2 * background writeback - scan btree for dirty data and write it to the backing
5 * Copyright 2010, 2011 Kent Overstreet <kent.overstreet@gmail.com>
6 * Copyright 2012 Google, Inc.
12 #include "writeback.h"
14 #include <linux/delay.h>
15 #include <linux/kthread.h>
16 #include <trace/events/bcache.h>
20 static void __update_writeback_rate(struct cached_dev *dc)
22 struct cache_set *c = dc->disk.c;
23 uint64_t cache_sectors = c->nbuckets * c->sb.bucket_size -
24 bcache_flash_devs_sectors_dirty(c);
25 uint64_t cache_dirty_target =
26 div_u64(cache_sectors * dc->writeback_percent, 100);
28 int64_t target = div64_u64(cache_dirty_target * bdev_sectors(dc->bdev),
29 c->cached_dev_sectors);
33 int64_t dirty = bcache_dev_sectors_dirty(&dc->disk);
34 int64_t derivative = dirty - dc->disk.sectors_dirty_last;
35 int64_t proportional = dirty - target;
38 dc->disk.sectors_dirty_last = dirty;
40 /* Scale to sectors per second */
42 proportional *= dc->writeback_rate_update_seconds;
43 proportional = div_s64(proportional, dc->writeback_rate_p_term_inverse);
45 derivative = div_s64(derivative, dc->writeback_rate_update_seconds);
47 derivative = ewma_add(dc->disk.sectors_dirty_derivative, derivative,
48 (dc->writeback_rate_d_term /
49 dc->writeback_rate_update_seconds) ?: 1, 0);
51 derivative *= dc->writeback_rate_d_term;
52 derivative = div_s64(derivative, dc->writeback_rate_p_term_inverse);
54 change = proportional + derivative;
56 /* Don't increase writeback rate if the device isn't keeping up */
58 time_after64(local_clock(),
59 dc->writeback_rate.next + NSEC_PER_MSEC))
62 dc->writeback_rate.rate =
63 clamp_t(int64_t, (int64_t) dc->writeback_rate.rate + change,
66 dc->writeback_rate_proportional = proportional;
67 dc->writeback_rate_derivative = derivative;
68 dc->writeback_rate_change = change;
69 dc->writeback_rate_target = target;
72 static void update_writeback_rate(struct work_struct *work)
74 struct cached_dev *dc = container_of(to_delayed_work(work),
76 writeback_rate_update);
78 down_read(&dc->writeback_lock);
80 if (atomic_read(&dc->has_dirty) &&
81 dc->writeback_percent)
82 __update_writeback_rate(dc);
84 up_read(&dc->writeback_lock);
86 schedule_delayed_work(&dc->writeback_rate_update,
87 dc->writeback_rate_update_seconds * HZ);
90 static unsigned writeback_delay(struct cached_dev *dc, unsigned sectors)
92 if (test_bit(BCACHE_DEV_DETACHING, &dc->disk.flags) ||
93 !dc->writeback_percent)
96 return bch_next_delay(&dc->writeback_rate, sectors);
101 struct cached_dev *dc;
105 static void dirty_init(struct keybuf_key *w)
107 struct dirty_io *io = w->private;
108 struct bio *bio = &io->bio;
111 if (!io->dc->writeback_percent)
112 bio_set_prio(bio, IOPRIO_PRIO_VALUE(IOPRIO_CLASS_IDLE, 0));
114 bio->bi_iter.bi_size = KEY_SIZE(&w->key) << 9;
115 bio->bi_max_vecs = DIV_ROUND_UP(KEY_SIZE(&w->key), PAGE_SECTORS);
117 bio->bi_io_vec = bio->bi_inline_vecs;
118 bch_bio_map(bio, NULL);
121 static void dirty_io_destructor(struct closure *cl)
123 struct dirty_io *io = container_of(cl, struct dirty_io, cl);
127 static void write_dirty_finish(struct closure *cl)
129 struct dirty_io *io = container_of(cl, struct dirty_io, cl);
130 struct keybuf_key *w = io->bio.bi_private;
131 struct cached_dev *dc = io->dc;
133 bio_free_pages(&io->bio);
135 /* This is kind of a dumb way of signalling errors. */
136 if (KEY_DIRTY(&w->key)) {
141 bch_keylist_init(&keys);
143 bkey_copy(keys.top, &w->key);
144 SET_KEY_DIRTY(keys.top, false);
145 bch_keylist_push(&keys);
147 for (i = 0; i < KEY_PTRS(&w->key); i++)
148 atomic_inc(&PTR_BUCKET(dc->disk.c, &w->key, i)->pin);
150 ret = bch_btree_insert(dc->disk.c, &keys, NULL, &w->key);
153 trace_bcache_writeback_collision(&w->key);
156 ? &dc->disk.c->writeback_keys_failed
157 : &dc->disk.c->writeback_keys_done);
160 bch_keybuf_del(&dc->writeback_keys, w);
163 closure_return_with_destructor(cl, dirty_io_destructor);
166 static void dirty_endio(struct bio *bio)
168 struct keybuf_key *w = bio->bi_private;
169 struct dirty_io *io = w->private;
172 SET_KEY_DIRTY(&w->key, false);
174 closure_put(&io->cl);
177 static void write_dirty(struct closure *cl)
179 struct dirty_io *io = container_of(cl, struct dirty_io, cl);
180 struct keybuf_key *w = io->bio.bi_private;
183 bio_set_op_attrs(&io->bio, REQ_OP_WRITE, 0);
184 io->bio.bi_iter.bi_sector = KEY_START(&w->key);
185 io->bio.bi_bdev = io->dc->bdev;
186 io->bio.bi_end_io = dirty_endio;
188 closure_bio_submit(&io->bio, cl);
190 continue_at(cl, write_dirty_finish, io->dc->writeback_write_wq);
193 static void read_dirty_endio(struct bio *bio)
195 struct keybuf_key *w = bio->bi_private;
196 struct dirty_io *io = w->private;
198 bch_count_io_errors(PTR_CACHE(io->dc->disk.c, &w->key, 0),
199 bio->bi_error, "reading dirty data from cache");
204 static void read_dirty_submit(struct closure *cl)
206 struct dirty_io *io = container_of(cl, struct dirty_io, cl);
208 closure_bio_submit(&io->bio, cl);
210 continue_at(cl, write_dirty, io->dc->writeback_write_wq);
213 static void read_dirty(struct cached_dev *dc)
216 struct keybuf_key *w;
220 closure_init_stack(&cl);
223 * XXX: if we error, background writeback just spins. Should use some
227 while (!kthread_should_stop()) {
229 w = bch_keybuf_next(&dc->writeback_keys);
233 BUG_ON(ptr_stale(dc->disk.c, &w->key, 0));
235 if (KEY_START(&w->key) != dc->last_read ||
236 jiffies_to_msecs(delay) > 50)
237 while (!kthread_should_stop() && delay)
238 delay = schedule_timeout_interruptible(delay);
240 dc->last_read = KEY_OFFSET(&w->key);
242 io = kzalloc(sizeof(struct dirty_io) + sizeof(struct bio_vec)
243 * DIV_ROUND_UP(KEY_SIZE(&w->key), PAGE_SECTORS),
252 bio_set_op_attrs(&io->bio, REQ_OP_READ, 0);
253 io->bio.bi_iter.bi_sector = PTR_OFFSET(&w->key, 0);
254 io->bio.bi_bdev = PTR_CACHE(dc->disk.c,
256 io->bio.bi_end_io = read_dirty_endio;
258 if (bio_alloc_pages(&io->bio, GFP_KERNEL))
261 trace_bcache_writeback(&w->key);
263 down(&dc->in_flight);
264 closure_call(&io->cl, read_dirty_submit, NULL, &cl);
266 delay = writeback_delay(dc, KEY_SIZE(&w->key));
273 bch_keybuf_del(&dc->writeback_keys, w);
277 * Wait for outstanding writeback IOs to finish (and keybuf slots to be
278 * freed) before refilling again
283 /* Scan for dirty data */
285 void bcache_dev_sectors_dirty_add(struct cache_set *c, unsigned inode,
286 uint64_t offset, int nr_sectors)
288 struct bcache_device *d = c->devices[inode];
289 unsigned stripe_offset, stripe, sectors_dirty;
294 stripe = offset_to_stripe(d, offset);
295 stripe_offset = offset & (d->stripe_size - 1);
298 int s = min_t(unsigned, abs(nr_sectors),
299 d->stripe_size - stripe_offset);
304 if (stripe >= d->nr_stripes)
307 sectors_dirty = atomic_add_return(s,
308 d->stripe_sectors_dirty + stripe);
309 if (sectors_dirty == d->stripe_size)
310 set_bit(stripe, d->full_dirty_stripes);
312 clear_bit(stripe, d->full_dirty_stripes);
320 static bool dirty_pred(struct keybuf *buf, struct bkey *k)
322 struct cached_dev *dc = container_of(buf, struct cached_dev, writeback_keys);
324 BUG_ON(KEY_INODE(k) != dc->disk.id);
329 static void refill_full_stripes(struct cached_dev *dc)
331 struct keybuf *buf = &dc->writeback_keys;
332 unsigned start_stripe, stripe, next_stripe;
333 bool wrapped = false;
335 stripe = offset_to_stripe(&dc->disk, KEY_OFFSET(&buf->last_scanned));
337 if (stripe >= dc->disk.nr_stripes)
340 start_stripe = stripe;
343 stripe = find_next_bit(dc->disk.full_dirty_stripes,
344 dc->disk.nr_stripes, stripe);
346 if (stripe == dc->disk.nr_stripes)
349 next_stripe = find_next_zero_bit(dc->disk.full_dirty_stripes,
350 dc->disk.nr_stripes, stripe);
352 buf->last_scanned = KEY(dc->disk.id,
353 stripe * dc->disk.stripe_size, 0);
355 bch_refill_keybuf(dc->disk.c, buf,
357 next_stripe * dc->disk.stripe_size, 0),
360 if (array_freelist_empty(&buf->freelist))
363 stripe = next_stripe;
365 if (wrapped && stripe > start_stripe)
368 if (stripe == dc->disk.nr_stripes) {
376 * Returns true if we scanned the entire disk
378 static bool refill_dirty(struct cached_dev *dc)
380 struct keybuf *buf = &dc->writeback_keys;
381 struct bkey start = KEY(dc->disk.id, 0, 0);
382 struct bkey end = KEY(dc->disk.id, MAX_KEY_OFFSET, 0);
383 struct bkey start_pos;
386 * make sure keybuf pos is inside the range for this disk - at bringup
387 * we might not be attached yet so this disk's inode nr isn't
390 if (bkey_cmp(&buf->last_scanned, &start) < 0 ||
391 bkey_cmp(&buf->last_scanned, &end) > 0)
392 buf->last_scanned = start;
394 if (dc->partial_stripes_expensive) {
395 refill_full_stripes(dc);
396 if (array_freelist_empty(&buf->freelist))
400 start_pos = buf->last_scanned;
401 bch_refill_keybuf(dc->disk.c, buf, &end, dirty_pred);
403 if (bkey_cmp(&buf->last_scanned, &end) < 0)
407 * If we get to the end start scanning again from the beginning, and
408 * only scan up to where we initially started scanning from:
410 buf->last_scanned = start;
411 bch_refill_keybuf(dc->disk.c, buf, &start_pos, dirty_pred);
413 return bkey_cmp(&buf->last_scanned, &start_pos) >= 0;
416 static int bch_writeback_thread(void *arg)
418 struct cached_dev *dc = arg;
419 bool searched_full_index;
421 while (!kthread_should_stop()) {
422 down_write(&dc->writeback_lock);
423 set_current_state(TASK_INTERRUPTIBLE);
425 * If the bache device is detaching, skip here and continue
426 * to perform writeback. Otherwise, if no dirty data on cache,
427 * or there is dirty data on cache but writeback is disabled,
428 * the writeback thread should sleep here and wait for others
431 if (!test_bit(BCACHE_DEV_DETACHING, &dc->disk.flags) &&
432 (!atomic_read(&dc->has_dirty) || !dc->writeback_running)) {
433 up_write(&dc->writeback_lock);
435 if (kthread_should_stop()) {
436 set_current_state(TASK_RUNNING);
443 set_current_state(TASK_RUNNING);
445 searched_full_index = refill_dirty(dc);
447 if (searched_full_index &&
448 RB_EMPTY_ROOT(&dc->writeback_keys.keys)) {
449 atomic_set(&dc->has_dirty, 0);
451 SET_BDEV_STATE(&dc->sb, BDEV_STATE_CLEAN);
452 bch_write_bdev_super(dc, NULL);
454 * If bcache device is detaching via sysfs interface,
455 * writeback thread should stop after there is no dirty
456 * data on cache. BCACHE_DEV_DETACHING flag is set in
457 * bch_cached_dev_detach().
459 if (test_bit(BCACHE_DEV_DETACHING, &dc->disk.flags)) {
460 up_write(&dc->writeback_lock);
465 up_write(&dc->writeback_lock);
467 bch_ratelimit_reset(&dc->writeback_rate);
470 if (searched_full_index) {
471 unsigned delay = dc->writeback_delay * HZ;
474 !kthread_should_stop() &&
475 !test_bit(BCACHE_DEV_DETACHING, &dc->disk.flags))
476 delay = schedule_timeout_interruptible(delay);
485 struct sectors_dirty_init {
490 static int sectors_dirty_init_fn(struct btree_op *_op, struct btree *b,
493 struct sectors_dirty_init *op = container_of(_op,
494 struct sectors_dirty_init, op);
495 if (KEY_INODE(k) > op->inode)
499 bcache_dev_sectors_dirty_add(b->c, KEY_INODE(k),
500 KEY_START(k), KEY_SIZE(k));
505 void bch_sectors_dirty_init(struct bcache_device *d)
507 struct sectors_dirty_init op;
509 bch_btree_op_init(&op.op, -1);
512 bch_btree_map_keys(&op.op, d->c, &KEY(op.inode, 0, 0),
513 sectors_dirty_init_fn, 0);
515 d->sectors_dirty_last = bcache_dev_sectors_dirty(d);
518 void bch_cached_dev_writeback_init(struct cached_dev *dc)
520 sema_init(&dc->in_flight, 64);
521 init_rwsem(&dc->writeback_lock);
522 bch_keybuf_init(&dc->writeback_keys);
524 dc->writeback_metadata = true;
525 dc->writeback_running = true;
526 dc->writeback_percent = 10;
527 dc->writeback_delay = 30;
528 dc->writeback_rate.rate = 1024;
530 dc->writeback_rate_update_seconds = 5;
531 dc->writeback_rate_d_term = 30;
532 dc->writeback_rate_p_term_inverse = 6000;
534 INIT_DELAYED_WORK(&dc->writeback_rate_update, update_writeback_rate);
537 int bch_cached_dev_writeback_start(struct cached_dev *dc)
539 dc->writeback_write_wq = alloc_workqueue("bcache_writeback_wq",
541 if (!dc->writeback_write_wq)
544 dc->writeback_thread = kthread_create(bch_writeback_thread, dc,
546 if (IS_ERR(dc->writeback_thread))
547 return PTR_ERR(dc->writeback_thread);
549 schedule_delayed_work(&dc->writeback_rate_update,
550 dc->writeback_rate_update_seconds * HZ);
552 bch_writeback_queue(dc);