1 // SPDX-License-Identifier: GPL-2.0
3 * background writeback - scan btree for dirty data and write it to the backing
6 * Copyright 2010, 2011 Kent Overstreet <kent.overstreet@gmail.com>
7 * Copyright 2012 Google, Inc.
13 #include "writeback.h"
15 #include <linux/delay.h>
16 #include <linux/kthread.h>
17 #include <linux/sched/clock.h>
18 #include <trace/events/bcache.h>
20 static void update_gc_after_writeback(struct cache_set *c)
22 if (c->gc_after_writeback != (BCH_ENABLE_AUTO_GC) ||
23 c->gc_stats.in_use < BCH_AUTO_GC_DIRTY_THRESHOLD)
26 c->gc_after_writeback |= BCH_DO_AUTO_GC;
30 static uint64_t __calc_target_rate(struct cached_dev *dc)
32 struct cache_set *c = dc->disk.c;
35 * This is the size of the cache, minus the amount used for
38 uint64_t cache_sectors = c->nbuckets * c->cache->sb.bucket_size -
39 atomic_long_read(&c->flash_dev_dirty_sectors);
42 * Unfortunately there is no control of global dirty data. If the
43 * user states that they want 10% dirty data in the cache, and has,
44 * e.g., 5 backing volumes of equal size, we try and ensure each
45 * backing volume uses about 2% of the cache for dirty data.
48 div64_u64(bdev_sectors(dc->bdev) << WRITEBACK_SHARE_SHIFT,
49 c->cached_dev_sectors);
51 uint64_t cache_dirty_target =
52 div_u64(cache_sectors * dc->writeback_percent, 100);
54 /* Ensure each backing dev gets at least one dirty share */
58 return (cache_dirty_target * bdev_share) >> WRITEBACK_SHARE_SHIFT;
61 static void __update_writeback_rate(struct cached_dev *dc)
65 * Figures out the amount that should be written per second.
67 * First, the error (number of sectors that are dirty beyond our
68 * target) is calculated. The error is accumulated (numerically
71 * Then, the proportional value and integral value are scaled
72 * based on configured values. These are stored as inverses to
73 * avoid fixed point math and to make configuration easy-- e.g.
74 * the default value of 40 for writeback_rate_p_term_inverse
75 * attempts to write at a rate that would retire all the dirty
76 * blocks in 40 seconds.
78 * The writeback_rate_i_inverse value of 10000 means that 1/10000th
79 * of the error is accumulated in the integral term per second.
80 * This acts as a slow, long-term average that is not subject to
81 * variations in usage like the p term.
83 int64_t target = __calc_target_rate(dc);
84 int64_t dirty = bcache_dev_sectors_dirty(&dc->disk);
85 int64_t error = dirty - target;
86 int64_t proportional_scaled =
87 div_s64(error, dc->writeback_rate_p_term_inverse);
88 int64_t integral_scaled;
92 * We need to consider the number of dirty buckets as well
93 * when calculating the proportional_scaled, Otherwise we might
94 * have an unreasonable small writeback rate at a highly fragmented situation
95 * when very few dirty sectors consumed a lot dirty buckets, the
96 * worst case is when dirty buckets reached cutoff_writeback_sync and
97 * dirty data is still not even reached to writeback percent, so the rate
98 * still will be at the minimum value, which will cause the write
99 * stuck at a non-writeback mode.
101 struct cache_set *c = dc->disk.c;
103 int64_t dirty_buckets = c->nbuckets - c->avail_nbuckets;
105 if (dc->writeback_consider_fragment &&
106 c->gc_stats.in_use > BCH_WRITEBACK_FRAGMENT_THRESHOLD_LOW && dirty > 0) {
108 div_s64((dirty_buckets * c->cache->sb.bucket_size), dirty);
112 if (c->gc_stats.in_use <= BCH_WRITEBACK_FRAGMENT_THRESHOLD_MID) {
113 fp_term = (int64_t)dc->writeback_rate_fp_term_low *
114 (c->gc_stats.in_use - BCH_WRITEBACK_FRAGMENT_THRESHOLD_LOW);
115 } else if (c->gc_stats.in_use <= BCH_WRITEBACK_FRAGMENT_THRESHOLD_HIGH) {
116 fp_term = (int64_t)dc->writeback_rate_fp_term_mid *
117 (c->gc_stats.in_use - BCH_WRITEBACK_FRAGMENT_THRESHOLD_MID);
119 fp_term = (int64_t)dc->writeback_rate_fp_term_high *
120 (c->gc_stats.in_use - BCH_WRITEBACK_FRAGMENT_THRESHOLD_HIGH);
122 fps = div_s64(dirty, dirty_buckets) * fp_term;
123 if (fragment > 3 && fps > proportional_scaled) {
124 /* Only overrite the p when fragment > 3 */
125 proportional_scaled = fps;
129 if ((error < 0 && dc->writeback_rate_integral > 0) ||
130 (error > 0 && time_before64(local_clock(),
131 dc->writeback_rate.next + NSEC_PER_MSEC))) {
133 * Only decrease the integral term if it's more than
134 * zero. Only increase the integral term if the device
135 * is keeping up. (Don't wind up the integral
136 * ineffectively in either case).
138 * It's necessary to scale this by
139 * writeback_rate_update_seconds to keep the integral
140 * term dimensioned properly.
142 dc->writeback_rate_integral += error *
143 dc->writeback_rate_update_seconds;
146 integral_scaled = div_s64(dc->writeback_rate_integral,
147 dc->writeback_rate_i_term_inverse);
149 new_rate = clamp_t(int32_t, (proportional_scaled + integral_scaled),
150 dc->writeback_rate_minimum, NSEC_PER_SEC);
152 dc->writeback_rate_proportional = proportional_scaled;
153 dc->writeback_rate_integral_scaled = integral_scaled;
154 dc->writeback_rate_change = new_rate -
155 atomic_long_read(&dc->writeback_rate.rate);
156 atomic_long_set(&dc->writeback_rate.rate, new_rate);
157 dc->writeback_rate_target = target;
160 static bool set_at_max_writeback_rate(struct cache_set *c,
161 struct cached_dev *dc)
163 /* Don't sst max writeback rate if it is disabled */
164 if (!c->idle_max_writeback_rate_enabled)
167 /* Don't set max writeback rate if gc is running */
168 if (!c->gc_mark_valid)
171 * Idle_counter is increased everytime when update_writeback_rate() is
172 * called. If all backing devices attached to the same cache set have
173 * identical dc->writeback_rate_update_seconds values, it is about 6
174 * rounds of update_writeback_rate() on each backing device before
175 * c->at_max_writeback_rate is set to 1, and then max wrteback rate set
176 * to each dc->writeback_rate.rate.
177 * In order to avoid extra locking cost for counting exact dirty cached
178 * devices number, c->attached_dev_nr is used to calculate the idle
179 * throushold. It might be bigger if not all cached device are in write-
180 * back mode, but it still works well with limited extra rounds of
181 * update_writeback_rate().
183 if (atomic_inc_return(&c->idle_counter) <
184 atomic_read(&c->attached_dev_nr) * 6)
187 if (atomic_read(&c->at_max_writeback_rate) != 1)
188 atomic_set(&c->at_max_writeback_rate, 1);
190 atomic_long_set(&dc->writeback_rate.rate, INT_MAX);
192 /* keep writeback_rate_target as existing value */
193 dc->writeback_rate_proportional = 0;
194 dc->writeback_rate_integral_scaled = 0;
195 dc->writeback_rate_change = 0;
198 * Check c->idle_counter and c->at_max_writeback_rate agagain in case
199 * new I/O arrives during before set_at_max_writeback_rate() returns.
200 * Then the writeback rate is set to 1, and its new value should be
201 * decided via __update_writeback_rate().
203 if ((atomic_read(&c->idle_counter) <
204 atomic_read(&c->attached_dev_nr) * 6) ||
205 !atomic_read(&c->at_max_writeback_rate))
211 static void update_writeback_rate(struct work_struct *work)
213 struct cached_dev *dc = container_of(to_delayed_work(work),
215 writeback_rate_update);
216 struct cache_set *c = dc->disk.c;
219 * should check BCACHE_DEV_RATE_DW_RUNNING before calling
220 * cancel_delayed_work_sync().
222 set_bit(BCACHE_DEV_RATE_DW_RUNNING, &dc->disk.flags);
223 /* paired with where BCACHE_DEV_RATE_DW_RUNNING is tested */
224 smp_mb__after_atomic();
227 * CACHE_SET_IO_DISABLE might be set via sysfs interface,
230 if (!test_bit(BCACHE_DEV_WB_RUNNING, &dc->disk.flags) ||
231 test_bit(CACHE_SET_IO_DISABLE, &c->flags)) {
232 clear_bit(BCACHE_DEV_RATE_DW_RUNNING, &dc->disk.flags);
233 /* paired with where BCACHE_DEV_RATE_DW_RUNNING is tested */
234 smp_mb__after_atomic();
238 if (atomic_read(&dc->has_dirty) && dc->writeback_percent) {
240 * If the whole cache set is idle, set_at_max_writeback_rate()
241 * will set writeback rate to a max number. Then it is
242 * unncessary to update writeback rate for an idle cache set
243 * in maximum writeback rate number(s).
245 if (!set_at_max_writeback_rate(c, dc)) {
246 down_read(&dc->writeback_lock);
247 __update_writeback_rate(dc);
248 update_gc_after_writeback(c);
249 up_read(&dc->writeback_lock);
255 * CACHE_SET_IO_DISABLE might be set via sysfs interface,
258 if (test_bit(BCACHE_DEV_WB_RUNNING, &dc->disk.flags) &&
259 !test_bit(CACHE_SET_IO_DISABLE, &c->flags)) {
260 schedule_delayed_work(&dc->writeback_rate_update,
261 dc->writeback_rate_update_seconds * HZ);
265 * should check BCACHE_DEV_RATE_DW_RUNNING before calling
266 * cancel_delayed_work_sync().
268 clear_bit(BCACHE_DEV_RATE_DW_RUNNING, &dc->disk.flags);
269 /* paired with where BCACHE_DEV_RATE_DW_RUNNING is tested */
270 smp_mb__after_atomic();
273 static unsigned int writeback_delay(struct cached_dev *dc,
274 unsigned int sectors)
276 if (test_bit(BCACHE_DEV_DETACHING, &dc->disk.flags) ||
277 !dc->writeback_percent)
280 return bch_next_delay(&dc->writeback_rate, sectors);
285 struct cached_dev *dc;
290 static void dirty_init(struct keybuf_key *w)
292 struct dirty_io *io = w->private;
293 struct bio *bio = &io->bio;
295 bio_init(bio, bio->bi_inline_vecs,
296 DIV_ROUND_UP(KEY_SIZE(&w->key), PAGE_SECTORS));
297 if (!io->dc->writeback_percent)
298 bio_set_prio(bio, IOPRIO_PRIO_VALUE(IOPRIO_CLASS_IDLE, 0));
300 bio->bi_iter.bi_size = KEY_SIZE(&w->key) << 9;
302 bch_bio_map(bio, NULL);
305 static void dirty_io_destructor(struct closure *cl)
307 struct dirty_io *io = container_of(cl, struct dirty_io, cl);
312 static void write_dirty_finish(struct closure *cl)
314 struct dirty_io *io = container_of(cl, struct dirty_io, cl);
315 struct keybuf_key *w = io->bio.bi_private;
316 struct cached_dev *dc = io->dc;
318 bio_free_pages(&io->bio);
320 /* This is kind of a dumb way of signalling errors. */
321 if (KEY_DIRTY(&w->key)) {
326 bch_keylist_init(&keys);
328 bkey_copy(keys.top, &w->key);
329 SET_KEY_DIRTY(keys.top, false);
330 bch_keylist_push(&keys);
332 for (i = 0; i < KEY_PTRS(&w->key); i++)
333 atomic_inc(&PTR_BUCKET(dc->disk.c, &w->key, i)->pin);
335 ret = bch_btree_insert(dc->disk.c, &keys, NULL, &w->key);
338 trace_bcache_writeback_collision(&w->key);
341 ? &dc->disk.c->writeback_keys_failed
342 : &dc->disk.c->writeback_keys_done);
345 bch_keybuf_del(&dc->writeback_keys, w);
348 closure_return_with_destructor(cl, dirty_io_destructor);
351 static void dirty_endio(struct bio *bio)
353 struct keybuf_key *w = bio->bi_private;
354 struct dirty_io *io = w->private;
356 if (bio->bi_status) {
357 SET_KEY_DIRTY(&w->key, false);
358 bch_count_backing_io_errors(io->dc, bio);
361 closure_put(&io->cl);
364 static void write_dirty(struct closure *cl)
366 struct dirty_io *io = container_of(cl, struct dirty_io, cl);
367 struct keybuf_key *w = io->bio.bi_private;
368 struct cached_dev *dc = io->dc;
370 uint16_t next_sequence;
372 if (atomic_read(&dc->writeback_sequence_next) != io->sequence) {
373 /* Not our turn to write; wait for a write to complete */
374 closure_wait(&dc->writeback_ordering_wait, cl);
376 if (atomic_read(&dc->writeback_sequence_next) == io->sequence) {
378 * Edge case-- it happened in indeterminate order
379 * relative to when we were added to wait list..
381 closure_wake_up(&dc->writeback_ordering_wait);
384 continue_at(cl, write_dirty, io->dc->writeback_write_wq);
388 next_sequence = io->sequence + 1;
391 * IO errors are signalled using the dirty bit on the key.
392 * If we failed to read, we should not attempt to write to the
393 * backing device. Instead, immediately go to write_dirty_finish
396 if (KEY_DIRTY(&w->key)) {
398 bio_set_op_attrs(&io->bio, REQ_OP_WRITE, 0);
399 io->bio.bi_iter.bi_sector = KEY_START(&w->key);
400 bio_set_dev(&io->bio, io->dc->bdev);
401 io->bio.bi_end_io = dirty_endio;
403 /* I/O request sent to backing device */
404 closure_bio_submit(io->dc->disk.c, &io->bio, cl);
407 atomic_set(&dc->writeback_sequence_next, next_sequence);
408 closure_wake_up(&dc->writeback_ordering_wait);
410 continue_at(cl, write_dirty_finish, io->dc->writeback_write_wq);
413 static void read_dirty_endio(struct bio *bio)
415 struct keybuf_key *w = bio->bi_private;
416 struct dirty_io *io = w->private;
419 bch_count_io_errors(io->dc->disk.c->cache,
421 "reading dirty data from cache");
426 static void read_dirty_submit(struct closure *cl)
428 struct dirty_io *io = container_of(cl, struct dirty_io, cl);
430 closure_bio_submit(io->dc->disk.c, &io->bio, cl);
432 continue_at(cl, write_dirty, io->dc->writeback_write_wq);
435 static void read_dirty(struct cached_dev *dc)
437 unsigned int delay = 0;
438 struct keybuf_key *next, *keys[MAX_WRITEBACKS_IN_PASS], *w;
443 uint16_t sequence = 0;
445 BUG_ON(!llist_empty(&dc->writeback_ordering_wait.list));
446 atomic_set(&dc->writeback_sequence_next, sequence);
447 closure_init_stack(&cl);
450 * XXX: if we error, background writeback just spins. Should use some
454 next = bch_keybuf_next(&dc->writeback_keys);
456 while (!kthread_should_stop() &&
457 !test_bit(CACHE_SET_IO_DISABLE, &dc->disk.c->flags) &&
463 BUG_ON(ptr_stale(dc->disk.c, &next->key, 0));
466 * Don't combine too many operations, even if they
469 if (nk >= MAX_WRITEBACKS_IN_PASS)
473 * If the current operation is very large, don't
474 * further combine operations.
476 if (size >= MAX_WRITESIZE_IN_PASS)
480 * Operations are only eligible to be combined
481 * if they are contiguous.
483 * TODO: add a heuristic willing to fire a
484 * certain amount of non-contiguous IO per pass,
485 * so that we can benefit from backing device
488 if ((nk != 0) && bkey_cmp(&keys[nk-1]->key,
489 &START_KEY(&next->key)))
492 size += KEY_SIZE(&next->key);
494 } while ((next = bch_keybuf_next(&dc->writeback_keys)));
496 /* Now we have gathered a set of 1..5 keys to write back. */
497 for (i = 0; i < nk; i++) {
500 io = kzalloc(struct_size(io, bio.bi_inline_vecs,
501 DIV_ROUND_UP(KEY_SIZE(&w->key), PAGE_SECTORS)),
508 io->sequence = sequence++;
511 bio_set_op_attrs(&io->bio, REQ_OP_READ, 0);
512 io->bio.bi_iter.bi_sector = PTR_OFFSET(&w->key, 0);
513 bio_set_dev(&io->bio, dc->disk.c->cache->bdev);
514 io->bio.bi_end_io = read_dirty_endio;
516 if (bch_bio_alloc_pages(&io->bio, GFP_KERNEL))
519 trace_bcache_writeback(&w->key);
521 down(&dc->in_flight);
524 * We've acquired a semaphore for the maximum
525 * simultaneous number of writebacks; from here
526 * everything happens asynchronously.
528 closure_call(&io->cl, read_dirty_submit, NULL, &cl);
531 delay = writeback_delay(dc, size);
533 while (!kthread_should_stop() &&
534 !test_bit(CACHE_SET_IO_DISABLE, &dc->disk.c->flags) &&
536 schedule_timeout_interruptible(delay);
537 delay = writeback_delay(dc, 0);
545 bch_keybuf_del(&dc->writeback_keys, w);
549 * Wait for outstanding writeback IOs to finish (and keybuf slots to be
550 * freed) before refilling again
555 /* Scan for dirty data */
557 void bcache_dev_sectors_dirty_add(struct cache_set *c, unsigned int inode,
558 uint64_t offset, int nr_sectors)
560 struct bcache_device *d = c->devices[inode];
561 unsigned int stripe_offset, sectors_dirty;
567 stripe = offset_to_stripe(d, offset);
571 if (UUID_FLASH_ONLY(&c->uuids[inode]))
572 atomic_long_add(nr_sectors, &c->flash_dev_dirty_sectors);
574 stripe_offset = offset & (d->stripe_size - 1);
577 int s = min_t(unsigned int, abs(nr_sectors),
578 d->stripe_size - stripe_offset);
583 if (stripe >= d->nr_stripes)
586 sectors_dirty = atomic_add_return(s,
587 d->stripe_sectors_dirty + stripe);
588 if (sectors_dirty == d->stripe_size)
589 set_bit(stripe, d->full_dirty_stripes);
591 clear_bit(stripe, d->full_dirty_stripes);
599 static bool dirty_pred(struct keybuf *buf, struct bkey *k)
601 struct cached_dev *dc = container_of(buf,
605 BUG_ON(KEY_INODE(k) != dc->disk.id);
610 static void refill_full_stripes(struct cached_dev *dc)
612 struct keybuf *buf = &dc->writeback_keys;
613 unsigned int start_stripe, next_stripe;
615 bool wrapped = false;
617 stripe = offset_to_stripe(&dc->disk, KEY_OFFSET(&buf->last_scanned));
621 start_stripe = stripe;
624 stripe = find_next_bit(dc->disk.full_dirty_stripes,
625 dc->disk.nr_stripes, stripe);
627 if (stripe == dc->disk.nr_stripes)
630 next_stripe = find_next_zero_bit(dc->disk.full_dirty_stripes,
631 dc->disk.nr_stripes, stripe);
633 buf->last_scanned = KEY(dc->disk.id,
634 stripe * dc->disk.stripe_size, 0);
636 bch_refill_keybuf(dc->disk.c, buf,
638 next_stripe * dc->disk.stripe_size, 0),
641 if (array_freelist_empty(&buf->freelist))
644 stripe = next_stripe;
646 if (wrapped && stripe > start_stripe)
649 if (stripe == dc->disk.nr_stripes) {
657 * Returns true if we scanned the entire disk
659 static bool refill_dirty(struct cached_dev *dc)
661 struct keybuf *buf = &dc->writeback_keys;
662 struct bkey start = KEY(dc->disk.id, 0, 0);
663 struct bkey end = KEY(dc->disk.id, MAX_KEY_OFFSET, 0);
664 struct bkey start_pos;
667 * make sure keybuf pos is inside the range for this disk - at bringup
668 * we might not be attached yet so this disk's inode nr isn't
671 if (bkey_cmp(&buf->last_scanned, &start) < 0 ||
672 bkey_cmp(&buf->last_scanned, &end) > 0)
673 buf->last_scanned = start;
675 if (dc->partial_stripes_expensive) {
676 refill_full_stripes(dc);
677 if (array_freelist_empty(&buf->freelist))
681 start_pos = buf->last_scanned;
682 bch_refill_keybuf(dc->disk.c, buf, &end, dirty_pred);
684 if (bkey_cmp(&buf->last_scanned, &end) < 0)
688 * If we get to the end start scanning again from the beginning, and
689 * only scan up to where we initially started scanning from:
691 buf->last_scanned = start;
692 bch_refill_keybuf(dc->disk.c, buf, &start_pos, dirty_pred);
694 return bkey_cmp(&buf->last_scanned, &start_pos) >= 0;
697 static int bch_writeback_thread(void *arg)
699 struct cached_dev *dc = arg;
700 struct cache_set *c = dc->disk.c;
701 bool searched_full_index;
703 bch_ratelimit_reset(&dc->writeback_rate);
705 while (!kthread_should_stop() &&
706 !test_bit(CACHE_SET_IO_DISABLE, &c->flags)) {
707 down_write(&dc->writeback_lock);
708 set_current_state(TASK_INTERRUPTIBLE);
710 * If the bache device is detaching, skip here and continue
711 * to perform writeback. Otherwise, if no dirty data on cache,
712 * or there is dirty data on cache but writeback is disabled,
713 * the writeback thread should sleep here and wait for others
716 if (!test_bit(BCACHE_DEV_DETACHING, &dc->disk.flags) &&
717 (!atomic_read(&dc->has_dirty) || !dc->writeback_running)) {
718 up_write(&dc->writeback_lock);
720 if (kthread_should_stop() ||
721 test_bit(CACHE_SET_IO_DISABLE, &c->flags)) {
722 set_current_state(TASK_RUNNING);
729 set_current_state(TASK_RUNNING);
731 searched_full_index = refill_dirty(dc);
733 if (searched_full_index &&
734 RB_EMPTY_ROOT(&dc->writeback_keys.keys)) {
735 atomic_set(&dc->has_dirty, 0);
736 SET_BDEV_STATE(&dc->sb, BDEV_STATE_CLEAN);
737 bch_write_bdev_super(dc, NULL);
739 * If bcache device is detaching via sysfs interface,
740 * writeback thread should stop after there is no dirty
741 * data on cache. BCACHE_DEV_DETACHING flag is set in
742 * bch_cached_dev_detach().
744 if (test_bit(BCACHE_DEV_DETACHING, &dc->disk.flags)) {
747 closure_init_stack(&cl);
748 memset(&dc->sb.set_uuid, 0, 16);
749 SET_BDEV_STATE(&dc->sb, BDEV_STATE_NONE);
751 bch_write_bdev_super(dc, &cl);
754 up_write(&dc->writeback_lock);
759 * When dirty data rate is high (e.g. 50%+), there might
760 * be heavy buckets fragmentation after writeback
761 * finished, which hurts following write performance.
762 * If users really care about write performance they
763 * may set BCH_ENABLE_AUTO_GC via sysfs, then when
764 * BCH_DO_AUTO_GC is set, garbage collection thread
765 * will be wake up here. After moving gc, the shrunk
766 * btree and discarded free buckets SSD space may be
767 * helpful for following write requests.
769 if (c->gc_after_writeback ==
770 (BCH_ENABLE_AUTO_GC|BCH_DO_AUTO_GC)) {
771 c->gc_after_writeback &= ~BCH_DO_AUTO_GC;
776 up_write(&dc->writeback_lock);
780 if (searched_full_index) {
781 unsigned int delay = dc->writeback_delay * HZ;
784 !kthread_should_stop() &&
785 !test_bit(CACHE_SET_IO_DISABLE, &c->flags) &&
786 !test_bit(BCACHE_DEV_DETACHING, &dc->disk.flags))
787 delay = schedule_timeout_interruptible(delay);
789 bch_ratelimit_reset(&dc->writeback_rate);
793 if (dc->writeback_write_wq) {
794 flush_workqueue(dc->writeback_write_wq);
795 destroy_workqueue(dc->writeback_write_wq);
798 wait_for_kthread_stop();
804 #define INIT_KEYS_EACH_TIME 500000
805 #define INIT_KEYS_SLEEP_MS 100
807 struct sectors_dirty_init {
814 static int sectors_dirty_init_fn(struct btree_op *_op, struct btree *b,
817 struct sectors_dirty_init *op = container_of(_op,
818 struct sectors_dirty_init, op);
819 if (KEY_INODE(k) > op->inode)
823 bcache_dev_sectors_dirty_add(b->c, KEY_INODE(k),
824 KEY_START(k), KEY_SIZE(k));
827 if (atomic_read(&b->c->search_inflight) &&
828 !(op->count % INIT_KEYS_EACH_TIME)) {
829 bkey_copy_key(&op->start, k);
836 static int bch_root_node_dirty_init(struct cache_set *c,
837 struct bcache_device *d,
840 struct sectors_dirty_init op;
843 bch_btree_op_init(&op.op, -1);
846 op.start = KEY(op.inode, 0, 0);
849 ret = bcache_btree(map_keys_recurse,
854 sectors_dirty_init_fn,
857 schedule_timeout_interruptible(
858 msecs_to_jiffies(INIT_KEYS_SLEEP_MS));
860 pr_warn("sectors dirty init failed, ret=%d!\n", ret);
863 } while (ret == -EAGAIN);
868 static int bch_dirty_init_thread(void *arg)
870 struct dirty_init_thrd_info *info = arg;
871 struct bch_dirty_init_state *state = info->state;
872 struct cache_set *c = state->c;
873 struct btree_iter iter;
875 int cur_idx, prev_idx, skip_nr;
878 cur_idx = prev_idx = 0;
880 bch_btree_iter_init(&c->root->keys, &iter, NULL);
881 k = bch_btree_iter_next_filter(&iter, &c->root->keys, bch_ptr_bad);
887 spin_lock(&state->idx_lock);
888 cur_idx = state->key_idx;
890 spin_unlock(&state->idx_lock);
892 skip_nr = cur_idx - prev_idx;
895 k = bch_btree_iter_next_filter(&iter,
901 atomic_set(&state->enough, 1);
902 /* Update state->enough earlier */
903 smp_mb__after_atomic();
911 if (bch_root_node_dirty_init(c, state->d, p) < 0)
921 /* In order to wake up state->wait in time */
922 smp_mb__before_atomic();
923 if (atomic_dec_and_test(&state->started))
924 wake_up(&state->wait);
929 static int bch_btre_dirty_init_thread_nr(void)
931 int n = num_online_cpus()/2;
935 else if (n > BCH_DIRTY_INIT_THRD_MAX)
936 n = BCH_DIRTY_INIT_THRD_MAX;
941 void bch_sectors_dirty_init(struct bcache_device *d)
944 struct bkey *k = NULL;
945 struct btree_iter iter;
946 struct sectors_dirty_init op;
947 struct cache_set *c = d->c;
948 struct bch_dirty_init_state *state;
951 /* Just count root keys if no leaf node */
952 if (c->root->level == 0) {
953 bch_btree_op_init(&op.op, -1);
956 op.start = KEY(op.inode, 0, 0);
958 for_each_key_filter(&c->root->keys,
959 k, &iter, bch_ptr_invalid)
960 sectors_dirty_init_fn(&op.op, c->root, k);
964 state = kzalloc(sizeof(struct bch_dirty_init_state), GFP_KERNEL);
966 pr_warn("sectors dirty init failed: cannot allocate memory\n");
972 state->total_threads = bch_btre_dirty_init_thread_nr();
974 spin_lock_init(&state->idx_lock);
975 atomic_set(&state->started, 0);
976 atomic_set(&state->enough, 0);
977 init_waitqueue_head(&state->wait);
979 for (i = 0; i < state->total_threads; i++) {
980 /* Fetch latest state->enough earlier */
981 smp_mb__before_atomic();
982 if (atomic_read(&state->enough))
985 state->infos[i].state = state;
986 atomic_inc(&state->started);
987 snprintf(name, sizeof(name), "bch_dirty_init[%d]", i);
989 state->infos[i].thread =
990 kthread_run(bch_dirty_init_thread,
993 if (IS_ERR(state->infos[i].thread)) {
994 pr_err("fails to run thread bch_dirty_init[%d]\n", i);
995 for (--i; i >= 0; i--)
996 kthread_stop(state->infos[i].thread);
1001 wait_event_interruptible(state->wait,
1002 atomic_read(&state->started) == 0 ||
1003 test_bit(CACHE_SET_IO_DISABLE, &c->flags));
1009 void bch_cached_dev_writeback_init(struct cached_dev *dc)
1011 sema_init(&dc->in_flight, 64);
1012 init_rwsem(&dc->writeback_lock);
1013 bch_keybuf_init(&dc->writeback_keys);
1015 dc->writeback_metadata = true;
1016 dc->writeback_running = false;
1017 dc->writeback_consider_fragment = true;
1018 dc->writeback_percent = 10;
1019 dc->writeback_delay = 30;
1020 atomic_long_set(&dc->writeback_rate.rate, 1024);
1021 dc->writeback_rate_minimum = 8;
1023 dc->writeback_rate_update_seconds = WRITEBACK_RATE_UPDATE_SECS_DEFAULT;
1024 dc->writeback_rate_p_term_inverse = 40;
1025 dc->writeback_rate_fp_term_low = 1;
1026 dc->writeback_rate_fp_term_mid = 10;
1027 dc->writeback_rate_fp_term_high = 1000;
1028 dc->writeback_rate_i_term_inverse = 10000;
1030 WARN_ON(test_and_clear_bit(BCACHE_DEV_WB_RUNNING, &dc->disk.flags));
1031 INIT_DELAYED_WORK(&dc->writeback_rate_update, update_writeback_rate);
1034 int bch_cached_dev_writeback_start(struct cached_dev *dc)
1036 dc->writeback_write_wq = alloc_workqueue("bcache_writeback_wq",
1038 if (!dc->writeback_write_wq)
1042 dc->writeback_thread = kthread_create(bch_writeback_thread, dc,
1043 "bcache_writeback");
1044 if (IS_ERR(dc->writeback_thread)) {
1046 destroy_workqueue(dc->writeback_write_wq);
1047 return PTR_ERR(dc->writeback_thread);
1049 dc->writeback_running = true;
1051 WARN_ON(test_and_set_bit(BCACHE_DEV_WB_RUNNING, &dc->disk.flags));
1052 schedule_delayed_work(&dc->writeback_rate_update,
1053 dc->writeback_rate_update_seconds * HZ);
1055 bch_writeback_queue(dc);