xref: /openbmc/linux/drivers/md/bcache/writeback.c (revision 930beb5a)
1 /*
2  * background writeback - scan btree for dirty data and write it to the backing
3  * device
4  *
5  * Copyright 2010, 2011 Kent Overstreet <kent.overstreet@gmail.com>
6  * Copyright 2012 Google, Inc.
7  */
8 
9 #include "bcache.h"
10 #include "btree.h"
11 #include "debug.h"
12 #include "writeback.h"
13 
14 #include <linux/delay.h>
15 #include <linux/freezer.h>
16 #include <linux/kthread.h>
17 #include <trace/events/bcache.h>
18 
19 /* Rate limiting */
20 
21 static void __update_writeback_rate(struct cached_dev *dc)
22 {
23 	struct cache_set *c = dc->disk.c;
24 	uint64_t cache_sectors = c->nbuckets * c->sb.bucket_size;
25 	uint64_t cache_dirty_target =
26 		div_u64(cache_sectors * dc->writeback_percent, 100);
27 
28 	int64_t target = div64_u64(cache_dirty_target * bdev_sectors(dc->bdev),
29 				   c->cached_dev_sectors);
30 
31 	/* PD controller */
32 
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;
36 	int64_t change;
37 
38 	dc->disk.sectors_dirty_last = dirty;
39 
40 	/* Scale to sectors per second */
41 
42 	proportional *= dc->writeback_rate_update_seconds;
43 	proportional = div_s64(proportional, dc->writeback_rate_p_term_inverse);
44 
45 	derivative = div_s64(derivative, dc->writeback_rate_update_seconds);
46 
47 	derivative = ewma_add(dc->disk.sectors_dirty_derivative, derivative,
48 			      (dc->writeback_rate_d_term /
49 			       dc->writeback_rate_update_seconds) ?: 1, 0);
50 
51 	derivative *= dc->writeback_rate_d_term;
52 	derivative = div_s64(derivative, dc->writeback_rate_p_term_inverse);
53 
54 	change = proportional + derivative;
55 
56 	/* Don't increase writeback rate if the device isn't keeping up */
57 	if (change > 0 &&
58 	    time_after64(local_clock(),
59 			 dc->writeback_rate.next + NSEC_PER_MSEC))
60 		change = 0;
61 
62 	dc->writeback_rate.rate =
63 		clamp_t(int64_t, (int64_t) dc->writeback_rate.rate + change,
64 			1, NSEC_PER_MSEC);
65 
66 	dc->writeback_rate_proportional = proportional;
67 	dc->writeback_rate_derivative = derivative;
68 	dc->writeback_rate_change = change;
69 	dc->writeback_rate_target = target;
70 }
71 
72 static void update_writeback_rate(struct work_struct *work)
73 {
74 	struct cached_dev *dc = container_of(to_delayed_work(work),
75 					     struct cached_dev,
76 					     writeback_rate_update);
77 
78 	down_read(&dc->writeback_lock);
79 
80 	if (atomic_read(&dc->has_dirty) &&
81 	    dc->writeback_percent)
82 		__update_writeback_rate(dc);
83 
84 	up_read(&dc->writeback_lock);
85 
86 	schedule_delayed_work(&dc->writeback_rate_update,
87 			      dc->writeback_rate_update_seconds * HZ);
88 }
89 
90 static unsigned writeback_delay(struct cached_dev *dc, unsigned sectors)
91 {
92 	if (test_bit(BCACHE_DEV_DETACHING, &dc->disk.flags) ||
93 	    !dc->writeback_percent)
94 		return 0;
95 
96 	return bch_next_delay(&dc->writeback_rate, sectors);
97 }
98 
99 struct dirty_io {
100 	struct closure		cl;
101 	struct cached_dev	*dc;
102 	struct bio		bio;
103 };
104 
105 static void dirty_init(struct keybuf_key *w)
106 {
107 	struct dirty_io *io = w->private;
108 	struct bio *bio = &io->bio;
109 
110 	bio_init(bio);
111 	if (!io->dc->writeback_percent)
112 		bio_set_prio(bio, IOPRIO_PRIO_VALUE(IOPRIO_CLASS_IDLE, 0));
113 
114 	bio->bi_size		= KEY_SIZE(&w->key) << 9;
115 	bio->bi_max_vecs	= DIV_ROUND_UP(KEY_SIZE(&w->key), PAGE_SECTORS);
116 	bio->bi_private		= w;
117 	bio->bi_io_vec		= bio->bi_inline_vecs;
118 	bch_bio_map(bio, NULL);
119 }
120 
121 static void dirty_io_destructor(struct closure *cl)
122 {
123 	struct dirty_io *io = container_of(cl, struct dirty_io, cl);
124 	kfree(io);
125 }
126 
127 static void write_dirty_finish(struct closure *cl)
128 {
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;
132 	struct bio_vec *bv;
133 	int i;
134 
135 	bio_for_each_segment_all(bv, &io->bio, i)
136 		__free_page(bv->bv_page);
137 
138 	/* This is kind of a dumb way of signalling errors. */
139 	if (KEY_DIRTY(&w->key)) {
140 		int ret;
141 		unsigned i;
142 		struct keylist keys;
143 
144 		bch_keylist_init(&keys);
145 
146 		bkey_copy(keys.top, &w->key);
147 		SET_KEY_DIRTY(keys.top, false);
148 		bch_keylist_push(&keys);
149 
150 		for (i = 0; i < KEY_PTRS(&w->key); i++)
151 			atomic_inc(&PTR_BUCKET(dc->disk.c, &w->key, i)->pin);
152 
153 		ret = bch_btree_insert(dc->disk.c, &keys, NULL, &w->key);
154 
155 		if (ret)
156 			trace_bcache_writeback_collision(&w->key);
157 
158 		atomic_long_inc(ret
159 				? &dc->disk.c->writeback_keys_failed
160 				: &dc->disk.c->writeback_keys_done);
161 	}
162 
163 	bch_keybuf_del(&dc->writeback_keys, w);
164 	up(&dc->in_flight);
165 
166 	closure_return_with_destructor(cl, dirty_io_destructor);
167 }
168 
169 static void dirty_endio(struct bio *bio, int error)
170 {
171 	struct keybuf_key *w = bio->bi_private;
172 	struct dirty_io *io = w->private;
173 
174 	if (error)
175 		SET_KEY_DIRTY(&w->key, false);
176 
177 	closure_put(&io->cl);
178 }
179 
180 static void write_dirty(struct closure *cl)
181 {
182 	struct dirty_io *io = container_of(cl, struct dirty_io, cl);
183 	struct keybuf_key *w = io->bio.bi_private;
184 
185 	dirty_init(w);
186 	io->bio.bi_rw		= WRITE;
187 	io->bio.bi_sector	= KEY_START(&w->key);
188 	io->bio.bi_bdev		= io->dc->bdev;
189 	io->bio.bi_end_io	= dirty_endio;
190 
191 	closure_bio_submit(&io->bio, cl, &io->dc->disk);
192 
193 	continue_at(cl, write_dirty_finish, system_wq);
194 }
195 
196 static void read_dirty_endio(struct bio *bio, int error)
197 {
198 	struct keybuf_key *w = bio->bi_private;
199 	struct dirty_io *io = w->private;
200 
201 	bch_count_io_errors(PTR_CACHE(io->dc->disk.c, &w->key, 0),
202 			    error, "reading dirty data from cache");
203 
204 	dirty_endio(bio, error);
205 }
206 
207 static void read_dirty_submit(struct closure *cl)
208 {
209 	struct dirty_io *io = container_of(cl, struct dirty_io, cl);
210 
211 	closure_bio_submit(&io->bio, cl, &io->dc->disk);
212 
213 	continue_at(cl, write_dirty, system_wq);
214 }
215 
216 static void read_dirty(struct cached_dev *dc)
217 {
218 	unsigned delay = 0;
219 	struct keybuf_key *w;
220 	struct dirty_io *io;
221 	struct closure cl;
222 
223 	closure_init_stack(&cl);
224 
225 	/*
226 	 * XXX: if we error, background writeback just spins. Should use some
227 	 * mempools.
228 	 */
229 
230 	while (!kthread_should_stop()) {
231 		try_to_freeze();
232 
233 		w = bch_keybuf_next(&dc->writeback_keys);
234 		if (!w)
235 			break;
236 
237 		BUG_ON(ptr_stale(dc->disk.c, &w->key, 0));
238 
239 		if (KEY_START(&w->key) != dc->last_read ||
240 		    jiffies_to_msecs(delay) > 50)
241 			while (!kthread_should_stop() && delay)
242 				delay = schedule_timeout_uninterruptible(delay);
243 
244 		dc->last_read	= KEY_OFFSET(&w->key);
245 
246 		io = kzalloc(sizeof(struct dirty_io) + sizeof(struct bio_vec)
247 			     * DIV_ROUND_UP(KEY_SIZE(&w->key), PAGE_SECTORS),
248 			     GFP_KERNEL);
249 		if (!io)
250 			goto err;
251 
252 		w->private	= io;
253 		io->dc		= dc;
254 
255 		dirty_init(w);
256 		io->bio.bi_sector	= PTR_OFFSET(&w->key, 0);
257 		io->bio.bi_bdev		= PTR_CACHE(dc->disk.c,
258 						    &w->key, 0)->bdev;
259 		io->bio.bi_rw		= READ;
260 		io->bio.bi_end_io	= read_dirty_endio;
261 
262 		if (bio_alloc_pages(&io->bio, GFP_KERNEL))
263 			goto err_free;
264 
265 		trace_bcache_writeback(&w->key);
266 
267 		down(&dc->in_flight);
268 		closure_call(&io->cl, read_dirty_submit, NULL, &cl);
269 
270 		delay = writeback_delay(dc, KEY_SIZE(&w->key));
271 	}
272 
273 	if (0) {
274 err_free:
275 		kfree(w->private);
276 err:
277 		bch_keybuf_del(&dc->writeback_keys, w);
278 	}
279 
280 	/*
281 	 * Wait for outstanding writeback IOs to finish (and keybuf slots to be
282 	 * freed) before refilling again
283 	 */
284 	closure_sync(&cl);
285 }
286 
287 /* Scan for dirty data */
288 
289 void bcache_dev_sectors_dirty_add(struct cache_set *c, unsigned inode,
290 				  uint64_t offset, int nr_sectors)
291 {
292 	struct bcache_device *d = c->devices[inode];
293 	unsigned stripe_offset, stripe, sectors_dirty;
294 
295 	if (!d)
296 		return;
297 
298 	stripe = offset_to_stripe(d, offset);
299 	stripe_offset = offset & (d->stripe_size - 1);
300 
301 	while (nr_sectors) {
302 		int s = min_t(unsigned, abs(nr_sectors),
303 			      d->stripe_size - stripe_offset);
304 
305 		if (nr_sectors < 0)
306 			s = -s;
307 
308 		if (stripe >= d->nr_stripes)
309 			return;
310 
311 		sectors_dirty = atomic_add_return(s,
312 					d->stripe_sectors_dirty + stripe);
313 		if (sectors_dirty == d->stripe_size)
314 			set_bit(stripe, d->full_dirty_stripes);
315 		else
316 			clear_bit(stripe, d->full_dirty_stripes);
317 
318 		nr_sectors -= s;
319 		stripe_offset = 0;
320 		stripe++;
321 	}
322 }
323 
324 static bool dirty_pred(struct keybuf *buf, struct bkey *k)
325 {
326 	return KEY_DIRTY(k);
327 }
328 
329 static void refill_full_stripes(struct cached_dev *dc)
330 {
331 	struct keybuf *buf = &dc->writeback_keys;
332 	unsigned start_stripe, stripe, next_stripe;
333 	bool wrapped = false;
334 
335 	stripe = offset_to_stripe(&dc->disk, KEY_OFFSET(&buf->last_scanned));
336 
337 	if (stripe >= dc->disk.nr_stripes)
338 		stripe = 0;
339 
340 	start_stripe = stripe;
341 
342 	while (1) {
343 		stripe = find_next_bit(dc->disk.full_dirty_stripes,
344 				       dc->disk.nr_stripes, stripe);
345 
346 		if (stripe == dc->disk.nr_stripes)
347 			goto next;
348 
349 		next_stripe = find_next_zero_bit(dc->disk.full_dirty_stripes,
350 						 dc->disk.nr_stripes, stripe);
351 
352 		buf->last_scanned = KEY(dc->disk.id,
353 					stripe * dc->disk.stripe_size, 0);
354 
355 		bch_refill_keybuf(dc->disk.c, buf,
356 				  &KEY(dc->disk.id,
357 				       next_stripe * dc->disk.stripe_size, 0),
358 				  dirty_pred);
359 
360 		if (array_freelist_empty(&buf->freelist))
361 			return;
362 
363 		stripe = next_stripe;
364 next:
365 		if (wrapped && stripe > start_stripe)
366 			return;
367 
368 		if (stripe == dc->disk.nr_stripes) {
369 			stripe = 0;
370 			wrapped = true;
371 		}
372 	}
373 }
374 
375 static bool refill_dirty(struct cached_dev *dc)
376 {
377 	struct keybuf *buf = &dc->writeback_keys;
378 	struct bkey end = KEY(dc->disk.id, MAX_KEY_OFFSET, 0);
379 	bool searched_from_start = false;
380 
381 	if (dc->partial_stripes_expensive) {
382 		refill_full_stripes(dc);
383 		if (array_freelist_empty(&buf->freelist))
384 			return false;
385 	}
386 
387 	if (bkey_cmp(&buf->last_scanned, &end) >= 0) {
388 		buf->last_scanned = KEY(dc->disk.id, 0, 0);
389 		searched_from_start = true;
390 	}
391 
392 	bch_refill_keybuf(dc->disk.c, buf, &end, dirty_pred);
393 
394 	return bkey_cmp(&buf->last_scanned, &end) >= 0 && searched_from_start;
395 }
396 
397 static int bch_writeback_thread(void *arg)
398 {
399 	struct cached_dev *dc = arg;
400 	bool searched_full_index;
401 
402 	while (!kthread_should_stop()) {
403 		down_write(&dc->writeback_lock);
404 		if (!atomic_read(&dc->has_dirty) ||
405 		    (!test_bit(BCACHE_DEV_DETACHING, &dc->disk.flags) &&
406 		     !dc->writeback_running)) {
407 			up_write(&dc->writeback_lock);
408 			set_current_state(TASK_INTERRUPTIBLE);
409 
410 			if (kthread_should_stop())
411 				return 0;
412 
413 			try_to_freeze();
414 			schedule();
415 			continue;
416 		}
417 
418 		searched_full_index = refill_dirty(dc);
419 
420 		if (searched_full_index &&
421 		    RB_EMPTY_ROOT(&dc->writeback_keys.keys)) {
422 			atomic_set(&dc->has_dirty, 0);
423 			cached_dev_put(dc);
424 			SET_BDEV_STATE(&dc->sb, BDEV_STATE_CLEAN);
425 			bch_write_bdev_super(dc, NULL);
426 		}
427 
428 		up_write(&dc->writeback_lock);
429 
430 		bch_ratelimit_reset(&dc->writeback_rate);
431 		read_dirty(dc);
432 
433 		if (searched_full_index) {
434 			unsigned delay = dc->writeback_delay * HZ;
435 
436 			while (delay &&
437 			       !kthread_should_stop() &&
438 			       !test_bit(BCACHE_DEV_DETACHING, &dc->disk.flags))
439 				delay = schedule_timeout_uninterruptible(delay);
440 		}
441 	}
442 
443 	return 0;
444 }
445 
446 /* Init */
447 
448 struct sectors_dirty_init {
449 	struct btree_op	op;
450 	unsigned	inode;
451 };
452 
453 static int sectors_dirty_init_fn(struct btree_op *_op, struct btree *b,
454 				 struct bkey *k)
455 {
456 	struct sectors_dirty_init *op = container_of(_op,
457 						struct sectors_dirty_init, op);
458 	if (KEY_INODE(k) > op->inode)
459 		return MAP_DONE;
460 
461 	if (KEY_DIRTY(k))
462 		bcache_dev_sectors_dirty_add(b->c, KEY_INODE(k),
463 					     KEY_START(k), KEY_SIZE(k));
464 
465 	return MAP_CONTINUE;
466 }
467 
468 void bch_sectors_dirty_init(struct cached_dev *dc)
469 {
470 	struct sectors_dirty_init op;
471 
472 	bch_btree_op_init(&op.op, -1);
473 	op.inode = dc->disk.id;
474 
475 	bch_btree_map_keys(&op.op, dc->disk.c, &KEY(op.inode, 0, 0),
476 			   sectors_dirty_init_fn, 0);
477 
478 	dc->disk.sectors_dirty_last = bcache_dev_sectors_dirty(&dc->disk);
479 }
480 
481 int bch_cached_dev_writeback_init(struct cached_dev *dc)
482 {
483 	sema_init(&dc->in_flight, 64);
484 	init_rwsem(&dc->writeback_lock);
485 	bch_keybuf_init(&dc->writeback_keys);
486 
487 	dc->writeback_metadata		= true;
488 	dc->writeback_running		= true;
489 	dc->writeback_percent		= 10;
490 	dc->writeback_delay		= 30;
491 	dc->writeback_rate.rate		= 1024;
492 
493 	dc->writeback_rate_update_seconds = 5;
494 	dc->writeback_rate_d_term	= 30;
495 	dc->writeback_rate_p_term_inverse = 6000;
496 
497 	dc->writeback_thread = kthread_create(bch_writeback_thread, dc,
498 					      "bcache_writeback");
499 	if (IS_ERR(dc->writeback_thread))
500 		return PTR_ERR(dc->writeback_thread);
501 
502 	INIT_DELAYED_WORK(&dc->writeback_rate_update, update_writeback_rate);
503 	schedule_delayed_work(&dc->writeback_rate_update,
504 			      dc->writeback_rate_update_seconds * HZ);
505 
506 	return 0;
507 }
508