xref: /linux/drivers/md/bcache/writeback.c (revision fdfc374af5dc345fbb9686921fa60176c1c41da0)
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 
20 static 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 */
30 static 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 
61 static 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 
160 static 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  */
207 static 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 
242 static 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 
312 static 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 
322 struct dirty_io {
323 	struct closure		cl;
324 	struct cached_dev	*dc;
325 	uint16_t		sequence;
326 	struct bio		bio;
327 };
328 
329 static 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 
344 static void dirty_io_destructor(struct closure *cl)
345 {
346 	struct dirty_io *io = container_of(cl, struct dirty_io, cl);
347 
348 	kfree(io);
349 }
350 
351 static void write_dirty_finish(struct closure *cl)
352 {
353 	struct dirty_io *io = container_of(cl, 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 
390 static 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 
403 static void write_dirty(struct closure *cl)
404 {
405 	struct dirty_io *io = container_of(cl, 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 
452 static 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 
465 static void read_dirty_submit(struct closure *cl)
466 {
467 	struct dirty_io *io = container_of(cl, 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 
474 static 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) {
581 err_free:
582 		kfree(w->private);
583 err:
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 
596 void 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 
641 static 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 
652 static 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;
687 next:
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  */
701 static 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 
739 static 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 
847 struct sectors_dirty_init {
848 	struct btree_op	op;
849 	unsigned int	inode;
850 	size_t		count;
851 };
852 
853 static 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 
872 static 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 
906 static 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 	cur_idx = 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 
956 out:
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 
965 static 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 
977 void bch_sectors_dirty_init(struct bcache_device *d)
978 {
979 	int i;
980 	struct bkey *k = NULL;
981 	struct btree_iter iter;
982 	struct sectors_dirty_init op;
983 	struct cache_set *c = d->c;
984 	struct bch_dirty_init_state state;
985 
986 	/* Just count root keys if no leaf node */
987 	rw_lock(0, c->root, c->root->level);
988 	if (c->root->level == 0) {
989 		bch_btree_op_init(&op.op, -1);
990 		op.inode = d->id;
991 		op.count = 0;
992 
993 		for_each_key_filter(&c->root->keys,
994 				    k, &iter, bch_ptr_invalid)
995 			sectors_dirty_init_fn(&op.op, c->root, k);
996 
997 		rw_unlock(0, c->root);
998 		return;
999 	}
1000 
1001 	memset(&state, 0, sizeof(struct bch_dirty_init_state));
1002 	state.c = c;
1003 	state.d = d;
1004 	state.total_threads = bch_btre_dirty_init_thread_nr();
1005 	state.key_idx = 0;
1006 	spin_lock_init(&state.idx_lock);
1007 	atomic_set(&state.started, 0);
1008 	atomic_set(&state.enough, 0);
1009 	init_waitqueue_head(&state.wait);
1010 
1011 	for (i = 0; i < state.total_threads; i++) {
1012 		/* Fetch latest state.enough earlier */
1013 		smp_mb__before_atomic();
1014 		if (atomic_read(&state.enough))
1015 			break;
1016 
1017 		state.infos[i].state = &state;
1018 		state.infos[i].thread =
1019 			kthread_run(bch_dirty_init_thread, &state.infos[i],
1020 				    "bch_dirtcnt[%d]", i);
1021 		if (IS_ERR(state.infos[i].thread)) {
1022 			pr_err("fails to run thread bch_dirty_init[%d]\n", i);
1023 			for (--i; i >= 0; i--)
1024 				kthread_stop(state.infos[i].thread);
1025 			goto out;
1026 		}
1027 		atomic_inc(&state.started);
1028 	}
1029 
1030 out:
1031 	/* Must wait for all threads to stop. */
1032 	wait_event(state.wait, atomic_read(&state.started) == 0);
1033 	rw_unlock(0, c->root);
1034 }
1035 
1036 void bch_cached_dev_writeback_init(struct cached_dev *dc)
1037 {
1038 	sema_init(&dc->in_flight, 64);
1039 	init_rwsem(&dc->writeback_lock);
1040 	bch_keybuf_init(&dc->writeback_keys);
1041 
1042 	dc->writeback_metadata		= true;
1043 	dc->writeback_running		= false;
1044 	dc->writeback_consider_fragment = true;
1045 	dc->writeback_percent		= 10;
1046 	dc->writeback_delay		= 30;
1047 	atomic_long_set(&dc->writeback_rate.rate, 1024);
1048 	dc->writeback_rate_minimum	= 8;
1049 
1050 	dc->writeback_rate_update_seconds = WRITEBACK_RATE_UPDATE_SECS_DEFAULT;
1051 	dc->writeback_rate_p_term_inverse = 40;
1052 	dc->writeback_rate_fp_term_low = 1;
1053 	dc->writeback_rate_fp_term_mid = 10;
1054 	dc->writeback_rate_fp_term_high = 1000;
1055 	dc->writeback_rate_i_term_inverse = 10000;
1056 
1057 	/* For dc->writeback_lock contention in update_writeback_rate() */
1058 	dc->rate_update_retry = 0;
1059 
1060 	WARN_ON(test_and_clear_bit(BCACHE_DEV_WB_RUNNING, &dc->disk.flags));
1061 	INIT_DELAYED_WORK(&dc->writeback_rate_update, update_writeback_rate);
1062 }
1063 
1064 int bch_cached_dev_writeback_start(struct cached_dev *dc)
1065 {
1066 	dc->writeback_write_wq = alloc_workqueue("bcache_writeback_wq",
1067 						WQ_MEM_RECLAIM, 0);
1068 	if (!dc->writeback_write_wq)
1069 		return -ENOMEM;
1070 
1071 	cached_dev_get(dc);
1072 	dc->writeback_thread = kthread_create(bch_writeback_thread, dc,
1073 					      "bcache_writeback");
1074 	if (IS_ERR(dc->writeback_thread)) {
1075 		cached_dev_put(dc);
1076 		destroy_workqueue(dc->writeback_write_wq);
1077 		return PTR_ERR(dc->writeback_thread);
1078 	}
1079 	dc->writeback_running = true;
1080 
1081 	WARN_ON(test_and_set_bit(BCACHE_DEV_WB_RUNNING, &dc->disk.flags));
1082 	schedule_delayed_work(&dc->writeback_rate_update,
1083 			      dc->writeback_rate_update_seconds * HZ);
1084 
1085 	bch_writeback_queue(dc);
1086 
1087 	return 0;
1088 }
1089