned, max_entry_size, ca->mi.bucket_size);
206 j->can_discard = can_discard;
208 if (nr_online < metadata_replicas_required(c)) {
209 ret = JOURNAL_ERR_insufficient_devices;
213 nr_devs_want = min_t(unsigned, nr_online, c->opts.metadata_replicas);
215 for (unsigned i = 0; i < journal_space_nr; i++)
216 j->space[i] = __journal_space_available(j, nr_devs_want, i);
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;
222 if (!j->space[journal_space_discarded].next_entry)
223 ret = JOURNAL_ERR_journal_full;
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);
231 clear_bit(JOURNAL_MAY_SKIP_FLUSH, &j->flags);
233 bch2_journal_set_watermark(j);
235 j->cur_entry_sectors = !ret ? j->space[journal_space_discarded].next_entry : 0;
236 j->cur_entry_error = ret;
242 /* Discards - last part of journal reclaim: */
244 static bool should_discard_bucket(struct journal *j, struct journal_device *ja)
249 ret = ja->discard_idx != ja->dirty_idx_ondisk;
250 spin_unlock(&j->lock);
256 * Advance ja->discard_idx as long as it points to buckets that are no longer
257 * dirty, issuing discards if necessary:
259 void bch2_journal_do_discards(struct journal *j)
261 struct bch_fs *c = container_of(j, struct bch_fs, journal);
263 mutex_lock(&j->discard_lock);
265 for_each_rw_member(c, ca) {
266 struct journal_device *ja = &ca->journal;
268 while (should_discard_bucket(j, ja)) {
269 if (!c->opts.nochanges &&
271 bdev_max_discard_sectors(ca->disk_sb.bdev))
272 blkdev_issue_discard(ca->disk_sb.bdev,
274 ja->buckets[ja->discard_idx]),
275 ca->mi.bucket_size, GFP_NOFS);
278 ja->discard_idx = (ja->discard_idx + 1) % ja->nr;
280 bch2_journal_space_available(j);
281 spin_unlock(&j->lock);
285 mutex_unlock(&j->discard_lock);
289 * Journal entry pinning - machinery for holding a reference on a given journal
290 * entry, holding it open to ensure it gets replayed during recovery:
293 void bch2_journal_reclaim_fast(struct journal *j)
297 lockdep_assert_held(&j->lock);
300 * Unpin journal entries whose reference counts reached zero, meaning
301 * all btree nodes got written out
303 while (!fifo_empty(&j->pin) &&
304 j->pin.front <= j->seq_ondisk &&
305 !atomic_read(&fifo_peek_front(&j->pin).count)) {
311 bch2_journal_space_available(j);
314 bool __bch2_journal_pin_put(struct journal *j, u64 seq)
316 struct journal_entry_pin_list *pin_list = journal_seq_pin(j, seq);
318 return atomic_dec_and_test(&pin_list->count);
321 void bch2_journal_pin_put(struct journal *j, u64 seq)
323 if (__bch2_journal_pin_put(j, seq)) {
325 bch2_journal_reclaim_fast(j);
326 spin_unlock(&j->lock);
330 static inline bool __journal_pin_drop(struct journal *j,
331 struct journal_entry_pin *pin)
333 struct journal_entry_pin_list *pin_list;
335 if (!journal_pin_active(pin))
338 if (j->flush_in_progress == pin)
339 j->flush_in_progress_dropped = true;
341 pin_list = journal_seq_pin(j, pin->seq);
343 list_del_init(&pin->list);
346 * Unpinning a journal entry may make journal_next_bucket() succeed, if
347 * writing a new last_seq will now make another bucket available:
349 return atomic_dec_and_test(&pin_list->count) &&
350 pin_list == &fifo_peek_front(&j->pin);
353 void bch2_journal_pin_drop(struct journal *j,
354 struct journal_entry_pin *pin)
357 if (__journal_pin_drop(j, pin))
358 bch2_journal_reclaim_fast(j);
359 spin_unlock(&j->lock);
362 static enum journal_pin_type journal_pin_type(journal_pin_flush_fn fn)
364 if (fn == bch2_btree_node_flush0 ||
365 fn == bch2_btree_node_flush1)
366 return JOURNAL_PIN_btree;
367 else if (fn == bch2_btree_key_cache_journal_flush)
368 return JOURNAL_PIN_key_cache;
370 return JOURNAL_PIN_other;
373 static inline void bch2_journal_pin_set_locked(struct journal *j, u64 seq,
374 struct journal_entry_pin *pin,
375 journal_pin_flush_fn flush_fn,
376 enum journal_pin_type type)
378 struct journal_entry_pin_list *pin_list = journal_seq_pin(j, seq);
381 * flush_fn is how we identify journal pins in debugfs, so must always
382 * exist, even if it doesn't do anything:
386 atomic_inc(&pin_list->count);
388 pin->flush = flush_fn;
389 list_add(&pin->list, &pin_list->list[type]);
392 void bch2_journal_pin_copy(struct journal *j,
393 struct journal_entry_pin *dst,
394 struct journal_entry_pin *src,
395 journal_pin_flush_fn flush_fn)
401 u64 seq = READ_ONCE(src->seq);
403 if (seq < journal_last_seq(j)) {
405 * bch2_journal_pin_copy() raced with bch2_journal_pin_drop() on
406 * the src pin - with the pin dropped, the entry to pin might no
407 * longer to exist, but that means there's no longer anything to
408 * copy and we can bail out here:
410 spin_unlock(&j->lock);
414 reclaim = __journal_pin_drop(j, dst);
416 bch2_journal_pin_set_locked(j, seq, dst, flush_fn, journal_pin_type(flush_fn));
419 bch2_journal_reclaim_fast(j);
420 spin_unlock(&j->lock);
423 * If the journal is currently full, we might want to call flush_fn
429 void bch2_journal_pin_set(struct journal *j, u64 seq,
430 struct journal_entry_pin *pin,
431 journal_pin_flush_fn flush_fn)
437 BUG_ON(seq < journal_last_seq(j));
439 reclaim = __journal_pin_drop(j, pin);
441 bch2_journal_pin_set_locked(j, seq, pin, flush_fn, journal_pin_type(flush_fn));
444 bch2_journal_reclaim_fast(j);
445 spin_unlock(&j->lock);
448 * If the journal is currently full, we might want to call flush_fn
455 * bch2_journal_pin_flush: ensure journal pin callback is no longer running
459 void bch2_journal_pin_flush(struct journal *j, struct journal_entry_pin *pin)
461 BUG_ON(journal_pin_active(pin));
463 wait_event(j->pin_flush_wait, j->flush_in_progress != pin);
467 * Journal reclaim: flush references to open journal entries to reclaim space in
470 * May be done by the journal code in the background as needed to free up space
471 * for more journal entries, or as part of doing a clean shutdown, or to migrate
472 * data off of a specific device:
475 static struct journal_entry_pin *
476 journal_get_next_pin(struct journal *j,
478 unsigned allowed_below_seq,
479 unsigned allowed_above_seq,
482 struct journal_entry_pin_list *pin_list;
483 struct journal_entry_pin *ret = NULL;
486 fifo_for_each_entry_ptr(pin_list, &j->pin, *seq) {
487 if (*seq > seq_to_flush && !allowed_above_seq)
490 for (i = 0; i < JOURNAL_PIN_NR; i++)
491 if ((((1U << i) & allowed_below_seq) && *seq <= seq_to_flush) ||
492 ((1U << i) & allowed_above_seq)) {
493 ret = list_first_entry_or_null(&pin_list->list[i],
494 struct journal_entry_pin, list);
503 /* returns true if we did work */
504 static size_t journal_flush_pins(struct journal *j,
506 unsigned allowed_below_seq,
507 unsigned allowed_above_seq,
509 unsigned min_key_cache)
511 struct journal_entry_pin *pin;
512 size_t nr_flushed = 0;
513 journal_pin_flush_fn flush_fn;
517 lockdep_assert_held(&j->reclaim_lock);
520 unsigned allowed_above = allowed_above_seq;
521 unsigned allowed_below = allowed_below_seq;
529 allowed_above |= 1U << JOURNAL_PIN_key_cache;
530 allowed_below |= 1U << JOURNAL_PIN_key_cache;
535 j->last_flushed = jiffies;
538 pin = journal_get_next_pin(j, seq_to_flush, allowed_below, allowed_above, &seq);
540 BUG_ON(j->flush_in_progress);
541 j->flush_in_progress = pin;
542 j->flush_in_progress_dropped = false;
543 flush_fn = pin->flush;
545 spin_unlock(&j->lock);
550 if (min_key_cache && pin->flush == bch2_btree_key_cache_journal_flush)
556 err = flush_fn(j, pin, seq);
559 /* Pin might have been dropped or rearmed: */
560 if (likely(!err && !j->flush_in_progress_dropped))
561 list_move(&pin->list, &journal_seq_pin(j, seq)->flushed);
562 j->flush_in_progress = NULL;
563 j->flush_in_progress_dropped = false;
564 spin_unlock(&j->lock);
566 wake_up(&j->pin_flush_wait);
577 static u64 journal_seq_to_flush(struct journal *j)
579 struct bch_fs *c = container_of(j, struct bch_fs, journal);
580 u64 seq_to_flush = 0;
584 for_each_rw_member(c, ca) {
585 struct journal_device *ja = &ca->journal;
586 unsigned nr_buckets, bucket_to_flush;
591 /* Try to keep the journal at most half full: */
592 nr_buckets = ja->nr / 2;
594 nr_buckets = min(nr_buckets, ja->nr);
596 bucket_to_flush = (ja->cur_idx + nr_buckets) % ja->nr;
597 seq_to_flush = max(seq_to_flush,
598 ja->bucket_seq[bucket_to_flush]);
601 /* Also flush if the pin fifo is more than half full */
602 seq_to_flush = max_t(s64, seq_to_flush,
603 (s64) journal_cur_seq(j) -
605 spin_unlock(&j->lock);
611 * __bch2_journal_reclaim - free up journal buckets
613 * @direct: direct or background reclaim?
614 * @kicked: requested to run since we last ran?
615 * Returns: 0 on success, or -EIO if the journal has been shutdown
617 * Background journal reclaim writes out btree nodes. It should be run
618 * early enough so that we never completely run out of journal buckets.
620 * High watermarks for triggering background reclaim:
621 * - FIFO has fewer than 512 entries left
622 * - fewer than 25% journal buckets free
624 * Background reclaim runs until low watermarks are reached:
625 * - FIFO has more than 1024 entries left
626 * - more than 50% journal buckets free
628 * As long as a reclaim can complete in the time it takes to fill up
629 * 512 journal entries or 25% of all journal buckets, then
630 * journal_next_bucket() should not stall.
632 static int __bch2_journal_reclaim(struct journal *j, bool direct, bool kicked)
634 struct bch_fs *c = container_of(j, struct bch_fs, journal);
635 bool kthread = (current->flags & PF_KTHREAD) != 0;
637 size_t min_nr, min_key_cache, nr_flushed;
642 * We can't invoke memory reclaim while holding the reclaim_lock -
643 * journal reclaim is required to make progress for memory reclaim
644 * (cleaning the caches), so we can't get stuck in memory reclaim while
645 * we're holding the reclaim lock:
647 lockdep_assert_held(&j->reclaim_lock);
648 flags = memalloc_noreclaim_save();
651 if (kthread && kthread_should_stop())
654 if (bch2_journal_error(j)) {
659 bch2_journal_do_discards(j);
661 seq_to_flush = journal_seq_to_flush(j);
665 * If it's been longer than j->reclaim_delay_ms since we last flushed,
666 * make sure to flush at least one journal pin:
668 if (time_after(jiffies, j->last_flushed +
669 msecs_to_jiffies(c->opts.journal_reclaim_delay)))
672 if (j->watermark != BCH_WATERMARK_stripe)
675 if (atomic_read(&c->btree_cache.dirty) * 2 > c->btree_cache.used)
678 min_key_cache = min(bch2_nr_btree_keys_need_flush(c), (size_t) 128);
680 trace_and_count(c, journal_reclaim_start, c,
682 min_nr, min_key_cache,
683 atomic_read(&c->btree_cache.dirty),
685 atomic_long_read(&c->btree_key_cache.nr_dirty),
686 atomic_long_read(&c->btree_key_cache.nr_keys));
688 nr_flushed = journal_flush_pins(j, seq_to_flush,
690 min_nr, min_key_cache);
693 j->nr_direct_reclaim += nr_flushed;
695 j->nr_background_reclaim += nr_flushed;
696 trace_and_count(c, journal_reclaim_finish, c, nr_flushed);
699 wake_up(&j->reclaim_wait);
700 } while ((min_nr || min_key_cache) && nr_flushed && !direct);
702 memalloc_noreclaim_restore(flags);
707 int bch2_journal_reclaim(struct journal *j)
709 return __bch2_journal_reclaim(j, true, true);
712 static int bch2_journal_reclaim_thread(void *arg)
714 struct journal *j = arg;
715 struct bch_fs *c = container_of(j, struct bch_fs, journal);
716 unsigned long delay, now;
722 j->last_flushed = jiffies;
724 while (!ret && !kthread_should_stop()) {
725 bool kicked = j->reclaim_kicked;
727 j->reclaim_kicked = false;
729 mutex_lock(&j->reclaim_lock);
730 ret = __bch2_journal_reclaim(j, false, kicked);
731 mutex_unlock(&j->reclaim_lock);
734 delay = msecs_to_jiffies(c->opts.journal_reclaim_delay);
735 j->next_reclaim = j->last_flushed + delay;
737 if (!time_in_range(j->next_reclaim, now, now + delay))
738 j->next_reclaim = now + delay;
741 set_current_state(TASK_INTERRUPTIBLE|TASK_FREEZABLE);
742 if (kthread_should_stop())
744 if (j->reclaim_kicked)
748 journal_empty = fifo_empty(&j->pin);
749 spin_unlock(&j->lock);
753 else if (time_after(j->next_reclaim, jiffies))
754 schedule_timeout(j->next_reclaim - jiffies);
758 __set_current_state(TASK_RUNNING);
764 void bch2_journal_reclaim_stop(struct journal *j)
766 struct task_struct *p = j->reclaim_thread;
768 j->reclaim_thread = NULL;
776 int bch2_journal_reclaim_start(struct journal *j)
778 struct bch_fs *c = container_of(j, struct bch_fs, journal);
779 struct task_struct *p;
782 if (j->reclaim_thread)
785 p = kthread_create(bch2_journal_reclaim_thread, j,
786 "bch-reclaim/%s", c->name);
787 ret = PTR_ERR_OR_ZERO(p);
788 bch_err_msg(c, ret, "creating journal reclaim thread");
793 j->reclaim_thread = p;
798 static int journal_flush_done(struct journal *j, u64 seq_to_flush,
803 ret = bch2_journal_error(j);
807 mutex_lock(&j->reclaim_lock);
809 if (journal_flush_pins(j, seq_to_flush,
810 (1U << JOURNAL_PIN_key_cache)|
811 (1U << JOURNAL_PIN_other), 0, 0, 0) ||
812 journal_flush_pins(j, seq_to_flush,
813 (1U << JOURNAL_PIN_btree), 0, 0, 0))
816 if (seq_to_flush > journal_cur_seq(j))
817 bch2_journal_entry_close(j);
821 * If journal replay hasn't completed, the unreplayed journal entries
822 * hold refs on their corresponding sequence numbers
824 ret = !test_bit(JOURNAL_REPLAY_DONE, &j->flags) ||
825 journal_last_seq(j) > seq_to_flush ||
828 spin_unlock(&j->lock);
829 mutex_unlock(&j->reclaim_lock);
834 bool bch2_journal_flush_pins(struct journal *j, u64 seq_to_flush)
836 /* time_stats this */
837 bool did_work = false;
839 if (!test_bit(JOURNAL_STARTED, &j->flags))
842 closure_wait_event(&j->async_wait,
843 journal_flush_done(j, seq_to_flush, &did_work));
848 int bch2_journal_flush_device_pins(struct journal *j, int dev_idx)
850 struct bch_fs *c = container_of(j, struct bch_fs, journal);
851 struct journal_entry_pin_list *p;
856 fifo_for_each_entry_ptr(p, &j->pin, iter)
858 ? bch2_dev_list_has_dev(p->devs, dev_idx)
859 : p->devs.nr < c->opts.metadata_replicas)
861 spin_unlock(&j->lock);
863 bch2_journal_flush_pins(j, seq);
865 ret = bch2_journal_error(j);
869 mutex_lock(&c->replicas_gc_lock);
870 bch2_replicas_gc_start(c, 1 << BCH_DATA_journal);
873 * Now that we've populated replicas_gc, write to the journal to mark
874 * active journal devices. This handles the case where the journal might
875 * be empty. Otherwise we could clear all journal replicas and
876 * temporarily put the fs into an unrecoverable state. Journal recovery
877 * expects to find devices marked for journal data on unclean mount.
879 ret = bch2_journal_meta(&c->journal);
886 struct bch_replicas_padded replicas;
888 seq = max(seq, journal_last_seq(j));
889 if (seq >= j->pin.back)
891 bch2_devlist_to_replicas(&replicas.e, BCH_DATA_journal,
892 journal_seq_pin(j, seq)->devs);
895 if (replicas.e.nr_devs) {
896 spin_unlock(&j->lock);
897 ret = bch2_mark_replicas(c, &replicas.e);
901 spin_unlock(&j->lock);
903 ret = bch2_replicas_gc_end(c, ret);
904 mutex_unlock(&c->replicas_gc_lock);