xref: /openbmc/linux/drivers/md/bcache/writeback.c (revision 29c37341)
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->sb.bucket_size -
39 				atomic_long_read(&c->flash_dev_dirty_sectors);
40 
41 	/*
42 	 * Unfortunately there is no control of global dirty data.  If the
43 	 * user states that they want 10% dirty data in the cache, and has,
44 	 * e.g., 5 backing volumes of equal size, we try and ensure each
45 	 * backing volume uses about 2% of the cache for dirty data.
46 	 */
47 	uint32_t bdev_share =
48 		div64_u64(bdev_sectors(dc->bdev) << WRITEBACK_SHARE_SHIFT,
49 				c->cached_dev_sectors);
50 
51 	uint64_t cache_dirty_target =
52 		div_u64(cache_sectors * dc->writeback_percent, 100);
53 
54 	/* Ensure each backing dev gets at least one dirty share */
55 	if (bdev_share < 1)
56 		bdev_share = 1;
57 
58 	return (cache_dirty_target * bdev_share) >> WRITEBACK_SHARE_SHIFT;
59 }
60 
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 	if ((error < 0 && dc->writeback_rate_integral > 0) ||
92 	    (error > 0 && time_before64(local_clock(),
93 			 dc->writeback_rate.next + NSEC_PER_MSEC))) {
94 		/*
95 		 * Only decrease the integral term if it's more than
96 		 * zero.  Only increase the integral term if the device
97 		 * is keeping up.  (Don't wind up the integral
98 		 * ineffectively in either case).
99 		 *
100 		 * It's necessary to scale this by
101 		 * writeback_rate_update_seconds to keep the integral
102 		 * term dimensioned properly.
103 		 */
104 		dc->writeback_rate_integral += error *
105 			dc->writeback_rate_update_seconds;
106 	}
107 
108 	integral_scaled = div_s64(dc->writeback_rate_integral,
109 			dc->writeback_rate_i_term_inverse);
110 
111 	new_rate = clamp_t(int32_t, (proportional_scaled + integral_scaled),
112 			dc->writeback_rate_minimum, NSEC_PER_SEC);
113 
114 	dc->writeback_rate_proportional = proportional_scaled;
115 	dc->writeback_rate_integral_scaled = integral_scaled;
116 	dc->writeback_rate_change = new_rate -
117 			atomic_long_read(&dc->writeback_rate.rate);
118 	atomic_long_set(&dc->writeback_rate.rate, new_rate);
119 	dc->writeback_rate_target = target;
120 }
121 
122 static bool set_at_max_writeback_rate(struct cache_set *c,
123 				       struct cached_dev *dc)
124 {
125 	/* Don't sst max writeback rate if it is disabled */
126 	if (!c->idle_max_writeback_rate_enabled)
127 		return false;
128 
129 	/* Don't set max writeback rate if gc is running */
130 	if (!c->gc_mark_valid)
131 		return false;
132 	/*
133 	 * Idle_counter is increased everytime when update_writeback_rate() is
134 	 * called. If all backing devices attached to the same cache set have
135 	 * identical dc->writeback_rate_update_seconds values, it is about 6
136 	 * rounds of update_writeback_rate() on each backing device before
137 	 * c->at_max_writeback_rate is set to 1, and then max wrteback rate set
138 	 * to each dc->writeback_rate.rate.
139 	 * In order to avoid extra locking cost for counting exact dirty cached
140 	 * devices number, c->attached_dev_nr is used to calculate the idle
141 	 * throushold. It might be bigger if not all cached device are in write-
142 	 * back mode, but it still works well with limited extra rounds of
143 	 * update_writeback_rate().
144 	 */
145 	if (atomic_inc_return(&c->idle_counter) <
146 	    atomic_read(&c->attached_dev_nr) * 6)
147 		return false;
148 
149 	if (atomic_read(&c->at_max_writeback_rate) != 1)
150 		atomic_set(&c->at_max_writeback_rate, 1);
151 
152 	atomic_long_set(&dc->writeback_rate.rate, INT_MAX);
153 
154 	/* keep writeback_rate_target as existing value */
155 	dc->writeback_rate_proportional = 0;
156 	dc->writeback_rate_integral_scaled = 0;
157 	dc->writeback_rate_change = 0;
158 
159 	/*
160 	 * Check c->idle_counter and c->at_max_writeback_rate agagain in case
161 	 * new I/O arrives during before set_at_max_writeback_rate() returns.
162 	 * Then the writeback rate is set to 1, and its new value should be
163 	 * decided via __update_writeback_rate().
164 	 */
165 	if ((atomic_read(&c->idle_counter) <
166 	     atomic_read(&c->attached_dev_nr) * 6) ||
167 	    !atomic_read(&c->at_max_writeback_rate))
168 		return false;
169 
170 	return true;
171 }
172 
173 static void update_writeback_rate(struct work_struct *work)
174 {
175 	struct cached_dev *dc = container_of(to_delayed_work(work),
176 					     struct cached_dev,
177 					     writeback_rate_update);
178 	struct cache_set *c = dc->disk.c;
179 
180 	/*
181 	 * should check BCACHE_DEV_RATE_DW_RUNNING before calling
182 	 * cancel_delayed_work_sync().
183 	 */
184 	set_bit(BCACHE_DEV_RATE_DW_RUNNING, &dc->disk.flags);
185 	/* paired with where BCACHE_DEV_RATE_DW_RUNNING is tested */
186 	smp_mb__after_atomic();
187 
188 	/*
189 	 * CACHE_SET_IO_DISABLE might be set via sysfs interface,
190 	 * check it here too.
191 	 */
192 	if (!test_bit(BCACHE_DEV_WB_RUNNING, &dc->disk.flags) ||
193 	    test_bit(CACHE_SET_IO_DISABLE, &c->flags)) {
194 		clear_bit(BCACHE_DEV_RATE_DW_RUNNING, &dc->disk.flags);
195 		/* paired with where BCACHE_DEV_RATE_DW_RUNNING is tested */
196 		smp_mb__after_atomic();
197 		return;
198 	}
199 
200 	if (atomic_read(&dc->has_dirty) && dc->writeback_percent) {
201 		/*
202 		 * If the whole cache set is idle, set_at_max_writeback_rate()
203 		 * will set writeback rate to a max number. Then it is
204 		 * unncessary to update writeback rate for an idle cache set
205 		 * in maximum writeback rate number(s).
206 		 */
207 		if (!set_at_max_writeback_rate(c, dc)) {
208 			down_read(&dc->writeback_lock);
209 			__update_writeback_rate(dc);
210 			update_gc_after_writeback(c);
211 			up_read(&dc->writeback_lock);
212 		}
213 	}
214 
215 
216 	/*
217 	 * CACHE_SET_IO_DISABLE might be set via sysfs interface,
218 	 * check it here too.
219 	 */
220 	if (test_bit(BCACHE_DEV_WB_RUNNING, &dc->disk.flags) &&
221 	    !test_bit(CACHE_SET_IO_DISABLE, &c->flags)) {
222 		schedule_delayed_work(&dc->writeback_rate_update,
223 			      dc->writeback_rate_update_seconds * HZ);
224 	}
225 
226 	/*
227 	 * should check BCACHE_DEV_RATE_DW_RUNNING before calling
228 	 * cancel_delayed_work_sync().
229 	 */
230 	clear_bit(BCACHE_DEV_RATE_DW_RUNNING, &dc->disk.flags);
231 	/* paired with where BCACHE_DEV_RATE_DW_RUNNING is tested */
232 	smp_mb__after_atomic();
233 }
234 
235 static unsigned int writeback_delay(struct cached_dev *dc,
236 				    unsigned int sectors)
237 {
238 	if (test_bit(BCACHE_DEV_DETACHING, &dc->disk.flags) ||
239 	    !dc->writeback_percent)
240 		return 0;
241 
242 	return bch_next_delay(&dc->writeback_rate, sectors);
243 }
244 
245 struct dirty_io {
246 	struct closure		cl;
247 	struct cached_dev	*dc;
248 	uint16_t		sequence;
249 	struct bio		bio;
250 };
251 
252 static void dirty_init(struct keybuf_key *w)
253 {
254 	struct dirty_io *io = w->private;
255 	struct bio *bio = &io->bio;
256 
257 	bio_init(bio, bio->bi_inline_vecs,
258 		 DIV_ROUND_UP(KEY_SIZE(&w->key), PAGE_SECTORS));
259 	if (!io->dc->writeback_percent)
260 		bio_set_prio(bio, IOPRIO_PRIO_VALUE(IOPRIO_CLASS_IDLE, 0));
261 
262 	bio->bi_iter.bi_size	= KEY_SIZE(&w->key) << 9;
263 	bio->bi_private		= w;
264 	bch_bio_map(bio, NULL);
265 }
266 
267 static void dirty_io_destructor(struct closure *cl)
268 {
269 	struct dirty_io *io = container_of(cl, struct dirty_io, cl);
270 
271 	kfree(io);
272 }
273 
274 static void write_dirty_finish(struct closure *cl)
275 {
276 	struct dirty_io *io = container_of(cl, struct dirty_io, cl);
277 	struct keybuf_key *w = io->bio.bi_private;
278 	struct cached_dev *dc = io->dc;
279 
280 	bio_free_pages(&io->bio);
281 
282 	/* This is kind of a dumb way of signalling errors. */
283 	if (KEY_DIRTY(&w->key)) {
284 		int ret;
285 		unsigned int i;
286 		struct keylist keys;
287 
288 		bch_keylist_init(&keys);
289 
290 		bkey_copy(keys.top, &w->key);
291 		SET_KEY_DIRTY(keys.top, false);
292 		bch_keylist_push(&keys);
293 
294 		for (i = 0; i < KEY_PTRS(&w->key); i++)
295 			atomic_inc(&PTR_BUCKET(dc->disk.c, &w->key, i)->pin);
296 
297 		ret = bch_btree_insert(dc->disk.c, &keys, NULL, &w->key);
298 
299 		if (ret)
300 			trace_bcache_writeback_collision(&w->key);
301 
302 		atomic_long_inc(ret
303 				? &dc->disk.c->writeback_keys_failed
304 				: &dc->disk.c->writeback_keys_done);
305 	}
306 
307 	bch_keybuf_del(&dc->writeback_keys, w);
308 	up(&dc->in_flight);
309 
310 	closure_return_with_destructor(cl, dirty_io_destructor);
311 }
312 
313 static void dirty_endio(struct bio *bio)
314 {
315 	struct keybuf_key *w = bio->bi_private;
316 	struct dirty_io *io = w->private;
317 
318 	if (bio->bi_status) {
319 		SET_KEY_DIRTY(&w->key, false);
320 		bch_count_backing_io_errors(io->dc, bio);
321 	}
322 
323 	closure_put(&io->cl);
324 }
325 
326 static void write_dirty(struct closure *cl)
327 {
328 	struct dirty_io *io = container_of(cl, struct dirty_io, cl);
329 	struct keybuf_key *w = io->bio.bi_private;
330 	struct cached_dev *dc = io->dc;
331 
332 	uint16_t next_sequence;
333 
334 	if (atomic_read(&dc->writeback_sequence_next) != io->sequence) {
335 		/* Not our turn to write; wait for a write to complete */
336 		closure_wait(&dc->writeback_ordering_wait, cl);
337 
338 		if (atomic_read(&dc->writeback_sequence_next) == io->sequence) {
339 			/*
340 			 * Edge case-- it happened in indeterminate order
341 			 * relative to when we were added to wait list..
342 			 */
343 			closure_wake_up(&dc->writeback_ordering_wait);
344 		}
345 
346 		continue_at(cl, write_dirty, io->dc->writeback_write_wq);
347 		return;
348 	}
349 
350 	next_sequence = io->sequence + 1;
351 
352 	/*
353 	 * IO errors are signalled using the dirty bit on the key.
354 	 * If we failed to read, we should not attempt to write to the
355 	 * backing device.  Instead, immediately go to write_dirty_finish
356 	 * to clean up.
357 	 */
358 	if (KEY_DIRTY(&w->key)) {
359 		dirty_init(w);
360 		bio_set_op_attrs(&io->bio, REQ_OP_WRITE, 0);
361 		io->bio.bi_iter.bi_sector = KEY_START(&w->key);
362 		bio_set_dev(&io->bio, io->dc->bdev);
363 		io->bio.bi_end_io	= dirty_endio;
364 
365 		/* I/O request sent to backing device */
366 		closure_bio_submit(io->dc->disk.c, &io->bio, cl);
367 	}
368 
369 	atomic_set(&dc->writeback_sequence_next, next_sequence);
370 	closure_wake_up(&dc->writeback_ordering_wait);
371 
372 	continue_at(cl, write_dirty_finish, io->dc->writeback_write_wq);
373 }
374 
375 static void read_dirty_endio(struct bio *bio)
376 {
377 	struct keybuf_key *w = bio->bi_private;
378 	struct dirty_io *io = w->private;
379 
380 	/* is_read = 1 */
381 	bch_count_io_errors(PTR_CACHE(io->dc->disk.c, &w->key, 0),
382 			    bio->bi_status, 1,
383 			    "reading dirty data from cache");
384 
385 	dirty_endio(bio);
386 }
387 
388 static void read_dirty_submit(struct closure *cl)
389 {
390 	struct dirty_io *io = container_of(cl, struct dirty_io, cl);
391 
392 	closure_bio_submit(io->dc->disk.c, &io->bio, cl);
393 
394 	continue_at(cl, write_dirty, io->dc->writeback_write_wq);
395 }
396 
397 static void read_dirty(struct cached_dev *dc)
398 {
399 	unsigned int delay = 0;
400 	struct keybuf_key *next, *keys[MAX_WRITEBACKS_IN_PASS], *w;
401 	size_t size;
402 	int nk, i;
403 	struct dirty_io *io;
404 	struct closure cl;
405 	uint16_t sequence = 0;
406 
407 	BUG_ON(!llist_empty(&dc->writeback_ordering_wait.list));
408 	atomic_set(&dc->writeback_sequence_next, sequence);
409 	closure_init_stack(&cl);
410 
411 	/*
412 	 * XXX: if we error, background writeback just spins. Should use some
413 	 * mempools.
414 	 */
415 
416 	next = bch_keybuf_next(&dc->writeback_keys);
417 
418 	while (!kthread_should_stop() &&
419 	       !test_bit(CACHE_SET_IO_DISABLE, &dc->disk.c->flags) &&
420 	       next) {
421 		size = 0;
422 		nk = 0;
423 
424 		do {
425 			BUG_ON(ptr_stale(dc->disk.c, &next->key, 0));
426 
427 			/*
428 			 * Don't combine too many operations, even if they
429 			 * are all small.
430 			 */
431 			if (nk >= MAX_WRITEBACKS_IN_PASS)
432 				break;
433 
434 			/*
435 			 * If the current operation is very large, don't
436 			 * further combine operations.
437 			 */
438 			if (size >= MAX_WRITESIZE_IN_PASS)
439 				break;
440 
441 			/*
442 			 * Operations are only eligible to be combined
443 			 * if they are contiguous.
444 			 *
445 			 * TODO: add a heuristic willing to fire a
446 			 * certain amount of non-contiguous IO per pass,
447 			 * so that we can benefit from backing device
448 			 * command queueing.
449 			 */
450 			if ((nk != 0) && bkey_cmp(&keys[nk-1]->key,
451 						&START_KEY(&next->key)))
452 				break;
453 
454 			size += KEY_SIZE(&next->key);
455 			keys[nk++] = next;
456 		} while ((next = bch_keybuf_next(&dc->writeback_keys)));
457 
458 		/* Now we have gathered a set of 1..5 keys to write back. */
459 		for (i = 0; i < nk; i++) {
460 			w = keys[i];
461 
462 			io = kzalloc(struct_size(io, bio.bi_inline_vecs,
463 						DIV_ROUND_UP(KEY_SIZE(&w->key), PAGE_SECTORS)),
464 				     GFP_KERNEL);
465 			if (!io)
466 				goto err;
467 
468 			w->private	= io;
469 			io->dc		= dc;
470 			io->sequence    = sequence++;
471 
472 			dirty_init(w);
473 			bio_set_op_attrs(&io->bio, REQ_OP_READ, 0);
474 			io->bio.bi_iter.bi_sector = PTR_OFFSET(&w->key, 0);
475 			bio_set_dev(&io->bio,
476 				    PTR_CACHE(dc->disk.c, &w->key, 0)->bdev);
477 			io->bio.bi_end_io	= read_dirty_endio;
478 
479 			if (bch_bio_alloc_pages(&io->bio, GFP_KERNEL))
480 				goto err_free;
481 
482 			trace_bcache_writeback(&w->key);
483 
484 			down(&dc->in_flight);
485 
486 			/*
487 			 * We've acquired a semaphore for the maximum
488 			 * simultaneous number of writebacks; from here
489 			 * everything happens asynchronously.
490 			 */
491 			closure_call(&io->cl, read_dirty_submit, NULL, &cl);
492 		}
493 
494 		delay = writeback_delay(dc, size);
495 
496 		while (!kthread_should_stop() &&
497 		       !test_bit(CACHE_SET_IO_DISABLE, &dc->disk.c->flags) &&
498 		       delay) {
499 			schedule_timeout_interruptible(delay);
500 			delay = writeback_delay(dc, 0);
501 		}
502 	}
503 
504 	if (0) {
505 err_free:
506 		kfree(w->private);
507 err:
508 		bch_keybuf_del(&dc->writeback_keys, w);
509 	}
510 
511 	/*
512 	 * Wait for outstanding writeback IOs to finish (and keybuf slots to be
513 	 * freed) before refilling again
514 	 */
515 	closure_sync(&cl);
516 }
517 
518 /* Scan for dirty data */
519 
520 void bcache_dev_sectors_dirty_add(struct cache_set *c, unsigned int inode,
521 				  uint64_t offset, int nr_sectors)
522 {
523 	struct bcache_device *d = c->devices[inode];
524 	unsigned int stripe_offset, sectors_dirty;
525 	int stripe;
526 
527 	if (!d)
528 		return;
529 
530 	stripe = offset_to_stripe(d, offset);
531 	if (stripe < 0)
532 		return;
533 
534 	if (UUID_FLASH_ONLY(&c->uuids[inode]))
535 		atomic_long_add(nr_sectors, &c->flash_dev_dirty_sectors);
536 
537 	stripe_offset = offset & (d->stripe_size - 1);
538 
539 	while (nr_sectors) {
540 		int s = min_t(unsigned int, abs(nr_sectors),
541 			      d->stripe_size - stripe_offset);
542 
543 		if (nr_sectors < 0)
544 			s = -s;
545 
546 		if (stripe >= d->nr_stripes)
547 			return;
548 
549 		sectors_dirty = atomic_add_return(s,
550 					d->stripe_sectors_dirty + stripe);
551 		if (sectors_dirty == d->stripe_size)
552 			set_bit(stripe, d->full_dirty_stripes);
553 		else
554 			clear_bit(stripe, d->full_dirty_stripes);
555 
556 		nr_sectors -= s;
557 		stripe_offset = 0;
558 		stripe++;
559 	}
560 }
561 
562 static bool dirty_pred(struct keybuf *buf, struct bkey *k)
563 {
564 	struct cached_dev *dc = container_of(buf,
565 					     struct cached_dev,
566 					     writeback_keys);
567 
568 	BUG_ON(KEY_INODE(k) != dc->disk.id);
569 
570 	return KEY_DIRTY(k);
571 }
572 
573 static void refill_full_stripes(struct cached_dev *dc)
574 {
575 	struct keybuf *buf = &dc->writeback_keys;
576 	unsigned int start_stripe, next_stripe;
577 	int stripe;
578 	bool wrapped = false;
579 
580 	stripe = offset_to_stripe(&dc->disk, KEY_OFFSET(&buf->last_scanned));
581 	if (stripe < 0)
582 		stripe = 0;
583 
584 	start_stripe = stripe;
585 
586 	while (1) {
587 		stripe = find_next_bit(dc->disk.full_dirty_stripes,
588 				       dc->disk.nr_stripes, stripe);
589 
590 		if (stripe == dc->disk.nr_stripes)
591 			goto next;
592 
593 		next_stripe = find_next_zero_bit(dc->disk.full_dirty_stripes,
594 						 dc->disk.nr_stripes, stripe);
595 
596 		buf->last_scanned = KEY(dc->disk.id,
597 					stripe * dc->disk.stripe_size, 0);
598 
599 		bch_refill_keybuf(dc->disk.c, buf,
600 				  &KEY(dc->disk.id,
601 				       next_stripe * dc->disk.stripe_size, 0),
602 				  dirty_pred);
603 
604 		if (array_freelist_empty(&buf->freelist))
605 			return;
606 
607 		stripe = next_stripe;
608 next:
609 		if (wrapped && stripe > start_stripe)
610 			return;
611 
612 		if (stripe == dc->disk.nr_stripes) {
613 			stripe = 0;
614 			wrapped = true;
615 		}
616 	}
617 }
618 
619 /*
620  * Returns true if we scanned the entire disk
621  */
622 static bool refill_dirty(struct cached_dev *dc)
623 {
624 	struct keybuf *buf = &dc->writeback_keys;
625 	struct bkey start = KEY(dc->disk.id, 0, 0);
626 	struct bkey end = KEY(dc->disk.id, MAX_KEY_OFFSET, 0);
627 	struct bkey start_pos;
628 
629 	/*
630 	 * make sure keybuf pos is inside the range for this disk - at bringup
631 	 * we might not be attached yet so this disk's inode nr isn't
632 	 * initialized then
633 	 */
634 	if (bkey_cmp(&buf->last_scanned, &start) < 0 ||
635 	    bkey_cmp(&buf->last_scanned, &end) > 0)
636 		buf->last_scanned = start;
637 
638 	if (dc->partial_stripes_expensive) {
639 		refill_full_stripes(dc);
640 		if (array_freelist_empty(&buf->freelist))
641 			return false;
642 	}
643 
644 	start_pos = buf->last_scanned;
645 	bch_refill_keybuf(dc->disk.c, buf, &end, dirty_pred);
646 
647 	if (bkey_cmp(&buf->last_scanned, &end) < 0)
648 		return false;
649 
650 	/*
651 	 * If we get to the end start scanning again from the beginning, and
652 	 * only scan up to where we initially started scanning from:
653 	 */
654 	buf->last_scanned = start;
655 	bch_refill_keybuf(dc->disk.c, buf, &start_pos, dirty_pred);
656 
657 	return bkey_cmp(&buf->last_scanned, &start_pos) >= 0;
658 }
659 
660 static int bch_writeback_thread(void *arg)
661 {
662 	struct cached_dev *dc = arg;
663 	struct cache_set *c = dc->disk.c;
664 	bool searched_full_index;
665 
666 	bch_ratelimit_reset(&dc->writeback_rate);
667 
668 	while (!kthread_should_stop() &&
669 	       !test_bit(CACHE_SET_IO_DISABLE, &c->flags)) {
670 		down_write(&dc->writeback_lock);
671 		set_current_state(TASK_INTERRUPTIBLE);
672 		/*
673 		 * If the bache device is detaching, skip here and continue
674 		 * to perform writeback. Otherwise, if no dirty data on cache,
675 		 * or there is dirty data on cache but writeback is disabled,
676 		 * the writeback thread should sleep here and wait for others
677 		 * to wake up it.
678 		 */
679 		if (!test_bit(BCACHE_DEV_DETACHING, &dc->disk.flags) &&
680 		    (!atomic_read(&dc->has_dirty) || !dc->writeback_running)) {
681 			up_write(&dc->writeback_lock);
682 
683 			if (kthread_should_stop() ||
684 			    test_bit(CACHE_SET_IO_DISABLE, &c->flags)) {
685 				set_current_state(TASK_RUNNING);
686 				break;
687 			}
688 
689 			schedule();
690 			continue;
691 		}
692 		set_current_state(TASK_RUNNING);
693 
694 		searched_full_index = refill_dirty(dc);
695 
696 		if (searched_full_index &&
697 		    RB_EMPTY_ROOT(&dc->writeback_keys.keys)) {
698 			atomic_set(&dc->has_dirty, 0);
699 			SET_BDEV_STATE(&dc->sb, BDEV_STATE_CLEAN);
700 			bch_write_bdev_super(dc, NULL);
701 			/*
702 			 * If bcache device is detaching via sysfs interface,
703 			 * writeback thread should stop after there is no dirty
704 			 * data on cache. BCACHE_DEV_DETACHING flag is set in
705 			 * bch_cached_dev_detach().
706 			 */
707 			if (test_bit(BCACHE_DEV_DETACHING, &dc->disk.flags)) {
708 				up_write(&dc->writeback_lock);
709 				break;
710 			}
711 
712 			/*
713 			 * When dirty data rate is high (e.g. 50%+), there might
714 			 * be heavy buckets fragmentation after writeback
715 			 * finished, which hurts following write performance.
716 			 * If users really care about write performance they
717 			 * may set BCH_ENABLE_AUTO_GC via sysfs, then when
718 			 * BCH_DO_AUTO_GC is set, garbage collection thread
719 			 * will be wake up here. After moving gc, the shrunk
720 			 * btree and discarded free buckets SSD space may be
721 			 * helpful for following write requests.
722 			 */
723 			if (c->gc_after_writeback ==
724 			    (BCH_ENABLE_AUTO_GC|BCH_DO_AUTO_GC)) {
725 				c->gc_after_writeback &= ~BCH_DO_AUTO_GC;
726 				force_wake_up_gc(c);
727 			}
728 		}
729 
730 		up_write(&dc->writeback_lock);
731 
732 		read_dirty(dc);
733 
734 		if (searched_full_index) {
735 			unsigned int delay = dc->writeback_delay * HZ;
736 
737 			while (delay &&
738 			       !kthread_should_stop() &&
739 			       !test_bit(CACHE_SET_IO_DISABLE, &c->flags) &&
740 			       !test_bit(BCACHE_DEV_DETACHING, &dc->disk.flags))
741 				delay = schedule_timeout_interruptible(delay);
742 
743 			bch_ratelimit_reset(&dc->writeback_rate);
744 		}
745 	}
746 
747 	if (dc->writeback_write_wq) {
748 		flush_workqueue(dc->writeback_write_wq);
749 		destroy_workqueue(dc->writeback_write_wq);
750 	}
751 	cached_dev_put(dc);
752 	wait_for_kthread_stop();
753 
754 	return 0;
755 }
756 
757 /* Init */
758 #define INIT_KEYS_EACH_TIME	500000
759 #define INIT_KEYS_SLEEP_MS	100
760 
761 struct sectors_dirty_init {
762 	struct btree_op	op;
763 	unsigned int	inode;
764 	size_t		count;
765 	struct bkey	start;
766 };
767 
768 static int sectors_dirty_init_fn(struct btree_op *_op, struct btree *b,
769 				 struct bkey *k)
770 {
771 	struct sectors_dirty_init *op = container_of(_op,
772 						struct sectors_dirty_init, op);
773 	if (KEY_INODE(k) > op->inode)
774 		return MAP_DONE;
775 
776 	if (KEY_DIRTY(k))
777 		bcache_dev_sectors_dirty_add(b->c, KEY_INODE(k),
778 					     KEY_START(k), KEY_SIZE(k));
779 
780 	op->count++;
781 	if (atomic_read(&b->c->search_inflight) &&
782 	    !(op->count % INIT_KEYS_EACH_TIME)) {
783 		bkey_copy_key(&op->start, k);
784 		return -EAGAIN;
785 	}
786 
787 	return MAP_CONTINUE;
788 }
789 
790 static int bch_root_node_dirty_init(struct cache_set *c,
791 				     struct bcache_device *d,
792 				     struct bkey *k)
793 {
794 	struct sectors_dirty_init op;
795 	int ret;
796 
797 	bch_btree_op_init(&op.op, -1);
798 	op.inode = d->id;
799 	op.count = 0;
800 	op.start = KEY(op.inode, 0, 0);
801 
802 	do {
803 		ret = bcache_btree(map_keys_recurse,
804 				   k,
805 				   c->root,
806 				   &op.op,
807 				   &op.start,
808 				   sectors_dirty_init_fn,
809 				   0);
810 		if (ret == -EAGAIN)
811 			schedule_timeout_interruptible(
812 				msecs_to_jiffies(INIT_KEYS_SLEEP_MS));
813 		else if (ret < 0) {
814 			pr_warn("sectors dirty init failed, ret=%d!\n", ret);
815 			break;
816 		}
817 	} while (ret == -EAGAIN);
818 
819 	return ret;
820 }
821 
822 static int bch_dirty_init_thread(void *arg)
823 {
824 	struct dirty_init_thrd_info *info = arg;
825 	struct bch_dirty_init_state *state = info->state;
826 	struct cache_set *c = state->c;
827 	struct btree_iter iter;
828 	struct bkey *k, *p;
829 	int cur_idx, prev_idx, skip_nr;
830 
831 	k = p = NULL;
832 	cur_idx = prev_idx = 0;
833 
834 	bch_btree_iter_init(&c->root->keys, &iter, NULL);
835 	k = bch_btree_iter_next_filter(&iter, &c->root->keys, bch_ptr_bad);
836 	BUG_ON(!k);
837 
838 	p = k;
839 
840 	while (k) {
841 		spin_lock(&state->idx_lock);
842 		cur_idx = state->key_idx;
843 		state->key_idx++;
844 		spin_unlock(&state->idx_lock);
845 
846 		skip_nr = cur_idx - prev_idx;
847 
848 		while (skip_nr) {
849 			k = bch_btree_iter_next_filter(&iter,
850 						       &c->root->keys,
851 						       bch_ptr_bad);
852 			if (k)
853 				p = k;
854 			else {
855 				atomic_set(&state->enough, 1);
856 				/* Update state->enough earlier */
857 				smp_mb__after_atomic();
858 				goto out;
859 			}
860 			skip_nr--;
861 			cond_resched();
862 		}
863 
864 		if (p) {
865 			if (bch_root_node_dirty_init(c, state->d, p) < 0)
866 				goto out;
867 		}
868 
869 		p = NULL;
870 		prev_idx = cur_idx;
871 		cond_resched();
872 	}
873 
874 out:
875 	/* In order to wake up state->wait in time */
876 	smp_mb__before_atomic();
877 	if (atomic_dec_and_test(&state->started))
878 		wake_up(&state->wait);
879 
880 	return 0;
881 }
882 
883 static int bch_btre_dirty_init_thread_nr(void)
884 {
885 	int n = num_online_cpus()/2;
886 
887 	if (n == 0)
888 		n = 1;
889 	else if (n > BCH_DIRTY_INIT_THRD_MAX)
890 		n = BCH_DIRTY_INIT_THRD_MAX;
891 
892 	return n;
893 }
894 
895 void bch_sectors_dirty_init(struct bcache_device *d)
896 {
897 	int i;
898 	struct bkey *k = NULL;
899 	struct btree_iter iter;
900 	struct sectors_dirty_init op;
901 	struct cache_set *c = d->c;
902 	struct bch_dirty_init_state *state;
903 	char name[32];
904 
905 	/* Just count root keys if no leaf node */
906 	if (c->root->level == 0) {
907 		bch_btree_op_init(&op.op, -1);
908 		op.inode = d->id;
909 		op.count = 0;
910 		op.start = KEY(op.inode, 0, 0);
911 
912 		for_each_key_filter(&c->root->keys,
913 				    k, &iter, bch_ptr_invalid)
914 			sectors_dirty_init_fn(&op.op, c->root, k);
915 		return;
916 	}
917 
918 	state = kzalloc(sizeof(struct bch_dirty_init_state), GFP_KERNEL);
919 	if (!state) {
920 		pr_warn("sectors dirty init failed: cannot allocate memory\n");
921 		return;
922 	}
923 
924 	state->c = c;
925 	state->d = d;
926 	state->total_threads = bch_btre_dirty_init_thread_nr();
927 	state->key_idx = 0;
928 	spin_lock_init(&state->idx_lock);
929 	atomic_set(&state->started, 0);
930 	atomic_set(&state->enough, 0);
931 	init_waitqueue_head(&state->wait);
932 
933 	for (i = 0; i < state->total_threads; i++) {
934 		/* Fetch latest state->enough earlier */
935 		smp_mb__before_atomic();
936 		if (atomic_read(&state->enough))
937 			break;
938 
939 		state->infos[i].state = state;
940 		atomic_inc(&state->started);
941 		snprintf(name, sizeof(name), "bch_dirty_init[%d]", i);
942 
943 		state->infos[i].thread =
944 			kthread_run(bch_dirty_init_thread,
945 				    &state->infos[i],
946 				    name);
947 		if (IS_ERR(state->infos[i].thread)) {
948 			pr_err("fails to run thread bch_dirty_init[%d]\n", i);
949 			for (--i; i >= 0; i--)
950 				kthread_stop(state->infos[i].thread);
951 			goto out;
952 		}
953 	}
954 
955 	wait_event_interruptible(state->wait,
956 		 atomic_read(&state->started) == 0 ||
957 		 test_bit(CACHE_SET_IO_DISABLE, &c->flags));
958 
959 out:
960 	kfree(state);
961 }
962 
963 void bch_cached_dev_writeback_init(struct cached_dev *dc)
964 {
965 	sema_init(&dc->in_flight, 64);
966 	init_rwsem(&dc->writeback_lock);
967 	bch_keybuf_init(&dc->writeback_keys);
968 
969 	dc->writeback_metadata		= true;
970 	dc->writeback_running		= false;
971 	dc->writeback_percent		= 10;
972 	dc->writeback_delay		= 30;
973 	atomic_long_set(&dc->writeback_rate.rate, 1024);
974 	dc->writeback_rate_minimum	= 8;
975 
976 	dc->writeback_rate_update_seconds = WRITEBACK_RATE_UPDATE_SECS_DEFAULT;
977 	dc->writeback_rate_p_term_inverse = 40;
978 	dc->writeback_rate_i_term_inverse = 10000;
979 
980 	WARN_ON(test_and_clear_bit(BCACHE_DEV_WB_RUNNING, &dc->disk.flags));
981 	INIT_DELAYED_WORK(&dc->writeback_rate_update, update_writeback_rate);
982 }
983 
984 int bch_cached_dev_writeback_start(struct cached_dev *dc)
985 {
986 	dc->writeback_write_wq = alloc_workqueue("bcache_writeback_wq",
987 						WQ_MEM_RECLAIM, 0);
988 	if (!dc->writeback_write_wq)
989 		return -ENOMEM;
990 
991 	cached_dev_get(dc);
992 	dc->writeback_thread = kthread_create(bch_writeback_thread, dc,
993 					      "bcache_writeback");
994 	if (IS_ERR(dc->writeback_thread)) {
995 		cached_dev_put(dc);
996 		destroy_workqueue(dc->writeback_write_wq);
997 		return PTR_ERR(dc->writeback_thread);
998 	}
999 	dc->writeback_running = true;
1000 
1001 	WARN_ON(test_and_set_bit(BCACHE_DEV_WB_RUNNING, &dc->disk.flags));
1002 	schedule_delayed_work(&dc->writeback_rate_update,
1003 			      dc->writeback_rate_update_seconds * HZ);
1004 
1005 	bch_writeback_queue(dc);
1006 
1007 	return 0;
1008 }
1009