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>
21 static uint64_t __calc_target_rate(struct cached_dev *dc)
23 struct cache_set *c = dc->disk.c;
26 * This is the size of the cache, minus the amount used for
29 uint64_t cache_sectors = c->nbuckets * c->sb.bucket_size -
30 atomic_long_read(&c->flash_dev_dirty_sectors);
33 * Unfortunately there is no control of global dirty data. If the
34 * user states that they want 10% dirty data in the cache, and has,
35 * e.g., 5 backing volumes of equal size, we try and ensure each
36 * backing volume uses about 2% of the cache for dirty data.
39 div64_u64(bdev_sectors(dc->bdev) << WRITEBACK_SHARE_SHIFT,
40 c->cached_dev_sectors);
42 uint64_t cache_dirty_target =
43 div_u64(cache_sectors * dc->writeback_percent, 100);
45 /* Ensure each backing dev gets at least one dirty share */
49 return (cache_dirty_target * bdev_share) >> WRITEBACK_SHARE_SHIFT;
52 static void __update_writeback_rate(struct cached_dev *dc)
56 * Figures out the amount that should be written per second.
58 * First, the error (number of sectors that are dirty beyond our
59 * target) is calculated. The error is accumulated (numerically
62 * Then, the proportional value and integral value are scaled
63 * based on configured values. These are stored as inverses to
64 * avoid fixed point math and to make configuration easy-- e.g.
65 * the default value of 40 for writeback_rate_p_term_inverse
66 * attempts to write at a rate that would retire all the dirty
67 * blocks in 40 seconds.
69 * The writeback_rate_i_inverse value of 10000 means that 1/10000th
70 * of the error is accumulated in the integral term per second.
71 * This acts as a slow, long-term average that is not subject to
72 * variations in usage like the p term.
74 int64_t target = __calc_target_rate(dc);
75 int64_t dirty = bcache_dev_sectors_dirty(&dc->disk);
76 int64_t error = dirty - target;
77 int64_t proportional_scaled =
78 div_s64(error, dc->writeback_rate_p_term_inverse);
79 int64_t integral_scaled;
82 if ((error < 0 && dc->writeback_rate_integral > 0) ||
83 (error > 0 && time_before64(local_clock(),
84 dc->writeback_rate.next + NSEC_PER_MSEC))) {
86 * Only decrease the integral term if it's more than
87 * zero. Only increase the integral term if the device
88 * is keeping up. (Don't wind up the integral
89 * ineffectively in either case).
91 * It's necessary to scale this by
92 * writeback_rate_update_seconds to keep the integral
93 * term dimensioned properly.
95 dc->writeback_rate_integral += error *
96 dc->writeback_rate_update_seconds;
99 integral_scaled = div_s64(dc->writeback_rate_integral,
100 dc->writeback_rate_i_term_inverse);
102 new_rate = clamp_t(int32_t, (proportional_scaled + integral_scaled),
103 dc->writeback_rate_minimum, NSEC_PER_SEC);
105 dc->writeback_rate_proportional = proportional_scaled;
106 dc->writeback_rate_integral_scaled = integral_scaled;
107 dc->writeback_rate_change = new_rate - dc->writeback_rate.rate;
108 dc->writeback_rate.rate = new_rate;
109 dc->writeback_rate_target = target;
112 static void update_writeback_rate(struct work_struct *work)
114 struct cached_dev *dc = container_of(to_delayed_work(work),
116 writeback_rate_update);
117 struct cache_set *c = dc->disk.c;
120 * should check BCACHE_DEV_RATE_DW_RUNNING before calling
121 * cancel_delayed_work_sync().
123 set_bit(BCACHE_DEV_RATE_DW_RUNNING, &dc->disk.flags);
124 /* paired with where BCACHE_DEV_RATE_DW_RUNNING is tested */
128 * CACHE_SET_IO_DISABLE might be set via sysfs interface,
131 if (!test_bit(BCACHE_DEV_WB_RUNNING, &dc->disk.flags) ||
132 test_bit(CACHE_SET_IO_DISABLE, &c->flags)) {
133 clear_bit(BCACHE_DEV_RATE_DW_RUNNING, &dc->disk.flags);
134 /* paired with where BCACHE_DEV_RATE_DW_RUNNING is tested */
139 down_read(&dc->writeback_lock);
141 if (atomic_read(&dc->has_dirty) &&
142 dc->writeback_percent)
143 __update_writeback_rate(dc);
145 up_read(&dc->writeback_lock);
148 * CACHE_SET_IO_DISABLE might be set via sysfs interface,
151 if (test_bit(BCACHE_DEV_WB_RUNNING, &dc->disk.flags) &&
152 !test_bit(CACHE_SET_IO_DISABLE, &c->flags)) {
153 schedule_delayed_work(&dc->writeback_rate_update,
154 dc->writeback_rate_update_seconds * HZ);
158 * should check BCACHE_DEV_RATE_DW_RUNNING before calling
159 * cancel_delayed_work_sync().
161 clear_bit(BCACHE_DEV_RATE_DW_RUNNING, &dc->disk.flags);
162 /* paired with where BCACHE_DEV_RATE_DW_RUNNING is tested */
166 static unsigned writeback_delay(struct cached_dev *dc, unsigned sectors)
168 if (test_bit(BCACHE_DEV_DETACHING, &dc->disk.flags) ||
169 !dc->writeback_percent)
172 return bch_next_delay(&dc->writeback_rate, sectors);
177 struct cached_dev *dc;
182 static void dirty_init(struct keybuf_key *w)
184 struct dirty_io *io = w->private;
185 struct bio *bio = &io->bio;
187 bio_init(bio, bio->bi_inline_vecs,
188 DIV_ROUND_UP(KEY_SIZE(&w->key), PAGE_SECTORS));
189 if (!io->dc->writeback_percent)
190 bio_set_prio(bio, IOPRIO_PRIO_VALUE(IOPRIO_CLASS_IDLE, 0));
192 bio->bi_iter.bi_size = KEY_SIZE(&w->key) << 9;
194 bch_bio_map(bio, NULL);
197 static void dirty_io_destructor(struct closure *cl)
199 struct dirty_io *io = container_of(cl, struct dirty_io, cl);
203 static void write_dirty_finish(struct closure *cl)
205 struct dirty_io *io = container_of(cl, struct dirty_io, cl);
206 struct keybuf_key *w = io->bio.bi_private;
207 struct cached_dev *dc = io->dc;
209 bio_free_pages(&io->bio);
211 /* This is kind of a dumb way of signalling errors. */
212 if (KEY_DIRTY(&w->key)) {
217 bch_keylist_init(&keys);
219 bkey_copy(keys.top, &w->key);
220 SET_KEY_DIRTY(keys.top, false);
221 bch_keylist_push(&keys);
223 for (i = 0; i < KEY_PTRS(&w->key); i++)
224 atomic_inc(&PTR_BUCKET(dc->disk.c, &w->key, i)->pin);
226 ret = bch_btree_insert(dc->disk.c, &keys, NULL, &w->key);
229 trace_bcache_writeback_collision(&w->key);
232 ? &dc->disk.c->writeback_keys_failed
233 : &dc->disk.c->writeback_keys_done);
236 bch_keybuf_del(&dc->writeback_keys, w);
239 closure_return_with_destructor(cl, dirty_io_destructor);
242 static void dirty_endio(struct bio *bio)
244 struct keybuf_key *w = bio->bi_private;
245 struct dirty_io *io = w->private;
247 if (bio->bi_status) {
248 SET_KEY_DIRTY(&w->key, false);
249 bch_count_backing_io_errors(io->dc, bio);
252 closure_put(&io->cl);
255 static void write_dirty(struct closure *cl)
257 struct dirty_io *io = container_of(cl, struct dirty_io, cl);
258 struct keybuf_key *w = io->bio.bi_private;
259 struct cached_dev *dc = io->dc;
261 uint16_t next_sequence;
263 if (atomic_read(&dc->writeback_sequence_next) != io->sequence) {
264 /* Not our turn to write; wait for a write to complete */
265 closure_wait(&dc->writeback_ordering_wait, cl);
267 if (atomic_read(&dc->writeback_sequence_next) == io->sequence) {
269 * Edge case-- it happened in indeterminate order
270 * relative to when we were added to wait list..
272 closure_wake_up(&dc->writeback_ordering_wait);
275 continue_at(cl, write_dirty, io->dc->writeback_write_wq);
279 next_sequence = io->sequence + 1;
282 * IO errors are signalled using the dirty bit on the key.
283 * If we failed to read, we should not attempt to write to the
284 * backing device. Instead, immediately go to write_dirty_finish
287 if (KEY_DIRTY(&w->key)) {
289 bio_set_op_attrs(&io->bio, REQ_OP_WRITE, 0);
290 io->bio.bi_iter.bi_sector = KEY_START(&w->key);
291 bio_set_dev(&io->bio, io->dc->bdev);
292 io->bio.bi_end_io = dirty_endio;
294 /* I/O request sent to backing device */
295 closure_bio_submit(io->dc->disk.c, &io->bio, cl);
298 atomic_set(&dc->writeback_sequence_next, next_sequence);
299 closure_wake_up(&dc->writeback_ordering_wait);
301 continue_at(cl, write_dirty_finish, io->dc->writeback_write_wq);
304 static void read_dirty_endio(struct bio *bio)
306 struct keybuf_key *w = bio->bi_private;
307 struct dirty_io *io = w->private;
310 bch_count_io_errors(PTR_CACHE(io->dc->disk.c, &w->key, 0),
312 "reading dirty data from cache");
317 static void read_dirty_submit(struct closure *cl)
319 struct dirty_io *io = container_of(cl, struct dirty_io, cl);
321 closure_bio_submit(io->dc->disk.c, &io->bio, cl);
323 continue_at(cl, write_dirty, io->dc->writeback_write_wq);
326 static void read_dirty(struct cached_dev *dc)
329 struct keybuf_key *next, *keys[MAX_WRITEBACKS_IN_PASS], *w;
334 uint16_t sequence = 0;
336 BUG_ON(!llist_empty(&dc->writeback_ordering_wait.list));
337 atomic_set(&dc->writeback_sequence_next, sequence);
338 closure_init_stack(&cl);
341 * XXX: if we error, background writeback just spins. Should use some
345 next = bch_keybuf_next(&dc->writeback_keys);
347 while (!kthread_should_stop() &&
348 !test_bit(CACHE_SET_IO_DISABLE, &dc->disk.c->flags) &&
354 BUG_ON(ptr_stale(dc->disk.c, &next->key, 0));
357 * Don't combine too many operations, even if they
360 if (nk >= MAX_WRITEBACKS_IN_PASS)
364 * If the current operation is very large, don't
365 * further combine operations.
367 if (size >= MAX_WRITESIZE_IN_PASS)
371 * Operations are only eligible to be combined
372 * if they are contiguous.
374 * TODO: add a heuristic willing to fire a
375 * certain amount of non-contiguous IO per pass,
376 * so that we can benefit from backing device
379 if ((nk != 0) && bkey_cmp(&keys[nk-1]->key,
380 &START_KEY(&next->key)))
383 size += KEY_SIZE(&next->key);
385 } while ((next = bch_keybuf_next(&dc->writeback_keys)));
387 /* Now we have gathered a set of 1..5 keys to write back. */
388 for (i = 0; i < nk; i++) {
391 io = kzalloc(sizeof(struct dirty_io) +
392 sizeof(struct bio_vec) *
393 DIV_ROUND_UP(KEY_SIZE(&w->key), PAGE_SECTORS),
400 io->sequence = sequence++;
403 bio_set_op_attrs(&io->bio, REQ_OP_READ, 0);
404 io->bio.bi_iter.bi_sector = PTR_OFFSET(&w->key, 0);
405 bio_set_dev(&io->bio,
406 PTR_CACHE(dc->disk.c, &w->key, 0)->bdev);
407 io->bio.bi_end_io = read_dirty_endio;
409 if (bch_bio_alloc_pages(&io->bio, GFP_KERNEL))
412 trace_bcache_writeback(&w->key);
414 down(&dc->in_flight);
416 /* We've acquired a semaphore for the maximum
417 * simultaneous number of writebacks; from here
418 * everything happens asynchronously.
420 closure_call(&io->cl, read_dirty_submit, NULL, &cl);
423 delay = writeback_delay(dc, size);
425 /* If the control system would wait for at least half a
426 * second, and there's been no reqs hitting the backing disk
427 * for awhile: use an alternate mode where we have at most
428 * one contiguous set of writebacks in flight at a time. If
429 * someone wants to do IO it will be quick, as it will only
430 * have to contend with one operation in flight, and we'll
431 * be round-tripping data to the backing disk as quickly as
434 if (delay >= HZ / 2) {
435 /* 3 means at least 1.5 seconds, up to 7.5 if we
436 * have slowed way down.
438 if (atomic_inc_return(&dc->backing_idle) >= 3) {
439 /* Wait for current I/Os to finish */
441 /* And immediately launch a new set. */
446 while (!kthread_should_stop() &&
447 !test_bit(CACHE_SET_IO_DISABLE, &dc->disk.c->flags) &&
449 schedule_timeout_interruptible(delay);
450 delay = writeback_delay(dc, 0);
458 bch_keybuf_del(&dc->writeback_keys, w);
462 * Wait for outstanding writeback IOs to finish (and keybuf slots to be
463 * freed) before refilling again
468 /* Scan for dirty data */
470 void bcache_dev_sectors_dirty_add(struct cache_set *c, unsigned inode,
471 uint64_t offset, int nr_sectors)
473 struct bcache_device *d = c->devices[inode];
474 unsigned stripe_offset, stripe, sectors_dirty;
479 if (UUID_FLASH_ONLY(&c->uuids[inode]))
480 atomic_long_add(nr_sectors, &c->flash_dev_dirty_sectors);
482 stripe = offset_to_stripe(d, offset);
483 stripe_offset = offset & (d->stripe_size - 1);
486 int s = min_t(unsigned, abs(nr_sectors),
487 d->stripe_size - stripe_offset);
492 if (stripe >= d->nr_stripes)
495 sectors_dirty = atomic_add_return(s,
496 d->stripe_sectors_dirty + stripe);
497 if (sectors_dirty == d->stripe_size)
498 set_bit(stripe, d->full_dirty_stripes);
500 clear_bit(stripe, d->full_dirty_stripes);
508 static bool dirty_pred(struct keybuf *buf, struct bkey *k)
510 struct cached_dev *dc = container_of(buf, struct cached_dev, writeback_keys);
512 BUG_ON(KEY_INODE(k) != dc->disk.id);
517 static void refill_full_stripes(struct cached_dev *dc)
519 struct keybuf *buf = &dc->writeback_keys;
520 unsigned start_stripe, stripe, next_stripe;
521 bool wrapped = false;
523 stripe = offset_to_stripe(&dc->disk, KEY_OFFSET(&buf->last_scanned));
525 if (stripe >= dc->disk.nr_stripes)
528 start_stripe = stripe;
531 stripe = find_next_bit(dc->disk.full_dirty_stripes,
532 dc->disk.nr_stripes, stripe);
534 if (stripe == dc->disk.nr_stripes)
537 next_stripe = find_next_zero_bit(dc->disk.full_dirty_stripes,
538 dc->disk.nr_stripes, stripe);
540 buf->last_scanned = KEY(dc->disk.id,
541 stripe * dc->disk.stripe_size, 0);
543 bch_refill_keybuf(dc->disk.c, buf,
545 next_stripe * dc->disk.stripe_size, 0),
548 if (array_freelist_empty(&buf->freelist))
551 stripe = next_stripe;
553 if (wrapped && stripe > start_stripe)
556 if (stripe == dc->disk.nr_stripes) {
564 * Returns true if we scanned the entire disk
566 static bool refill_dirty(struct cached_dev *dc)
568 struct keybuf *buf = &dc->writeback_keys;
569 struct bkey start = KEY(dc->disk.id, 0, 0);
570 struct bkey end = KEY(dc->disk.id, MAX_KEY_OFFSET, 0);
571 struct bkey start_pos;
574 * make sure keybuf pos is inside the range for this disk - at bringup
575 * we might not be attached yet so this disk's inode nr isn't
578 if (bkey_cmp(&buf->last_scanned, &start) < 0 ||
579 bkey_cmp(&buf->last_scanned, &end) > 0)
580 buf->last_scanned = start;
582 if (dc->partial_stripes_expensive) {
583 refill_full_stripes(dc);
584 if (array_freelist_empty(&buf->freelist))
588 start_pos = buf->last_scanned;
589 bch_refill_keybuf(dc->disk.c, buf, &end, dirty_pred);
591 if (bkey_cmp(&buf->last_scanned, &end) < 0)
595 * If we get to the end start scanning again from the beginning, and
596 * only scan up to where we initially started scanning from:
598 buf->last_scanned = start;
599 bch_refill_keybuf(dc->disk.c, buf, &start_pos, dirty_pred);
601 return bkey_cmp(&buf->last_scanned, &start_pos) >= 0;
604 static int bch_writeback_thread(void *arg)
606 struct cached_dev *dc = arg;
607 struct cache_set *c = dc->disk.c;
608 bool searched_full_index;
610 bch_ratelimit_reset(&dc->writeback_rate);
612 while (!kthread_should_stop() &&
613 !test_bit(CACHE_SET_IO_DISABLE, &c->flags)) {
614 down_write(&dc->writeback_lock);
615 set_current_state(TASK_INTERRUPTIBLE);
617 * If the bache device is detaching, skip here and continue
618 * to perform writeback. Otherwise, if no dirty data on cache,
619 * or there is dirty data on cache but writeback is disabled,
620 * the writeback thread should sleep here and wait for others
623 if (!test_bit(BCACHE_DEV_DETACHING, &dc->disk.flags) &&
624 (!atomic_read(&dc->has_dirty) || !dc->writeback_running)) {
625 up_write(&dc->writeback_lock);
627 if (kthread_should_stop() ||
628 test_bit(CACHE_SET_IO_DISABLE, &c->flags)) {
629 set_current_state(TASK_RUNNING);
636 set_current_state(TASK_RUNNING);
638 searched_full_index = refill_dirty(dc);
640 if (searched_full_index &&
641 RB_EMPTY_ROOT(&dc->writeback_keys.keys)) {
642 atomic_set(&dc->has_dirty, 0);
643 SET_BDEV_STATE(&dc->sb, BDEV_STATE_CLEAN);
644 bch_write_bdev_super(dc, NULL);
646 * If bcache device is detaching via sysfs interface,
647 * writeback thread should stop after there is no dirty
648 * data on cache. BCACHE_DEV_DETACHING flag is set in
649 * bch_cached_dev_detach().
651 if (test_bit(BCACHE_DEV_DETACHING, &dc->disk.flags))
655 up_write(&dc->writeback_lock);
659 if (searched_full_index) {
660 unsigned delay = dc->writeback_delay * HZ;
663 !kthread_should_stop() &&
664 !test_bit(CACHE_SET_IO_DISABLE, &c->flags) &&
665 !test_bit(BCACHE_DEV_DETACHING, &dc->disk.flags))
666 delay = schedule_timeout_interruptible(delay);
668 bch_ratelimit_reset(&dc->writeback_rate);
673 wait_for_kthread_stop();
679 #define INIT_KEYS_EACH_TIME 500000
680 #define INIT_KEYS_SLEEP_MS 100
682 struct sectors_dirty_init {
689 static int sectors_dirty_init_fn(struct btree_op *_op, struct btree *b,
692 struct sectors_dirty_init *op = container_of(_op,
693 struct sectors_dirty_init, op);
694 if (KEY_INODE(k) > op->inode)
698 bcache_dev_sectors_dirty_add(b->c, KEY_INODE(k),
699 KEY_START(k), KEY_SIZE(k));
702 if (atomic_read(&b->c->search_inflight) &&
703 !(op->count % INIT_KEYS_EACH_TIME)) {
704 bkey_copy_key(&op->start, k);
711 void bch_sectors_dirty_init(struct bcache_device *d)
713 struct sectors_dirty_init op;
716 bch_btree_op_init(&op.op, -1);
719 op.start = KEY(op.inode, 0, 0);
722 ret = bch_btree_map_keys(&op.op, d->c, &op.start,
723 sectors_dirty_init_fn, 0);
725 schedule_timeout_interruptible(
726 msecs_to_jiffies(INIT_KEYS_SLEEP_MS));
728 pr_warn("sectors dirty init failed, ret=%d!", ret);
731 } while (ret == -EAGAIN);
734 void bch_cached_dev_writeback_init(struct cached_dev *dc)
736 sema_init(&dc->in_flight, 64);
737 init_rwsem(&dc->writeback_lock);
738 bch_keybuf_init(&dc->writeback_keys);
740 dc->writeback_metadata = true;
741 dc->writeback_running = true;
742 dc->writeback_percent = 10;
743 dc->writeback_delay = 30;
744 dc->writeback_rate.rate = 1024;
745 dc->writeback_rate_minimum = 8;
747 dc->writeback_rate_update_seconds = WRITEBACK_RATE_UPDATE_SECS_DEFAULT;
748 dc->writeback_rate_p_term_inverse = 40;
749 dc->writeback_rate_i_term_inverse = 10000;
751 WARN_ON(test_and_clear_bit(BCACHE_DEV_WB_RUNNING, &dc->disk.flags));
752 INIT_DELAYED_WORK(&dc->writeback_rate_update, update_writeback_rate);
755 int bch_cached_dev_writeback_start(struct cached_dev *dc)
757 dc->writeback_write_wq = alloc_workqueue("bcache_writeback_wq",
759 if (!dc->writeback_write_wq)
763 dc->writeback_thread = kthread_create(bch_writeback_thread, dc,
765 if (IS_ERR(dc->writeback_thread)) {
767 return PTR_ERR(dc->writeback_thread);
770 WARN_ON(test_and_set_bit(BCACHE_DEV_WB_RUNNING, &dc->disk.flags));
771 schedule_delayed_work(&dc->writeback_rate_update,
772 dc->writeback_rate_update_seconds * HZ);
774 bch_writeback_queue(dc);