blob: 9ee0005874cda2621013edc4c73366690b8f1039 [file] [log] [blame]
Greg Kroah-Hartmanb2441312017-11-01 15:07:57 +01001// SPDX-License-Identifier: GPL-2.0
Kent Overstreetcafe5632013-03-23 16:11:31 -07002/*
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"
Kent Overstreet279afba2013-06-05 06:21:07 -070013#include "writeback.h"
Kent Overstreetcafe5632013-03-23 16:11:31 -070014
Kent Overstreet5e6926da2013-07-24 17:50:06 -070015#include <linux/delay.h>
Kent Overstreet5e6926da2013-07-24 17:50:06 -070016#include <linux/kthread.h>
Ingo Molnare6017572017-02-01 16:36:40 +010017#include <linux/sched/clock.h>
Kent Overstreetc37511b2013-04-26 15:39:55 -070018#include <trace/events/bcache.h>
19
Coly Li7a671d82018-12-13 22:53:53 +080020static 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
Kent Overstreetcafe5632013-03-23 16:11:31 -070029/* Rate limiting */
Michael Lyle616486a2018-01-08 12:21:30 -080030static 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 */
Coly Li4a784262020-10-01 14:50:56 +080038 uint64_t cache_sectors = c->nbuckets * c->cache->sb.bucket_size -
Tang Junhui99a27d52018-07-26 12:17:33 +080039 atomic_long_read(&c->flash_dev_dirty_sectors);
Michael Lyle616486a2018-01-08 12:21:30 -080040
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 =
Christoph Hellwigcda25b82021-10-18 12:11:03 +020048 div64_u64(bdev_nr_sectors(dc->bdev) << WRITEBACK_SHARE_SHIFT,
Michael Lyle616486a2018-01-08 12:21:30 -080049 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}
Kent Overstreetcafe5632013-03-23 16:11:31 -070060
61static void __update_writeback_rate(struct cached_dev *dc)
62{
Michael Lyle1d316e62017-10-13 16:35:36 -070063 /*
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 */
Michael Lyle616486a2018-01-08 12:21:30 -080083 int64_t target = __calc_target_rate(dc);
Kent Overstreet279afba2013-06-05 06:21:07 -070084 int64_t dirty = bcache_dev_sectors_dirty(&dc->disk);
Michael Lyle1d316e62017-10-13 16:35:36 -070085 int64_t error = dirty - target;
86 int64_t proportional_scaled =
87 div_s64(error, dc->writeback_rate_p_term_inverse);
Michael Lylee41166c2017-10-13 16:35:38 -070088 int64_t integral_scaled;
89 uint32_t new_rate;
Kent Overstreetcafe5632013-03-23 16:11:31 -070090
dongdong tao71dda2a2021-02-10 13:07:23 +080091 /*
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) {
Gustavo A. R. Silva62594f12021-04-11 21:43:15 +0800113 fp_term = (int64_t)dc->writeback_rate_fp_term_low *
dongdong tao71dda2a2021-02-10 13:07:23 +0800114 (c->gc_stats.in_use - BCH_WRITEBACK_FRAGMENT_THRESHOLD_LOW);
115 } else if (c->gc_stats.in_use <= BCH_WRITEBACK_FRAGMENT_THRESHOLD_HIGH) {
Gustavo A. R. Silva62594f12021-04-11 21:43:15 +0800116 fp_term = (int64_t)dc->writeback_rate_fp_term_mid *
dongdong tao71dda2a2021-02-10 13:07:23 +0800117 (c->gc_stats.in_use - BCH_WRITEBACK_FRAGMENT_THRESHOLD_MID);
118 } else {
Gustavo A. R. Silva62594f12021-04-11 21:43:15 +0800119 fp_term = (int64_t)dc->writeback_rate_fp_term_high *
dongdong tao71dda2a2021-02-10 13:07:23 +0800120 (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
Michael Lyle1d316e62017-10-13 16:35:36 -0700129 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 }
Kent Overstreetcafe5632013-03-23 16:11:31 -0700145
Michael Lyle1d316e62017-10-13 16:35:36 -0700146 integral_scaled = div_s64(dc->writeback_rate_integral,
147 dc->writeback_rate_i_term_inverse);
Kent Overstreet16749c22013-11-11 13:58:34 -0800148
Michael Lylee41166c2017-10-13 16:35:38 -0700149 new_rate = clamp_t(int32_t, (proportional_scaled + integral_scaled),
150 dc->writeback_rate_minimum, NSEC_PER_SEC);
Kent Overstreet16749c22013-11-11 13:58:34 -0800151
Michael Lyle1d316e62017-10-13 16:35:36 -0700152 dc->writeback_rate_proportional = proportional_scaled;
153 dc->writeback_rate_integral_scaled = integral_scaled;
Coly Liea8c53562018-08-09 15:48:49 +0800154 dc->writeback_rate_change = new_rate -
155 atomic_long_read(&dc->writeback_rate.rate);
156 atomic_long_set(&dc->writeback_rate.rate, new_rate);
Kent Overstreetcafe5632013-03-23 16:11:31 -0700157 dc->writeback_rate_target = target;
Kent Overstreetcafe5632013-03-23 16:11:31 -0700158}
159
Coly Liea8c53562018-08-09 15:48:49 +0800160static bool set_at_max_writeback_rate(struct cache_set *c,
161 struct cached_dev *dc)
162{
Coly Lic5fcded2019-11-13 16:03:23 +0800163 /* Don't sst max writeback rate if it is disabled */
164 if (!c->idle_max_writeback_rate_enabled)
165 return false;
166
Coly Li141df8b2019-06-28 19:59:24 +0800167 /* Don't set max writeback rate if gc is running */
168 if (!c->gc_mark_valid)
169 return false;
Coly Liea8c53562018-08-09 15:48:49 +0800170 /*
171 * Idle_counter is increased everytime when update_writeback_rate() is
172 * called. If all backing devices attached to the same cache set have
173 * identical dc->writeback_rate_update_seconds values, it is about 6
174 * rounds of update_writeback_rate() on each backing device before
175 * c->at_max_writeback_rate is set to 1, and then max wrteback rate set
176 * to each dc->writeback_rate.rate.
177 * In order to avoid extra locking cost for counting exact dirty cached
178 * devices number, c->attached_dev_nr is used to calculate the idle
179 * throushold. It might be bigger if not all cached device are in write-
180 * back mode, but it still works well with limited extra rounds of
181 * update_writeback_rate().
182 */
183 if (atomic_inc_return(&c->idle_counter) <
184 atomic_read(&c->attached_dev_nr) * 6)
185 return false;
186
187 if (atomic_read(&c->at_max_writeback_rate) != 1)
188 atomic_set(&c->at_max_writeback_rate, 1);
189
190 atomic_long_set(&dc->writeback_rate.rate, INT_MAX);
191
192 /* keep writeback_rate_target as existing value */
193 dc->writeback_rate_proportional = 0;
194 dc->writeback_rate_integral_scaled = 0;
195 dc->writeback_rate_change = 0;
196
197 /*
198 * Check c->idle_counter and c->at_max_writeback_rate agagain in case
199 * new I/O arrives during before set_at_max_writeback_rate() returns.
200 * Then the writeback rate is set to 1, and its new value should be
201 * decided via __update_writeback_rate().
202 */
203 if ((atomic_read(&c->idle_counter) <
204 atomic_read(&c->attached_dev_nr) * 6) ||
205 !atomic_read(&c->at_max_writeback_rate))
206 return false;
207
208 return true;
209}
210
Kent Overstreetcafe5632013-03-23 16:11:31 -0700211static void update_writeback_rate(struct work_struct *work)
212{
213 struct cached_dev *dc = container_of(to_delayed_work(work),
214 struct cached_dev,
215 writeback_rate_update);
Coly Li771f3932018-03-18 17:36:17 -0700216 struct cache_set *c = dc->disk.c;
Kent Overstreetcafe5632013-03-23 16:11:31 -0700217
Coly Li3fd47bf2018-03-18 17:36:16 -0700218 /*
219 * should check BCACHE_DEV_RATE_DW_RUNNING before calling
220 * cancel_delayed_work_sync().
221 */
222 set_bit(BCACHE_DEV_RATE_DW_RUNNING, &dc->disk.flags);
223 /* paired with where BCACHE_DEV_RATE_DW_RUNNING is tested */
Davidlohr Buesob004aa82020-03-22 14:03:04 +0800224 smp_mb__after_atomic();
Coly Li3fd47bf2018-03-18 17:36:16 -0700225
Coly Li771f3932018-03-18 17:36:17 -0700226 /*
227 * CACHE_SET_IO_DISABLE might be set via sysfs interface,
228 * check it here too.
229 */
230 if (!test_bit(BCACHE_DEV_WB_RUNNING, &dc->disk.flags) ||
231 test_bit(CACHE_SET_IO_DISABLE, &c->flags)) {
Coly Li3fd47bf2018-03-18 17:36:16 -0700232 clear_bit(BCACHE_DEV_RATE_DW_RUNNING, &dc->disk.flags);
233 /* paired with where BCACHE_DEV_RATE_DW_RUNNING is tested */
Davidlohr Buesob004aa82020-03-22 14:03:04 +0800234 smp_mb__after_atomic();
Coly Li3fd47bf2018-03-18 17:36:16 -0700235 return;
236 }
237
Coly Liea8c53562018-08-09 15:48:49 +0800238 if (atomic_read(&dc->has_dirty) && dc->writeback_percent) {
239 /*
240 * If the whole cache set is idle, set_at_max_writeback_rate()
241 * will set writeback rate to a max number. Then it is
242 * unncessary to update writeback rate for an idle cache set
243 * in maximum writeback rate number(s).
244 */
245 if (!set_at_max_writeback_rate(c, dc)) {
246 down_read(&dc->writeback_lock);
247 __update_writeback_rate(dc);
Coly Li7a671d82018-12-13 22:53:53 +0800248 update_gc_after_writeback(c);
Coly Liea8c53562018-08-09 15:48:49 +0800249 up_read(&dc->writeback_lock);
250 }
251 }
Kent Overstreetcafe5632013-03-23 16:11:31 -0700252
Kent Overstreet5e6926da2013-07-24 17:50:06 -0700253
Coly Li771f3932018-03-18 17:36:17 -0700254 /*
255 * CACHE_SET_IO_DISABLE might be set via sysfs interface,
256 * check it here too.
257 */
258 if (test_bit(BCACHE_DEV_WB_RUNNING, &dc->disk.flags) &&
259 !test_bit(CACHE_SET_IO_DISABLE, &c->flags)) {
Coly Li3fd47bf2018-03-18 17:36:16 -0700260 schedule_delayed_work(&dc->writeback_rate_update,
Kent Overstreet5e6926da2013-07-24 17:50:06 -0700261 dc->writeback_rate_update_seconds * HZ);
Coly Li3fd47bf2018-03-18 17:36:16 -0700262 }
263
264 /*
265 * should check BCACHE_DEV_RATE_DW_RUNNING before calling
266 * cancel_delayed_work_sync().
267 */
268 clear_bit(BCACHE_DEV_RATE_DW_RUNNING, &dc->disk.flags);
269 /* paired with where BCACHE_DEV_RATE_DW_RUNNING is tested */
Davidlohr Buesob004aa82020-03-22 14:03:04 +0800270 smp_mb__after_atomic();
Kent Overstreetcafe5632013-03-23 16:11:31 -0700271}
272
Coly Li6f10f7d2018-08-11 13:19:44 +0800273static unsigned int writeback_delay(struct cached_dev *dc,
274 unsigned int sectors)
Kent Overstreetcafe5632013-03-23 16:11:31 -0700275{
Kent Overstreetc4d951d2013-08-21 17:49:09 -0700276 if (test_bit(BCACHE_DEV_DETACHING, &dc->disk.flags) ||
Kent Overstreetcafe5632013-03-23 16:11:31 -0700277 !dc->writeback_percent)
278 return 0;
279
Kent Overstreet16749c22013-11-11 13:58:34 -0800280 return bch_next_delay(&dc->writeback_rate, sectors);
Kent Overstreetcafe5632013-03-23 16:11:31 -0700281}
282
Kent Overstreet5e6926da2013-07-24 17:50:06 -0700283struct dirty_io {
284 struct closure cl;
285 struct cached_dev *dc;
Michael Lyle6e6ccc62018-01-08 12:21:23 -0800286 uint16_t sequence;
Kent Overstreet5e6926da2013-07-24 17:50:06 -0700287 struct bio bio;
288};
Kent Overstreet72c27062013-06-05 06:24:39 -0700289
Kent Overstreetcafe5632013-03-23 16:11:31 -0700290static void dirty_init(struct keybuf_key *w)
291{
292 struct dirty_io *io = w->private;
293 struct bio *bio = &io->bio;
294
Christoph Hellwig49add492022-01-24 10:11:06 +0100295 bio_init(bio, NULL, bio->bi_inline_vecs,
296 DIV_ROUND_UP(KEY_SIZE(&w->key), PAGE_SECTORS), 0);
Kent Overstreetcafe5632013-03-23 16:11:31 -0700297 if (!io->dc->writeback_percent)
298 bio_set_prio(bio, IOPRIO_PRIO_VALUE(IOPRIO_CLASS_IDLE, 0));
299
Kent Overstreet4f024f32013-10-11 15:44:27 -0700300 bio->bi_iter.bi_size = KEY_SIZE(&w->key) << 9;
Kent Overstreetcafe5632013-03-23 16:11:31 -0700301 bio->bi_private = w;
Kent Overstreet169ef1c2013-03-28 12:50:55 -0600302 bch_bio_map(bio, NULL);
Kent Overstreetcafe5632013-03-23 16:11:31 -0700303}
304
Kent Overstreetcafe5632013-03-23 16:11:31 -0700305static void dirty_io_destructor(struct closure *cl)
306{
307 struct dirty_io *io = container_of(cl, struct dirty_io, cl);
Coly Li1fae7cf2018-08-11 13:19:45 +0800308
Kent Overstreetcafe5632013-03-23 16:11:31 -0700309 kfree(io);
310}
311
312static void write_dirty_finish(struct closure *cl)
313{
314 struct dirty_io *io = container_of(cl, struct dirty_io, cl);
315 struct keybuf_key *w = io->bio.bi_private;
316 struct cached_dev *dc = io->dc;
Kent Overstreetcafe5632013-03-23 16:11:31 -0700317
Guoqing Jiang491221f2016-09-22 03:10:01 -0400318 bio_free_pages(&io->bio);
Kent Overstreetcafe5632013-03-23 16:11:31 -0700319
320 /* This is kind of a dumb way of signalling errors. */
321 if (KEY_DIRTY(&w->key)) {
Kent Overstreet6054c6d2013-07-24 18:06:22 -0700322 int ret;
Coly Li6f10f7d2018-08-11 13:19:44 +0800323 unsigned int i;
Kent Overstreetcc7b88192013-07-24 18:07:22 -0700324 struct keylist keys;
Kent Overstreet0b932072013-07-24 17:26:51 -0700325
Kent Overstreet0b932072013-07-24 17:26:51 -0700326 bch_keylist_init(&keys);
Kent Overstreetcafe5632013-03-23 16:11:31 -0700327
Kent Overstreet1b207d82013-09-10 18:52:54 -0700328 bkey_copy(keys.top, &w->key);
329 SET_KEY_DIRTY(keys.top, false);
330 bch_keylist_push(&keys);
Kent Overstreetcafe5632013-03-23 16:11:31 -0700331
332 for (i = 0; i < KEY_PTRS(&w->key); i++)
333 atomic_inc(&PTR_BUCKET(dc->disk.c, &w->key, i)->pin);
334
Kent Overstreetcc7b88192013-07-24 18:07:22 -0700335 ret = bch_btree_insert(dc->disk.c, &keys, NULL, &w->key);
Kent Overstreetcafe5632013-03-23 16:11:31 -0700336
Kent Overstreet6054c6d2013-07-24 18:06:22 -0700337 if (ret)
Kent Overstreetc37511b2013-04-26 15:39:55 -0700338 trace_bcache_writeback_collision(&w->key);
339
Kent Overstreet6054c6d2013-07-24 18:06:22 -0700340 atomic_long_inc(ret
Kent Overstreetcafe5632013-03-23 16:11:31 -0700341 ? &dc->disk.c->writeback_keys_failed
342 : &dc->disk.c->writeback_keys_done);
343 }
344
345 bch_keybuf_del(&dc->writeback_keys, w);
Kent Overstreetc2a4f312013-09-23 23:17:31 -0700346 up(&dc->in_flight);
Kent Overstreetcafe5632013-03-23 16:11:31 -0700347
348 closure_return_with_destructor(cl, dirty_io_destructor);
349}
350
Christoph Hellwig4246a0b2015-07-20 15:29:37 +0200351static void dirty_endio(struct bio *bio)
Kent Overstreetcafe5632013-03-23 16:11:31 -0700352{
353 struct keybuf_key *w = bio->bi_private;
354 struct dirty_io *io = w->private;
355
Coly Libf789802018-05-03 18:51:34 +0800356 if (bio->bi_status) {
Kent Overstreetcafe5632013-03-23 16:11:31 -0700357 SET_KEY_DIRTY(&w->key, false);
Coly Libf789802018-05-03 18:51:34 +0800358 bch_count_backing_io_errors(io->dc, bio);
359 }
Kent Overstreetcafe5632013-03-23 16:11:31 -0700360
361 closure_put(&io->cl);
362}
363
364static void write_dirty(struct closure *cl)
365{
366 struct dirty_io *io = container_of(cl, struct dirty_io, cl);
367 struct keybuf_key *w = io->bio.bi_private;
Michael Lyle6e6ccc62018-01-08 12:21:23 -0800368 struct cached_dev *dc = io->dc;
369
370 uint16_t next_sequence;
371
372 if (atomic_read(&dc->writeback_sequence_next) != io->sequence) {
373 /* Not our turn to write; wait for a write to complete */
374 closure_wait(&dc->writeback_ordering_wait, cl);
375
376 if (atomic_read(&dc->writeback_sequence_next) == io->sequence) {
377 /*
378 * Edge case-- it happened in indeterminate order
379 * relative to when we were added to wait list..
380 */
381 closure_wake_up(&dc->writeback_ordering_wait);
382 }
383
384 continue_at(cl, write_dirty, io->dc->writeback_write_wq);
385 return;
386 }
387
388 next_sequence = io->sequence + 1;
Kent Overstreetcafe5632013-03-23 16:11:31 -0700389
Michael Lyle5fa89fb2017-10-13 16:35:35 -0700390 /*
391 * IO errors are signalled using the dirty bit on the key.
392 * If we failed to read, we should not attempt to write to the
393 * backing device. Instead, immediately go to write_dirty_finish
394 * to clean up.
395 */
396 if (KEY_DIRTY(&w->key)) {
397 dirty_init(w);
398 bio_set_op_attrs(&io->bio, REQ_OP_WRITE, 0);
399 io->bio.bi_iter.bi_sector = KEY_START(&w->key);
400 bio_set_dev(&io->bio, io->dc->bdev);
401 io->bio.bi_end_io = dirty_endio;
Kent Overstreetcafe5632013-03-23 16:11:31 -0700402
Coly Li27a40ab2018-03-18 17:36:24 -0700403 /* I/O request sent to backing device */
Coly Li771f3932018-03-18 17:36:17 -0700404 closure_bio_submit(io->dc->disk.c, &io->bio, cl);
Michael Lyle5fa89fb2017-10-13 16:35:35 -0700405 }
Kent Overstreetcafe5632013-03-23 16:11:31 -0700406
Michael Lyle6e6ccc62018-01-08 12:21:23 -0800407 atomic_set(&dc->writeback_sequence_next, next_sequence);
408 closure_wake_up(&dc->writeback_ordering_wait);
409
Tang Junhui9baf30972017-09-06 14:25:59 +0800410 continue_at(cl, write_dirty_finish, io->dc->writeback_write_wq);
Kent Overstreetcafe5632013-03-23 16:11:31 -0700411}
412
Christoph Hellwig4246a0b2015-07-20 15:29:37 +0200413static void read_dirty_endio(struct bio *bio)
Kent Overstreetcafe5632013-03-23 16:11:31 -0700414{
415 struct keybuf_key *w = bio->bi_private;
416 struct dirty_io *io = w->private;
417
Coly Li5138ac62018-01-08 12:21:29 -0800418 /* is_read = 1 */
Christoph Hellwig11e95602021-04-11 21:43:11 +0800419 bch_count_io_errors(io->dc->disk.c->cache,
Coly Li5138ac62018-01-08 12:21:29 -0800420 bio->bi_status, 1,
421 "reading dirty data from cache");
Kent Overstreetcafe5632013-03-23 16:11:31 -0700422
Christoph Hellwig4246a0b2015-07-20 15:29:37 +0200423 dirty_endio(bio);
Kent Overstreetcafe5632013-03-23 16:11:31 -0700424}
425
426static void read_dirty_submit(struct closure *cl)
427{
428 struct dirty_io *io = container_of(cl, struct dirty_io, cl);
429
Coly Li771f3932018-03-18 17:36:17 -0700430 closure_bio_submit(io->dc->disk.c, &io->bio, cl);
Kent Overstreetcafe5632013-03-23 16:11:31 -0700431
Tang Junhui9baf30972017-09-06 14:25:59 +0800432 continue_at(cl, write_dirty, io->dc->writeback_write_wq);
Kent Overstreetcafe5632013-03-23 16:11:31 -0700433}
434
Kent Overstreet5e6926da2013-07-24 17:50:06 -0700435static void read_dirty(struct cached_dev *dc)
Kent Overstreetcafe5632013-03-23 16:11:31 -0700436{
Coly Li6f10f7d2018-08-11 13:19:44 +0800437 unsigned int delay = 0;
Tang Junhui539d39e2018-01-08 12:21:22 -0800438 struct keybuf_key *next, *keys[MAX_WRITEBACKS_IN_PASS], *w;
439 size_t size;
440 int nk, i;
Kent Overstreetcafe5632013-03-23 16:11:31 -0700441 struct dirty_io *io;
Kent Overstreet5e6926da2013-07-24 17:50:06 -0700442 struct closure cl;
Michael Lyle6e6ccc62018-01-08 12:21:23 -0800443 uint16_t sequence = 0;
Kent Overstreet5e6926da2013-07-24 17:50:06 -0700444
Michael Lyle6e6ccc62018-01-08 12:21:23 -0800445 BUG_ON(!llist_empty(&dc->writeback_ordering_wait.list));
446 atomic_set(&dc->writeback_sequence_next, sequence);
Kent Overstreet5e6926da2013-07-24 17:50:06 -0700447 closure_init_stack(&cl);
Kent Overstreetcafe5632013-03-23 16:11:31 -0700448
449 /*
450 * XXX: if we error, background writeback just spins. Should use some
451 * mempools.
452 */
453
Tang Junhui539d39e2018-01-08 12:21:22 -0800454 next = bch_keybuf_next(&dc->writeback_keys);
Kent Overstreet5e6926da2013-07-24 17:50:06 -0700455
Coly Li771f3932018-03-18 17:36:17 -0700456 while (!kthread_should_stop() &&
457 !test_bit(CACHE_SET_IO_DISABLE, &dc->disk.c->flags) &&
458 next) {
Tang Junhui539d39e2018-01-08 12:21:22 -0800459 size = 0;
460 nk = 0;
Kent Overstreetcafe5632013-03-23 16:11:31 -0700461
Tang Junhui539d39e2018-01-08 12:21:22 -0800462 do {
463 BUG_ON(ptr_stale(dc->disk.c, &next->key, 0));
Kent Overstreetcafe5632013-03-23 16:11:31 -0700464
Tang Junhui539d39e2018-01-08 12:21:22 -0800465 /*
466 * Don't combine too many operations, even if they
467 * are all small.
468 */
469 if (nk >= MAX_WRITEBACKS_IN_PASS)
470 break;
Kent Overstreetcafe5632013-03-23 16:11:31 -0700471
Tang Junhui539d39e2018-01-08 12:21:22 -0800472 /*
473 * If the current operation is very large, don't
474 * further combine operations.
475 */
476 if (size >= MAX_WRITESIZE_IN_PASS)
477 break;
Kent Overstreetcafe5632013-03-23 16:11:31 -0700478
Tang Junhui539d39e2018-01-08 12:21:22 -0800479 /*
480 * Operations are only eligible to be combined
481 * if they are contiguous.
482 *
483 * TODO: add a heuristic willing to fire a
484 * certain amount of non-contiguous IO per pass,
485 * so that we can benefit from backing device
486 * command queueing.
487 */
488 if ((nk != 0) && bkey_cmp(&keys[nk-1]->key,
489 &START_KEY(&next->key)))
490 break;
Kent Overstreetcafe5632013-03-23 16:11:31 -0700491
Tang Junhui539d39e2018-01-08 12:21:22 -0800492 size += KEY_SIZE(&next->key);
493 keys[nk++] = next;
494 } while ((next = bch_keybuf_next(&dc->writeback_keys)));
Kent Overstreetcafe5632013-03-23 16:11:31 -0700495
Tang Junhui539d39e2018-01-08 12:21:22 -0800496 /* Now we have gathered a set of 1..5 keys to write back. */
497 for (i = 0; i < nk; i++) {
498 w = keys[i];
Kent Overstreetcafe5632013-03-23 16:11:31 -0700499
Gustavo A. R. Silva29f1d5c2020-07-25 20:00:20 +0800500 io = kzalloc(struct_size(io, bio.bi_inline_vecs,
501 DIV_ROUND_UP(KEY_SIZE(&w->key), PAGE_SECTORS)),
Tang Junhui539d39e2018-01-08 12:21:22 -0800502 GFP_KERNEL);
503 if (!io)
504 goto err;
Kent Overstreetcafe5632013-03-23 16:11:31 -0700505
Tang Junhui539d39e2018-01-08 12:21:22 -0800506 w->private = io;
507 io->dc = dc;
Michael Lyle6e6ccc62018-01-08 12:21:23 -0800508 io->sequence = sequence++;
Kent Overstreetcafe5632013-03-23 16:11:31 -0700509
Tang Junhui539d39e2018-01-08 12:21:22 -0800510 dirty_init(w);
511 bio_set_op_attrs(&io->bio, REQ_OP_READ, 0);
512 io->bio.bi_iter.bi_sector = PTR_OFFSET(&w->key, 0);
Christoph Hellwig11e95602021-04-11 21:43:11 +0800513 bio_set_dev(&io->bio, dc->disk.c->cache->bdev);
Tang Junhui539d39e2018-01-08 12:21:22 -0800514 io->bio.bi_end_io = read_dirty_endio;
Kent Overstreetcafe5632013-03-23 16:11:31 -0700515
Tang Junhui539d39e2018-01-08 12:21:22 -0800516 if (bch_bio_alloc_pages(&io->bio, GFP_KERNEL))
517 goto err_free;
518
519 trace_bcache_writeback(&w->key);
520
521 down(&dc->in_flight);
522
Coly Li3be11db2018-08-11 13:19:55 +0800523 /*
524 * We've acquired a semaphore for the maximum
Tang Junhui539d39e2018-01-08 12:21:22 -0800525 * simultaneous number of writebacks; from here
526 * everything happens asynchronously.
527 */
528 closure_call(&io->cl, read_dirty_submit, NULL, &cl);
529 }
530
531 delay = writeback_delay(dc, size);
532
Coly Li771f3932018-03-18 17:36:17 -0700533 while (!kthread_should_stop() &&
534 !test_bit(CACHE_SET_IO_DISABLE, &dc->disk.c->flags) &&
535 delay) {
Tang Junhui539d39e2018-01-08 12:21:22 -0800536 schedule_timeout_interruptible(delay);
537 delay = writeback_delay(dc, 0);
538 }
Kent Overstreetcafe5632013-03-23 16:11:31 -0700539 }
540
541 if (0) {
542err_free:
543 kfree(w->private);
544err:
545 bch_keybuf_del(&dc->writeback_keys, w);
546 }
547
Kent Overstreetc2a4f312013-09-23 23:17:31 -0700548 /*
549 * Wait for outstanding writeback IOs to finish (and keybuf slots to be
550 * freed) before refilling again
551 */
Kent Overstreet5e6926da2013-07-24 17:50:06 -0700552 closure_sync(&cl);
553}
554
555/* Scan for dirty data */
556
Coly Li6f10f7d2018-08-11 13:19:44 +0800557void bcache_dev_sectors_dirty_add(struct cache_set *c, unsigned int inode,
Kent Overstreet5e6926da2013-07-24 17:50:06 -0700558 uint64_t offset, int nr_sectors)
559{
560 struct bcache_device *d = c->devices[inode];
Coly Li7a148122020-07-25 20:00:22 +0800561 unsigned int stripe_offset, sectors_dirty;
562 int stripe;
Kent Overstreet5e6926da2013-07-24 17:50:06 -0700563
564 if (!d)
565 return;
566
Coly Li7a148122020-07-25 20:00:22 +0800567 stripe = offset_to_stripe(d, offset);
568 if (stripe < 0)
569 return;
570
Tang Junhui99a27d52018-07-26 12:17:33 +0800571 if (UUID_FLASH_ONLY(&c->uuids[inode]))
572 atomic_long_add(nr_sectors, &c->flash_dev_dirty_sectors);
573
Kent Overstreet5e6926da2013-07-24 17:50:06 -0700574 stripe_offset = offset & (d->stripe_size - 1);
575
576 while (nr_sectors) {
Coly Li6f10f7d2018-08-11 13:19:44 +0800577 int s = min_t(unsigned int, abs(nr_sectors),
Kent Overstreet5e6926da2013-07-24 17:50:06 -0700578 d->stripe_size - stripe_offset);
579
580 if (nr_sectors < 0)
581 s = -s;
582
Kent Overstreet48a915a2013-10-31 15:43:22 -0700583 if (stripe >= d->nr_stripes)
584 return;
585
586 sectors_dirty = atomic_add_return(s,
587 d->stripe_sectors_dirty + stripe);
Mingzhe Zou7b1002f2022-01-07 16:21:13 +0800588 if (sectors_dirty == d->stripe_size) {
589 if (!test_bit(stripe, d->full_dirty_stripes))
590 set_bit(stripe, d->full_dirty_stripes);
591 } else {
592 if (test_bit(stripe, d->full_dirty_stripes))
593 clear_bit(stripe, d->full_dirty_stripes);
594 }
Kent Overstreet48a915a2013-10-31 15:43:22 -0700595
Kent Overstreet5e6926da2013-07-24 17:50:06 -0700596 nr_sectors -= s;
597 stripe_offset = 0;
598 stripe++;
599 }
600}
601
602static bool dirty_pred(struct keybuf *buf, struct bkey *k)
603{
Coly Lib0d30982018-08-11 13:19:47 +0800604 struct cached_dev *dc = container_of(buf,
605 struct cached_dev,
606 writeback_keys);
Kent Overstreet627ccd22015-11-29 18:47:01 -0800607
608 BUG_ON(KEY_INODE(k) != dc->disk.id);
609
Kent Overstreet5e6926da2013-07-24 17:50:06 -0700610 return KEY_DIRTY(k);
611}
612
Kent Overstreet48a915a2013-10-31 15:43:22 -0700613static void refill_full_stripes(struct cached_dev *dc)
Kent Overstreet5e6926da2013-07-24 17:50:06 -0700614{
Kent Overstreet48a915a2013-10-31 15:43:22 -0700615 struct keybuf *buf = &dc->writeback_keys;
Coly Li7a148122020-07-25 20:00:22 +0800616 unsigned int start_stripe, next_stripe;
617 int stripe;
Kent Overstreet48a915a2013-10-31 15:43:22 -0700618 bool wrapped = false;
Kent Overstreet5e6926da2013-07-24 17:50:06 -0700619
Kent Overstreet48a915a2013-10-31 15:43:22 -0700620 stripe = offset_to_stripe(&dc->disk, KEY_OFFSET(&buf->last_scanned));
Coly Li7a148122020-07-25 20:00:22 +0800621 if (stripe < 0)
Kent Overstreet48a915a2013-10-31 15:43:22 -0700622 stripe = 0;
623
624 start_stripe = stripe;
Kent Overstreet5e6926da2013-07-24 17:50:06 -0700625
626 while (1) {
Kent Overstreet48a915a2013-10-31 15:43:22 -0700627 stripe = find_next_bit(dc->disk.full_dirty_stripes,
628 dc->disk.nr_stripes, stripe);
Kent Overstreet5e6926da2013-07-24 17:50:06 -0700629
Kent Overstreet48a915a2013-10-31 15:43:22 -0700630 if (stripe == dc->disk.nr_stripes)
631 goto next;
Kent Overstreet5e6926da2013-07-24 17:50:06 -0700632
Kent Overstreet48a915a2013-10-31 15:43:22 -0700633 next_stripe = find_next_zero_bit(dc->disk.full_dirty_stripes,
634 dc->disk.nr_stripes, stripe);
635
636 buf->last_scanned = KEY(dc->disk.id,
637 stripe * dc->disk.stripe_size, 0);
638
639 bch_refill_keybuf(dc->disk.c, buf,
640 &KEY(dc->disk.id,
641 next_stripe * dc->disk.stripe_size, 0),
642 dirty_pred);
643
644 if (array_freelist_empty(&buf->freelist))
645 return;
646
647 stripe = next_stripe;
648next:
649 if (wrapped && stripe > start_stripe)
650 return;
651
652 if (stripe == dc->disk.nr_stripes) {
653 stripe = 0;
654 wrapped = true;
655 }
Kent Overstreet5e6926da2013-07-24 17:50:06 -0700656 }
657}
658
Kent Overstreet627ccd22015-11-29 18:47:01 -0800659/*
660 * Returns true if we scanned the entire disk
661 */
Kent Overstreet5e6926da2013-07-24 17:50:06 -0700662static bool refill_dirty(struct cached_dev *dc)
663{
664 struct keybuf *buf = &dc->writeback_keys;
Kent Overstreet627ccd22015-11-29 18:47:01 -0800665 struct bkey start = KEY(dc->disk.id, 0, 0);
Kent Overstreet5e6926da2013-07-24 17:50:06 -0700666 struct bkey end = KEY(dc->disk.id, MAX_KEY_OFFSET, 0);
Kent Overstreet627ccd22015-11-29 18:47:01 -0800667 struct bkey start_pos;
668
669 /*
670 * make sure keybuf pos is inside the range for this disk - at bringup
671 * we might not be attached yet so this disk's inode nr isn't
672 * initialized then
673 */
674 if (bkey_cmp(&buf->last_scanned, &start) < 0 ||
675 bkey_cmp(&buf->last_scanned, &end) > 0)
676 buf->last_scanned = start;
Kent Overstreet48a915a2013-10-31 15:43:22 -0700677
678 if (dc->partial_stripes_expensive) {
679 refill_full_stripes(dc);
680 if (array_freelist_empty(&buf->freelist))
681 return false;
682 }
Kent Overstreet5e6926da2013-07-24 17:50:06 -0700683
Kent Overstreet627ccd22015-11-29 18:47:01 -0800684 start_pos = buf->last_scanned;
Kent Overstreet48a915a2013-10-31 15:43:22 -0700685 bch_refill_keybuf(dc->disk.c, buf, &end, dirty_pred);
Kent Overstreet5e6926da2013-07-24 17:50:06 -0700686
Kent Overstreet627ccd22015-11-29 18:47:01 -0800687 if (bkey_cmp(&buf->last_scanned, &end) < 0)
688 return false;
689
690 /*
691 * If we get to the end start scanning again from the beginning, and
692 * only scan up to where we initially started scanning from:
693 */
694 buf->last_scanned = start;
695 bch_refill_keybuf(dc->disk.c, buf, &start_pos, dirty_pred);
696
697 return bkey_cmp(&buf->last_scanned, &start_pos) >= 0;
Kent Overstreet5e6926da2013-07-24 17:50:06 -0700698}
699
700static int bch_writeback_thread(void *arg)
701{
702 struct cached_dev *dc = arg;
Coly Li771f3932018-03-18 17:36:17 -0700703 struct cache_set *c = dc->disk.c;
Kent Overstreet5e6926da2013-07-24 17:50:06 -0700704 bool searched_full_index;
705
Michael Lylea8500fc2017-10-13 16:35:39 -0700706 bch_ratelimit_reset(&dc->writeback_rate);
707
Coly Li771f3932018-03-18 17:36:17 -0700708 while (!kthread_should_stop() &&
709 !test_bit(CACHE_SET_IO_DISABLE, &c->flags)) {
Kent Overstreet5e6926da2013-07-24 17:50:06 -0700710 down_write(&dc->writeback_lock);
Coly Li99361bb2018-02-07 11:41:41 -0800711 set_current_state(TASK_INTERRUPTIBLE);
Coly Lifadd94e2018-03-18 17:36:15 -0700712 /*
713 * If the bache device is detaching, skip here and continue
714 * to perform writeback. Otherwise, if no dirty data on cache,
715 * or there is dirty data on cache but writeback is disabled,
716 * the writeback thread should sleep here and wait for others
717 * to wake up it.
718 */
719 if (!test_bit(BCACHE_DEV_DETACHING, &dc->disk.flags) &&
720 (!atomic_read(&dc->has_dirty) || !dc->writeback_running)) {
Kent Overstreet5e6926da2013-07-24 17:50:06 -0700721 up_write(&dc->writeback_lock);
Kent Overstreet5e6926da2013-07-24 17:50:06 -0700722
Coly Li771f3932018-03-18 17:36:17 -0700723 if (kthread_should_stop() ||
724 test_bit(CACHE_SET_IO_DISABLE, &c->flags)) {
Coly Li99361bb2018-02-07 11:41:41 -0800725 set_current_state(TASK_RUNNING);
Coly Li804f3c62018-03-18 17:36:14 -0700726 break;
Coly Li99361bb2018-02-07 11:41:41 -0800727 }
Kent Overstreet5e6926da2013-07-24 17:50:06 -0700728
Kent Overstreet5e6926da2013-07-24 17:50:06 -0700729 schedule();
730 continue;
731 }
Coly Li99361bb2018-02-07 11:41:41 -0800732 set_current_state(TASK_RUNNING);
Kent Overstreet5e6926da2013-07-24 17:50:06 -0700733
734 searched_full_index = refill_dirty(dc);
735
736 if (searched_full_index &&
737 RB_EMPTY_ROOT(&dc->writeback_keys.keys)) {
738 atomic_set(&dc->has_dirty, 0);
Kent Overstreet5e6926da2013-07-24 17:50:06 -0700739 SET_BDEV_STATE(&dc->sb, BDEV_STATE_CLEAN);
740 bch_write_bdev_super(dc, NULL);
Coly Lifadd94e2018-03-18 17:36:15 -0700741 /*
742 * If bcache device is detaching via sysfs interface,
743 * writeback thread should stop after there is no dirty
744 * data on cache. BCACHE_DEV_DETACHING flag is set in
745 * bch_cached_dev_detach().
746 */
Shan Hai3943b0402018-08-23 02:02:56 +0800747 if (test_bit(BCACHE_DEV_DETACHING, &dc->disk.flags)) {
Dongsheng Yangdf4ad532020-12-08 00:39:15 +0800748 struct closure cl;
749
750 closure_init_stack(&cl);
751 memset(&dc->sb.set_uuid, 0, 16);
752 SET_BDEV_STATE(&dc->sb, BDEV_STATE_NONE);
753
754 bch_write_bdev_super(dc, &cl);
755 closure_sync(&cl);
756
Shan Hai3943b0402018-08-23 02:02:56 +0800757 up_write(&dc->writeback_lock);
Coly Lifadd94e2018-03-18 17:36:15 -0700758 break;
Shan Hai3943b0402018-08-23 02:02:56 +0800759 }
Coly Li7a671d82018-12-13 22:53:53 +0800760
761 /*
762 * When dirty data rate is high (e.g. 50%+), there might
763 * be heavy buckets fragmentation after writeback
764 * finished, which hurts following write performance.
765 * If users really care about write performance they
766 * may set BCH_ENABLE_AUTO_GC via sysfs, then when
767 * BCH_DO_AUTO_GC is set, garbage collection thread
768 * will be wake up here. After moving gc, the shrunk
769 * btree and discarded free buckets SSD space may be
770 * helpful for following write requests.
771 */
772 if (c->gc_after_writeback ==
773 (BCH_ENABLE_AUTO_GC|BCH_DO_AUTO_GC)) {
774 c->gc_after_writeback &= ~BCH_DO_AUTO_GC;
775 force_wake_up_gc(c);
776 }
Kent Overstreet5e6926da2013-07-24 17:50:06 -0700777 }
778
779 up_write(&dc->writeback_lock);
780
Kent Overstreet5e6926da2013-07-24 17:50:06 -0700781 read_dirty(dc);
782
783 if (searched_full_index) {
Coly Li6f10f7d2018-08-11 13:19:44 +0800784 unsigned int delay = dc->writeback_delay * HZ;
Kent Overstreet5e6926da2013-07-24 17:50:06 -0700785
786 while (delay &&
787 !kthread_should_stop() &&
Coly Li771f3932018-03-18 17:36:17 -0700788 !test_bit(CACHE_SET_IO_DISABLE, &c->flags) &&
Kent Overstreetc4d951d2013-08-21 17:49:09 -0700789 !test_bit(BCACHE_DEV_DETACHING, &dc->disk.flags))
Slava Pestov9e5c3532014-05-01 13:48:57 -0700790 delay = schedule_timeout_interruptible(delay);
Michael Lylea8500fc2017-10-13 16:35:39 -0700791
792 bch_ratelimit_reset(&dc->writeback_rate);
Kent Overstreet5e6926da2013-07-24 17:50:06 -0700793 }
794 }
795
Coly Li7e865eb2019-06-28 19:59:49 +0800796 if (dc->writeback_write_wq) {
797 flush_workqueue(dc->writeback_write_wq);
798 destroy_workqueue(dc->writeback_write_wq);
799 }
Coly Li804f3c62018-03-18 17:36:14 -0700800 cached_dev_put(dc);
Coly Li771f3932018-03-18 17:36:17 -0700801 wait_for_kthread_stop();
Coly Li804f3c62018-03-18 17:36:14 -0700802
Kent Overstreet5e6926da2013-07-24 17:50:06 -0700803 return 0;
Kent Overstreetcafe5632013-03-23 16:11:31 -0700804}
805
Kent Overstreet444fc0b2013-05-11 17:07:26 -0700806/* Init */
Tang Junhui94f71c12018-07-26 12:17:36 +0800807#define INIT_KEYS_EACH_TIME 500000
808#define INIT_KEYS_SLEEP_MS 100
Kent Overstreet444fc0b2013-05-11 17:07:26 -0700809
Kent Overstreetc18536a2013-07-24 17:44:17 -0700810struct sectors_dirty_init {
811 struct btree_op op;
Coly Li6f10f7d2018-08-11 13:19:44 +0800812 unsigned int inode;
Tang Junhui94f71c12018-07-26 12:17:36 +0800813 size_t count;
814 struct bkey start;
Kent Overstreetc18536a2013-07-24 17:44:17 -0700815};
816
817static int sectors_dirty_init_fn(struct btree_op *_op, struct btree *b,
Kent Overstreet48dad8b2013-09-10 18:48:51 -0700818 struct bkey *k)
Kent Overstreet444fc0b2013-05-11 17:07:26 -0700819{
Kent Overstreetc18536a2013-07-24 17:44:17 -0700820 struct sectors_dirty_init *op = container_of(_op,
821 struct sectors_dirty_init, op);
Kent Overstreet48dad8b2013-09-10 18:48:51 -0700822 if (KEY_INODE(k) > op->inode)
823 return MAP_DONE;
Kent Overstreet444fc0b2013-05-11 17:07:26 -0700824
Kent Overstreet48dad8b2013-09-10 18:48:51 -0700825 if (KEY_DIRTY(k))
826 bcache_dev_sectors_dirty_add(b->c, KEY_INODE(k),
827 KEY_START(k), KEY_SIZE(k));
Kent Overstreet444fc0b2013-05-11 17:07:26 -0700828
Tang Junhui94f71c12018-07-26 12:17:36 +0800829 op->count++;
830 if (atomic_read(&b->c->search_inflight) &&
831 !(op->count % INIT_KEYS_EACH_TIME)) {
832 bkey_copy_key(&op->start, k);
833 return -EAGAIN;
834 }
835
Kent Overstreet48dad8b2013-09-10 18:48:51 -0700836 return MAP_CONTINUE;
Kent Overstreet444fc0b2013-05-11 17:07:26 -0700837}
838
Coly Lib144e452020-03-22 14:03:02 +0800839static int bch_root_node_dirty_init(struct cache_set *c,
840 struct bcache_device *d,
841 struct bkey *k)
Kent Overstreet444fc0b2013-05-11 17:07:26 -0700842{
Kent Overstreetc18536a2013-07-24 17:44:17 -0700843 struct sectors_dirty_init op;
Tang Junhui94f71c12018-07-26 12:17:36 +0800844 int ret;
Kent Overstreet444fc0b2013-05-11 17:07:26 -0700845
Kent Overstreetb54d6932013-07-24 18:04:18 -0700846 bch_btree_op_init(&op.op, -1);
Tang Junhui175206c2017-09-07 01:28:53 +0800847 op.inode = d->id;
Tang Junhui94f71c12018-07-26 12:17:36 +0800848 op.count = 0;
849 op.start = KEY(op.inode, 0, 0);
Kent Overstreet48dad8b2013-09-10 18:48:51 -0700850
Tang Junhui94f71c12018-07-26 12:17:36 +0800851 do {
Coly Lib144e452020-03-22 14:03:02 +0800852 ret = bcache_btree(map_keys_recurse,
853 k,
854 c->root,
855 &op.op,
856 &op.start,
857 sectors_dirty_init_fn,
858 0);
Tang Junhui94f71c12018-07-26 12:17:36 +0800859 if (ret == -EAGAIN)
860 schedule_timeout_interruptible(
861 msecs_to_jiffies(INIT_KEYS_SLEEP_MS));
862 else if (ret < 0) {
Joe Perches46f5aa82020-05-27 12:01:52 +0800863 pr_warn("sectors dirty init failed, ret=%d!\n", ret);
Tang Junhui94f71c12018-07-26 12:17:36 +0800864 break;
865 }
866 } while (ret == -EAGAIN);
Coly Lib144e452020-03-22 14:03:02 +0800867
868 return ret;
869}
870
871static int bch_dirty_init_thread(void *arg)
872{
873 struct dirty_init_thrd_info *info = arg;
874 struct bch_dirty_init_state *state = info->state;
875 struct cache_set *c = state->c;
876 struct btree_iter iter;
877 struct bkey *k, *p;
878 int cur_idx, prev_idx, skip_nr;
Coly Lib144e452020-03-22 14:03:02 +0800879
880 k = p = NULL;
Coly Lib144e452020-03-22 14:03:02 +0800881 cur_idx = prev_idx = 0;
882
883 bch_btree_iter_init(&c->root->keys, &iter, NULL);
884 k = bch_btree_iter_next_filter(&iter, &c->root->keys, bch_ptr_bad);
885 BUG_ON(!k);
886
887 p = k;
888
889 while (k) {
890 spin_lock(&state->idx_lock);
891 cur_idx = state->key_idx;
892 state->key_idx++;
893 spin_unlock(&state->idx_lock);
894
895 skip_nr = cur_idx - prev_idx;
896
897 while (skip_nr) {
898 k = bch_btree_iter_next_filter(&iter,
899 &c->root->keys,
900 bch_ptr_bad);
901 if (k)
902 p = k;
903 else {
904 atomic_set(&state->enough, 1);
905 /* Update state->enough earlier */
Coly Lieb9b6662020-03-22 14:03:05 +0800906 smp_mb__after_atomic();
Coly Lib144e452020-03-22 14:03:02 +0800907 goto out;
908 }
909 skip_nr--;
910 cond_resched();
911 }
912
913 if (p) {
914 if (bch_root_node_dirty_init(c, state->d, p) < 0)
915 goto out;
916 }
917
918 p = NULL;
919 prev_idx = cur_idx;
920 cond_resched();
921 }
922
923out:
924 /* In order to wake up state->wait in time */
Coly Lieb9b6662020-03-22 14:03:05 +0800925 smp_mb__before_atomic();
Coly Lib144e452020-03-22 14:03:02 +0800926 if (atomic_dec_and_test(&state->started))
927 wake_up(&state->wait);
928
929 return 0;
930}
931
932static int bch_btre_dirty_init_thread_nr(void)
933{
934 int n = num_online_cpus()/2;
935
936 if (n == 0)
937 n = 1;
938 else if (n > BCH_DIRTY_INIT_THRD_MAX)
939 n = BCH_DIRTY_INIT_THRD_MAX;
940
941 return n;
942}
943
944void bch_sectors_dirty_init(struct bcache_device *d)
945{
946 int i;
947 struct bkey *k = NULL;
948 struct btree_iter iter;
949 struct sectors_dirty_init op;
950 struct cache_set *c = d->c;
951 struct bch_dirty_init_state *state;
952 char name[32];
953
954 /* Just count root keys if no leaf node */
955 if (c->root->level == 0) {
956 bch_btree_op_init(&op.op, -1);
957 op.inode = d->id;
958 op.count = 0;
959 op.start = KEY(op.inode, 0, 0);
960
961 for_each_key_filter(&c->root->keys,
962 k, &iter, bch_ptr_invalid)
963 sectors_dirty_init_fn(&op.op, c->root, k);
964 return;
965 }
966
967 state = kzalloc(sizeof(struct bch_dirty_init_state), GFP_KERNEL);
968 if (!state) {
Joe Perches46f5aa82020-05-27 12:01:52 +0800969 pr_warn("sectors dirty init failed: cannot allocate memory\n");
Coly Lib144e452020-03-22 14:03:02 +0800970 return;
971 }
972
973 state->c = c;
974 state->d = d;
975 state->total_threads = bch_btre_dirty_init_thread_nr();
976 state->key_idx = 0;
977 spin_lock_init(&state->idx_lock);
978 atomic_set(&state->started, 0);
979 atomic_set(&state->enough, 0);
980 init_waitqueue_head(&state->wait);
981
982 for (i = 0; i < state->total_threads; i++) {
983 /* Fetch latest state->enough earlier */
Coly Lieb9b6662020-03-22 14:03:05 +0800984 smp_mb__before_atomic();
Coly Lib144e452020-03-22 14:03:02 +0800985 if (atomic_read(&state->enough))
986 break;
987
988 state->infos[i].state = state;
989 atomic_inc(&state->started);
990 snprintf(name, sizeof(name), "bch_dirty_init[%d]", i);
991
992 state->infos[i].thread =
993 kthread_run(bch_dirty_init_thread,
994 &state->infos[i],
995 name);
996 if (IS_ERR(state->infos[i].thread)) {
Joe Perches46f5aa82020-05-27 12:01:52 +0800997 pr_err("fails to run thread bch_dirty_init[%d]\n", i);
Coly Lib144e452020-03-22 14:03:02 +0800998 for (--i; i >= 0; i--)
999 kthread_stop(state->infos[i].thread);
1000 goto out;
1001 }
1002 }
1003
Mingzhe Zou887554a2022-02-11 14:39:15 +08001004 /*
1005 * Must wait for all threads to stop.
1006 */
Coly Lib144e452020-03-22 14:03:02 +08001007 wait_event_interruptible(state->wait,
Mingzhe Zou887554a2022-02-11 14:39:15 +08001008 atomic_read(&state->started) == 0);
Coly Lib144e452020-03-22 14:03:02 +08001009
1010out:
1011 kfree(state);
Kent Overstreet444fc0b2013-05-11 17:07:26 -07001012}
1013
Slava Pestov9e5c3532014-05-01 13:48:57 -07001014void bch_cached_dev_writeback_init(struct cached_dev *dc)
Kent Overstreetcafe5632013-03-23 16:11:31 -07001015{
Kent Overstreetc2a4f312013-09-23 23:17:31 -07001016 sema_init(&dc->in_flight, 64);
Kent Overstreetcafe5632013-03-23 16:11:31 -07001017 init_rwsem(&dc->writeback_lock);
Kent Overstreet72c27062013-06-05 06:24:39 -07001018 bch_keybuf_init(&dc->writeback_keys);
Kent Overstreetcafe5632013-03-23 16:11:31 -07001019
1020 dc->writeback_metadata = true;
Shenghui Wang79b79142018-12-13 22:53:50 +08001021 dc->writeback_running = false;
dongdong tao71dda2a2021-02-10 13:07:23 +08001022 dc->writeback_consider_fragment = true;
Kent Overstreetcafe5632013-03-23 16:11:31 -07001023 dc->writeback_percent = 10;
1024 dc->writeback_delay = 30;
Coly Liea8c53562018-08-09 15:48:49 +08001025 atomic_long_set(&dc->writeback_rate.rate, 1024);
Michael Lyleae82ddb2017-10-13 16:35:37 -07001026 dc->writeback_rate_minimum = 8;
Kent Overstreetcafe5632013-03-23 16:11:31 -07001027
Coly Li7a5e3ec2018-02-07 11:41:44 -08001028 dc->writeback_rate_update_seconds = WRITEBACK_RATE_UPDATE_SECS_DEFAULT;
Michael Lyle1d316e62017-10-13 16:35:36 -07001029 dc->writeback_rate_p_term_inverse = 40;
dongdong tao71dda2a2021-02-10 13:07:23 +08001030 dc->writeback_rate_fp_term_low = 1;
1031 dc->writeback_rate_fp_term_mid = 10;
1032 dc->writeback_rate_fp_term_high = 1000;
Michael Lyle1d316e62017-10-13 16:35:36 -07001033 dc->writeback_rate_i_term_inverse = 10000;
Kent Overstreetcafe5632013-03-23 16:11:31 -07001034
Coly Li3fd47bf2018-03-18 17:36:16 -07001035 WARN_ON(test_and_clear_bit(BCACHE_DEV_WB_RUNNING, &dc->disk.flags));
Slava Pestov9e5c3532014-05-01 13:48:57 -07001036 INIT_DELAYED_WORK(&dc->writeback_rate_update, update_writeback_rate);
1037}
1038
1039int bch_cached_dev_writeback_start(struct cached_dev *dc)
1040{
Tang Junhui9baf30972017-09-06 14:25:59 +08001041 dc->writeback_write_wq = alloc_workqueue("bcache_writeback_wq",
1042 WQ_MEM_RECLAIM, 0);
1043 if (!dc->writeback_write_wq)
1044 return -ENOMEM;
1045
Coly Li804f3c62018-03-18 17:36:14 -07001046 cached_dev_get(dc);
Kent Overstreet5e6926da2013-07-24 17:50:06 -07001047 dc->writeback_thread = kthread_create(bch_writeback_thread, dc,
1048 "bcache_writeback");
Coly Li804f3c62018-03-18 17:36:14 -07001049 if (IS_ERR(dc->writeback_thread)) {
1050 cached_dev_put(dc);
Coly Lif54d8012019-06-28 19:59:44 +08001051 destroy_workqueue(dc->writeback_write_wq);
Kent Overstreet5e6926da2013-07-24 17:50:06 -07001052 return PTR_ERR(dc->writeback_thread);
Coly Li804f3c62018-03-18 17:36:14 -07001053 }
Shenghui Wang79b79142018-12-13 22:53:50 +08001054 dc->writeback_running = true;
Kent Overstreet5e6926da2013-07-24 17:50:06 -07001055
Coly Li3fd47bf2018-03-18 17:36:16 -07001056 WARN_ON(test_and_set_bit(BCACHE_DEV_WB_RUNNING, &dc->disk.flags));
Kent Overstreetcafe5632013-03-23 16:11:31 -07001057 schedule_delayed_work(&dc->writeback_rate_update,
1058 dc->writeback_rate_update_seconds * HZ);
Kent Overstreetcafe5632013-03-23 16:11:31 -07001059
Slava Pestov9e5c3532014-05-01 13:48:57 -07001060 bch_writeback_queue(dc);
1061
Kent Overstreetcafe5632013-03-23 16:11:31 -07001062 return 0;
1063}