xref: /openbmc/linux/drivers/md/bcache/writeback.c (revision 920b0442)
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 set_at_max_writeback_rate(struct cache_set *c,
161 				       struct cached_dev *dc)
162 {
163 	/* Don't sst max writeback rate if it is disabled */
164 	if (!c->idle_max_writeback_rate_enabled)
165 		return false;
166 
167 	/* Don't set max writeback rate if gc is running */
168 	if (!c->gc_mark_valid)
169 		return false;
170 	/*
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 
211 static 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);
216 	struct cache_set *c = dc->disk.c;
217 
218 	/*
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 */
224 	smp_mb__after_atomic();
225 
226 	/*
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)) {
232 		clear_bit(BCACHE_DEV_RATE_DW_RUNNING, &dc->disk.flags);
233 		/* paired with where BCACHE_DEV_RATE_DW_RUNNING is tested */
234 		smp_mb__after_atomic();
235 		return;
236 	}
237 
238 	/*
239 	 * If the whole cache set is idle, set_at_max_writeback_rate()
240 	 * will set writeback rate to a max number. Then it is
241 	 * unncessary to update writeback rate for an idle cache set
242 	 * in maximum writeback rate number(s).
243 	 */
244 	if (atomic_read(&dc->has_dirty) && dc->writeback_percent &&
245 	    !set_at_max_writeback_rate(c, dc)) {
246 		do {
247 			if (!down_read_trylock((&dc->writeback_lock))) {
248 				dc->rate_update_retry++;
249 				if (dc->rate_update_retry <=
250 				    BCH_WBRATE_UPDATE_MAX_SKIPS)
251 					break;
252 				down_read(&dc->writeback_lock);
253 				dc->rate_update_retry = 0;
254 			}
255 			__update_writeback_rate(dc);
256 			update_gc_after_writeback(c);
257 			up_read(&dc->writeback_lock);
258 		} while (0);
259 	}
260 
261 
262 	/*
263 	 * CACHE_SET_IO_DISABLE might be set via sysfs interface,
264 	 * check it here too.
265 	 */
266 	if (test_bit(BCACHE_DEV_WB_RUNNING, &dc->disk.flags) &&
267 	    !test_bit(CACHE_SET_IO_DISABLE, &c->flags)) {
268 		schedule_delayed_work(&dc->writeback_rate_update,
269 			      dc->writeback_rate_update_seconds * HZ);
270 	}
271 
272 	/*
273 	 * should check BCACHE_DEV_RATE_DW_RUNNING before calling
274 	 * cancel_delayed_work_sync().
275 	 */
276 	clear_bit(BCACHE_DEV_RATE_DW_RUNNING, &dc->disk.flags);
277 	/* paired with where BCACHE_DEV_RATE_DW_RUNNING is tested */
278 	smp_mb__after_atomic();
279 }
280 
281 static unsigned int writeback_delay(struct cached_dev *dc,
282 				    unsigned int sectors)
283 {
284 	if (test_bit(BCACHE_DEV_DETACHING, &dc->disk.flags) ||
285 	    !dc->writeback_percent)
286 		return 0;
287 
288 	return bch_next_delay(&dc->writeback_rate, sectors);
289 }
290 
291 struct dirty_io {
292 	struct closure		cl;
293 	struct cached_dev	*dc;
294 	uint16_t		sequence;
295 	struct bio		bio;
296 };
297 
298 static void dirty_init(struct keybuf_key *w)
299 {
300 	struct dirty_io *io = w->private;
301 	struct bio *bio = &io->bio;
302 
303 	bio_init(bio, NULL, bio->bi_inline_vecs,
304 		 DIV_ROUND_UP(KEY_SIZE(&w->key), PAGE_SECTORS), 0);
305 	if (!io->dc->writeback_percent)
306 		bio_set_prio(bio, IOPRIO_PRIO_VALUE(IOPRIO_CLASS_IDLE, 0));
307 
308 	bio->bi_iter.bi_size	= KEY_SIZE(&w->key) << 9;
309 	bio->bi_private		= w;
310 	bch_bio_map(bio, NULL);
311 }
312 
313 static void dirty_io_destructor(struct closure *cl)
314 {
315 	struct dirty_io *io = container_of(cl, struct dirty_io, cl);
316 
317 	kfree(io);
318 }
319 
320 static void write_dirty_finish(struct closure *cl)
321 {
322 	struct dirty_io *io = container_of(cl, struct dirty_io, cl);
323 	struct keybuf_key *w = io->bio.bi_private;
324 	struct cached_dev *dc = io->dc;
325 
326 	bio_free_pages(&io->bio);
327 
328 	/* This is kind of a dumb way of signalling errors. */
329 	if (KEY_DIRTY(&w->key)) {
330 		int ret;
331 		unsigned int i;
332 		struct keylist keys;
333 
334 		bch_keylist_init(&keys);
335 
336 		bkey_copy(keys.top, &w->key);
337 		SET_KEY_DIRTY(keys.top, false);
338 		bch_keylist_push(&keys);
339 
340 		for (i = 0; i < KEY_PTRS(&w->key); i++)
341 			atomic_inc(&PTR_BUCKET(dc->disk.c, &w->key, i)->pin);
342 
343 		ret = bch_btree_insert(dc->disk.c, &keys, NULL, &w->key);
344 
345 		if (ret)
346 			trace_bcache_writeback_collision(&w->key);
347 
348 		atomic_long_inc(ret
349 				? &dc->disk.c->writeback_keys_failed
350 				: &dc->disk.c->writeback_keys_done);
351 	}
352 
353 	bch_keybuf_del(&dc->writeback_keys, w);
354 	up(&dc->in_flight);
355 
356 	closure_return_with_destructor(cl, dirty_io_destructor);
357 }
358 
359 static void dirty_endio(struct bio *bio)
360 {
361 	struct keybuf_key *w = bio->bi_private;
362 	struct dirty_io *io = w->private;
363 
364 	if (bio->bi_status) {
365 		SET_KEY_DIRTY(&w->key, false);
366 		bch_count_backing_io_errors(io->dc, bio);
367 	}
368 
369 	closure_put(&io->cl);
370 }
371 
372 static void write_dirty(struct closure *cl)
373 {
374 	struct dirty_io *io = container_of(cl, struct dirty_io, cl);
375 	struct keybuf_key *w = io->bio.bi_private;
376 	struct cached_dev *dc = io->dc;
377 
378 	uint16_t next_sequence;
379 
380 	if (atomic_read(&dc->writeback_sequence_next) != io->sequence) {
381 		/* Not our turn to write; wait for a write to complete */
382 		closure_wait(&dc->writeback_ordering_wait, cl);
383 
384 		if (atomic_read(&dc->writeback_sequence_next) == io->sequence) {
385 			/*
386 			 * Edge case-- it happened in indeterminate order
387 			 * relative to when we were added to wait list..
388 			 */
389 			closure_wake_up(&dc->writeback_ordering_wait);
390 		}
391 
392 		continue_at(cl, write_dirty, io->dc->writeback_write_wq);
393 		return;
394 	}
395 
396 	next_sequence = io->sequence + 1;
397 
398 	/*
399 	 * IO errors are signalled using the dirty bit on the key.
400 	 * If we failed to read, we should not attempt to write to the
401 	 * backing device.  Instead, immediately go to write_dirty_finish
402 	 * to clean up.
403 	 */
404 	if (KEY_DIRTY(&w->key)) {
405 		dirty_init(w);
406 		bio_set_op_attrs(&io->bio, REQ_OP_WRITE, 0);
407 		io->bio.bi_iter.bi_sector = KEY_START(&w->key);
408 		bio_set_dev(&io->bio, io->dc->bdev);
409 		io->bio.bi_end_io	= dirty_endio;
410 
411 		/* I/O request sent to backing device */
412 		closure_bio_submit(io->dc->disk.c, &io->bio, cl);
413 	}
414 
415 	atomic_set(&dc->writeback_sequence_next, next_sequence);
416 	closure_wake_up(&dc->writeback_ordering_wait);
417 
418 	continue_at(cl, write_dirty_finish, io->dc->writeback_write_wq);
419 }
420 
421 static void read_dirty_endio(struct bio *bio)
422 {
423 	struct keybuf_key *w = bio->bi_private;
424 	struct dirty_io *io = w->private;
425 
426 	/* is_read = 1 */
427 	bch_count_io_errors(io->dc->disk.c->cache,
428 			    bio->bi_status, 1,
429 			    "reading dirty data from cache");
430 
431 	dirty_endio(bio);
432 }
433 
434 static void read_dirty_submit(struct closure *cl)
435 {
436 	struct dirty_io *io = container_of(cl, struct dirty_io, cl);
437 
438 	closure_bio_submit(io->dc->disk.c, &io->bio, cl);
439 
440 	continue_at(cl, write_dirty, io->dc->writeback_write_wq);
441 }
442 
443 static void read_dirty(struct cached_dev *dc)
444 {
445 	unsigned int delay = 0;
446 	struct keybuf_key *next, *keys[MAX_WRITEBACKS_IN_PASS], *w;
447 	size_t size;
448 	int nk, i;
449 	struct dirty_io *io;
450 	struct closure cl;
451 	uint16_t sequence = 0;
452 
453 	BUG_ON(!llist_empty(&dc->writeback_ordering_wait.list));
454 	atomic_set(&dc->writeback_sequence_next, sequence);
455 	closure_init_stack(&cl);
456 
457 	/*
458 	 * XXX: if we error, background writeback just spins. Should use some
459 	 * mempools.
460 	 */
461 
462 	next = bch_keybuf_next(&dc->writeback_keys);
463 
464 	while (!kthread_should_stop() &&
465 	       !test_bit(CACHE_SET_IO_DISABLE, &dc->disk.c->flags) &&
466 	       next) {
467 		size = 0;
468 		nk = 0;
469 
470 		do {
471 			BUG_ON(ptr_stale(dc->disk.c, &next->key, 0));
472 
473 			/*
474 			 * Don't combine too many operations, even if they
475 			 * are all small.
476 			 */
477 			if (nk >= MAX_WRITEBACKS_IN_PASS)
478 				break;
479 
480 			/*
481 			 * If the current operation is very large, don't
482 			 * further combine operations.
483 			 */
484 			if (size >= MAX_WRITESIZE_IN_PASS)
485 				break;
486 
487 			/*
488 			 * Operations are only eligible to be combined
489 			 * if they are contiguous.
490 			 *
491 			 * TODO: add a heuristic willing to fire a
492 			 * certain amount of non-contiguous IO per pass,
493 			 * so that we can benefit from backing device
494 			 * command queueing.
495 			 */
496 			if ((nk != 0) && bkey_cmp(&keys[nk-1]->key,
497 						&START_KEY(&next->key)))
498 				break;
499 
500 			size += KEY_SIZE(&next->key);
501 			keys[nk++] = next;
502 		} while ((next = bch_keybuf_next(&dc->writeback_keys)));
503 
504 		/* Now we have gathered a set of 1..5 keys to write back. */
505 		for (i = 0; i < nk; i++) {
506 			w = keys[i];
507 
508 			io = kzalloc(struct_size(io, bio.bi_inline_vecs,
509 						DIV_ROUND_UP(KEY_SIZE(&w->key), PAGE_SECTORS)),
510 				     GFP_KERNEL);
511 			if (!io)
512 				goto err;
513 
514 			w->private	= io;
515 			io->dc		= dc;
516 			io->sequence    = sequence++;
517 
518 			dirty_init(w);
519 			bio_set_op_attrs(&io->bio, REQ_OP_READ, 0);
520 			io->bio.bi_iter.bi_sector = PTR_OFFSET(&w->key, 0);
521 			bio_set_dev(&io->bio, dc->disk.c->cache->bdev);
522 			io->bio.bi_end_io	= read_dirty_endio;
523 
524 			if (bch_bio_alloc_pages(&io->bio, GFP_KERNEL))
525 				goto err_free;
526 
527 			trace_bcache_writeback(&w->key);
528 
529 			down(&dc->in_flight);
530 
531 			/*
532 			 * We've acquired a semaphore for the maximum
533 			 * simultaneous number of writebacks; from here
534 			 * everything happens asynchronously.
535 			 */
536 			closure_call(&io->cl, read_dirty_submit, NULL, &cl);
537 		}
538 
539 		delay = writeback_delay(dc, size);
540 
541 		while (!kthread_should_stop() &&
542 		       !test_bit(CACHE_SET_IO_DISABLE, &dc->disk.c->flags) &&
543 		       delay) {
544 			schedule_timeout_interruptible(delay);
545 			delay = writeback_delay(dc, 0);
546 		}
547 	}
548 
549 	if (0) {
550 err_free:
551 		kfree(w->private);
552 err:
553 		bch_keybuf_del(&dc->writeback_keys, w);
554 	}
555 
556 	/*
557 	 * Wait for outstanding writeback IOs to finish (and keybuf slots to be
558 	 * freed) before refilling again
559 	 */
560 	closure_sync(&cl);
561 }
562 
563 /* Scan for dirty data */
564 
565 void bcache_dev_sectors_dirty_add(struct cache_set *c, unsigned int inode,
566 				  uint64_t offset, int nr_sectors)
567 {
568 	struct bcache_device *d = c->devices[inode];
569 	unsigned int stripe_offset, sectors_dirty;
570 	int stripe;
571 
572 	if (!d)
573 		return;
574 
575 	stripe = offset_to_stripe(d, offset);
576 	if (stripe < 0)
577 		return;
578 
579 	if (UUID_FLASH_ONLY(&c->uuids[inode]))
580 		atomic_long_add(nr_sectors, &c->flash_dev_dirty_sectors);
581 
582 	stripe_offset = offset & (d->stripe_size - 1);
583 
584 	while (nr_sectors) {
585 		int s = min_t(unsigned int, abs(nr_sectors),
586 			      d->stripe_size - stripe_offset);
587 
588 		if (nr_sectors < 0)
589 			s = -s;
590 
591 		if (stripe >= d->nr_stripes)
592 			return;
593 
594 		sectors_dirty = atomic_add_return(s,
595 					d->stripe_sectors_dirty + stripe);
596 		if (sectors_dirty == d->stripe_size) {
597 			if (!test_bit(stripe, d->full_dirty_stripes))
598 				set_bit(stripe, d->full_dirty_stripes);
599 		} else {
600 			if (test_bit(stripe, d->full_dirty_stripes))
601 				clear_bit(stripe, d->full_dirty_stripes);
602 		}
603 
604 		nr_sectors -= s;
605 		stripe_offset = 0;
606 		stripe++;
607 	}
608 }
609 
610 static bool dirty_pred(struct keybuf *buf, struct bkey *k)
611 {
612 	struct cached_dev *dc = container_of(buf,
613 					     struct cached_dev,
614 					     writeback_keys);
615 
616 	BUG_ON(KEY_INODE(k) != dc->disk.id);
617 
618 	return KEY_DIRTY(k);
619 }
620 
621 static void refill_full_stripes(struct cached_dev *dc)
622 {
623 	struct keybuf *buf = &dc->writeback_keys;
624 	unsigned int start_stripe, next_stripe;
625 	int stripe;
626 	bool wrapped = false;
627 
628 	stripe = offset_to_stripe(&dc->disk, KEY_OFFSET(&buf->last_scanned));
629 	if (stripe < 0)
630 		stripe = 0;
631 
632 	start_stripe = stripe;
633 
634 	while (1) {
635 		stripe = find_next_bit(dc->disk.full_dirty_stripes,
636 				       dc->disk.nr_stripes, stripe);
637 
638 		if (stripe == dc->disk.nr_stripes)
639 			goto next;
640 
641 		next_stripe = find_next_zero_bit(dc->disk.full_dirty_stripes,
642 						 dc->disk.nr_stripes, stripe);
643 
644 		buf->last_scanned = KEY(dc->disk.id,
645 					stripe * dc->disk.stripe_size, 0);
646 
647 		bch_refill_keybuf(dc->disk.c, buf,
648 				  &KEY(dc->disk.id,
649 				       next_stripe * dc->disk.stripe_size, 0),
650 				  dirty_pred);
651 
652 		if (array_freelist_empty(&buf->freelist))
653 			return;
654 
655 		stripe = next_stripe;
656 next:
657 		if (wrapped && stripe > start_stripe)
658 			return;
659 
660 		if (stripe == dc->disk.nr_stripes) {
661 			stripe = 0;
662 			wrapped = true;
663 		}
664 	}
665 }
666 
667 /*
668  * Returns true if we scanned the entire disk
669  */
670 static bool refill_dirty(struct cached_dev *dc)
671 {
672 	struct keybuf *buf = &dc->writeback_keys;
673 	struct bkey start = KEY(dc->disk.id, 0, 0);
674 	struct bkey end = KEY(dc->disk.id, MAX_KEY_OFFSET, 0);
675 	struct bkey start_pos;
676 
677 	/*
678 	 * make sure keybuf pos is inside the range for this disk - at bringup
679 	 * we might not be attached yet so this disk's inode nr isn't
680 	 * initialized then
681 	 */
682 	if (bkey_cmp(&buf->last_scanned, &start) < 0 ||
683 	    bkey_cmp(&buf->last_scanned, &end) > 0)
684 		buf->last_scanned = start;
685 
686 	if (dc->partial_stripes_expensive) {
687 		refill_full_stripes(dc);
688 		if (array_freelist_empty(&buf->freelist))
689 			return false;
690 	}
691 
692 	start_pos = buf->last_scanned;
693 	bch_refill_keybuf(dc->disk.c, buf, &end, dirty_pred);
694 
695 	if (bkey_cmp(&buf->last_scanned, &end) < 0)
696 		return false;
697 
698 	/*
699 	 * If we get to the end start scanning again from the beginning, and
700 	 * only scan up to where we initially started scanning from:
701 	 */
702 	buf->last_scanned = start;
703 	bch_refill_keybuf(dc->disk.c, buf, &start_pos, dirty_pred);
704 
705 	return bkey_cmp(&buf->last_scanned, &start_pos) >= 0;
706 }
707 
708 static int bch_writeback_thread(void *arg)
709 {
710 	struct cached_dev *dc = arg;
711 	struct cache_set *c = dc->disk.c;
712 	bool searched_full_index;
713 
714 	bch_ratelimit_reset(&dc->writeback_rate);
715 
716 	while (!kthread_should_stop() &&
717 	       !test_bit(CACHE_SET_IO_DISABLE, &c->flags)) {
718 		down_write(&dc->writeback_lock);
719 		set_current_state(TASK_INTERRUPTIBLE);
720 		/*
721 		 * If the bache device is detaching, skip here and continue
722 		 * to perform writeback. Otherwise, if no dirty data on cache,
723 		 * or there is dirty data on cache but writeback is disabled,
724 		 * the writeback thread should sleep here and wait for others
725 		 * to wake up it.
726 		 */
727 		if (!test_bit(BCACHE_DEV_DETACHING, &dc->disk.flags) &&
728 		    (!atomic_read(&dc->has_dirty) || !dc->writeback_running)) {
729 			up_write(&dc->writeback_lock);
730 
731 			if (kthread_should_stop() ||
732 			    test_bit(CACHE_SET_IO_DISABLE, &c->flags)) {
733 				set_current_state(TASK_RUNNING);
734 				break;
735 			}
736 
737 			schedule();
738 			continue;
739 		}
740 		set_current_state(TASK_RUNNING);
741 
742 		searched_full_index = refill_dirty(dc);
743 
744 		if (searched_full_index &&
745 		    RB_EMPTY_ROOT(&dc->writeback_keys.keys)) {
746 			atomic_set(&dc->has_dirty, 0);
747 			SET_BDEV_STATE(&dc->sb, BDEV_STATE_CLEAN);
748 			bch_write_bdev_super(dc, NULL);
749 			/*
750 			 * If bcache device is detaching via sysfs interface,
751 			 * writeback thread should stop after there is no dirty
752 			 * data on cache. BCACHE_DEV_DETACHING flag is set in
753 			 * bch_cached_dev_detach().
754 			 */
755 			if (test_bit(BCACHE_DEV_DETACHING, &dc->disk.flags)) {
756 				struct closure cl;
757 
758 				closure_init_stack(&cl);
759 				memset(&dc->sb.set_uuid, 0, 16);
760 				SET_BDEV_STATE(&dc->sb, BDEV_STATE_NONE);
761 
762 				bch_write_bdev_super(dc, &cl);
763 				closure_sync(&cl);
764 
765 				up_write(&dc->writeback_lock);
766 				break;
767 			}
768 
769 			/*
770 			 * When dirty data rate is high (e.g. 50%+), there might
771 			 * be heavy buckets fragmentation after writeback
772 			 * finished, which hurts following write performance.
773 			 * If users really care about write performance they
774 			 * may set BCH_ENABLE_AUTO_GC via sysfs, then when
775 			 * BCH_DO_AUTO_GC is set, garbage collection thread
776 			 * will be wake up here. After moving gc, the shrunk
777 			 * btree and discarded free buckets SSD space may be
778 			 * helpful for following write requests.
779 			 */
780 			if (c->gc_after_writeback ==
781 			    (BCH_ENABLE_AUTO_GC|BCH_DO_AUTO_GC)) {
782 				c->gc_after_writeback &= ~BCH_DO_AUTO_GC;
783 				force_wake_up_gc(c);
784 			}
785 		}
786 
787 		up_write(&dc->writeback_lock);
788 
789 		read_dirty(dc);
790 
791 		if (searched_full_index) {
792 			unsigned int delay = dc->writeback_delay * HZ;
793 
794 			while (delay &&
795 			       !kthread_should_stop() &&
796 			       !test_bit(CACHE_SET_IO_DISABLE, &c->flags) &&
797 			       !test_bit(BCACHE_DEV_DETACHING, &dc->disk.flags))
798 				delay = schedule_timeout_interruptible(delay);
799 
800 			bch_ratelimit_reset(&dc->writeback_rate);
801 		}
802 	}
803 
804 	if (dc->writeback_write_wq) {
805 		flush_workqueue(dc->writeback_write_wq);
806 		destroy_workqueue(dc->writeback_write_wq);
807 	}
808 	cached_dev_put(dc);
809 	wait_for_kthread_stop();
810 
811 	return 0;
812 }
813 
814 /* Init */
815 #define INIT_KEYS_EACH_TIME	500000
816 
817 struct sectors_dirty_init {
818 	struct btree_op	op;
819 	unsigned int	inode;
820 	size_t		count;
821 };
822 
823 static int sectors_dirty_init_fn(struct btree_op *_op, struct btree *b,
824 				 struct bkey *k)
825 {
826 	struct sectors_dirty_init *op = container_of(_op,
827 						struct sectors_dirty_init, op);
828 	if (KEY_INODE(k) > op->inode)
829 		return MAP_DONE;
830 
831 	if (KEY_DIRTY(k))
832 		bcache_dev_sectors_dirty_add(b->c, KEY_INODE(k),
833 					     KEY_START(k), KEY_SIZE(k));
834 
835 	op->count++;
836 	if (!(op->count % INIT_KEYS_EACH_TIME))
837 		cond_resched();
838 
839 	return MAP_CONTINUE;
840 }
841 
842 static int bch_root_node_dirty_init(struct cache_set *c,
843 				     struct bcache_device *d,
844 				     struct bkey *k)
845 {
846 	struct sectors_dirty_init op;
847 	int ret;
848 
849 	bch_btree_op_init(&op.op, -1);
850 	op.inode = d->id;
851 	op.count = 0;
852 
853 	ret = bcache_btree(map_keys_recurse,
854 			   k,
855 			   c->root,
856 			   &op.op,
857 			   &KEY(op.inode, 0, 0),
858 			   sectors_dirty_init_fn,
859 			   0);
860 	if (ret < 0)
861 		pr_warn("sectors dirty init failed, ret=%d!\n", ret);
862 
863 	return ret;
864 }
865 
866 static int bch_dirty_init_thread(void *arg)
867 {
868 	struct dirty_init_thrd_info *info = arg;
869 	struct bch_dirty_init_state *state = info->state;
870 	struct cache_set *c = state->c;
871 	struct btree_iter iter;
872 	struct bkey *k, *p;
873 	int cur_idx, prev_idx, skip_nr;
874 
875 	k = p = NULL;
876 	cur_idx = prev_idx = 0;
877 
878 	bch_btree_iter_init(&c->root->keys, &iter, NULL);
879 	k = bch_btree_iter_next_filter(&iter, &c->root->keys, bch_ptr_bad);
880 	BUG_ON(!k);
881 
882 	p = k;
883 
884 	while (k) {
885 		spin_lock(&state->idx_lock);
886 		cur_idx = state->key_idx;
887 		state->key_idx++;
888 		spin_unlock(&state->idx_lock);
889 
890 		skip_nr = cur_idx - prev_idx;
891 
892 		while (skip_nr) {
893 			k = bch_btree_iter_next_filter(&iter,
894 						       &c->root->keys,
895 						       bch_ptr_bad);
896 			if (k)
897 				p = k;
898 			else {
899 				atomic_set(&state->enough, 1);
900 				/* Update state->enough earlier */
901 				smp_mb__after_atomic();
902 				goto out;
903 			}
904 			skip_nr--;
905 		}
906 
907 		if (p) {
908 			if (bch_root_node_dirty_init(c, state->d, p) < 0)
909 				goto out;
910 		}
911 
912 		p = NULL;
913 		prev_idx = cur_idx;
914 	}
915 
916 out:
917 	/* In order to wake up state->wait in time */
918 	smp_mb__before_atomic();
919 	if (atomic_dec_and_test(&state->started))
920 		wake_up(&state->wait);
921 
922 	return 0;
923 }
924 
925 static int bch_btre_dirty_init_thread_nr(void)
926 {
927 	int n = num_online_cpus()/2;
928 
929 	if (n == 0)
930 		n = 1;
931 	else if (n > BCH_DIRTY_INIT_THRD_MAX)
932 		n = BCH_DIRTY_INIT_THRD_MAX;
933 
934 	return n;
935 }
936 
937 void bch_sectors_dirty_init(struct bcache_device *d)
938 {
939 	int i;
940 	struct bkey *k = NULL;
941 	struct btree_iter iter;
942 	struct sectors_dirty_init op;
943 	struct cache_set *c = d->c;
944 	struct bch_dirty_init_state state;
945 
946 	/* Just count root keys if no leaf node */
947 	rw_lock(0, c->root, c->root->level);
948 	if (c->root->level == 0) {
949 		bch_btree_op_init(&op.op, -1);
950 		op.inode = d->id;
951 		op.count = 0;
952 
953 		for_each_key_filter(&c->root->keys,
954 				    k, &iter, bch_ptr_invalid)
955 			sectors_dirty_init_fn(&op.op, c->root, k);
956 
957 		rw_unlock(0, c->root);
958 		return;
959 	}
960 
961 	memset(&state, 0, sizeof(struct bch_dirty_init_state));
962 	state.c = c;
963 	state.d = d;
964 	state.total_threads = bch_btre_dirty_init_thread_nr();
965 	state.key_idx = 0;
966 	spin_lock_init(&state.idx_lock);
967 	atomic_set(&state.started, 0);
968 	atomic_set(&state.enough, 0);
969 	init_waitqueue_head(&state.wait);
970 
971 	for (i = 0; i < state.total_threads; i++) {
972 		/* Fetch latest state.enough earlier */
973 		smp_mb__before_atomic();
974 		if (atomic_read(&state.enough))
975 			break;
976 
977 		state.infos[i].state = &state;
978 		state.infos[i].thread =
979 			kthread_run(bch_dirty_init_thread, &state.infos[i],
980 				    "bch_dirtcnt[%d]", i);
981 		if (IS_ERR(state.infos[i].thread)) {
982 			pr_err("fails to run thread bch_dirty_init[%d]\n", i);
983 			for (--i; i >= 0; i--)
984 				kthread_stop(state.infos[i].thread);
985 			goto out;
986 		}
987 		atomic_inc(&state.started);
988 	}
989 
990 out:
991 	/* Must wait for all threads to stop. */
992 	wait_event(state.wait, atomic_read(&state.started) == 0);
993 	rw_unlock(0, c->root);
994 }
995 
996 void bch_cached_dev_writeback_init(struct cached_dev *dc)
997 {
998 	sema_init(&dc->in_flight, 64);
999 	init_rwsem(&dc->writeback_lock);
1000 	bch_keybuf_init(&dc->writeback_keys);
1001 
1002 	dc->writeback_metadata		= true;
1003 	dc->writeback_running		= false;
1004 	dc->writeback_consider_fragment = true;
1005 	dc->writeback_percent		= 10;
1006 	dc->writeback_delay		= 30;
1007 	atomic_long_set(&dc->writeback_rate.rate, 1024);
1008 	dc->writeback_rate_minimum	= 8;
1009 
1010 	dc->writeback_rate_update_seconds = WRITEBACK_RATE_UPDATE_SECS_DEFAULT;
1011 	dc->writeback_rate_p_term_inverse = 40;
1012 	dc->writeback_rate_fp_term_low = 1;
1013 	dc->writeback_rate_fp_term_mid = 10;
1014 	dc->writeback_rate_fp_term_high = 1000;
1015 	dc->writeback_rate_i_term_inverse = 10000;
1016 
1017 	/* For dc->writeback_lock contention in update_writeback_rate() */
1018 	dc->rate_update_retry = 0;
1019 
1020 	WARN_ON(test_and_clear_bit(BCACHE_DEV_WB_RUNNING, &dc->disk.flags));
1021 	INIT_DELAYED_WORK(&dc->writeback_rate_update, update_writeback_rate);
1022 }
1023 
1024 int bch_cached_dev_writeback_start(struct cached_dev *dc)
1025 {
1026 	dc->writeback_write_wq = alloc_workqueue("bcache_writeback_wq",
1027 						WQ_MEM_RECLAIM, 0);
1028 	if (!dc->writeback_write_wq)
1029 		return -ENOMEM;
1030 
1031 	cached_dev_get(dc);
1032 	dc->writeback_thread = kthread_create(bch_writeback_thread, dc,
1033 					      "bcache_writeback");
1034 	if (IS_ERR(dc->writeback_thread)) {
1035 		cached_dev_put(dc);
1036 		destroy_workqueue(dc->writeback_write_wq);
1037 		return PTR_ERR(dc->writeback_thread);
1038 	}
1039 	dc->writeback_running = true;
1040 
1041 	WARN_ON(test_and_set_bit(BCACHE_DEV_WB_RUNNING, &dc->disk.flags));
1042 	schedule_delayed_work(&dc->writeback_rate_update,
1043 			      dc->writeback_rate_update_seconds * HZ);
1044 
1045 	bch_writeback_queue(dc);
1046 
1047 	return 0;
1048 }
1049