Lines Matching +full:idle +full:- +full:wait +full:- +full:delay
1 // SPDX-License-Identifier: GPL-2.0
3 * background writeback - scan btree for dirty data and write it to the backing
15 #include <linux/delay.h>
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;
32 struct cache_set *c = dc->disk.c;
36 * flash-only devices
38 uint64_t cache_sectors = c->nbuckets * c->cache->sb.bucket_size -
39 atomic_long_read(&c->flash_dev_dirty_sectors);
48 div64_u64(bdev_nr_sectors(dc->bdev) << WRITEBACK_SHARE_SHIFT,
49 c->cached_dev_sectors);
52 div_u64(cache_sectors * dc->writeback_percent, 100);
73 * avoid fixed point math and to make configuration easy-- e.g.
80 * This acts as a slow, long-term average that is not subject to
84 int64_t dirty = bcache_dev_sectors_dirty(&dc->disk);
85 int64_t error = dirty - target;
87 div_s64(error, dc->writeback_rate_p_term_inverse);
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);
129 if ((error < 0 && dc->writeback_rate_integral > 0) ||
131 dc->writeback_rate.next + NSEC_PER_MSEC))) {
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);
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;
165 * If c->idle_counter is overflow (idel for really long time),
169 counter = atomic_inc_return(&c->idle_counter);
171 atomic_set(&c->idle_counter, 0);
175 dev_nr = atomic_read(&c->attached_dev_nr);
180 * c->idle_counter is increased by writeback thread of all
183 * Otherwise the idle time cannot be larger with more backing
197 * identical dc->writeback_rate_update_seconds values, it is about 6
199 * c->at_max_writeback_rate is set to 1, and then max wrteback rate set
200 * to each dc->writeback_rate.rate.
202 * devices number, c->attached_dev_nr is used to calculate the idle
203 * throushold. It might be bigger if not all cached device are in write-
211 if (!c->idle_max_writeback_rate_enabled)
215 if (!c->gc_mark_valid)
221 if (atomic_read(&c->at_max_writeback_rate) != 1)
222 atomic_set(&c->at_max_writeback_rate, 1);
224 atomic_long_set(&dc->writeback_rate.rate, INT_MAX);
227 dc->writeback_rate_proportional = 0;
228 dc->writeback_rate_integral_scaled = 0;
229 dc->writeback_rate_change = 0;
236 !atomic_read(&c->at_max_writeback_rate))
247 struct cache_set *c = dc->disk.c;
253 set_bit(BCACHE_DEV_RATE_DW_RUNNING, &dc->disk.flags);
261 if (!test_bit(BCACHE_DEV_WB_RUNNING, &dc->disk.flags) ||
262 test_bit(CACHE_SET_IO_DISABLE, &c->flags)) {
263 clear_bit(BCACHE_DEV_RATE_DW_RUNNING, &dc->disk.flags);
270 * If the whole cache set is idle, set_at_max_writeback_rate()
272 * unncessary to update writeback rate for an idle cache set
275 if (atomic_read(&dc->has_dirty) && dc->writeback_percent &&
278 if (!down_read_trylock((&dc->writeback_lock))) {
279 dc->rate_update_retry++;
280 if (dc->rate_update_retry <=
283 down_read(&dc->writeback_lock);
284 dc->rate_update_retry = 0;
288 up_read(&dc->writeback_lock);
297 if (test_bit(BCACHE_DEV_WB_RUNNING, &dc->disk.flags) &&
298 !test_bit(CACHE_SET_IO_DISABLE, &c->flags)) {
299 schedule_delayed_work(&dc->writeback_rate_update,
300 dc->writeback_rate_update_seconds * HZ);
307 clear_bit(BCACHE_DEV_RATE_DW_RUNNING, &dc->disk.flags);
315 if (test_bit(BCACHE_DEV_DETACHING, &dc->disk.flags) ||
316 !dc->writeback_percent)
319 return bch_next_delay(&dc->writeback_rate, sectors);
331 struct dirty_io *io = w->private;
332 struct bio *bio = &io->bio;
334 bio_init(bio, NULL, bio->bi_inline_vecs,
335 DIV_ROUND_UP(KEY_SIZE(&w->key), PAGE_SECTORS), 0);
336 if (!io->dc->writeback_percent)
337 bio->bi_ioprio = IOPRIO_PRIO_VALUE(IOPRIO_CLASS_IDLE, 0);
339 bio->bi_iter.bi_size = KEY_SIZE(&w->key) << 9;
340 bio->bi_private = w;
354 struct keybuf_key *w = io->bio.bi_private;
355 struct cached_dev *dc = io->dc;
357 bio_free_pages(&io->bio);
360 if (KEY_DIRTY(&w->key)) {
367 bkey_copy(keys.top, &w->key);
371 for (i = 0; i < KEY_PTRS(&w->key); i++)
372 atomic_inc(&PTR_BUCKET(dc->disk.c, &w->key, i)->pin);
374 ret = bch_btree_insert(dc->disk.c, &keys, NULL, &w->key);
377 trace_bcache_writeback_collision(&w->key);
380 ? &dc->disk.c->writeback_keys_failed
381 : &dc->disk.c->writeback_keys_done);
384 bch_keybuf_del(&dc->writeback_keys, w);
385 up(&dc->in_flight);
392 struct keybuf_key *w = bio->bi_private;
393 struct dirty_io *io = w->private;
395 if (bio->bi_status) {
396 SET_KEY_DIRTY(&w->key, false);
397 bch_count_backing_io_errors(io->dc, bio);
400 closure_put(&io->cl);
406 struct keybuf_key *w = io->bio.bi_private;
407 struct cached_dev *dc = io->dc;
411 if (atomic_read(&dc->writeback_sequence_next) != io->sequence) {
412 /* Not our turn to write; wait for a write to complete */
413 closure_wait(&dc->writeback_ordering_wait, cl);
415 if (atomic_read(&dc->writeback_sequence_next) == io->sequence) {
417 * Edge case-- it happened in indeterminate order
418 * relative to when we were added to wait list..
420 closure_wake_up(&dc->writeback_ordering_wait);
423 continue_at(cl, write_dirty, io->dc->writeback_write_wq);
427 next_sequence = io->sequence + 1;
435 if (KEY_DIRTY(&w->key)) {
437 io->bio.bi_opf = REQ_OP_WRITE;
438 io->bio.bi_iter.bi_sector = KEY_START(&w->key);
439 bio_set_dev(&io->bio, io->dc->bdev);
440 io->bio.bi_end_io = dirty_endio;
443 closure_bio_submit(io->dc->disk.c, &io->bio, cl);
446 atomic_set(&dc->writeback_sequence_next, next_sequence);
447 closure_wake_up(&dc->writeback_ordering_wait);
449 continue_at(cl, write_dirty_finish, io->dc->writeback_write_wq);
454 struct keybuf_key *w = bio->bi_private;
455 struct dirty_io *io = w->private;
458 bch_count_io_errors(io->dc->disk.c->cache,
459 bio->bi_status, 1,
469 closure_bio_submit(io->dc->disk.c, &io->bio, cl);
471 continue_at(cl, write_dirty, io->dc->writeback_write_wq);
476 unsigned int delay = 0;
484 BUG_ON(!llist_empty(&dc->writeback_ordering_wait.list));
485 atomic_set(&dc->writeback_sequence_next, sequence);
493 next = bch_keybuf_next(&dc->writeback_keys);
496 !test_bit(CACHE_SET_IO_DISABLE, &dc->disk.c->flags) &&
502 BUG_ON(ptr_stale(dc->disk.c, &next->key, 0));
523 * certain amount of non-contiguous IO per pass,
527 if ((nk != 0) && bkey_cmp(&keys[nk-1]->key,
528 &START_KEY(&next->key)))
531 size += KEY_SIZE(&next->key);
533 } while ((next = bch_keybuf_next(&dc->writeback_keys)));
540 DIV_ROUND_UP(KEY_SIZE(&w->key), PAGE_SECTORS)),
545 w->private = io;
546 io->dc = dc;
547 io->sequence = sequence++;
550 io->bio.bi_opf = REQ_OP_READ;
551 io->bio.bi_iter.bi_sector = PTR_OFFSET(&w->key, 0);
552 bio_set_dev(&io->bio, dc->disk.c->cache->bdev);
553 io->bio.bi_end_io = read_dirty_endio;
555 if (bch_bio_alloc_pages(&io->bio, GFP_KERNEL))
558 trace_bcache_writeback(&w->key);
560 down(&dc->in_flight);
567 closure_call(&io->cl, read_dirty_submit, NULL, &cl);
570 delay = writeback_delay(dc, size);
573 !test_bit(CACHE_SET_IO_DISABLE, &dc->disk.c->flags) &&
574 delay) {
575 schedule_timeout_interruptible(delay);
576 delay = writeback_delay(dc, 0);
582 kfree(w->private);
584 bch_keybuf_del(&dc->writeback_keys, w);
588 * Wait for outstanding writeback IOs to finish (and keybuf slots to be
599 struct bcache_device *d = c->devices[inode];
610 if (UUID_FLASH_ONLY(&c->uuids[inode]))
611 atomic_long_add(nr_sectors, &c->flash_dev_dirty_sectors);
613 stripe_offset = offset & (d->stripe_size - 1);
617 d->stripe_size - stripe_offset);
620 s = -s;
622 if (stripe >= d->nr_stripes)
626 d->stripe_sectors_dirty + stripe);
627 if (sectors_dirty == d->stripe_size) {
628 if (!test_bit(stripe, d->full_dirty_stripes))
629 set_bit(stripe, d->full_dirty_stripes);
631 if (test_bit(stripe, d->full_dirty_stripes))
632 clear_bit(stripe, d->full_dirty_stripes);
635 nr_sectors -= s;
647 BUG_ON(KEY_INODE(k) != dc->disk.id);
654 struct keybuf *buf = &dc->writeback_keys;
659 stripe = offset_to_stripe(&dc->disk, KEY_OFFSET(&buf->last_scanned));
666 stripe = find_next_bit(dc->disk.full_dirty_stripes,
667 dc->disk.nr_stripes, stripe);
669 if (stripe == dc->disk.nr_stripes)
672 next_stripe = find_next_zero_bit(dc->disk.full_dirty_stripes,
673 dc->disk.nr_stripes, stripe);
675 buf->last_scanned = KEY(dc->disk.id,
676 stripe * dc->disk.stripe_size, 0);
678 bch_refill_keybuf(dc->disk.c, buf,
679 &KEY(dc->disk.id,
680 next_stripe * dc->disk.stripe_size, 0),
683 if (array_freelist_empty(&buf->freelist))
691 if (stripe == dc->disk.nr_stripes) {
703 struct keybuf *buf = &dc->writeback_keys;
704 struct bkey start = KEY(dc->disk.id, 0, 0);
705 struct bkey end = KEY(dc->disk.id, MAX_KEY_OFFSET, 0);
709 * make sure keybuf pos is inside the range for this disk - at bringup
713 if (bkey_cmp(&buf->last_scanned, &start) < 0 ||
714 bkey_cmp(&buf->last_scanned, &end) > 0)
715 buf->last_scanned = start;
717 if (dc->partial_stripes_expensive) {
719 if (array_freelist_empty(&buf->freelist))
723 start_pos = buf->last_scanned;
724 bch_refill_keybuf(dc->disk.c, buf, &end, dirty_pred);
726 if (bkey_cmp(&buf->last_scanned, &end) < 0)
733 buf->last_scanned = start;
734 bch_refill_keybuf(dc->disk.c, buf, &start_pos, dirty_pred);
736 return bkey_cmp(&buf->last_scanned, &start_pos) >= 0;
742 struct cache_set *c = dc->disk.c;
745 bch_ratelimit_reset(&dc->writeback_rate);
748 !test_bit(CACHE_SET_IO_DISABLE, &c->flags)) {
749 down_write(&dc->writeback_lock);
755 * the writeback thread should sleep here and wait for others
758 if (!test_bit(BCACHE_DEV_DETACHING, &dc->disk.flags) &&
759 (!atomic_read(&dc->has_dirty) || !dc->writeback_running)) {
760 up_write(&dc->writeback_lock);
763 test_bit(CACHE_SET_IO_DISABLE, &c->flags)) {
776 RB_EMPTY_ROOT(&dc->writeback_keys.keys)) {
777 atomic_set(&dc->has_dirty, 0);
778 SET_BDEV_STATE(&dc->sb, BDEV_STATE_CLEAN);
786 if (test_bit(BCACHE_DEV_DETACHING, &dc->disk.flags)) {
790 memset(&dc->sb.set_uuid, 0, 16);
791 SET_BDEV_STATE(&dc->sb, BDEV_STATE_NONE);
796 up_write(&dc->writeback_lock);
811 if (c->gc_after_writeback ==
813 c->gc_after_writeback &= ~BCH_DO_AUTO_GC;
818 up_write(&dc->writeback_lock);
823 unsigned int delay = dc->writeback_delay * HZ;
825 while (delay &&
827 !test_bit(CACHE_SET_IO_DISABLE, &c->flags) &&
828 !test_bit(BCACHE_DEV_DETACHING, &dc->disk.flags))
829 delay = schedule_timeout_interruptible(delay);
831 bch_ratelimit_reset(&dc->writeback_rate);
835 if (dc->writeback_write_wq)
836 destroy_workqueue(dc->writeback_write_wq);
858 if (KEY_INODE(k) > op->inode)
862 bcache_dev_sectors_dirty_add(b->c, KEY_INODE(k),
865 op->count++;
866 if (!(op->count % INIT_KEYS_EACH_TIME))
879 bch_btree_op_init(&op.op, -1);
880 op.inode = d->id;
885 c->root,
901 finish_wait(&c->btree_cache_wait, &(&op.op)->wait);
909 struct bch_dirty_init_state *state = info->state;
910 struct cache_set *c = state->c;
919 bch_btree_iter_init(&c->root->keys, &iter, NULL);
920 k = bch_btree_iter_next_filter(&iter, &c->root->keys, bch_ptr_bad);
926 spin_lock(&state->idx_lock);
927 cur_idx = state->key_idx;
928 state->key_idx++;
929 spin_unlock(&state->idx_lock);
931 skip_nr = cur_idx - prev_idx;
935 &c->root->keys,
940 atomic_set(&state->enough, 1);
941 /* Update state->enough earlier */
945 skip_nr--;
949 if (bch_root_node_dirty_init(c, state->d, p) < 0)
958 /* In order to wake up state->wait in time */
960 if (atomic_dec_and_test(&state->started))
961 wake_up(&state->wait);
985 struct cache_set *c = d->c;
991 b = c->root;
992 rw_lock(0, b, b->level);
993 if (b != c->root) {
999 if (c->root->level == 0) {
1000 bch_btree_op_init(&op.op, -1);
1001 op.inode = d->id;
1004 for_each_key_filter(&c->root->keys,
1008 sectors_dirty_init_fn(&op.op, c->root, k);
1023 init_waitqueue_head(&state.wait);
1039 for (--i; i >= 0; i--)
1046 /* Must wait for all threads to stop. */
1047 wait_event(state.wait, atomic_read(&state.started) == 0);
1053 sema_init(&dc->in_flight, 64);
1054 init_rwsem(&dc->writeback_lock);
1055 bch_keybuf_init(&dc->writeback_keys);
1057 dc->writeback_metadata = true;
1058 dc->writeback_running = false;
1059 dc->writeback_consider_fragment = true;
1060 dc->writeback_percent = 10;
1061 dc->writeback_delay = 30;
1062 atomic_long_set(&dc->writeback_rate.rate, 1024);
1063 dc->writeback_rate_minimum = 8;
1065 dc->writeback_rate_update_seconds = WRITEBACK_RATE_UPDATE_SECS_DEFAULT;
1066 dc->writeback_rate_p_term_inverse = 40;
1067 dc->writeback_rate_fp_term_low = 1;
1068 dc->writeback_rate_fp_term_mid = 10;
1069 dc->writeback_rate_fp_term_high = 1000;
1070 dc->writeback_rate_i_term_inverse = 10000;
1072 /* For dc->writeback_lock contention in update_writeback_rate() */
1073 dc->rate_update_retry = 0;
1075 WARN_ON(test_and_clear_bit(BCACHE_DEV_WB_RUNNING, &dc->disk.flags));
1076 INIT_DELAYED_WORK(&dc->writeback_rate_update, update_writeback_rate);
1081 dc->writeback_write_wq = alloc_workqueue("bcache_writeback_wq",
1083 if (!dc->writeback_write_wq)
1084 return -ENOMEM;
1087 dc->writeback_thread = kthread_create(bch_writeback_thread, dc,
1089 if (IS_ERR(dc->writeback_thread)) {
1091 destroy_workqueue(dc->writeback_write_wq);
1092 return PTR_ERR(dc->writeback_thread);
1094 dc->writeback_running = true;
1096 WARN_ON(test_and_set_bit(BCACHE_DEV_WB_RUNNING, &dc->disk.flags));
1097 schedule_delayed_work(&dc->writeback_rate_update,
1098 dc->writeback_rate_update_seconds * HZ);