1 // SPDX-License-Identifier: GPL-2.0
4 #include "btree_key_cache.h"
7 #include "journal_io.h"
8 #include "journal_reclaim.h"
12 #include <linux/kthread.h>
13 #include <linux/sched/mm.h>
14 #include <trace/events/bcachefs.h>
16 /* Free space calculations: */
18 static unsigned journal_space_from(struct journal_device *ja,
19 enum journal_space_from from)
22 case journal_space_discarded:
23 return ja->discard_idx;
24 case journal_space_clean_ondisk:
25 return ja->dirty_idx_ondisk;
26 case journal_space_clean:
33 unsigned bch2_journal_dev_buckets_available(struct journal *j,
34 struct journal_device *ja,
35 enum journal_space_from from)
37 unsigned available = (journal_space_from(ja, from) -
38 ja->cur_idx - 1 + ja->nr) % ja->nr;
41 * Don't use the last bucket unless writing the new last_seq
42 * will make another bucket available:
44 if (available && ja->dirty_idx_ondisk == ja->dirty_idx)
50 static void journal_set_remaining(struct journal *j, unsigned u64s_remaining)
52 union journal_preres_state old, new;
53 u64 v = atomic64_read(&j->prereserved.counter);
57 new.remaining = u64s_remaining;
58 } while ((v = atomic64_cmpxchg(&j->prereserved.counter,
59 old.v, new.v)) != old.v);
62 static inline unsigned get_unwritten_sectors(struct journal *j, unsigned *idx)
66 while (!sectors && *idx != j->reservations.idx) {
67 sectors = j->buf[*idx].sectors;
69 *idx = (*idx + 1) & JOURNAL_BUF_MASK;
75 static struct journal_space
76 journal_dev_space_available(struct journal *j, struct bch_dev *ca,
77 enum journal_space_from from)
79 struct journal_device *ja = &ca->journal;
80 unsigned sectors, buckets, unwritten, idx = j->reservations.unwritten_idx;
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,
88 buckets = bch2_journal_dev_buckets_available(j, ja, from);
89 sectors = ja->sectors_free;
92 * We that we don't allocate the space for a journal entry
93 * until we write it out - thus, account for it here:
95 while ((unwritten = get_unwritten_sectors(j, &idx))) {
96 if (unwritten >= sectors) {
103 sectors = ca->mi.bucket_size;
106 sectors -= unwritten;
109 if (sectors < ca->mi.bucket_size && buckets) {
111 sectors = ca->mi.bucket_size;
114 return (struct journal_space) {
115 .next_entry = sectors,
116 .total = sectors + buckets * ca->mi.bucket_size,
120 static struct journal_space __journal_space_available(struct journal *j, unsigned nr_devs_want,
121 enum journal_space_from from)
123 struct bch_fs *c = container_of(j, struct bch_fs, journal);
125 unsigned i, pos, nr_devs = 0;
126 struct journal_space space, dev_space[BCH_SB_MEMBERS_MAX];
128 BUG_ON(nr_devs_want > ARRAY_SIZE(dev_space));
131 for_each_member_device_rcu(ca, c, i,
132 &c->rw_devs[BCH_DATA_journal]) {
136 space = journal_dev_space_available(j, ca, from);
137 if (!space.next_entry)
140 for (pos = 0; pos < nr_devs; pos++)
141 if (space.total > dev_space[pos].total)
144 array_insert_item(dev_space, nr_devs, pos, space);
148 if (nr_devs < nr_devs_want)
149 return (struct journal_space) { 0, 0 };
152 * We sorted largest to smallest, and we want the smallest out of the
153 * @nr_devs_want largest devices:
155 return dev_space[nr_devs_want - 1];
158 void bch2_journal_space_available(struct journal *j)
160 struct bch_fs *c = container_of(j, struct bch_fs, journal);
162 unsigned clean, clean_ondisk, total;
163 s64 u64s_remaining = 0;
164 unsigned max_entry_size = min(j->buf[0].buf_size >> 9,
165 j->buf[1].buf_size >> 9);
166 unsigned i, nr_online = 0, nr_devs_want;
167 bool can_discard = false;
170 lockdep_assert_held(&j->lock);
173 for_each_member_device_rcu(ca, c, i,
174 &c->rw_devs[BCH_DATA_journal]) {
175 struct journal_device *ja = &ca->journal;
180 while (ja->dirty_idx != ja->cur_idx &&
181 ja->bucket_seq[ja->dirty_idx] < journal_last_seq(j))
182 ja->dirty_idx = (ja->dirty_idx + 1) % ja->nr;
184 while (ja->dirty_idx_ondisk != ja->dirty_idx &&
185 ja->bucket_seq[ja->dirty_idx_ondisk] < j->last_seq_ondisk)
186 ja->dirty_idx_ondisk = (ja->dirty_idx_ondisk + 1) % ja->nr;
188 if (ja->discard_idx != ja->dirty_idx_ondisk)
191 max_entry_size = min_t(unsigned, max_entry_size, ca->mi.bucket_size);
196 j->can_discard = can_discard;
198 if (nr_online < c->opts.metadata_replicas_required) {
199 ret = cur_entry_insufficient_devices;
203 nr_devs_want = min_t(unsigned, nr_online, c->opts.metadata_replicas);
205 for (i = 0; i < journal_space_nr; i++)
206 j->space[i] = __journal_space_available(j, nr_devs_want, i);
208 clean_ondisk = j->space[journal_space_clean_ondisk].total;
209 clean = j->space[journal_space_clean].total;
210 total = j->space[journal_space_total].total;
213 j->reservations.idx ==
214 j->reservations.unwritten_idx) {
215 char *buf = kmalloc(4096, GFP_ATOMIC);
217 bch_err(c, "journal stuck");
219 __bch2_journal_debug_to_text(&_PBUF(buf, 4096), j);
225 ret = cur_entry_journal_stuck;
226 } else if (!j->space[journal_space_discarded].next_entry)
227 ret = cur_entry_journal_full;
228 else if (!fifo_free(&j->pin))
229 ret = cur_entry_journal_pin_full;
231 if ((j->space[journal_space_clean_ondisk].next_entry <
232 j->space[journal_space_clean_ondisk].total) &&
233 (clean - clean_ondisk <= total / 8) &&
234 (clean_ondisk * 2 > clean ))
235 set_bit(JOURNAL_MAY_SKIP_FLUSH, &j->flags);
237 clear_bit(JOURNAL_MAY_SKIP_FLUSH, &j->flags);
239 u64s_remaining = (u64) clean << 6;
240 u64s_remaining -= (u64) total << 3;
241 u64s_remaining = max(0LL, u64s_remaining);
243 u64s_remaining = min_t(u64, u64s_remaining, U32_MAX);
245 j->cur_entry_sectors = !ret ? j->space[journal_space_discarded].next_entry : 0;
246 j->cur_entry_error = ret;
247 journal_set_remaining(j, u64s_remaining);
248 journal_check_may_get_unreserved(j);
254 /* Discards - last part of journal reclaim: */
256 static bool should_discard_bucket(struct journal *j, struct journal_device *ja)
261 ret = ja->discard_idx != ja->dirty_idx_ondisk;
262 spin_unlock(&j->lock);
268 * Advance ja->discard_idx as long as it points to buckets that are no longer
269 * dirty, issuing discards if necessary:
271 void bch2_journal_do_discards(struct journal *j)
273 struct bch_fs *c = container_of(j, struct bch_fs, journal);
277 mutex_lock(&j->discard_lock);
279 for_each_rw_member(ca, c, iter) {
280 struct journal_device *ja = &ca->journal;
282 while (should_discard_bucket(j, ja)) {
283 if (ca->mi.discard &&
284 blk_queue_discard(bdev_get_queue(ca->disk_sb.bdev)))
285 blkdev_issue_discard(ca->disk_sb.bdev,
287 ja->buckets[ja->discard_idx]),
288 ca->mi.bucket_size, GFP_NOIO, 0);
291 ja->discard_idx = (ja->discard_idx + 1) % ja->nr;
293 bch2_journal_space_available(j);
294 spin_unlock(&j->lock);
298 mutex_unlock(&j->discard_lock);
302 * Journal entry pinning - machinery for holding a reference on a given journal
303 * entry, holding it open to ensure it gets replayed during recovery:
306 static void bch2_journal_reclaim_fast(struct journal *j)
308 struct journal_entry_pin_list temp;
311 lockdep_assert_held(&j->lock);
314 * Unpin journal entries whose reference counts reached zero, meaning
315 * all btree nodes got written out
317 while (!fifo_empty(&j->pin) &&
318 !atomic_read(&fifo_peek_front(&j->pin).count)) {
319 BUG_ON(!list_empty(&fifo_peek_front(&j->pin).list));
320 BUG_ON(!list_empty(&fifo_peek_front(&j->pin).flushed));
321 BUG_ON(!fifo_pop(&j->pin, temp));
326 bch2_journal_space_available(j);
329 void __bch2_journal_pin_put(struct journal *j, u64 seq)
331 struct journal_entry_pin_list *pin_list = journal_seq_pin(j, seq);
333 if (atomic_dec_and_test(&pin_list->count))
334 bch2_journal_reclaim_fast(j);
337 void bch2_journal_pin_put(struct journal *j, u64 seq)
339 struct journal_entry_pin_list *pin_list = journal_seq_pin(j, seq);
341 if (atomic_dec_and_test(&pin_list->count)) {
343 bch2_journal_reclaim_fast(j);
344 spin_unlock(&j->lock);
348 static inline void __journal_pin_drop(struct journal *j,
349 struct journal_entry_pin *pin)
351 struct journal_entry_pin_list *pin_list;
353 if (!journal_pin_active(pin))
356 pin_list = journal_seq_pin(j, pin->seq);
358 list_del_init(&pin->list);
361 * Unpinning a journal entry make make journal_next_bucket() succeed, if
362 * writing a new last_seq will now make another bucket available:
364 if (atomic_dec_and_test(&pin_list->count) &&
365 pin_list == &fifo_peek_front(&j->pin))
366 bch2_journal_reclaim_fast(j);
367 else if (fifo_used(&j->pin) == 1 &&
368 atomic_read(&pin_list->count) == 1)
372 void bch2_journal_pin_drop(struct journal *j,
373 struct journal_entry_pin *pin)
376 __journal_pin_drop(j, pin);
377 spin_unlock(&j->lock);
380 void bch2_journal_pin_set(struct journal *j, u64 seq,
381 struct journal_entry_pin *pin,
382 journal_pin_flush_fn flush_fn)
384 struct journal_entry_pin_list *pin_list;
388 if (seq < journal_last_seq(j)) {
390 * bch2_journal_pin_copy() raced with bch2_journal_pin_drop() on
391 * the src pin - with the pin dropped, the entry to pin might no
392 * longer to exist, but that means there's no longer anything to
393 * copy and we can bail out here:
395 spin_unlock(&j->lock);
399 pin_list = journal_seq_pin(j, seq);
401 __journal_pin_drop(j, pin);
403 atomic_inc(&pin_list->count);
405 pin->flush = flush_fn;
407 list_add(&pin->list, flush_fn ? &pin_list->list : &pin_list->flushed);
408 spin_unlock(&j->lock);
411 * If the journal is currently full, we might want to call flush_fn
418 * bch2_journal_pin_flush: ensure journal pin callback is no longer running
420 void bch2_journal_pin_flush(struct journal *j, struct journal_entry_pin *pin)
422 BUG_ON(journal_pin_active(pin));
424 wait_event(j->pin_flush_wait, j->flush_in_progress != pin);
428 * Journal reclaim: flush references to open journal entries to reclaim space in
431 * May be done by the journal code in the background as needed to free up space
432 * for more journal entries, or as part of doing a clean shutdown, or to migrate
433 * data off of a specific device:
436 static struct journal_entry_pin *
437 journal_get_next_pin(struct journal *j, u64 max_seq, u64 *seq)
439 struct journal_entry_pin_list *pin_list;
440 struct journal_entry_pin *ret = NULL;
442 if (!test_bit(JOURNAL_RECLAIM_STARTED, &j->flags))
447 fifo_for_each_entry_ptr(pin_list, &j->pin, *seq)
448 if (*seq > max_seq ||
449 (ret = list_first_entry_or_null(&pin_list->list,
450 struct journal_entry_pin, list)))
454 list_move(&ret->list, &pin_list->flushed);
455 BUG_ON(j->flush_in_progress);
456 j->flush_in_progress = ret;
459 spin_unlock(&j->lock);
464 /* returns true if we did work */
465 static u64 journal_flush_pins(struct journal *j, u64 seq_to_flush,
468 struct journal_entry_pin *pin;
471 lockdep_assert_held(&j->reclaim_lock);
476 j->last_flushed = jiffies;
478 pin = journal_get_next_pin(j, min_nr
479 ? U64_MAX : seq_to_flush, &seq);
486 pin->flush(j, pin, seq);
488 BUG_ON(j->flush_in_progress != pin);
489 j->flush_in_progress = NULL;
490 wake_up(&j->pin_flush_wait);
497 static u64 journal_seq_to_flush(struct journal *j)
499 struct bch_fs *c = container_of(j, struct bch_fs, journal);
501 u64 seq_to_flush = 0;
506 for_each_rw_member(ca, c, iter) {
507 struct journal_device *ja = &ca->journal;
508 unsigned nr_buckets, bucket_to_flush;
513 /* Try to keep the journal at most half full: */
514 nr_buckets = ja->nr / 2;
516 /* And include pre-reservations: */
517 nr_buckets += DIV_ROUND_UP(j->prereserved.reserved,
518 (ca->mi.bucket_size << 6) -
519 journal_entry_overhead(j));
521 nr_buckets = min(nr_buckets, ja->nr);
523 bucket_to_flush = (ja->cur_idx + nr_buckets) % ja->nr;
524 seq_to_flush = max(seq_to_flush,
525 ja->bucket_seq[bucket_to_flush]);
528 /* Also flush if the pin fifo is more than half full */
529 seq_to_flush = max_t(s64, seq_to_flush,
530 (s64) journal_cur_seq(j) -
532 spin_unlock(&j->lock);
538 * bch2_journal_reclaim - free up journal buckets
540 * Background journal reclaim writes out btree nodes. It should be run
541 * early enough so that we never completely run out of journal buckets.
543 * High watermarks for triggering background reclaim:
544 * - FIFO has fewer than 512 entries left
545 * - fewer than 25% journal buckets free
547 * Background reclaim runs until low watermarks are reached:
548 * - FIFO has more than 1024 entries left
549 * - more than 50% journal buckets free
551 * As long as a reclaim can complete in the time it takes to fill up
552 * 512 journal entries or 25% of all journal buckets, then
553 * journal_next_bucket() should not stall.
555 static int __bch2_journal_reclaim(struct journal *j, bool direct)
557 struct bch_fs *c = container_of(j, struct bch_fs, journal);
558 bool kthread = (current->flags & PF_KTHREAD) != 0;
559 u64 seq_to_flush, nr_flushed = 0;
565 * We can't invoke memory reclaim while holding the reclaim_lock -
566 * journal reclaim is required to make progress for memory reclaim
567 * (cleaning the caches), so we can't get stuck in memory reclaim while
568 * we're holding the reclaim lock:
570 lockdep_assert_held(&j->reclaim_lock);
571 flags = memalloc_noreclaim_save();
574 if (kthread && kthread_should_stop())
577 if (bch2_journal_error(j)) {
582 bch2_journal_do_discards(j);
584 seq_to_flush = journal_seq_to_flush(j);
588 * If it's been longer than j->reclaim_delay_ms since we last flushed,
589 * make sure to flush at least one journal pin:
591 if (time_after(jiffies, j->last_flushed +
592 msecs_to_jiffies(j->reclaim_delay_ms)))
595 if (j->prereserved.reserved * 2 > j->prereserved.remaining)
598 if (atomic_read(&c->btree_cache.dirty) * 4 >
599 c->btree_cache.used * 3)
602 if (fifo_free(&j->pin) <= 32)
605 min_nr = max(min_nr, bch2_nr_btree_keys_need_flush(c));
607 trace_journal_reclaim_start(c,
609 j->prereserved.reserved,
610 j->prereserved.remaining,
611 atomic_read(&c->btree_cache.dirty),
613 atomic_long_read(&c->btree_key_cache.nr_dirty),
614 atomic_long_read(&c->btree_key_cache.nr_keys));
616 nr_flushed = journal_flush_pins(j, seq_to_flush, min_nr);
619 j->nr_direct_reclaim += nr_flushed;
621 j->nr_background_reclaim += nr_flushed;
622 trace_journal_reclaim_finish(c, nr_flushed);
623 } while (min_nr && nr_flushed);
625 memalloc_noreclaim_restore(flags);
630 int bch2_journal_reclaim(struct journal *j)
632 return __bch2_journal_reclaim(j, true);
635 static int bch2_journal_reclaim_thread(void *arg)
637 struct journal *j = arg;
643 kthread_wait_freezable(test_bit(JOURNAL_RECLAIM_STARTED, &j->flags));
645 while (!ret && !kthread_should_stop()) {
646 j->reclaim_kicked = false;
648 mutex_lock(&j->reclaim_lock);
649 ret = __bch2_journal_reclaim(j, false);
650 mutex_unlock(&j->reclaim_lock);
652 next = j->last_flushed + msecs_to_jiffies(j->reclaim_delay_ms);
655 set_current_state(TASK_INTERRUPTIBLE);
656 if (kthread_should_stop())
658 if (j->reclaim_kicked)
660 if (time_after_eq(jiffies, next))
662 schedule_timeout(next - jiffies);
666 __set_current_state(TASK_RUNNING);
672 void bch2_journal_reclaim_stop(struct journal *j)
674 struct task_struct *p = j->reclaim_thread;
676 j->reclaim_thread = NULL;
684 int bch2_journal_reclaim_start(struct journal *j)
686 struct bch_fs *c = container_of(j, struct bch_fs, journal);
687 struct task_struct *p;
689 if (j->reclaim_thread)
692 p = kthread_create(bch2_journal_reclaim_thread, j,
693 "bch-reclaim/%s", c->name);
695 bch_err(c, "error creating journal reclaim thread: %li", PTR_ERR(p));
700 j->reclaim_thread = p;
705 static int journal_flush_done(struct journal *j, u64 seq_to_flush,
710 ret = bch2_journal_error(j);
714 mutex_lock(&j->reclaim_lock);
716 *did_work = journal_flush_pins(j, seq_to_flush, 0) != 0;
720 * If journal replay hasn't completed, the unreplayed journal entries
721 * hold refs on their corresponding sequence numbers
723 ret = !test_bit(JOURNAL_REPLAY_DONE, &j->flags) ||
724 journal_last_seq(j) > seq_to_flush ||
725 (fifo_used(&j->pin) == 1 &&
726 atomic_read(&fifo_peek_front(&j->pin).count) == 1);
728 spin_unlock(&j->lock);
729 mutex_unlock(&j->reclaim_lock);
734 bool bch2_journal_flush_pins(struct journal *j, u64 seq_to_flush)
736 bool did_work = false;
738 if (!test_bit(JOURNAL_STARTED, &j->flags))
741 closure_wait_event(&j->async_wait,
742 journal_flush_done(j, seq_to_flush, &did_work));
747 int bch2_journal_flush_device_pins(struct journal *j, int dev_idx)
749 struct bch_fs *c = container_of(j, struct bch_fs, journal);
750 struct journal_entry_pin_list *p;
755 fifo_for_each_entry_ptr(p, &j->pin, iter)
757 ? bch2_dev_list_has_dev(p->devs, dev_idx)
758 : p->devs.nr < c->opts.metadata_replicas)
760 spin_unlock(&j->lock);
762 bch2_journal_flush_pins(j, seq);
764 ret = bch2_journal_error(j);
768 mutex_lock(&c->replicas_gc_lock);
769 bch2_replicas_gc_start(c, 1 << BCH_DATA_journal);
774 while (!ret && seq < j->pin.back) {
775 struct bch_replicas_padded replicas;
777 seq = max(seq, journal_last_seq(j));
778 bch2_devlist_to_replicas(&replicas.e, BCH_DATA_journal,
779 journal_seq_pin(j, seq)->devs);
782 spin_unlock(&j->lock);
783 ret = bch2_mark_replicas(c, &replicas.e);
786 spin_unlock(&j->lock);
788 ret = bch2_replicas_gc_end(c, ret);
789 mutex_unlock(&c->replicas_gc_lock);