xref: /linux/drivers/md/bcache/super.c (revision 37744feebc086908fd89760650f458ab19071750)
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/debugfs.h>
19 #include <linux/genhd.h>
20 #include <linux/idr.h>
21 #include <linux/kthread.h>
22 #include <linux/module.h>
23 #include <linux/random.h>
24 #include <linux/reboot.h>
25 #include <linux/sysfs.h>
26 
27 unsigned int bch_cutoff_writeback;
28 unsigned int bch_cutoff_writeback_sync;
29 
30 static const char bcache_magic[] = {
31 	0xc6, 0x85, 0x73, 0xf6, 0x4e, 0x1a, 0x45, 0xca,
32 	0x82, 0x65, 0xf5, 0x7f, 0x48, 0xba, 0x6d, 0x81
33 };
34 
35 static const char invalid_uuid[] = {
36 	0xa0, 0x3e, 0xf8, 0xed, 0x3e, 0xe1, 0xb8, 0x78,
37 	0xc8, 0x50, 0xfc, 0x5e, 0xcb, 0x16, 0xcd, 0x99
38 };
39 
40 static struct kobject *bcache_kobj;
41 struct mutex bch_register_lock;
42 bool bcache_is_reboot;
43 LIST_HEAD(bch_cache_sets);
44 static LIST_HEAD(uncached_devices);
45 
46 static int bcache_major;
47 static DEFINE_IDA(bcache_device_idx);
48 static wait_queue_head_t unregister_wait;
49 struct workqueue_struct *bcache_wq;
50 struct workqueue_struct *bch_journal_wq;
51 
52 
53 #define BTREE_MAX_PAGES		(256 * 1024 / PAGE_SIZE)
54 /* limitation of partitions number on single bcache device */
55 #define BCACHE_MINORS		128
56 /* limitation of bcache devices number on single system */
57 #define BCACHE_DEVICE_IDX_MAX	((1U << MINORBITS)/BCACHE_MINORS)
58 
59 /* Superblock */
60 
61 static const char *read_super(struct cache_sb *sb, struct block_device *bdev,
62 			      struct cache_sb_disk **res)
63 {
64 	const char *err;
65 	struct cache_sb_disk *s;
66 	struct page *page;
67 	unsigned int i;
68 
69 	page = read_cache_page_gfp(bdev->bd_inode->i_mapping,
70 				   SB_OFFSET >> PAGE_SHIFT, GFP_KERNEL);
71 	if (IS_ERR(page))
72 		return "IO error";
73 	s = page_address(page) + offset_in_page(SB_OFFSET);
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 (bad offset)";
96 	if (sb->offset != SB_SECTOR)
97 		goto err;
98 
99 	err = "Not a bcache superblock (bad magic)";
100 	if (memcmp(sb->magic, bcache_magic, 16))
101 		goto err;
102 
103 	err = "Too many journal buckets";
104 	if (sb->keys > SB_JOURNAL_BUCKETS)
105 		goto err;
106 
107 	err = "Bad checksum";
108 	if (s->csum != csum_set(s))
109 		goto err;
110 
111 	err = "Bad UUID";
112 	if (bch_is_zero(sb->uuid, 16))
113 		goto err;
114 
115 	sb->block_size	= le16_to_cpu(s->block_size);
116 
117 	err = "Superblock block size smaller than device block size";
118 	if (sb->block_size << 9 < bdev_logical_block_size(bdev))
119 		goto err;
120 
121 	switch (sb->version) {
122 	case BCACHE_SB_VERSION_BDEV:
123 		sb->data_offset	= BDEV_DATA_START_DEFAULT;
124 		break;
125 	case BCACHE_SB_VERSION_BDEV_WITH_OFFSET:
126 		sb->data_offset	= le64_to_cpu(s->data_offset);
127 
128 		err = "Bad data offset";
129 		if (sb->data_offset < BDEV_DATA_START_DEFAULT)
130 			goto err;
131 
132 		break;
133 	case BCACHE_SB_VERSION_CDEV:
134 	case BCACHE_SB_VERSION_CDEV_WITH_UUID:
135 		sb->nbuckets	= le64_to_cpu(s->nbuckets);
136 		sb->bucket_size	= le16_to_cpu(s->bucket_size);
137 
138 		sb->nr_in_set	= le16_to_cpu(s->nr_in_set);
139 		sb->nr_this_dev	= le16_to_cpu(s->nr_this_dev);
140 
141 		err = "Too many buckets";
142 		if (sb->nbuckets > LONG_MAX)
143 			goto err;
144 
145 		err = "Not enough buckets";
146 		if (sb->nbuckets < 1 << 7)
147 			goto err;
148 
149 		err = "Bad block/bucket size";
150 		if (!is_power_of_2(sb->block_size) ||
151 		    sb->block_size > PAGE_SECTORS ||
152 		    !is_power_of_2(sb->bucket_size) ||
153 		    sb->bucket_size < PAGE_SECTORS)
154 			goto err;
155 
156 		err = "Invalid superblock: device too small";
157 		if (get_capacity(bdev->bd_disk) <
158 		    sb->bucket_size * sb->nbuckets)
159 			goto err;
160 
161 		err = "Bad UUID";
162 		if (bch_is_zero(sb->set_uuid, 16))
163 			goto err;
164 
165 		err = "Bad cache device number in set";
166 		if (!sb->nr_in_set ||
167 		    sb->nr_in_set <= sb->nr_this_dev ||
168 		    sb->nr_in_set > MAX_CACHES_PER_SET)
169 			goto err;
170 
171 		err = "Journal buckets not sequential";
172 		for (i = 0; i < sb->keys; i++)
173 			if (sb->d[i] != sb->first_bucket + i)
174 				goto err;
175 
176 		err = "Too many journal buckets";
177 		if (sb->first_bucket + sb->keys > sb->nbuckets)
178 			goto err;
179 
180 		err = "Invalid superblock: first bucket comes before end of super";
181 		if (sb->first_bucket * sb->bucket_size < 16)
182 			goto err;
183 
184 		break;
185 	default:
186 		err = "Unsupported superblock version";
187 		goto err;
188 	}
189 
190 	sb->last_mount = (u32)ktime_get_real_seconds();
191 	*res = s;
192 	return NULL;
193 err:
194 	put_page(page);
195 	return err;
196 }
197 
198 static void write_bdev_super_endio(struct bio *bio)
199 {
200 	struct cached_dev *dc = bio->bi_private;
201 
202 	if (bio->bi_status)
203 		bch_count_backing_io_errors(dc, bio);
204 
205 	closure_put(&dc->sb_write);
206 }
207 
208 static void __write_super(struct cache_sb *sb, struct cache_sb_disk *out,
209 		struct bio *bio)
210 {
211 	unsigned int i;
212 
213 	bio->bi_opf = REQ_OP_WRITE | REQ_SYNC | REQ_META;
214 	bio->bi_iter.bi_sector	= SB_SECTOR;
215 	__bio_add_page(bio, virt_to_page(out), SB_SIZE,
216 			offset_in_page(out));
217 
218 	out->offset		= cpu_to_le64(sb->offset);
219 	out->version		= cpu_to_le64(sb->version);
220 
221 	memcpy(out->uuid,	sb->uuid, 16);
222 	memcpy(out->set_uuid,	sb->set_uuid, 16);
223 	memcpy(out->label,	sb->label, SB_LABEL_SIZE);
224 
225 	out->flags		= cpu_to_le64(sb->flags);
226 	out->seq		= cpu_to_le64(sb->seq);
227 
228 	out->last_mount		= cpu_to_le32(sb->last_mount);
229 	out->first_bucket	= cpu_to_le16(sb->first_bucket);
230 	out->keys		= cpu_to_le16(sb->keys);
231 
232 	for (i = 0; i < sb->keys; i++)
233 		out->d[i] = cpu_to_le64(sb->d[i]);
234 
235 	out->csum = csum_set(out);
236 
237 	pr_debug("ver %llu, flags %llu, seq %llu",
238 		 sb->version, sb->flags, sb->seq);
239 
240 	submit_bio(bio);
241 }
242 
243 static void bch_write_bdev_super_unlock(struct closure *cl)
244 {
245 	struct cached_dev *dc = container_of(cl, struct cached_dev, sb_write);
246 
247 	up(&dc->sb_write_mutex);
248 }
249 
250 void bch_write_bdev_super(struct cached_dev *dc, struct closure *parent)
251 {
252 	struct closure *cl = &dc->sb_write;
253 	struct bio *bio = &dc->sb_bio;
254 
255 	down(&dc->sb_write_mutex);
256 	closure_init(cl, parent);
257 
258 	bio_init(bio, dc->sb_bv, 1);
259 	bio_set_dev(bio, dc->bdev);
260 	bio->bi_end_io	= write_bdev_super_endio;
261 	bio->bi_private = dc;
262 
263 	closure_get(cl);
264 	/* I/O request sent to backing device */
265 	__write_super(&dc->sb, dc->sb_disk, bio);
266 
267 	closure_return_with_destructor(cl, bch_write_bdev_super_unlock);
268 }
269 
270 static void write_super_endio(struct bio *bio)
271 {
272 	struct cache *ca = bio->bi_private;
273 
274 	/* is_read = 0 */
275 	bch_count_io_errors(ca, bio->bi_status, 0,
276 			    "writing superblock");
277 	closure_put(&ca->set->sb_write);
278 }
279 
280 static void bcache_write_super_unlock(struct closure *cl)
281 {
282 	struct cache_set *c = container_of(cl, struct cache_set, sb_write);
283 
284 	up(&c->sb_write_mutex);
285 }
286 
287 void bcache_write_super(struct cache_set *c)
288 {
289 	struct closure *cl = &c->sb_write;
290 	struct cache *ca;
291 	unsigned int i;
292 
293 	down(&c->sb_write_mutex);
294 	closure_init(cl, &c->cl);
295 
296 	c->sb.seq++;
297 
298 	for_each_cache(ca, c, i) {
299 		struct bio *bio = &ca->sb_bio;
300 
301 		ca->sb.version		= BCACHE_SB_VERSION_CDEV_WITH_UUID;
302 		ca->sb.seq		= c->sb.seq;
303 		ca->sb.last_mount	= c->sb.last_mount;
304 
305 		SET_CACHE_SYNC(&ca->sb, CACHE_SYNC(&c->sb));
306 
307 		bio_init(bio, ca->sb_bv, 1);
308 		bio_set_dev(bio, ca->bdev);
309 		bio->bi_end_io	= write_super_endio;
310 		bio->bi_private = ca;
311 
312 		closure_get(cl);
313 		__write_super(&ca->sb, ca->sb_disk, bio);
314 	}
315 
316 	closure_return_with_destructor(cl, bcache_write_super_unlock);
317 }
318 
319 /* UUID io */
320 
321 static void uuid_endio(struct bio *bio)
322 {
323 	struct closure *cl = bio->bi_private;
324 	struct cache_set *c = container_of(cl, struct cache_set, uuid_write);
325 
326 	cache_set_err_on(bio->bi_status, c, "accessing uuids");
327 	bch_bbio_free(bio, c);
328 	closure_put(cl);
329 }
330 
331 static void uuid_io_unlock(struct closure *cl)
332 {
333 	struct cache_set *c = container_of(cl, struct cache_set, uuid_write);
334 
335 	up(&c->uuid_write_mutex);
336 }
337 
338 static void uuid_io(struct cache_set *c, int op, unsigned long op_flags,
339 		    struct bkey *k, struct closure *parent)
340 {
341 	struct closure *cl = &c->uuid_write;
342 	struct uuid_entry *u;
343 	unsigned int i;
344 	char buf[80];
345 
346 	BUG_ON(!parent);
347 	down(&c->uuid_write_mutex);
348 	closure_init(cl, parent);
349 
350 	for (i = 0; i < KEY_PTRS(k); i++) {
351 		struct bio *bio = bch_bbio_alloc(c);
352 
353 		bio->bi_opf = REQ_SYNC | REQ_META | op_flags;
354 		bio->bi_iter.bi_size = KEY_SIZE(k) << 9;
355 
356 		bio->bi_end_io	= uuid_endio;
357 		bio->bi_private = cl;
358 		bio_set_op_attrs(bio, op, REQ_SYNC|REQ_META|op_flags);
359 		bch_bio_map(bio, c->uuids);
360 
361 		bch_submit_bbio(bio, c, k, i);
362 
363 		if (op != REQ_OP_WRITE)
364 			break;
365 	}
366 
367 	bch_extent_to_text(buf, sizeof(buf), k);
368 	pr_debug("%s UUIDs at %s", op == REQ_OP_WRITE ? "wrote" : "read", buf);
369 
370 	for (u = c->uuids; u < c->uuids + c->nr_uuids; u++)
371 		if (!bch_is_zero(u->uuid, 16))
372 			pr_debug("Slot %zi: %pU: %s: 1st: %u last: %u inv: %u",
373 				 u - c->uuids, u->uuid, u->label,
374 				 u->first_reg, u->last_reg, u->invalidated);
375 
376 	closure_return_with_destructor(cl, uuid_io_unlock);
377 }
378 
379 static char *uuid_read(struct cache_set *c, struct jset *j, struct closure *cl)
380 {
381 	struct bkey *k = &j->uuid_bucket;
382 
383 	if (__bch_btree_ptr_invalid(c, k))
384 		return "bad uuid pointer";
385 
386 	bkey_copy(&c->uuid_bucket, k);
387 	uuid_io(c, REQ_OP_READ, 0, k, cl);
388 
389 	if (j->version < BCACHE_JSET_VERSION_UUIDv1) {
390 		struct uuid_entry_v0	*u0 = (void *) c->uuids;
391 		struct uuid_entry	*u1 = (void *) c->uuids;
392 		int i;
393 
394 		closure_sync(cl);
395 
396 		/*
397 		 * Since the new uuid entry is bigger than the old, we have to
398 		 * convert starting at the highest memory address and work down
399 		 * in order to do it in place
400 		 */
401 
402 		for (i = c->nr_uuids - 1;
403 		     i >= 0;
404 		     --i) {
405 			memcpy(u1[i].uuid,	u0[i].uuid, 16);
406 			memcpy(u1[i].label,	u0[i].label, 32);
407 
408 			u1[i].first_reg		= u0[i].first_reg;
409 			u1[i].last_reg		= u0[i].last_reg;
410 			u1[i].invalidated	= u0[i].invalidated;
411 
412 			u1[i].flags	= 0;
413 			u1[i].sectors	= 0;
414 		}
415 	}
416 
417 	return NULL;
418 }
419 
420 static int __uuid_write(struct cache_set *c)
421 {
422 	BKEY_PADDED(key) k;
423 	struct closure cl;
424 	struct cache *ca;
425 
426 	closure_init_stack(&cl);
427 	lockdep_assert_held(&bch_register_lock);
428 
429 	if (bch_bucket_alloc_set(c, RESERVE_BTREE, &k.key, 1, true))
430 		return 1;
431 
432 	SET_KEY_SIZE(&k.key, c->sb.bucket_size);
433 	uuid_io(c, REQ_OP_WRITE, 0, &k.key, &cl);
434 	closure_sync(&cl);
435 
436 	/* Only one bucket used for uuid write */
437 	ca = PTR_CACHE(c, &k.key, 0);
438 	atomic_long_add(ca->sb.bucket_size, &ca->meta_sectors_written);
439 
440 	bkey_copy(&c->uuid_bucket, &k.key);
441 	bkey_put(c, &k.key);
442 	return 0;
443 }
444 
445 int bch_uuid_write(struct cache_set *c)
446 {
447 	int ret = __uuid_write(c);
448 
449 	if (!ret)
450 		bch_journal_meta(c, NULL);
451 
452 	return ret;
453 }
454 
455 static struct uuid_entry *uuid_find(struct cache_set *c, const char *uuid)
456 {
457 	struct uuid_entry *u;
458 
459 	for (u = c->uuids;
460 	     u < c->uuids + c->nr_uuids; u++)
461 		if (!memcmp(u->uuid, uuid, 16))
462 			return u;
463 
464 	return NULL;
465 }
466 
467 static struct uuid_entry *uuid_find_empty(struct cache_set *c)
468 {
469 	static const char zero_uuid[16] = "\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0";
470 
471 	return uuid_find(c, zero_uuid);
472 }
473 
474 /*
475  * Bucket priorities/gens:
476  *
477  * For each bucket, we store on disk its
478  *   8 bit gen
479  *  16 bit priority
480  *
481  * See alloc.c for an explanation of the gen. The priority is used to implement
482  * lru (and in the future other) cache replacement policies; for most purposes
483  * it's just an opaque integer.
484  *
485  * The gens and the priorities don't have a whole lot to do with each other, and
486  * it's actually the gens that must be written out at specific times - it's no
487  * big deal if the priorities don't get written, if we lose them we just reuse
488  * buckets in suboptimal order.
489  *
490  * On disk they're stored in a packed array, and in as many buckets are required
491  * to fit them all. The buckets we use to store them form a list; the journal
492  * header points to the first bucket, the first bucket points to the second
493  * bucket, et cetera.
494  *
495  * This code is used by the allocation code; periodically (whenever it runs out
496  * of buckets to allocate from) the allocation code will invalidate some
497  * buckets, but it can't use those buckets until their new gens are safely on
498  * disk.
499  */
500 
501 static void prio_endio(struct bio *bio)
502 {
503 	struct cache *ca = bio->bi_private;
504 
505 	cache_set_err_on(bio->bi_status, ca->set, "accessing priorities");
506 	bch_bbio_free(bio, ca->set);
507 	closure_put(&ca->prio);
508 }
509 
510 static void prio_io(struct cache *ca, uint64_t bucket, int op,
511 		    unsigned long op_flags)
512 {
513 	struct closure *cl = &ca->prio;
514 	struct bio *bio = bch_bbio_alloc(ca->set);
515 
516 	closure_init_stack(cl);
517 
518 	bio->bi_iter.bi_sector	= bucket * ca->sb.bucket_size;
519 	bio_set_dev(bio, ca->bdev);
520 	bio->bi_iter.bi_size	= bucket_bytes(ca);
521 
522 	bio->bi_end_io	= prio_endio;
523 	bio->bi_private = ca;
524 	bio_set_op_attrs(bio, op, REQ_SYNC|REQ_META|op_flags);
525 	bch_bio_map(bio, ca->disk_buckets);
526 
527 	closure_bio_submit(ca->set, bio, &ca->prio);
528 	closure_sync(cl);
529 }
530 
531 int bch_prio_write(struct cache *ca, bool wait)
532 {
533 	int i;
534 	struct bucket *b;
535 	struct closure cl;
536 
537 	pr_debug("free_prio=%zu, free_none=%zu, free_inc=%zu",
538 		 fifo_used(&ca->free[RESERVE_PRIO]),
539 		 fifo_used(&ca->free[RESERVE_NONE]),
540 		 fifo_used(&ca->free_inc));
541 
542 	/*
543 	 * Pre-check if there are enough free buckets. In the non-blocking
544 	 * scenario it's better to fail early rather than starting to allocate
545 	 * buckets and do a cleanup later in case of failure.
546 	 */
547 	if (!wait) {
548 		size_t avail = fifo_used(&ca->free[RESERVE_PRIO]) +
549 			       fifo_used(&ca->free[RESERVE_NONE]);
550 		if (prio_buckets(ca) > avail)
551 			return -ENOMEM;
552 	}
553 
554 	closure_init_stack(&cl);
555 
556 	lockdep_assert_held(&ca->set->bucket_lock);
557 
558 	ca->disk_buckets->seq++;
559 
560 	atomic_long_add(ca->sb.bucket_size * prio_buckets(ca),
561 			&ca->meta_sectors_written);
562 
563 	for (i = prio_buckets(ca) - 1; i >= 0; --i) {
564 		long bucket;
565 		struct prio_set *p = ca->disk_buckets;
566 		struct bucket_disk *d = p->data;
567 		struct bucket_disk *end = d + prios_per_bucket(ca);
568 
569 		for (b = ca->buckets + i * prios_per_bucket(ca);
570 		     b < ca->buckets + ca->sb.nbuckets && d < end;
571 		     b++, d++) {
572 			d->prio = cpu_to_le16(b->prio);
573 			d->gen = b->gen;
574 		}
575 
576 		p->next_bucket	= ca->prio_buckets[i + 1];
577 		p->magic	= pset_magic(&ca->sb);
578 		p->csum		= bch_crc64(&p->magic, bucket_bytes(ca) - 8);
579 
580 		bucket = bch_bucket_alloc(ca, RESERVE_PRIO, wait);
581 		BUG_ON(bucket == -1);
582 
583 		mutex_unlock(&ca->set->bucket_lock);
584 		prio_io(ca, bucket, REQ_OP_WRITE, 0);
585 		mutex_lock(&ca->set->bucket_lock);
586 
587 		ca->prio_buckets[i] = bucket;
588 		atomic_dec_bug(&ca->buckets[bucket].pin);
589 	}
590 
591 	mutex_unlock(&ca->set->bucket_lock);
592 
593 	bch_journal_meta(ca->set, &cl);
594 	closure_sync(&cl);
595 
596 	mutex_lock(&ca->set->bucket_lock);
597 
598 	/*
599 	 * Don't want the old priorities to get garbage collected until after we
600 	 * finish writing the new ones, and they're journalled
601 	 */
602 	for (i = 0; i < prio_buckets(ca); i++) {
603 		if (ca->prio_last_buckets[i])
604 			__bch_bucket_free(ca,
605 				&ca->buckets[ca->prio_last_buckets[i]]);
606 
607 		ca->prio_last_buckets[i] = ca->prio_buckets[i];
608 	}
609 	return 0;
610 }
611 
612 static int prio_read(struct cache *ca, uint64_t bucket)
613 {
614 	struct prio_set *p = ca->disk_buckets;
615 	struct bucket_disk *d = p->data + prios_per_bucket(ca), *end = d;
616 	struct bucket *b;
617 	unsigned int bucket_nr = 0;
618 	int ret = -EIO;
619 
620 	for (b = ca->buckets;
621 	     b < ca->buckets + ca->sb.nbuckets;
622 	     b++, d++) {
623 		if (d == end) {
624 			ca->prio_buckets[bucket_nr] = bucket;
625 			ca->prio_last_buckets[bucket_nr] = bucket;
626 			bucket_nr++;
627 
628 			prio_io(ca, bucket, REQ_OP_READ, 0);
629 
630 			if (p->csum !=
631 			    bch_crc64(&p->magic, bucket_bytes(ca) - 8)) {
632 				pr_warn("bad csum reading priorities");
633 				goto out;
634 			}
635 
636 			if (p->magic != pset_magic(&ca->sb)) {
637 				pr_warn("bad magic reading priorities");
638 				goto out;
639 			}
640 
641 			bucket = p->next_bucket;
642 			d = p->data;
643 		}
644 
645 		b->prio = le16_to_cpu(d->prio);
646 		b->gen = b->last_gc = d->gen;
647 	}
648 
649 	ret = 0;
650 out:
651 	return ret;
652 }
653 
654 /* Bcache device */
655 
656 static int open_dev(struct block_device *b, fmode_t mode)
657 {
658 	struct bcache_device *d = b->bd_disk->private_data;
659 
660 	if (test_bit(BCACHE_DEV_CLOSING, &d->flags))
661 		return -ENXIO;
662 
663 	closure_get(&d->cl);
664 	return 0;
665 }
666 
667 static void release_dev(struct gendisk *b, fmode_t mode)
668 {
669 	struct bcache_device *d = b->private_data;
670 
671 	closure_put(&d->cl);
672 }
673 
674 static int ioctl_dev(struct block_device *b, fmode_t mode,
675 		     unsigned int cmd, unsigned long arg)
676 {
677 	struct bcache_device *d = b->bd_disk->private_data;
678 
679 	return d->ioctl(d, mode, cmd, arg);
680 }
681 
682 static const struct block_device_operations bcache_ops = {
683 	.open		= open_dev,
684 	.release	= release_dev,
685 	.ioctl		= ioctl_dev,
686 	.owner		= THIS_MODULE,
687 };
688 
689 void bcache_device_stop(struct bcache_device *d)
690 {
691 	if (!test_and_set_bit(BCACHE_DEV_CLOSING, &d->flags))
692 		/*
693 		 * closure_fn set to
694 		 * - cached device: cached_dev_flush()
695 		 * - flash dev: flash_dev_flush()
696 		 */
697 		closure_queue(&d->cl);
698 }
699 
700 static void bcache_device_unlink(struct bcache_device *d)
701 {
702 	lockdep_assert_held(&bch_register_lock);
703 
704 	if (d->c && !test_and_set_bit(BCACHE_DEV_UNLINK_DONE, &d->flags)) {
705 		unsigned int i;
706 		struct cache *ca;
707 
708 		sysfs_remove_link(&d->c->kobj, d->name);
709 		sysfs_remove_link(&d->kobj, "cache");
710 
711 		for_each_cache(ca, d->c, i)
712 			bd_unlink_disk_holder(ca->bdev, d->disk);
713 	}
714 }
715 
716 static void bcache_device_link(struct bcache_device *d, struct cache_set *c,
717 			       const char *name)
718 {
719 	unsigned int i;
720 	struct cache *ca;
721 	int ret;
722 
723 	for_each_cache(ca, d->c, i)
724 		bd_link_disk_holder(ca->bdev, d->disk);
725 
726 	snprintf(d->name, BCACHEDEVNAME_SIZE,
727 		 "%s%u", name, d->id);
728 
729 	ret = sysfs_create_link(&d->kobj, &c->kobj, "cache");
730 	if (ret < 0)
731 		pr_err("Couldn't create device -> cache set symlink");
732 
733 	ret = sysfs_create_link(&c->kobj, &d->kobj, d->name);
734 	if (ret < 0)
735 		pr_err("Couldn't create cache set -> device symlink");
736 
737 	clear_bit(BCACHE_DEV_UNLINK_DONE, &d->flags);
738 }
739 
740 static void bcache_device_detach(struct bcache_device *d)
741 {
742 	lockdep_assert_held(&bch_register_lock);
743 
744 	atomic_dec(&d->c->attached_dev_nr);
745 
746 	if (test_bit(BCACHE_DEV_DETACHING, &d->flags)) {
747 		struct uuid_entry *u = d->c->uuids + d->id;
748 
749 		SET_UUID_FLASH_ONLY(u, 0);
750 		memcpy(u->uuid, invalid_uuid, 16);
751 		u->invalidated = cpu_to_le32((u32)ktime_get_real_seconds());
752 		bch_uuid_write(d->c);
753 	}
754 
755 	bcache_device_unlink(d);
756 
757 	d->c->devices[d->id] = NULL;
758 	closure_put(&d->c->caching);
759 	d->c = NULL;
760 }
761 
762 static void bcache_device_attach(struct bcache_device *d, struct cache_set *c,
763 				 unsigned int id)
764 {
765 	d->id = id;
766 	d->c = c;
767 	c->devices[id] = d;
768 
769 	if (id >= c->devices_max_used)
770 		c->devices_max_used = id + 1;
771 
772 	closure_get(&c->caching);
773 }
774 
775 static inline int first_minor_to_idx(int first_minor)
776 {
777 	return (first_minor/BCACHE_MINORS);
778 }
779 
780 static inline int idx_to_first_minor(int idx)
781 {
782 	return (idx * BCACHE_MINORS);
783 }
784 
785 static void bcache_device_free(struct bcache_device *d)
786 {
787 	struct gendisk *disk = d->disk;
788 
789 	lockdep_assert_held(&bch_register_lock);
790 
791 	if (disk)
792 		pr_info("%s stopped", disk->disk_name);
793 	else
794 		pr_err("bcache device (NULL gendisk) stopped");
795 
796 	if (d->c)
797 		bcache_device_detach(d);
798 
799 	if (disk) {
800 		if (disk->flags & GENHD_FL_UP)
801 			del_gendisk(disk);
802 
803 		if (disk->queue)
804 			blk_cleanup_queue(disk->queue);
805 
806 		ida_simple_remove(&bcache_device_idx,
807 				  first_minor_to_idx(disk->first_minor));
808 		put_disk(disk);
809 	}
810 
811 	bioset_exit(&d->bio_split);
812 	kvfree(d->full_dirty_stripes);
813 	kvfree(d->stripe_sectors_dirty);
814 
815 	closure_debug_destroy(&d->cl);
816 }
817 
818 static int bcache_device_init(struct bcache_device *d, unsigned int block_size,
819 			      sector_t sectors, make_request_fn make_request_fn)
820 {
821 	struct request_queue *q;
822 	const size_t max_stripes = min_t(size_t, INT_MAX,
823 					 SIZE_MAX / sizeof(atomic_t));
824 	size_t n;
825 	int idx;
826 
827 	if (!d->stripe_size)
828 		d->stripe_size = 1 << 31;
829 
830 	d->nr_stripes = DIV_ROUND_UP_ULL(sectors, d->stripe_size);
831 
832 	if (!d->nr_stripes || d->nr_stripes > max_stripes) {
833 		pr_err("nr_stripes too large or invalid: %u (start sector beyond end of disk?)",
834 			(unsigned int)d->nr_stripes);
835 		return -ENOMEM;
836 	}
837 
838 	n = d->nr_stripes * sizeof(atomic_t);
839 	d->stripe_sectors_dirty = kvzalloc(n, GFP_KERNEL);
840 	if (!d->stripe_sectors_dirty)
841 		return -ENOMEM;
842 
843 	n = BITS_TO_LONGS(d->nr_stripes) * sizeof(unsigned long);
844 	d->full_dirty_stripes = kvzalloc(n, GFP_KERNEL);
845 	if (!d->full_dirty_stripes)
846 		return -ENOMEM;
847 
848 	idx = ida_simple_get(&bcache_device_idx, 0,
849 				BCACHE_DEVICE_IDX_MAX, GFP_KERNEL);
850 	if (idx < 0)
851 		return idx;
852 
853 	if (bioset_init(&d->bio_split, 4, offsetof(struct bbio, bio),
854 			BIOSET_NEED_BVECS|BIOSET_NEED_RESCUER))
855 		goto err;
856 
857 	d->disk = alloc_disk(BCACHE_MINORS);
858 	if (!d->disk)
859 		goto err;
860 
861 	set_capacity(d->disk, sectors);
862 	snprintf(d->disk->disk_name, DISK_NAME_LEN, "bcache%i", idx);
863 
864 	d->disk->major		= bcache_major;
865 	d->disk->first_minor	= idx_to_first_minor(idx);
866 	d->disk->fops		= &bcache_ops;
867 	d->disk->private_data	= d;
868 
869 	q = blk_alloc_queue(make_request_fn, NUMA_NO_NODE);
870 	if (!q)
871 		return -ENOMEM;
872 
873 	d->disk->queue			= q;
874 	q->queuedata			= d;
875 	q->backing_dev_info->congested_data = d;
876 	q->limits.max_hw_sectors	= UINT_MAX;
877 	q->limits.max_sectors		= UINT_MAX;
878 	q->limits.max_segment_size	= UINT_MAX;
879 	q->limits.max_segments		= BIO_MAX_PAGES;
880 	blk_queue_max_discard_sectors(q, UINT_MAX);
881 	q->limits.discard_granularity	= 512;
882 	q->limits.io_min		= block_size;
883 	q->limits.logical_block_size	= block_size;
884 	q->limits.physical_block_size	= block_size;
885 	blk_queue_flag_set(QUEUE_FLAG_NONROT, d->disk->queue);
886 	blk_queue_flag_clear(QUEUE_FLAG_ADD_RANDOM, d->disk->queue);
887 	blk_queue_flag_set(QUEUE_FLAG_DISCARD, d->disk->queue);
888 
889 	blk_queue_write_cache(q, true, true);
890 
891 	return 0;
892 
893 err:
894 	ida_simple_remove(&bcache_device_idx, idx);
895 	return -ENOMEM;
896 
897 }
898 
899 /* Cached device */
900 
901 static void calc_cached_dev_sectors(struct cache_set *c)
902 {
903 	uint64_t sectors = 0;
904 	struct cached_dev *dc;
905 
906 	list_for_each_entry(dc, &c->cached_devs, list)
907 		sectors += bdev_sectors(dc->bdev);
908 
909 	c->cached_dev_sectors = sectors;
910 }
911 
912 #define BACKING_DEV_OFFLINE_TIMEOUT 5
913 static int cached_dev_status_update(void *arg)
914 {
915 	struct cached_dev *dc = arg;
916 	struct request_queue *q;
917 
918 	/*
919 	 * If this delayed worker is stopping outside, directly quit here.
920 	 * dc->io_disable might be set via sysfs interface, so check it
921 	 * here too.
922 	 */
923 	while (!kthread_should_stop() && !dc->io_disable) {
924 		q = bdev_get_queue(dc->bdev);
925 		if (blk_queue_dying(q))
926 			dc->offline_seconds++;
927 		else
928 			dc->offline_seconds = 0;
929 
930 		if (dc->offline_seconds >= BACKING_DEV_OFFLINE_TIMEOUT) {
931 			pr_err("%s: device offline for %d seconds",
932 			       dc->backing_dev_name,
933 			       BACKING_DEV_OFFLINE_TIMEOUT);
934 			pr_err("%s: disable I/O request due to backing "
935 			       "device offline", dc->disk.name);
936 			dc->io_disable = true;
937 			/* let others know earlier that io_disable is true */
938 			smp_mb();
939 			bcache_device_stop(&dc->disk);
940 			break;
941 		}
942 		schedule_timeout_interruptible(HZ);
943 	}
944 
945 	wait_for_kthread_stop();
946 	return 0;
947 }
948 
949 
950 int bch_cached_dev_run(struct cached_dev *dc)
951 {
952 	struct bcache_device *d = &dc->disk;
953 	char *buf = kmemdup_nul(dc->sb.label, SB_LABEL_SIZE, GFP_KERNEL);
954 	char *env[] = {
955 		"DRIVER=bcache",
956 		kasprintf(GFP_KERNEL, "CACHED_UUID=%pU", dc->sb.uuid),
957 		kasprintf(GFP_KERNEL, "CACHED_LABEL=%s", buf ? : ""),
958 		NULL,
959 	};
960 
961 	if (dc->io_disable) {
962 		pr_err("I/O disabled on cached dev %s",
963 		       dc->backing_dev_name);
964 		kfree(env[1]);
965 		kfree(env[2]);
966 		kfree(buf);
967 		return -EIO;
968 	}
969 
970 	if (atomic_xchg(&dc->running, 1)) {
971 		kfree(env[1]);
972 		kfree(env[2]);
973 		kfree(buf);
974 		pr_info("cached dev %s is running already",
975 		       dc->backing_dev_name);
976 		return -EBUSY;
977 	}
978 
979 	if (!d->c &&
980 	    BDEV_STATE(&dc->sb) != BDEV_STATE_NONE) {
981 		struct closure cl;
982 
983 		closure_init_stack(&cl);
984 
985 		SET_BDEV_STATE(&dc->sb, BDEV_STATE_STALE);
986 		bch_write_bdev_super(dc, &cl);
987 		closure_sync(&cl);
988 	}
989 
990 	add_disk(d->disk);
991 	bd_link_disk_holder(dc->bdev, dc->disk.disk);
992 	/*
993 	 * won't show up in the uevent file, use udevadm monitor -e instead
994 	 * only class / kset properties are persistent
995 	 */
996 	kobject_uevent_env(&disk_to_dev(d->disk)->kobj, KOBJ_CHANGE, env);
997 	kfree(env[1]);
998 	kfree(env[2]);
999 	kfree(buf);
1000 
1001 	if (sysfs_create_link(&d->kobj, &disk_to_dev(d->disk)->kobj, "dev") ||
1002 	    sysfs_create_link(&disk_to_dev(d->disk)->kobj,
1003 			      &d->kobj, "bcache")) {
1004 		pr_err("Couldn't create bcache dev <-> disk sysfs symlinks");
1005 		return -ENOMEM;
1006 	}
1007 
1008 	dc->status_update_thread = kthread_run(cached_dev_status_update,
1009 					       dc, "bcache_status_update");
1010 	if (IS_ERR(dc->status_update_thread)) {
1011 		pr_warn("failed to create bcache_status_update kthread, "
1012 			"continue to run without monitoring backing "
1013 			"device status");
1014 	}
1015 
1016 	return 0;
1017 }
1018 
1019 /*
1020  * If BCACHE_DEV_RATE_DW_RUNNING is set, it means routine of the delayed
1021  * work dc->writeback_rate_update is running. Wait until the routine
1022  * quits (BCACHE_DEV_RATE_DW_RUNNING is clear), then continue to
1023  * cancel it. If BCACHE_DEV_RATE_DW_RUNNING is not clear after time_out
1024  * seconds, give up waiting here and continue to cancel it too.
1025  */
1026 static void cancel_writeback_rate_update_dwork(struct cached_dev *dc)
1027 {
1028 	int time_out = WRITEBACK_RATE_UPDATE_SECS_MAX * HZ;
1029 
1030 	do {
1031 		if (!test_bit(BCACHE_DEV_RATE_DW_RUNNING,
1032 			      &dc->disk.flags))
1033 			break;
1034 		time_out--;
1035 		schedule_timeout_interruptible(1);
1036 	} while (time_out > 0);
1037 
1038 	if (time_out == 0)
1039 		pr_warn("give up waiting for dc->writeback_write_update to quit");
1040 
1041 	cancel_delayed_work_sync(&dc->writeback_rate_update);
1042 }
1043 
1044 static void cached_dev_detach_finish(struct work_struct *w)
1045 {
1046 	struct cached_dev *dc = container_of(w, struct cached_dev, detach);
1047 	struct closure cl;
1048 
1049 	closure_init_stack(&cl);
1050 
1051 	BUG_ON(!test_bit(BCACHE_DEV_DETACHING, &dc->disk.flags));
1052 	BUG_ON(refcount_read(&dc->count));
1053 
1054 
1055 	if (test_and_clear_bit(BCACHE_DEV_WB_RUNNING, &dc->disk.flags))
1056 		cancel_writeback_rate_update_dwork(dc);
1057 
1058 	if (!IS_ERR_OR_NULL(dc->writeback_thread)) {
1059 		kthread_stop(dc->writeback_thread);
1060 		dc->writeback_thread = NULL;
1061 	}
1062 
1063 	memset(&dc->sb.set_uuid, 0, 16);
1064 	SET_BDEV_STATE(&dc->sb, BDEV_STATE_NONE);
1065 
1066 	bch_write_bdev_super(dc, &cl);
1067 	closure_sync(&cl);
1068 
1069 	mutex_lock(&bch_register_lock);
1070 
1071 	calc_cached_dev_sectors(dc->disk.c);
1072 	bcache_device_detach(&dc->disk);
1073 	list_move(&dc->list, &uncached_devices);
1074 
1075 	clear_bit(BCACHE_DEV_DETACHING, &dc->disk.flags);
1076 	clear_bit(BCACHE_DEV_UNLINK_DONE, &dc->disk.flags);
1077 
1078 	mutex_unlock(&bch_register_lock);
1079 
1080 	pr_info("Caching disabled for %s", dc->backing_dev_name);
1081 
1082 	/* Drop ref we took in cached_dev_detach() */
1083 	closure_put(&dc->disk.cl);
1084 }
1085 
1086 void bch_cached_dev_detach(struct cached_dev *dc)
1087 {
1088 	lockdep_assert_held(&bch_register_lock);
1089 
1090 	if (test_bit(BCACHE_DEV_CLOSING, &dc->disk.flags))
1091 		return;
1092 
1093 	if (test_and_set_bit(BCACHE_DEV_DETACHING, &dc->disk.flags))
1094 		return;
1095 
1096 	/*
1097 	 * Block the device from being closed and freed until we're finished
1098 	 * detaching
1099 	 */
1100 	closure_get(&dc->disk.cl);
1101 
1102 	bch_writeback_queue(dc);
1103 
1104 	cached_dev_put(dc);
1105 }
1106 
1107 int bch_cached_dev_attach(struct cached_dev *dc, struct cache_set *c,
1108 			  uint8_t *set_uuid)
1109 {
1110 	uint32_t rtime = cpu_to_le32((u32)ktime_get_real_seconds());
1111 	struct uuid_entry *u;
1112 	struct cached_dev *exist_dc, *t;
1113 	int ret = 0;
1114 
1115 	if ((set_uuid && memcmp(set_uuid, c->sb.set_uuid, 16)) ||
1116 	    (!set_uuid && memcmp(dc->sb.set_uuid, c->sb.set_uuid, 16)))
1117 		return -ENOENT;
1118 
1119 	if (dc->disk.c) {
1120 		pr_err("Can't attach %s: already attached",
1121 		       dc->backing_dev_name);
1122 		return -EINVAL;
1123 	}
1124 
1125 	if (test_bit(CACHE_SET_STOPPING, &c->flags)) {
1126 		pr_err("Can't attach %s: shutting down",
1127 		       dc->backing_dev_name);
1128 		return -EINVAL;
1129 	}
1130 
1131 	if (dc->sb.block_size < c->sb.block_size) {
1132 		/* Will die */
1133 		pr_err("Couldn't attach %s: block size less than set's block size",
1134 		       dc->backing_dev_name);
1135 		return -EINVAL;
1136 	}
1137 
1138 	/* Check whether already attached */
1139 	list_for_each_entry_safe(exist_dc, t, &c->cached_devs, list) {
1140 		if (!memcmp(dc->sb.uuid, exist_dc->sb.uuid, 16)) {
1141 			pr_err("Tried to attach %s but duplicate UUID already attached",
1142 				dc->backing_dev_name);
1143 
1144 			return -EINVAL;
1145 		}
1146 	}
1147 
1148 	u = uuid_find(c, dc->sb.uuid);
1149 
1150 	if (u &&
1151 	    (BDEV_STATE(&dc->sb) == BDEV_STATE_STALE ||
1152 	     BDEV_STATE(&dc->sb) == BDEV_STATE_NONE)) {
1153 		memcpy(u->uuid, invalid_uuid, 16);
1154 		u->invalidated = cpu_to_le32((u32)ktime_get_real_seconds());
1155 		u = NULL;
1156 	}
1157 
1158 	if (!u) {
1159 		if (BDEV_STATE(&dc->sb) == BDEV_STATE_DIRTY) {
1160 			pr_err("Couldn't find uuid for %s in set",
1161 			       dc->backing_dev_name);
1162 			return -ENOENT;
1163 		}
1164 
1165 		u = uuid_find_empty(c);
1166 		if (!u) {
1167 			pr_err("Not caching %s, no room for UUID",
1168 			       dc->backing_dev_name);
1169 			return -EINVAL;
1170 		}
1171 	}
1172 
1173 	/*
1174 	 * Deadlocks since we're called via sysfs...
1175 	 * sysfs_remove_file(&dc->kobj, &sysfs_attach);
1176 	 */
1177 
1178 	if (bch_is_zero(u->uuid, 16)) {
1179 		struct closure cl;
1180 
1181 		closure_init_stack(&cl);
1182 
1183 		memcpy(u->uuid, dc->sb.uuid, 16);
1184 		memcpy(u->label, dc->sb.label, SB_LABEL_SIZE);
1185 		u->first_reg = u->last_reg = rtime;
1186 		bch_uuid_write(c);
1187 
1188 		memcpy(dc->sb.set_uuid, c->sb.set_uuid, 16);
1189 		SET_BDEV_STATE(&dc->sb, BDEV_STATE_CLEAN);
1190 
1191 		bch_write_bdev_super(dc, &cl);
1192 		closure_sync(&cl);
1193 	} else {
1194 		u->last_reg = rtime;
1195 		bch_uuid_write(c);
1196 	}
1197 
1198 	bcache_device_attach(&dc->disk, c, u - c->uuids);
1199 	list_move(&dc->list, &c->cached_devs);
1200 	calc_cached_dev_sectors(c);
1201 
1202 	/*
1203 	 * dc->c must be set before dc->count != 0 - paired with the mb in
1204 	 * cached_dev_get()
1205 	 */
1206 	smp_wmb();
1207 	refcount_set(&dc->count, 1);
1208 
1209 	/* Block writeback thread, but spawn it */
1210 	down_write(&dc->writeback_lock);
1211 	if (bch_cached_dev_writeback_start(dc)) {
1212 		up_write(&dc->writeback_lock);
1213 		pr_err("Couldn't start writeback facilities for %s",
1214 		       dc->disk.disk->disk_name);
1215 		return -ENOMEM;
1216 	}
1217 
1218 	if (BDEV_STATE(&dc->sb) == BDEV_STATE_DIRTY) {
1219 		atomic_set(&dc->has_dirty, 1);
1220 		bch_writeback_queue(dc);
1221 	}
1222 
1223 	bch_sectors_dirty_init(&dc->disk);
1224 
1225 	ret = bch_cached_dev_run(dc);
1226 	if (ret && (ret != -EBUSY)) {
1227 		up_write(&dc->writeback_lock);
1228 		/*
1229 		 * bch_register_lock is held, bcache_device_stop() is not
1230 		 * able to be directly called. The kthread and kworker
1231 		 * created previously in bch_cached_dev_writeback_start()
1232 		 * have to be stopped manually here.
1233 		 */
1234 		kthread_stop(dc->writeback_thread);
1235 		cancel_writeback_rate_update_dwork(dc);
1236 		pr_err("Couldn't run cached device %s",
1237 		       dc->backing_dev_name);
1238 		return ret;
1239 	}
1240 
1241 	bcache_device_link(&dc->disk, c, "bdev");
1242 	atomic_inc(&c->attached_dev_nr);
1243 
1244 	/* Allow the writeback thread to proceed */
1245 	up_write(&dc->writeback_lock);
1246 
1247 	pr_info("Caching %s as %s on set %pU",
1248 		dc->backing_dev_name,
1249 		dc->disk.disk->disk_name,
1250 		dc->disk.c->sb.set_uuid);
1251 	return 0;
1252 }
1253 
1254 /* when dc->disk.kobj released */
1255 void bch_cached_dev_release(struct kobject *kobj)
1256 {
1257 	struct cached_dev *dc = container_of(kobj, struct cached_dev,
1258 					     disk.kobj);
1259 	kfree(dc);
1260 	module_put(THIS_MODULE);
1261 }
1262 
1263 static void cached_dev_free(struct closure *cl)
1264 {
1265 	struct cached_dev *dc = container_of(cl, struct cached_dev, disk.cl);
1266 
1267 	if (test_and_clear_bit(BCACHE_DEV_WB_RUNNING, &dc->disk.flags))
1268 		cancel_writeback_rate_update_dwork(dc);
1269 
1270 	if (!IS_ERR_OR_NULL(dc->writeback_thread))
1271 		kthread_stop(dc->writeback_thread);
1272 	if (!IS_ERR_OR_NULL(dc->status_update_thread))
1273 		kthread_stop(dc->status_update_thread);
1274 
1275 	mutex_lock(&bch_register_lock);
1276 
1277 	if (atomic_read(&dc->running))
1278 		bd_unlink_disk_holder(dc->bdev, dc->disk.disk);
1279 	bcache_device_free(&dc->disk);
1280 	list_del(&dc->list);
1281 
1282 	mutex_unlock(&bch_register_lock);
1283 
1284 	if (dc->sb_disk)
1285 		put_page(virt_to_page(dc->sb_disk));
1286 
1287 	if (!IS_ERR_OR_NULL(dc->bdev))
1288 		blkdev_put(dc->bdev, FMODE_READ|FMODE_WRITE|FMODE_EXCL);
1289 
1290 	wake_up(&unregister_wait);
1291 
1292 	kobject_put(&dc->disk.kobj);
1293 }
1294 
1295 static void cached_dev_flush(struct closure *cl)
1296 {
1297 	struct cached_dev *dc = container_of(cl, struct cached_dev, disk.cl);
1298 	struct bcache_device *d = &dc->disk;
1299 
1300 	mutex_lock(&bch_register_lock);
1301 	bcache_device_unlink(d);
1302 	mutex_unlock(&bch_register_lock);
1303 
1304 	bch_cache_accounting_destroy(&dc->accounting);
1305 	kobject_del(&d->kobj);
1306 
1307 	continue_at(cl, cached_dev_free, system_wq);
1308 }
1309 
1310 static int cached_dev_init(struct cached_dev *dc, unsigned int block_size)
1311 {
1312 	int ret;
1313 	struct io *io;
1314 	struct request_queue *q = bdev_get_queue(dc->bdev);
1315 
1316 	__module_get(THIS_MODULE);
1317 	INIT_LIST_HEAD(&dc->list);
1318 	closure_init(&dc->disk.cl, NULL);
1319 	set_closure_fn(&dc->disk.cl, cached_dev_flush, system_wq);
1320 	kobject_init(&dc->disk.kobj, &bch_cached_dev_ktype);
1321 	INIT_WORK(&dc->detach, cached_dev_detach_finish);
1322 	sema_init(&dc->sb_write_mutex, 1);
1323 	INIT_LIST_HEAD(&dc->io_lru);
1324 	spin_lock_init(&dc->io_lock);
1325 	bch_cache_accounting_init(&dc->accounting, &dc->disk.cl);
1326 
1327 	dc->sequential_cutoff		= 4 << 20;
1328 
1329 	for (io = dc->io; io < dc->io + RECENT_IO; io++) {
1330 		list_add(&io->lru, &dc->io_lru);
1331 		hlist_add_head(&io->hash, dc->io_hash + RECENT_IO);
1332 	}
1333 
1334 	dc->disk.stripe_size = q->limits.io_opt >> 9;
1335 
1336 	if (dc->disk.stripe_size)
1337 		dc->partial_stripes_expensive =
1338 			q->limits.raid_partial_stripes_expensive;
1339 
1340 	ret = bcache_device_init(&dc->disk, block_size,
1341 			 dc->bdev->bd_part->nr_sects - dc->sb.data_offset,
1342 			 cached_dev_make_request);
1343 	if (ret)
1344 		return ret;
1345 
1346 	dc->disk.disk->queue->backing_dev_info->ra_pages =
1347 		max(dc->disk.disk->queue->backing_dev_info->ra_pages,
1348 		    q->backing_dev_info->ra_pages);
1349 
1350 	atomic_set(&dc->io_errors, 0);
1351 	dc->io_disable = false;
1352 	dc->error_limit = DEFAULT_CACHED_DEV_ERROR_LIMIT;
1353 	/* default to auto */
1354 	dc->stop_when_cache_set_failed = BCH_CACHED_DEV_STOP_AUTO;
1355 
1356 	bch_cached_dev_request_init(dc);
1357 	bch_cached_dev_writeback_init(dc);
1358 	return 0;
1359 }
1360 
1361 /* Cached device - bcache superblock */
1362 
1363 static int register_bdev(struct cache_sb *sb, struct cache_sb_disk *sb_disk,
1364 				 struct block_device *bdev,
1365 				 struct cached_dev *dc)
1366 {
1367 	const char *err = "cannot allocate memory";
1368 	struct cache_set *c;
1369 	int ret = -ENOMEM;
1370 
1371 	bdevname(bdev, dc->backing_dev_name);
1372 	memcpy(&dc->sb, sb, sizeof(struct cache_sb));
1373 	dc->bdev = bdev;
1374 	dc->bdev->bd_holder = dc;
1375 	dc->sb_disk = sb_disk;
1376 
1377 	if (cached_dev_init(dc, sb->block_size << 9))
1378 		goto err;
1379 
1380 	err = "error creating kobject";
1381 	if (kobject_add(&dc->disk.kobj, &part_to_dev(bdev->bd_part)->kobj,
1382 			"bcache"))
1383 		goto err;
1384 	if (bch_cache_accounting_add_kobjs(&dc->accounting, &dc->disk.kobj))
1385 		goto err;
1386 
1387 	pr_info("registered backing device %s", dc->backing_dev_name);
1388 
1389 	list_add(&dc->list, &uncached_devices);
1390 	/* attach to a matched cache set if it exists */
1391 	list_for_each_entry(c, &bch_cache_sets, list)
1392 		bch_cached_dev_attach(dc, c, NULL);
1393 
1394 	if (BDEV_STATE(&dc->sb) == BDEV_STATE_NONE ||
1395 	    BDEV_STATE(&dc->sb) == BDEV_STATE_STALE) {
1396 		err = "failed to run cached device";
1397 		ret = bch_cached_dev_run(dc);
1398 		if (ret)
1399 			goto err;
1400 	}
1401 
1402 	return 0;
1403 err:
1404 	pr_notice("error %s: %s", dc->backing_dev_name, err);
1405 	bcache_device_stop(&dc->disk);
1406 	return ret;
1407 }
1408 
1409 /* Flash only volumes */
1410 
1411 /* When d->kobj released */
1412 void bch_flash_dev_release(struct kobject *kobj)
1413 {
1414 	struct bcache_device *d = container_of(kobj, struct bcache_device,
1415 					       kobj);
1416 	kfree(d);
1417 }
1418 
1419 static void flash_dev_free(struct closure *cl)
1420 {
1421 	struct bcache_device *d = container_of(cl, struct bcache_device, cl);
1422 
1423 	mutex_lock(&bch_register_lock);
1424 	atomic_long_sub(bcache_dev_sectors_dirty(d),
1425 			&d->c->flash_dev_dirty_sectors);
1426 	bcache_device_free(d);
1427 	mutex_unlock(&bch_register_lock);
1428 	kobject_put(&d->kobj);
1429 }
1430 
1431 static void flash_dev_flush(struct closure *cl)
1432 {
1433 	struct bcache_device *d = container_of(cl, struct bcache_device, cl);
1434 
1435 	mutex_lock(&bch_register_lock);
1436 	bcache_device_unlink(d);
1437 	mutex_unlock(&bch_register_lock);
1438 	kobject_del(&d->kobj);
1439 	continue_at(cl, flash_dev_free, system_wq);
1440 }
1441 
1442 static int flash_dev_run(struct cache_set *c, struct uuid_entry *u)
1443 {
1444 	struct bcache_device *d = kzalloc(sizeof(struct bcache_device),
1445 					  GFP_KERNEL);
1446 	if (!d)
1447 		return -ENOMEM;
1448 
1449 	closure_init(&d->cl, NULL);
1450 	set_closure_fn(&d->cl, flash_dev_flush, system_wq);
1451 
1452 	kobject_init(&d->kobj, &bch_flash_dev_ktype);
1453 
1454 	if (bcache_device_init(d, block_bytes(c), u->sectors,
1455 			flash_dev_make_request))
1456 		goto err;
1457 
1458 	bcache_device_attach(d, c, u - c->uuids);
1459 	bch_sectors_dirty_init(d);
1460 	bch_flash_dev_request_init(d);
1461 	add_disk(d->disk);
1462 
1463 	if (kobject_add(&d->kobj, &disk_to_dev(d->disk)->kobj, "bcache"))
1464 		goto err;
1465 
1466 	bcache_device_link(d, c, "volume");
1467 
1468 	return 0;
1469 err:
1470 	kobject_put(&d->kobj);
1471 	return -ENOMEM;
1472 }
1473 
1474 static int flash_devs_run(struct cache_set *c)
1475 {
1476 	int ret = 0;
1477 	struct uuid_entry *u;
1478 
1479 	for (u = c->uuids;
1480 	     u < c->uuids + c->nr_uuids && !ret;
1481 	     u++)
1482 		if (UUID_FLASH_ONLY(u))
1483 			ret = flash_dev_run(c, u);
1484 
1485 	return ret;
1486 }
1487 
1488 int bch_flash_dev_create(struct cache_set *c, uint64_t size)
1489 {
1490 	struct uuid_entry *u;
1491 
1492 	if (test_bit(CACHE_SET_STOPPING, &c->flags))
1493 		return -EINTR;
1494 
1495 	if (!test_bit(CACHE_SET_RUNNING, &c->flags))
1496 		return -EPERM;
1497 
1498 	u = uuid_find_empty(c);
1499 	if (!u) {
1500 		pr_err("Can't create volume, no room for UUID");
1501 		return -EINVAL;
1502 	}
1503 
1504 	get_random_bytes(u->uuid, 16);
1505 	memset(u->label, 0, 32);
1506 	u->first_reg = u->last_reg = cpu_to_le32((u32)ktime_get_real_seconds());
1507 
1508 	SET_UUID_FLASH_ONLY(u, 1);
1509 	u->sectors = size >> 9;
1510 
1511 	bch_uuid_write(c);
1512 
1513 	return flash_dev_run(c, u);
1514 }
1515 
1516 bool bch_cached_dev_error(struct cached_dev *dc)
1517 {
1518 	if (!dc || test_bit(BCACHE_DEV_CLOSING, &dc->disk.flags))
1519 		return false;
1520 
1521 	dc->io_disable = true;
1522 	/* make others know io_disable is true earlier */
1523 	smp_mb();
1524 
1525 	pr_err("stop %s: too many IO errors on backing device %s\n",
1526 		dc->disk.disk->disk_name, dc->backing_dev_name);
1527 
1528 	bcache_device_stop(&dc->disk);
1529 	return true;
1530 }
1531 
1532 /* Cache set */
1533 
1534 __printf(2, 3)
1535 bool bch_cache_set_error(struct cache_set *c, const char *fmt, ...)
1536 {
1537 	va_list args;
1538 
1539 	if (c->on_error != ON_ERROR_PANIC &&
1540 	    test_bit(CACHE_SET_STOPPING, &c->flags))
1541 		return false;
1542 
1543 	if (test_and_set_bit(CACHE_SET_IO_DISABLE, &c->flags))
1544 		pr_info("CACHE_SET_IO_DISABLE already set");
1545 
1546 	/*
1547 	 * XXX: we can be called from atomic context
1548 	 * acquire_console_sem();
1549 	 */
1550 
1551 	pr_err("bcache: error on %pU: ", c->sb.set_uuid);
1552 
1553 	va_start(args, fmt);
1554 	vprintk(fmt, args);
1555 	va_end(args);
1556 
1557 	pr_err(", disabling caching\n");
1558 
1559 	if (c->on_error == ON_ERROR_PANIC)
1560 		panic("panic forced after error\n");
1561 
1562 	bch_cache_set_unregister(c);
1563 	return true;
1564 }
1565 
1566 /* When c->kobj released */
1567 void bch_cache_set_release(struct kobject *kobj)
1568 {
1569 	struct cache_set *c = container_of(kobj, struct cache_set, kobj);
1570 
1571 	kfree(c);
1572 	module_put(THIS_MODULE);
1573 }
1574 
1575 static void cache_set_free(struct closure *cl)
1576 {
1577 	struct cache_set *c = container_of(cl, struct cache_set, cl);
1578 	struct cache *ca;
1579 	unsigned int i;
1580 
1581 	debugfs_remove(c->debug);
1582 
1583 	bch_open_buckets_free(c);
1584 	bch_btree_cache_free(c);
1585 	bch_journal_free(c);
1586 
1587 	mutex_lock(&bch_register_lock);
1588 	for_each_cache(ca, c, i)
1589 		if (ca) {
1590 			ca->set = NULL;
1591 			c->cache[ca->sb.nr_this_dev] = NULL;
1592 			kobject_put(&ca->kobj);
1593 		}
1594 
1595 	bch_bset_sort_state_free(&c->sort);
1596 	free_pages((unsigned long) c->uuids, ilog2(bucket_pages(c)));
1597 
1598 	if (c->moving_gc_wq)
1599 		destroy_workqueue(c->moving_gc_wq);
1600 	bioset_exit(&c->bio_split);
1601 	mempool_exit(&c->fill_iter);
1602 	mempool_exit(&c->bio_meta);
1603 	mempool_exit(&c->search);
1604 	kfree(c->devices);
1605 
1606 	list_del(&c->list);
1607 	mutex_unlock(&bch_register_lock);
1608 
1609 	pr_info("Cache set %pU unregistered", c->sb.set_uuid);
1610 	wake_up(&unregister_wait);
1611 
1612 	closure_debug_destroy(&c->cl);
1613 	kobject_put(&c->kobj);
1614 }
1615 
1616 static void cache_set_flush(struct closure *cl)
1617 {
1618 	struct cache_set *c = container_of(cl, struct cache_set, caching);
1619 	struct cache *ca;
1620 	struct btree *b;
1621 	unsigned int i;
1622 
1623 	bch_cache_accounting_destroy(&c->accounting);
1624 
1625 	kobject_put(&c->internal);
1626 	kobject_del(&c->kobj);
1627 
1628 	if (!IS_ERR_OR_NULL(c->gc_thread))
1629 		kthread_stop(c->gc_thread);
1630 
1631 	if (!IS_ERR_OR_NULL(c->root))
1632 		list_add(&c->root->list, &c->btree_cache);
1633 
1634 	/*
1635 	 * Avoid flushing cached nodes if cache set is retiring
1636 	 * due to too many I/O errors detected.
1637 	 */
1638 	if (!test_bit(CACHE_SET_IO_DISABLE, &c->flags))
1639 		list_for_each_entry(b, &c->btree_cache, list) {
1640 			mutex_lock(&b->write_lock);
1641 			if (btree_node_dirty(b))
1642 				__bch_btree_node_write(b, NULL);
1643 			mutex_unlock(&b->write_lock);
1644 		}
1645 
1646 	for_each_cache(ca, c, i)
1647 		if (ca->alloc_thread)
1648 			kthread_stop(ca->alloc_thread);
1649 
1650 	if (c->journal.cur) {
1651 		cancel_delayed_work_sync(&c->journal.work);
1652 		/* flush last journal entry if needed */
1653 		c->journal.work.work.func(&c->journal.work.work);
1654 	}
1655 
1656 	closure_return(cl);
1657 }
1658 
1659 /*
1660  * This function is only called when CACHE_SET_IO_DISABLE is set, which means
1661  * cache set is unregistering due to too many I/O errors. In this condition,
1662  * the bcache device might be stopped, it depends on stop_when_cache_set_failed
1663  * value and whether the broken cache has dirty data:
1664  *
1665  * dc->stop_when_cache_set_failed    dc->has_dirty   stop bcache device
1666  *  BCH_CACHED_STOP_AUTO               0               NO
1667  *  BCH_CACHED_STOP_AUTO               1               YES
1668  *  BCH_CACHED_DEV_STOP_ALWAYS         0               YES
1669  *  BCH_CACHED_DEV_STOP_ALWAYS         1               YES
1670  *
1671  * The expected behavior is, if stop_when_cache_set_failed is configured to
1672  * "auto" via sysfs interface, the bcache device will not be stopped if the
1673  * backing device is clean on the broken cache device.
1674  */
1675 static void conditional_stop_bcache_device(struct cache_set *c,
1676 					   struct bcache_device *d,
1677 					   struct cached_dev *dc)
1678 {
1679 	if (dc->stop_when_cache_set_failed == BCH_CACHED_DEV_STOP_ALWAYS) {
1680 		pr_warn("stop_when_cache_set_failed of %s is \"always\", stop it for failed cache set %pU.",
1681 			d->disk->disk_name, c->sb.set_uuid);
1682 		bcache_device_stop(d);
1683 	} else if (atomic_read(&dc->has_dirty)) {
1684 		/*
1685 		 * dc->stop_when_cache_set_failed == BCH_CACHED_STOP_AUTO
1686 		 * and dc->has_dirty == 1
1687 		 */
1688 		pr_warn("stop_when_cache_set_failed of %s is \"auto\" and cache is dirty, stop it to avoid potential data corruption.",
1689 			d->disk->disk_name);
1690 		/*
1691 		 * There might be a small time gap that cache set is
1692 		 * released but bcache device is not. Inside this time
1693 		 * gap, regular I/O requests will directly go into
1694 		 * backing device as no cache set attached to. This
1695 		 * behavior may also introduce potential inconsistence
1696 		 * data in writeback mode while cache is dirty.
1697 		 * Therefore before calling bcache_device_stop() due
1698 		 * to a broken cache device, dc->io_disable should be
1699 		 * explicitly set to true.
1700 		 */
1701 		dc->io_disable = true;
1702 		/* make others know io_disable is true earlier */
1703 		smp_mb();
1704 		bcache_device_stop(d);
1705 	} else {
1706 		/*
1707 		 * dc->stop_when_cache_set_failed == BCH_CACHED_STOP_AUTO
1708 		 * and dc->has_dirty == 0
1709 		 */
1710 		pr_warn("stop_when_cache_set_failed of %s is \"auto\" and cache is clean, keep it alive.",
1711 			d->disk->disk_name);
1712 	}
1713 }
1714 
1715 static void __cache_set_unregister(struct closure *cl)
1716 {
1717 	struct cache_set *c = container_of(cl, struct cache_set, caching);
1718 	struct cached_dev *dc;
1719 	struct bcache_device *d;
1720 	size_t i;
1721 
1722 	mutex_lock(&bch_register_lock);
1723 
1724 	for (i = 0; i < c->devices_max_used; i++) {
1725 		d = c->devices[i];
1726 		if (!d)
1727 			continue;
1728 
1729 		if (!UUID_FLASH_ONLY(&c->uuids[i]) &&
1730 		    test_bit(CACHE_SET_UNREGISTERING, &c->flags)) {
1731 			dc = container_of(d, struct cached_dev, disk);
1732 			bch_cached_dev_detach(dc);
1733 			if (test_bit(CACHE_SET_IO_DISABLE, &c->flags))
1734 				conditional_stop_bcache_device(c, d, dc);
1735 		} else {
1736 			bcache_device_stop(d);
1737 		}
1738 	}
1739 
1740 	mutex_unlock(&bch_register_lock);
1741 
1742 	continue_at(cl, cache_set_flush, system_wq);
1743 }
1744 
1745 void bch_cache_set_stop(struct cache_set *c)
1746 {
1747 	if (!test_and_set_bit(CACHE_SET_STOPPING, &c->flags))
1748 		/* closure_fn set to __cache_set_unregister() */
1749 		closure_queue(&c->caching);
1750 }
1751 
1752 void bch_cache_set_unregister(struct cache_set *c)
1753 {
1754 	set_bit(CACHE_SET_UNREGISTERING, &c->flags);
1755 	bch_cache_set_stop(c);
1756 }
1757 
1758 #define alloc_bucket_pages(gfp, c)			\
1759 	((void *) __get_free_pages(__GFP_ZERO|gfp, ilog2(bucket_pages(c))))
1760 
1761 struct cache_set *bch_cache_set_alloc(struct cache_sb *sb)
1762 {
1763 	int iter_size;
1764 	struct cache_set *c = kzalloc(sizeof(struct cache_set), GFP_KERNEL);
1765 
1766 	if (!c)
1767 		return NULL;
1768 
1769 	__module_get(THIS_MODULE);
1770 	closure_init(&c->cl, NULL);
1771 	set_closure_fn(&c->cl, cache_set_free, system_wq);
1772 
1773 	closure_init(&c->caching, &c->cl);
1774 	set_closure_fn(&c->caching, __cache_set_unregister, system_wq);
1775 
1776 	/* Maybe create continue_at_noreturn() and use it here? */
1777 	closure_set_stopped(&c->cl);
1778 	closure_put(&c->cl);
1779 
1780 	kobject_init(&c->kobj, &bch_cache_set_ktype);
1781 	kobject_init(&c->internal, &bch_cache_set_internal_ktype);
1782 
1783 	bch_cache_accounting_init(&c->accounting, &c->cl);
1784 
1785 	memcpy(c->sb.set_uuid, sb->set_uuid, 16);
1786 	c->sb.block_size	= sb->block_size;
1787 	c->sb.bucket_size	= sb->bucket_size;
1788 	c->sb.nr_in_set		= sb->nr_in_set;
1789 	c->sb.last_mount	= sb->last_mount;
1790 	c->bucket_bits		= ilog2(sb->bucket_size);
1791 	c->block_bits		= ilog2(sb->block_size);
1792 	c->nr_uuids		= bucket_bytes(c) / sizeof(struct uuid_entry);
1793 	c->devices_max_used	= 0;
1794 	atomic_set(&c->attached_dev_nr, 0);
1795 	c->btree_pages		= bucket_pages(c);
1796 	if (c->btree_pages > BTREE_MAX_PAGES)
1797 		c->btree_pages = max_t(int, c->btree_pages / 4,
1798 				       BTREE_MAX_PAGES);
1799 
1800 	sema_init(&c->sb_write_mutex, 1);
1801 	mutex_init(&c->bucket_lock);
1802 	init_waitqueue_head(&c->btree_cache_wait);
1803 	spin_lock_init(&c->btree_cannibalize_lock);
1804 	init_waitqueue_head(&c->bucket_wait);
1805 	init_waitqueue_head(&c->gc_wait);
1806 	sema_init(&c->uuid_write_mutex, 1);
1807 
1808 	spin_lock_init(&c->btree_gc_time.lock);
1809 	spin_lock_init(&c->btree_split_time.lock);
1810 	spin_lock_init(&c->btree_read_time.lock);
1811 
1812 	bch_moving_init_cache_set(c);
1813 
1814 	INIT_LIST_HEAD(&c->list);
1815 	INIT_LIST_HEAD(&c->cached_devs);
1816 	INIT_LIST_HEAD(&c->btree_cache);
1817 	INIT_LIST_HEAD(&c->btree_cache_freeable);
1818 	INIT_LIST_HEAD(&c->btree_cache_freed);
1819 	INIT_LIST_HEAD(&c->data_buckets);
1820 
1821 	iter_size = (sb->bucket_size / sb->block_size + 1) *
1822 		sizeof(struct btree_iter_set);
1823 
1824 	if (!(c->devices = kcalloc(c->nr_uuids, sizeof(void *), GFP_KERNEL)) ||
1825 	    mempool_init_slab_pool(&c->search, 32, bch_search_cache) ||
1826 	    mempool_init_kmalloc_pool(&c->bio_meta, 2,
1827 				sizeof(struct bbio) + sizeof(struct bio_vec) *
1828 				bucket_pages(c)) ||
1829 	    mempool_init_kmalloc_pool(&c->fill_iter, 1, iter_size) ||
1830 	    bioset_init(&c->bio_split, 4, offsetof(struct bbio, bio),
1831 			BIOSET_NEED_BVECS|BIOSET_NEED_RESCUER) ||
1832 	    !(c->uuids = alloc_bucket_pages(GFP_KERNEL, c)) ||
1833 	    !(c->moving_gc_wq = alloc_workqueue("bcache_gc",
1834 						WQ_MEM_RECLAIM, 0)) ||
1835 	    bch_journal_alloc(c) ||
1836 	    bch_btree_cache_alloc(c) ||
1837 	    bch_open_buckets_alloc(c) ||
1838 	    bch_bset_sort_state_init(&c->sort, ilog2(c->btree_pages)))
1839 		goto err;
1840 
1841 	c->congested_read_threshold_us	= 2000;
1842 	c->congested_write_threshold_us	= 20000;
1843 	c->error_limit	= DEFAULT_IO_ERROR_LIMIT;
1844 	c->idle_max_writeback_rate_enabled = 1;
1845 	WARN_ON(test_and_clear_bit(CACHE_SET_IO_DISABLE, &c->flags));
1846 
1847 	return c;
1848 err:
1849 	bch_cache_set_unregister(c);
1850 	return NULL;
1851 }
1852 
1853 static int run_cache_set(struct cache_set *c)
1854 {
1855 	const char *err = "cannot allocate memory";
1856 	struct cached_dev *dc, *t;
1857 	struct cache *ca;
1858 	struct closure cl;
1859 	unsigned int i;
1860 	LIST_HEAD(journal);
1861 	struct journal_replay *l;
1862 
1863 	closure_init_stack(&cl);
1864 
1865 	for_each_cache(ca, c, i)
1866 		c->nbuckets += ca->sb.nbuckets;
1867 	set_gc_sectors(c);
1868 
1869 	if (CACHE_SYNC(&c->sb)) {
1870 		struct bkey *k;
1871 		struct jset *j;
1872 
1873 		err = "cannot allocate memory for journal";
1874 		if (bch_journal_read(c, &journal))
1875 			goto err;
1876 
1877 		pr_debug("btree_journal_read() done");
1878 
1879 		err = "no journal entries found";
1880 		if (list_empty(&journal))
1881 			goto err;
1882 
1883 		j = &list_entry(journal.prev, struct journal_replay, list)->j;
1884 
1885 		err = "IO error reading priorities";
1886 		for_each_cache(ca, c, i) {
1887 			if (prio_read(ca, j->prio_bucket[ca->sb.nr_this_dev]))
1888 				goto err;
1889 		}
1890 
1891 		/*
1892 		 * If prio_read() fails it'll call cache_set_error and we'll
1893 		 * tear everything down right away, but if we perhaps checked
1894 		 * sooner we could avoid journal replay.
1895 		 */
1896 
1897 		k = &j->btree_root;
1898 
1899 		err = "bad btree root";
1900 		if (__bch_btree_ptr_invalid(c, k))
1901 			goto err;
1902 
1903 		err = "error reading btree root";
1904 		c->root = bch_btree_node_get(c, NULL, k,
1905 					     j->btree_level,
1906 					     true, NULL);
1907 		if (IS_ERR_OR_NULL(c->root))
1908 			goto err;
1909 
1910 		list_del_init(&c->root->list);
1911 		rw_unlock(true, c->root);
1912 
1913 		err = uuid_read(c, j, &cl);
1914 		if (err)
1915 			goto err;
1916 
1917 		err = "error in recovery";
1918 		if (bch_btree_check(c))
1919 			goto err;
1920 
1921 		bch_journal_mark(c, &journal);
1922 		bch_initial_gc_finish(c);
1923 		pr_debug("btree_check() done");
1924 
1925 		/*
1926 		 * bcache_journal_next() can't happen sooner, or
1927 		 * btree_gc_finish() will give spurious errors about last_gc >
1928 		 * gc_gen - this is a hack but oh well.
1929 		 */
1930 		bch_journal_next(&c->journal);
1931 
1932 		err = "error starting allocator thread";
1933 		for_each_cache(ca, c, i)
1934 			if (bch_cache_allocator_start(ca))
1935 				goto err;
1936 
1937 		/*
1938 		 * First place it's safe to allocate: btree_check() and
1939 		 * btree_gc_finish() have to run before we have buckets to
1940 		 * allocate, and bch_bucket_alloc_set() might cause a journal
1941 		 * entry to be written so bcache_journal_next() has to be called
1942 		 * first.
1943 		 *
1944 		 * If the uuids were in the old format we have to rewrite them
1945 		 * before the next journal entry is written:
1946 		 */
1947 		if (j->version < BCACHE_JSET_VERSION_UUID)
1948 			__uuid_write(c);
1949 
1950 		err = "bcache: replay journal failed";
1951 		if (bch_journal_replay(c, &journal))
1952 			goto err;
1953 	} else {
1954 		pr_notice("invalidating existing data");
1955 
1956 		for_each_cache(ca, c, i) {
1957 			unsigned int j;
1958 
1959 			ca->sb.keys = clamp_t(int, ca->sb.nbuckets >> 7,
1960 					      2, SB_JOURNAL_BUCKETS);
1961 
1962 			for (j = 0; j < ca->sb.keys; j++)
1963 				ca->sb.d[j] = ca->sb.first_bucket + j;
1964 		}
1965 
1966 		bch_initial_gc_finish(c);
1967 
1968 		err = "error starting allocator thread";
1969 		for_each_cache(ca, c, i)
1970 			if (bch_cache_allocator_start(ca))
1971 				goto err;
1972 
1973 		mutex_lock(&c->bucket_lock);
1974 		for_each_cache(ca, c, i)
1975 			bch_prio_write(ca, true);
1976 		mutex_unlock(&c->bucket_lock);
1977 
1978 		err = "cannot allocate new UUID bucket";
1979 		if (__uuid_write(c))
1980 			goto err;
1981 
1982 		err = "cannot allocate new btree root";
1983 		c->root = __bch_btree_node_alloc(c, NULL, 0, true, NULL);
1984 		if (IS_ERR_OR_NULL(c->root))
1985 			goto err;
1986 
1987 		mutex_lock(&c->root->write_lock);
1988 		bkey_copy_key(&c->root->key, &MAX_KEY);
1989 		bch_btree_node_write(c->root, &cl);
1990 		mutex_unlock(&c->root->write_lock);
1991 
1992 		bch_btree_set_root(c->root);
1993 		rw_unlock(true, c->root);
1994 
1995 		/*
1996 		 * We don't want to write the first journal entry until
1997 		 * everything is set up - fortunately journal entries won't be
1998 		 * written until the SET_CACHE_SYNC() here:
1999 		 */
2000 		SET_CACHE_SYNC(&c->sb, true);
2001 
2002 		bch_journal_next(&c->journal);
2003 		bch_journal_meta(c, &cl);
2004 	}
2005 
2006 	err = "error starting gc thread";
2007 	if (bch_gc_thread_start(c))
2008 		goto err;
2009 
2010 	closure_sync(&cl);
2011 	c->sb.last_mount = (u32)ktime_get_real_seconds();
2012 	bcache_write_super(c);
2013 
2014 	list_for_each_entry_safe(dc, t, &uncached_devices, list)
2015 		bch_cached_dev_attach(dc, c, NULL);
2016 
2017 	flash_devs_run(c);
2018 
2019 	set_bit(CACHE_SET_RUNNING, &c->flags);
2020 	return 0;
2021 err:
2022 	while (!list_empty(&journal)) {
2023 		l = list_first_entry(&journal, struct journal_replay, list);
2024 		list_del(&l->list);
2025 		kfree(l);
2026 	}
2027 
2028 	closure_sync(&cl);
2029 
2030 	bch_cache_set_error(c, "%s", err);
2031 
2032 	return -EIO;
2033 }
2034 
2035 static bool can_attach_cache(struct cache *ca, struct cache_set *c)
2036 {
2037 	return ca->sb.block_size	== c->sb.block_size &&
2038 		ca->sb.bucket_size	== c->sb.bucket_size &&
2039 		ca->sb.nr_in_set	== c->sb.nr_in_set;
2040 }
2041 
2042 static const char *register_cache_set(struct cache *ca)
2043 {
2044 	char buf[12];
2045 	const char *err = "cannot allocate memory";
2046 	struct cache_set *c;
2047 
2048 	list_for_each_entry(c, &bch_cache_sets, list)
2049 		if (!memcmp(c->sb.set_uuid, ca->sb.set_uuid, 16)) {
2050 			if (c->cache[ca->sb.nr_this_dev])
2051 				return "duplicate cache set member";
2052 
2053 			if (!can_attach_cache(ca, c))
2054 				return "cache sb does not match set";
2055 
2056 			if (!CACHE_SYNC(&ca->sb))
2057 				SET_CACHE_SYNC(&c->sb, false);
2058 
2059 			goto found;
2060 		}
2061 
2062 	c = bch_cache_set_alloc(&ca->sb);
2063 	if (!c)
2064 		return err;
2065 
2066 	err = "error creating kobject";
2067 	if (kobject_add(&c->kobj, bcache_kobj, "%pU", c->sb.set_uuid) ||
2068 	    kobject_add(&c->internal, &c->kobj, "internal"))
2069 		goto err;
2070 
2071 	if (bch_cache_accounting_add_kobjs(&c->accounting, &c->kobj))
2072 		goto err;
2073 
2074 	bch_debug_init_cache_set(c);
2075 
2076 	list_add(&c->list, &bch_cache_sets);
2077 found:
2078 	sprintf(buf, "cache%i", ca->sb.nr_this_dev);
2079 	if (sysfs_create_link(&ca->kobj, &c->kobj, "set") ||
2080 	    sysfs_create_link(&c->kobj, &ca->kobj, buf))
2081 		goto err;
2082 
2083 	if (ca->sb.seq > c->sb.seq) {
2084 		c->sb.version		= ca->sb.version;
2085 		memcpy(c->sb.set_uuid, ca->sb.set_uuid, 16);
2086 		c->sb.flags             = ca->sb.flags;
2087 		c->sb.seq		= ca->sb.seq;
2088 		pr_debug("set version = %llu", c->sb.version);
2089 	}
2090 
2091 	kobject_get(&ca->kobj);
2092 	ca->set = c;
2093 	ca->set->cache[ca->sb.nr_this_dev] = ca;
2094 	c->cache_by_alloc[c->caches_loaded++] = ca;
2095 
2096 	if (c->caches_loaded == c->sb.nr_in_set) {
2097 		err = "failed to run cache set";
2098 		if (run_cache_set(c) < 0)
2099 			goto err;
2100 	}
2101 
2102 	return NULL;
2103 err:
2104 	bch_cache_set_unregister(c);
2105 	return err;
2106 }
2107 
2108 /* Cache device */
2109 
2110 /* When ca->kobj released */
2111 void bch_cache_release(struct kobject *kobj)
2112 {
2113 	struct cache *ca = container_of(kobj, struct cache, kobj);
2114 	unsigned int i;
2115 
2116 	if (ca->set) {
2117 		BUG_ON(ca->set->cache[ca->sb.nr_this_dev] != ca);
2118 		ca->set->cache[ca->sb.nr_this_dev] = NULL;
2119 	}
2120 
2121 	free_pages((unsigned long) ca->disk_buckets, ilog2(bucket_pages(ca)));
2122 	kfree(ca->prio_buckets);
2123 	vfree(ca->buckets);
2124 
2125 	free_heap(&ca->heap);
2126 	free_fifo(&ca->free_inc);
2127 
2128 	for (i = 0; i < RESERVE_NR; i++)
2129 		free_fifo(&ca->free[i]);
2130 
2131 	if (ca->sb_disk)
2132 		put_page(virt_to_page(ca->sb_disk));
2133 
2134 	if (!IS_ERR_OR_NULL(ca->bdev))
2135 		blkdev_put(ca->bdev, FMODE_READ|FMODE_WRITE|FMODE_EXCL);
2136 
2137 	kfree(ca);
2138 	module_put(THIS_MODULE);
2139 }
2140 
2141 static int cache_alloc(struct cache *ca)
2142 {
2143 	size_t free;
2144 	size_t btree_buckets;
2145 	struct bucket *b;
2146 	int ret = -ENOMEM;
2147 	const char *err = NULL;
2148 
2149 	__module_get(THIS_MODULE);
2150 	kobject_init(&ca->kobj, &bch_cache_ktype);
2151 
2152 	bio_init(&ca->journal.bio, ca->journal.bio.bi_inline_vecs, 8);
2153 
2154 	/*
2155 	 * when ca->sb.njournal_buckets is not zero, journal exists,
2156 	 * and in bch_journal_replay(), tree node may split,
2157 	 * so bucket of RESERVE_BTREE type is needed,
2158 	 * the worst situation is all journal buckets are valid journal,
2159 	 * and all the keys need to replay,
2160 	 * so the number of  RESERVE_BTREE type buckets should be as much
2161 	 * as journal buckets
2162 	 */
2163 	btree_buckets = ca->sb.njournal_buckets ?: 8;
2164 	free = roundup_pow_of_two(ca->sb.nbuckets) >> 10;
2165 	if (!free) {
2166 		ret = -EPERM;
2167 		err = "ca->sb.nbuckets is too small";
2168 		goto err_free;
2169 	}
2170 
2171 	if (!init_fifo(&ca->free[RESERVE_BTREE], btree_buckets,
2172 						GFP_KERNEL)) {
2173 		err = "ca->free[RESERVE_BTREE] alloc failed";
2174 		goto err_btree_alloc;
2175 	}
2176 
2177 	if (!init_fifo_exact(&ca->free[RESERVE_PRIO], prio_buckets(ca),
2178 							GFP_KERNEL)) {
2179 		err = "ca->free[RESERVE_PRIO] alloc failed";
2180 		goto err_prio_alloc;
2181 	}
2182 
2183 	if (!init_fifo(&ca->free[RESERVE_MOVINGGC], free, GFP_KERNEL)) {
2184 		err = "ca->free[RESERVE_MOVINGGC] alloc failed";
2185 		goto err_movinggc_alloc;
2186 	}
2187 
2188 	if (!init_fifo(&ca->free[RESERVE_NONE], free, GFP_KERNEL)) {
2189 		err = "ca->free[RESERVE_NONE] alloc failed";
2190 		goto err_none_alloc;
2191 	}
2192 
2193 	if (!init_fifo(&ca->free_inc, free << 2, GFP_KERNEL)) {
2194 		err = "ca->free_inc alloc failed";
2195 		goto err_free_inc_alloc;
2196 	}
2197 
2198 	if (!init_heap(&ca->heap, free << 3, GFP_KERNEL)) {
2199 		err = "ca->heap alloc failed";
2200 		goto err_heap_alloc;
2201 	}
2202 
2203 	ca->buckets = vzalloc(array_size(sizeof(struct bucket),
2204 			      ca->sb.nbuckets));
2205 	if (!ca->buckets) {
2206 		err = "ca->buckets alloc failed";
2207 		goto err_buckets_alloc;
2208 	}
2209 
2210 	ca->prio_buckets = kzalloc(array3_size(sizeof(uint64_t),
2211 				   prio_buckets(ca), 2),
2212 				   GFP_KERNEL);
2213 	if (!ca->prio_buckets) {
2214 		err = "ca->prio_buckets alloc failed";
2215 		goto err_prio_buckets_alloc;
2216 	}
2217 
2218 	ca->disk_buckets = alloc_bucket_pages(GFP_KERNEL, ca);
2219 	if (!ca->disk_buckets) {
2220 		err = "ca->disk_buckets alloc failed";
2221 		goto err_disk_buckets_alloc;
2222 	}
2223 
2224 	ca->prio_last_buckets = ca->prio_buckets + prio_buckets(ca);
2225 
2226 	for_each_bucket(b, ca)
2227 		atomic_set(&b->pin, 0);
2228 	return 0;
2229 
2230 err_disk_buckets_alloc:
2231 	kfree(ca->prio_buckets);
2232 err_prio_buckets_alloc:
2233 	vfree(ca->buckets);
2234 err_buckets_alloc:
2235 	free_heap(&ca->heap);
2236 err_heap_alloc:
2237 	free_fifo(&ca->free_inc);
2238 err_free_inc_alloc:
2239 	free_fifo(&ca->free[RESERVE_NONE]);
2240 err_none_alloc:
2241 	free_fifo(&ca->free[RESERVE_MOVINGGC]);
2242 err_movinggc_alloc:
2243 	free_fifo(&ca->free[RESERVE_PRIO]);
2244 err_prio_alloc:
2245 	free_fifo(&ca->free[RESERVE_BTREE]);
2246 err_btree_alloc:
2247 err_free:
2248 	module_put(THIS_MODULE);
2249 	if (err)
2250 		pr_notice("error %s: %s", ca->cache_dev_name, err);
2251 	return ret;
2252 }
2253 
2254 static int register_cache(struct cache_sb *sb, struct cache_sb_disk *sb_disk,
2255 				struct block_device *bdev, struct cache *ca)
2256 {
2257 	const char *err = NULL; /* must be set for any error case */
2258 	int ret = 0;
2259 
2260 	bdevname(bdev, ca->cache_dev_name);
2261 	memcpy(&ca->sb, sb, sizeof(struct cache_sb));
2262 	ca->bdev = bdev;
2263 	ca->bdev->bd_holder = ca;
2264 	ca->sb_disk = sb_disk;
2265 
2266 	if (blk_queue_discard(bdev_get_queue(bdev)))
2267 		ca->discard = CACHE_DISCARD(&ca->sb);
2268 
2269 	ret = cache_alloc(ca);
2270 	if (ret != 0) {
2271 		/*
2272 		 * If we failed here, it means ca->kobj is not initialized yet,
2273 		 * kobject_put() won't be called and there is no chance to
2274 		 * call blkdev_put() to bdev in bch_cache_release(). So we
2275 		 * explicitly call blkdev_put() here.
2276 		 */
2277 		blkdev_put(bdev, FMODE_READ|FMODE_WRITE|FMODE_EXCL);
2278 		if (ret == -ENOMEM)
2279 			err = "cache_alloc(): -ENOMEM";
2280 		else if (ret == -EPERM)
2281 			err = "cache_alloc(): cache device is too small";
2282 		else
2283 			err = "cache_alloc(): unknown error";
2284 		goto err;
2285 	}
2286 
2287 	if (kobject_add(&ca->kobj,
2288 			&part_to_dev(bdev->bd_part)->kobj,
2289 			"bcache")) {
2290 		err = "error calling kobject_add";
2291 		ret = -ENOMEM;
2292 		goto out;
2293 	}
2294 
2295 	mutex_lock(&bch_register_lock);
2296 	err = register_cache_set(ca);
2297 	mutex_unlock(&bch_register_lock);
2298 
2299 	if (err) {
2300 		ret = -ENODEV;
2301 		goto out;
2302 	}
2303 
2304 	pr_info("registered cache device %s", ca->cache_dev_name);
2305 
2306 out:
2307 	kobject_put(&ca->kobj);
2308 
2309 err:
2310 	if (err)
2311 		pr_notice("error %s: %s", ca->cache_dev_name, err);
2312 
2313 	return ret;
2314 }
2315 
2316 /* Global interfaces/init */
2317 
2318 static ssize_t register_bcache(struct kobject *k, struct kobj_attribute *attr,
2319 			       const char *buffer, size_t size);
2320 static ssize_t bch_pending_bdevs_cleanup(struct kobject *k,
2321 					 struct kobj_attribute *attr,
2322 					 const char *buffer, size_t size);
2323 
2324 kobj_attribute_write(register,		register_bcache);
2325 kobj_attribute_write(register_quiet,	register_bcache);
2326 kobj_attribute_write(pendings_cleanup,	bch_pending_bdevs_cleanup);
2327 
2328 static bool bch_is_open_backing(struct block_device *bdev)
2329 {
2330 	struct cache_set *c, *tc;
2331 	struct cached_dev *dc, *t;
2332 
2333 	list_for_each_entry_safe(c, tc, &bch_cache_sets, list)
2334 		list_for_each_entry_safe(dc, t, &c->cached_devs, list)
2335 			if (dc->bdev == bdev)
2336 				return true;
2337 	list_for_each_entry_safe(dc, t, &uncached_devices, list)
2338 		if (dc->bdev == bdev)
2339 			return true;
2340 	return false;
2341 }
2342 
2343 static bool bch_is_open_cache(struct block_device *bdev)
2344 {
2345 	struct cache_set *c, *tc;
2346 	struct cache *ca;
2347 	unsigned int i;
2348 
2349 	list_for_each_entry_safe(c, tc, &bch_cache_sets, list)
2350 		for_each_cache(ca, c, i)
2351 			if (ca->bdev == bdev)
2352 				return true;
2353 	return false;
2354 }
2355 
2356 static bool bch_is_open(struct block_device *bdev)
2357 {
2358 	return bch_is_open_cache(bdev) || bch_is_open_backing(bdev);
2359 }
2360 
2361 static ssize_t register_bcache(struct kobject *k, struct kobj_attribute *attr,
2362 			       const char *buffer, size_t size)
2363 {
2364 	const char *err;
2365 	char *path = NULL;
2366 	struct cache_sb *sb;
2367 	struct cache_sb_disk *sb_disk;
2368 	struct block_device *bdev;
2369 	ssize_t ret;
2370 
2371 	ret = -EBUSY;
2372 	err = "failed to reference bcache module";
2373 	if (!try_module_get(THIS_MODULE))
2374 		goto out;
2375 
2376 	/* For latest state of bcache_is_reboot */
2377 	smp_mb();
2378 	err = "bcache is in reboot";
2379 	if (bcache_is_reboot)
2380 		goto out_module_put;
2381 
2382 	ret = -ENOMEM;
2383 	err = "cannot allocate memory";
2384 	path = kstrndup(buffer, size, GFP_KERNEL);
2385 	if (!path)
2386 		goto out_module_put;
2387 
2388 	sb = kmalloc(sizeof(struct cache_sb), GFP_KERNEL);
2389 	if (!sb)
2390 		goto out_free_path;
2391 
2392 	ret = -EINVAL;
2393 	err = "failed to open device";
2394 	bdev = blkdev_get_by_path(strim(path),
2395 				  FMODE_READ|FMODE_WRITE|FMODE_EXCL,
2396 				  sb);
2397 	if (IS_ERR(bdev)) {
2398 		if (bdev == ERR_PTR(-EBUSY)) {
2399 			bdev = lookup_bdev(strim(path));
2400 			mutex_lock(&bch_register_lock);
2401 			if (!IS_ERR(bdev) && bch_is_open(bdev))
2402 				err = "device already registered";
2403 			else
2404 				err = "device busy";
2405 			mutex_unlock(&bch_register_lock);
2406 			if (!IS_ERR(bdev))
2407 				bdput(bdev);
2408 			if (attr == &ksysfs_register_quiet)
2409 				goto done;
2410 		}
2411 		goto out_free_sb;
2412 	}
2413 
2414 	err = "failed to set blocksize";
2415 	if (set_blocksize(bdev, 4096))
2416 		goto out_blkdev_put;
2417 
2418 	err = read_super(sb, bdev, &sb_disk);
2419 	if (err)
2420 		goto out_blkdev_put;
2421 
2422 	err = "failed to register device";
2423 	if (SB_IS_BDEV(sb)) {
2424 		struct cached_dev *dc = kzalloc(sizeof(*dc), GFP_KERNEL);
2425 
2426 		if (!dc)
2427 			goto out_put_sb_page;
2428 
2429 		mutex_lock(&bch_register_lock);
2430 		ret = register_bdev(sb, sb_disk, bdev, dc);
2431 		mutex_unlock(&bch_register_lock);
2432 		/* blkdev_put() will be called in cached_dev_free() */
2433 		if (ret < 0)
2434 			goto out_free_sb;
2435 	} else {
2436 		struct cache *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
2437 
2438 		if (!ca)
2439 			goto out_put_sb_page;
2440 
2441 		/* blkdev_put() will be called in bch_cache_release() */
2442 		if (register_cache(sb, sb_disk, bdev, ca) != 0)
2443 			goto out_free_sb;
2444 	}
2445 
2446 done:
2447 	kfree(sb);
2448 	kfree(path);
2449 	module_put(THIS_MODULE);
2450 	return size;
2451 
2452 out_put_sb_page:
2453 	put_page(virt_to_page(sb_disk));
2454 out_blkdev_put:
2455 	blkdev_put(bdev, FMODE_READ | FMODE_WRITE | FMODE_EXCL);
2456 out_free_sb:
2457 	kfree(sb);
2458 out_free_path:
2459 	kfree(path);
2460 	path = NULL;
2461 out_module_put:
2462 	module_put(THIS_MODULE);
2463 out:
2464 	pr_info("error %s: %s", path?path:"", err);
2465 	return ret;
2466 }
2467 
2468 
2469 struct pdev {
2470 	struct list_head list;
2471 	struct cached_dev *dc;
2472 };
2473 
2474 static ssize_t bch_pending_bdevs_cleanup(struct kobject *k,
2475 					 struct kobj_attribute *attr,
2476 					 const char *buffer,
2477 					 size_t size)
2478 {
2479 	LIST_HEAD(pending_devs);
2480 	ssize_t ret = size;
2481 	struct cached_dev *dc, *tdc;
2482 	struct pdev *pdev, *tpdev;
2483 	struct cache_set *c, *tc;
2484 
2485 	mutex_lock(&bch_register_lock);
2486 	list_for_each_entry_safe(dc, tdc, &uncached_devices, list) {
2487 		pdev = kmalloc(sizeof(struct pdev), GFP_KERNEL);
2488 		if (!pdev)
2489 			break;
2490 		pdev->dc = dc;
2491 		list_add(&pdev->list, &pending_devs);
2492 	}
2493 
2494 	list_for_each_entry_safe(pdev, tpdev, &pending_devs, list) {
2495 		list_for_each_entry_safe(c, tc, &bch_cache_sets, list) {
2496 			char *pdev_set_uuid = pdev->dc->sb.set_uuid;
2497 			char *set_uuid = c->sb.uuid;
2498 
2499 			if (!memcmp(pdev_set_uuid, set_uuid, 16)) {
2500 				list_del(&pdev->list);
2501 				kfree(pdev);
2502 				break;
2503 			}
2504 		}
2505 	}
2506 	mutex_unlock(&bch_register_lock);
2507 
2508 	list_for_each_entry_safe(pdev, tpdev, &pending_devs, list) {
2509 		pr_info("delete pdev %p", pdev);
2510 		list_del(&pdev->list);
2511 		bcache_device_stop(&pdev->dc->disk);
2512 		kfree(pdev);
2513 	}
2514 
2515 	return ret;
2516 }
2517 
2518 static int bcache_reboot(struct notifier_block *n, unsigned long code, void *x)
2519 {
2520 	if (bcache_is_reboot)
2521 		return NOTIFY_DONE;
2522 
2523 	if (code == SYS_DOWN ||
2524 	    code == SYS_HALT ||
2525 	    code == SYS_POWER_OFF) {
2526 		DEFINE_WAIT(wait);
2527 		unsigned long start = jiffies;
2528 		bool stopped = false;
2529 
2530 		struct cache_set *c, *tc;
2531 		struct cached_dev *dc, *tdc;
2532 
2533 		mutex_lock(&bch_register_lock);
2534 
2535 		if (bcache_is_reboot)
2536 			goto out;
2537 
2538 		/* New registration is rejected since now */
2539 		bcache_is_reboot = true;
2540 		/*
2541 		 * Make registering caller (if there is) on other CPU
2542 		 * core know bcache_is_reboot set to true earlier
2543 		 */
2544 		smp_mb();
2545 
2546 		if (list_empty(&bch_cache_sets) &&
2547 		    list_empty(&uncached_devices))
2548 			goto out;
2549 
2550 		mutex_unlock(&bch_register_lock);
2551 
2552 		pr_info("Stopping all devices:");
2553 
2554 		/*
2555 		 * The reason bch_register_lock is not held to call
2556 		 * bch_cache_set_stop() and bcache_device_stop() is to
2557 		 * avoid potential deadlock during reboot, because cache
2558 		 * set or bcache device stopping process will acqurie
2559 		 * bch_register_lock too.
2560 		 *
2561 		 * We are safe here because bcache_is_reboot sets to
2562 		 * true already, register_bcache() will reject new
2563 		 * registration now. bcache_is_reboot also makes sure
2564 		 * bcache_reboot() won't be re-entered on by other thread,
2565 		 * so there is no race in following list iteration by
2566 		 * list_for_each_entry_safe().
2567 		 */
2568 		list_for_each_entry_safe(c, tc, &bch_cache_sets, list)
2569 			bch_cache_set_stop(c);
2570 
2571 		list_for_each_entry_safe(dc, tdc, &uncached_devices, list)
2572 			bcache_device_stop(&dc->disk);
2573 
2574 
2575 		/*
2576 		 * Give an early chance for other kthreads and
2577 		 * kworkers to stop themselves
2578 		 */
2579 		schedule();
2580 
2581 		/* What's a condition variable? */
2582 		while (1) {
2583 			long timeout = start + 10 * HZ - jiffies;
2584 
2585 			mutex_lock(&bch_register_lock);
2586 			stopped = list_empty(&bch_cache_sets) &&
2587 				list_empty(&uncached_devices);
2588 
2589 			if (timeout < 0 || stopped)
2590 				break;
2591 
2592 			prepare_to_wait(&unregister_wait, &wait,
2593 					TASK_UNINTERRUPTIBLE);
2594 
2595 			mutex_unlock(&bch_register_lock);
2596 			schedule_timeout(timeout);
2597 		}
2598 
2599 		finish_wait(&unregister_wait, &wait);
2600 
2601 		if (stopped)
2602 			pr_info("All devices stopped");
2603 		else
2604 			pr_notice("Timeout waiting for devices to be closed");
2605 out:
2606 		mutex_unlock(&bch_register_lock);
2607 	}
2608 
2609 	return NOTIFY_DONE;
2610 }
2611 
2612 static struct notifier_block reboot = {
2613 	.notifier_call	= bcache_reboot,
2614 	.priority	= INT_MAX, /* before any real devices */
2615 };
2616 
2617 static void bcache_exit(void)
2618 {
2619 	bch_debug_exit();
2620 	bch_request_exit();
2621 	if (bcache_kobj)
2622 		kobject_put(bcache_kobj);
2623 	if (bcache_wq)
2624 		destroy_workqueue(bcache_wq);
2625 	if (bch_journal_wq)
2626 		destroy_workqueue(bch_journal_wq);
2627 
2628 	if (bcache_major)
2629 		unregister_blkdev(bcache_major, "bcache");
2630 	unregister_reboot_notifier(&reboot);
2631 	mutex_destroy(&bch_register_lock);
2632 }
2633 
2634 /* Check and fixup module parameters */
2635 static void check_module_parameters(void)
2636 {
2637 	if (bch_cutoff_writeback_sync == 0)
2638 		bch_cutoff_writeback_sync = CUTOFF_WRITEBACK_SYNC;
2639 	else if (bch_cutoff_writeback_sync > CUTOFF_WRITEBACK_SYNC_MAX) {
2640 		pr_warn("set bch_cutoff_writeback_sync (%u) to max value %u",
2641 			bch_cutoff_writeback_sync, CUTOFF_WRITEBACK_SYNC_MAX);
2642 		bch_cutoff_writeback_sync = CUTOFF_WRITEBACK_SYNC_MAX;
2643 	}
2644 
2645 	if (bch_cutoff_writeback == 0)
2646 		bch_cutoff_writeback = CUTOFF_WRITEBACK;
2647 	else if (bch_cutoff_writeback > CUTOFF_WRITEBACK_MAX) {
2648 		pr_warn("set bch_cutoff_writeback (%u) to max value %u",
2649 			bch_cutoff_writeback, CUTOFF_WRITEBACK_MAX);
2650 		bch_cutoff_writeback = CUTOFF_WRITEBACK_MAX;
2651 	}
2652 
2653 	if (bch_cutoff_writeback > bch_cutoff_writeback_sync) {
2654 		pr_warn("set bch_cutoff_writeback (%u) to %u",
2655 			bch_cutoff_writeback, bch_cutoff_writeback_sync);
2656 		bch_cutoff_writeback = bch_cutoff_writeback_sync;
2657 	}
2658 }
2659 
2660 static int __init bcache_init(void)
2661 {
2662 	static const struct attribute *files[] = {
2663 		&ksysfs_register.attr,
2664 		&ksysfs_register_quiet.attr,
2665 		&ksysfs_pendings_cleanup.attr,
2666 		NULL
2667 	};
2668 
2669 	check_module_parameters();
2670 
2671 	mutex_init(&bch_register_lock);
2672 	init_waitqueue_head(&unregister_wait);
2673 	register_reboot_notifier(&reboot);
2674 
2675 	bcache_major = register_blkdev(0, "bcache");
2676 	if (bcache_major < 0) {
2677 		unregister_reboot_notifier(&reboot);
2678 		mutex_destroy(&bch_register_lock);
2679 		return bcache_major;
2680 	}
2681 
2682 	bcache_wq = alloc_workqueue("bcache", WQ_MEM_RECLAIM, 0);
2683 	if (!bcache_wq)
2684 		goto err;
2685 
2686 	bch_journal_wq = alloc_workqueue("bch_journal", WQ_MEM_RECLAIM, 0);
2687 	if (!bch_journal_wq)
2688 		goto err;
2689 
2690 	bcache_kobj = kobject_create_and_add("bcache", fs_kobj);
2691 	if (!bcache_kobj)
2692 		goto err;
2693 
2694 	if (bch_request_init() ||
2695 	    sysfs_create_files(bcache_kobj, files))
2696 		goto err;
2697 
2698 	bch_debug_init();
2699 	closure_debug_init();
2700 
2701 	bcache_is_reboot = false;
2702 
2703 	return 0;
2704 err:
2705 	bcache_exit();
2706 	return -ENOMEM;
2707 }
2708 
2709 /*
2710  * Module hooks
2711  */
2712 module_exit(bcache_exit);
2713 module_init(bcache_init);
2714 
2715 module_param(bch_cutoff_writeback, uint, 0);
2716 MODULE_PARM_DESC(bch_cutoff_writeback, "threshold to cutoff writeback");
2717 
2718 module_param(bch_cutoff_writeback_sync, uint, 0);
2719 MODULE_PARM_DESC(bch_cutoff_writeback_sync, "hard threshold to cutoff writeback");
2720 
2721 MODULE_DESCRIPTION("Bcache: a Linux block layer cache");
2722 MODULE_AUTHOR("Kent Overstreet <kent.overstreet@gmail.com>");
2723 MODULE_LICENSE("GPL");
2724