xref: /linux/fs/bcachefs/journal_reclaim.c (revision 24168c5e6dfbdd5b414f048f47f75d64533296ca)
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 		ret = JOURNAL_ERR_insufficient_devices;
209 		goto out;
210 	}
211 
212 	nr_devs_want = min_t(unsigned, nr_online, c->opts.metadata_replicas);
213 
214 	for (unsigned i = 0; i < journal_space_nr; i++)
215 		j->space[i] = __journal_space_available(j, nr_devs_want, i);
216 
217 	clean_ondisk	= j->space[journal_space_clean_ondisk].total;
218 	clean		= j->space[journal_space_clean].total;
219 	total		= j->space[journal_space_total].total;
220 
221 	if (!j->space[journal_space_discarded].next_entry)
222 		ret = JOURNAL_ERR_journal_full;
223 
224 	if ((j->space[journal_space_clean_ondisk].next_entry <
225 	     j->space[journal_space_clean_ondisk].total) &&
226 	    (clean - clean_ondisk <= total / 8) &&
227 	    (clean_ondisk * 2 > clean))
228 		set_bit(JOURNAL_may_skip_flush, &j->flags);
229 	else
230 		clear_bit(JOURNAL_may_skip_flush, &j->flags);
231 
232 	bch2_journal_set_watermark(j);
233 out:
234 	j->cur_entry_sectors	= !ret ? j->space[journal_space_discarded].next_entry : 0;
235 	j->cur_entry_error	= ret;
236 
237 	if (!ret)
238 		journal_wake(j);
239 }
240 
241 /* Discards - last part of journal reclaim: */
242 
243 static bool should_discard_bucket(struct journal *j, struct journal_device *ja)
244 {
245 	bool ret;
246 
247 	spin_lock(&j->lock);
248 	ret = ja->discard_idx != ja->dirty_idx_ondisk;
249 	spin_unlock(&j->lock);
250 
251 	return ret;
252 }
253 
254 /*
255  * Advance ja->discard_idx as long as it points to buckets that are no longer
256  * dirty, issuing discards if necessary:
257  */
258 void bch2_journal_do_discards(struct journal *j)
259 {
260 	struct bch_fs *c = container_of(j, struct bch_fs, journal);
261 
262 	mutex_lock(&j->discard_lock);
263 
264 	for_each_rw_member(c, ca) {
265 		struct journal_device *ja = &ca->journal;
266 
267 		while (should_discard_bucket(j, ja)) {
268 			if (!c->opts.nochanges &&
269 			    ca->mi.discard &&
270 			    bdev_max_discard_sectors(ca->disk_sb.bdev))
271 				blkdev_issue_discard(ca->disk_sb.bdev,
272 					bucket_to_sector(ca,
273 						ja->buckets[ja->discard_idx]),
274 					ca->mi.bucket_size, GFP_NOFS);
275 
276 			spin_lock(&j->lock);
277 			ja->discard_idx = (ja->discard_idx + 1) % ja->nr;
278 
279 			bch2_journal_space_available(j);
280 			spin_unlock(&j->lock);
281 		}
282 	}
283 
284 	mutex_unlock(&j->discard_lock);
285 }
286 
287 /*
288  * Journal entry pinning - machinery for holding a reference on a given journal
289  * entry, holding it open to ensure it gets replayed during recovery:
290  */
291 
292 void bch2_journal_reclaim_fast(struct journal *j)
293 {
294 	bool popped = false;
295 
296 	lockdep_assert_held(&j->lock);
297 
298 	/*
299 	 * Unpin journal entries whose reference counts reached zero, meaning
300 	 * all btree nodes got written out
301 	 */
302 	while (!fifo_empty(&j->pin) &&
303 	       j->pin.front <= j->seq_ondisk &&
304 	       !atomic_read(&fifo_peek_front(&j->pin).count)) {
305 		j->pin.front++;
306 		popped = true;
307 	}
308 
309 	if (popped)
310 		bch2_journal_space_available(j);
311 }
312 
313 bool __bch2_journal_pin_put(struct journal *j, u64 seq)
314 {
315 	struct journal_entry_pin_list *pin_list = journal_seq_pin(j, seq);
316 
317 	return atomic_dec_and_test(&pin_list->count);
318 }
319 
320 void bch2_journal_pin_put(struct journal *j, u64 seq)
321 {
322 	if (__bch2_journal_pin_put(j, seq)) {
323 		spin_lock(&j->lock);
324 		bch2_journal_reclaim_fast(j);
325 		spin_unlock(&j->lock);
326 	}
327 }
328 
329 static inline bool __journal_pin_drop(struct journal *j,
330 				      struct journal_entry_pin *pin)
331 {
332 	struct journal_entry_pin_list *pin_list;
333 
334 	if (!journal_pin_active(pin))
335 		return false;
336 
337 	if (j->flush_in_progress == pin)
338 		j->flush_in_progress_dropped = true;
339 
340 	pin_list = journal_seq_pin(j, pin->seq);
341 	pin->seq = 0;
342 	list_del_init(&pin->list);
343 
344 	/*
345 	 * Unpinning a journal entry may make journal_next_bucket() succeed, if
346 	 * writing a new last_seq will now make another bucket available:
347 	 */
348 	return atomic_dec_and_test(&pin_list->count) &&
349 		pin_list == &fifo_peek_front(&j->pin);
350 }
351 
352 void bch2_journal_pin_drop(struct journal *j,
353 			   struct journal_entry_pin *pin)
354 {
355 	spin_lock(&j->lock);
356 	if (__journal_pin_drop(j, pin))
357 		bch2_journal_reclaim_fast(j);
358 	spin_unlock(&j->lock);
359 }
360 
361 static enum journal_pin_type journal_pin_type(journal_pin_flush_fn fn)
362 {
363 	if (fn == bch2_btree_node_flush0 ||
364 	    fn == bch2_btree_node_flush1)
365 		return JOURNAL_PIN_btree;
366 	else if (fn == bch2_btree_key_cache_journal_flush)
367 		return JOURNAL_PIN_key_cache;
368 	else
369 		return JOURNAL_PIN_other;
370 }
371 
372 static inline void bch2_journal_pin_set_locked(struct journal *j, u64 seq,
373 			  struct journal_entry_pin *pin,
374 			  journal_pin_flush_fn flush_fn,
375 			  enum journal_pin_type type)
376 {
377 	struct journal_entry_pin_list *pin_list = journal_seq_pin(j, seq);
378 
379 	/*
380 	 * flush_fn is how we identify journal pins in debugfs, so must always
381 	 * exist, even if it doesn't do anything:
382 	 */
383 	BUG_ON(!flush_fn);
384 
385 	atomic_inc(&pin_list->count);
386 	pin->seq	= seq;
387 	pin->flush	= flush_fn;
388 	list_add(&pin->list, &pin_list->list[type]);
389 }
390 
391 void bch2_journal_pin_copy(struct journal *j,
392 			   struct journal_entry_pin *dst,
393 			   struct journal_entry_pin *src,
394 			   journal_pin_flush_fn flush_fn)
395 {
396 	spin_lock(&j->lock);
397 
398 	u64 seq = READ_ONCE(src->seq);
399 
400 	if (seq < journal_last_seq(j)) {
401 		/*
402 		 * bch2_journal_pin_copy() raced with bch2_journal_pin_drop() on
403 		 * the src pin - with the pin dropped, the entry to pin might no
404 		 * longer to exist, but that means there's no longer anything to
405 		 * copy and we can bail out here:
406 		 */
407 		spin_unlock(&j->lock);
408 		return;
409 	}
410 
411 	bool reclaim = __journal_pin_drop(j, dst);
412 
413 	bch2_journal_pin_set_locked(j, seq, dst, flush_fn, journal_pin_type(flush_fn));
414 
415 	if (reclaim)
416 		bch2_journal_reclaim_fast(j);
417 
418 	/*
419 	 * If the journal is currently full,  we might want to call flush_fn
420 	 * immediately:
421 	 */
422 	if (seq == journal_last_seq(j))
423 		journal_wake(j);
424 	spin_unlock(&j->lock);
425 }
426 
427 void bch2_journal_pin_set(struct journal *j, u64 seq,
428 			  struct journal_entry_pin *pin,
429 			  journal_pin_flush_fn flush_fn)
430 {
431 	spin_lock(&j->lock);
432 
433 	BUG_ON(seq < journal_last_seq(j));
434 
435 	bool reclaim = __journal_pin_drop(j, pin);
436 
437 	bch2_journal_pin_set_locked(j, seq, pin, flush_fn, journal_pin_type(flush_fn));
438 
439 	if (reclaim)
440 		bch2_journal_reclaim_fast(j);
441 	/*
442 	 * If the journal is currently full,  we might want to call flush_fn
443 	 * immediately:
444 	 */
445 	if (seq == journal_last_seq(j))
446 		journal_wake(j);
447 
448 	spin_unlock(&j->lock);
449 }
450 
451 /**
452  * bch2_journal_pin_flush: ensure journal pin callback is no longer running
453  * @j:		journal object
454  * @pin:	pin to flush
455  */
456 void bch2_journal_pin_flush(struct journal *j, struct journal_entry_pin *pin)
457 {
458 	BUG_ON(journal_pin_active(pin));
459 
460 	wait_event(j->pin_flush_wait, j->flush_in_progress != pin);
461 }
462 
463 /*
464  * Journal reclaim: flush references to open journal entries to reclaim space in
465  * the journal
466  *
467  * May be done by the journal code in the background as needed to free up space
468  * for more journal entries, or as part of doing a clean shutdown, or to migrate
469  * data off of a specific device:
470  */
471 
472 static struct journal_entry_pin *
473 journal_get_next_pin(struct journal *j,
474 		     u64 seq_to_flush,
475 		     unsigned allowed_below_seq,
476 		     unsigned allowed_above_seq,
477 		     u64 *seq)
478 {
479 	struct journal_entry_pin_list *pin_list;
480 	struct journal_entry_pin *ret = NULL;
481 	unsigned i;
482 
483 	fifo_for_each_entry_ptr(pin_list, &j->pin, *seq) {
484 		if (*seq > seq_to_flush && !allowed_above_seq)
485 			break;
486 
487 		for (i = 0; i < JOURNAL_PIN_NR; i++)
488 			if ((((1U << i) & allowed_below_seq) && *seq <= seq_to_flush) ||
489 			    ((1U << i) & allowed_above_seq)) {
490 				ret = list_first_entry_or_null(&pin_list->list[i],
491 					struct journal_entry_pin, list);
492 				if (ret)
493 					return ret;
494 			}
495 	}
496 
497 	return NULL;
498 }
499 
500 /* returns true if we did work */
501 static size_t journal_flush_pins(struct journal *j,
502 				 u64 seq_to_flush,
503 				 unsigned allowed_below_seq,
504 				 unsigned allowed_above_seq,
505 				 unsigned min_any,
506 				 unsigned min_key_cache)
507 {
508 	struct journal_entry_pin *pin;
509 	size_t nr_flushed = 0;
510 	journal_pin_flush_fn flush_fn;
511 	u64 seq;
512 	int err;
513 
514 	lockdep_assert_held(&j->reclaim_lock);
515 
516 	while (1) {
517 		unsigned allowed_above = allowed_above_seq;
518 		unsigned allowed_below = allowed_below_seq;
519 
520 		if (min_any) {
521 			allowed_above |= ~0;
522 			allowed_below |= ~0;
523 		}
524 
525 		if (min_key_cache) {
526 			allowed_above |= 1U << JOURNAL_PIN_key_cache;
527 			allowed_below |= 1U << JOURNAL_PIN_key_cache;
528 		}
529 
530 		cond_resched();
531 
532 		j->last_flushed = jiffies;
533 
534 		spin_lock(&j->lock);
535 		pin = journal_get_next_pin(j, seq_to_flush, allowed_below, allowed_above, &seq);
536 		if (pin) {
537 			BUG_ON(j->flush_in_progress);
538 			j->flush_in_progress = pin;
539 			j->flush_in_progress_dropped = false;
540 			flush_fn = pin->flush;
541 		}
542 		spin_unlock(&j->lock);
543 
544 		if (!pin)
545 			break;
546 
547 		if (min_key_cache && pin->flush == bch2_btree_key_cache_journal_flush)
548 			min_key_cache--;
549 
550 		if (min_any)
551 			min_any--;
552 
553 		err = flush_fn(j, pin, seq);
554 
555 		spin_lock(&j->lock);
556 		/* Pin might have been dropped or rearmed: */
557 		if (likely(!err && !j->flush_in_progress_dropped))
558 			list_move(&pin->list, &journal_seq_pin(j, seq)->flushed);
559 		j->flush_in_progress = NULL;
560 		j->flush_in_progress_dropped = false;
561 		spin_unlock(&j->lock);
562 
563 		wake_up(&j->pin_flush_wait);
564 
565 		if (err)
566 			break;
567 
568 		nr_flushed++;
569 	}
570 
571 	return nr_flushed;
572 }
573 
574 static u64 journal_seq_to_flush(struct journal *j)
575 {
576 	struct bch_fs *c = container_of(j, struct bch_fs, journal);
577 	u64 seq_to_flush = 0;
578 
579 	spin_lock(&j->lock);
580 
581 	for_each_rw_member(c, ca) {
582 		struct journal_device *ja = &ca->journal;
583 		unsigned nr_buckets, bucket_to_flush;
584 
585 		if (!ja->nr)
586 			continue;
587 
588 		/* Try to keep the journal at most half full: */
589 		nr_buckets = ja->nr / 2;
590 
591 		nr_buckets = min(nr_buckets, ja->nr);
592 
593 		bucket_to_flush = (ja->cur_idx + nr_buckets) % ja->nr;
594 		seq_to_flush = max(seq_to_flush,
595 				   ja->bucket_seq[bucket_to_flush]);
596 	}
597 
598 	/* Also flush if the pin fifo is more than half full */
599 	seq_to_flush = max_t(s64, seq_to_flush,
600 			     (s64) journal_cur_seq(j) -
601 			     (j->pin.size >> 1));
602 	spin_unlock(&j->lock);
603 
604 	return seq_to_flush;
605 }
606 
607 /**
608  * __bch2_journal_reclaim - free up journal buckets
609  * @j:		journal object
610  * @direct:	direct or background reclaim?
611  * @kicked:	requested to run since we last ran?
612  * Returns:	0 on success, or -EIO if the journal has been shutdown
613  *
614  * Background journal reclaim writes out btree nodes. It should be run
615  * early enough so that we never completely run out of journal buckets.
616  *
617  * High watermarks for triggering background reclaim:
618  * - FIFO has fewer than 512 entries left
619  * - fewer than 25% journal buckets free
620  *
621  * Background reclaim runs until low watermarks are reached:
622  * - FIFO has more than 1024 entries left
623  * - more than 50% journal buckets free
624  *
625  * As long as a reclaim can complete in the time it takes to fill up
626  * 512 journal entries or 25% of all journal buckets, then
627  * journal_next_bucket() should not stall.
628  */
629 static int __bch2_journal_reclaim(struct journal *j, bool direct, bool kicked)
630 {
631 	struct bch_fs *c = container_of(j, struct bch_fs, journal);
632 	bool kthread = (current->flags & PF_KTHREAD) != 0;
633 	u64 seq_to_flush;
634 	size_t min_nr, min_key_cache, nr_flushed;
635 	unsigned flags;
636 	int ret = 0;
637 
638 	/*
639 	 * We can't invoke memory reclaim while holding the reclaim_lock -
640 	 * journal reclaim is required to make progress for memory reclaim
641 	 * (cleaning the caches), so we can't get stuck in memory reclaim while
642 	 * we're holding the reclaim lock:
643 	 */
644 	lockdep_assert_held(&j->reclaim_lock);
645 	flags = memalloc_noreclaim_save();
646 
647 	do {
648 		if (kthread && kthread_should_stop())
649 			break;
650 
651 		if (bch2_journal_error(j)) {
652 			ret = -EIO;
653 			break;
654 		}
655 
656 		bch2_journal_do_discards(j);
657 
658 		seq_to_flush = journal_seq_to_flush(j);
659 		min_nr = 0;
660 
661 		/*
662 		 * If it's been longer than j->reclaim_delay_ms since we last flushed,
663 		 * make sure to flush at least one journal pin:
664 		 */
665 		if (time_after(jiffies, j->last_flushed +
666 			       msecs_to_jiffies(c->opts.journal_reclaim_delay)))
667 			min_nr = 1;
668 
669 		if (j->watermark != BCH_WATERMARK_stripe)
670 			min_nr = 1;
671 
672 		if (atomic_read(&c->btree_cache.dirty) * 2 > c->btree_cache.used)
673 			min_nr = 1;
674 
675 		min_key_cache = min(bch2_nr_btree_keys_need_flush(c), (size_t) 128);
676 
677 		trace_and_count(c, journal_reclaim_start, c,
678 				direct, kicked,
679 				min_nr, min_key_cache,
680 				atomic_read(&c->btree_cache.dirty),
681 				c->btree_cache.used,
682 				atomic_long_read(&c->btree_key_cache.nr_dirty),
683 				atomic_long_read(&c->btree_key_cache.nr_keys));
684 
685 		nr_flushed = journal_flush_pins(j, seq_to_flush,
686 						~0, 0,
687 						min_nr, min_key_cache);
688 
689 		if (direct)
690 			j->nr_direct_reclaim += nr_flushed;
691 		else
692 			j->nr_background_reclaim += nr_flushed;
693 		trace_and_count(c, journal_reclaim_finish, c, nr_flushed);
694 
695 		if (nr_flushed)
696 			wake_up(&j->reclaim_wait);
697 	} while ((min_nr || min_key_cache) && nr_flushed && !direct);
698 
699 	memalloc_noreclaim_restore(flags);
700 
701 	return ret;
702 }
703 
704 int bch2_journal_reclaim(struct journal *j)
705 {
706 	return __bch2_journal_reclaim(j, true, true);
707 }
708 
709 static int bch2_journal_reclaim_thread(void *arg)
710 {
711 	struct journal *j = arg;
712 	struct bch_fs *c = container_of(j, struct bch_fs, journal);
713 	unsigned long delay, now;
714 	bool journal_empty;
715 	int ret = 0;
716 
717 	set_freezable();
718 
719 	j->last_flushed = jiffies;
720 
721 	while (!ret && !kthread_should_stop()) {
722 		bool kicked = j->reclaim_kicked;
723 
724 		j->reclaim_kicked = false;
725 
726 		mutex_lock(&j->reclaim_lock);
727 		ret = __bch2_journal_reclaim(j, false, kicked);
728 		mutex_unlock(&j->reclaim_lock);
729 
730 		now = jiffies;
731 		delay = msecs_to_jiffies(c->opts.journal_reclaim_delay);
732 		j->next_reclaim = j->last_flushed + delay;
733 
734 		if (!time_in_range(j->next_reclaim, now, now + delay))
735 			j->next_reclaim = now + delay;
736 
737 		while (1) {
738 			set_current_state(TASK_INTERRUPTIBLE|TASK_FREEZABLE);
739 			if (kthread_should_stop())
740 				break;
741 			if (j->reclaim_kicked)
742 				break;
743 
744 			spin_lock(&j->lock);
745 			journal_empty = fifo_empty(&j->pin);
746 			spin_unlock(&j->lock);
747 
748 			if (journal_empty)
749 				schedule();
750 			else if (time_after(j->next_reclaim, jiffies))
751 				schedule_timeout(j->next_reclaim - jiffies);
752 			else
753 				break;
754 		}
755 		__set_current_state(TASK_RUNNING);
756 	}
757 
758 	return 0;
759 }
760 
761 void bch2_journal_reclaim_stop(struct journal *j)
762 {
763 	struct task_struct *p = j->reclaim_thread;
764 
765 	j->reclaim_thread = NULL;
766 
767 	if (p) {
768 		kthread_stop(p);
769 		put_task_struct(p);
770 	}
771 }
772 
773 int bch2_journal_reclaim_start(struct journal *j)
774 {
775 	struct bch_fs *c = container_of(j, struct bch_fs, journal);
776 	struct task_struct *p;
777 	int ret;
778 
779 	if (j->reclaim_thread)
780 		return 0;
781 
782 	p = kthread_create(bch2_journal_reclaim_thread, j,
783 			   "bch-reclaim/%s", c->name);
784 	ret = PTR_ERR_OR_ZERO(p);
785 	bch_err_msg(c, ret, "creating journal reclaim thread");
786 	if (ret)
787 		return ret;
788 
789 	get_task_struct(p);
790 	j->reclaim_thread = p;
791 	wake_up_process(p);
792 	return 0;
793 }
794 
795 static int journal_flush_done(struct journal *j, u64 seq_to_flush,
796 			      bool *did_work)
797 {
798 	int ret;
799 
800 	ret = bch2_journal_error(j);
801 	if (ret)
802 		return ret;
803 
804 	mutex_lock(&j->reclaim_lock);
805 
806 	if (journal_flush_pins(j, seq_to_flush,
807 			       (1U << JOURNAL_PIN_key_cache)|
808 			       (1U << JOURNAL_PIN_other), 0, 0, 0) ||
809 	    journal_flush_pins(j, seq_to_flush,
810 			       (1U << JOURNAL_PIN_btree), 0, 0, 0))
811 		*did_work = true;
812 
813 	if (seq_to_flush > journal_cur_seq(j))
814 		bch2_journal_entry_close(j);
815 
816 	spin_lock(&j->lock);
817 	/*
818 	 * If journal replay hasn't completed, the unreplayed journal entries
819 	 * hold refs on their corresponding sequence numbers
820 	 */
821 	ret = !test_bit(JOURNAL_replay_done, &j->flags) ||
822 		journal_last_seq(j) > seq_to_flush ||
823 		!fifo_used(&j->pin);
824 
825 	spin_unlock(&j->lock);
826 	mutex_unlock(&j->reclaim_lock);
827 
828 	return ret;
829 }
830 
831 bool bch2_journal_flush_pins(struct journal *j, u64 seq_to_flush)
832 {
833 	/* time_stats this */
834 	bool did_work = false;
835 
836 	if (!test_bit(JOURNAL_running, &j->flags))
837 		return false;
838 
839 	closure_wait_event(&j->async_wait,
840 		journal_flush_done(j, seq_to_flush, &did_work));
841 
842 	return did_work;
843 }
844 
845 int bch2_journal_flush_device_pins(struct journal *j, int dev_idx)
846 {
847 	struct bch_fs *c = container_of(j, struct bch_fs, journal);
848 	struct journal_entry_pin_list *p;
849 	u64 iter, seq = 0;
850 	int ret = 0;
851 
852 	spin_lock(&j->lock);
853 	fifo_for_each_entry_ptr(p, &j->pin, iter)
854 		if (dev_idx >= 0
855 		    ? bch2_dev_list_has_dev(p->devs, dev_idx)
856 		    : p->devs.nr < c->opts.metadata_replicas)
857 			seq = iter;
858 	spin_unlock(&j->lock);
859 
860 	bch2_journal_flush_pins(j, seq);
861 
862 	ret = bch2_journal_error(j);
863 	if (ret)
864 		return ret;
865 
866 	mutex_lock(&c->replicas_gc_lock);
867 	bch2_replicas_gc_start(c, 1 << BCH_DATA_journal);
868 
869 	/*
870 	 * Now that we've populated replicas_gc, write to the journal to mark
871 	 * active journal devices. This handles the case where the journal might
872 	 * be empty. Otherwise we could clear all journal replicas and
873 	 * temporarily put the fs into an unrecoverable state. Journal recovery
874 	 * expects to find devices marked for journal data on unclean mount.
875 	 */
876 	ret = bch2_journal_meta(&c->journal);
877 	if (ret)
878 		goto err;
879 
880 	seq = 0;
881 	spin_lock(&j->lock);
882 	while (!ret) {
883 		struct bch_replicas_padded replicas;
884 
885 		seq = max(seq, journal_last_seq(j));
886 		if (seq >= j->pin.back)
887 			break;
888 		bch2_devlist_to_replicas(&replicas.e, BCH_DATA_journal,
889 					 journal_seq_pin(j, seq)->devs);
890 		seq++;
891 
892 		if (replicas.e.nr_devs) {
893 			spin_unlock(&j->lock);
894 			ret = bch2_mark_replicas(c, &replicas.e);
895 			spin_lock(&j->lock);
896 		}
897 	}
898 	spin_unlock(&j->lock);
899 err:
900 	ret = bch2_replicas_gc_end(c, ret);
901 	mutex_unlock(&c->replicas_gc_lock);
902 
903 	return ret;
904 }
905