xref: /openbmc/linux/drivers/md/bcache/super.c (revision b830f94f)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * bcache setup/teardown code, and some metadata io - read a superblock and
4  * figure out what to do with it.
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 "extents.h"
14 #include "request.h"
15 #include "writeback.h"
16 
17 #include <linux/blkdev.h>
18 #include <linux/buffer_head.h>
19 #include <linux/debugfs.h>
20 #include <linux/genhd.h>
21 #include <linux/idr.h>
22 #include <linux/kthread.h>
23 #include <linux/module.h>
24 #include <linux/random.h>
25 #include <linux/reboot.h>
26 #include <linux/sysfs.h>
27 
28 unsigned int bch_cutoff_writeback;
29 unsigned int bch_cutoff_writeback_sync;
30 
31 static const char bcache_magic[] = {
32 	0xc6, 0x85, 0x73, 0xf6, 0x4e, 0x1a, 0x45, 0xca,
33 	0x82, 0x65, 0xf5, 0x7f, 0x48, 0xba, 0x6d, 0x81
34 };
35 
36 static const char invalid_uuid[] = {
37 	0xa0, 0x3e, 0xf8, 0xed, 0x3e, 0xe1, 0xb8, 0x78,
38 	0xc8, 0x50, 0xfc, 0x5e, 0xcb, 0x16, 0xcd, 0x99
39 };
40 
41 static struct kobject *bcache_kobj;
42 struct mutex bch_register_lock;
43 bool bcache_is_reboot;
44 LIST_HEAD(bch_cache_sets);
45 static LIST_HEAD(uncached_devices);
46 
47 static int bcache_major;
48 static DEFINE_IDA(bcache_device_idx);
49 static wait_queue_head_t unregister_wait;
50 struct workqueue_struct *bcache_wq;
51 struct workqueue_struct *bch_journal_wq;
52 
53 
54 #define BTREE_MAX_PAGES		(256 * 1024 / PAGE_SIZE)
55 /* limitation of partitions number on single bcache device */
56 #define BCACHE_MINORS		128
57 /* limitation of bcache devices number on single system */
58 #define BCACHE_DEVICE_IDX_MAX	((1U << MINORBITS)/BCACHE_MINORS)
59 
60 /* Superblock */
61 
62 static const char *read_super(struct cache_sb *sb, struct block_device *bdev,
63 			      struct page **res)
64 {
65 	const char *err;
66 	struct cache_sb *s;
67 	struct buffer_head *bh = __bread(bdev, 1, SB_SIZE);
68 	unsigned int i;
69 
70 	if (!bh)
71 		return "IO error";
72 
73 	s = (struct cache_sb *) bh->b_data;
74 
75 	sb->offset		= le64_to_cpu(s->offset);
76 	sb->version		= le64_to_cpu(s->version);
77 
78 	memcpy(sb->magic,	s->magic, 16);
79 	memcpy(sb->uuid,	s->uuid, 16);
80 	memcpy(sb->set_uuid,	s->set_uuid, 16);
81 	memcpy(sb->label,	s->label, SB_LABEL_SIZE);
82 
83 	sb->flags		= le64_to_cpu(s->flags);
84 	sb->seq			= le64_to_cpu(s->seq);
85 	sb->last_mount		= le32_to_cpu(s->last_mount);
86 	sb->first_bucket	= le16_to_cpu(s->first_bucket);
87 	sb->keys		= le16_to_cpu(s->keys);
88 
89 	for (i = 0; i < SB_JOURNAL_BUCKETS; i++)
90 		sb->d[i] = le64_to_cpu(s->d[i]);
91 
92 	pr_debug("read sb version %llu, flags %llu, seq %llu, journal size %u",
93 		 sb->version, sb->flags, sb->seq, sb->keys);
94 
95 	err = "Not a bcache superblock";
96 	if (sb->offset != SB_SECTOR)
97 		goto err;
98 
99 	if (memcmp(sb->magic, bcache_magic, 16))
100 		goto err;
101 
102 	err = "Too many journal buckets";
103 	if (sb->keys > SB_JOURNAL_BUCKETS)
104 		goto err;
105 
106 	err = "Bad checksum";
107 	if (s->csum != csum_set(s))
108 		goto err;
109 
110 	err = "Bad UUID";
111 	if (bch_is_zero(sb->uuid, 16))
112 		goto err;
113 
114 	sb->block_size	= le16_to_cpu(s->block_size);
115 
116 	err = "Superblock block size smaller than device block size";
117 	if (sb->block_size << 9 < bdev_logical_block_size(bdev))
118 		goto err;
119 
120 	switch (sb->version) {
121 	case BCACHE_SB_VERSION_BDEV:
122 		sb->data_offset	= BDEV_DATA_START_DEFAULT;
123 		break;
124 	case BCACHE_SB_VERSION_BDEV_WITH_OFFSET:
125 		sb->data_offset	= le64_to_cpu(s->data_offset);
126 
127 		err = "Bad data offset";
128 		if (sb->data_offset < BDEV_DATA_START_DEFAULT)
129 			goto err;
130 
131 		break;
132 	case BCACHE_SB_VERSION_CDEV:
133 	case BCACHE_SB_VERSION_CDEV_WITH_UUID:
134 		sb->nbuckets	= le64_to_cpu(s->nbuckets);
135 		sb->bucket_size	= le16_to_cpu(s->bucket_size);
136 
137 		sb->nr_in_set	= le16_to_cpu(s->nr_in_set);
138 		sb->nr_this_dev	= le16_to_cpu(s->nr_this_dev);
139 
140 		err = "Too many buckets";
141 		if (sb->nbuckets > LONG_MAX)
142 			goto err;
143 
144 		err = "Not enough buckets";
145 		if (sb->nbuckets < 1 << 7)
146 			goto err;
147 
148 		err = "Bad block/bucket size";
149 		if (!is_power_of_2(sb->block_size) ||
150 		    sb->block_size > PAGE_SECTORS ||
151 		    !is_power_of_2(sb->bucket_size) ||
152 		    sb->bucket_size < PAGE_SECTORS)
153 			goto err;
154 
155 		err = "Invalid superblock: device too small";
156 		if (get_capacity(bdev->bd_disk) <
157 		    sb->bucket_size * sb->nbuckets)
158 			goto err;
159 
160 		err = "Bad UUID";
161 		if (bch_is_zero(sb->set_uuid, 16))
162 			goto err;
163 
164 		err = "Bad cache device number in set";
165 		if (!sb->nr_in_set ||
166 		    sb->nr_in_set <= sb->nr_this_dev ||
167 		    sb->nr_in_set > MAX_CACHES_PER_SET)
168 			goto err;
169 
170 		err = "Journal buckets not sequential";
171 		for (i = 0; i < sb->keys; i++)
172 			if (sb->d[i] != sb->first_bucket + i)
173 				goto err;
174 
175 		err = "Too many journal buckets";
176 		if (sb->first_bucket + sb->keys > sb->nbuckets)
177 			goto err;
178 
179 		err = "Invalid superblock: first bucket comes before end of super";
180 		if (sb->first_bucket * sb->bucket_size < 16)
181 			goto err;
182 
183 		break;
184 	default:
185 		err = "Unsupported superblock version";
186 		goto err;
187 	}
188 
189 	sb->last_mount = (u32)ktime_get_real_seconds();
190 	err = NULL;
191 
192 	get_page(bh->b_page);
193 	*res = bh->b_page;
194 err:
195 	put_bh(bh);
196 	return err;
197 }
198 
199 static void write_bdev_super_endio(struct bio *bio)
200 {
201 	struct cached_dev *dc = bio->bi_private;
202 
203 	if (bio->bi_status)
204 		bch_count_backing_io_errors(dc, bio);
205 
206 	closure_put(&dc->sb_write);
207 }
208 
209 static void __write_super(struct cache_sb *sb, struct bio *bio)
210 {
211 	struct cache_sb *out = page_address(bio_first_page_all(bio));
212 	unsigned int i;
213 
214 	bio->bi_iter.bi_sector	= SB_SECTOR;
215 	bio->bi_iter.bi_size	= SB_SIZE;
216 	bio_set_op_attrs(bio, REQ_OP_WRITE, REQ_SYNC|REQ_META);
217 	bch_bio_map(bio, NULL);
218 
219 	out->offset		= cpu_to_le64(sb->offset);
220 	out->version		= cpu_to_le64(sb->version);
221 
222 	memcpy(out->uuid,	sb->uuid, 16);
223 	memcpy(out->set_uuid,	sb->set_uuid, 16);
224 	memcpy(out->label,	sb->label, SB_LABEL_SIZE);
225 
226 	out->flags		= cpu_to_le64(sb->flags);
227 	out->seq		= cpu_to_le64(sb->seq);
228 
229 	out->last_mount		= cpu_to_le32(sb->last_mount);
230 	out->first_bucket	= cpu_to_le16(sb->first_bucket);
231 	out->keys		= cpu_to_le16(sb->keys);
232 
233 	for (i = 0; i < sb->keys; i++)
234 		out->d[i] = cpu_to_le64(sb->d[i]);
235 
236 	out->csum = csum_set(out);
237 
238 	pr_debug("ver %llu, flags %llu, seq %llu",
239 		 sb->version, sb->flags, sb->seq);
240 
241 	submit_bio(bio);
242 }
243 
244 static void bch_write_bdev_super_unlock(struct closure *cl)
245 {
246 	struct cached_dev *dc = container_of(cl, struct cached_dev, sb_write);
247 
248 	up(&dc->sb_write_mutex);
249 }
250 
251 void bch_write_bdev_super(struct cached_dev *dc, struct closure *parent)
252 {
253 	struct closure *cl = &dc->sb_write;
254 	struct bio *bio = &dc->sb_bio;
255 
256 	down(&dc->sb_write_mutex);
257 	closure_init(cl, parent);
258 
259 	bio_reset(bio);
260 	bio_set_dev(bio, dc->bdev);
261 	bio->bi_end_io	= write_bdev_super_endio;
262 	bio->bi_private = dc;
263 
264 	closure_get(cl);
265 	/* I/O request sent to backing device */
266 	__write_super(&dc->sb, bio);
267 
268 	closure_return_with_destructor(cl, bch_write_bdev_super_unlock);
269 }
270 
271 static void write_super_endio(struct bio *bio)
272 {
273 	struct cache *ca = bio->bi_private;
274 
275 	/* is_read = 0 */
276 	bch_count_io_errors(ca, bio->bi_status, 0,
277 			    "writing superblock");
278 	closure_put(&ca->set->sb_write);
279 }
280 
281 static void bcache_write_super_unlock(struct closure *cl)
282 {
283 	struct cache_set *c = container_of(cl, struct cache_set, sb_write);
284 
285 	up(&c->sb_write_mutex);
286 }
287 
288 void bcache_write_super(struct cache_set *c)
289 {
290 	struct closure *cl = &c->sb_write;
291 	struct cache *ca;
292 	unsigned int i;
293 
294 	down(&c->sb_write_mutex);
295 	closure_init(cl, &c->cl);
296 
297 	c->sb.seq++;
298 
299 	for_each_cache(ca, c, i) {
300 		struct bio *bio = &ca->sb_bio;
301 
302 		ca->sb.version		= BCACHE_SB_VERSION_CDEV_WITH_UUID;
303 		ca->sb.seq		= c->sb.seq;
304 		ca->sb.last_mount	= c->sb.last_mount;
305 
306 		SET_CACHE_SYNC(&ca->sb, CACHE_SYNC(&c->sb));
307 
308 		bio_reset(bio);
309 		bio_set_dev(bio, ca->bdev);
310 		bio->bi_end_io	= write_super_endio;
311 		bio->bi_private = ca;
312 
313 		closure_get(cl);
314 		__write_super(&ca->sb, bio);
315 	}
316 
317 	closure_return_with_destructor(cl, bcache_write_super_unlock);
318 }
319 
320 /* UUID io */
321 
322 static void uuid_endio(struct bio *bio)
323 {
324 	struct closure *cl = bio->bi_private;
325 	struct cache_set *c = container_of(cl, struct cache_set, uuid_write);
326 
327 	cache_set_err_on(bio->bi_status, c, "accessing uuids");
328 	bch_bbio_free(bio, c);
329 	closure_put(cl);
330 }
331 
332 static void uuid_io_unlock(struct closure *cl)
333 {
334 	struct cache_set *c = container_of(cl, struct cache_set, uuid_write);
335 
336 	up(&c->uuid_write_mutex);
337 }
338 
339 static void uuid_io(struct cache_set *c, int op, unsigned long op_flags,
340 		    struct bkey *k, struct closure *parent)
341 {
342 	struct closure *cl = &c->uuid_write;
343 	struct uuid_entry *u;
344 	unsigned int i;
345 	char buf[80];
346 
347 	BUG_ON(!parent);
348 	down(&c->uuid_write_mutex);
349 	closure_init(cl, parent);
350 
351 	for (i = 0; i < KEY_PTRS(k); i++) {
352 		struct bio *bio = bch_bbio_alloc(c);
353 
354 		bio->bi_opf = REQ_SYNC | REQ_META | op_flags;
355 		bio->bi_iter.bi_size = KEY_SIZE(k) << 9;
356 
357 		bio->bi_end_io	= uuid_endio;
358 		bio->bi_private = cl;
359 		bio_set_op_attrs(bio, op, REQ_SYNC|REQ_META|op_flags);
360 		bch_bio_map(bio, c->uuids);
361 
362 		bch_submit_bbio(bio, c, k, i);
363 
364 		if (op != REQ_OP_WRITE)
365 			break;
366 	}
367 
368 	bch_extent_to_text(buf, sizeof(buf), k);
369 	pr_debug("%s UUIDs at %s", op == REQ_OP_WRITE ? "wrote" : "read", buf);
370 
371 	for (u = c->uuids; u < c->uuids + c->nr_uuids; u++)
372 		if (!bch_is_zero(u->uuid, 16))
373 			pr_debug("Slot %zi: %pU: %s: 1st: %u last: %u inv: %u",
374 				 u - c->uuids, u->uuid, u->label,
375 				 u->first_reg, u->last_reg, u->invalidated);
376 
377 	closure_return_with_destructor(cl, uuid_io_unlock);
378 }
379 
380 static char *uuid_read(struct cache_set *c, struct jset *j, struct closure *cl)
381 {
382 	struct bkey *k = &j->uuid_bucket;
383 
384 	if (__bch_btree_ptr_invalid(c, k))
385 		return "bad uuid pointer";
386 
387 	bkey_copy(&c->uuid_bucket, k);
388 	uuid_io(c, REQ_OP_READ, 0, k, cl);
389 
390 	if (j->version < BCACHE_JSET_VERSION_UUIDv1) {
391 		struct uuid_entry_v0	*u0 = (void *) c->uuids;
392 		struct uuid_entry	*u1 = (void *) c->uuids;
393 		int i;
394 
395 		closure_sync(cl);
396 
397 		/*
398 		 * Since the new uuid entry is bigger than the old, we have to
399 		 * convert starting at the highest memory address and work down
400 		 * in order to do it in place
401 		 */
402 
403 		for (i = c->nr_uuids - 1;
404 		     i >= 0;
405 		     --i) {
406 			memcpy(u1[i].uuid,	u0[i].uuid, 16);
407 			memcpy(u1[i].label,	u0[i].label, 32);
408 
409 			u1[i].first_reg		= u0[i].first_reg;
410 			u1[i].last_reg		= u0[i].last_reg;
411 			u1[i].invalidated	= u0[i].invalidated;
412 
413 			u1[i].flags	= 0;
414 			u1[i].sectors	= 0;
415 		}
416 	}
417 
418 	return NULL;
419 }
420 
421 static int __uuid_write(struct cache_set *c)
422 {
423 	BKEY_PADDED(key) k;
424 	struct closure cl;
425 	struct cache *ca;
426 
427 	closure_init_stack(&cl);
428 	lockdep_assert_held(&bch_register_lock);
429 
430 	if (bch_bucket_alloc_set(c, RESERVE_BTREE, &k.key, 1, true))
431 		return 1;
432 
433 	SET_KEY_SIZE(&k.key, c->sb.bucket_size);
434 	uuid_io(c, REQ_OP_WRITE, 0, &k.key, &cl);
435 	closure_sync(&cl);
436 
437 	/* Only one bucket used for uuid write */
438 	ca = PTR_CACHE(c, &k.key, 0);
439 	atomic_long_add(ca->sb.bucket_size, &ca->meta_sectors_written);
440 
441 	bkey_copy(&c->uuid_bucket, &k.key);
442 	bkey_put(c, &k.key);
443 	return 0;
444 }
445 
446 int bch_uuid_write(struct cache_set *c)
447 {
448 	int ret = __uuid_write(c);
449 
450 	if (!ret)
451 		bch_journal_meta(c, NULL);
452 
453 	return ret;
454 }
455 
456 static struct uuid_entry *uuid_find(struct cache_set *c, const char *uuid)
457 {
458 	struct uuid_entry *u;
459 
460 	for (u = c->uuids;
461 	     u < c->uuids + c->nr_uuids; u++)
462 		if (!memcmp(u->uuid, uuid, 16))
463 			return u;
464 
465 	return NULL;
466 }
467 
468 static struct uuid_entry *uuid_find_empty(struct cache_set *c)
469 {
470 	static const char zero_uuid[16] = "\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0";
471 
472 	return uuid_find(c, zero_uuid);
473 }
474 
475 /*
476  * Bucket priorities/gens:
477  *
478  * For each bucket, we store on disk its
479  *   8 bit gen
480  *  16 bit priority
481  *
482  * See alloc.c for an explanation of the gen. The priority is used to implement
483  * lru (and in the future other) cache replacement policies; for most purposes
484  * it's just an opaque integer.
485  *
486  * The gens and the priorities don't have a whole lot to do with each other, and
487  * it's actually the gens that must be written out at specific times - it's no
488  * big deal if the priorities don't get written, if we lose them we just reuse
489  * buckets in suboptimal order.
490  *
491  * On disk they're stored in a packed array, and in as many buckets are required
492  * to fit them all. The buckets we use to store them form a list; the journal
493  * header points to the first bucket, the first bucket points to the second
494  * bucket, et cetera.
495  *
496  * This code is used by the allocation code; periodically (whenever it runs out
497  * of buckets to allocate from) the allocation code will invalidate some
498  * buckets, but it can't use those buckets until their new gens are safely on
499  * disk.
500  */
501 
502 static void prio_endio(struct bio *bio)
503 {
504 	struct cache *ca = bio->bi_private;
505 
506 	cache_set_err_on(bio->bi_status, ca->set, "accessing priorities");
507 	bch_bbio_free(bio, ca->set);
508 	closure_put(&ca->prio);
509 }
510 
511 static void prio_io(struct cache *ca, uint64_t bucket, int op,
512 		    unsigned long op_flags)
513 {
514 	struct closure *cl = &ca->prio;
515 	struct bio *bio = bch_bbio_alloc(ca->set);
516 
517 	closure_init_stack(cl);
518 
519 	bio->bi_iter.bi_sector	= bucket * ca->sb.bucket_size;
520 	bio_set_dev(bio, ca->bdev);
521 	bio->bi_iter.bi_size	= bucket_bytes(ca);
522 
523 	bio->bi_end_io	= prio_endio;
524 	bio->bi_private = ca;
525 	bio_set_op_attrs(bio, op, REQ_SYNC|REQ_META|op_flags);
526 	bch_bio_map(bio, ca->disk_buckets);
527 
528 	closure_bio_submit(ca->set, bio, &ca->prio);
529 	closure_sync(cl);
530 }
531 
532 void bch_prio_write(struct cache *ca)
533 {
534 	int i;
535 	struct bucket *b;
536 	struct closure cl;
537 
538 	closure_init_stack(&cl);
539 
540 	lockdep_assert_held(&ca->set->bucket_lock);
541 
542 	ca->disk_buckets->seq++;
543 
544 	atomic_long_add(ca->sb.bucket_size * prio_buckets(ca),
545 			&ca->meta_sectors_written);
546 
547 	//pr_debug("free %zu, free_inc %zu, unused %zu", fifo_used(&ca->free),
548 	//	 fifo_used(&ca->free_inc), fifo_used(&ca->unused));
549 
550 	for (i = prio_buckets(ca) - 1; i >= 0; --i) {
551 		long bucket;
552 		struct prio_set *p = ca->disk_buckets;
553 		struct bucket_disk *d = p->data;
554 		struct bucket_disk *end = d + prios_per_bucket(ca);
555 
556 		for (b = ca->buckets + i * prios_per_bucket(ca);
557 		     b < ca->buckets + ca->sb.nbuckets && d < end;
558 		     b++, d++) {
559 			d->prio = cpu_to_le16(b->prio);
560 			d->gen = b->gen;
561 		}
562 
563 		p->next_bucket	= ca->prio_buckets[i + 1];
564 		p->magic	= pset_magic(&ca->sb);
565 		p->csum		= bch_crc64(&p->magic, bucket_bytes(ca) - 8);
566 
567 		bucket = bch_bucket_alloc(ca, RESERVE_PRIO, true);
568 		BUG_ON(bucket == -1);
569 
570 		mutex_unlock(&ca->set->bucket_lock);
571 		prio_io(ca, bucket, REQ_OP_WRITE, 0);
572 		mutex_lock(&ca->set->bucket_lock);
573 
574 		ca->prio_buckets[i] = bucket;
575 		atomic_dec_bug(&ca->buckets[bucket].pin);
576 	}
577 
578 	mutex_unlock(&ca->set->bucket_lock);
579 
580 	bch_journal_meta(ca->set, &cl);
581 	closure_sync(&cl);
582 
583 	mutex_lock(&ca->set->bucket_lock);
584 
585 	/*
586 	 * Don't want the old priorities to get garbage collected until after we
587 	 * finish writing the new ones, and they're journalled
588 	 */
589 	for (i = 0; i < prio_buckets(ca); i++) {
590 		if (ca->prio_last_buckets[i])
591 			__bch_bucket_free(ca,
592 				&ca->buckets[ca->prio_last_buckets[i]]);
593 
594 		ca->prio_last_buckets[i] = ca->prio_buckets[i];
595 	}
596 }
597 
598 static void prio_read(struct cache *ca, uint64_t bucket)
599 {
600 	struct prio_set *p = ca->disk_buckets;
601 	struct bucket_disk *d = p->data + prios_per_bucket(ca), *end = d;
602 	struct bucket *b;
603 	unsigned int bucket_nr = 0;
604 
605 	for (b = ca->buckets;
606 	     b < ca->buckets + ca->sb.nbuckets;
607 	     b++, d++) {
608 		if (d == end) {
609 			ca->prio_buckets[bucket_nr] = bucket;
610 			ca->prio_last_buckets[bucket_nr] = bucket;
611 			bucket_nr++;
612 
613 			prio_io(ca, bucket, REQ_OP_READ, 0);
614 
615 			if (p->csum !=
616 			    bch_crc64(&p->magic, bucket_bytes(ca) - 8))
617 				pr_warn("bad csum reading priorities");
618 
619 			if (p->magic != pset_magic(&ca->sb))
620 				pr_warn("bad magic reading priorities");
621 
622 			bucket = p->next_bucket;
623 			d = p->data;
624 		}
625 
626 		b->prio = le16_to_cpu(d->prio);
627 		b->gen = b->last_gc = d->gen;
628 	}
629 }
630 
631 /* Bcache device */
632 
633 static int open_dev(struct block_device *b, fmode_t mode)
634 {
635 	struct bcache_device *d = b->bd_disk->private_data;
636 
637 	if (test_bit(BCACHE_DEV_CLOSING, &d->flags))
638 		return -ENXIO;
639 
640 	closure_get(&d->cl);
641 	return 0;
642 }
643 
644 static void release_dev(struct gendisk *b, fmode_t mode)
645 {
646 	struct bcache_device *d = b->private_data;
647 
648 	closure_put(&d->cl);
649 }
650 
651 static int ioctl_dev(struct block_device *b, fmode_t mode,
652 		     unsigned int cmd, unsigned long arg)
653 {
654 	struct bcache_device *d = b->bd_disk->private_data;
655 
656 	return d->ioctl(d, mode, cmd, arg);
657 }
658 
659 static const struct block_device_operations bcache_ops = {
660 	.open		= open_dev,
661 	.release	= release_dev,
662 	.ioctl		= ioctl_dev,
663 	.owner		= THIS_MODULE,
664 };
665 
666 void bcache_device_stop(struct bcache_device *d)
667 {
668 	if (!test_and_set_bit(BCACHE_DEV_CLOSING, &d->flags))
669 		/*
670 		 * closure_fn set to
671 		 * - cached device: cached_dev_flush()
672 		 * - flash dev: flash_dev_flush()
673 		 */
674 		closure_queue(&d->cl);
675 }
676 
677 static void bcache_device_unlink(struct bcache_device *d)
678 {
679 	lockdep_assert_held(&bch_register_lock);
680 
681 	if (d->c && !test_and_set_bit(BCACHE_DEV_UNLINK_DONE, &d->flags)) {
682 		unsigned int i;
683 		struct cache *ca;
684 
685 		sysfs_remove_link(&d->c->kobj, d->name);
686 		sysfs_remove_link(&d->kobj, "cache");
687 
688 		for_each_cache(ca, d->c, i)
689 			bd_unlink_disk_holder(ca->bdev, d->disk);
690 	}
691 }
692 
693 static void bcache_device_link(struct bcache_device *d, struct cache_set *c,
694 			       const char *name)
695 {
696 	unsigned int i;
697 	struct cache *ca;
698 	int ret;
699 
700 	for_each_cache(ca, d->c, i)
701 		bd_link_disk_holder(ca->bdev, d->disk);
702 
703 	snprintf(d->name, BCACHEDEVNAME_SIZE,
704 		 "%s%u", name, d->id);
705 
706 	ret = sysfs_create_link(&d->kobj, &c->kobj, "cache");
707 	if (ret < 0)
708 		pr_err("Couldn't create device -> cache set symlink");
709 
710 	ret = sysfs_create_link(&c->kobj, &d->kobj, d->name);
711 	if (ret < 0)
712 		pr_err("Couldn't create cache set -> device symlink");
713 
714 	clear_bit(BCACHE_DEV_UNLINK_DONE, &d->flags);
715 }
716 
717 static void bcache_device_detach(struct bcache_device *d)
718 {
719 	lockdep_assert_held(&bch_register_lock);
720 
721 	atomic_dec(&d->c->attached_dev_nr);
722 
723 	if (test_bit(BCACHE_DEV_DETACHING, &d->flags)) {
724 		struct uuid_entry *u = d->c->uuids + d->id;
725 
726 		SET_UUID_FLASH_ONLY(u, 0);
727 		memcpy(u->uuid, invalid_uuid, 16);
728 		u->invalidated = cpu_to_le32((u32)ktime_get_real_seconds());
729 		bch_uuid_write(d->c);
730 	}
731 
732 	bcache_device_unlink(d);
733 
734 	d->c->devices[d->id] = NULL;
735 	closure_put(&d->c->caching);
736 	d->c = NULL;
737 }
738 
739 static void bcache_device_attach(struct bcache_device *d, struct cache_set *c,
740 				 unsigned int id)
741 {
742 	d->id = id;
743 	d->c = c;
744 	c->devices[id] = d;
745 
746 	if (id >= c->devices_max_used)
747 		c->devices_max_used = id + 1;
748 
749 	closure_get(&c->caching);
750 }
751 
752 static inline int first_minor_to_idx(int first_minor)
753 {
754 	return (first_minor/BCACHE_MINORS);
755 }
756 
757 static inline int idx_to_first_minor(int idx)
758 {
759 	return (idx * BCACHE_MINORS);
760 }
761 
762 static void bcache_device_free(struct bcache_device *d)
763 {
764 	lockdep_assert_held(&bch_register_lock);
765 
766 	pr_info("%s stopped", d->disk->disk_name);
767 
768 	if (d->c)
769 		bcache_device_detach(d);
770 	if (d->disk && d->disk->flags & GENHD_FL_UP)
771 		del_gendisk(d->disk);
772 	if (d->disk && d->disk->queue)
773 		blk_cleanup_queue(d->disk->queue);
774 	if (d->disk) {
775 		ida_simple_remove(&bcache_device_idx,
776 				  first_minor_to_idx(d->disk->first_minor));
777 		put_disk(d->disk);
778 	}
779 
780 	bioset_exit(&d->bio_split);
781 	kvfree(d->full_dirty_stripes);
782 	kvfree(d->stripe_sectors_dirty);
783 
784 	closure_debug_destroy(&d->cl);
785 }
786 
787 static int bcache_device_init(struct bcache_device *d, unsigned int block_size,
788 			      sector_t sectors)
789 {
790 	struct request_queue *q;
791 	const size_t max_stripes = min_t(size_t, INT_MAX,
792 					 SIZE_MAX / sizeof(atomic_t));
793 	size_t n;
794 	int idx;
795 
796 	if (!d->stripe_size)
797 		d->stripe_size = 1 << 31;
798 
799 	d->nr_stripes = DIV_ROUND_UP_ULL(sectors, d->stripe_size);
800 
801 	if (!d->nr_stripes || d->nr_stripes > max_stripes) {
802 		pr_err("nr_stripes too large or invalid: %u (start sector beyond end of disk?)",
803 			(unsigned int)d->nr_stripes);
804 		return -ENOMEM;
805 	}
806 
807 	n = d->nr_stripes * sizeof(atomic_t);
808 	d->stripe_sectors_dirty = kvzalloc(n, GFP_KERNEL);
809 	if (!d->stripe_sectors_dirty)
810 		return -ENOMEM;
811 
812 	n = BITS_TO_LONGS(d->nr_stripes) * sizeof(unsigned long);
813 	d->full_dirty_stripes = kvzalloc(n, GFP_KERNEL);
814 	if (!d->full_dirty_stripes)
815 		return -ENOMEM;
816 
817 	idx = ida_simple_get(&bcache_device_idx, 0,
818 				BCACHE_DEVICE_IDX_MAX, GFP_KERNEL);
819 	if (idx < 0)
820 		return idx;
821 
822 	if (bioset_init(&d->bio_split, 4, offsetof(struct bbio, bio),
823 			BIOSET_NEED_BVECS|BIOSET_NEED_RESCUER))
824 		goto err;
825 
826 	d->disk = alloc_disk(BCACHE_MINORS);
827 	if (!d->disk)
828 		goto err;
829 
830 	set_capacity(d->disk, sectors);
831 	snprintf(d->disk->disk_name, DISK_NAME_LEN, "bcache%i", idx);
832 
833 	d->disk->major		= bcache_major;
834 	d->disk->first_minor	= idx_to_first_minor(idx);
835 	d->disk->fops		= &bcache_ops;
836 	d->disk->private_data	= d;
837 
838 	q = blk_alloc_queue(GFP_KERNEL);
839 	if (!q)
840 		return -ENOMEM;
841 
842 	blk_queue_make_request(q, NULL);
843 	d->disk->queue			= q;
844 	q->queuedata			= d;
845 	q->backing_dev_info->congested_data = d;
846 	q->limits.max_hw_sectors	= UINT_MAX;
847 	q->limits.max_sectors		= UINT_MAX;
848 	q->limits.max_segment_size	= UINT_MAX;
849 	q->limits.max_segments		= BIO_MAX_PAGES;
850 	blk_queue_max_discard_sectors(q, UINT_MAX);
851 	q->limits.discard_granularity	= 512;
852 	q->limits.io_min		= block_size;
853 	q->limits.logical_block_size	= block_size;
854 	q->limits.physical_block_size	= block_size;
855 	blk_queue_flag_set(QUEUE_FLAG_NONROT, d->disk->queue);
856 	blk_queue_flag_clear(QUEUE_FLAG_ADD_RANDOM, d->disk->queue);
857 	blk_queue_flag_set(QUEUE_FLAG_DISCARD, d->disk->queue);
858 
859 	blk_queue_write_cache(q, true, true);
860 
861 	return 0;
862 
863 err:
864 	ida_simple_remove(&bcache_device_idx, idx);
865 	return -ENOMEM;
866 
867 }
868 
869 /* Cached device */
870 
871 static void calc_cached_dev_sectors(struct cache_set *c)
872 {
873 	uint64_t sectors = 0;
874 	struct cached_dev *dc;
875 
876 	list_for_each_entry(dc, &c->cached_devs, list)
877 		sectors += bdev_sectors(dc->bdev);
878 
879 	c->cached_dev_sectors = sectors;
880 }
881 
882 #define BACKING_DEV_OFFLINE_TIMEOUT 5
883 static int cached_dev_status_update(void *arg)
884 {
885 	struct cached_dev *dc = arg;
886 	struct request_queue *q;
887 
888 	/*
889 	 * If this delayed worker is stopping outside, directly quit here.
890 	 * dc->io_disable might be set via sysfs interface, so check it
891 	 * here too.
892 	 */
893 	while (!kthread_should_stop() && !dc->io_disable) {
894 		q = bdev_get_queue(dc->bdev);
895 		if (blk_queue_dying(q))
896 			dc->offline_seconds++;
897 		else
898 			dc->offline_seconds = 0;
899 
900 		if (dc->offline_seconds >= BACKING_DEV_OFFLINE_TIMEOUT) {
901 			pr_err("%s: device offline for %d seconds",
902 			       dc->backing_dev_name,
903 			       BACKING_DEV_OFFLINE_TIMEOUT);
904 			pr_err("%s: disable I/O request due to backing "
905 			       "device offline", dc->disk.name);
906 			dc->io_disable = true;
907 			/* let others know earlier that io_disable is true */
908 			smp_mb();
909 			bcache_device_stop(&dc->disk);
910 			break;
911 		}
912 		schedule_timeout_interruptible(HZ);
913 	}
914 
915 	wait_for_kthread_stop();
916 	return 0;
917 }
918 
919 
920 int bch_cached_dev_run(struct cached_dev *dc)
921 {
922 	struct bcache_device *d = &dc->disk;
923 	char *buf = kmemdup_nul(dc->sb.label, SB_LABEL_SIZE, GFP_KERNEL);
924 	char *env[] = {
925 		"DRIVER=bcache",
926 		kasprintf(GFP_KERNEL, "CACHED_UUID=%pU", dc->sb.uuid),
927 		kasprintf(GFP_KERNEL, "CACHED_LABEL=%s", buf ? : ""),
928 		NULL,
929 	};
930 
931 	if (dc->io_disable) {
932 		pr_err("I/O disabled on cached dev %s",
933 		       dc->backing_dev_name);
934 		return -EIO;
935 	}
936 
937 	if (atomic_xchg(&dc->running, 1)) {
938 		kfree(env[1]);
939 		kfree(env[2]);
940 		kfree(buf);
941 		pr_info("cached dev %s is running already",
942 		       dc->backing_dev_name);
943 		return -EBUSY;
944 	}
945 
946 	if (!d->c &&
947 	    BDEV_STATE(&dc->sb) != BDEV_STATE_NONE) {
948 		struct closure cl;
949 
950 		closure_init_stack(&cl);
951 
952 		SET_BDEV_STATE(&dc->sb, BDEV_STATE_STALE);
953 		bch_write_bdev_super(dc, &cl);
954 		closure_sync(&cl);
955 	}
956 
957 	add_disk(d->disk);
958 	bd_link_disk_holder(dc->bdev, dc->disk.disk);
959 	/*
960 	 * won't show up in the uevent file, use udevadm monitor -e instead
961 	 * only class / kset properties are persistent
962 	 */
963 	kobject_uevent_env(&disk_to_dev(d->disk)->kobj, KOBJ_CHANGE, env);
964 	kfree(env[1]);
965 	kfree(env[2]);
966 	kfree(buf);
967 
968 	if (sysfs_create_link(&d->kobj, &disk_to_dev(d->disk)->kobj, "dev") ||
969 	    sysfs_create_link(&disk_to_dev(d->disk)->kobj,
970 			      &d->kobj, "bcache")) {
971 		pr_err("Couldn't create bcache dev <-> disk sysfs symlinks");
972 		return -ENOMEM;
973 	}
974 
975 	dc->status_update_thread = kthread_run(cached_dev_status_update,
976 					       dc, "bcache_status_update");
977 	if (IS_ERR(dc->status_update_thread)) {
978 		pr_warn("failed to create bcache_status_update kthread, "
979 			"continue to run without monitoring backing "
980 			"device status");
981 	}
982 
983 	return 0;
984 }
985 
986 /*
987  * If BCACHE_DEV_RATE_DW_RUNNING is set, it means routine of the delayed
988  * work dc->writeback_rate_update is running. Wait until the routine
989  * quits (BCACHE_DEV_RATE_DW_RUNNING is clear), then continue to
990  * cancel it. If BCACHE_DEV_RATE_DW_RUNNING is not clear after time_out
991  * seconds, give up waiting here and continue to cancel it too.
992  */
993 static void cancel_writeback_rate_update_dwork(struct cached_dev *dc)
994 {
995 	int time_out = WRITEBACK_RATE_UPDATE_SECS_MAX * HZ;
996 
997 	do {
998 		if (!test_bit(BCACHE_DEV_RATE_DW_RUNNING,
999 			      &dc->disk.flags))
1000 			break;
1001 		time_out--;
1002 		schedule_timeout_interruptible(1);
1003 	} while (time_out > 0);
1004 
1005 	if (time_out == 0)
1006 		pr_warn("give up waiting for dc->writeback_write_update to quit");
1007 
1008 	cancel_delayed_work_sync(&dc->writeback_rate_update);
1009 }
1010 
1011 static void cached_dev_detach_finish(struct work_struct *w)
1012 {
1013 	struct cached_dev *dc = container_of(w, struct cached_dev, detach);
1014 	struct closure cl;
1015 
1016 	closure_init_stack(&cl);
1017 
1018 	BUG_ON(!test_bit(BCACHE_DEV_DETACHING, &dc->disk.flags));
1019 	BUG_ON(refcount_read(&dc->count));
1020 
1021 
1022 	if (test_and_clear_bit(BCACHE_DEV_WB_RUNNING, &dc->disk.flags))
1023 		cancel_writeback_rate_update_dwork(dc);
1024 
1025 	if (!IS_ERR_OR_NULL(dc->writeback_thread)) {
1026 		kthread_stop(dc->writeback_thread);
1027 		dc->writeback_thread = NULL;
1028 	}
1029 
1030 	memset(&dc->sb.set_uuid, 0, 16);
1031 	SET_BDEV_STATE(&dc->sb, BDEV_STATE_NONE);
1032 
1033 	bch_write_bdev_super(dc, &cl);
1034 	closure_sync(&cl);
1035 
1036 	mutex_lock(&bch_register_lock);
1037 
1038 	calc_cached_dev_sectors(dc->disk.c);
1039 	bcache_device_detach(&dc->disk);
1040 	list_move(&dc->list, &uncached_devices);
1041 
1042 	clear_bit(BCACHE_DEV_DETACHING, &dc->disk.flags);
1043 	clear_bit(BCACHE_DEV_UNLINK_DONE, &dc->disk.flags);
1044 
1045 	mutex_unlock(&bch_register_lock);
1046 
1047 	pr_info("Caching disabled for %s", dc->backing_dev_name);
1048 
1049 	/* Drop ref we took in cached_dev_detach() */
1050 	closure_put(&dc->disk.cl);
1051 }
1052 
1053 void bch_cached_dev_detach(struct cached_dev *dc)
1054 {
1055 	lockdep_assert_held(&bch_register_lock);
1056 
1057 	if (test_bit(BCACHE_DEV_CLOSING, &dc->disk.flags))
1058 		return;
1059 
1060 	if (test_and_set_bit(BCACHE_DEV_DETACHING, &dc->disk.flags))
1061 		return;
1062 
1063 	/*
1064 	 * Block the device from being closed and freed until we're finished
1065 	 * detaching
1066 	 */
1067 	closure_get(&dc->disk.cl);
1068 
1069 	bch_writeback_queue(dc);
1070 
1071 	cached_dev_put(dc);
1072 }
1073 
1074 int bch_cached_dev_attach(struct cached_dev *dc, struct cache_set *c,
1075 			  uint8_t *set_uuid)
1076 {
1077 	uint32_t rtime = cpu_to_le32((u32)ktime_get_real_seconds());
1078 	struct uuid_entry *u;
1079 	struct cached_dev *exist_dc, *t;
1080 	int ret = 0;
1081 
1082 	if ((set_uuid && memcmp(set_uuid, c->sb.set_uuid, 16)) ||
1083 	    (!set_uuid && memcmp(dc->sb.set_uuid, c->sb.set_uuid, 16)))
1084 		return -ENOENT;
1085 
1086 	if (dc->disk.c) {
1087 		pr_err("Can't attach %s: already attached",
1088 		       dc->backing_dev_name);
1089 		return -EINVAL;
1090 	}
1091 
1092 	if (test_bit(CACHE_SET_STOPPING, &c->flags)) {
1093 		pr_err("Can't attach %s: shutting down",
1094 		       dc->backing_dev_name);
1095 		return -EINVAL;
1096 	}
1097 
1098 	if (dc->sb.block_size < c->sb.block_size) {
1099 		/* Will die */
1100 		pr_err("Couldn't attach %s: block size less than set's block size",
1101 		       dc->backing_dev_name);
1102 		return -EINVAL;
1103 	}
1104 
1105 	/* Check whether already attached */
1106 	list_for_each_entry_safe(exist_dc, t, &c->cached_devs, list) {
1107 		if (!memcmp(dc->sb.uuid, exist_dc->sb.uuid, 16)) {
1108 			pr_err("Tried to attach %s but duplicate UUID already attached",
1109 				dc->backing_dev_name);
1110 
1111 			return -EINVAL;
1112 		}
1113 	}
1114 
1115 	u = uuid_find(c, dc->sb.uuid);
1116 
1117 	if (u &&
1118 	    (BDEV_STATE(&dc->sb) == BDEV_STATE_STALE ||
1119 	     BDEV_STATE(&dc->sb) == BDEV_STATE_NONE)) {
1120 		memcpy(u->uuid, invalid_uuid, 16);
1121 		u->invalidated = cpu_to_le32((u32)ktime_get_real_seconds());
1122 		u = NULL;
1123 	}
1124 
1125 	if (!u) {
1126 		if (BDEV_STATE(&dc->sb) == BDEV_STATE_DIRTY) {
1127 			pr_err("Couldn't find uuid for %s in set",
1128 			       dc->backing_dev_name);
1129 			return -ENOENT;
1130 		}
1131 
1132 		u = uuid_find_empty(c);
1133 		if (!u) {
1134 			pr_err("Not caching %s, no room for UUID",
1135 			       dc->backing_dev_name);
1136 			return -EINVAL;
1137 		}
1138 	}
1139 
1140 	/*
1141 	 * Deadlocks since we're called via sysfs...
1142 	 * sysfs_remove_file(&dc->kobj, &sysfs_attach);
1143 	 */
1144 
1145 	if (bch_is_zero(u->uuid, 16)) {
1146 		struct closure cl;
1147 
1148 		closure_init_stack(&cl);
1149 
1150 		memcpy(u->uuid, dc->sb.uuid, 16);
1151 		memcpy(u->label, dc->sb.label, SB_LABEL_SIZE);
1152 		u->first_reg = u->last_reg = rtime;
1153 		bch_uuid_write(c);
1154 
1155 		memcpy(dc->sb.set_uuid, c->sb.set_uuid, 16);
1156 		SET_BDEV_STATE(&dc->sb, BDEV_STATE_CLEAN);
1157 
1158 		bch_write_bdev_super(dc, &cl);
1159 		closure_sync(&cl);
1160 	} else {
1161 		u->last_reg = rtime;
1162 		bch_uuid_write(c);
1163 	}
1164 
1165 	bcache_device_attach(&dc->disk, c, u - c->uuids);
1166 	list_move(&dc->list, &c->cached_devs);
1167 	calc_cached_dev_sectors(c);
1168 
1169 	/*
1170 	 * dc->c must be set before dc->count != 0 - paired with the mb in
1171 	 * cached_dev_get()
1172 	 */
1173 	smp_wmb();
1174 	refcount_set(&dc->count, 1);
1175 
1176 	/* Block writeback thread, but spawn it */
1177 	down_write(&dc->writeback_lock);
1178 	if (bch_cached_dev_writeback_start(dc)) {
1179 		up_write(&dc->writeback_lock);
1180 		pr_err("Couldn't start writeback facilities for %s",
1181 		       dc->disk.disk->disk_name);
1182 		return -ENOMEM;
1183 	}
1184 
1185 	if (BDEV_STATE(&dc->sb) == BDEV_STATE_DIRTY) {
1186 		atomic_set(&dc->has_dirty, 1);
1187 		bch_writeback_queue(dc);
1188 	}
1189 
1190 	bch_sectors_dirty_init(&dc->disk);
1191 
1192 	ret = bch_cached_dev_run(dc);
1193 	if (ret && (ret != -EBUSY)) {
1194 		up_write(&dc->writeback_lock);
1195 		/*
1196 		 * bch_register_lock is held, bcache_device_stop() is not
1197 		 * able to be directly called. The kthread and kworker
1198 		 * created previously in bch_cached_dev_writeback_start()
1199 		 * have to be stopped manually here.
1200 		 */
1201 		kthread_stop(dc->writeback_thread);
1202 		cancel_writeback_rate_update_dwork(dc);
1203 		pr_err("Couldn't run cached device %s",
1204 		       dc->backing_dev_name);
1205 		return ret;
1206 	}
1207 
1208 	bcache_device_link(&dc->disk, c, "bdev");
1209 	atomic_inc(&c->attached_dev_nr);
1210 
1211 	/* Allow the writeback thread to proceed */
1212 	up_write(&dc->writeback_lock);
1213 
1214 	pr_info("Caching %s as %s on set %pU",
1215 		dc->backing_dev_name,
1216 		dc->disk.disk->disk_name,
1217 		dc->disk.c->sb.set_uuid);
1218 	return 0;
1219 }
1220 
1221 /* when dc->disk.kobj released */
1222 void bch_cached_dev_release(struct kobject *kobj)
1223 {
1224 	struct cached_dev *dc = container_of(kobj, struct cached_dev,
1225 					     disk.kobj);
1226 	kfree(dc);
1227 	module_put(THIS_MODULE);
1228 }
1229 
1230 static void cached_dev_free(struct closure *cl)
1231 {
1232 	struct cached_dev *dc = container_of(cl, struct cached_dev, disk.cl);
1233 
1234 	if (test_and_clear_bit(BCACHE_DEV_WB_RUNNING, &dc->disk.flags))
1235 		cancel_writeback_rate_update_dwork(dc);
1236 
1237 	if (!IS_ERR_OR_NULL(dc->writeback_thread))
1238 		kthread_stop(dc->writeback_thread);
1239 	if (!IS_ERR_OR_NULL(dc->status_update_thread))
1240 		kthread_stop(dc->status_update_thread);
1241 
1242 	mutex_lock(&bch_register_lock);
1243 
1244 	if (atomic_read(&dc->running))
1245 		bd_unlink_disk_holder(dc->bdev, dc->disk.disk);
1246 	bcache_device_free(&dc->disk);
1247 	list_del(&dc->list);
1248 
1249 	mutex_unlock(&bch_register_lock);
1250 
1251 	if (!IS_ERR_OR_NULL(dc->bdev))
1252 		blkdev_put(dc->bdev, FMODE_READ|FMODE_WRITE|FMODE_EXCL);
1253 
1254 	wake_up(&unregister_wait);
1255 
1256 	kobject_put(&dc->disk.kobj);
1257 }
1258 
1259 static void cached_dev_flush(struct closure *cl)
1260 {
1261 	struct cached_dev *dc = container_of(cl, struct cached_dev, disk.cl);
1262 	struct bcache_device *d = &dc->disk;
1263 
1264 	mutex_lock(&bch_register_lock);
1265 	bcache_device_unlink(d);
1266 	mutex_unlock(&bch_register_lock);
1267 
1268 	bch_cache_accounting_destroy(&dc->accounting);
1269 	kobject_del(&d->kobj);
1270 
1271 	continue_at(cl, cached_dev_free, system_wq);
1272 }
1273 
1274 static int cached_dev_init(struct cached_dev *dc, unsigned int block_size)
1275 {
1276 	int ret;
1277 	struct io *io;
1278 	struct request_queue *q = bdev_get_queue(dc->bdev);
1279 
1280 	__module_get(THIS_MODULE);
1281 	INIT_LIST_HEAD(&dc->list);
1282 	closure_init(&dc->disk.cl, NULL);
1283 	set_closure_fn(&dc->disk.cl, cached_dev_flush, system_wq);
1284 	kobject_init(&dc->disk.kobj, &bch_cached_dev_ktype);
1285 	INIT_WORK(&dc->detach, cached_dev_detach_finish);
1286 	sema_init(&dc->sb_write_mutex, 1);
1287 	INIT_LIST_HEAD(&dc->io_lru);
1288 	spin_lock_init(&dc->io_lock);
1289 	bch_cache_accounting_init(&dc->accounting, &dc->disk.cl);
1290 
1291 	dc->sequential_cutoff		= 4 << 20;
1292 
1293 	for (io = dc->io; io < dc->io + RECENT_IO; io++) {
1294 		list_add(&io->lru, &dc->io_lru);
1295 		hlist_add_head(&io->hash, dc->io_hash + RECENT_IO);
1296 	}
1297 
1298 	dc->disk.stripe_size = q->limits.io_opt >> 9;
1299 
1300 	if (dc->disk.stripe_size)
1301 		dc->partial_stripes_expensive =
1302 			q->limits.raid_partial_stripes_expensive;
1303 
1304 	ret = bcache_device_init(&dc->disk, block_size,
1305 			 dc->bdev->bd_part->nr_sects - dc->sb.data_offset);
1306 	if (ret)
1307 		return ret;
1308 
1309 	dc->disk.disk->queue->backing_dev_info->ra_pages =
1310 		max(dc->disk.disk->queue->backing_dev_info->ra_pages,
1311 		    q->backing_dev_info->ra_pages);
1312 
1313 	atomic_set(&dc->io_errors, 0);
1314 	dc->io_disable = false;
1315 	dc->error_limit = DEFAULT_CACHED_DEV_ERROR_LIMIT;
1316 	/* default to auto */
1317 	dc->stop_when_cache_set_failed = BCH_CACHED_DEV_STOP_AUTO;
1318 
1319 	bch_cached_dev_request_init(dc);
1320 	bch_cached_dev_writeback_init(dc);
1321 	return 0;
1322 }
1323 
1324 /* Cached device - bcache superblock */
1325 
1326 static int register_bdev(struct cache_sb *sb, struct page *sb_page,
1327 				 struct block_device *bdev,
1328 				 struct cached_dev *dc)
1329 {
1330 	const char *err = "cannot allocate memory";
1331 	struct cache_set *c;
1332 	int ret = -ENOMEM;
1333 
1334 	bdevname(bdev, dc->backing_dev_name);
1335 	memcpy(&dc->sb, sb, sizeof(struct cache_sb));
1336 	dc->bdev = bdev;
1337 	dc->bdev->bd_holder = dc;
1338 
1339 	bio_init(&dc->sb_bio, dc->sb_bio.bi_inline_vecs, 1);
1340 	bio_first_bvec_all(&dc->sb_bio)->bv_page = sb_page;
1341 	get_page(sb_page);
1342 
1343 
1344 	if (cached_dev_init(dc, sb->block_size << 9))
1345 		goto err;
1346 
1347 	err = "error creating kobject";
1348 	if (kobject_add(&dc->disk.kobj, &part_to_dev(bdev->bd_part)->kobj,
1349 			"bcache"))
1350 		goto err;
1351 	if (bch_cache_accounting_add_kobjs(&dc->accounting, &dc->disk.kobj))
1352 		goto err;
1353 
1354 	pr_info("registered backing device %s", dc->backing_dev_name);
1355 
1356 	list_add(&dc->list, &uncached_devices);
1357 	/* attach to a matched cache set if it exists */
1358 	list_for_each_entry(c, &bch_cache_sets, list)
1359 		bch_cached_dev_attach(dc, c, NULL);
1360 
1361 	if (BDEV_STATE(&dc->sb) == BDEV_STATE_NONE ||
1362 	    BDEV_STATE(&dc->sb) == BDEV_STATE_STALE) {
1363 		err = "failed to run cached device";
1364 		ret = bch_cached_dev_run(dc);
1365 		if (ret)
1366 			goto err;
1367 	}
1368 
1369 	return 0;
1370 err:
1371 	pr_notice("error %s: %s", dc->backing_dev_name, err);
1372 	bcache_device_stop(&dc->disk);
1373 	return ret;
1374 }
1375 
1376 /* Flash only volumes */
1377 
1378 /* When d->kobj released */
1379 void bch_flash_dev_release(struct kobject *kobj)
1380 {
1381 	struct bcache_device *d = container_of(kobj, struct bcache_device,
1382 					       kobj);
1383 	kfree(d);
1384 }
1385 
1386 static void flash_dev_free(struct closure *cl)
1387 {
1388 	struct bcache_device *d = container_of(cl, struct bcache_device, cl);
1389 
1390 	mutex_lock(&bch_register_lock);
1391 	atomic_long_sub(bcache_dev_sectors_dirty(d),
1392 			&d->c->flash_dev_dirty_sectors);
1393 	bcache_device_free(d);
1394 	mutex_unlock(&bch_register_lock);
1395 	kobject_put(&d->kobj);
1396 }
1397 
1398 static void flash_dev_flush(struct closure *cl)
1399 {
1400 	struct bcache_device *d = container_of(cl, struct bcache_device, cl);
1401 
1402 	mutex_lock(&bch_register_lock);
1403 	bcache_device_unlink(d);
1404 	mutex_unlock(&bch_register_lock);
1405 	kobject_del(&d->kobj);
1406 	continue_at(cl, flash_dev_free, system_wq);
1407 }
1408 
1409 static int flash_dev_run(struct cache_set *c, struct uuid_entry *u)
1410 {
1411 	struct bcache_device *d = kzalloc(sizeof(struct bcache_device),
1412 					  GFP_KERNEL);
1413 	if (!d)
1414 		return -ENOMEM;
1415 
1416 	closure_init(&d->cl, NULL);
1417 	set_closure_fn(&d->cl, flash_dev_flush, system_wq);
1418 
1419 	kobject_init(&d->kobj, &bch_flash_dev_ktype);
1420 
1421 	if (bcache_device_init(d, block_bytes(c), u->sectors))
1422 		goto err;
1423 
1424 	bcache_device_attach(d, c, u - c->uuids);
1425 	bch_sectors_dirty_init(d);
1426 	bch_flash_dev_request_init(d);
1427 	add_disk(d->disk);
1428 
1429 	if (kobject_add(&d->kobj, &disk_to_dev(d->disk)->kobj, "bcache"))
1430 		goto err;
1431 
1432 	bcache_device_link(d, c, "volume");
1433 
1434 	return 0;
1435 err:
1436 	kobject_put(&d->kobj);
1437 	return -ENOMEM;
1438 }
1439 
1440 static int flash_devs_run(struct cache_set *c)
1441 {
1442 	int ret = 0;
1443 	struct uuid_entry *u;
1444 
1445 	for (u = c->uuids;
1446 	     u < c->uuids + c->nr_uuids && !ret;
1447 	     u++)
1448 		if (UUID_FLASH_ONLY(u))
1449 			ret = flash_dev_run(c, u);
1450 
1451 	return ret;
1452 }
1453 
1454 int bch_flash_dev_create(struct cache_set *c, uint64_t size)
1455 {
1456 	struct uuid_entry *u;
1457 
1458 	if (test_bit(CACHE_SET_STOPPING, &c->flags))
1459 		return -EINTR;
1460 
1461 	if (!test_bit(CACHE_SET_RUNNING, &c->flags))
1462 		return -EPERM;
1463 
1464 	u = uuid_find_empty(c);
1465 	if (!u) {
1466 		pr_err("Can't create volume, no room for UUID");
1467 		return -EINVAL;
1468 	}
1469 
1470 	get_random_bytes(u->uuid, 16);
1471 	memset(u->label, 0, 32);
1472 	u->first_reg = u->last_reg = cpu_to_le32((u32)ktime_get_real_seconds());
1473 
1474 	SET_UUID_FLASH_ONLY(u, 1);
1475 	u->sectors = size >> 9;
1476 
1477 	bch_uuid_write(c);
1478 
1479 	return flash_dev_run(c, u);
1480 }
1481 
1482 bool bch_cached_dev_error(struct cached_dev *dc)
1483 {
1484 	if (!dc || test_bit(BCACHE_DEV_CLOSING, &dc->disk.flags))
1485 		return false;
1486 
1487 	dc->io_disable = true;
1488 	/* make others know io_disable is true earlier */
1489 	smp_mb();
1490 
1491 	pr_err("stop %s: too many IO errors on backing device %s\n",
1492 		dc->disk.disk->disk_name, dc->backing_dev_name);
1493 
1494 	bcache_device_stop(&dc->disk);
1495 	return true;
1496 }
1497 
1498 /* Cache set */
1499 
1500 __printf(2, 3)
1501 bool bch_cache_set_error(struct cache_set *c, const char *fmt, ...)
1502 {
1503 	va_list args;
1504 
1505 	if (c->on_error != ON_ERROR_PANIC &&
1506 	    test_bit(CACHE_SET_STOPPING, &c->flags))
1507 		return false;
1508 
1509 	if (test_and_set_bit(CACHE_SET_IO_DISABLE, &c->flags))
1510 		pr_info("CACHE_SET_IO_DISABLE already set");
1511 
1512 	/*
1513 	 * XXX: we can be called from atomic context
1514 	 * acquire_console_sem();
1515 	 */
1516 
1517 	pr_err("bcache: error on %pU: ", c->sb.set_uuid);
1518 
1519 	va_start(args, fmt);
1520 	vprintk(fmt, args);
1521 	va_end(args);
1522 
1523 	pr_err(", disabling caching\n");
1524 
1525 	if (c->on_error == ON_ERROR_PANIC)
1526 		panic("panic forced after error\n");
1527 
1528 	bch_cache_set_unregister(c);
1529 	return true;
1530 }
1531 
1532 /* When c->kobj released */
1533 void bch_cache_set_release(struct kobject *kobj)
1534 {
1535 	struct cache_set *c = container_of(kobj, struct cache_set, kobj);
1536 
1537 	kfree(c);
1538 	module_put(THIS_MODULE);
1539 }
1540 
1541 static void cache_set_free(struct closure *cl)
1542 {
1543 	struct cache_set *c = container_of(cl, struct cache_set, cl);
1544 	struct cache *ca;
1545 	unsigned int i;
1546 
1547 	debugfs_remove(c->debug);
1548 
1549 	bch_open_buckets_free(c);
1550 	bch_btree_cache_free(c);
1551 	bch_journal_free(c);
1552 
1553 	mutex_lock(&bch_register_lock);
1554 	for_each_cache(ca, c, i)
1555 		if (ca) {
1556 			ca->set = NULL;
1557 			c->cache[ca->sb.nr_this_dev] = NULL;
1558 			kobject_put(&ca->kobj);
1559 		}
1560 
1561 	bch_bset_sort_state_free(&c->sort);
1562 	free_pages((unsigned long) c->uuids, ilog2(bucket_pages(c)));
1563 
1564 	if (c->moving_gc_wq)
1565 		destroy_workqueue(c->moving_gc_wq);
1566 	bioset_exit(&c->bio_split);
1567 	mempool_exit(&c->fill_iter);
1568 	mempool_exit(&c->bio_meta);
1569 	mempool_exit(&c->search);
1570 	kfree(c->devices);
1571 
1572 	list_del(&c->list);
1573 	mutex_unlock(&bch_register_lock);
1574 
1575 	pr_info("Cache set %pU unregistered", c->sb.set_uuid);
1576 	wake_up(&unregister_wait);
1577 
1578 	closure_debug_destroy(&c->cl);
1579 	kobject_put(&c->kobj);
1580 }
1581 
1582 static void cache_set_flush(struct closure *cl)
1583 {
1584 	struct cache_set *c = container_of(cl, struct cache_set, caching);
1585 	struct cache *ca;
1586 	struct btree *b;
1587 	unsigned int i;
1588 
1589 	bch_cache_accounting_destroy(&c->accounting);
1590 
1591 	kobject_put(&c->internal);
1592 	kobject_del(&c->kobj);
1593 
1594 	if (!IS_ERR_OR_NULL(c->gc_thread))
1595 		kthread_stop(c->gc_thread);
1596 
1597 	if (!IS_ERR_OR_NULL(c->root))
1598 		list_add(&c->root->list, &c->btree_cache);
1599 
1600 	/*
1601 	 * Avoid flushing cached nodes if cache set is retiring
1602 	 * due to too many I/O errors detected.
1603 	 */
1604 	if (!test_bit(CACHE_SET_IO_DISABLE, &c->flags))
1605 		list_for_each_entry(b, &c->btree_cache, list) {
1606 			mutex_lock(&b->write_lock);
1607 			if (btree_node_dirty(b))
1608 				__bch_btree_node_write(b, NULL);
1609 			mutex_unlock(&b->write_lock);
1610 		}
1611 
1612 	for_each_cache(ca, c, i)
1613 		if (ca->alloc_thread)
1614 			kthread_stop(ca->alloc_thread);
1615 
1616 	if (c->journal.cur) {
1617 		cancel_delayed_work_sync(&c->journal.work);
1618 		/* flush last journal entry if needed */
1619 		c->journal.work.work.func(&c->journal.work.work);
1620 	}
1621 
1622 	closure_return(cl);
1623 }
1624 
1625 /*
1626  * This function is only called when CACHE_SET_IO_DISABLE is set, which means
1627  * cache set is unregistering due to too many I/O errors. In this condition,
1628  * the bcache device might be stopped, it depends on stop_when_cache_set_failed
1629  * value and whether the broken cache has dirty data:
1630  *
1631  * dc->stop_when_cache_set_failed    dc->has_dirty   stop bcache device
1632  *  BCH_CACHED_STOP_AUTO               0               NO
1633  *  BCH_CACHED_STOP_AUTO               1               YES
1634  *  BCH_CACHED_DEV_STOP_ALWAYS         0               YES
1635  *  BCH_CACHED_DEV_STOP_ALWAYS         1               YES
1636  *
1637  * The expected behavior is, if stop_when_cache_set_failed is configured to
1638  * "auto" via sysfs interface, the bcache device will not be stopped if the
1639  * backing device is clean on the broken cache device.
1640  */
1641 static void conditional_stop_bcache_device(struct cache_set *c,
1642 					   struct bcache_device *d,
1643 					   struct cached_dev *dc)
1644 {
1645 	if (dc->stop_when_cache_set_failed == BCH_CACHED_DEV_STOP_ALWAYS) {
1646 		pr_warn("stop_when_cache_set_failed of %s is \"always\", stop it for failed cache set %pU.",
1647 			d->disk->disk_name, c->sb.set_uuid);
1648 		bcache_device_stop(d);
1649 	} else if (atomic_read(&dc->has_dirty)) {
1650 		/*
1651 		 * dc->stop_when_cache_set_failed == BCH_CACHED_STOP_AUTO
1652 		 * and dc->has_dirty == 1
1653 		 */
1654 		pr_warn("stop_when_cache_set_failed of %s is \"auto\" and cache is dirty, stop it to avoid potential data corruption.",
1655 			d->disk->disk_name);
1656 		/*
1657 		 * There might be a small time gap that cache set is
1658 		 * released but bcache device is not. Inside this time
1659 		 * gap, regular I/O requests will directly go into
1660 		 * backing device as no cache set attached to. This
1661 		 * behavior may also introduce potential inconsistence
1662 		 * data in writeback mode while cache is dirty.
1663 		 * Therefore before calling bcache_device_stop() due
1664 		 * to a broken cache device, dc->io_disable should be
1665 		 * explicitly set to true.
1666 		 */
1667 		dc->io_disable = true;
1668 		/* make others know io_disable is true earlier */
1669 		smp_mb();
1670 		bcache_device_stop(d);
1671 	} else {
1672 		/*
1673 		 * dc->stop_when_cache_set_failed == BCH_CACHED_STOP_AUTO
1674 		 * and dc->has_dirty == 0
1675 		 */
1676 		pr_warn("stop_when_cache_set_failed of %s is \"auto\" and cache is clean, keep it alive.",
1677 			d->disk->disk_name);
1678 	}
1679 }
1680 
1681 static void __cache_set_unregister(struct closure *cl)
1682 {
1683 	struct cache_set *c = container_of(cl, struct cache_set, caching);
1684 	struct cached_dev *dc;
1685 	struct bcache_device *d;
1686 	size_t i;
1687 
1688 	mutex_lock(&bch_register_lock);
1689 
1690 	for (i = 0; i < c->devices_max_used; i++) {
1691 		d = c->devices[i];
1692 		if (!d)
1693 			continue;
1694 
1695 		if (!UUID_FLASH_ONLY(&c->uuids[i]) &&
1696 		    test_bit(CACHE_SET_UNREGISTERING, &c->flags)) {
1697 			dc = container_of(d, struct cached_dev, disk);
1698 			bch_cached_dev_detach(dc);
1699 			if (test_bit(CACHE_SET_IO_DISABLE, &c->flags))
1700 				conditional_stop_bcache_device(c, d, dc);
1701 		} else {
1702 			bcache_device_stop(d);
1703 		}
1704 	}
1705 
1706 	mutex_unlock(&bch_register_lock);
1707 
1708 	continue_at(cl, cache_set_flush, system_wq);
1709 }
1710 
1711 void bch_cache_set_stop(struct cache_set *c)
1712 {
1713 	if (!test_and_set_bit(CACHE_SET_STOPPING, &c->flags))
1714 		/* closure_fn set to __cache_set_unregister() */
1715 		closure_queue(&c->caching);
1716 }
1717 
1718 void bch_cache_set_unregister(struct cache_set *c)
1719 {
1720 	set_bit(CACHE_SET_UNREGISTERING, &c->flags);
1721 	bch_cache_set_stop(c);
1722 }
1723 
1724 #define alloc_bucket_pages(gfp, c)			\
1725 	((void *) __get_free_pages(__GFP_ZERO|gfp, ilog2(bucket_pages(c))))
1726 
1727 struct cache_set *bch_cache_set_alloc(struct cache_sb *sb)
1728 {
1729 	int iter_size;
1730 	struct cache_set *c = kzalloc(sizeof(struct cache_set), GFP_KERNEL);
1731 
1732 	if (!c)
1733 		return NULL;
1734 
1735 	__module_get(THIS_MODULE);
1736 	closure_init(&c->cl, NULL);
1737 	set_closure_fn(&c->cl, cache_set_free, system_wq);
1738 
1739 	closure_init(&c->caching, &c->cl);
1740 	set_closure_fn(&c->caching, __cache_set_unregister, system_wq);
1741 
1742 	/* Maybe create continue_at_noreturn() and use it here? */
1743 	closure_set_stopped(&c->cl);
1744 	closure_put(&c->cl);
1745 
1746 	kobject_init(&c->kobj, &bch_cache_set_ktype);
1747 	kobject_init(&c->internal, &bch_cache_set_internal_ktype);
1748 
1749 	bch_cache_accounting_init(&c->accounting, &c->cl);
1750 
1751 	memcpy(c->sb.set_uuid, sb->set_uuid, 16);
1752 	c->sb.block_size	= sb->block_size;
1753 	c->sb.bucket_size	= sb->bucket_size;
1754 	c->sb.nr_in_set		= sb->nr_in_set;
1755 	c->sb.last_mount	= sb->last_mount;
1756 	c->bucket_bits		= ilog2(sb->bucket_size);
1757 	c->block_bits		= ilog2(sb->block_size);
1758 	c->nr_uuids		= bucket_bytes(c) / sizeof(struct uuid_entry);
1759 	c->devices_max_used	= 0;
1760 	atomic_set(&c->attached_dev_nr, 0);
1761 	c->btree_pages		= bucket_pages(c);
1762 	if (c->btree_pages > BTREE_MAX_PAGES)
1763 		c->btree_pages = max_t(int, c->btree_pages / 4,
1764 				       BTREE_MAX_PAGES);
1765 
1766 	sema_init(&c->sb_write_mutex, 1);
1767 	mutex_init(&c->bucket_lock);
1768 	init_waitqueue_head(&c->btree_cache_wait);
1769 	init_waitqueue_head(&c->bucket_wait);
1770 	init_waitqueue_head(&c->gc_wait);
1771 	sema_init(&c->uuid_write_mutex, 1);
1772 
1773 	spin_lock_init(&c->btree_gc_time.lock);
1774 	spin_lock_init(&c->btree_split_time.lock);
1775 	spin_lock_init(&c->btree_read_time.lock);
1776 
1777 	bch_moving_init_cache_set(c);
1778 
1779 	INIT_LIST_HEAD(&c->list);
1780 	INIT_LIST_HEAD(&c->cached_devs);
1781 	INIT_LIST_HEAD(&c->btree_cache);
1782 	INIT_LIST_HEAD(&c->btree_cache_freeable);
1783 	INIT_LIST_HEAD(&c->btree_cache_freed);
1784 	INIT_LIST_HEAD(&c->data_buckets);
1785 
1786 	iter_size = (sb->bucket_size / sb->block_size + 1) *
1787 		sizeof(struct btree_iter_set);
1788 
1789 	if (!(c->devices = kcalloc(c->nr_uuids, sizeof(void *), GFP_KERNEL)) ||
1790 	    mempool_init_slab_pool(&c->search, 32, bch_search_cache) ||
1791 	    mempool_init_kmalloc_pool(&c->bio_meta, 2,
1792 				sizeof(struct bbio) + sizeof(struct bio_vec) *
1793 				bucket_pages(c)) ||
1794 	    mempool_init_kmalloc_pool(&c->fill_iter, 1, iter_size) ||
1795 	    bioset_init(&c->bio_split, 4, offsetof(struct bbio, bio),
1796 			BIOSET_NEED_BVECS|BIOSET_NEED_RESCUER) ||
1797 	    !(c->uuids = alloc_bucket_pages(GFP_KERNEL, c)) ||
1798 	    !(c->moving_gc_wq = alloc_workqueue("bcache_gc",
1799 						WQ_MEM_RECLAIM, 0)) ||
1800 	    bch_journal_alloc(c) ||
1801 	    bch_btree_cache_alloc(c) ||
1802 	    bch_open_buckets_alloc(c) ||
1803 	    bch_bset_sort_state_init(&c->sort, ilog2(c->btree_pages)))
1804 		goto err;
1805 
1806 	c->congested_read_threshold_us	= 2000;
1807 	c->congested_write_threshold_us	= 20000;
1808 	c->error_limit	= DEFAULT_IO_ERROR_LIMIT;
1809 	WARN_ON(test_and_clear_bit(CACHE_SET_IO_DISABLE, &c->flags));
1810 
1811 	return c;
1812 err:
1813 	bch_cache_set_unregister(c);
1814 	return NULL;
1815 }
1816 
1817 static int run_cache_set(struct cache_set *c)
1818 {
1819 	const char *err = "cannot allocate memory";
1820 	struct cached_dev *dc, *t;
1821 	struct cache *ca;
1822 	struct closure cl;
1823 	unsigned int i;
1824 	LIST_HEAD(journal);
1825 	struct journal_replay *l;
1826 
1827 	closure_init_stack(&cl);
1828 
1829 	for_each_cache(ca, c, i)
1830 		c->nbuckets += ca->sb.nbuckets;
1831 	set_gc_sectors(c);
1832 
1833 	if (CACHE_SYNC(&c->sb)) {
1834 		struct bkey *k;
1835 		struct jset *j;
1836 
1837 		err = "cannot allocate memory for journal";
1838 		if (bch_journal_read(c, &journal))
1839 			goto err;
1840 
1841 		pr_debug("btree_journal_read() done");
1842 
1843 		err = "no journal entries found";
1844 		if (list_empty(&journal))
1845 			goto err;
1846 
1847 		j = &list_entry(journal.prev, struct journal_replay, list)->j;
1848 
1849 		err = "IO error reading priorities";
1850 		for_each_cache(ca, c, i)
1851 			prio_read(ca, j->prio_bucket[ca->sb.nr_this_dev]);
1852 
1853 		/*
1854 		 * If prio_read() fails it'll call cache_set_error and we'll
1855 		 * tear everything down right away, but if we perhaps checked
1856 		 * sooner we could avoid journal replay.
1857 		 */
1858 
1859 		k = &j->btree_root;
1860 
1861 		err = "bad btree root";
1862 		if (__bch_btree_ptr_invalid(c, k))
1863 			goto err;
1864 
1865 		err = "error reading btree root";
1866 		c->root = bch_btree_node_get(c, NULL, k,
1867 					     j->btree_level,
1868 					     true, NULL);
1869 		if (IS_ERR_OR_NULL(c->root))
1870 			goto err;
1871 
1872 		list_del_init(&c->root->list);
1873 		rw_unlock(true, c->root);
1874 
1875 		err = uuid_read(c, j, &cl);
1876 		if (err)
1877 			goto err;
1878 
1879 		err = "error in recovery";
1880 		if (bch_btree_check(c))
1881 			goto err;
1882 
1883 		/*
1884 		 * bch_btree_check() may occupy too much system memory which
1885 		 * has negative effects to user space application (e.g. data
1886 		 * base) performance. Shrink the mca cache memory proactively
1887 		 * here to avoid competing memory with user space workloads..
1888 		 */
1889 		if (!c->shrinker_disabled) {
1890 			struct shrink_control sc;
1891 
1892 			sc.gfp_mask = GFP_KERNEL;
1893 			sc.nr_to_scan = c->btree_cache_used * c->btree_pages;
1894 			/* first run to clear b->accessed tag */
1895 			c->shrink.scan_objects(&c->shrink, &sc);
1896 			/* second run to reap non-accessed nodes */
1897 			c->shrink.scan_objects(&c->shrink, &sc);
1898 		}
1899 
1900 		bch_journal_mark(c, &journal);
1901 		bch_initial_gc_finish(c);
1902 		pr_debug("btree_check() done");
1903 
1904 		/*
1905 		 * bcache_journal_next() can't happen sooner, or
1906 		 * btree_gc_finish() will give spurious errors about last_gc >
1907 		 * gc_gen - this is a hack but oh well.
1908 		 */
1909 		bch_journal_next(&c->journal);
1910 
1911 		err = "error starting allocator thread";
1912 		for_each_cache(ca, c, i)
1913 			if (bch_cache_allocator_start(ca))
1914 				goto err;
1915 
1916 		/*
1917 		 * First place it's safe to allocate: btree_check() and
1918 		 * btree_gc_finish() have to run before we have buckets to
1919 		 * allocate, and bch_bucket_alloc_set() might cause a journal
1920 		 * entry to be written so bcache_journal_next() has to be called
1921 		 * first.
1922 		 *
1923 		 * If the uuids were in the old format we have to rewrite them
1924 		 * before the next journal entry is written:
1925 		 */
1926 		if (j->version < BCACHE_JSET_VERSION_UUID)
1927 			__uuid_write(c);
1928 
1929 		err = "bcache: replay journal failed";
1930 		if (bch_journal_replay(c, &journal))
1931 			goto err;
1932 	} else {
1933 		pr_notice("invalidating existing data");
1934 
1935 		for_each_cache(ca, c, i) {
1936 			unsigned int j;
1937 
1938 			ca->sb.keys = clamp_t(int, ca->sb.nbuckets >> 7,
1939 					      2, SB_JOURNAL_BUCKETS);
1940 
1941 			for (j = 0; j < ca->sb.keys; j++)
1942 				ca->sb.d[j] = ca->sb.first_bucket + j;
1943 		}
1944 
1945 		bch_initial_gc_finish(c);
1946 
1947 		err = "error starting allocator thread";
1948 		for_each_cache(ca, c, i)
1949 			if (bch_cache_allocator_start(ca))
1950 				goto err;
1951 
1952 		mutex_lock(&c->bucket_lock);
1953 		for_each_cache(ca, c, i)
1954 			bch_prio_write(ca);
1955 		mutex_unlock(&c->bucket_lock);
1956 
1957 		err = "cannot allocate new UUID bucket";
1958 		if (__uuid_write(c))
1959 			goto err;
1960 
1961 		err = "cannot allocate new btree root";
1962 		c->root = __bch_btree_node_alloc(c, NULL, 0, true, NULL);
1963 		if (IS_ERR_OR_NULL(c->root))
1964 			goto err;
1965 
1966 		mutex_lock(&c->root->write_lock);
1967 		bkey_copy_key(&c->root->key, &MAX_KEY);
1968 		bch_btree_node_write(c->root, &cl);
1969 		mutex_unlock(&c->root->write_lock);
1970 
1971 		bch_btree_set_root(c->root);
1972 		rw_unlock(true, c->root);
1973 
1974 		/*
1975 		 * We don't want to write the first journal entry until
1976 		 * everything is set up - fortunately journal entries won't be
1977 		 * written until the SET_CACHE_SYNC() here:
1978 		 */
1979 		SET_CACHE_SYNC(&c->sb, true);
1980 
1981 		bch_journal_next(&c->journal);
1982 		bch_journal_meta(c, &cl);
1983 	}
1984 
1985 	err = "error starting gc thread";
1986 	if (bch_gc_thread_start(c))
1987 		goto err;
1988 
1989 	closure_sync(&cl);
1990 	c->sb.last_mount = (u32)ktime_get_real_seconds();
1991 	bcache_write_super(c);
1992 
1993 	list_for_each_entry_safe(dc, t, &uncached_devices, list)
1994 		bch_cached_dev_attach(dc, c, NULL);
1995 
1996 	flash_devs_run(c);
1997 
1998 	set_bit(CACHE_SET_RUNNING, &c->flags);
1999 	return 0;
2000 err:
2001 	while (!list_empty(&journal)) {
2002 		l = list_first_entry(&journal, struct journal_replay, list);
2003 		list_del(&l->list);
2004 		kfree(l);
2005 	}
2006 
2007 	closure_sync(&cl);
2008 
2009 	bch_cache_set_error(c, "%s", err);
2010 
2011 	return -EIO;
2012 }
2013 
2014 static bool can_attach_cache(struct cache *ca, struct cache_set *c)
2015 {
2016 	return ca->sb.block_size	== c->sb.block_size &&
2017 		ca->sb.bucket_size	== c->sb.bucket_size &&
2018 		ca->sb.nr_in_set	== c->sb.nr_in_set;
2019 }
2020 
2021 static const char *register_cache_set(struct cache *ca)
2022 {
2023 	char buf[12];
2024 	const char *err = "cannot allocate memory";
2025 	struct cache_set *c;
2026 
2027 	list_for_each_entry(c, &bch_cache_sets, list)
2028 		if (!memcmp(c->sb.set_uuid, ca->sb.set_uuid, 16)) {
2029 			if (c->cache[ca->sb.nr_this_dev])
2030 				return "duplicate cache set member";
2031 
2032 			if (!can_attach_cache(ca, c))
2033 				return "cache sb does not match set";
2034 
2035 			if (!CACHE_SYNC(&ca->sb))
2036 				SET_CACHE_SYNC(&c->sb, false);
2037 
2038 			goto found;
2039 		}
2040 
2041 	c = bch_cache_set_alloc(&ca->sb);
2042 	if (!c)
2043 		return err;
2044 
2045 	err = "error creating kobject";
2046 	if (kobject_add(&c->kobj, bcache_kobj, "%pU", c->sb.set_uuid) ||
2047 	    kobject_add(&c->internal, &c->kobj, "internal"))
2048 		goto err;
2049 
2050 	if (bch_cache_accounting_add_kobjs(&c->accounting, &c->kobj))
2051 		goto err;
2052 
2053 	bch_debug_init_cache_set(c);
2054 
2055 	list_add(&c->list, &bch_cache_sets);
2056 found:
2057 	sprintf(buf, "cache%i", ca->sb.nr_this_dev);
2058 	if (sysfs_create_link(&ca->kobj, &c->kobj, "set") ||
2059 	    sysfs_create_link(&c->kobj, &ca->kobj, buf))
2060 		goto err;
2061 
2062 	if (ca->sb.seq > c->sb.seq) {
2063 		c->sb.version		= ca->sb.version;
2064 		memcpy(c->sb.set_uuid, ca->sb.set_uuid, 16);
2065 		c->sb.flags             = ca->sb.flags;
2066 		c->sb.seq		= ca->sb.seq;
2067 		pr_debug("set version = %llu", c->sb.version);
2068 	}
2069 
2070 	kobject_get(&ca->kobj);
2071 	ca->set = c;
2072 	ca->set->cache[ca->sb.nr_this_dev] = ca;
2073 	c->cache_by_alloc[c->caches_loaded++] = ca;
2074 
2075 	if (c->caches_loaded == c->sb.nr_in_set) {
2076 		err = "failed to run cache set";
2077 		if (run_cache_set(c) < 0)
2078 			goto err;
2079 	}
2080 
2081 	return NULL;
2082 err:
2083 	bch_cache_set_unregister(c);
2084 	return err;
2085 }
2086 
2087 /* Cache device */
2088 
2089 /* When ca->kobj released */
2090 void bch_cache_release(struct kobject *kobj)
2091 {
2092 	struct cache *ca = container_of(kobj, struct cache, kobj);
2093 	unsigned int i;
2094 
2095 	if (ca->set) {
2096 		BUG_ON(ca->set->cache[ca->sb.nr_this_dev] != ca);
2097 		ca->set->cache[ca->sb.nr_this_dev] = NULL;
2098 	}
2099 
2100 	free_pages((unsigned long) ca->disk_buckets, ilog2(bucket_pages(ca)));
2101 	kfree(ca->prio_buckets);
2102 	vfree(ca->buckets);
2103 
2104 	free_heap(&ca->heap);
2105 	free_fifo(&ca->free_inc);
2106 
2107 	for (i = 0; i < RESERVE_NR; i++)
2108 		free_fifo(&ca->free[i]);
2109 
2110 	if (ca->sb_bio.bi_inline_vecs[0].bv_page)
2111 		put_page(bio_first_page_all(&ca->sb_bio));
2112 
2113 	if (!IS_ERR_OR_NULL(ca->bdev))
2114 		blkdev_put(ca->bdev, FMODE_READ|FMODE_WRITE|FMODE_EXCL);
2115 
2116 	kfree(ca);
2117 	module_put(THIS_MODULE);
2118 }
2119 
2120 static int cache_alloc(struct cache *ca)
2121 {
2122 	size_t free;
2123 	size_t btree_buckets;
2124 	struct bucket *b;
2125 	int ret = -ENOMEM;
2126 	const char *err = NULL;
2127 
2128 	__module_get(THIS_MODULE);
2129 	kobject_init(&ca->kobj, &bch_cache_ktype);
2130 
2131 	bio_init(&ca->journal.bio, ca->journal.bio.bi_inline_vecs, 8);
2132 
2133 	/*
2134 	 * when ca->sb.njournal_buckets is not zero, journal exists,
2135 	 * and in bch_journal_replay(), tree node may split,
2136 	 * so bucket of RESERVE_BTREE type is needed,
2137 	 * the worst situation is all journal buckets are valid journal,
2138 	 * and all the keys need to replay,
2139 	 * so the number of  RESERVE_BTREE type buckets should be as much
2140 	 * as journal buckets
2141 	 */
2142 	btree_buckets = ca->sb.njournal_buckets ?: 8;
2143 	free = roundup_pow_of_two(ca->sb.nbuckets) >> 10;
2144 	if (!free) {
2145 		ret = -EPERM;
2146 		err = "ca->sb.nbuckets is too small";
2147 		goto err_free;
2148 	}
2149 
2150 	if (!init_fifo(&ca->free[RESERVE_BTREE], btree_buckets,
2151 						GFP_KERNEL)) {
2152 		err = "ca->free[RESERVE_BTREE] alloc failed";
2153 		goto err_btree_alloc;
2154 	}
2155 
2156 	if (!init_fifo_exact(&ca->free[RESERVE_PRIO], prio_buckets(ca),
2157 							GFP_KERNEL)) {
2158 		err = "ca->free[RESERVE_PRIO] alloc failed";
2159 		goto err_prio_alloc;
2160 	}
2161 
2162 	if (!init_fifo(&ca->free[RESERVE_MOVINGGC], free, GFP_KERNEL)) {
2163 		err = "ca->free[RESERVE_MOVINGGC] alloc failed";
2164 		goto err_movinggc_alloc;
2165 	}
2166 
2167 	if (!init_fifo(&ca->free[RESERVE_NONE], free, GFP_KERNEL)) {
2168 		err = "ca->free[RESERVE_NONE] alloc failed";
2169 		goto err_none_alloc;
2170 	}
2171 
2172 	if (!init_fifo(&ca->free_inc, free << 2, GFP_KERNEL)) {
2173 		err = "ca->free_inc alloc failed";
2174 		goto err_free_inc_alloc;
2175 	}
2176 
2177 	if (!init_heap(&ca->heap, free << 3, GFP_KERNEL)) {
2178 		err = "ca->heap alloc failed";
2179 		goto err_heap_alloc;
2180 	}
2181 
2182 	ca->buckets = vzalloc(array_size(sizeof(struct bucket),
2183 			      ca->sb.nbuckets));
2184 	if (!ca->buckets) {
2185 		err = "ca->buckets alloc failed";
2186 		goto err_buckets_alloc;
2187 	}
2188 
2189 	ca->prio_buckets = kzalloc(array3_size(sizeof(uint64_t),
2190 				   prio_buckets(ca), 2),
2191 				   GFP_KERNEL);
2192 	if (!ca->prio_buckets) {
2193 		err = "ca->prio_buckets alloc failed";
2194 		goto err_prio_buckets_alloc;
2195 	}
2196 
2197 	ca->disk_buckets = alloc_bucket_pages(GFP_KERNEL, ca);
2198 	if (!ca->disk_buckets) {
2199 		err = "ca->disk_buckets alloc failed";
2200 		goto err_disk_buckets_alloc;
2201 	}
2202 
2203 	ca->prio_last_buckets = ca->prio_buckets + prio_buckets(ca);
2204 
2205 	for_each_bucket(b, ca)
2206 		atomic_set(&b->pin, 0);
2207 	return 0;
2208 
2209 err_disk_buckets_alloc:
2210 	kfree(ca->prio_buckets);
2211 err_prio_buckets_alloc:
2212 	vfree(ca->buckets);
2213 err_buckets_alloc:
2214 	free_heap(&ca->heap);
2215 err_heap_alloc:
2216 	free_fifo(&ca->free_inc);
2217 err_free_inc_alloc:
2218 	free_fifo(&ca->free[RESERVE_NONE]);
2219 err_none_alloc:
2220 	free_fifo(&ca->free[RESERVE_MOVINGGC]);
2221 err_movinggc_alloc:
2222 	free_fifo(&ca->free[RESERVE_PRIO]);
2223 err_prio_alloc:
2224 	free_fifo(&ca->free[RESERVE_BTREE]);
2225 err_btree_alloc:
2226 err_free:
2227 	module_put(THIS_MODULE);
2228 	if (err)
2229 		pr_notice("error %s: %s", ca->cache_dev_name, err);
2230 	return ret;
2231 }
2232 
2233 static int register_cache(struct cache_sb *sb, struct page *sb_page,
2234 				struct block_device *bdev, struct cache *ca)
2235 {
2236 	const char *err = NULL; /* must be set for any error case */
2237 	int ret = 0;
2238 
2239 	bdevname(bdev, ca->cache_dev_name);
2240 	memcpy(&ca->sb, sb, sizeof(struct cache_sb));
2241 	ca->bdev = bdev;
2242 	ca->bdev->bd_holder = ca;
2243 
2244 	bio_init(&ca->sb_bio, ca->sb_bio.bi_inline_vecs, 1);
2245 	bio_first_bvec_all(&ca->sb_bio)->bv_page = sb_page;
2246 	get_page(sb_page);
2247 
2248 	if (blk_queue_discard(bdev_get_queue(bdev)))
2249 		ca->discard = CACHE_DISCARD(&ca->sb);
2250 
2251 	ret = cache_alloc(ca);
2252 	if (ret != 0) {
2253 		/*
2254 		 * If we failed here, it means ca->kobj is not initialized yet,
2255 		 * kobject_put() won't be called and there is no chance to
2256 		 * call blkdev_put() to bdev in bch_cache_release(). So we
2257 		 * explicitly call blkdev_put() here.
2258 		 */
2259 		blkdev_put(bdev, FMODE_READ|FMODE_WRITE|FMODE_EXCL);
2260 		if (ret == -ENOMEM)
2261 			err = "cache_alloc(): -ENOMEM";
2262 		else if (ret == -EPERM)
2263 			err = "cache_alloc(): cache device is too small";
2264 		else
2265 			err = "cache_alloc(): unknown error";
2266 		goto err;
2267 	}
2268 
2269 	if (kobject_add(&ca->kobj,
2270 			&part_to_dev(bdev->bd_part)->kobj,
2271 			"bcache")) {
2272 		err = "error calling kobject_add";
2273 		ret = -ENOMEM;
2274 		goto out;
2275 	}
2276 
2277 	mutex_lock(&bch_register_lock);
2278 	err = register_cache_set(ca);
2279 	mutex_unlock(&bch_register_lock);
2280 
2281 	if (err) {
2282 		ret = -ENODEV;
2283 		goto out;
2284 	}
2285 
2286 	pr_info("registered cache device %s", ca->cache_dev_name);
2287 
2288 out:
2289 	kobject_put(&ca->kobj);
2290 
2291 err:
2292 	if (err)
2293 		pr_notice("error %s: %s", ca->cache_dev_name, err);
2294 
2295 	return ret;
2296 }
2297 
2298 /* Global interfaces/init */
2299 
2300 static ssize_t register_bcache(struct kobject *k, struct kobj_attribute *attr,
2301 			       const char *buffer, size_t size);
2302 static ssize_t bch_pending_bdevs_cleanup(struct kobject *k,
2303 					 struct kobj_attribute *attr,
2304 					 const char *buffer, size_t size);
2305 
2306 kobj_attribute_write(register,		register_bcache);
2307 kobj_attribute_write(register_quiet,	register_bcache);
2308 kobj_attribute_write(pendings_cleanup,	bch_pending_bdevs_cleanup);
2309 
2310 static bool bch_is_open_backing(struct block_device *bdev)
2311 {
2312 	struct cache_set *c, *tc;
2313 	struct cached_dev *dc, *t;
2314 
2315 	list_for_each_entry_safe(c, tc, &bch_cache_sets, list)
2316 		list_for_each_entry_safe(dc, t, &c->cached_devs, list)
2317 			if (dc->bdev == bdev)
2318 				return true;
2319 	list_for_each_entry_safe(dc, t, &uncached_devices, list)
2320 		if (dc->bdev == bdev)
2321 			return true;
2322 	return false;
2323 }
2324 
2325 static bool bch_is_open_cache(struct block_device *bdev)
2326 {
2327 	struct cache_set *c, *tc;
2328 	struct cache *ca;
2329 	unsigned int i;
2330 
2331 	list_for_each_entry_safe(c, tc, &bch_cache_sets, list)
2332 		for_each_cache(ca, c, i)
2333 			if (ca->bdev == bdev)
2334 				return true;
2335 	return false;
2336 }
2337 
2338 static bool bch_is_open(struct block_device *bdev)
2339 {
2340 	return bch_is_open_cache(bdev) || bch_is_open_backing(bdev);
2341 }
2342 
2343 static ssize_t register_bcache(struct kobject *k, struct kobj_attribute *attr,
2344 			       const char *buffer, size_t size)
2345 {
2346 	ssize_t ret = -EINVAL;
2347 	const char *err = "cannot allocate memory";
2348 	char *path = NULL;
2349 	struct cache_sb *sb = NULL;
2350 	struct block_device *bdev = NULL;
2351 	struct page *sb_page = NULL;
2352 
2353 	if (!try_module_get(THIS_MODULE))
2354 		return -EBUSY;
2355 
2356 	/* For latest state of bcache_is_reboot */
2357 	smp_mb();
2358 	if (bcache_is_reboot)
2359 		return -EBUSY;
2360 
2361 	path = kstrndup(buffer, size, GFP_KERNEL);
2362 	if (!path)
2363 		goto err;
2364 
2365 	sb = kmalloc(sizeof(struct cache_sb), GFP_KERNEL);
2366 	if (!sb)
2367 		goto err;
2368 
2369 	err = "failed to open device";
2370 	bdev = blkdev_get_by_path(strim(path),
2371 				  FMODE_READ|FMODE_WRITE|FMODE_EXCL,
2372 				  sb);
2373 	if (IS_ERR(bdev)) {
2374 		if (bdev == ERR_PTR(-EBUSY)) {
2375 			bdev = lookup_bdev(strim(path));
2376 			mutex_lock(&bch_register_lock);
2377 			if (!IS_ERR(bdev) && bch_is_open(bdev))
2378 				err = "device already registered";
2379 			else
2380 				err = "device busy";
2381 			mutex_unlock(&bch_register_lock);
2382 			if (!IS_ERR(bdev))
2383 				bdput(bdev);
2384 			if (attr == &ksysfs_register_quiet)
2385 				goto quiet_out;
2386 		}
2387 		goto err;
2388 	}
2389 
2390 	err = "failed to set blocksize";
2391 	if (set_blocksize(bdev, 4096))
2392 		goto err_close;
2393 
2394 	err = read_super(sb, bdev, &sb_page);
2395 	if (err)
2396 		goto err_close;
2397 
2398 	err = "failed to register device";
2399 	if (SB_IS_BDEV(sb)) {
2400 		struct cached_dev *dc = kzalloc(sizeof(*dc), GFP_KERNEL);
2401 
2402 		if (!dc)
2403 			goto err_close;
2404 
2405 		mutex_lock(&bch_register_lock);
2406 		ret = register_bdev(sb, sb_page, bdev, dc);
2407 		mutex_unlock(&bch_register_lock);
2408 		/* blkdev_put() will be called in cached_dev_free() */
2409 		if (ret < 0)
2410 			goto err;
2411 	} else {
2412 		struct cache *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
2413 
2414 		if (!ca)
2415 			goto err_close;
2416 
2417 		/* blkdev_put() will be called in bch_cache_release() */
2418 		if (register_cache(sb, sb_page, bdev, ca) != 0)
2419 			goto err;
2420 	}
2421 quiet_out:
2422 	ret = size;
2423 out:
2424 	if (sb_page)
2425 		put_page(sb_page);
2426 	kfree(sb);
2427 	kfree(path);
2428 	module_put(THIS_MODULE);
2429 	return ret;
2430 
2431 err_close:
2432 	blkdev_put(bdev, FMODE_READ|FMODE_WRITE|FMODE_EXCL);
2433 err:
2434 	pr_info("error %s: %s", path, err);
2435 	goto out;
2436 }
2437 
2438 
2439 struct pdev {
2440 	struct list_head list;
2441 	struct cached_dev *dc;
2442 };
2443 
2444 static ssize_t bch_pending_bdevs_cleanup(struct kobject *k,
2445 					 struct kobj_attribute *attr,
2446 					 const char *buffer,
2447 					 size_t size)
2448 {
2449 	LIST_HEAD(pending_devs);
2450 	ssize_t ret = size;
2451 	struct cached_dev *dc, *tdc;
2452 	struct pdev *pdev, *tpdev;
2453 	struct cache_set *c, *tc;
2454 
2455 	mutex_lock(&bch_register_lock);
2456 	list_for_each_entry_safe(dc, tdc, &uncached_devices, list) {
2457 		pdev = kmalloc(sizeof(struct pdev), GFP_KERNEL);
2458 		if (!pdev)
2459 			break;
2460 		pdev->dc = dc;
2461 		list_add(&pdev->list, &pending_devs);
2462 	}
2463 
2464 	list_for_each_entry_safe(pdev, tpdev, &pending_devs, list) {
2465 		list_for_each_entry_safe(c, tc, &bch_cache_sets, list) {
2466 			char *pdev_set_uuid = pdev->dc->sb.set_uuid;
2467 			char *set_uuid = c->sb.uuid;
2468 
2469 			if (!memcmp(pdev_set_uuid, set_uuid, 16)) {
2470 				list_del(&pdev->list);
2471 				kfree(pdev);
2472 				break;
2473 			}
2474 		}
2475 	}
2476 	mutex_unlock(&bch_register_lock);
2477 
2478 	list_for_each_entry_safe(pdev, tpdev, &pending_devs, list) {
2479 		pr_info("delete pdev %p", pdev);
2480 		list_del(&pdev->list);
2481 		bcache_device_stop(&pdev->dc->disk);
2482 		kfree(pdev);
2483 	}
2484 
2485 	return ret;
2486 }
2487 
2488 static int bcache_reboot(struct notifier_block *n, unsigned long code, void *x)
2489 {
2490 	if (bcache_is_reboot)
2491 		return NOTIFY_DONE;
2492 
2493 	if (code == SYS_DOWN ||
2494 	    code == SYS_HALT ||
2495 	    code == SYS_POWER_OFF) {
2496 		DEFINE_WAIT(wait);
2497 		unsigned long start = jiffies;
2498 		bool stopped = false;
2499 
2500 		struct cache_set *c, *tc;
2501 		struct cached_dev *dc, *tdc;
2502 
2503 		mutex_lock(&bch_register_lock);
2504 
2505 		if (bcache_is_reboot)
2506 			goto out;
2507 
2508 		/* New registration is rejected since now */
2509 		bcache_is_reboot = true;
2510 		/*
2511 		 * Make registering caller (if there is) on other CPU
2512 		 * core know bcache_is_reboot set to true earlier
2513 		 */
2514 		smp_mb();
2515 
2516 		if (list_empty(&bch_cache_sets) &&
2517 		    list_empty(&uncached_devices))
2518 			goto out;
2519 
2520 		mutex_unlock(&bch_register_lock);
2521 
2522 		pr_info("Stopping all devices:");
2523 
2524 		/*
2525 		 * The reason bch_register_lock is not held to call
2526 		 * bch_cache_set_stop() and bcache_device_stop() is to
2527 		 * avoid potential deadlock during reboot, because cache
2528 		 * set or bcache device stopping process will acqurie
2529 		 * bch_register_lock too.
2530 		 *
2531 		 * We are safe here because bcache_is_reboot sets to
2532 		 * true already, register_bcache() will reject new
2533 		 * registration now. bcache_is_reboot also makes sure
2534 		 * bcache_reboot() won't be re-entered on by other thread,
2535 		 * so there is no race in following list iteration by
2536 		 * list_for_each_entry_safe().
2537 		 */
2538 		list_for_each_entry_safe(c, tc, &bch_cache_sets, list)
2539 			bch_cache_set_stop(c);
2540 
2541 		list_for_each_entry_safe(dc, tdc, &uncached_devices, list)
2542 			bcache_device_stop(&dc->disk);
2543 
2544 
2545 		/*
2546 		 * Give an early chance for other kthreads and
2547 		 * kworkers to stop themselves
2548 		 */
2549 		schedule();
2550 
2551 		/* What's a condition variable? */
2552 		while (1) {
2553 			long timeout = start + 10 * HZ - jiffies;
2554 
2555 			mutex_lock(&bch_register_lock);
2556 			stopped = list_empty(&bch_cache_sets) &&
2557 				list_empty(&uncached_devices);
2558 
2559 			if (timeout < 0 || stopped)
2560 				break;
2561 
2562 			prepare_to_wait(&unregister_wait, &wait,
2563 					TASK_UNINTERRUPTIBLE);
2564 
2565 			mutex_unlock(&bch_register_lock);
2566 			schedule_timeout(timeout);
2567 		}
2568 
2569 		finish_wait(&unregister_wait, &wait);
2570 
2571 		if (stopped)
2572 			pr_info("All devices stopped");
2573 		else
2574 			pr_notice("Timeout waiting for devices to be closed");
2575 out:
2576 		mutex_unlock(&bch_register_lock);
2577 	}
2578 
2579 	return NOTIFY_DONE;
2580 }
2581 
2582 static struct notifier_block reboot = {
2583 	.notifier_call	= bcache_reboot,
2584 	.priority	= INT_MAX, /* before any real devices */
2585 };
2586 
2587 static void bcache_exit(void)
2588 {
2589 	bch_debug_exit();
2590 	bch_request_exit();
2591 	if (bcache_kobj)
2592 		kobject_put(bcache_kobj);
2593 	if (bcache_wq)
2594 		destroy_workqueue(bcache_wq);
2595 	if (bch_journal_wq)
2596 		destroy_workqueue(bch_journal_wq);
2597 
2598 	if (bcache_major)
2599 		unregister_blkdev(bcache_major, "bcache");
2600 	unregister_reboot_notifier(&reboot);
2601 	mutex_destroy(&bch_register_lock);
2602 }
2603 
2604 /* Check and fixup module parameters */
2605 static void check_module_parameters(void)
2606 {
2607 	if (bch_cutoff_writeback_sync == 0)
2608 		bch_cutoff_writeback_sync = CUTOFF_WRITEBACK_SYNC;
2609 	else if (bch_cutoff_writeback_sync > CUTOFF_WRITEBACK_SYNC_MAX) {
2610 		pr_warn("set bch_cutoff_writeback_sync (%u) to max value %u",
2611 			bch_cutoff_writeback_sync, CUTOFF_WRITEBACK_SYNC_MAX);
2612 		bch_cutoff_writeback_sync = CUTOFF_WRITEBACK_SYNC_MAX;
2613 	}
2614 
2615 	if (bch_cutoff_writeback == 0)
2616 		bch_cutoff_writeback = CUTOFF_WRITEBACK;
2617 	else if (bch_cutoff_writeback > CUTOFF_WRITEBACK_MAX) {
2618 		pr_warn("set bch_cutoff_writeback (%u) to max value %u",
2619 			bch_cutoff_writeback, CUTOFF_WRITEBACK_MAX);
2620 		bch_cutoff_writeback = CUTOFF_WRITEBACK_MAX;
2621 	}
2622 
2623 	if (bch_cutoff_writeback > bch_cutoff_writeback_sync) {
2624 		pr_warn("set bch_cutoff_writeback (%u) to %u",
2625 			bch_cutoff_writeback, bch_cutoff_writeback_sync);
2626 		bch_cutoff_writeback = bch_cutoff_writeback_sync;
2627 	}
2628 }
2629 
2630 static int __init bcache_init(void)
2631 {
2632 	static const struct attribute *files[] = {
2633 		&ksysfs_register.attr,
2634 		&ksysfs_register_quiet.attr,
2635 		&ksysfs_pendings_cleanup.attr,
2636 		NULL
2637 	};
2638 
2639 	check_module_parameters();
2640 
2641 	mutex_init(&bch_register_lock);
2642 	init_waitqueue_head(&unregister_wait);
2643 	register_reboot_notifier(&reboot);
2644 
2645 	bcache_major = register_blkdev(0, "bcache");
2646 	if (bcache_major < 0) {
2647 		unregister_reboot_notifier(&reboot);
2648 		mutex_destroy(&bch_register_lock);
2649 		return bcache_major;
2650 	}
2651 
2652 	bcache_wq = alloc_workqueue("bcache", WQ_MEM_RECLAIM, 0);
2653 	if (!bcache_wq)
2654 		goto err;
2655 
2656 	bch_journal_wq = alloc_workqueue("bch_journal", WQ_MEM_RECLAIM, 0);
2657 	if (!bch_journal_wq)
2658 		goto err;
2659 
2660 	bcache_kobj = kobject_create_and_add("bcache", fs_kobj);
2661 	if (!bcache_kobj)
2662 		goto err;
2663 
2664 	if (bch_request_init() ||
2665 	    sysfs_create_files(bcache_kobj, files))
2666 		goto err;
2667 
2668 	bch_debug_init();
2669 	closure_debug_init();
2670 
2671 	bcache_is_reboot = false;
2672 
2673 	return 0;
2674 err:
2675 	bcache_exit();
2676 	return -ENOMEM;
2677 }
2678 
2679 /*
2680  * Module hooks
2681  */
2682 module_exit(bcache_exit);
2683 module_init(bcache_init);
2684 
2685 module_param(bch_cutoff_writeback, uint, 0);
2686 MODULE_PARM_DESC(bch_cutoff_writeback, "threshold to cutoff writeback");
2687 
2688 module_param(bch_cutoff_writeback_sync, uint, 0);
2689 MODULE_PARM_DESC(bch_cutoff_writeback_sync, "hard threshold to cutoff writeback");
2690 
2691 MODULE_DESCRIPTION("Bcache: a Linux block layer cache");
2692 MODULE_AUTHOR("Kent Overstreet <kent.overstreet@gmail.com>");
2693 MODULE_LICENSE("GPL");
2694