Loading...
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * background writeback - scan btree for dirty data and write it to the backing
4 * device
5 *
6 * Copyright 2010, 2011 Kent Overstreet <kent.overstreet@gmail.com>
7 * Copyright 2012 Google, Inc.
8 */
9
10#include "bcache.h"
11#include "btree.h"
12#include "debug.h"
13#include "writeback.h"
14
15#include <linux/delay.h>
16#include <linux/kthread.h>
17#include <linux/sched/clock.h>
18#include <trace/events/bcache.h>
19
20static void update_gc_after_writeback(struct cache_set *c)
21{
22 if (c->gc_after_writeback != (BCH_ENABLE_AUTO_GC) ||
23 c->gc_stats.in_use < BCH_AUTO_GC_DIRTY_THRESHOLD)
24 return;
25
26 c->gc_after_writeback |= BCH_DO_AUTO_GC;
27}
28
29/* Rate limiting */
30static uint64_t __calc_target_rate(struct cached_dev *dc)
31{
32 struct cache_set *c = dc->disk.c;
33
34 /*
35 * This is the size of the cache, minus the amount used for
36 * flash-only devices
37 */
38 uint64_t cache_sectors = c->nbuckets * c->cache->sb.bucket_size -
39 atomic_long_read(&c->flash_dev_dirty_sectors);
40
41 /*
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.
46 */
47 uint32_t bdev_share =
48 div64_u64(bdev_nr_sectors(dc->bdev) << WRITEBACK_SHARE_SHIFT,
49 c->cached_dev_sectors);
50
51 uint64_t cache_dirty_target =
52 div_u64(cache_sectors * dc->writeback_percent, 100);
53
54 /* Ensure each backing dev gets at least one dirty share */
55 if (bdev_share < 1)
56 bdev_share = 1;
57
58 return (cache_dirty_target * bdev_share) >> WRITEBACK_SHARE_SHIFT;
59}
60
61static void __update_writeback_rate(struct cached_dev *dc)
62{
63 /*
64 * PI controller:
65 * Figures out the amount that should be written per second.
66 *
67 * First, the error (number of sectors that are dirty beyond our
68 * target) is calculated. The error is accumulated (numerically
69 * integrated).
70 *
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.
77 *
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.
82 */
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;
89 uint32_t new_rate;
90
91 /*
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.
100 */
101 struct cache_set *c = dc->disk.c;
102
103 int64_t dirty_buckets = c->nbuckets - c->avail_nbuckets;
104
105 if (dc->writeback_consider_fragment &&
106 c->gc_stats.in_use > BCH_WRITEBACK_FRAGMENT_THRESHOLD_LOW && dirty > 0) {
107 int64_t fragment =
108 div_s64((dirty_buckets * c->cache->sb.bucket_size), dirty);
109 int64_t fp_term;
110 int64_t fps;
111
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);
118 } else {
119 fp_term = (int64_t)dc->writeback_rate_fp_term_high *
120 (c->gc_stats.in_use - BCH_WRITEBACK_FRAGMENT_THRESHOLD_HIGH);
121 }
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;
126 }
127 }
128
129 if ((error < 0 && dc->writeback_rate_integral > 0) ||
130 (error > 0 && time_before64(local_clock(),
131 dc->writeback_rate.next + NSEC_PER_MSEC))) {
132 /*
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).
137 *
138 * It's necessary to scale this by
139 * writeback_rate_update_seconds to keep the integral
140 * term dimensioned properly.
141 */
142 dc->writeback_rate_integral += error *
143 dc->writeback_rate_update_seconds;
144 }
145
146 integral_scaled = div_s64(dc->writeback_rate_integral,
147 dc->writeback_rate_i_term_inverse);
148
149 new_rate = clamp_t(int32_t, (proportional_scaled + integral_scaled),
150 dc->writeback_rate_minimum, NSEC_PER_SEC);
151
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;
158}
159
160static bool idle_counter_exceeded(struct cache_set *c)
161{
162 int counter, dev_nr;
163
164 /*
165 * If c->idle_counter is overflow (idel for really long time),
166 * reset as 0 and not set maximum rate this time for code
167 * simplicity.
168 */
169 counter = atomic_inc_return(&c->idle_counter);
170 if (counter <= 0) {
171 atomic_set(&c->idle_counter, 0);
172 return false;
173 }
174
175 dev_nr = atomic_read(&c->attached_dev_nr);
176 if (dev_nr == 0)
177 return false;
178
179 /*
180 * c->idle_counter is increased by writeback thread of all
181 * attached backing devices, in order to represent a rough
182 * time period, counter should be divided by dev_nr.
183 * Otherwise the idle time cannot be larger with more backing
184 * device attached.
185 * The following calculation equals to checking
186 * (counter / dev_nr) < (dev_nr * 6)
187 */
188 if (counter < (dev_nr * dev_nr * 6))
189 return false;
190
191 return true;
192}
193
194/*
195 * Idle_counter is increased every time when update_writeback_rate() is
196 * called. If all backing devices attached to the same cache set have
197 * identical dc->writeback_rate_update_seconds values, it is about 6
198 * rounds of update_writeback_rate() on each backing device before
199 * c->at_max_writeback_rate is set to 1, and then max wrteback rate set
200 * to each dc->writeback_rate.rate.
201 * In order to avoid extra locking cost for counting exact dirty cached
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-
204 * back mode, but it still works well with limited extra rounds of
205 * update_writeback_rate().
206 */
207static bool set_at_max_writeback_rate(struct cache_set *c,
208 struct cached_dev *dc)
209{
210 /* Don't sst max writeback rate if it is disabled */
211 if (!c->idle_max_writeback_rate_enabled)
212 return false;
213
214 /* Don't set max writeback rate if gc is running */
215 if (!c->gc_mark_valid)
216 return false;
217
218 if (!idle_counter_exceeded(c))
219 return false;
220
221 if (atomic_read(&c->at_max_writeback_rate) != 1)
222 atomic_set(&c->at_max_writeback_rate, 1);
223
224 atomic_long_set(&dc->writeback_rate.rate, INT_MAX);
225
226 /* keep writeback_rate_target as existing value */
227 dc->writeback_rate_proportional = 0;
228 dc->writeback_rate_integral_scaled = 0;
229 dc->writeback_rate_change = 0;
230
231 /*
232 * In case new I/O arrives during before
233 * set_at_max_writeback_rate() returns.
234 */
235 if (!idle_counter_exceeded(c) ||
236 !atomic_read(&c->at_max_writeback_rate))
237 return false;
238
239 return true;
240}
241
242static void update_writeback_rate(struct work_struct *work)
243{
244 struct cached_dev *dc = container_of(to_delayed_work(work),
245 struct cached_dev,
246 writeback_rate_update);
247 struct cache_set *c = dc->disk.c;
248
249 /*
250 * should check BCACHE_DEV_RATE_DW_RUNNING before calling
251 * cancel_delayed_work_sync().
252 */
253 set_bit(BCACHE_DEV_RATE_DW_RUNNING, &dc->disk.flags);
254 /* paired with where BCACHE_DEV_RATE_DW_RUNNING is tested */
255 smp_mb__after_atomic();
256
257 /*
258 * CACHE_SET_IO_DISABLE might be set via sysfs interface,
259 * check it here too.
260 */
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);
264 /* paired with where BCACHE_DEV_RATE_DW_RUNNING is tested */
265 smp_mb__after_atomic();
266 return;
267 }
268
269 /*
270 * If the whole cache set is idle, set_at_max_writeback_rate()
271 * will set writeback rate to a max number. Then it is
272 * unncessary to update writeback rate for an idle cache set
273 * in maximum writeback rate number(s).
274 */
275 if (atomic_read(&dc->has_dirty) && dc->writeback_percent &&
276 !set_at_max_writeback_rate(c, dc)) {
277 do {
278 if (!down_read_trylock((&dc->writeback_lock))) {
279 dc->rate_update_retry++;
280 if (dc->rate_update_retry <=
281 BCH_WBRATE_UPDATE_MAX_SKIPS)
282 break;
283 down_read(&dc->writeback_lock);
284 dc->rate_update_retry = 0;
285 }
286 __update_writeback_rate(dc);
287 update_gc_after_writeback(c);
288 up_read(&dc->writeback_lock);
289 } while (0);
290 }
291
292
293 /*
294 * CACHE_SET_IO_DISABLE might be set via sysfs interface,
295 * check it here too.
296 */
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);
301 }
302
303 /*
304 * should check BCACHE_DEV_RATE_DW_RUNNING before calling
305 * cancel_delayed_work_sync().
306 */
307 clear_bit(BCACHE_DEV_RATE_DW_RUNNING, &dc->disk.flags);
308 /* paired with where BCACHE_DEV_RATE_DW_RUNNING is tested */
309 smp_mb__after_atomic();
310}
311
312static unsigned int writeback_delay(struct cached_dev *dc,
313 unsigned int sectors)
314{
315 if (test_bit(BCACHE_DEV_DETACHING, &dc->disk.flags) ||
316 !dc->writeback_percent)
317 return 0;
318
319 return bch_next_delay(&dc->writeback_rate, sectors);
320}
321
322struct dirty_io {
323 struct closure cl;
324 struct cached_dev *dc;
325 uint16_t sequence;
326 struct bio bio;
327};
328
329static void dirty_init(struct keybuf_key *w)
330{
331 struct dirty_io *io = w->private;
332 struct bio *bio = &io->bio;
333
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_set_prio(bio, IOPRIO_PRIO_VALUE(IOPRIO_CLASS_IDLE, 0));
338
339 bio->bi_iter.bi_size = KEY_SIZE(&w->key) << 9;
340 bio->bi_private = w;
341 bch_bio_map(bio, NULL);
342}
343
344static CLOSURE_CALLBACK(dirty_io_destructor)
345{
346 closure_type(io, struct dirty_io, cl);
347
348 kfree(io);
349}
350
351static CLOSURE_CALLBACK(write_dirty_finish)
352{
353 closure_type(io, struct dirty_io, cl);
354 struct keybuf_key *w = io->bio.bi_private;
355 struct cached_dev *dc = io->dc;
356
357 bio_free_pages(&io->bio);
358
359 /* This is kind of a dumb way of signalling errors. */
360 if (KEY_DIRTY(&w->key)) {
361 int ret;
362 unsigned int i;
363 struct keylist keys;
364
365 bch_keylist_init(&keys);
366
367 bkey_copy(keys.top, &w->key);
368 SET_KEY_DIRTY(keys.top, false);
369 bch_keylist_push(&keys);
370
371 for (i = 0; i < KEY_PTRS(&w->key); i++)
372 atomic_inc(&PTR_BUCKET(dc->disk.c, &w->key, i)->pin);
373
374 ret = bch_btree_insert(dc->disk.c, &keys, NULL, &w->key);
375
376 if (ret)
377 trace_bcache_writeback_collision(&w->key);
378
379 atomic_long_inc(ret
380 ? &dc->disk.c->writeback_keys_failed
381 : &dc->disk.c->writeback_keys_done);
382 }
383
384 bch_keybuf_del(&dc->writeback_keys, w);
385 up(&dc->in_flight);
386
387 closure_return_with_destructor(cl, dirty_io_destructor);
388}
389
390static void dirty_endio(struct bio *bio)
391{
392 struct keybuf_key *w = bio->bi_private;
393 struct dirty_io *io = w->private;
394
395 if (bio->bi_status) {
396 SET_KEY_DIRTY(&w->key, false);
397 bch_count_backing_io_errors(io->dc, bio);
398 }
399
400 closure_put(&io->cl);
401}
402
403static CLOSURE_CALLBACK(write_dirty)
404{
405 closure_type(io, struct dirty_io, cl);
406 struct keybuf_key *w = io->bio.bi_private;
407 struct cached_dev *dc = io->dc;
408
409 uint16_t next_sequence;
410
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);
414
415 if (atomic_read(&dc->writeback_sequence_next) == io->sequence) {
416 /*
417 * Edge case-- it happened in indeterminate order
418 * relative to when we were added to wait list..
419 */
420 closure_wake_up(&dc->writeback_ordering_wait);
421 }
422
423 continue_at(cl, write_dirty, io->dc->writeback_write_wq);
424 return;
425 }
426
427 next_sequence = io->sequence + 1;
428
429 /*
430 * IO errors are signalled using the dirty bit on the key.
431 * If we failed to read, we should not attempt to write to the
432 * backing device. Instead, immediately go to write_dirty_finish
433 * to clean up.
434 */
435 if (KEY_DIRTY(&w->key)) {
436 dirty_init(w);
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;
441
442 /* I/O request sent to backing device */
443 closure_bio_submit(io->dc->disk.c, &io->bio, cl);
444 }
445
446 atomic_set(&dc->writeback_sequence_next, next_sequence);
447 closure_wake_up(&dc->writeback_ordering_wait);
448
449 continue_at(cl, write_dirty_finish, io->dc->writeback_write_wq);
450}
451
452static void read_dirty_endio(struct bio *bio)
453{
454 struct keybuf_key *w = bio->bi_private;
455 struct dirty_io *io = w->private;
456
457 /* is_read = 1 */
458 bch_count_io_errors(io->dc->disk.c->cache,
459 bio->bi_status, 1,
460 "reading dirty data from cache");
461
462 dirty_endio(bio);
463}
464
465static CLOSURE_CALLBACK(read_dirty_submit)
466{
467 closure_type(io, struct dirty_io, cl);
468
469 closure_bio_submit(io->dc->disk.c, &io->bio, cl);
470
471 continue_at(cl, write_dirty, io->dc->writeback_write_wq);
472}
473
474static void read_dirty(struct cached_dev *dc)
475{
476 unsigned int delay = 0;
477 struct keybuf_key *next, *keys[MAX_WRITEBACKS_IN_PASS], *w;
478 size_t size;
479 int nk, i;
480 struct dirty_io *io;
481 struct closure cl;
482 uint16_t sequence = 0;
483
484 BUG_ON(!llist_empty(&dc->writeback_ordering_wait.list));
485 atomic_set(&dc->writeback_sequence_next, sequence);
486 closure_init_stack(&cl);
487
488 /*
489 * XXX: if we error, background writeback just spins. Should use some
490 * mempools.
491 */
492
493 next = bch_keybuf_next(&dc->writeback_keys);
494
495 while (!kthread_should_stop() &&
496 !test_bit(CACHE_SET_IO_DISABLE, &dc->disk.c->flags) &&
497 next) {
498 size = 0;
499 nk = 0;
500
501 do {
502 BUG_ON(ptr_stale(dc->disk.c, &next->key, 0));
503
504 /*
505 * Don't combine too many operations, even if they
506 * are all small.
507 */
508 if (nk >= MAX_WRITEBACKS_IN_PASS)
509 break;
510
511 /*
512 * If the current operation is very large, don't
513 * further combine operations.
514 */
515 if (size >= MAX_WRITESIZE_IN_PASS)
516 break;
517
518 /*
519 * Operations are only eligible to be combined
520 * if they are contiguous.
521 *
522 * TODO: add a heuristic willing to fire a
523 * certain amount of non-contiguous IO per pass,
524 * so that we can benefit from backing device
525 * command queueing.
526 */
527 if ((nk != 0) && bkey_cmp(&keys[nk-1]->key,
528 &START_KEY(&next->key)))
529 break;
530
531 size += KEY_SIZE(&next->key);
532 keys[nk++] = next;
533 } while ((next = bch_keybuf_next(&dc->writeback_keys)));
534
535 /* Now we have gathered a set of 1..5 keys to write back. */
536 for (i = 0; i < nk; i++) {
537 w = keys[i];
538
539 io = kzalloc(struct_size(io, bio.bi_inline_vecs,
540 DIV_ROUND_UP(KEY_SIZE(&w->key), PAGE_SECTORS)),
541 GFP_KERNEL);
542 if (!io)
543 goto err;
544
545 w->private = io;
546 io->dc = dc;
547 io->sequence = sequence++;
548
549 dirty_init(w);
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;
554
555 if (bch_bio_alloc_pages(&io->bio, GFP_KERNEL))
556 goto err_free;
557
558 trace_bcache_writeback(&w->key);
559
560 down(&dc->in_flight);
561
562 /*
563 * We've acquired a semaphore for the maximum
564 * simultaneous number of writebacks; from here
565 * everything happens asynchronously.
566 */
567 closure_call(&io->cl, read_dirty_submit, NULL, &cl);
568 }
569
570 delay = writeback_delay(dc, size);
571
572 while (!kthread_should_stop() &&
573 !test_bit(CACHE_SET_IO_DISABLE, &dc->disk.c->flags) &&
574 delay) {
575 schedule_timeout_interruptible(delay);
576 delay = writeback_delay(dc, 0);
577 }
578 }
579
580 if (0) {
581err_free:
582 kfree(w->private);
583err:
584 bch_keybuf_del(&dc->writeback_keys, w);
585 }
586
587 /*
588 * Wait for outstanding writeback IOs to finish (and keybuf slots to be
589 * freed) before refilling again
590 */
591 closure_sync(&cl);
592}
593
594/* Scan for dirty data */
595
596void bcache_dev_sectors_dirty_add(struct cache_set *c, unsigned int inode,
597 uint64_t offset, int nr_sectors)
598{
599 struct bcache_device *d = c->devices[inode];
600 unsigned int stripe_offset, sectors_dirty;
601 int stripe;
602
603 if (!d)
604 return;
605
606 stripe = offset_to_stripe(d, offset);
607 if (stripe < 0)
608 return;
609
610 if (UUID_FLASH_ONLY(&c->uuids[inode]))
611 atomic_long_add(nr_sectors, &c->flash_dev_dirty_sectors);
612
613 stripe_offset = offset & (d->stripe_size - 1);
614
615 while (nr_sectors) {
616 int s = min_t(unsigned int, abs(nr_sectors),
617 d->stripe_size - stripe_offset);
618
619 if (nr_sectors < 0)
620 s = -s;
621
622 if (stripe >= d->nr_stripes)
623 return;
624
625 sectors_dirty = atomic_add_return(s,
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);
630 } else {
631 if (test_bit(stripe, d->full_dirty_stripes))
632 clear_bit(stripe, d->full_dirty_stripes);
633 }
634
635 nr_sectors -= s;
636 stripe_offset = 0;
637 stripe++;
638 }
639}
640
641static bool dirty_pred(struct keybuf *buf, struct bkey *k)
642{
643 struct cached_dev *dc = container_of(buf,
644 struct cached_dev,
645 writeback_keys);
646
647 BUG_ON(KEY_INODE(k) != dc->disk.id);
648
649 return KEY_DIRTY(k);
650}
651
652static void refill_full_stripes(struct cached_dev *dc)
653{
654 struct keybuf *buf = &dc->writeback_keys;
655 unsigned int start_stripe, next_stripe;
656 int stripe;
657 bool wrapped = false;
658
659 stripe = offset_to_stripe(&dc->disk, KEY_OFFSET(&buf->last_scanned));
660 if (stripe < 0)
661 stripe = 0;
662
663 start_stripe = stripe;
664
665 while (1) {
666 stripe = find_next_bit(dc->disk.full_dirty_stripes,
667 dc->disk.nr_stripes, stripe);
668
669 if (stripe == dc->disk.nr_stripes)
670 goto next;
671
672 next_stripe = find_next_zero_bit(dc->disk.full_dirty_stripes,
673 dc->disk.nr_stripes, stripe);
674
675 buf->last_scanned = KEY(dc->disk.id,
676 stripe * dc->disk.stripe_size, 0);
677
678 bch_refill_keybuf(dc->disk.c, buf,
679 &KEY(dc->disk.id,
680 next_stripe * dc->disk.stripe_size, 0),
681 dirty_pred);
682
683 if (array_freelist_empty(&buf->freelist))
684 return;
685
686 stripe = next_stripe;
687next:
688 if (wrapped && stripe > start_stripe)
689 return;
690
691 if (stripe == dc->disk.nr_stripes) {
692 stripe = 0;
693 wrapped = true;
694 }
695 }
696}
697
698/*
699 * Returns true if we scanned the entire disk
700 */
701static bool refill_dirty(struct cached_dev *dc)
702{
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);
706 struct bkey start_pos;
707
708 /*
709 * make sure keybuf pos is inside the range for this disk - at bringup
710 * we might not be attached yet so this disk's inode nr isn't
711 * initialized then
712 */
713 if (bkey_cmp(&buf->last_scanned, &start) < 0 ||
714 bkey_cmp(&buf->last_scanned, &end) > 0)
715 buf->last_scanned = start;
716
717 if (dc->partial_stripes_expensive) {
718 refill_full_stripes(dc);
719 if (array_freelist_empty(&buf->freelist))
720 return false;
721 }
722
723 start_pos = buf->last_scanned;
724 bch_refill_keybuf(dc->disk.c, buf, &end, dirty_pred);
725
726 if (bkey_cmp(&buf->last_scanned, &end) < 0)
727 return false;
728
729 /*
730 * If we get to the end start scanning again from the beginning, and
731 * only scan up to where we initially started scanning from:
732 */
733 buf->last_scanned = start;
734 bch_refill_keybuf(dc->disk.c, buf, &start_pos, dirty_pred);
735
736 return bkey_cmp(&buf->last_scanned, &start_pos) >= 0;
737}
738
739static int bch_writeback_thread(void *arg)
740{
741 struct cached_dev *dc = arg;
742 struct cache_set *c = dc->disk.c;
743 bool searched_full_index;
744
745 bch_ratelimit_reset(&dc->writeback_rate);
746
747 while (!kthread_should_stop() &&
748 !test_bit(CACHE_SET_IO_DISABLE, &c->flags)) {
749 down_write(&dc->writeback_lock);
750 set_current_state(TASK_INTERRUPTIBLE);
751 /*
752 * If the bache device is detaching, skip here and continue
753 * to perform writeback. Otherwise, if no dirty data on cache,
754 * or there is dirty data on cache but writeback is disabled,
755 * the writeback thread should sleep here and wait for others
756 * to wake up it.
757 */
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);
761
762 if (kthread_should_stop() ||
763 test_bit(CACHE_SET_IO_DISABLE, &c->flags)) {
764 set_current_state(TASK_RUNNING);
765 break;
766 }
767
768 schedule();
769 continue;
770 }
771 set_current_state(TASK_RUNNING);
772
773 searched_full_index = refill_dirty(dc);
774
775 if (searched_full_index &&
776 RB_EMPTY_ROOT(&dc->writeback_keys.keys)) {
777 atomic_set(&dc->has_dirty, 0);
778 SET_BDEV_STATE(&dc->sb, BDEV_STATE_CLEAN);
779 bch_write_bdev_super(dc, NULL);
780 /*
781 * If bcache device is detaching via sysfs interface,
782 * writeback thread should stop after there is no dirty
783 * data on cache. BCACHE_DEV_DETACHING flag is set in
784 * bch_cached_dev_detach().
785 */
786 if (test_bit(BCACHE_DEV_DETACHING, &dc->disk.flags)) {
787 struct closure cl;
788
789 closure_init_stack(&cl);
790 memset(&dc->sb.set_uuid, 0, 16);
791 SET_BDEV_STATE(&dc->sb, BDEV_STATE_NONE);
792
793 bch_write_bdev_super(dc, &cl);
794 closure_sync(&cl);
795
796 up_write(&dc->writeback_lock);
797 break;
798 }
799
800 /*
801 * When dirty data rate is high (e.g. 50%+), there might
802 * be heavy buckets fragmentation after writeback
803 * finished, which hurts following write performance.
804 * If users really care about write performance they
805 * may set BCH_ENABLE_AUTO_GC via sysfs, then when
806 * BCH_DO_AUTO_GC is set, garbage collection thread
807 * will be wake up here. After moving gc, the shrunk
808 * btree and discarded free buckets SSD space may be
809 * helpful for following write requests.
810 */
811 if (c->gc_after_writeback ==
812 (BCH_ENABLE_AUTO_GC|BCH_DO_AUTO_GC)) {
813 c->gc_after_writeback &= ~BCH_DO_AUTO_GC;
814 force_wake_up_gc(c);
815 }
816 }
817
818 up_write(&dc->writeback_lock);
819
820 read_dirty(dc);
821
822 if (searched_full_index) {
823 unsigned int delay = dc->writeback_delay * HZ;
824
825 while (delay &&
826 !kthread_should_stop() &&
827 !test_bit(CACHE_SET_IO_DISABLE, &c->flags) &&
828 !test_bit(BCACHE_DEV_DETACHING, &dc->disk.flags))
829 delay = schedule_timeout_interruptible(delay);
830
831 bch_ratelimit_reset(&dc->writeback_rate);
832 }
833 }
834
835 if (dc->writeback_write_wq)
836 destroy_workqueue(dc->writeback_write_wq);
837
838 cached_dev_put(dc);
839 wait_for_kthread_stop();
840
841 return 0;
842}
843
844/* Init */
845#define INIT_KEYS_EACH_TIME 500000
846
847struct sectors_dirty_init {
848 struct btree_op op;
849 unsigned int inode;
850 size_t count;
851};
852
853static int sectors_dirty_init_fn(struct btree_op *_op, struct btree *b,
854 struct bkey *k)
855{
856 struct sectors_dirty_init *op = container_of(_op,
857 struct sectors_dirty_init, op);
858 if (KEY_INODE(k) > op->inode)
859 return MAP_DONE;
860
861 if (KEY_DIRTY(k))
862 bcache_dev_sectors_dirty_add(b->c, KEY_INODE(k),
863 KEY_START(k), KEY_SIZE(k));
864
865 op->count++;
866 if (!(op->count % INIT_KEYS_EACH_TIME))
867 cond_resched();
868
869 return MAP_CONTINUE;
870}
871
872static int bch_root_node_dirty_init(struct cache_set *c,
873 struct bcache_device *d,
874 struct bkey *k)
875{
876 struct sectors_dirty_init op;
877 int ret;
878
879 bch_btree_op_init(&op.op, -1);
880 op.inode = d->id;
881 op.count = 0;
882
883 ret = bcache_btree(map_keys_recurse,
884 k,
885 c->root,
886 &op.op,
887 &KEY(op.inode, 0, 0),
888 sectors_dirty_init_fn,
889 0);
890 if (ret < 0)
891 pr_warn("sectors dirty init failed, ret=%d!\n", ret);
892
893 /*
894 * The op may be added to cache_set's btree_cache_wait
895 * in mca_cannibalize(), must ensure it is removed from
896 * the list and release btree_cache_alloc_lock before
897 * free op memory.
898 * Otherwise, the btree_cache_wait will be damaged.
899 */
900 bch_cannibalize_unlock(c);
901 finish_wait(&c->btree_cache_wait, &(&op.op)->wait);
902
903 return ret;
904}
905
906static int bch_dirty_init_thread(void *arg)
907{
908 struct dirty_init_thrd_info *info = arg;
909 struct bch_dirty_init_state *state = info->state;
910 struct cache_set *c = state->c;
911 struct btree_iter iter;
912 struct bkey *k, *p;
913 int cur_idx, prev_idx, skip_nr;
914
915 k = p = NULL;
916 prev_idx = 0;
917
918 bch_btree_iter_init(&c->root->keys, &iter, NULL);
919 k = bch_btree_iter_next_filter(&iter, &c->root->keys, bch_ptr_bad);
920 BUG_ON(!k);
921
922 p = k;
923
924 while (k) {
925 spin_lock(&state->idx_lock);
926 cur_idx = state->key_idx;
927 state->key_idx++;
928 spin_unlock(&state->idx_lock);
929
930 skip_nr = cur_idx - prev_idx;
931
932 while (skip_nr) {
933 k = bch_btree_iter_next_filter(&iter,
934 &c->root->keys,
935 bch_ptr_bad);
936 if (k)
937 p = k;
938 else {
939 atomic_set(&state->enough, 1);
940 /* Update state->enough earlier */
941 smp_mb__after_atomic();
942 goto out;
943 }
944 skip_nr--;
945 }
946
947 if (p) {
948 if (bch_root_node_dirty_init(c, state->d, p) < 0)
949 goto out;
950 }
951
952 p = NULL;
953 prev_idx = cur_idx;
954 }
955
956out:
957 /* In order to wake up state->wait in time */
958 smp_mb__before_atomic();
959 if (atomic_dec_and_test(&state->started))
960 wake_up(&state->wait);
961
962 return 0;
963}
964
965static int bch_btre_dirty_init_thread_nr(void)
966{
967 int n = num_online_cpus()/2;
968
969 if (n == 0)
970 n = 1;
971 else if (n > BCH_DIRTY_INIT_THRD_MAX)
972 n = BCH_DIRTY_INIT_THRD_MAX;
973
974 return n;
975}
976
977void bch_sectors_dirty_init(struct bcache_device *d)
978{
979 int i;
980 struct btree *b = NULL;
981 struct bkey *k = NULL;
982 struct btree_iter iter;
983 struct sectors_dirty_init op;
984 struct cache_set *c = d->c;
985 struct bch_dirty_init_state state;
986
987retry_lock:
988 b = c->root;
989 rw_lock(0, b, b->level);
990 if (b != c->root) {
991 rw_unlock(0, b);
992 goto retry_lock;
993 }
994
995 /* Just count root keys if no leaf node */
996 if (c->root->level == 0) {
997 bch_btree_op_init(&op.op, -1);
998 op.inode = d->id;
999 op.count = 0;
1000
1001 for_each_key_filter(&c->root->keys,
1002 k, &iter, bch_ptr_invalid) {
1003 if (KEY_INODE(k) != op.inode)
1004 continue;
1005 sectors_dirty_init_fn(&op.op, c->root, k);
1006 }
1007
1008 rw_unlock(0, b);
1009 return;
1010 }
1011
1012 memset(&state, 0, sizeof(struct bch_dirty_init_state));
1013 state.c = c;
1014 state.d = d;
1015 state.total_threads = bch_btre_dirty_init_thread_nr();
1016 state.key_idx = 0;
1017 spin_lock_init(&state.idx_lock);
1018 atomic_set(&state.started, 0);
1019 atomic_set(&state.enough, 0);
1020 init_waitqueue_head(&state.wait);
1021
1022 for (i = 0; i < state.total_threads; i++) {
1023 /* Fetch latest state.enough earlier */
1024 smp_mb__before_atomic();
1025 if (atomic_read(&state.enough))
1026 break;
1027
1028 atomic_inc(&state.started);
1029 state.infos[i].state = &state;
1030 state.infos[i].thread =
1031 kthread_run(bch_dirty_init_thread, &state.infos[i],
1032 "bch_dirtcnt[%d]", i);
1033 if (IS_ERR(state.infos[i].thread)) {
1034 pr_err("fails to run thread bch_dirty_init[%d]\n", i);
1035 atomic_dec(&state.started);
1036 for (--i; i >= 0; i--)
1037 kthread_stop(state.infos[i].thread);
1038 goto out;
1039 }
1040 }
1041
1042out:
1043 /* Must wait for all threads to stop. */
1044 wait_event(state.wait, atomic_read(&state.started) == 0);
1045 rw_unlock(0, b);
1046}
1047
1048void bch_cached_dev_writeback_init(struct cached_dev *dc)
1049{
1050 sema_init(&dc->in_flight, 64);
1051 init_rwsem(&dc->writeback_lock);
1052 bch_keybuf_init(&dc->writeback_keys);
1053
1054 dc->writeback_metadata = true;
1055 dc->writeback_running = false;
1056 dc->writeback_consider_fragment = true;
1057 dc->writeback_percent = 10;
1058 dc->writeback_delay = 30;
1059 atomic_long_set(&dc->writeback_rate.rate, 1024);
1060 dc->writeback_rate_minimum = 8;
1061
1062 dc->writeback_rate_update_seconds = WRITEBACK_RATE_UPDATE_SECS_DEFAULT;
1063 dc->writeback_rate_p_term_inverse = 40;
1064 dc->writeback_rate_fp_term_low = 1;
1065 dc->writeback_rate_fp_term_mid = 10;
1066 dc->writeback_rate_fp_term_high = 1000;
1067 dc->writeback_rate_i_term_inverse = 10000;
1068
1069 /* For dc->writeback_lock contention in update_writeback_rate() */
1070 dc->rate_update_retry = 0;
1071
1072 WARN_ON(test_and_clear_bit(BCACHE_DEV_WB_RUNNING, &dc->disk.flags));
1073 INIT_DELAYED_WORK(&dc->writeback_rate_update, update_writeback_rate);
1074}
1075
1076int bch_cached_dev_writeback_start(struct cached_dev *dc)
1077{
1078 dc->writeback_write_wq = alloc_workqueue("bcache_writeback_wq",
1079 WQ_MEM_RECLAIM, 0);
1080 if (!dc->writeback_write_wq)
1081 return -ENOMEM;
1082
1083 cached_dev_get(dc);
1084 dc->writeback_thread = kthread_create(bch_writeback_thread, dc,
1085 "bcache_writeback");
1086 if (IS_ERR(dc->writeback_thread)) {
1087 cached_dev_put(dc);
1088 destroy_workqueue(dc->writeback_write_wq);
1089 return PTR_ERR(dc->writeback_thread);
1090 }
1091 dc->writeback_running = true;
1092
1093 WARN_ON(test_and_set_bit(BCACHE_DEV_WB_RUNNING, &dc->disk.flags));
1094 schedule_delayed_work(&dc->writeback_rate_update,
1095 dc->writeback_rate_update_seconds * HZ);
1096
1097 bch_writeback_queue(dc);
1098
1099 return 0;
1100}
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * background writeback - scan btree for dirty data and write it to the backing
4 * device
5 *
6 * Copyright 2010, 2011 Kent Overstreet <kent.overstreet@gmail.com>
7 * Copyright 2012 Google, Inc.
8 */
9
10#include "bcache.h"
11#include "btree.h"
12#include "debug.h"
13#include "writeback.h"
14
15#include <linux/delay.h>
16#include <linux/kthread.h>
17#include <linux/sched/clock.h>
18#include <trace/events/bcache.h>
19
20static void update_gc_after_writeback(struct cache_set *c)
21{
22 if (c->gc_after_writeback != (BCH_ENABLE_AUTO_GC) ||
23 c->gc_stats.in_use < BCH_AUTO_GC_DIRTY_THRESHOLD)
24 return;
25
26 c->gc_after_writeback |= BCH_DO_AUTO_GC;
27}
28
29/* Rate limiting */
30static uint64_t __calc_target_rate(struct cached_dev *dc)
31{
32 struct cache_set *c = dc->disk.c;
33
34 /*
35 * This is the size of the cache, minus the amount used for
36 * flash-only devices
37 */
38 uint64_t cache_sectors = c->nbuckets * c->sb.bucket_size -
39 atomic_long_read(&c->flash_dev_dirty_sectors);
40
41 /*
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.
46 */
47 uint32_t bdev_share =
48 div64_u64(bdev_sectors(dc->bdev) << WRITEBACK_SHARE_SHIFT,
49 c->cached_dev_sectors);
50
51 uint64_t cache_dirty_target =
52 div_u64(cache_sectors * dc->writeback_percent, 100);
53
54 /* Ensure each backing dev gets at least one dirty share */
55 if (bdev_share < 1)
56 bdev_share = 1;
57
58 return (cache_dirty_target * bdev_share) >> WRITEBACK_SHARE_SHIFT;
59}
60
61static void __update_writeback_rate(struct cached_dev *dc)
62{
63 /*
64 * PI controller:
65 * Figures out the amount that should be written per second.
66 *
67 * First, the error (number of sectors that are dirty beyond our
68 * target) is calculated. The error is accumulated (numerically
69 * integrated).
70 *
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.
77 *
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.
82 */
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;
89 uint32_t new_rate;
90
91 if ((error < 0 && dc->writeback_rate_integral > 0) ||
92 (error > 0 && time_before64(local_clock(),
93 dc->writeback_rate.next + NSEC_PER_MSEC))) {
94 /*
95 * Only decrease the integral term if it's more than
96 * zero. Only increase the integral term if the device
97 * is keeping up. (Don't wind up the integral
98 * ineffectively in either case).
99 *
100 * It's necessary to scale this by
101 * writeback_rate_update_seconds to keep the integral
102 * term dimensioned properly.
103 */
104 dc->writeback_rate_integral += error *
105 dc->writeback_rate_update_seconds;
106 }
107
108 integral_scaled = div_s64(dc->writeback_rate_integral,
109 dc->writeback_rate_i_term_inverse);
110
111 new_rate = clamp_t(int32_t, (proportional_scaled + integral_scaled),
112 dc->writeback_rate_minimum, NSEC_PER_SEC);
113
114 dc->writeback_rate_proportional = proportional_scaled;
115 dc->writeback_rate_integral_scaled = integral_scaled;
116 dc->writeback_rate_change = new_rate -
117 atomic_long_read(&dc->writeback_rate.rate);
118 atomic_long_set(&dc->writeback_rate.rate, new_rate);
119 dc->writeback_rate_target = target;
120}
121
122static bool set_at_max_writeback_rate(struct cache_set *c,
123 struct cached_dev *dc)
124{
125 /* Don't sst max writeback rate if it is disabled */
126 if (!c->idle_max_writeback_rate_enabled)
127 return false;
128
129 /* Don't set max writeback rate if gc is running */
130 if (!c->gc_mark_valid)
131 return false;
132 /*
133 * Idle_counter is increased everytime when update_writeback_rate() is
134 * called. If all backing devices attached to the same cache set have
135 * identical dc->writeback_rate_update_seconds values, it is about 6
136 * rounds of update_writeback_rate() on each backing device before
137 * c->at_max_writeback_rate is set to 1, and then max wrteback rate set
138 * to each dc->writeback_rate.rate.
139 * In order to avoid extra locking cost for counting exact dirty cached
140 * devices number, c->attached_dev_nr is used to calculate the idle
141 * throushold. It might be bigger if not all cached device are in write-
142 * back mode, but it still works well with limited extra rounds of
143 * update_writeback_rate().
144 */
145 if (atomic_inc_return(&c->idle_counter) <
146 atomic_read(&c->attached_dev_nr) * 6)
147 return false;
148
149 if (atomic_read(&c->at_max_writeback_rate) != 1)
150 atomic_set(&c->at_max_writeback_rate, 1);
151
152 atomic_long_set(&dc->writeback_rate.rate, INT_MAX);
153
154 /* keep writeback_rate_target as existing value */
155 dc->writeback_rate_proportional = 0;
156 dc->writeback_rate_integral_scaled = 0;
157 dc->writeback_rate_change = 0;
158
159 /*
160 * Check c->idle_counter and c->at_max_writeback_rate agagain in case
161 * new I/O arrives during before set_at_max_writeback_rate() returns.
162 * Then the writeback rate is set to 1, and its new value should be
163 * decided via __update_writeback_rate().
164 */
165 if ((atomic_read(&c->idle_counter) <
166 atomic_read(&c->attached_dev_nr) * 6) ||
167 !atomic_read(&c->at_max_writeback_rate))
168 return false;
169
170 return true;
171}
172
173static void update_writeback_rate(struct work_struct *work)
174{
175 struct cached_dev *dc = container_of(to_delayed_work(work),
176 struct cached_dev,
177 writeback_rate_update);
178 struct cache_set *c = dc->disk.c;
179
180 /*
181 * should check BCACHE_DEV_RATE_DW_RUNNING before calling
182 * cancel_delayed_work_sync().
183 */
184 set_bit(BCACHE_DEV_RATE_DW_RUNNING, &dc->disk.flags);
185 /* paired with where BCACHE_DEV_RATE_DW_RUNNING is tested */
186 smp_mb__after_atomic();
187
188 /*
189 * CACHE_SET_IO_DISABLE might be set via sysfs interface,
190 * check it here too.
191 */
192 if (!test_bit(BCACHE_DEV_WB_RUNNING, &dc->disk.flags) ||
193 test_bit(CACHE_SET_IO_DISABLE, &c->flags)) {
194 clear_bit(BCACHE_DEV_RATE_DW_RUNNING, &dc->disk.flags);
195 /* paired with where BCACHE_DEV_RATE_DW_RUNNING is tested */
196 smp_mb__after_atomic();
197 return;
198 }
199
200 if (atomic_read(&dc->has_dirty) && dc->writeback_percent) {
201 /*
202 * If the whole cache set is idle, set_at_max_writeback_rate()
203 * will set writeback rate to a max number. Then it is
204 * unncessary to update writeback rate for an idle cache set
205 * in maximum writeback rate number(s).
206 */
207 if (!set_at_max_writeback_rate(c, dc)) {
208 down_read(&dc->writeback_lock);
209 __update_writeback_rate(dc);
210 update_gc_after_writeback(c);
211 up_read(&dc->writeback_lock);
212 }
213 }
214
215
216 /*
217 * CACHE_SET_IO_DISABLE might be set via sysfs interface,
218 * check it here too.
219 */
220 if (test_bit(BCACHE_DEV_WB_RUNNING, &dc->disk.flags) &&
221 !test_bit(CACHE_SET_IO_DISABLE, &c->flags)) {
222 schedule_delayed_work(&dc->writeback_rate_update,
223 dc->writeback_rate_update_seconds * HZ);
224 }
225
226 /*
227 * should check BCACHE_DEV_RATE_DW_RUNNING before calling
228 * cancel_delayed_work_sync().
229 */
230 clear_bit(BCACHE_DEV_RATE_DW_RUNNING, &dc->disk.flags);
231 /* paired with where BCACHE_DEV_RATE_DW_RUNNING is tested */
232 smp_mb__after_atomic();
233}
234
235static unsigned int writeback_delay(struct cached_dev *dc,
236 unsigned int sectors)
237{
238 if (test_bit(BCACHE_DEV_DETACHING, &dc->disk.flags) ||
239 !dc->writeback_percent)
240 return 0;
241
242 return bch_next_delay(&dc->writeback_rate, sectors);
243}
244
245struct dirty_io {
246 struct closure cl;
247 struct cached_dev *dc;
248 uint16_t sequence;
249 struct bio bio;
250};
251
252static void dirty_init(struct keybuf_key *w)
253{
254 struct dirty_io *io = w->private;
255 struct bio *bio = &io->bio;
256
257 bio_init(bio, bio->bi_inline_vecs,
258 DIV_ROUND_UP(KEY_SIZE(&w->key), PAGE_SECTORS));
259 if (!io->dc->writeback_percent)
260 bio_set_prio(bio, IOPRIO_PRIO_VALUE(IOPRIO_CLASS_IDLE, 0));
261
262 bio->bi_iter.bi_size = KEY_SIZE(&w->key) << 9;
263 bio->bi_private = w;
264 bch_bio_map(bio, NULL);
265}
266
267static void dirty_io_destructor(struct closure *cl)
268{
269 struct dirty_io *io = container_of(cl, struct dirty_io, cl);
270
271 kfree(io);
272}
273
274static void write_dirty_finish(struct closure *cl)
275{
276 struct dirty_io *io = container_of(cl, struct dirty_io, cl);
277 struct keybuf_key *w = io->bio.bi_private;
278 struct cached_dev *dc = io->dc;
279
280 bio_free_pages(&io->bio);
281
282 /* This is kind of a dumb way of signalling errors. */
283 if (KEY_DIRTY(&w->key)) {
284 int ret;
285 unsigned int i;
286 struct keylist keys;
287
288 bch_keylist_init(&keys);
289
290 bkey_copy(keys.top, &w->key);
291 SET_KEY_DIRTY(keys.top, false);
292 bch_keylist_push(&keys);
293
294 for (i = 0; i < KEY_PTRS(&w->key); i++)
295 atomic_inc(&PTR_BUCKET(dc->disk.c, &w->key, i)->pin);
296
297 ret = bch_btree_insert(dc->disk.c, &keys, NULL, &w->key);
298
299 if (ret)
300 trace_bcache_writeback_collision(&w->key);
301
302 atomic_long_inc(ret
303 ? &dc->disk.c->writeback_keys_failed
304 : &dc->disk.c->writeback_keys_done);
305 }
306
307 bch_keybuf_del(&dc->writeback_keys, w);
308 up(&dc->in_flight);
309
310 closure_return_with_destructor(cl, dirty_io_destructor);
311}
312
313static void dirty_endio(struct bio *bio)
314{
315 struct keybuf_key *w = bio->bi_private;
316 struct dirty_io *io = w->private;
317
318 if (bio->bi_status) {
319 SET_KEY_DIRTY(&w->key, false);
320 bch_count_backing_io_errors(io->dc, bio);
321 }
322
323 closure_put(&io->cl);
324}
325
326static void write_dirty(struct closure *cl)
327{
328 struct dirty_io *io = container_of(cl, struct dirty_io, cl);
329 struct keybuf_key *w = io->bio.bi_private;
330 struct cached_dev *dc = io->dc;
331
332 uint16_t next_sequence;
333
334 if (atomic_read(&dc->writeback_sequence_next) != io->sequence) {
335 /* Not our turn to write; wait for a write to complete */
336 closure_wait(&dc->writeback_ordering_wait, cl);
337
338 if (atomic_read(&dc->writeback_sequence_next) == io->sequence) {
339 /*
340 * Edge case-- it happened in indeterminate order
341 * relative to when we were added to wait list..
342 */
343 closure_wake_up(&dc->writeback_ordering_wait);
344 }
345
346 continue_at(cl, write_dirty, io->dc->writeback_write_wq);
347 return;
348 }
349
350 next_sequence = io->sequence + 1;
351
352 /*
353 * IO errors are signalled using the dirty bit on the key.
354 * If we failed to read, we should not attempt to write to the
355 * backing device. Instead, immediately go to write_dirty_finish
356 * to clean up.
357 */
358 if (KEY_DIRTY(&w->key)) {
359 dirty_init(w);
360 bio_set_op_attrs(&io->bio, REQ_OP_WRITE, 0);
361 io->bio.bi_iter.bi_sector = KEY_START(&w->key);
362 bio_set_dev(&io->bio, io->dc->bdev);
363 io->bio.bi_end_io = dirty_endio;
364
365 /* I/O request sent to backing device */
366 closure_bio_submit(io->dc->disk.c, &io->bio, cl);
367 }
368
369 atomic_set(&dc->writeback_sequence_next, next_sequence);
370 closure_wake_up(&dc->writeback_ordering_wait);
371
372 continue_at(cl, write_dirty_finish, io->dc->writeback_write_wq);
373}
374
375static void read_dirty_endio(struct bio *bio)
376{
377 struct keybuf_key *w = bio->bi_private;
378 struct dirty_io *io = w->private;
379
380 /* is_read = 1 */
381 bch_count_io_errors(PTR_CACHE(io->dc->disk.c, &w->key, 0),
382 bio->bi_status, 1,
383 "reading dirty data from cache");
384
385 dirty_endio(bio);
386}
387
388static void read_dirty_submit(struct closure *cl)
389{
390 struct dirty_io *io = container_of(cl, struct dirty_io, cl);
391
392 closure_bio_submit(io->dc->disk.c, &io->bio, cl);
393
394 continue_at(cl, write_dirty, io->dc->writeback_write_wq);
395}
396
397static void read_dirty(struct cached_dev *dc)
398{
399 unsigned int delay = 0;
400 struct keybuf_key *next, *keys[MAX_WRITEBACKS_IN_PASS], *w;
401 size_t size;
402 int nk, i;
403 struct dirty_io *io;
404 struct closure cl;
405 uint16_t sequence = 0;
406
407 BUG_ON(!llist_empty(&dc->writeback_ordering_wait.list));
408 atomic_set(&dc->writeback_sequence_next, sequence);
409 closure_init_stack(&cl);
410
411 /*
412 * XXX: if we error, background writeback just spins. Should use some
413 * mempools.
414 */
415
416 next = bch_keybuf_next(&dc->writeback_keys);
417
418 while (!kthread_should_stop() &&
419 !test_bit(CACHE_SET_IO_DISABLE, &dc->disk.c->flags) &&
420 next) {
421 size = 0;
422 nk = 0;
423
424 do {
425 BUG_ON(ptr_stale(dc->disk.c, &next->key, 0));
426
427 /*
428 * Don't combine too many operations, even if they
429 * are all small.
430 */
431 if (nk >= MAX_WRITEBACKS_IN_PASS)
432 break;
433
434 /*
435 * If the current operation is very large, don't
436 * further combine operations.
437 */
438 if (size >= MAX_WRITESIZE_IN_PASS)
439 break;
440
441 /*
442 * Operations are only eligible to be combined
443 * if they are contiguous.
444 *
445 * TODO: add a heuristic willing to fire a
446 * certain amount of non-contiguous IO per pass,
447 * so that we can benefit from backing device
448 * command queueing.
449 */
450 if ((nk != 0) && bkey_cmp(&keys[nk-1]->key,
451 &START_KEY(&next->key)))
452 break;
453
454 size += KEY_SIZE(&next->key);
455 keys[nk++] = next;
456 } while ((next = bch_keybuf_next(&dc->writeback_keys)));
457
458 /* Now we have gathered a set of 1..5 keys to write back. */
459 for (i = 0; i < nk; i++) {
460 w = keys[i];
461
462 io = kzalloc(struct_size(io, bio.bi_inline_vecs,
463 DIV_ROUND_UP(KEY_SIZE(&w->key), PAGE_SECTORS)),
464 GFP_KERNEL);
465 if (!io)
466 goto err;
467
468 w->private = io;
469 io->dc = dc;
470 io->sequence = sequence++;
471
472 dirty_init(w);
473 bio_set_op_attrs(&io->bio, REQ_OP_READ, 0);
474 io->bio.bi_iter.bi_sector = PTR_OFFSET(&w->key, 0);
475 bio_set_dev(&io->bio,
476 PTR_CACHE(dc->disk.c, &w->key, 0)->bdev);
477 io->bio.bi_end_io = read_dirty_endio;
478
479 if (bch_bio_alloc_pages(&io->bio, GFP_KERNEL))
480 goto err_free;
481
482 trace_bcache_writeback(&w->key);
483
484 down(&dc->in_flight);
485
486 /*
487 * We've acquired a semaphore for the maximum
488 * simultaneous number of writebacks; from here
489 * everything happens asynchronously.
490 */
491 closure_call(&io->cl, read_dirty_submit, NULL, &cl);
492 }
493
494 delay = writeback_delay(dc, size);
495
496 while (!kthread_should_stop() &&
497 !test_bit(CACHE_SET_IO_DISABLE, &dc->disk.c->flags) &&
498 delay) {
499 schedule_timeout_interruptible(delay);
500 delay = writeback_delay(dc, 0);
501 }
502 }
503
504 if (0) {
505err_free:
506 kfree(w->private);
507err:
508 bch_keybuf_del(&dc->writeback_keys, w);
509 }
510
511 /*
512 * Wait for outstanding writeback IOs to finish (and keybuf slots to be
513 * freed) before refilling again
514 */
515 closure_sync(&cl);
516}
517
518/* Scan for dirty data */
519
520void bcache_dev_sectors_dirty_add(struct cache_set *c, unsigned int inode,
521 uint64_t offset, int nr_sectors)
522{
523 struct bcache_device *d = c->devices[inode];
524 unsigned int stripe_offset, sectors_dirty;
525 int stripe;
526
527 if (!d)
528 return;
529
530 stripe = offset_to_stripe(d, offset);
531 if (stripe < 0)
532 return;
533
534 if (UUID_FLASH_ONLY(&c->uuids[inode]))
535 atomic_long_add(nr_sectors, &c->flash_dev_dirty_sectors);
536
537 stripe_offset = offset & (d->stripe_size - 1);
538
539 while (nr_sectors) {
540 int s = min_t(unsigned int, abs(nr_sectors),
541 d->stripe_size - stripe_offset);
542
543 if (nr_sectors < 0)
544 s = -s;
545
546 if (stripe >= d->nr_stripes)
547 return;
548
549 sectors_dirty = atomic_add_return(s,
550 d->stripe_sectors_dirty + stripe);
551 if (sectors_dirty == d->stripe_size)
552 set_bit(stripe, d->full_dirty_stripes);
553 else
554 clear_bit(stripe, d->full_dirty_stripes);
555
556 nr_sectors -= s;
557 stripe_offset = 0;
558 stripe++;
559 }
560}
561
562static bool dirty_pred(struct keybuf *buf, struct bkey *k)
563{
564 struct cached_dev *dc = container_of(buf,
565 struct cached_dev,
566 writeback_keys);
567
568 BUG_ON(KEY_INODE(k) != dc->disk.id);
569
570 return KEY_DIRTY(k);
571}
572
573static void refill_full_stripes(struct cached_dev *dc)
574{
575 struct keybuf *buf = &dc->writeback_keys;
576 unsigned int start_stripe, next_stripe;
577 int stripe;
578 bool wrapped = false;
579
580 stripe = offset_to_stripe(&dc->disk, KEY_OFFSET(&buf->last_scanned));
581 if (stripe < 0)
582 stripe = 0;
583
584 start_stripe = stripe;
585
586 while (1) {
587 stripe = find_next_bit(dc->disk.full_dirty_stripes,
588 dc->disk.nr_stripes, stripe);
589
590 if (stripe == dc->disk.nr_stripes)
591 goto next;
592
593 next_stripe = find_next_zero_bit(dc->disk.full_dirty_stripes,
594 dc->disk.nr_stripes, stripe);
595
596 buf->last_scanned = KEY(dc->disk.id,
597 stripe * dc->disk.stripe_size, 0);
598
599 bch_refill_keybuf(dc->disk.c, buf,
600 &KEY(dc->disk.id,
601 next_stripe * dc->disk.stripe_size, 0),
602 dirty_pred);
603
604 if (array_freelist_empty(&buf->freelist))
605 return;
606
607 stripe = next_stripe;
608next:
609 if (wrapped && stripe > start_stripe)
610 return;
611
612 if (stripe == dc->disk.nr_stripes) {
613 stripe = 0;
614 wrapped = true;
615 }
616 }
617}
618
619/*
620 * Returns true if we scanned the entire disk
621 */
622static bool refill_dirty(struct cached_dev *dc)
623{
624 struct keybuf *buf = &dc->writeback_keys;
625 struct bkey start = KEY(dc->disk.id, 0, 0);
626 struct bkey end = KEY(dc->disk.id, MAX_KEY_OFFSET, 0);
627 struct bkey start_pos;
628
629 /*
630 * make sure keybuf pos is inside the range for this disk - at bringup
631 * we might not be attached yet so this disk's inode nr isn't
632 * initialized then
633 */
634 if (bkey_cmp(&buf->last_scanned, &start) < 0 ||
635 bkey_cmp(&buf->last_scanned, &end) > 0)
636 buf->last_scanned = start;
637
638 if (dc->partial_stripes_expensive) {
639 refill_full_stripes(dc);
640 if (array_freelist_empty(&buf->freelist))
641 return false;
642 }
643
644 start_pos = buf->last_scanned;
645 bch_refill_keybuf(dc->disk.c, buf, &end, dirty_pred);
646
647 if (bkey_cmp(&buf->last_scanned, &end) < 0)
648 return false;
649
650 /*
651 * If we get to the end start scanning again from the beginning, and
652 * only scan up to where we initially started scanning from:
653 */
654 buf->last_scanned = start;
655 bch_refill_keybuf(dc->disk.c, buf, &start_pos, dirty_pred);
656
657 return bkey_cmp(&buf->last_scanned, &start_pos) >= 0;
658}
659
660static int bch_writeback_thread(void *arg)
661{
662 struct cached_dev *dc = arg;
663 struct cache_set *c = dc->disk.c;
664 bool searched_full_index;
665
666 bch_ratelimit_reset(&dc->writeback_rate);
667
668 while (!kthread_should_stop() &&
669 !test_bit(CACHE_SET_IO_DISABLE, &c->flags)) {
670 down_write(&dc->writeback_lock);
671 set_current_state(TASK_INTERRUPTIBLE);
672 /*
673 * If the bache device is detaching, skip here and continue
674 * to perform writeback. Otherwise, if no dirty data on cache,
675 * or there is dirty data on cache but writeback is disabled,
676 * the writeback thread should sleep here and wait for others
677 * to wake up it.
678 */
679 if (!test_bit(BCACHE_DEV_DETACHING, &dc->disk.flags) &&
680 (!atomic_read(&dc->has_dirty) || !dc->writeback_running)) {
681 up_write(&dc->writeback_lock);
682
683 if (kthread_should_stop() ||
684 test_bit(CACHE_SET_IO_DISABLE, &c->flags)) {
685 set_current_state(TASK_RUNNING);
686 break;
687 }
688
689 schedule();
690 continue;
691 }
692 set_current_state(TASK_RUNNING);
693
694 searched_full_index = refill_dirty(dc);
695
696 if (searched_full_index &&
697 RB_EMPTY_ROOT(&dc->writeback_keys.keys)) {
698 atomic_set(&dc->has_dirty, 0);
699 SET_BDEV_STATE(&dc->sb, BDEV_STATE_CLEAN);
700 bch_write_bdev_super(dc, NULL);
701 /*
702 * If bcache device is detaching via sysfs interface,
703 * writeback thread should stop after there is no dirty
704 * data on cache. BCACHE_DEV_DETACHING flag is set in
705 * bch_cached_dev_detach().
706 */
707 if (test_bit(BCACHE_DEV_DETACHING, &dc->disk.flags)) {
708 up_write(&dc->writeback_lock);
709 break;
710 }
711
712 /*
713 * When dirty data rate is high (e.g. 50%+), there might
714 * be heavy buckets fragmentation after writeback
715 * finished, which hurts following write performance.
716 * If users really care about write performance they
717 * may set BCH_ENABLE_AUTO_GC via sysfs, then when
718 * BCH_DO_AUTO_GC is set, garbage collection thread
719 * will be wake up here. After moving gc, the shrunk
720 * btree and discarded free buckets SSD space may be
721 * helpful for following write requests.
722 */
723 if (c->gc_after_writeback ==
724 (BCH_ENABLE_AUTO_GC|BCH_DO_AUTO_GC)) {
725 c->gc_after_writeback &= ~BCH_DO_AUTO_GC;
726 force_wake_up_gc(c);
727 }
728 }
729
730 up_write(&dc->writeback_lock);
731
732 read_dirty(dc);
733
734 if (searched_full_index) {
735 unsigned int delay = dc->writeback_delay * HZ;
736
737 while (delay &&
738 !kthread_should_stop() &&
739 !test_bit(CACHE_SET_IO_DISABLE, &c->flags) &&
740 !test_bit(BCACHE_DEV_DETACHING, &dc->disk.flags))
741 delay = schedule_timeout_interruptible(delay);
742
743 bch_ratelimit_reset(&dc->writeback_rate);
744 }
745 }
746
747 if (dc->writeback_write_wq) {
748 flush_workqueue(dc->writeback_write_wq);
749 destroy_workqueue(dc->writeback_write_wq);
750 }
751 cached_dev_put(dc);
752 wait_for_kthread_stop();
753
754 return 0;
755}
756
757/* Init */
758#define INIT_KEYS_EACH_TIME 500000
759#define INIT_KEYS_SLEEP_MS 100
760
761struct sectors_dirty_init {
762 struct btree_op op;
763 unsigned int inode;
764 size_t count;
765 struct bkey start;
766};
767
768static int sectors_dirty_init_fn(struct btree_op *_op, struct btree *b,
769 struct bkey *k)
770{
771 struct sectors_dirty_init *op = container_of(_op,
772 struct sectors_dirty_init, op);
773 if (KEY_INODE(k) > op->inode)
774 return MAP_DONE;
775
776 if (KEY_DIRTY(k))
777 bcache_dev_sectors_dirty_add(b->c, KEY_INODE(k),
778 KEY_START(k), KEY_SIZE(k));
779
780 op->count++;
781 if (atomic_read(&b->c->search_inflight) &&
782 !(op->count % INIT_KEYS_EACH_TIME)) {
783 bkey_copy_key(&op->start, k);
784 return -EAGAIN;
785 }
786
787 return MAP_CONTINUE;
788}
789
790static int bch_root_node_dirty_init(struct cache_set *c,
791 struct bcache_device *d,
792 struct bkey *k)
793{
794 struct sectors_dirty_init op;
795 int ret;
796
797 bch_btree_op_init(&op.op, -1);
798 op.inode = d->id;
799 op.count = 0;
800 op.start = KEY(op.inode, 0, 0);
801
802 do {
803 ret = bcache_btree(map_keys_recurse,
804 k,
805 c->root,
806 &op.op,
807 &op.start,
808 sectors_dirty_init_fn,
809 0);
810 if (ret == -EAGAIN)
811 schedule_timeout_interruptible(
812 msecs_to_jiffies(INIT_KEYS_SLEEP_MS));
813 else if (ret < 0) {
814 pr_warn("sectors dirty init failed, ret=%d!\n", ret);
815 break;
816 }
817 } while (ret == -EAGAIN);
818
819 return ret;
820}
821
822static int bch_dirty_init_thread(void *arg)
823{
824 struct dirty_init_thrd_info *info = arg;
825 struct bch_dirty_init_state *state = info->state;
826 struct cache_set *c = state->c;
827 struct btree_iter iter;
828 struct bkey *k, *p;
829 int cur_idx, prev_idx, skip_nr;
830
831 k = p = NULL;
832 cur_idx = prev_idx = 0;
833
834 bch_btree_iter_init(&c->root->keys, &iter, NULL);
835 k = bch_btree_iter_next_filter(&iter, &c->root->keys, bch_ptr_bad);
836 BUG_ON(!k);
837
838 p = k;
839
840 while (k) {
841 spin_lock(&state->idx_lock);
842 cur_idx = state->key_idx;
843 state->key_idx++;
844 spin_unlock(&state->idx_lock);
845
846 skip_nr = cur_idx - prev_idx;
847
848 while (skip_nr) {
849 k = bch_btree_iter_next_filter(&iter,
850 &c->root->keys,
851 bch_ptr_bad);
852 if (k)
853 p = k;
854 else {
855 atomic_set(&state->enough, 1);
856 /* Update state->enough earlier */
857 smp_mb__after_atomic();
858 goto out;
859 }
860 skip_nr--;
861 cond_resched();
862 }
863
864 if (p) {
865 if (bch_root_node_dirty_init(c, state->d, p) < 0)
866 goto out;
867 }
868
869 p = NULL;
870 prev_idx = cur_idx;
871 cond_resched();
872 }
873
874out:
875 /* In order to wake up state->wait in time */
876 smp_mb__before_atomic();
877 if (atomic_dec_and_test(&state->started))
878 wake_up(&state->wait);
879
880 return 0;
881}
882
883static int bch_btre_dirty_init_thread_nr(void)
884{
885 int n = num_online_cpus()/2;
886
887 if (n == 0)
888 n = 1;
889 else if (n > BCH_DIRTY_INIT_THRD_MAX)
890 n = BCH_DIRTY_INIT_THRD_MAX;
891
892 return n;
893}
894
895void bch_sectors_dirty_init(struct bcache_device *d)
896{
897 int i;
898 struct bkey *k = NULL;
899 struct btree_iter iter;
900 struct sectors_dirty_init op;
901 struct cache_set *c = d->c;
902 struct bch_dirty_init_state *state;
903 char name[32];
904
905 /* Just count root keys if no leaf node */
906 if (c->root->level == 0) {
907 bch_btree_op_init(&op.op, -1);
908 op.inode = d->id;
909 op.count = 0;
910 op.start = KEY(op.inode, 0, 0);
911
912 for_each_key_filter(&c->root->keys,
913 k, &iter, bch_ptr_invalid)
914 sectors_dirty_init_fn(&op.op, c->root, k);
915 return;
916 }
917
918 state = kzalloc(sizeof(struct bch_dirty_init_state), GFP_KERNEL);
919 if (!state) {
920 pr_warn("sectors dirty init failed: cannot allocate memory\n");
921 return;
922 }
923
924 state->c = c;
925 state->d = d;
926 state->total_threads = bch_btre_dirty_init_thread_nr();
927 state->key_idx = 0;
928 spin_lock_init(&state->idx_lock);
929 atomic_set(&state->started, 0);
930 atomic_set(&state->enough, 0);
931 init_waitqueue_head(&state->wait);
932
933 for (i = 0; i < state->total_threads; i++) {
934 /* Fetch latest state->enough earlier */
935 smp_mb__before_atomic();
936 if (atomic_read(&state->enough))
937 break;
938
939 state->infos[i].state = state;
940 atomic_inc(&state->started);
941 snprintf(name, sizeof(name), "bch_dirty_init[%d]", i);
942
943 state->infos[i].thread =
944 kthread_run(bch_dirty_init_thread,
945 &state->infos[i],
946 name);
947 if (IS_ERR(state->infos[i].thread)) {
948 pr_err("fails to run thread bch_dirty_init[%d]\n", i);
949 for (--i; i >= 0; i--)
950 kthread_stop(state->infos[i].thread);
951 goto out;
952 }
953 }
954
955 wait_event_interruptible(state->wait,
956 atomic_read(&state->started) == 0 ||
957 test_bit(CACHE_SET_IO_DISABLE, &c->flags));
958
959out:
960 kfree(state);
961}
962
963void bch_cached_dev_writeback_init(struct cached_dev *dc)
964{
965 sema_init(&dc->in_flight, 64);
966 init_rwsem(&dc->writeback_lock);
967 bch_keybuf_init(&dc->writeback_keys);
968
969 dc->writeback_metadata = true;
970 dc->writeback_running = false;
971 dc->writeback_percent = 10;
972 dc->writeback_delay = 30;
973 atomic_long_set(&dc->writeback_rate.rate, 1024);
974 dc->writeback_rate_minimum = 8;
975
976 dc->writeback_rate_update_seconds = WRITEBACK_RATE_UPDATE_SECS_DEFAULT;
977 dc->writeback_rate_p_term_inverse = 40;
978 dc->writeback_rate_i_term_inverse = 10000;
979
980 WARN_ON(test_and_clear_bit(BCACHE_DEV_WB_RUNNING, &dc->disk.flags));
981 INIT_DELAYED_WORK(&dc->writeback_rate_update, update_writeback_rate);
982}
983
984int bch_cached_dev_writeback_start(struct cached_dev *dc)
985{
986 dc->writeback_write_wq = alloc_workqueue("bcache_writeback_wq",
987 WQ_MEM_RECLAIM, 0);
988 if (!dc->writeback_write_wq)
989 return -ENOMEM;
990
991 cached_dev_get(dc);
992 dc->writeback_thread = kthread_create(bch_writeback_thread, dc,
993 "bcache_writeback");
994 if (IS_ERR(dc->writeback_thread)) {
995 cached_dev_put(dc);
996 destroy_workqueue(dc->writeback_write_wq);
997 return PTR_ERR(dc->writeback_thread);
998 }
999 dc->writeback_running = true;
1000
1001 WARN_ON(test_and_set_bit(BCACHE_DEV_WB_RUNNING, &dc->disk.flags));
1002 schedule_delayed_work(&dc->writeback_rate_update,
1003 dc->writeback_rate_update_seconds * HZ);
1004
1005 bch_writeback_queue(dc);
1006
1007 return 0;
1008}