brintos

brintos / linux-shallow public Read only

0
0
Text · 28.4 KiB · c1d28e3 Raw
1104 lines · c
1// SPDX-License-Identifier: GPL-2.02/*3 * background writeback - scan btree for dirty data and write it to the backing4 * device5 *6 * Copyright 2010, 2011 Kent Overstreet <kent.overstreet@gmail.com>7 * Copyright 2012 Google, Inc.8 */9 10#include "bcache.h"11#include "btree.h"12#include "debug.h"13#include "writeback.h"14 15#include <linux/delay.h>16#include <linux/kthread.h>17#include <linux/sched/clock.h>18#include <trace/events/bcache.h>19 20static void update_gc_after_writeback(struct cache_set *c)21{22	if (c->gc_after_writeback != (BCH_ENABLE_AUTO_GC) ||23	    c->gc_stats.in_use < BCH_AUTO_GC_DIRTY_THRESHOLD)24		return;25 26	c->gc_after_writeback |= BCH_DO_AUTO_GC;27}28 29/* Rate limiting */30static uint64_t __calc_target_rate(struct cached_dev *dc)31{32	struct cache_set *c = dc->disk.c;33 34	/*35	 * This is the size of the cache, minus the amount used for36	 * flash-only devices37	 */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 the43	 * 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 each45	 * backing volume uses about 2% of the cache for dirty data.46	 */47	uint32_t bdev_share =48		div64_u64(bdev_nr_sectors(dc->bdev) << WRITEBACK_SHARE_SHIFT,49				c->cached_dev_sectors);50 51	uint64_t cache_dirty_target =52		div_u64(cache_sectors * dc->writeback_percent, 100);53 54	/* Ensure each backing dev gets at least one dirty share */55	if (bdev_share < 1)56		bdev_share = 1;57 58	return (cache_dirty_target * bdev_share) >> WRITEBACK_SHARE_SHIFT;59}60 61static void __update_writeback_rate(struct cached_dev *dc)62{63	/*64	 * PI controller:65	 * Figures out the amount that should be written per second.66	 *67	 * First, the error (number of sectors that are dirty beyond our68	 * target) is calculated.  The error is accumulated (numerically69	 * integrated).70	 *71	 * Then, the proportional value and integral value are scaled72	 * based on configured values.  These are stored as inverses to73	 * avoid fixed point math and to make configuration easy-- e.g.74	 * the default value of 40 for writeback_rate_p_term_inverse75	 * attempts to write at a rate that would retire all the dirty76	 * blocks in 40 seconds.77	 *78	 * The writeback_rate_i_inverse value of 10000 means that 1/10000th79	 * of the error is accumulated in the integral term per second.80	 * This acts as a slow, long-term average that is not subject to81	 * 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 well93	 * when calculating the proportional_scaled, Otherwise we might94	 * have an unreasonable small writeback rate at a highly fragmented situation95	 * when very few dirty sectors consumed a lot dirty buckets, the96	 * worst case is when dirty buckets reached cutoff_writeback_sync and97	 * dirty data is still not even reached to writeback percent, so the rate98	 * still will be at the minimum value, which will cause the write99	 * 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 than134		 * zero.  Only increase the integral term if the device135		 * is keeping up.  (Don't wind up the integral136		 * ineffectively in either case).137		 *138		 * It's necessary to scale this by139		 * writeback_rate_update_seconds to keep the integral140		 * term dimensioned properly.141		 */142		dc->writeback_rate_integral += error *143			dc->writeback_rate_update_seconds;144	}145 146	integral_scaled = div_s64(dc->writeback_rate_integral,147			dc->writeback_rate_i_term_inverse);148 149	new_rate = clamp_t(int32_t, (proportional_scaled + integral_scaled),150			dc->writeback_rate_minimum, NSEC_PER_SEC);151 152	dc->writeback_rate_proportional = proportional_scaled;153	dc->writeback_rate_integral_scaled = integral_scaled;154	dc->writeback_rate_change = new_rate -155			atomic_long_read(&dc->writeback_rate.rate);156	atomic_long_set(&dc->writeback_rate.rate, new_rate);157	dc->writeback_rate_target = target;158}159 160static bool idle_counter_exceeded(struct cache_set *c)161{162	int counter, dev_nr;163 164	/*165	 * If c->idle_counter is overflow (idel for really long time),166	 * reset as 0 and not set maximum rate this time for code167	 * 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 all181	 * attached backing devices, in order to represent a rough182	 * time period, counter should be divided by dev_nr.183	 * Otherwise the idle time cannot be larger with more backing184	 * device attached.185	 * The following calculation equals to checking186	 *	(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() is196 * called. If all backing devices attached to the same cache set have197 * identical dc->writeback_rate_update_seconds values, it is about 6198 * rounds of update_writeback_rate() on each backing device before199 * c->at_max_writeback_rate is set to 1, and then max wrteback rate set200 * to each dc->writeback_rate.rate.201 * In order to avoid extra locking cost for counting exact dirty cached202 * devices number, c->attached_dev_nr is used to calculate the idle203 * 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 of205 * update_writeback_rate().206 */207static bool set_at_max_writeback_rate(struct cache_set *c,208				       struct cached_dev *dc)209{210	/* Don't sst max writeback rate if it is disabled */211	if (!c->idle_max_writeback_rate_enabled)212		return false;213 214	/* Don't set max writeback rate if gc is running */215	if (!c->gc_mark_valid)216		return false;217 218	if (!idle_counter_exceeded(c))219		return false;220 221	if (atomic_read(&c->at_max_writeback_rate) != 1)222		atomic_set(&c->at_max_writeback_rate, 1);223 224	atomic_long_set(&dc->writeback_rate.rate, INT_MAX);225 226	/* keep writeback_rate_target as existing value */227	dc->writeback_rate_proportional = 0;228	dc->writeback_rate_integral_scaled = 0;229	dc->writeback_rate_change = 0;230 231	/*232	 * In case new I/O arrives during before233	 * set_at_max_writeback_rate() returns.234	 */235	if (!idle_counter_exceeded(c) ||236	    !atomic_read(&c->at_max_writeback_rate))237		return false;238 239	return true;240}241 242static void update_writeback_rate(struct work_struct *work)243{244	struct cached_dev *dc = container_of(to_delayed_work(work),245					     struct cached_dev,246					     writeback_rate_update);247	struct cache_set *c = dc->disk.c;248 249	/*250	 * should check BCACHE_DEV_RATE_DW_RUNNING before calling251	 * 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 is272	 * unncessary to update writeback rate for an idle cache set273	 * 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 calling305	 * cancel_delayed_work_sync().306	 */307	clear_bit(BCACHE_DEV_RATE_DW_RUNNING, &dc->disk.flags);308	/* paired with where BCACHE_DEV_RATE_DW_RUNNING is tested */309	smp_mb__after_atomic();310}311 312static unsigned int writeback_delay(struct cached_dev *dc,313				    unsigned int sectors)314{315	if (test_bit(BCACHE_DEV_DETACHING, &dc->disk.flags) ||316	    !dc->writeback_percent)317		return 0;318 319	return bch_next_delay(&dc->writeback_rate, sectors);320}321 322struct dirty_io {323	struct closure		cl;324	struct cached_dev	*dc;325	uint16_t		sequence;326	struct bio		bio;327};328 329static void dirty_init(struct keybuf_key *w)330{331	struct dirty_io *io = w->private;332	struct bio *bio = &io->bio;333 334	bio_init(bio, NULL, bio->bi_inline_vecs,335		 DIV_ROUND_UP(KEY_SIZE(&w->key), PAGE_SECTORS), 0);336	if (!io->dc->writeback_percent)337		bio_set_prio(bio, IOPRIO_PRIO_VALUE(IOPRIO_CLASS_IDLE, 0));338 339	bio->bi_iter.bi_size	= KEY_SIZE(&w->key) << 9;340	bio->bi_private		= w;341	bch_bio_map(bio, NULL);342}343 344static CLOSURE_CALLBACK(dirty_io_destructor)345{346	closure_type(io, struct dirty_io, cl);347 348	kfree(io);349}350 351static CLOSURE_CALLBACK(write_dirty_finish)352{353	closure_type(io, struct dirty_io, cl);354	struct keybuf_key *w = io->bio.bi_private;355	struct cached_dev *dc = io->dc;356 357	bio_free_pages(&io->bio);358 359	/* This is kind of a dumb way of signalling errors. */360	if (KEY_DIRTY(&w->key)) {361		int ret;362		unsigned int i;363		struct keylist keys;364 365		bch_keylist_init(&keys);366 367		bkey_copy(keys.top, &w->key);368		SET_KEY_DIRTY(keys.top, false);369		bch_keylist_push(&keys);370 371		for (i = 0; i < KEY_PTRS(&w->key); i++)372			atomic_inc(&PTR_BUCKET(dc->disk.c, &w->key, i)->pin);373 374		ret = bch_btree_insert(dc->disk.c, &keys, NULL, &w->key);375 376		if (ret)377			trace_bcache_writeback_collision(&w->key);378 379		atomic_long_inc(ret380				? &dc->disk.c->writeback_keys_failed381				: &dc->disk.c->writeback_keys_done);382	}383 384	bch_keybuf_del(&dc->writeback_keys, w);385	up(&dc->in_flight);386 387	closure_return_with_destructor(cl, dirty_io_destructor);388}389 390static void dirty_endio(struct bio *bio)391{392	struct keybuf_key *w = bio->bi_private;393	struct dirty_io *io = w->private;394 395	if (bio->bi_status) {396		SET_KEY_DIRTY(&w->key, false);397		bch_count_backing_io_errors(io->dc, bio);398	}399 400	closure_put(&io->cl);401}402 403static CLOSURE_CALLBACK(write_dirty)404{405	closure_type(io, struct dirty_io, cl);406	struct keybuf_key *w = io->bio.bi_private;407	struct cached_dev *dc = io->dc;408 409	uint16_t next_sequence;410 411	if (atomic_read(&dc->writeback_sequence_next) != io->sequence) {412		/* Not our turn to write; wait for a write to complete */413		closure_wait(&dc->writeback_ordering_wait, cl);414 415		if (atomic_read(&dc->writeback_sequence_next) == io->sequence) {416			/*417			 * Edge case-- it happened in indeterminate order418			 * 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 the432	 * backing device.  Instead, immediately go to write_dirty_finish433	 * to clean up.434	 */435	if (KEY_DIRTY(&w->key)) {436		dirty_init(w);437		io->bio.bi_opf = REQ_OP_WRITE;438		io->bio.bi_iter.bi_sector = KEY_START(&w->key);439		bio_set_dev(&io->bio, io->dc->bdev);440		io->bio.bi_end_io	= dirty_endio;441 442		/* I/O request sent to backing device */443		closure_bio_submit(io->dc->disk.c, &io->bio, cl);444	}445 446	atomic_set(&dc->writeback_sequence_next, next_sequence);447	closure_wake_up(&dc->writeback_ordering_wait);448 449	continue_at(cl, write_dirty_finish, io->dc->writeback_write_wq);450}451 452static void read_dirty_endio(struct bio *bio)453{454	struct keybuf_key *w = bio->bi_private;455	struct dirty_io *io = w->private;456 457	/* is_read = 1 */458	bch_count_io_errors(io->dc->disk.c->cache,459			    bio->bi_status, 1,460			    "reading dirty data from cache");461 462	dirty_endio(bio);463}464 465static CLOSURE_CALLBACK(read_dirty_submit)466{467	closure_type(io, struct dirty_io, cl);468 469	closure_bio_submit(io->dc->disk.c, &io->bio, cl);470 471	continue_at(cl, write_dirty, io->dc->writeback_write_wq);472}473 474static void read_dirty(struct cached_dev *dc)475{476	unsigned int delay = 0;477	struct keybuf_key *next, *keys[MAX_WRITEBACKS_IN_PASS], *w;478	size_t size;479	int nk, i;480	struct dirty_io *io;481	struct closure cl;482	uint16_t sequence = 0;483 484	BUG_ON(!llist_empty(&dc->writeback_ordering_wait.list));485	atomic_set(&dc->writeback_sequence_next, sequence);486	closure_init_stack(&cl);487 488	/*489	 * XXX: if we error, background writeback just spins. Should use some490	 * 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 they506			 * are all small.507			 */508			if (nk >= MAX_WRITEBACKS_IN_PASS)509				break;510 511			/*512			 * If the current operation is very large, don't513			 * further combine operations.514			 */515			if (size >= MAX_WRITESIZE_IN_PASS)516				break;517 518			/*519			 * Operations are only eligible to be combined520			 * if they are contiguous.521			 *522			 * TODO: add a heuristic willing to fire a523			 * certain amount of non-contiguous IO per pass,524			 * so that we can benefit from backing device525			 * 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 maximum564			 * simultaneous number of writebacks; from here565			 * everything happens asynchronously.566			 */567			closure_call(&io->cl, read_dirty_submit, NULL, &cl);568		}569 570		delay = writeback_delay(dc, size);571 572		while (!kthread_should_stop() &&573		       !test_bit(CACHE_SET_IO_DISABLE, &dc->disk.c->flags) &&574		       delay) {575			schedule_timeout_interruptible(delay);576			delay = writeback_delay(dc, 0);577		}578	}579 580	if (0) {581err_free:582		kfree(w->private);583err:584		bch_keybuf_del(&dc->writeback_keys, w);585	}586 587	/*588	 * Wait for outstanding writeback IOs to finish (and keybuf slots to be589	 * freed) before refilling again590	 */591	closure_sync(&cl);592}593 594/* Scan for dirty data */595 596void bcache_dev_sectors_dirty_add(struct cache_set *c, unsigned int inode,597				  uint64_t offset, int nr_sectors)598{599	struct bcache_device *d = c->devices[inode];600	unsigned int stripe_offset, sectors_dirty;601	int stripe;602 603	if (!d)604		return;605 606	stripe = offset_to_stripe(d, offset);607	if (stripe < 0)608		return;609 610	if (UUID_FLASH_ONLY(&c->uuids[inode]))611		atomic_long_add(nr_sectors, &c->flash_dev_dirty_sectors);612 613	stripe_offset = offset & (d->stripe_size - 1);614 615	while (nr_sectors) {616		int s = min_t(unsigned int, abs(nr_sectors),617			      d->stripe_size - stripe_offset);618 619		if (nr_sectors < 0)620			s = -s;621 622		if (stripe >= d->nr_stripes)623			return;624 625		sectors_dirty = atomic_add_return(s,626					d->stripe_sectors_dirty + stripe);627		if (sectors_dirty == d->stripe_size) {628			if (!test_bit(stripe, d->full_dirty_stripes))629				set_bit(stripe, d->full_dirty_stripes);630		} else {631			if (test_bit(stripe, d->full_dirty_stripes))632				clear_bit(stripe, d->full_dirty_stripes);633		}634 635		nr_sectors -= s;636		stripe_offset = 0;637		stripe++;638	}639}640 641static bool dirty_pred(struct keybuf *buf, struct bkey *k)642{643	struct cached_dev *dc = container_of(buf,644					     struct cached_dev,645					     writeback_keys);646 647	BUG_ON(KEY_INODE(k) != dc->disk.id);648 649	return KEY_DIRTY(k);650}651 652static void refill_full_stripes(struct cached_dev *dc)653{654	struct keybuf *buf = &dc->writeback_keys;655	unsigned int start_stripe, next_stripe;656	int stripe;657	bool wrapped = false;658 659	stripe = offset_to_stripe(&dc->disk, KEY_OFFSET(&buf->last_scanned));660	if (stripe < 0)661		stripe = 0;662 663	start_stripe = stripe;664 665	while (1) {666		stripe = find_next_bit(dc->disk.full_dirty_stripes,667				       dc->disk.nr_stripes, stripe);668 669		if (stripe == dc->disk.nr_stripes)670			goto next;671 672		next_stripe = find_next_zero_bit(dc->disk.full_dirty_stripes,673						 dc->disk.nr_stripes, stripe);674 675		buf->last_scanned = KEY(dc->disk.id,676					stripe * dc->disk.stripe_size, 0);677 678		bch_refill_keybuf(dc->disk.c, buf,679				  &KEY(dc->disk.id,680				       next_stripe * dc->disk.stripe_size, 0),681				  dirty_pred);682 683		if (array_freelist_empty(&buf->freelist))684			return;685 686		stripe = next_stripe;687next:688		if (wrapped && stripe > start_stripe)689			return;690 691		if (stripe == dc->disk.nr_stripes) {692			stripe = 0;693			wrapped = true;694		}695	}696}697 698/*699 * Returns true if we scanned the entire disk700 */701static bool refill_dirty(struct cached_dev *dc)702{703	struct keybuf *buf = &dc->writeback_keys;704	struct bkey start = KEY(dc->disk.id, 0, 0);705	struct bkey end = KEY(dc->disk.id, MAX_KEY_OFFSET, 0);706	struct bkey start_pos;707 708	/*709	 * make sure keybuf pos is inside the range for this disk - at bringup710	 * we might not be attached yet so this disk's inode nr isn't711	 * initialized then712	 */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, and731	 * only scan up to where we initially started scanning from:732	 */733	buf->last_scanned = start;734	bch_refill_keybuf(dc->disk.c, buf, &start_pos, dirty_pred);735 736	return bkey_cmp(&buf->last_scanned, &start_pos) >= 0;737}738 739static int bch_writeback_thread(void *arg)740{741	struct cached_dev *dc = arg;742	struct cache_set *c = dc->disk.c;743	bool searched_full_index;744 745	bch_ratelimit_reset(&dc->writeback_rate);746 747	while (!kthread_should_stop() &&748	       !test_bit(CACHE_SET_IO_DISABLE, &c->flags)) {749		down_write(&dc->writeback_lock);750		set_current_state(TASK_INTERRUPTIBLE);751		/*752		 * If the bache device is detaching, skip here and continue753		 * 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 others756		 * 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 dirty783			 * data on cache. BCACHE_DEV_DETACHING flag is set in784			 * 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 might802			 * be heavy buckets fragmentation after writeback803			 * finished, which hurts following write performance.804			 * If users really care about write performance they805			 * may set BCH_ENABLE_AUTO_GC via sysfs, then when806			 * BCH_DO_AUTO_GC is set, garbage collection thread807			 * will be wake up here. After moving gc, the shrunk808			 * btree and discarded free buckets SSD space may be809			 * 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	500000846 847struct sectors_dirty_init {848	struct btree_op	op;849	unsigned int	inode;850	size_t		count;851};852 853static int sectors_dirty_init_fn(struct btree_op *_op, struct btree *b,854				 struct bkey *k)855{856	struct sectors_dirty_init *op = container_of(_op,857						struct sectors_dirty_init, op);858	if (KEY_INODE(k) > op->inode)859		return MAP_DONE;860 861	if (KEY_DIRTY(k))862		bcache_dev_sectors_dirty_add(b->c, KEY_INODE(k),863					     KEY_START(k), KEY_SIZE(k));864 865	op->count++;866	if (!(op->count % INIT_KEYS_EACH_TIME))867		cond_resched();868 869	return MAP_CONTINUE;870}871 872static int bch_root_node_dirty_init(struct cache_set *c,873				     struct bcache_device *d,874				     struct bkey *k)875{876	struct sectors_dirty_init op;877	int ret;878 879	bch_btree_op_init(&op.op, -1);880	op.inode = d->id;881	op.count = 0;882 883	ret = bcache_btree(map_keys_recurse,884			   k,885			   c->root,886			   &op.op,887			   &KEY(op.inode, 0, 0),888			   sectors_dirty_init_fn,889			   0);890	if (ret < 0)891		pr_warn("sectors dirty init failed, ret=%d!\n", ret);892 893	/*894	 * The op may be added to cache_set's btree_cache_wait895	 * in mca_cannibalize(), must ensure it is removed from896	 * the list and release btree_cache_alloc_lock before897	 * free op memory.898	 * Otherwise, the btree_cache_wait will be damaged.899	 */900	bch_cannibalize_unlock(c);901	finish_wait(&c->btree_cache_wait, &(&op.op)->wait);902 903	return ret;904}905 906static int bch_dirty_init_thread(void *arg)907{908	struct dirty_init_thrd_info *info = arg;909	struct bch_dirty_init_state *state = info->state;910	struct cache_set *c = state->c;911	struct btree_iter iter;912	struct bkey *k, *p;913	int cur_idx, prev_idx, skip_nr;914 915	k = p = NULL;916	prev_idx = 0;917 918	min_heap_init(&iter.heap, NULL, MAX_BSETS);919	bch_btree_iter_init(&c->root->keys, &iter, NULL);920	k = bch_btree_iter_next_filter(&iter, &c->root->keys, bch_ptr_bad);921	BUG_ON(!k);922 923	p = k;924 925	while (k) {926		spin_lock(&state->idx_lock);927		cur_idx = state->key_idx;928		state->key_idx++;929		spin_unlock(&state->idx_lock);930 931		skip_nr = cur_idx - prev_idx;932 933		while (skip_nr) {934			k = bch_btree_iter_next_filter(&iter,935						       &c->root->keys,936						       bch_ptr_bad);937			if (k)938				p = k;939			else {940				atomic_set(&state->enough, 1);941				/* Update state->enough earlier */942				smp_mb__after_atomic();943				goto out;944			}945			skip_nr--;946		}947 948		if (p) {949			if (bch_root_node_dirty_init(c, state->d, p) < 0)950				goto out;951		}952 953		p = NULL;954		prev_idx = cur_idx;955	}956 957out:958	/* In order to wake up state->wait in time */959	smp_mb__before_atomic();960	if (atomic_dec_and_test(&state->started))961		wake_up(&state->wait);962 963	return 0;964}965 966static int bch_btre_dirty_init_thread_nr(void)967{968	int n = num_online_cpus()/2;969 970	if (n == 0)971		n = 1;972	else if (n > BCH_DIRTY_INIT_THRD_MAX)973		n = BCH_DIRTY_INIT_THRD_MAX;974 975	return n;976}977 978void bch_sectors_dirty_init(struct bcache_device *d)979{980	int i;981	struct btree *b = NULL;982	struct bkey *k = NULL;983	struct btree_iter iter;984	struct sectors_dirty_init op;985	struct cache_set *c = d->c;986	struct bch_dirty_init_state state;987 988	min_heap_init(&iter.heap, NULL, MAX_BSETS);989 990retry_lock:991	b = c->root;992	rw_lock(0, b, b->level);993	if (b != c->root) {994		rw_unlock(0, b);995		goto retry_lock;996	}997 998	/* Just count root keys if no leaf node */999	if (c->root->level == 0) {1000		bch_btree_op_init(&op.op, -1);1001		op.inode = d->id;1002		op.count = 0;1003 1004		for_each_key_filter(&c->root->keys,1005				    k, &iter, bch_ptr_invalid) {1006			if (KEY_INODE(k) != op.inode)1007				continue;1008			sectors_dirty_init_fn(&op.op, c->root, k);1009		}1010 1011		rw_unlock(0, b);1012		return;1013	}1014 1015	memset(&state, 0, sizeof(struct bch_dirty_init_state));1016	state.c = c;1017	state.d = d;1018	state.total_threads = bch_btre_dirty_init_thread_nr();1019	state.key_idx = 0;1020	spin_lock_init(&state.idx_lock);1021	atomic_set(&state.started, 0);1022	atomic_set(&state.enough, 0);1023	init_waitqueue_head(&state.wait);1024 1025	for (i = 0; i < state.total_threads; i++) {1026		/* Fetch latest state.enough earlier */1027		smp_mb__before_atomic();1028		if (atomic_read(&state.enough))1029			break;1030 1031		atomic_inc(&state.started);1032		state.infos[i].state = &state;1033		state.infos[i].thread =1034			kthread_run(bch_dirty_init_thread, &state.infos[i],1035				    "bch_dirtcnt[%d]", i);1036		if (IS_ERR(state.infos[i].thread)) {1037			pr_err("fails to run thread bch_dirty_init[%d]\n", i);1038			atomic_dec(&state.started);1039			for (--i; i >= 0; i--)1040				kthread_stop(state.infos[i].thread);1041			goto out;1042		}1043	}1044 1045out:1046	/* Must wait for all threads to stop. */1047	wait_event(state.wait, atomic_read(&state.started) == 0);1048	rw_unlock(0, b);1049}1050 1051void bch_cached_dev_writeback_init(struct cached_dev *dc)1052{1053	sema_init(&dc->in_flight, 64);1054	init_rwsem(&dc->writeback_lock);1055	bch_keybuf_init(&dc->writeback_keys);1056 1057	dc->writeback_metadata		= true;1058	dc->writeback_running		= false;1059	dc->writeback_consider_fragment = true;1060	dc->writeback_percent		= 10;1061	dc->writeback_delay		= 30;1062	atomic_long_set(&dc->writeback_rate.rate, 1024);1063	dc->writeback_rate_minimum	= 8;1064 1065	dc->writeback_rate_update_seconds = WRITEBACK_RATE_UPDATE_SECS_DEFAULT;1066	dc->writeback_rate_p_term_inverse = 40;1067	dc->writeback_rate_fp_term_low = 1;1068	dc->writeback_rate_fp_term_mid = 10;1069	dc->writeback_rate_fp_term_high = 1000;1070	dc->writeback_rate_i_term_inverse = 10000;1071 1072	/* For dc->writeback_lock contention in update_writeback_rate() */1073	dc->rate_update_retry = 0;1074 1075	WARN_ON(test_and_clear_bit(BCACHE_DEV_WB_RUNNING, &dc->disk.flags));1076	INIT_DELAYED_WORK(&dc->writeback_rate_update, update_writeback_rate);1077}1078 1079int bch_cached_dev_writeback_start(struct cached_dev *dc)1080{1081	dc->writeback_write_wq = alloc_workqueue("bcache_writeback_wq",1082						WQ_MEM_RECLAIM, 0);1083	if (!dc->writeback_write_wq)1084		return -ENOMEM;1085 1086	cached_dev_get(dc);1087	dc->writeback_thread = kthread_create(bch_writeback_thread, dc,1088					      "bcache_writeback");1089	if (IS_ERR(dc->writeback_thread)) {1090		cached_dev_put(dc);1091		destroy_workqueue(dc->writeback_write_wq);1092		return PTR_ERR(dc->writeback_thread);1093	}1094	dc->writeback_running = true;1095 1096	WARN_ON(test_and_set_bit(BCACHE_DEV_WB_RUNNING, &dc->disk.flags));1097	schedule_delayed_work(&dc->writeback_rate_update,1098			      dc->writeback_rate_update_seconds * HZ);1099 1100	bch_writeback_queue(dc);1101 1102	return 0;1103}1104