Linux Audio

Check our new training course

Loading...
v6.2
   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
  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 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
  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 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
 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 code
 167	 * simplicity.
 168	 */
 169	counter = atomic_inc_return(&c->idle_counter);
 170	if (counter <= 0) {
 171		atomic_set(&c->idle_counter, 0);
 172		return false;
 173	}
 174
 175	dev_nr = atomic_read(&c->attached_dev_nr);
 176	if (dev_nr == 0)
 177		return false;
 178
 179	/*
 180	 * c->idle_counter is increased by writeback thread of all
 181	 * attached backing devices, in order to represent a rough
 182	 * time period, counter should be divided by dev_nr.
 183	 * Otherwise the idle time cannot be larger with more backing
 184	 * device attached.
 185	 * The following calculation equals to checking
 186	 *	(counter / dev_nr) < (dev_nr * 6)
 187	 */
 188	if (counter < (dev_nr * dev_nr * 6))
 189		return false;
 190
 191	return true;
 192}
 193
 194/*
 195 * Idle_counter is increased every time when update_writeback_rate() is
 196 * called. If all backing devices attached to the same cache set have
 197 * identical dc->writeback_rate_update_seconds values, it is about 6
 198 * rounds of update_writeback_rate() on each backing device before
 199 * c->at_max_writeback_rate is set to 1, and then max wrteback rate set
 200 * to each dc->writeback_rate.rate.
 201 * In order to avoid extra locking cost for counting exact dirty cached
 202 * devices number, c->attached_dev_nr is used to calculate the idle
 203 * throushold. It might be bigger if not all cached device are in write-
 204 * back mode, but it still works well with limited extra rounds of
 205 * update_writeback_rate().
 206 */
 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 before
 233	 * set_at_max_writeback_rate() returns.
 
 
 234	 */
 235	if (!idle_counter_exceeded(c) ||
 
 236	    !atomic_read(&c->at_max_writeback_rate))
 237		return false;
 238
 239	return true;
 240}
 241
 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 calling
 251	 * cancel_delayed_work_sync().
 252	 */
 253	set_bit(BCACHE_DEV_RATE_DW_RUNNING, &dc->disk.flags);
 254	/* paired with where BCACHE_DEV_RATE_DW_RUNNING is tested */
 255	smp_mb__after_atomic();
 256
 257	/*
 258	 * CACHE_SET_IO_DISABLE might be set via sysfs interface,
 259	 * check it here too.
 260	 */
 261	if (!test_bit(BCACHE_DEV_WB_RUNNING, &dc->disk.flags) ||
 262	    test_bit(CACHE_SET_IO_DISABLE, &c->flags)) {
 263		clear_bit(BCACHE_DEV_RATE_DW_RUNNING, &dc->disk.flags);
 264		/* paired with where BCACHE_DEV_RATE_DW_RUNNING is tested */
 265		smp_mb__after_atomic();
 266		return;
 267	}
 268
 269	/*
 270	 * If the whole cache set is idle, set_at_max_writeback_rate()
 271	 * will set writeback rate to a max number. Then it is
 272	 * unncessary to update writeback rate for an idle cache set
 273	 * in maximum writeback rate number(s).
 274	 */
 275	if (atomic_read(&dc->has_dirty) && dc->writeback_percent &&
 276	    !set_at_max_writeback_rate(c, dc)) {
 277		do {
 278			if (!down_read_trylock((&dc->writeback_lock))) {
 279				dc->rate_update_retry++;
 280				if (dc->rate_update_retry <=
 281				    BCH_WBRATE_UPDATE_MAX_SKIPS)
 282					break;
 283				down_read(&dc->writeback_lock);
 284				dc->rate_update_retry = 0;
 285			}
 286			__update_writeback_rate(dc);
 287			update_gc_after_writeback(c);
 288			up_read(&dc->writeback_lock);
 289		} while (0);
 290	}
 291
 292
 293	/*
 294	 * CACHE_SET_IO_DISABLE might be set via sysfs interface,
 295	 * check it here too.
 296	 */
 297	if (test_bit(BCACHE_DEV_WB_RUNNING, &dc->disk.flags) &&
 298	    !test_bit(CACHE_SET_IO_DISABLE, &c->flags)) {
 299		schedule_delayed_work(&dc->writeback_rate_update,
 300			      dc->writeback_rate_update_seconds * HZ);
 301	}
 302
 303	/*
 304	 * should check BCACHE_DEV_RATE_DW_RUNNING before calling
 305	 * cancel_delayed_work_sync().
 306	 */
 307	clear_bit(BCACHE_DEV_RATE_DW_RUNNING, &dc->disk.flags);
 308	/* paired with where BCACHE_DEV_RATE_DW_RUNNING is tested */
 309	smp_mb__after_atomic();
 310}
 311
 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 void dirty_io_destructor(struct closure *cl)
 345{
 346	struct dirty_io *io = container_of(cl, struct dirty_io, cl);
 347
 348	kfree(io);
 349}
 350
 351static void write_dirty_finish(struct closure *cl)
 352{
 353	struct dirty_io *io = container_of(cl, struct dirty_io, cl);
 354	struct keybuf_key *w = io->bio.bi_private;
 355	struct cached_dev *dc = io->dc;
 356
 357	bio_free_pages(&io->bio);
 358
 359	/* This is kind of a dumb way of signalling errors. */
 360	if (KEY_DIRTY(&w->key)) {
 361		int ret;
 362		unsigned int i;
 363		struct keylist keys;
 364
 365		bch_keylist_init(&keys);
 366
 367		bkey_copy(keys.top, &w->key);
 368		SET_KEY_DIRTY(keys.top, false);
 369		bch_keylist_push(&keys);
 370
 371		for (i = 0; i < KEY_PTRS(&w->key); i++)
 372			atomic_inc(&PTR_BUCKET(dc->disk.c, &w->key, i)->pin);
 373
 374		ret = bch_btree_insert(dc->disk.c, &keys, NULL, &w->key);
 375
 376		if (ret)
 377			trace_bcache_writeback_collision(&w->key);
 378
 379		atomic_long_inc(ret
 380				? &dc->disk.c->writeback_keys_failed
 381				: &dc->disk.c->writeback_keys_done);
 382	}
 383
 384	bch_keybuf_del(&dc->writeback_keys, w);
 385	up(&dc->in_flight);
 386
 387	closure_return_with_destructor(cl, dirty_io_destructor);
 388}
 389
 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 void write_dirty(struct closure *cl)
 404{
 405	struct dirty_io *io = container_of(cl, struct dirty_io, cl);
 406	struct keybuf_key *w = io->bio.bi_private;
 407	struct cached_dev *dc = io->dc;
 408
 409	uint16_t next_sequence;
 410
 411	if (atomic_read(&dc->writeback_sequence_next) != io->sequence) {
 412		/* Not our turn to write; wait for a write to complete */
 413		closure_wait(&dc->writeback_ordering_wait, cl);
 414
 415		if (atomic_read(&dc->writeback_sequence_next) == io->sequence) {
 416			/*
 417			 * Edge case-- it happened in indeterminate order
 418			 * relative to when we were added to wait list..
 419			 */
 420			closure_wake_up(&dc->writeback_ordering_wait);
 421		}
 422
 423		continue_at(cl, write_dirty, io->dc->writeback_write_wq);
 424		return;
 425	}
 426
 427	next_sequence = io->sequence + 1;
 428
 429	/*
 430	 * IO errors are signalled using the dirty bit on the key.
 431	 * If we failed to read, we should not attempt to write to the
 432	 * backing device.  Instead, immediately go to write_dirty_finish
 433	 * to clean up.
 434	 */
 435	if (KEY_DIRTY(&w->key)) {
 436		dirty_init(w);
 437		io->bio.bi_opf = REQ_OP_WRITE;
 438		io->bio.bi_iter.bi_sector = KEY_START(&w->key);
 439		bio_set_dev(&io->bio, io->dc->bdev);
 440		io->bio.bi_end_io	= dirty_endio;
 441
 442		/* I/O request sent to backing device */
 443		closure_bio_submit(io->dc->disk.c, &io->bio, cl);
 444	}
 445
 446	atomic_set(&dc->writeback_sequence_next, next_sequence);
 447	closure_wake_up(&dc->writeback_ordering_wait);
 448
 449	continue_at(cl, write_dirty_finish, io->dc->writeback_write_wq);
 450}
 451
 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 void read_dirty_submit(struct closure *cl)
 466{
 467	struct dirty_io *io = container_of(cl, struct dirty_io, cl);
 468
 469	closure_bio_submit(io->dc->disk.c, &io->bio, cl);
 470
 471	continue_at(cl, write_dirty, io->dc->writeback_write_wq);
 472}
 473
 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 some
 490	 * mempools.
 491	 */
 492
 493	next = bch_keybuf_next(&dc->writeback_keys);
 494
 495	while (!kthread_should_stop() &&
 496	       !test_bit(CACHE_SET_IO_DISABLE, &dc->disk.c->flags) &&
 497	       next) {
 498		size = 0;
 499		nk = 0;
 500
 501		do {
 502			BUG_ON(ptr_stale(dc->disk.c, &next->key, 0));
 503
 504			/*
 505			 * Don't combine too many operations, even if they
 506			 * are all small.
 507			 */
 508			if (nk >= MAX_WRITEBACKS_IN_PASS)
 509				break;
 510
 511			/*
 512			 * If the current operation is very large, don't
 513			 * further combine operations.
 514			 */
 515			if (size >= MAX_WRITESIZE_IN_PASS)
 516				break;
 517
 518			/*
 519			 * Operations are only eligible to be combined
 520			 * if they are contiguous.
 521			 *
 522			 * TODO: add a heuristic willing to fire a
 523			 * certain amount of non-contiguous IO per pass,
 524			 * so that we can benefit from backing device
 525			 * command queueing.
 526			 */
 527			if ((nk != 0) && bkey_cmp(&keys[nk-1]->key,
 528						&START_KEY(&next->key)))
 529				break;
 530
 531			size += KEY_SIZE(&next->key);
 532			keys[nk++] = next;
 533		} while ((next = bch_keybuf_next(&dc->writeback_keys)));
 534
 535		/* Now we have gathered a set of 1..5 keys to write back. */
 536		for (i = 0; i < nk; i++) {
 537			w = keys[i];
 538
 539			io = kzalloc(struct_size(io, bio.bi_inline_vecs,
 540						DIV_ROUND_UP(KEY_SIZE(&w->key), PAGE_SECTORS)),
 
 
 541				     GFP_KERNEL);
 542			if (!io)
 543				goto err;
 544
 545			w->private	= io;
 546			io->dc		= dc;
 547			io->sequence    = sequence++;
 548
 549			dirty_init(w);
 550			io->bio.bi_opf = REQ_OP_READ;
 551			io->bio.bi_iter.bi_sector = PTR_OFFSET(&w->key, 0);
 552			bio_set_dev(&io->bio, dc->disk.c->cache->bdev);
 
 553			io->bio.bi_end_io	= read_dirty_endio;
 554
 555			if (bch_bio_alloc_pages(&io->bio, GFP_KERNEL))
 556				goto err_free;
 557
 558			trace_bcache_writeback(&w->key);
 559
 560			down(&dc->in_flight);
 561
 562			/*
 563			 * We've acquired a semaphore for the maximum
 564			 * simultaneous number of writebacks; from here
 565			 * everything happens asynchronously.
 566			 */
 567			closure_call(&io->cl, read_dirty_submit, NULL, &cl);
 568		}
 569
 570		delay = writeback_delay(dc, size);
 571
 572		while (!kthread_should_stop() &&
 573		       !test_bit(CACHE_SET_IO_DISABLE, &dc->disk.c->flags) &&
 574		       delay) {
 575			schedule_timeout_interruptible(delay);
 576			delay = writeback_delay(dc, 0);
 577		}
 578	}
 579
 580	if (0) {
 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 be
 589	 * freed) before refilling again
 590	 */
 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 disk
 700 */
 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 bringup
 710	 * we might not be attached yet so this disk's inode nr isn't
 711	 * initialized then
 712	 */
 713	if (bkey_cmp(&buf->last_scanned, &start) < 0 ||
 714	    bkey_cmp(&buf->last_scanned, &end) > 0)
 715		buf->last_scanned = start;
 716
 717	if (dc->partial_stripes_expensive) {
 718		refill_full_stripes(dc);
 719		if (array_freelist_empty(&buf->freelist))
 720			return false;
 721	}
 722
 723	start_pos = buf->last_scanned;
 724	bch_refill_keybuf(dc->disk.c, buf, &end, dirty_pred);
 725
 726	if (bkey_cmp(&buf->last_scanned, &end) < 0)
 727		return false;
 728
 729	/*
 730	 * If we get to the end start scanning again from the beginning, and
 731	 * only scan up to where we initially started scanning from:
 732	 */
 733	buf->last_scanned = start;
 734	bch_refill_keybuf(dc->disk.c, buf, &start_pos, dirty_pred);
 735
 736	return bkey_cmp(&buf->last_scanned, &start_pos) >= 0;
 737}
 738
 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 continue
 753		 * to perform writeback. Otherwise, if no dirty data on cache,
 754		 * or there is dirty data on cache but writeback is disabled,
 755		 * the writeback thread should sleep here and wait for others
 756		 * to wake up it.
 757		 */
 758		if (!test_bit(BCACHE_DEV_DETACHING, &dc->disk.flags) &&
 759		    (!atomic_read(&dc->has_dirty) || !dc->writeback_running)) {
 760			up_write(&dc->writeback_lock);
 761
 762			if (kthread_should_stop() ||
 763			    test_bit(CACHE_SET_IO_DISABLE, &c->flags)) {
 764				set_current_state(TASK_RUNNING);
 765				break;
 766			}
 767
 768			schedule();
 769			continue;
 770		}
 771		set_current_state(TASK_RUNNING);
 772
 773		searched_full_index = refill_dirty(dc);
 774
 775		if (searched_full_index &&
 776		    RB_EMPTY_ROOT(&dc->writeback_keys.keys)) {
 777			atomic_set(&dc->has_dirty, 0);
 778			SET_BDEV_STATE(&dc->sb, BDEV_STATE_CLEAN);
 779			bch_write_bdev_super(dc, NULL);
 780			/*
 781			 * If bcache device is detaching via sysfs interface,
 782			 * writeback thread should stop after there is no dirty
 783			 * data on cache. BCACHE_DEV_DETACHING flag is set in
 784			 * bch_cached_dev_detach().
 785			 */
 786			if (test_bit(BCACHE_DEV_DETACHING, &dc->disk.flags)) {
 787				struct closure cl;
 788
 789				closure_init_stack(&cl);
 790				memset(&dc->sb.set_uuid, 0, 16);
 791				SET_BDEV_STATE(&dc->sb, BDEV_STATE_NONE);
 792
 793				bch_write_bdev_super(dc, &cl);
 794				closure_sync(&cl);
 795
 796				up_write(&dc->writeback_lock);
 797				break;
 798			}
 799
 800			/*
 801			 * When dirty data rate is high (e.g. 50%+), there might
 802			 * be heavy buckets fragmentation after writeback
 803			 * finished, which hurts following write performance.
 804			 * If users really care about write performance they
 805			 * may set BCH_ENABLE_AUTO_GC via sysfs, then when
 806			 * BCH_DO_AUTO_GC is set, garbage collection thread
 807			 * will be wake up here. After moving gc, the shrunk
 808			 * btree and discarded free buckets SSD space may be
 809			 * helpful for following write requests.
 810			 */
 811			if (c->gc_after_writeback ==
 812			    (BCH_ENABLE_AUTO_GC|BCH_DO_AUTO_GC)) {
 813				c->gc_after_writeback &= ~BCH_DO_AUTO_GC;
 814				force_wake_up_gc(c);
 815			}
 816		}
 817
 818		up_write(&dc->writeback_lock);
 819
 820		read_dirty(dc);
 821
 822		if (searched_full_index) {
 823			unsigned int delay = dc->writeback_delay * HZ;
 824
 825			while (delay &&
 826			       !kthread_should_stop() &&
 827			       !test_bit(CACHE_SET_IO_DISABLE, &c->flags) &&
 828			       !test_bit(BCACHE_DEV_DETACHING, &dc->disk.flags))
 829				delay = schedule_timeout_interruptible(delay);
 830
 831			bch_ratelimit_reset(&dc->writeback_rate);
 832		}
 833	}
 834
 835	if (dc->writeback_write_wq)
 
 836		destroy_workqueue(dc->writeback_write_wq);
 837
 838	cached_dev_put(dc);
 839	wait_for_kthread_stop();
 840
 841	return 0;
 842}
 843
 844/* Init */
 845#define INIT_KEYS_EACH_TIME	500000
 
 846
 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	return ret;
 894}
 895
 896static int bch_dirty_init_thread(void *arg)
 897{
 898	struct dirty_init_thrd_info *info = arg;
 899	struct bch_dirty_init_state *state = info->state;
 900	struct cache_set *c = state->c;
 901	struct btree_iter iter;
 902	struct bkey *k, *p;
 903	int cur_idx, prev_idx, skip_nr;
 904
 905	k = p = NULL;
 906	cur_idx = prev_idx = 0;
 907
 908	bch_btree_iter_init(&c->root->keys, &iter, NULL);
 909	k = bch_btree_iter_next_filter(&iter, &c->root->keys, bch_ptr_bad);
 910	BUG_ON(!k);
 911
 912	p = k;
 913
 914	while (k) {
 915		spin_lock(&state->idx_lock);
 916		cur_idx = state->key_idx;
 917		state->key_idx++;
 918		spin_unlock(&state->idx_lock);
 919
 920		skip_nr = cur_idx - prev_idx;
 921
 922		while (skip_nr) {
 923			k = bch_btree_iter_next_filter(&iter,
 924						       &c->root->keys,
 925						       bch_ptr_bad);
 926			if (k)
 927				p = k;
 928			else {
 929				atomic_set(&state->enough, 1);
 930				/* Update state->enough earlier */
 931				smp_mb__after_atomic();
 932				goto out;
 933			}
 934			skip_nr--;
 935		}
 936
 937		if (p) {
 938			if (bch_root_node_dirty_init(c, state->d, p) < 0)
 939				goto out;
 940		}
 941
 942		p = NULL;
 943		prev_idx = cur_idx;
 944	}
 945
 946out:
 947	/* In order to wake up state->wait in time */
 948	smp_mb__before_atomic();
 949	if (atomic_dec_and_test(&state->started))
 950		wake_up(&state->wait);
 951
 952	return 0;
 953}
 954
 955static int bch_btre_dirty_init_thread_nr(void)
 956{
 957	int n = num_online_cpus()/2;
 958
 959	if (n == 0)
 960		n = 1;
 961	else if (n > BCH_DIRTY_INIT_THRD_MAX)
 962		n = BCH_DIRTY_INIT_THRD_MAX;
 963
 964	return n;
 965}
 966
 967void bch_sectors_dirty_init(struct bcache_device *d)
 968{
 969	int i;
 970	struct bkey *k = NULL;
 971	struct btree_iter iter;
 972	struct sectors_dirty_init op;
 973	struct cache_set *c = d->c;
 974	struct bch_dirty_init_state state;
 975
 976	/* Just count root keys if no leaf node */
 977	rw_lock(0, c->root, c->root->level);
 978	if (c->root->level == 0) {
 979		bch_btree_op_init(&op.op, -1);
 980		op.inode = d->id;
 981		op.count = 0;
 982
 983		for_each_key_filter(&c->root->keys,
 984				    k, &iter, bch_ptr_invalid)
 985			sectors_dirty_init_fn(&op.op, c->root, k);
 986
 987		rw_unlock(0, c->root);
 988		return;
 989	}
 990
 991	memset(&state, 0, sizeof(struct bch_dirty_init_state));
 992	state.c = c;
 993	state.d = d;
 994	state.total_threads = bch_btre_dirty_init_thread_nr();
 995	state.key_idx = 0;
 996	spin_lock_init(&state.idx_lock);
 997	atomic_set(&state.started, 0);
 998	atomic_set(&state.enough, 0);
 999	init_waitqueue_head(&state.wait);
1000
1001	for (i = 0; i < state.total_threads; i++) {
1002		/* Fetch latest state.enough earlier */
1003		smp_mb__before_atomic();
1004		if (atomic_read(&state.enough))
1005			break;
1006
1007		state.infos[i].state = &state;
1008		state.infos[i].thread =
1009			kthread_run(bch_dirty_init_thread, &state.infos[i],
1010				    "bch_dirtcnt[%d]", i);
1011		if (IS_ERR(state.infos[i].thread)) {
1012			pr_err("fails to run thread bch_dirty_init[%d]\n", i);
1013			for (--i; i >= 0; i--)
1014				kthread_stop(state.infos[i].thread);
1015			goto out;
1016		}
1017		atomic_inc(&state.started);
1018	}
1019
1020out:
1021	/* Must wait for all threads to stop. */
1022	wait_event(state.wait, atomic_read(&state.started) == 0);
1023	rw_unlock(0, c->root);
1024}
1025
1026void bch_cached_dev_writeback_init(struct cached_dev *dc)
1027{
1028	sema_init(&dc->in_flight, 64);
1029	init_rwsem(&dc->writeback_lock);
1030	bch_keybuf_init(&dc->writeback_keys);
1031
1032	dc->writeback_metadata		= true;
1033	dc->writeback_running		= false;
1034	dc->writeback_consider_fragment = true;
1035	dc->writeback_percent		= 10;
1036	dc->writeback_delay		= 30;
1037	atomic_long_set(&dc->writeback_rate.rate, 1024);
1038	dc->writeback_rate_minimum	= 8;
1039
1040	dc->writeback_rate_update_seconds = WRITEBACK_RATE_UPDATE_SECS_DEFAULT;
1041	dc->writeback_rate_p_term_inverse = 40;
1042	dc->writeback_rate_fp_term_low = 1;
1043	dc->writeback_rate_fp_term_mid = 10;
1044	dc->writeback_rate_fp_term_high = 1000;
1045	dc->writeback_rate_i_term_inverse = 10000;
1046
1047	/* For dc->writeback_lock contention in update_writeback_rate() */
1048	dc->rate_update_retry = 0;
1049
1050	WARN_ON(test_and_clear_bit(BCACHE_DEV_WB_RUNNING, &dc->disk.flags));
1051	INIT_DELAYED_WORK(&dc->writeback_rate_update, update_writeback_rate);
1052}
1053
1054int bch_cached_dev_writeback_start(struct cached_dev *dc)
1055{
1056	dc->writeback_write_wq = alloc_workqueue("bcache_writeback_wq",
1057						WQ_MEM_RECLAIM, 0);
1058	if (!dc->writeback_write_wq)
1059		return -ENOMEM;
1060
1061	cached_dev_get(dc);
1062	dc->writeback_thread = kthread_create(bch_writeback_thread, dc,
1063					      "bcache_writeback");
1064	if (IS_ERR(dc->writeback_thread)) {
1065		cached_dev_put(dc);
1066		destroy_workqueue(dc->writeback_write_wq);
1067		return PTR_ERR(dc->writeback_thread);
1068	}
1069	dc->writeback_running = true;
1070
1071	WARN_ON(test_and_set_bit(BCACHE_DEV_WB_RUNNING, &dc->disk.flags));
1072	schedule_delayed_work(&dc->writeback_rate_update,
1073			      dc->writeback_rate_update_seconds * HZ);
1074
1075	bch_writeback_queue(dc);
1076
1077	return 0;
1078}
v5.4
  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
 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 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
 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 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
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
122static bool set_at_max_writeback_rate(struct cache_set *c,
123				       struct cached_dev *dc)
124{
 
 
 
 
125	/* Don't set max writeback rate if gc is running */
126	if (!c->gc_mark_valid)
127		return false;
128	/*
129	 * Idle_counter is increased everytime when update_writeback_rate() is
130	 * called. If all backing devices attached to the same cache set have
131	 * identical dc->writeback_rate_update_seconds values, it is about 6
132	 * rounds of update_writeback_rate() on each backing device before
133	 * c->at_max_writeback_rate is set to 1, and then max wrteback rate set
134	 * to each dc->writeback_rate.rate.
135	 * In order to avoid extra locking cost for counting exact dirty cached
136	 * devices number, c->attached_dev_nr is used to calculate the idle
137	 * throushold. It might be bigger if not all cached device are in write-
138	 * back mode, but it still works well with limited extra rounds of
139	 * update_writeback_rate().
140	 */
141	if (atomic_inc_return(&c->idle_counter) <
142	    atomic_read(&c->attached_dev_nr) * 6)
143		return false;
144
145	if (atomic_read(&c->at_max_writeback_rate) != 1)
146		atomic_set(&c->at_max_writeback_rate, 1);
147
148	atomic_long_set(&dc->writeback_rate.rate, INT_MAX);
149
150	/* keep writeback_rate_target as existing value */
151	dc->writeback_rate_proportional = 0;
152	dc->writeback_rate_integral_scaled = 0;
153	dc->writeback_rate_change = 0;
154
155	/*
156	 * Check c->idle_counter and c->at_max_writeback_rate agagain in case
157	 * new I/O arrives during before set_at_max_writeback_rate() returns.
158	 * Then the writeback rate is set to 1, and its new value should be
159	 * decided via __update_writeback_rate().
160	 */
161	if ((atomic_read(&c->idle_counter) <
162	     atomic_read(&c->attached_dev_nr) * 6) ||
163	    !atomic_read(&c->at_max_writeback_rate))
164		return false;
165
166	return true;
167}
168
169static void update_writeback_rate(struct work_struct *work)
170{
171	struct cached_dev *dc = container_of(to_delayed_work(work),
172					     struct cached_dev,
173					     writeback_rate_update);
174	struct cache_set *c = dc->disk.c;
175
176	/*
177	 * should check BCACHE_DEV_RATE_DW_RUNNING before calling
178	 * cancel_delayed_work_sync().
179	 */
180	set_bit(BCACHE_DEV_RATE_DW_RUNNING, &dc->disk.flags);
181	/* paired with where BCACHE_DEV_RATE_DW_RUNNING is tested */
182	smp_mb();
183
184	/*
185	 * CACHE_SET_IO_DISABLE might be set via sysfs interface,
186	 * check it here too.
187	 */
188	if (!test_bit(BCACHE_DEV_WB_RUNNING, &dc->disk.flags) ||
189	    test_bit(CACHE_SET_IO_DISABLE, &c->flags)) {
190		clear_bit(BCACHE_DEV_RATE_DW_RUNNING, &dc->disk.flags);
191		/* paired with where BCACHE_DEV_RATE_DW_RUNNING is tested */
192		smp_mb();
193		return;
194	}
195
196	if (atomic_read(&dc->has_dirty) && dc->writeback_percent) {
197		/*
198		 * If the whole cache set is idle, set_at_max_writeback_rate()
199		 * will set writeback rate to a max number. Then it is
200		 * unncessary to update writeback rate for an idle cache set
201		 * in maximum writeback rate number(s).
202		 */
203		if (!set_at_max_writeback_rate(c, dc)) {
204			down_read(&dc->writeback_lock);
 
 
 
 
 
 
 
 
205			__update_writeback_rate(dc);
206			update_gc_after_writeback(c);
207			up_read(&dc->writeback_lock);
208		}
209	}
210
211
212	/*
213	 * CACHE_SET_IO_DISABLE might be set via sysfs interface,
214	 * check it here too.
215	 */
216	if (test_bit(BCACHE_DEV_WB_RUNNING, &dc->disk.flags) &&
217	    !test_bit(CACHE_SET_IO_DISABLE, &c->flags)) {
218		schedule_delayed_work(&dc->writeback_rate_update,
219			      dc->writeback_rate_update_seconds * HZ);
220	}
221
222	/*
223	 * should check BCACHE_DEV_RATE_DW_RUNNING before calling
224	 * cancel_delayed_work_sync().
225	 */
226	clear_bit(BCACHE_DEV_RATE_DW_RUNNING, &dc->disk.flags);
227	/* paired with where BCACHE_DEV_RATE_DW_RUNNING is tested */
228	smp_mb();
229}
230
231static unsigned int writeback_delay(struct cached_dev *dc,
232				    unsigned int sectors)
233{
234	if (test_bit(BCACHE_DEV_DETACHING, &dc->disk.flags) ||
235	    !dc->writeback_percent)
236		return 0;
237
238	return bch_next_delay(&dc->writeback_rate, sectors);
239}
240
241struct dirty_io {
242	struct closure		cl;
243	struct cached_dev	*dc;
244	uint16_t		sequence;
245	struct bio		bio;
246};
247
248static void dirty_init(struct keybuf_key *w)
249{
250	struct dirty_io *io = w->private;
251	struct bio *bio = &io->bio;
252
253	bio_init(bio, bio->bi_inline_vecs,
254		 DIV_ROUND_UP(KEY_SIZE(&w->key), PAGE_SECTORS));
255	if (!io->dc->writeback_percent)
256		bio_set_prio(bio, IOPRIO_PRIO_VALUE(IOPRIO_CLASS_IDLE, 0));
257
258	bio->bi_iter.bi_size	= KEY_SIZE(&w->key) << 9;
259	bio->bi_private		= w;
260	bch_bio_map(bio, NULL);
261}
262
263static void dirty_io_destructor(struct closure *cl)
264{
265	struct dirty_io *io = container_of(cl, struct dirty_io, cl);
266
267	kfree(io);
268}
269
270static void write_dirty_finish(struct closure *cl)
271{
272	struct dirty_io *io = container_of(cl, struct dirty_io, cl);
273	struct keybuf_key *w = io->bio.bi_private;
274	struct cached_dev *dc = io->dc;
275
276	bio_free_pages(&io->bio);
277
278	/* This is kind of a dumb way of signalling errors. */
279	if (KEY_DIRTY(&w->key)) {
280		int ret;
281		unsigned int i;
282		struct keylist keys;
283
284		bch_keylist_init(&keys);
285
286		bkey_copy(keys.top, &w->key);
287		SET_KEY_DIRTY(keys.top, false);
288		bch_keylist_push(&keys);
289
290		for (i = 0; i < KEY_PTRS(&w->key); i++)
291			atomic_inc(&PTR_BUCKET(dc->disk.c, &w->key, i)->pin);
292
293		ret = bch_btree_insert(dc->disk.c, &keys, NULL, &w->key);
294
295		if (ret)
296			trace_bcache_writeback_collision(&w->key);
297
298		atomic_long_inc(ret
299				? &dc->disk.c->writeback_keys_failed
300				: &dc->disk.c->writeback_keys_done);
301	}
302
303	bch_keybuf_del(&dc->writeback_keys, w);
304	up(&dc->in_flight);
305
306	closure_return_with_destructor(cl, dirty_io_destructor);
307}
308
309static void dirty_endio(struct bio *bio)
310{
311	struct keybuf_key *w = bio->bi_private;
312	struct dirty_io *io = w->private;
313
314	if (bio->bi_status) {
315		SET_KEY_DIRTY(&w->key, false);
316		bch_count_backing_io_errors(io->dc, bio);
317	}
318
319	closure_put(&io->cl);
320}
321
322static void write_dirty(struct closure *cl)
323{
324	struct dirty_io *io = container_of(cl, struct dirty_io, cl);
325	struct keybuf_key *w = io->bio.bi_private;
326	struct cached_dev *dc = io->dc;
327
328	uint16_t next_sequence;
329
330	if (atomic_read(&dc->writeback_sequence_next) != io->sequence) {
331		/* Not our turn to write; wait for a write to complete */
332		closure_wait(&dc->writeback_ordering_wait, cl);
333
334		if (atomic_read(&dc->writeback_sequence_next) == io->sequence) {
335			/*
336			 * Edge case-- it happened in indeterminate order
337			 * relative to when we were added to wait list..
338			 */
339			closure_wake_up(&dc->writeback_ordering_wait);
340		}
341
342		continue_at(cl, write_dirty, io->dc->writeback_write_wq);
343		return;
344	}
345
346	next_sequence = io->sequence + 1;
347
348	/*
349	 * IO errors are signalled using the dirty bit on the key.
350	 * If we failed to read, we should not attempt to write to the
351	 * backing device.  Instead, immediately go to write_dirty_finish
352	 * to clean up.
353	 */
354	if (KEY_DIRTY(&w->key)) {
355		dirty_init(w);
356		bio_set_op_attrs(&io->bio, REQ_OP_WRITE, 0);
357		io->bio.bi_iter.bi_sector = KEY_START(&w->key);
358		bio_set_dev(&io->bio, io->dc->bdev);
359		io->bio.bi_end_io	= dirty_endio;
360
361		/* I/O request sent to backing device */
362		closure_bio_submit(io->dc->disk.c, &io->bio, cl);
363	}
364
365	atomic_set(&dc->writeback_sequence_next, next_sequence);
366	closure_wake_up(&dc->writeback_ordering_wait);
367
368	continue_at(cl, write_dirty_finish, io->dc->writeback_write_wq);
369}
370
371static void read_dirty_endio(struct bio *bio)
372{
373	struct keybuf_key *w = bio->bi_private;
374	struct dirty_io *io = w->private;
375
376	/* is_read = 1 */
377	bch_count_io_errors(PTR_CACHE(io->dc->disk.c, &w->key, 0),
378			    bio->bi_status, 1,
379			    "reading dirty data from cache");
380
381	dirty_endio(bio);
382}
383
384static void read_dirty_submit(struct closure *cl)
385{
386	struct dirty_io *io = container_of(cl, struct dirty_io, cl);
387
388	closure_bio_submit(io->dc->disk.c, &io->bio, cl);
389
390	continue_at(cl, write_dirty, io->dc->writeback_write_wq);
391}
392
393static void read_dirty(struct cached_dev *dc)
394{
395	unsigned int delay = 0;
396	struct keybuf_key *next, *keys[MAX_WRITEBACKS_IN_PASS], *w;
397	size_t size;
398	int nk, i;
399	struct dirty_io *io;
400	struct closure cl;
401	uint16_t sequence = 0;
402
403	BUG_ON(!llist_empty(&dc->writeback_ordering_wait.list));
404	atomic_set(&dc->writeback_sequence_next, sequence);
405	closure_init_stack(&cl);
406
407	/*
408	 * XXX: if we error, background writeback just spins. Should use some
409	 * mempools.
410	 */
411
412	next = bch_keybuf_next(&dc->writeback_keys);
413
414	while (!kthread_should_stop() &&
415	       !test_bit(CACHE_SET_IO_DISABLE, &dc->disk.c->flags) &&
416	       next) {
417		size = 0;
418		nk = 0;
419
420		do {
421			BUG_ON(ptr_stale(dc->disk.c, &next->key, 0));
422
423			/*
424			 * Don't combine too many operations, even if they
425			 * are all small.
426			 */
427			if (nk >= MAX_WRITEBACKS_IN_PASS)
428				break;
429
430			/*
431			 * If the current operation is very large, don't
432			 * further combine operations.
433			 */
434			if (size >= MAX_WRITESIZE_IN_PASS)
435				break;
436
437			/*
438			 * Operations are only eligible to be combined
439			 * if they are contiguous.
440			 *
441			 * TODO: add a heuristic willing to fire a
442			 * certain amount of non-contiguous IO per pass,
443			 * so that we can benefit from backing device
444			 * command queueing.
445			 */
446			if ((nk != 0) && bkey_cmp(&keys[nk-1]->key,
447						&START_KEY(&next->key)))
448				break;
449
450			size += KEY_SIZE(&next->key);
451			keys[nk++] = next;
452		} while ((next = bch_keybuf_next(&dc->writeback_keys)));
453
454		/* Now we have gathered a set of 1..5 keys to write back. */
455		for (i = 0; i < nk; i++) {
456			w = keys[i];
457
458			io = kzalloc(sizeof(struct dirty_io) +
459				     sizeof(struct bio_vec) *
460				     DIV_ROUND_UP(KEY_SIZE(&w->key),
461						  PAGE_SECTORS),
462				     GFP_KERNEL);
463			if (!io)
464				goto err;
465
466			w->private	= io;
467			io->dc		= dc;
468			io->sequence    = sequence++;
469
470			dirty_init(w);
471			bio_set_op_attrs(&io->bio, REQ_OP_READ, 0);
472			io->bio.bi_iter.bi_sector = PTR_OFFSET(&w->key, 0);
473			bio_set_dev(&io->bio,
474				    PTR_CACHE(dc->disk.c, &w->key, 0)->bdev);
475			io->bio.bi_end_io	= read_dirty_endio;
476
477			if (bch_bio_alloc_pages(&io->bio, GFP_KERNEL))
478				goto err_free;
479
480			trace_bcache_writeback(&w->key);
481
482			down(&dc->in_flight);
483
484			/*
485			 * We've acquired a semaphore for the maximum
486			 * simultaneous number of writebacks; from here
487			 * everything happens asynchronously.
488			 */
489			closure_call(&io->cl, read_dirty_submit, NULL, &cl);
490		}
491
492		delay = writeback_delay(dc, size);
493
494		while (!kthread_should_stop() &&
495		       !test_bit(CACHE_SET_IO_DISABLE, &dc->disk.c->flags) &&
496		       delay) {
497			schedule_timeout_interruptible(delay);
498			delay = writeback_delay(dc, 0);
499		}
500	}
501
502	if (0) {
503err_free:
504		kfree(w->private);
505err:
506		bch_keybuf_del(&dc->writeback_keys, w);
507	}
508
509	/*
510	 * Wait for outstanding writeback IOs to finish (and keybuf slots to be
511	 * freed) before refilling again
512	 */
513	closure_sync(&cl);
514}
515
516/* Scan for dirty data */
517
518void bcache_dev_sectors_dirty_add(struct cache_set *c, unsigned int inode,
519				  uint64_t offset, int nr_sectors)
520{
521	struct bcache_device *d = c->devices[inode];
522	unsigned int stripe_offset, stripe, sectors_dirty;
 
523
524	if (!d)
525		return;
526
 
 
 
 
527	if (UUID_FLASH_ONLY(&c->uuids[inode]))
528		atomic_long_add(nr_sectors, &c->flash_dev_dirty_sectors);
529
530	stripe = offset_to_stripe(d, offset);
531	stripe_offset = offset & (d->stripe_size - 1);
532
533	while (nr_sectors) {
534		int s = min_t(unsigned int, abs(nr_sectors),
535			      d->stripe_size - stripe_offset);
536
537		if (nr_sectors < 0)
538			s = -s;
539
540		if (stripe >= d->nr_stripes)
541			return;
542
543		sectors_dirty = atomic_add_return(s,
544					d->stripe_sectors_dirty + stripe);
545		if (sectors_dirty == d->stripe_size)
546			set_bit(stripe, d->full_dirty_stripes);
547		else
548			clear_bit(stripe, d->full_dirty_stripes);
 
 
 
549
550		nr_sectors -= s;
551		stripe_offset = 0;
552		stripe++;
553	}
554}
555
556static bool dirty_pred(struct keybuf *buf, struct bkey *k)
557{
558	struct cached_dev *dc = container_of(buf,
559					     struct cached_dev,
560					     writeback_keys);
561
562	BUG_ON(KEY_INODE(k) != dc->disk.id);
563
564	return KEY_DIRTY(k);
565}
566
567static void refill_full_stripes(struct cached_dev *dc)
568{
569	struct keybuf *buf = &dc->writeback_keys;
570	unsigned int start_stripe, stripe, next_stripe;
 
571	bool wrapped = false;
572
573	stripe = offset_to_stripe(&dc->disk, KEY_OFFSET(&buf->last_scanned));
574
575	if (stripe >= dc->disk.nr_stripes)
576		stripe = 0;
577
578	start_stripe = stripe;
579
580	while (1) {
581		stripe = find_next_bit(dc->disk.full_dirty_stripes,
582				       dc->disk.nr_stripes, stripe);
583
584		if (stripe == dc->disk.nr_stripes)
585			goto next;
586
587		next_stripe = find_next_zero_bit(dc->disk.full_dirty_stripes,
588						 dc->disk.nr_stripes, stripe);
589
590		buf->last_scanned = KEY(dc->disk.id,
591					stripe * dc->disk.stripe_size, 0);
592
593		bch_refill_keybuf(dc->disk.c, buf,
594				  &KEY(dc->disk.id,
595				       next_stripe * dc->disk.stripe_size, 0),
596				  dirty_pred);
597
598		if (array_freelist_empty(&buf->freelist))
599			return;
600
601		stripe = next_stripe;
602next:
603		if (wrapped && stripe > start_stripe)
604			return;
605
606		if (stripe == dc->disk.nr_stripes) {
607			stripe = 0;
608			wrapped = true;
609		}
610	}
611}
612
613/*
614 * Returns true if we scanned the entire disk
615 */
616static bool refill_dirty(struct cached_dev *dc)
617{
618	struct keybuf *buf = &dc->writeback_keys;
619	struct bkey start = KEY(dc->disk.id, 0, 0);
620	struct bkey end = KEY(dc->disk.id, MAX_KEY_OFFSET, 0);
621	struct bkey start_pos;
622
623	/*
624	 * make sure keybuf pos is inside the range for this disk - at bringup
625	 * we might not be attached yet so this disk's inode nr isn't
626	 * initialized then
627	 */
628	if (bkey_cmp(&buf->last_scanned, &start) < 0 ||
629	    bkey_cmp(&buf->last_scanned, &end) > 0)
630		buf->last_scanned = start;
631
632	if (dc->partial_stripes_expensive) {
633		refill_full_stripes(dc);
634		if (array_freelist_empty(&buf->freelist))
635			return false;
636	}
637
638	start_pos = buf->last_scanned;
639	bch_refill_keybuf(dc->disk.c, buf, &end, dirty_pred);
640
641	if (bkey_cmp(&buf->last_scanned, &end) < 0)
642		return false;
643
644	/*
645	 * If we get to the end start scanning again from the beginning, and
646	 * only scan up to where we initially started scanning from:
647	 */
648	buf->last_scanned = start;
649	bch_refill_keybuf(dc->disk.c, buf, &start_pos, dirty_pred);
650
651	return bkey_cmp(&buf->last_scanned, &start_pos) >= 0;
652}
653
654static int bch_writeback_thread(void *arg)
655{
656	struct cached_dev *dc = arg;
657	struct cache_set *c = dc->disk.c;
658	bool searched_full_index;
659
660	bch_ratelimit_reset(&dc->writeback_rate);
661
662	while (!kthread_should_stop() &&
663	       !test_bit(CACHE_SET_IO_DISABLE, &c->flags)) {
664		down_write(&dc->writeback_lock);
665		set_current_state(TASK_INTERRUPTIBLE);
666		/*
667		 * If the bache device is detaching, skip here and continue
668		 * to perform writeback. Otherwise, if no dirty data on cache,
669		 * or there is dirty data on cache but writeback is disabled,
670		 * the writeback thread should sleep here and wait for others
671		 * to wake up it.
672		 */
673		if (!test_bit(BCACHE_DEV_DETACHING, &dc->disk.flags) &&
674		    (!atomic_read(&dc->has_dirty) || !dc->writeback_running)) {
675			up_write(&dc->writeback_lock);
676
677			if (kthread_should_stop() ||
678			    test_bit(CACHE_SET_IO_DISABLE, &c->flags)) {
679				set_current_state(TASK_RUNNING);
680				break;
681			}
682
683			schedule();
684			continue;
685		}
686		set_current_state(TASK_RUNNING);
687
688		searched_full_index = refill_dirty(dc);
689
690		if (searched_full_index &&
691		    RB_EMPTY_ROOT(&dc->writeback_keys.keys)) {
692			atomic_set(&dc->has_dirty, 0);
693			SET_BDEV_STATE(&dc->sb, BDEV_STATE_CLEAN);
694			bch_write_bdev_super(dc, NULL);
695			/*
696			 * If bcache device is detaching via sysfs interface,
697			 * writeback thread should stop after there is no dirty
698			 * data on cache. BCACHE_DEV_DETACHING flag is set in
699			 * bch_cached_dev_detach().
700			 */
701			if (test_bit(BCACHE_DEV_DETACHING, &dc->disk.flags)) {
 
 
 
 
 
 
 
 
 
702				up_write(&dc->writeback_lock);
703				break;
704			}
705
706			/*
707			 * When dirty data rate is high (e.g. 50%+), there might
708			 * be heavy buckets fragmentation after writeback
709			 * finished, which hurts following write performance.
710			 * If users really care about write performance they
711			 * may set BCH_ENABLE_AUTO_GC via sysfs, then when
712			 * BCH_DO_AUTO_GC is set, garbage collection thread
713			 * will be wake up here. After moving gc, the shrunk
714			 * btree and discarded free buckets SSD space may be
715			 * helpful for following write requests.
716			 */
717			if (c->gc_after_writeback ==
718			    (BCH_ENABLE_AUTO_GC|BCH_DO_AUTO_GC)) {
719				c->gc_after_writeback &= ~BCH_DO_AUTO_GC;
720				force_wake_up_gc(c);
721			}
722		}
723
724		up_write(&dc->writeback_lock);
725
726		read_dirty(dc);
727
728		if (searched_full_index) {
729			unsigned int delay = dc->writeback_delay * HZ;
730
731			while (delay &&
732			       !kthread_should_stop() &&
733			       !test_bit(CACHE_SET_IO_DISABLE, &c->flags) &&
734			       !test_bit(BCACHE_DEV_DETACHING, &dc->disk.flags))
735				delay = schedule_timeout_interruptible(delay);
736
737			bch_ratelimit_reset(&dc->writeback_rate);
738		}
739	}
740
741	if (dc->writeback_write_wq) {
742		flush_workqueue(dc->writeback_write_wq);
743		destroy_workqueue(dc->writeback_write_wq);
744	}
745	cached_dev_put(dc);
746	wait_for_kthread_stop();
747
748	return 0;
749}
750
751/* Init */
752#define INIT_KEYS_EACH_TIME	500000
753#define INIT_KEYS_SLEEP_MS	100
754
755struct sectors_dirty_init {
756	struct btree_op	op;
757	unsigned int	inode;
758	size_t		count;
759	struct bkey	start;
760};
761
762static int sectors_dirty_init_fn(struct btree_op *_op, struct btree *b,
763				 struct bkey *k)
764{
765	struct sectors_dirty_init *op = container_of(_op,
766						struct sectors_dirty_init, op);
767	if (KEY_INODE(k) > op->inode)
768		return MAP_DONE;
769
770	if (KEY_DIRTY(k))
771		bcache_dev_sectors_dirty_add(b->c, KEY_INODE(k),
772					     KEY_START(k), KEY_SIZE(k));
773
774	op->count++;
775	if (atomic_read(&b->c->search_inflight) &&
776	    !(op->count % INIT_KEYS_EACH_TIME)) {
777		bkey_copy_key(&op->start, k);
778		return -EAGAIN;
779	}
780
781	return MAP_CONTINUE;
782}
783
784void bch_sectors_dirty_init(struct bcache_device *d)
 
 
785{
786	struct sectors_dirty_init op;
787	int ret;
788
789	bch_btree_op_init(&op.op, -1);
790	op.inode = d->id;
791	op.count = 0;
792	op.start = KEY(op.inode, 0, 0);
793
794	do {
795		ret = bch_btree_map_keys(&op.op, d->c, &op.start,
796					 sectors_dirty_init_fn, 0);
797		if (ret == -EAGAIN)
798			schedule_timeout_interruptible(
799				msecs_to_jiffies(INIT_KEYS_SLEEP_MS));
800		else if (ret < 0) {
801			pr_warn("sectors dirty init failed, ret=%d!", ret);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
802			break;
 
 
 
 
 
 
 
 
 
 
803		}
804	} while (ret == -EAGAIN);
 
 
 
 
 
 
805}
806
807void bch_cached_dev_writeback_init(struct cached_dev *dc)
808{
809	sema_init(&dc->in_flight, 64);
810	init_rwsem(&dc->writeback_lock);
811	bch_keybuf_init(&dc->writeback_keys);
812
813	dc->writeback_metadata		= true;
814	dc->writeback_running		= false;
 
815	dc->writeback_percent		= 10;
816	dc->writeback_delay		= 30;
817	atomic_long_set(&dc->writeback_rate.rate, 1024);
818	dc->writeback_rate_minimum	= 8;
819
820	dc->writeback_rate_update_seconds = WRITEBACK_RATE_UPDATE_SECS_DEFAULT;
821	dc->writeback_rate_p_term_inverse = 40;
 
 
 
822	dc->writeback_rate_i_term_inverse = 10000;
 
 
 
823
824	WARN_ON(test_and_clear_bit(BCACHE_DEV_WB_RUNNING, &dc->disk.flags));
825	INIT_DELAYED_WORK(&dc->writeback_rate_update, update_writeback_rate);
826}
827
828int bch_cached_dev_writeback_start(struct cached_dev *dc)
829{
830	dc->writeback_write_wq = alloc_workqueue("bcache_writeback_wq",
831						WQ_MEM_RECLAIM, 0);
832	if (!dc->writeback_write_wq)
833		return -ENOMEM;
834
835	cached_dev_get(dc);
836	dc->writeback_thread = kthread_create(bch_writeback_thread, dc,
837					      "bcache_writeback");
838	if (IS_ERR(dc->writeback_thread)) {
839		cached_dev_put(dc);
840		destroy_workqueue(dc->writeback_write_wq);
841		return PTR_ERR(dc->writeback_thread);
842	}
843	dc->writeback_running = true;
844
845	WARN_ON(test_and_set_bit(BCACHE_DEV_WB_RUNNING, &dc->disk.flags));
846	schedule_delayed_work(&dc->writeback_rate_update,
847			      dc->writeback_rate_update_seconds * HZ);
848
849	bch_writeback_queue(dc);
850
851	return 0;
852}