xref: /openbmc/linux/drivers/md/bcache/super.c (revision 5ef12cb4a3a78ffb331c03a795a15eea4ae35155)
1 /*
2  * bcache setup/teardown code, and some metadata io - read a superblock and
3  * figure out what to do with it.
4  *
5  * Copyright 2010, 2011 Kent Overstreet <kent.overstreet@gmail.com>
6  * Copyright 2012 Google, Inc.
7  */
8 
9 #include "bcache.h"
10 #include "btree.h"
11 #include "debug.h"
12 #include "extents.h"
13 #include "request.h"
14 #include "writeback.h"
15 
16 #include <linux/blkdev.h>
17 #include <linux/buffer_head.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 MODULE_LICENSE("GPL");
28 MODULE_AUTHOR("Kent Overstreet <kent.overstreet@gmail.com>");
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 /* Default is -1; we skip past it for struct cached_dev's cache mode */
41 const char * const bch_cache_modes[] = {
42 	"default",
43 	"writethrough",
44 	"writeback",
45 	"writearound",
46 	"none",
47 	NULL
48 };
49 
50 /* Default is -1; we skip past it for stop_when_cache_set_failed */
51 const char * const bch_stop_on_failure_modes[] = {
52 	"default",
53 	"auto",
54 	"always",
55 	NULL
56 };
57 
58 static struct kobject *bcache_kobj;
59 struct mutex bch_register_lock;
60 LIST_HEAD(bch_cache_sets);
61 static LIST_HEAD(uncached_devices);
62 
63 static int bcache_major;
64 static DEFINE_IDA(bcache_device_idx);
65 static wait_queue_head_t unregister_wait;
66 struct workqueue_struct *bcache_wq;
67 
68 #define BTREE_MAX_PAGES		(256 * 1024 / PAGE_SIZE)
69 /* limitation of partitions number on single bcache device */
70 #define BCACHE_MINORS		128
71 /* limitation of bcache devices number on single system */
72 #define BCACHE_DEVICE_IDX_MAX	((1U << MINORBITS)/BCACHE_MINORS)
73 
74 /* Superblock */
75 
76 static const char *read_super(struct cache_sb *sb, struct block_device *bdev,
77 			      struct page **res)
78 {
79 	const char *err;
80 	struct cache_sb *s;
81 	struct buffer_head *bh = __bread(bdev, 1, SB_SIZE);
82 	unsigned i;
83 
84 	if (!bh)
85 		return "IO error";
86 
87 	s = (struct cache_sb *) bh->b_data;
88 
89 	sb->offset		= le64_to_cpu(s->offset);
90 	sb->version		= le64_to_cpu(s->version);
91 
92 	memcpy(sb->magic,	s->magic, 16);
93 	memcpy(sb->uuid,	s->uuid, 16);
94 	memcpy(sb->set_uuid,	s->set_uuid, 16);
95 	memcpy(sb->label,	s->label, SB_LABEL_SIZE);
96 
97 	sb->flags		= le64_to_cpu(s->flags);
98 	sb->seq			= le64_to_cpu(s->seq);
99 	sb->last_mount		= le32_to_cpu(s->last_mount);
100 	sb->first_bucket	= le16_to_cpu(s->first_bucket);
101 	sb->keys		= le16_to_cpu(s->keys);
102 
103 	for (i = 0; i < SB_JOURNAL_BUCKETS; i++)
104 		sb->d[i] = le64_to_cpu(s->d[i]);
105 
106 	pr_debug("read sb version %llu, flags %llu, seq %llu, journal size %u",
107 		 sb->version, sb->flags, sb->seq, sb->keys);
108 
109 	err = "Not a bcache superblock";
110 	if (sb->offset != SB_SECTOR)
111 		goto err;
112 
113 	if (memcmp(sb->magic, bcache_magic, 16))
114 		goto err;
115 
116 	err = "Too many journal buckets";
117 	if (sb->keys > SB_JOURNAL_BUCKETS)
118 		goto err;
119 
120 	err = "Bad checksum";
121 	if (s->csum != csum_set(s))
122 		goto err;
123 
124 	err = "Bad UUID";
125 	if (bch_is_zero(sb->uuid, 16))
126 		goto err;
127 
128 	sb->block_size	= le16_to_cpu(s->block_size);
129 
130 	err = "Superblock block size smaller than device block size";
131 	if (sb->block_size << 9 < bdev_logical_block_size(bdev))
132 		goto err;
133 
134 	switch (sb->version) {
135 	case BCACHE_SB_VERSION_BDEV:
136 		sb->data_offset	= BDEV_DATA_START_DEFAULT;
137 		break;
138 	case BCACHE_SB_VERSION_BDEV_WITH_OFFSET:
139 		sb->data_offset	= le64_to_cpu(s->data_offset);
140 
141 		err = "Bad data offset";
142 		if (sb->data_offset < BDEV_DATA_START_DEFAULT)
143 			goto err;
144 
145 		break;
146 	case BCACHE_SB_VERSION_CDEV:
147 	case BCACHE_SB_VERSION_CDEV_WITH_UUID:
148 		sb->nbuckets	= le64_to_cpu(s->nbuckets);
149 		sb->bucket_size	= le16_to_cpu(s->bucket_size);
150 
151 		sb->nr_in_set	= le16_to_cpu(s->nr_in_set);
152 		sb->nr_this_dev	= le16_to_cpu(s->nr_this_dev);
153 
154 		err = "Too many buckets";
155 		if (sb->nbuckets > LONG_MAX)
156 			goto err;
157 
158 		err = "Not enough buckets";
159 		if (sb->nbuckets < 1 << 7)
160 			goto err;
161 
162 		err = "Bad block/bucket size";
163 		if (!is_power_of_2(sb->block_size) ||
164 		    sb->block_size > PAGE_SECTORS ||
165 		    !is_power_of_2(sb->bucket_size) ||
166 		    sb->bucket_size < PAGE_SECTORS)
167 			goto err;
168 
169 		err = "Invalid superblock: device too small";
170 		if (get_capacity(bdev->bd_disk) < sb->bucket_size * sb->nbuckets)
171 			goto err;
172 
173 		err = "Bad UUID";
174 		if (bch_is_zero(sb->set_uuid, 16))
175 			goto err;
176 
177 		err = "Bad cache device number in set";
178 		if (!sb->nr_in_set ||
179 		    sb->nr_in_set <= sb->nr_this_dev ||
180 		    sb->nr_in_set > MAX_CACHES_PER_SET)
181 			goto err;
182 
183 		err = "Journal buckets not sequential";
184 		for (i = 0; i < sb->keys; i++)
185 			if (sb->d[i] != sb->first_bucket + i)
186 				goto err;
187 
188 		err = "Too many journal buckets";
189 		if (sb->first_bucket + sb->keys > sb->nbuckets)
190 			goto err;
191 
192 		err = "Invalid superblock: first bucket comes before end of super";
193 		if (sb->first_bucket * sb->bucket_size < 16)
194 			goto err;
195 
196 		break;
197 	default:
198 		err = "Unsupported superblock version";
199 		goto err;
200 	}
201 
202 	sb->last_mount = get_seconds();
203 	err = NULL;
204 
205 	get_page(bh->b_page);
206 	*res = bh->b_page;
207 err:
208 	put_bh(bh);
209 	return err;
210 }
211 
212 static void write_bdev_super_endio(struct bio *bio)
213 {
214 	struct cached_dev *dc = bio->bi_private;
215 	/* XXX: error checking */
216 
217 	closure_put(&dc->sb_write);
218 }
219 
220 static void __write_super(struct cache_sb *sb, struct bio *bio)
221 {
222 	struct cache_sb *out = page_address(bio_first_page_all(bio));
223 	unsigned i;
224 
225 	bio->bi_iter.bi_sector	= SB_SECTOR;
226 	bio->bi_iter.bi_size	= SB_SIZE;
227 	bio_set_op_attrs(bio, REQ_OP_WRITE, REQ_SYNC|REQ_META);
228 	bch_bio_map(bio, NULL);
229 
230 	out->offset		= cpu_to_le64(sb->offset);
231 	out->version		= cpu_to_le64(sb->version);
232 
233 	memcpy(out->uuid,	sb->uuid, 16);
234 	memcpy(out->set_uuid,	sb->set_uuid, 16);
235 	memcpy(out->label,	sb->label, SB_LABEL_SIZE);
236 
237 	out->flags		= cpu_to_le64(sb->flags);
238 	out->seq		= cpu_to_le64(sb->seq);
239 
240 	out->last_mount		= cpu_to_le32(sb->last_mount);
241 	out->first_bucket	= cpu_to_le16(sb->first_bucket);
242 	out->keys		= cpu_to_le16(sb->keys);
243 
244 	for (i = 0; i < sb->keys; i++)
245 		out->d[i] = cpu_to_le64(sb->d[i]);
246 
247 	out->csum = csum_set(out);
248 
249 	pr_debug("ver %llu, flags %llu, seq %llu",
250 		 sb->version, sb->flags, sb->seq);
251 
252 	submit_bio(bio);
253 }
254 
255 static void bch_write_bdev_super_unlock(struct closure *cl)
256 {
257 	struct cached_dev *dc = container_of(cl, struct cached_dev, sb_write);
258 
259 	up(&dc->sb_write_mutex);
260 }
261 
262 void bch_write_bdev_super(struct cached_dev *dc, struct closure *parent)
263 {
264 	struct closure *cl = &dc->sb_write;
265 	struct bio *bio = &dc->sb_bio;
266 
267 	down(&dc->sb_write_mutex);
268 	closure_init(cl, parent);
269 
270 	bio_reset(bio);
271 	bio_set_dev(bio, dc->bdev);
272 	bio->bi_end_io	= write_bdev_super_endio;
273 	bio->bi_private = dc;
274 
275 	closure_get(cl);
276 	/* I/O request sent to backing device */
277 	__write_super(&dc->sb, bio);
278 
279 	closure_return_with_destructor(cl, bch_write_bdev_super_unlock);
280 }
281 
282 static void write_super_endio(struct bio *bio)
283 {
284 	struct cache *ca = bio->bi_private;
285 
286 	/* is_read = 0 */
287 	bch_count_io_errors(ca, bio->bi_status, 0,
288 			    "writing superblock");
289 	closure_put(&ca->set->sb_write);
290 }
291 
292 static void bcache_write_super_unlock(struct closure *cl)
293 {
294 	struct cache_set *c = container_of(cl, struct cache_set, sb_write);
295 
296 	up(&c->sb_write_mutex);
297 }
298 
299 void bcache_write_super(struct cache_set *c)
300 {
301 	struct closure *cl = &c->sb_write;
302 	struct cache *ca;
303 	unsigned i;
304 
305 	down(&c->sb_write_mutex);
306 	closure_init(cl, &c->cl);
307 
308 	c->sb.seq++;
309 
310 	for_each_cache(ca, c, i) {
311 		struct bio *bio = &ca->sb_bio;
312 
313 		ca->sb.version		= BCACHE_SB_VERSION_CDEV_WITH_UUID;
314 		ca->sb.seq		= c->sb.seq;
315 		ca->sb.last_mount	= c->sb.last_mount;
316 
317 		SET_CACHE_SYNC(&ca->sb, CACHE_SYNC(&c->sb));
318 
319 		bio_reset(bio);
320 		bio_set_dev(bio, ca->bdev);
321 		bio->bi_end_io	= write_super_endio;
322 		bio->bi_private = ca;
323 
324 		closure_get(cl);
325 		__write_super(&ca->sb, bio);
326 	}
327 
328 	closure_return_with_destructor(cl, bcache_write_super_unlock);
329 }
330 
331 /* UUID io */
332 
333 static void uuid_endio(struct bio *bio)
334 {
335 	struct closure *cl = bio->bi_private;
336 	struct cache_set *c = container_of(cl, struct cache_set, uuid_write);
337 
338 	cache_set_err_on(bio->bi_status, c, "accessing uuids");
339 	bch_bbio_free(bio, c);
340 	closure_put(cl);
341 }
342 
343 static void uuid_io_unlock(struct closure *cl)
344 {
345 	struct cache_set *c = container_of(cl, struct cache_set, uuid_write);
346 
347 	up(&c->uuid_write_mutex);
348 }
349 
350 static void uuid_io(struct cache_set *c, int op, unsigned long op_flags,
351 		    struct bkey *k, struct closure *parent)
352 {
353 	struct closure *cl = &c->uuid_write;
354 	struct uuid_entry *u;
355 	unsigned i;
356 	char buf[80];
357 
358 	BUG_ON(!parent);
359 	down(&c->uuid_write_mutex);
360 	closure_init(cl, parent);
361 
362 	for (i = 0; i < KEY_PTRS(k); i++) {
363 		struct bio *bio = bch_bbio_alloc(c);
364 
365 		bio->bi_opf = REQ_SYNC | REQ_META | op_flags;
366 		bio->bi_iter.bi_size = KEY_SIZE(k) << 9;
367 
368 		bio->bi_end_io	= uuid_endio;
369 		bio->bi_private = cl;
370 		bio_set_op_attrs(bio, op, REQ_SYNC|REQ_META|op_flags);
371 		bch_bio_map(bio, c->uuids);
372 
373 		bch_submit_bbio(bio, c, k, i);
374 
375 		if (op != REQ_OP_WRITE)
376 			break;
377 	}
378 
379 	bch_extent_to_text(buf, sizeof(buf), k);
380 	pr_debug("%s UUIDs at %s", op == REQ_OP_WRITE ? "wrote" : "read", buf);
381 
382 	for (u = c->uuids; u < c->uuids + c->nr_uuids; u++)
383 		if (!bch_is_zero(u->uuid, 16))
384 			pr_debug("Slot %zi: %pU: %s: 1st: %u last: %u inv: %u",
385 				 u - c->uuids, u->uuid, u->label,
386 				 u->first_reg, u->last_reg, u->invalidated);
387 
388 	closure_return_with_destructor(cl, uuid_io_unlock);
389 }
390 
391 static char *uuid_read(struct cache_set *c, struct jset *j, struct closure *cl)
392 {
393 	struct bkey *k = &j->uuid_bucket;
394 
395 	if (__bch_btree_ptr_invalid(c, k))
396 		return "bad uuid pointer";
397 
398 	bkey_copy(&c->uuid_bucket, k);
399 	uuid_io(c, REQ_OP_READ, 0, k, cl);
400 
401 	if (j->version < BCACHE_JSET_VERSION_UUIDv1) {
402 		struct uuid_entry_v0	*u0 = (void *) c->uuids;
403 		struct uuid_entry	*u1 = (void *) c->uuids;
404 		int i;
405 
406 		closure_sync(cl);
407 
408 		/*
409 		 * Since the new uuid entry is bigger than the old, we have to
410 		 * convert starting at the highest memory address and work down
411 		 * in order to do it in place
412 		 */
413 
414 		for (i = c->nr_uuids - 1;
415 		     i >= 0;
416 		     --i) {
417 			memcpy(u1[i].uuid,	u0[i].uuid, 16);
418 			memcpy(u1[i].label,	u0[i].label, 32);
419 
420 			u1[i].first_reg		= u0[i].first_reg;
421 			u1[i].last_reg		= u0[i].last_reg;
422 			u1[i].invalidated	= u0[i].invalidated;
423 
424 			u1[i].flags	= 0;
425 			u1[i].sectors	= 0;
426 		}
427 	}
428 
429 	return NULL;
430 }
431 
432 static int __uuid_write(struct cache_set *c)
433 {
434 	BKEY_PADDED(key) k;
435 	struct closure cl;
436 	closure_init_stack(&cl);
437 
438 	lockdep_assert_held(&bch_register_lock);
439 
440 	if (bch_bucket_alloc_set(c, RESERVE_BTREE, &k.key, 1, true))
441 		return 1;
442 
443 	SET_KEY_SIZE(&k.key, c->sb.bucket_size);
444 	uuid_io(c, REQ_OP_WRITE, 0, &k.key, &cl);
445 	closure_sync(&cl);
446 
447 	bkey_copy(&c->uuid_bucket, &k.key);
448 	bkey_put(c, &k.key);
449 	return 0;
450 }
451 
452 int bch_uuid_write(struct cache_set *c)
453 {
454 	int ret = __uuid_write(c);
455 
456 	if (!ret)
457 		bch_journal_meta(c, NULL);
458 
459 	return ret;
460 }
461 
462 static struct uuid_entry *uuid_find(struct cache_set *c, const char *uuid)
463 {
464 	struct uuid_entry *u;
465 
466 	for (u = c->uuids;
467 	     u < c->uuids + c->nr_uuids; u++)
468 		if (!memcmp(u->uuid, uuid, 16))
469 			return u;
470 
471 	return NULL;
472 }
473 
474 static struct uuid_entry *uuid_find_empty(struct cache_set *c)
475 {
476 	static const char zero_uuid[16] = "\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0";
477 	return uuid_find(c, zero_uuid);
478 }
479 
480 /*
481  * Bucket priorities/gens:
482  *
483  * For each bucket, we store on disk its
484    * 8 bit gen
485    * 16 bit priority
486  *
487  * See alloc.c for an explanation of the gen. The priority is used to implement
488  * lru (and in the future other) cache replacement policies; for most purposes
489  * it's just an opaque integer.
490  *
491  * The gens and the priorities don't have a whole lot to do with each other, and
492  * it's actually the gens that must be written out at specific times - it's no
493  * big deal if the priorities don't get written, if we lose them we just reuse
494  * buckets in suboptimal order.
495  *
496  * On disk they're stored in a packed array, and in as many buckets are required
497  * to fit them all. The buckets we use to store them form a list; the journal
498  * header points to the first bucket, the first bucket points to the second
499  * bucket, et cetera.
500  *
501  * This code is used by the allocation code; periodically (whenever it runs out
502  * of buckets to allocate from) the allocation code will invalidate some
503  * buckets, but it can't use those buckets until their new gens are safely on
504  * disk.
505  */
506 
507 static void prio_endio(struct bio *bio)
508 {
509 	struct cache *ca = bio->bi_private;
510 
511 	cache_set_err_on(bio->bi_status, ca->set, "accessing priorities");
512 	bch_bbio_free(bio, ca->set);
513 	closure_put(&ca->prio);
514 }
515 
516 static void prio_io(struct cache *ca, uint64_t bucket, int op,
517 		    unsigned long op_flags)
518 {
519 	struct closure *cl = &ca->prio;
520 	struct bio *bio = bch_bbio_alloc(ca->set);
521 
522 	closure_init_stack(cl);
523 
524 	bio->bi_iter.bi_sector	= bucket * ca->sb.bucket_size;
525 	bio_set_dev(bio, ca->bdev);
526 	bio->bi_iter.bi_size	= bucket_bytes(ca);
527 
528 	bio->bi_end_io	= prio_endio;
529 	bio->bi_private = ca;
530 	bio_set_op_attrs(bio, op, REQ_SYNC|REQ_META|op_flags);
531 	bch_bio_map(bio, ca->disk_buckets);
532 
533 	closure_bio_submit(ca->set, bio, &ca->prio);
534 	closure_sync(cl);
535 }
536 
537 void bch_prio_write(struct cache *ca)
538 {
539 	int i;
540 	struct bucket *b;
541 	struct closure cl;
542 
543 	closure_init_stack(&cl);
544 
545 	lockdep_assert_held(&ca->set->bucket_lock);
546 
547 	ca->disk_buckets->seq++;
548 
549 	atomic_long_add(ca->sb.bucket_size * prio_buckets(ca),
550 			&ca->meta_sectors_written);
551 
552 	//pr_debug("free %zu, free_inc %zu, unused %zu", fifo_used(&ca->free),
553 	//	 fifo_used(&ca->free_inc), fifo_used(&ca->unused));
554 
555 	for (i = prio_buckets(ca) - 1; i >= 0; --i) {
556 		long bucket;
557 		struct prio_set *p = ca->disk_buckets;
558 		struct bucket_disk *d = p->data;
559 		struct bucket_disk *end = d + prios_per_bucket(ca);
560 
561 		for (b = ca->buckets + i * prios_per_bucket(ca);
562 		     b < ca->buckets + ca->sb.nbuckets && d < end;
563 		     b++, d++) {
564 			d->prio = cpu_to_le16(b->prio);
565 			d->gen = b->gen;
566 		}
567 
568 		p->next_bucket	= ca->prio_buckets[i + 1];
569 		p->magic	= pset_magic(&ca->sb);
570 		p->csum		= bch_crc64(&p->magic, bucket_bytes(ca) - 8);
571 
572 		bucket = bch_bucket_alloc(ca, RESERVE_PRIO, true);
573 		BUG_ON(bucket == -1);
574 
575 		mutex_unlock(&ca->set->bucket_lock);
576 		prio_io(ca, bucket, REQ_OP_WRITE, 0);
577 		mutex_lock(&ca->set->bucket_lock);
578 
579 		ca->prio_buckets[i] = bucket;
580 		atomic_dec_bug(&ca->buckets[bucket].pin);
581 	}
582 
583 	mutex_unlock(&ca->set->bucket_lock);
584 
585 	bch_journal_meta(ca->set, &cl);
586 	closure_sync(&cl);
587 
588 	mutex_lock(&ca->set->bucket_lock);
589 
590 	/*
591 	 * Don't want the old priorities to get garbage collected until after we
592 	 * finish writing the new ones, and they're journalled
593 	 */
594 	for (i = 0; i < prio_buckets(ca); i++) {
595 		if (ca->prio_last_buckets[i])
596 			__bch_bucket_free(ca,
597 				&ca->buckets[ca->prio_last_buckets[i]]);
598 
599 		ca->prio_last_buckets[i] = ca->prio_buckets[i];
600 	}
601 }
602 
603 static void prio_read(struct cache *ca, uint64_t bucket)
604 {
605 	struct prio_set *p = ca->disk_buckets;
606 	struct bucket_disk *d = p->data + prios_per_bucket(ca), *end = d;
607 	struct bucket *b;
608 	unsigned bucket_nr = 0;
609 
610 	for (b = ca->buckets;
611 	     b < ca->buckets + ca->sb.nbuckets;
612 	     b++, d++) {
613 		if (d == end) {
614 			ca->prio_buckets[bucket_nr] = bucket;
615 			ca->prio_last_buckets[bucket_nr] = bucket;
616 			bucket_nr++;
617 
618 			prio_io(ca, bucket, REQ_OP_READ, 0);
619 
620 			if (p->csum != bch_crc64(&p->magic, bucket_bytes(ca) - 8))
621 				pr_warn("bad csum reading priorities");
622 
623 			if (p->magic != pset_magic(&ca->sb))
624 				pr_warn("bad magic reading priorities");
625 
626 			bucket = p->next_bucket;
627 			d = p->data;
628 		}
629 
630 		b->prio = le16_to_cpu(d->prio);
631 		b->gen = b->last_gc = d->gen;
632 	}
633 }
634 
635 /* Bcache device */
636 
637 static int open_dev(struct block_device *b, fmode_t mode)
638 {
639 	struct bcache_device *d = b->bd_disk->private_data;
640 	if (test_bit(BCACHE_DEV_CLOSING, &d->flags))
641 		return -ENXIO;
642 
643 	closure_get(&d->cl);
644 	return 0;
645 }
646 
647 static void release_dev(struct gendisk *b, fmode_t mode)
648 {
649 	struct bcache_device *d = b->private_data;
650 	closure_put(&d->cl);
651 }
652 
653 static int ioctl_dev(struct block_device *b, fmode_t mode,
654 		     unsigned int cmd, unsigned long arg)
655 {
656 	struct bcache_device *d = b->bd_disk->private_data;
657 	return d->ioctl(d, mode, cmd, arg);
658 }
659 
660 static const struct block_device_operations bcache_ops = {
661 	.open		= open_dev,
662 	.release	= release_dev,
663 	.ioctl		= ioctl_dev,
664 	.owner		= THIS_MODULE,
665 };
666 
667 void bcache_device_stop(struct bcache_device *d)
668 {
669 	if (!test_and_set_bit(BCACHE_DEV_CLOSING, &d->flags))
670 		closure_queue(&d->cl);
671 }
672 
673 static void bcache_device_unlink(struct bcache_device *d)
674 {
675 	lockdep_assert_held(&bch_register_lock);
676 
677 	if (d->c && !test_and_set_bit(BCACHE_DEV_UNLINK_DONE, &d->flags)) {
678 		unsigned i;
679 		struct cache *ca;
680 
681 		sysfs_remove_link(&d->c->kobj, d->name);
682 		sysfs_remove_link(&d->kobj, "cache");
683 
684 		for_each_cache(ca, d->c, i)
685 			bd_unlink_disk_holder(ca->bdev, d->disk);
686 	}
687 }
688 
689 static void bcache_device_link(struct bcache_device *d, struct cache_set *c,
690 			       const char *name)
691 {
692 	unsigned i;
693 	struct cache *ca;
694 
695 	for_each_cache(ca, d->c, i)
696 		bd_link_disk_holder(ca->bdev, d->disk);
697 
698 	snprintf(d->name, BCACHEDEVNAME_SIZE,
699 		 "%s%u", name, d->id);
700 
701 	WARN(sysfs_create_link(&d->kobj, &c->kobj, "cache") ||
702 	     sysfs_create_link(&c->kobj, &d->kobj, d->name),
703 	     "Couldn't create device <-> cache set symlinks");
704 
705 	clear_bit(BCACHE_DEV_UNLINK_DONE, &d->flags);
706 }
707 
708 static void bcache_device_detach(struct bcache_device *d)
709 {
710 	lockdep_assert_held(&bch_register_lock);
711 
712 	if (test_bit(BCACHE_DEV_DETACHING, &d->flags)) {
713 		struct uuid_entry *u = d->c->uuids + d->id;
714 
715 		SET_UUID_FLASH_ONLY(u, 0);
716 		memcpy(u->uuid, invalid_uuid, 16);
717 		u->invalidated = cpu_to_le32(get_seconds());
718 		bch_uuid_write(d->c);
719 	}
720 
721 	bcache_device_unlink(d);
722 
723 	d->c->devices[d->id] = NULL;
724 	closure_put(&d->c->caching);
725 	d->c = NULL;
726 }
727 
728 static void bcache_device_attach(struct bcache_device *d, struct cache_set *c,
729 				 unsigned id)
730 {
731 	d->id = id;
732 	d->c = c;
733 	c->devices[id] = d;
734 
735 	if (id >= c->devices_max_used)
736 		c->devices_max_used = id + 1;
737 
738 	closure_get(&c->caching);
739 }
740 
741 static inline int first_minor_to_idx(int first_minor)
742 {
743 	return (first_minor/BCACHE_MINORS);
744 }
745 
746 static inline int idx_to_first_minor(int idx)
747 {
748 	return (idx * BCACHE_MINORS);
749 }
750 
751 static void bcache_device_free(struct bcache_device *d)
752 {
753 	lockdep_assert_held(&bch_register_lock);
754 
755 	pr_info("%s stopped", d->disk->disk_name);
756 
757 	if (d->c)
758 		bcache_device_detach(d);
759 	if (d->disk && d->disk->flags & GENHD_FL_UP)
760 		del_gendisk(d->disk);
761 	if (d->disk && d->disk->queue)
762 		blk_cleanup_queue(d->disk->queue);
763 	if (d->disk) {
764 		ida_simple_remove(&bcache_device_idx,
765 				  first_minor_to_idx(d->disk->first_minor));
766 		put_disk(d->disk);
767 	}
768 
769 	if (d->bio_split)
770 		bioset_free(d->bio_split);
771 	kvfree(d->full_dirty_stripes);
772 	kvfree(d->stripe_sectors_dirty);
773 
774 	closure_debug_destroy(&d->cl);
775 }
776 
777 static int bcache_device_init(struct bcache_device *d, unsigned block_size,
778 			      sector_t sectors)
779 {
780 	struct request_queue *q;
781 	const size_t max_stripes = min_t(size_t, INT_MAX,
782 					 SIZE_MAX / sizeof(atomic_t));
783 	size_t n;
784 	int idx;
785 
786 	if (!d->stripe_size)
787 		d->stripe_size = 1 << 31;
788 
789 	d->nr_stripes = DIV_ROUND_UP_ULL(sectors, d->stripe_size);
790 
791 	if (!d->nr_stripes || d->nr_stripes > max_stripes) {
792 		pr_err("nr_stripes too large or invalid: %u (start sector beyond end of disk?)",
793 			(unsigned)d->nr_stripes);
794 		return -ENOMEM;
795 	}
796 
797 	n = d->nr_stripes * sizeof(atomic_t);
798 	d->stripe_sectors_dirty = kvzalloc(n, GFP_KERNEL);
799 	if (!d->stripe_sectors_dirty)
800 		return -ENOMEM;
801 
802 	n = BITS_TO_LONGS(d->nr_stripes) * sizeof(unsigned long);
803 	d->full_dirty_stripes = kvzalloc(n, GFP_KERNEL);
804 	if (!d->full_dirty_stripes)
805 		return -ENOMEM;
806 
807 	idx = ida_simple_get(&bcache_device_idx, 0,
808 				BCACHE_DEVICE_IDX_MAX, GFP_KERNEL);
809 	if (idx < 0)
810 		return idx;
811 
812 	if (!(d->bio_split = bioset_create(4, offsetof(struct bbio, bio),
813 					   BIOSET_NEED_BVECS |
814 					   BIOSET_NEED_RESCUER)) ||
815 	    !(d->disk = alloc_disk(BCACHE_MINORS))) {
816 		ida_simple_remove(&bcache_device_idx, idx);
817 		return -ENOMEM;
818 	}
819 
820 	set_capacity(d->disk, sectors);
821 	snprintf(d->disk->disk_name, DISK_NAME_LEN, "bcache%i", idx);
822 
823 	d->disk->major		= bcache_major;
824 	d->disk->first_minor	= idx_to_first_minor(idx);
825 	d->disk->fops		= &bcache_ops;
826 	d->disk->private_data	= d;
827 
828 	q = blk_alloc_queue(GFP_KERNEL);
829 	if (!q)
830 		return -ENOMEM;
831 
832 	blk_queue_make_request(q, NULL);
833 	d->disk->queue			= q;
834 	q->queuedata			= d;
835 	q->backing_dev_info->congested_data = d;
836 	q->limits.max_hw_sectors	= UINT_MAX;
837 	q->limits.max_sectors		= UINT_MAX;
838 	q->limits.max_segment_size	= UINT_MAX;
839 	q->limits.max_segments		= BIO_MAX_PAGES;
840 	blk_queue_max_discard_sectors(q, UINT_MAX);
841 	q->limits.discard_granularity	= 512;
842 	q->limits.io_min		= block_size;
843 	q->limits.logical_block_size	= block_size;
844 	q->limits.physical_block_size	= block_size;
845 	blk_queue_flag_set(QUEUE_FLAG_NONROT, d->disk->queue);
846 	blk_queue_flag_clear(QUEUE_FLAG_ADD_RANDOM, d->disk->queue);
847 	blk_queue_flag_set(QUEUE_FLAG_DISCARD, d->disk->queue);
848 
849 	blk_queue_write_cache(q, true, true);
850 
851 	return 0;
852 }
853 
854 /* Cached device */
855 
856 static void calc_cached_dev_sectors(struct cache_set *c)
857 {
858 	uint64_t sectors = 0;
859 	struct cached_dev *dc;
860 
861 	list_for_each_entry(dc, &c->cached_devs, list)
862 		sectors += bdev_sectors(dc->bdev);
863 
864 	c->cached_dev_sectors = sectors;
865 }
866 
867 void bch_cached_dev_run(struct cached_dev *dc)
868 {
869 	struct bcache_device *d = &dc->disk;
870 	char buf[SB_LABEL_SIZE + 1];
871 	char *env[] = {
872 		"DRIVER=bcache",
873 		kasprintf(GFP_KERNEL, "CACHED_UUID=%pU", dc->sb.uuid),
874 		NULL,
875 		NULL,
876 	};
877 
878 	memcpy(buf, dc->sb.label, SB_LABEL_SIZE);
879 	buf[SB_LABEL_SIZE] = '\0';
880 	env[2] = kasprintf(GFP_KERNEL, "CACHED_LABEL=%s", buf);
881 
882 	if (atomic_xchg(&dc->running, 1)) {
883 		kfree(env[1]);
884 		kfree(env[2]);
885 		return;
886 	}
887 
888 	if (!d->c &&
889 	    BDEV_STATE(&dc->sb) != BDEV_STATE_NONE) {
890 		struct closure cl;
891 		closure_init_stack(&cl);
892 
893 		SET_BDEV_STATE(&dc->sb, BDEV_STATE_STALE);
894 		bch_write_bdev_super(dc, &cl);
895 		closure_sync(&cl);
896 	}
897 
898 	add_disk(d->disk);
899 	bd_link_disk_holder(dc->bdev, dc->disk.disk);
900 	/* won't show up in the uevent file, use udevadm monitor -e instead
901 	 * only class / kset properties are persistent */
902 	kobject_uevent_env(&disk_to_dev(d->disk)->kobj, KOBJ_CHANGE, env);
903 	kfree(env[1]);
904 	kfree(env[2]);
905 
906 	if (sysfs_create_link(&d->kobj, &disk_to_dev(d->disk)->kobj, "dev") ||
907 	    sysfs_create_link(&disk_to_dev(d->disk)->kobj, &d->kobj, "bcache"))
908 		pr_debug("error creating sysfs link");
909 }
910 
911 /*
912  * If BCACHE_DEV_RATE_DW_RUNNING is set, it means routine of the delayed
913  * work dc->writeback_rate_update is running. Wait until the routine
914  * quits (BCACHE_DEV_RATE_DW_RUNNING is clear), then continue to
915  * cancel it. If BCACHE_DEV_RATE_DW_RUNNING is not clear after time_out
916  * seconds, give up waiting here and continue to cancel it too.
917  */
918 static void cancel_writeback_rate_update_dwork(struct cached_dev *dc)
919 {
920 	int time_out = WRITEBACK_RATE_UPDATE_SECS_MAX * HZ;
921 
922 	do {
923 		if (!test_bit(BCACHE_DEV_RATE_DW_RUNNING,
924 			      &dc->disk.flags))
925 			break;
926 		time_out--;
927 		schedule_timeout_interruptible(1);
928 	} while (time_out > 0);
929 
930 	if (time_out == 0)
931 		pr_warn("give up waiting for dc->writeback_write_update to quit");
932 
933 	cancel_delayed_work_sync(&dc->writeback_rate_update);
934 }
935 
936 static void cached_dev_detach_finish(struct work_struct *w)
937 {
938 	struct cached_dev *dc = container_of(w, struct cached_dev, detach);
939 	struct closure cl;
940 	closure_init_stack(&cl);
941 
942 	BUG_ON(!test_bit(BCACHE_DEV_DETACHING, &dc->disk.flags));
943 	BUG_ON(refcount_read(&dc->count));
944 
945 	mutex_lock(&bch_register_lock);
946 
947 	if (test_and_clear_bit(BCACHE_DEV_WB_RUNNING, &dc->disk.flags))
948 		cancel_writeback_rate_update_dwork(dc);
949 
950 	if (!IS_ERR_OR_NULL(dc->writeback_thread)) {
951 		kthread_stop(dc->writeback_thread);
952 		dc->writeback_thread = NULL;
953 	}
954 
955 	memset(&dc->sb.set_uuid, 0, 16);
956 	SET_BDEV_STATE(&dc->sb, BDEV_STATE_NONE);
957 
958 	bch_write_bdev_super(dc, &cl);
959 	closure_sync(&cl);
960 
961 	bcache_device_detach(&dc->disk);
962 	list_move(&dc->list, &uncached_devices);
963 
964 	clear_bit(BCACHE_DEV_DETACHING, &dc->disk.flags);
965 	clear_bit(BCACHE_DEV_UNLINK_DONE, &dc->disk.flags);
966 
967 	mutex_unlock(&bch_register_lock);
968 
969 	pr_info("Caching disabled for %s", dc->backing_dev_name);
970 
971 	/* Drop ref we took in cached_dev_detach() */
972 	closure_put(&dc->disk.cl);
973 }
974 
975 void bch_cached_dev_detach(struct cached_dev *dc)
976 {
977 	lockdep_assert_held(&bch_register_lock);
978 
979 	if (test_bit(BCACHE_DEV_CLOSING, &dc->disk.flags))
980 		return;
981 
982 	if (test_and_set_bit(BCACHE_DEV_DETACHING, &dc->disk.flags))
983 		return;
984 
985 	/*
986 	 * Block the device from being closed and freed until we're finished
987 	 * detaching
988 	 */
989 	closure_get(&dc->disk.cl);
990 
991 	bch_writeback_queue(dc);
992 
993 	cached_dev_put(dc);
994 }
995 
996 int bch_cached_dev_attach(struct cached_dev *dc, struct cache_set *c,
997 			  uint8_t *set_uuid)
998 {
999 	uint32_t rtime = cpu_to_le32(get_seconds());
1000 	struct uuid_entry *u;
1001 	struct cached_dev *exist_dc, *t;
1002 
1003 	if ((set_uuid && memcmp(set_uuid, c->sb.set_uuid, 16)) ||
1004 	    (!set_uuid && memcmp(dc->sb.set_uuid, c->sb.set_uuid, 16)))
1005 		return -ENOENT;
1006 
1007 	if (dc->disk.c) {
1008 		pr_err("Can't attach %s: already attached",
1009 		       dc->backing_dev_name);
1010 		return -EINVAL;
1011 	}
1012 
1013 	if (test_bit(CACHE_SET_STOPPING, &c->flags)) {
1014 		pr_err("Can't attach %s: shutting down",
1015 		       dc->backing_dev_name);
1016 		return -EINVAL;
1017 	}
1018 
1019 	if (dc->sb.block_size < c->sb.block_size) {
1020 		/* Will die */
1021 		pr_err("Couldn't attach %s: block size less than set's block size",
1022 		       dc->backing_dev_name);
1023 		return -EINVAL;
1024 	}
1025 
1026 	/* Check whether already attached */
1027 	list_for_each_entry_safe(exist_dc, t, &c->cached_devs, list) {
1028 		if (!memcmp(dc->sb.uuid, exist_dc->sb.uuid, 16)) {
1029 			pr_err("Tried to attach %s but duplicate UUID already attached",
1030 				dc->backing_dev_name);
1031 
1032 			return -EINVAL;
1033 		}
1034 	}
1035 
1036 	u = uuid_find(c, dc->sb.uuid);
1037 
1038 	if (u &&
1039 	    (BDEV_STATE(&dc->sb) == BDEV_STATE_STALE ||
1040 	     BDEV_STATE(&dc->sb) == BDEV_STATE_NONE)) {
1041 		memcpy(u->uuid, invalid_uuid, 16);
1042 		u->invalidated = cpu_to_le32(get_seconds());
1043 		u = NULL;
1044 	}
1045 
1046 	if (!u) {
1047 		if (BDEV_STATE(&dc->sb) == BDEV_STATE_DIRTY) {
1048 			pr_err("Couldn't find uuid for %s in set",
1049 			       dc->backing_dev_name);
1050 			return -ENOENT;
1051 		}
1052 
1053 		u = uuid_find_empty(c);
1054 		if (!u) {
1055 			pr_err("Not caching %s, no room for UUID",
1056 			       dc->backing_dev_name);
1057 			return -EINVAL;
1058 		}
1059 	}
1060 
1061 	/* Deadlocks since we're called via sysfs...
1062 	sysfs_remove_file(&dc->kobj, &sysfs_attach);
1063 	 */
1064 
1065 	if (bch_is_zero(u->uuid, 16)) {
1066 		struct closure cl;
1067 		closure_init_stack(&cl);
1068 
1069 		memcpy(u->uuid, dc->sb.uuid, 16);
1070 		memcpy(u->label, dc->sb.label, SB_LABEL_SIZE);
1071 		u->first_reg = u->last_reg = rtime;
1072 		bch_uuid_write(c);
1073 
1074 		memcpy(dc->sb.set_uuid, c->sb.set_uuid, 16);
1075 		SET_BDEV_STATE(&dc->sb, BDEV_STATE_CLEAN);
1076 
1077 		bch_write_bdev_super(dc, &cl);
1078 		closure_sync(&cl);
1079 	} else {
1080 		u->last_reg = rtime;
1081 		bch_uuid_write(c);
1082 	}
1083 
1084 	bcache_device_attach(&dc->disk, c, u - c->uuids);
1085 	list_move(&dc->list, &c->cached_devs);
1086 	calc_cached_dev_sectors(c);
1087 
1088 	smp_wmb();
1089 	/*
1090 	 * dc->c must be set before dc->count != 0 - paired with the mb in
1091 	 * cached_dev_get()
1092 	 */
1093 	refcount_set(&dc->count, 1);
1094 
1095 	/* Block writeback thread, but spawn it */
1096 	down_write(&dc->writeback_lock);
1097 	if (bch_cached_dev_writeback_start(dc)) {
1098 		up_write(&dc->writeback_lock);
1099 		return -ENOMEM;
1100 	}
1101 
1102 	if (BDEV_STATE(&dc->sb) == BDEV_STATE_DIRTY) {
1103 		bch_sectors_dirty_init(&dc->disk);
1104 		atomic_set(&dc->has_dirty, 1);
1105 		bch_writeback_queue(dc);
1106 	}
1107 
1108 	bch_cached_dev_run(dc);
1109 	bcache_device_link(&dc->disk, c, "bdev");
1110 
1111 	/* Allow the writeback thread to proceed */
1112 	up_write(&dc->writeback_lock);
1113 
1114 	pr_info("Caching %s as %s on set %pU",
1115 		dc->backing_dev_name,
1116 		dc->disk.disk->disk_name,
1117 		dc->disk.c->sb.set_uuid);
1118 	return 0;
1119 }
1120 
1121 void bch_cached_dev_release(struct kobject *kobj)
1122 {
1123 	struct cached_dev *dc = container_of(kobj, struct cached_dev,
1124 					     disk.kobj);
1125 	kfree(dc);
1126 	module_put(THIS_MODULE);
1127 }
1128 
1129 static void cached_dev_free(struct closure *cl)
1130 {
1131 	struct cached_dev *dc = container_of(cl, struct cached_dev, disk.cl);
1132 
1133 	mutex_lock(&bch_register_lock);
1134 
1135 	if (test_and_clear_bit(BCACHE_DEV_WB_RUNNING, &dc->disk.flags))
1136 		cancel_writeback_rate_update_dwork(dc);
1137 
1138 	if (!IS_ERR_OR_NULL(dc->writeback_thread))
1139 		kthread_stop(dc->writeback_thread);
1140 	if (dc->writeback_write_wq)
1141 		destroy_workqueue(dc->writeback_write_wq);
1142 
1143 	if (atomic_read(&dc->running))
1144 		bd_unlink_disk_holder(dc->bdev, dc->disk.disk);
1145 	bcache_device_free(&dc->disk);
1146 	list_del(&dc->list);
1147 
1148 	mutex_unlock(&bch_register_lock);
1149 
1150 	if (!IS_ERR_OR_NULL(dc->bdev))
1151 		blkdev_put(dc->bdev, FMODE_READ|FMODE_WRITE|FMODE_EXCL);
1152 
1153 	wake_up(&unregister_wait);
1154 
1155 	kobject_put(&dc->disk.kobj);
1156 }
1157 
1158 static void cached_dev_flush(struct closure *cl)
1159 {
1160 	struct cached_dev *dc = container_of(cl, struct cached_dev, disk.cl);
1161 	struct bcache_device *d = &dc->disk;
1162 
1163 	mutex_lock(&bch_register_lock);
1164 	bcache_device_unlink(d);
1165 	mutex_unlock(&bch_register_lock);
1166 
1167 	bch_cache_accounting_destroy(&dc->accounting);
1168 	kobject_del(&d->kobj);
1169 
1170 	continue_at(cl, cached_dev_free, system_wq);
1171 }
1172 
1173 static int cached_dev_init(struct cached_dev *dc, unsigned block_size)
1174 {
1175 	int ret;
1176 	struct io *io;
1177 	struct request_queue *q = bdev_get_queue(dc->bdev);
1178 
1179 	__module_get(THIS_MODULE);
1180 	INIT_LIST_HEAD(&dc->list);
1181 	closure_init(&dc->disk.cl, NULL);
1182 	set_closure_fn(&dc->disk.cl, cached_dev_flush, system_wq);
1183 	kobject_init(&dc->disk.kobj, &bch_cached_dev_ktype);
1184 	INIT_WORK(&dc->detach, cached_dev_detach_finish);
1185 	sema_init(&dc->sb_write_mutex, 1);
1186 	INIT_LIST_HEAD(&dc->io_lru);
1187 	spin_lock_init(&dc->io_lock);
1188 	bch_cache_accounting_init(&dc->accounting, &dc->disk.cl);
1189 
1190 	dc->sequential_cutoff		= 4 << 20;
1191 
1192 	for (io = dc->io; io < dc->io + RECENT_IO; io++) {
1193 		list_add(&io->lru, &dc->io_lru);
1194 		hlist_add_head(&io->hash, dc->io_hash + RECENT_IO);
1195 	}
1196 
1197 	dc->disk.stripe_size = q->limits.io_opt >> 9;
1198 
1199 	if (dc->disk.stripe_size)
1200 		dc->partial_stripes_expensive =
1201 			q->limits.raid_partial_stripes_expensive;
1202 
1203 	ret = bcache_device_init(&dc->disk, block_size,
1204 			 dc->bdev->bd_part->nr_sects - dc->sb.data_offset);
1205 	if (ret)
1206 		return ret;
1207 
1208 	dc->disk.disk->queue->backing_dev_info->ra_pages =
1209 		max(dc->disk.disk->queue->backing_dev_info->ra_pages,
1210 		    q->backing_dev_info->ra_pages);
1211 
1212 	atomic_set(&dc->io_errors, 0);
1213 	dc->io_disable = false;
1214 	dc->error_limit = DEFAULT_CACHED_DEV_ERROR_LIMIT;
1215 	/* default to auto */
1216 	dc->stop_when_cache_set_failed = BCH_CACHED_DEV_STOP_AUTO;
1217 
1218 	bch_cached_dev_request_init(dc);
1219 	bch_cached_dev_writeback_init(dc);
1220 	return 0;
1221 }
1222 
1223 /* Cached device - bcache superblock */
1224 
1225 static void register_bdev(struct cache_sb *sb, struct page *sb_page,
1226 				 struct block_device *bdev,
1227 				 struct cached_dev *dc)
1228 {
1229 	const char *err = "cannot allocate memory";
1230 	struct cache_set *c;
1231 
1232 	bdevname(bdev, dc->backing_dev_name);
1233 	memcpy(&dc->sb, sb, sizeof(struct cache_sb));
1234 	dc->bdev = bdev;
1235 	dc->bdev->bd_holder = dc;
1236 
1237 	bio_init(&dc->sb_bio, dc->sb_bio.bi_inline_vecs, 1);
1238 	bio_first_bvec_all(&dc->sb_bio)->bv_page = sb_page;
1239 	get_page(sb_page);
1240 
1241 
1242 	if (cached_dev_init(dc, sb->block_size << 9))
1243 		goto err;
1244 
1245 	err = "error creating kobject";
1246 	if (kobject_add(&dc->disk.kobj, &part_to_dev(bdev->bd_part)->kobj,
1247 			"bcache"))
1248 		goto err;
1249 	if (bch_cache_accounting_add_kobjs(&dc->accounting, &dc->disk.kobj))
1250 		goto err;
1251 
1252 	pr_info("registered backing device %s", dc->backing_dev_name);
1253 
1254 	list_add(&dc->list, &uncached_devices);
1255 	list_for_each_entry(c, &bch_cache_sets, list)
1256 		bch_cached_dev_attach(dc, c, NULL);
1257 
1258 	if (BDEV_STATE(&dc->sb) == BDEV_STATE_NONE ||
1259 	    BDEV_STATE(&dc->sb) == BDEV_STATE_STALE)
1260 		bch_cached_dev_run(dc);
1261 
1262 	return;
1263 err:
1264 	pr_notice("error %s: %s", dc->backing_dev_name, err);
1265 	bcache_device_stop(&dc->disk);
1266 }
1267 
1268 /* Flash only volumes */
1269 
1270 void bch_flash_dev_release(struct kobject *kobj)
1271 {
1272 	struct bcache_device *d = container_of(kobj, struct bcache_device,
1273 					       kobj);
1274 	kfree(d);
1275 }
1276 
1277 static void flash_dev_free(struct closure *cl)
1278 {
1279 	struct bcache_device *d = container_of(cl, struct bcache_device, cl);
1280 	mutex_lock(&bch_register_lock);
1281 	bcache_device_free(d);
1282 	mutex_unlock(&bch_register_lock);
1283 	kobject_put(&d->kobj);
1284 }
1285 
1286 static void flash_dev_flush(struct closure *cl)
1287 {
1288 	struct bcache_device *d = container_of(cl, struct bcache_device, cl);
1289 
1290 	mutex_lock(&bch_register_lock);
1291 	bcache_device_unlink(d);
1292 	mutex_unlock(&bch_register_lock);
1293 	kobject_del(&d->kobj);
1294 	continue_at(cl, flash_dev_free, system_wq);
1295 }
1296 
1297 static int flash_dev_run(struct cache_set *c, struct uuid_entry *u)
1298 {
1299 	struct bcache_device *d = kzalloc(sizeof(struct bcache_device),
1300 					  GFP_KERNEL);
1301 	if (!d)
1302 		return -ENOMEM;
1303 
1304 	closure_init(&d->cl, NULL);
1305 	set_closure_fn(&d->cl, flash_dev_flush, system_wq);
1306 
1307 	kobject_init(&d->kobj, &bch_flash_dev_ktype);
1308 
1309 	if (bcache_device_init(d, block_bytes(c), u->sectors))
1310 		goto err;
1311 
1312 	bcache_device_attach(d, c, u - c->uuids);
1313 	bch_sectors_dirty_init(d);
1314 	bch_flash_dev_request_init(d);
1315 	add_disk(d->disk);
1316 
1317 	if (kobject_add(&d->kobj, &disk_to_dev(d->disk)->kobj, "bcache"))
1318 		goto err;
1319 
1320 	bcache_device_link(d, c, "volume");
1321 
1322 	return 0;
1323 err:
1324 	kobject_put(&d->kobj);
1325 	return -ENOMEM;
1326 }
1327 
1328 static int flash_devs_run(struct cache_set *c)
1329 {
1330 	int ret = 0;
1331 	struct uuid_entry *u;
1332 
1333 	for (u = c->uuids;
1334 	     u < c->uuids + c->nr_uuids && !ret;
1335 	     u++)
1336 		if (UUID_FLASH_ONLY(u))
1337 			ret = flash_dev_run(c, u);
1338 
1339 	return ret;
1340 }
1341 
1342 int bch_flash_dev_create(struct cache_set *c, uint64_t size)
1343 {
1344 	struct uuid_entry *u;
1345 
1346 	if (test_bit(CACHE_SET_STOPPING, &c->flags))
1347 		return -EINTR;
1348 
1349 	if (!test_bit(CACHE_SET_RUNNING, &c->flags))
1350 		return -EPERM;
1351 
1352 	u = uuid_find_empty(c);
1353 	if (!u) {
1354 		pr_err("Can't create volume, no room for UUID");
1355 		return -EINVAL;
1356 	}
1357 
1358 	get_random_bytes(u->uuid, 16);
1359 	memset(u->label, 0, 32);
1360 	u->first_reg = u->last_reg = cpu_to_le32(get_seconds());
1361 
1362 	SET_UUID_FLASH_ONLY(u, 1);
1363 	u->sectors = size >> 9;
1364 
1365 	bch_uuid_write(c);
1366 
1367 	return flash_dev_run(c, u);
1368 }
1369 
1370 bool bch_cached_dev_error(struct cached_dev *dc)
1371 {
1372 	struct cache_set *c;
1373 
1374 	if (!dc || test_bit(BCACHE_DEV_CLOSING, &dc->disk.flags))
1375 		return false;
1376 
1377 	dc->io_disable = true;
1378 	/* make others know io_disable is true earlier */
1379 	smp_mb();
1380 
1381 	pr_err("stop %s: too many IO errors on backing device %s\n",
1382 		dc->disk.disk->disk_name, dc->backing_dev_name);
1383 
1384 	/*
1385 	 * If the cached device is still attached to a cache set,
1386 	 * even dc->io_disable is true and no more I/O requests
1387 	 * accepted, cache device internal I/O (writeback scan or
1388 	 * garbage collection) may still prevent bcache device from
1389 	 * being stopped. So here CACHE_SET_IO_DISABLE should be
1390 	 * set to c->flags too, to make the internal I/O to cache
1391 	 * device rejected and stopped immediately.
1392 	 * If c is NULL, that means the bcache device is not attached
1393 	 * to any cache set, then no CACHE_SET_IO_DISABLE bit to set.
1394 	 */
1395 	c = dc->disk.c;
1396 	if (c && test_and_set_bit(CACHE_SET_IO_DISABLE, &c->flags))
1397 		pr_info("CACHE_SET_IO_DISABLE already set");
1398 
1399 	bcache_device_stop(&dc->disk);
1400 	return true;
1401 }
1402 
1403 /* Cache set */
1404 
1405 __printf(2, 3)
1406 bool bch_cache_set_error(struct cache_set *c, const char *fmt, ...)
1407 {
1408 	va_list args;
1409 
1410 	if (c->on_error != ON_ERROR_PANIC &&
1411 	    test_bit(CACHE_SET_STOPPING, &c->flags))
1412 		return false;
1413 
1414 	if (test_and_set_bit(CACHE_SET_IO_DISABLE, &c->flags))
1415 		pr_info("CACHE_SET_IO_DISABLE already set");
1416 
1417 	/* XXX: we can be called from atomic context
1418 	acquire_console_sem();
1419 	*/
1420 
1421 	printk(KERN_ERR "bcache: error on %pU: ", c->sb.set_uuid);
1422 
1423 	va_start(args, fmt);
1424 	vprintk(fmt, args);
1425 	va_end(args);
1426 
1427 	printk(", disabling caching\n");
1428 
1429 	if (c->on_error == ON_ERROR_PANIC)
1430 		panic("panic forced after error\n");
1431 
1432 	bch_cache_set_unregister(c);
1433 	return true;
1434 }
1435 
1436 void bch_cache_set_release(struct kobject *kobj)
1437 {
1438 	struct cache_set *c = container_of(kobj, struct cache_set, kobj);
1439 	kfree(c);
1440 	module_put(THIS_MODULE);
1441 }
1442 
1443 static void cache_set_free(struct closure *cl)
1444 {
1445 	struct cache_set *c = container_of(cl, struct cache_set, cl);
1446 	struct cache *ca;
1447 	unsigned i;
1448 
1449 	if (!IS_ERR_OR_NULL(c->debug))
1450 		debugfs_remove(c->debug);
1451 
1452 	bch_open_buckets_free(c);
1453 	bch_btree_cache_free(c);
1454 	bch_journal_free(c);
1455 
1456 	for_each_cache(ca, c, i)
1457 		if (ca) {
1458 			ca->set = NULL;
1459 			c->cache[ca->sb.nr_this_dev] = NULL;
1460 			kobject_put(&ca->kobj);
1461 		}
1462 
1463 	bch_bset_sort_state_free(&c->sort);
1464 	free_pages((unsigned long) c->uuids, ilog2(bucket_pages(c)));
1465 
1466 	if (c->moving_gc_wq)
1467 		destroy_workqueue(c->moving_gc_wq);
1468 	if (c->bio_split)
1469 		bioset_free(c->bio_split);
1470 	if (c->fill_iter)
1471 		mempool_destroy(c->fill_iter);
1472 	if (c->bio_meta)
1473 		mempool_destroy(c->bio_meta);
1474 	if (c->search)
1475 		mempool_destroy(c->search);
1476 	kfree(c->devices);
1477 
1478 	mutex_lock(&bch_register_lock);
1479 	list_del(&c->list);
1480 	mutex_unlock(&bch_register_lock);
1481 
1482 	pr_info("Cache set %pU unregistered", c->sb.set_uuid);
1483 	wake_up(&unregister_wait);
1484 
1485 	closure_debug_destroy(&c->cl);
1486 	kobject_put(&c->kobj);
1487 }
1488 
1489 static void cache_set_flush(struct closure *cl)
1490 {
1491 	struct cache_set *c = container_of(cl, struct cache_set, caching);
1492 	struct cache *ca;
1493 	struct btree *b;
1494 	unsigned i;
1495 
1496 	bch_cache_accounting_destroy(&c->accounting);
1497 
1498 	kobject_put(&c->internal);
1499 	kobject_del(&c->kobj);
1500 
1501 	if (c->gc_thread)
1502 		kthread_stop(c->gc_thread);
1503 
1504 	if (!IS_ERR_OR_NULL(c->root))
1505 		list_add(&c->root->list, &c->btree_cache);
1506 
1507 	/* Should skip this if we're unregistering because of an error */
1508 	list_for_each_entry(b, &c->btree_cache, list) {
1509 		mutex_lock(&b->write_lock);
1510 		if (btree_node_dirty(b))
1511 			__bch_btree_node_write(b, NULL);
1512 		mutex_unlock(&b->write_lock);
1513 	}
1514 
1515 	for_each_cache(ca, c, i)
1516 		if (ca->alloc_thread)
1517 			kthread_stop(ca->alloc_thread);
1518 
1519 	if (c->journal.cur) {
1520 		cancel_delayed_work_sync(&c->journal.work);
1521 		/* flush last journal entry if needed */
1522 		c->journal.work.work.func(&c->journal.work.work);
1523 	}
1524 
1525 	closure_return(cl);
1526 }
1527 
1528 /*
1529  * This function is only called when CACHE_SET_IO_DISABLE is set, which means
1530  * cache set is unregistering due to too many I/O errors. In this condition,
1531  * the bcache device might be stopped, it depends on stop_when_cache_set_failed
1532  * value and whether the broken cache has dirty data:
1533  *
1534  * dc->stop_when_cache_set_failed    dc->has_dirty   stop bcache device
1535  *  BCH_CACHED_STOP_AUTO               0               NO
1536  *  BCH_CACHED_STOP_AUTO               1               YES
1537  *  BCH_CACHED_DEV_STOP_ALWAYS         0               YES
1538  *  BCH_CACHED_DEV_STOP_ALWAYS         1               YES
1539  *
1540  * The expected behavior is, if stop_when_cache_set_failed is configured to
1541  * "auto" via sysfs interface, the bcache device will not be stopped if the
1542  * backing device is clean on the broken cache device.
1543  */
1544 static void conditional_stop_bcache_device(struct cache_set *c,
1545 					   struct bcache_device *d,
1546 					   struct cached_dev *dc)
1547 {
1548 	if (dc->stop_when_cache_set_failed == BCH_CACHED_DEV_STOP_ALWAYS) {
1549 		pr_warn("stop_when_cache_set_failed of %s is \"always\", stop it for failed cache set %pU.",
1550 			d->disk->disk_name, c->sb.set_uuid);
1551 		bcache_device_stop(d);
1552 	} else if (atomic_read(&dc->has_dirty)) {
1553 		/*
1554 		 * dc->stop_when_cache_set_failed == BCH_CACHED_STOP_AUTO
1555 		 * and dc->has_dirty == 1
1556 		 */
1557 		pr_warn("stop_when_cache_set_failed of %s is \"auto\" and cache is dirty, stop it to avoid potential data corruption.",
1558 			d->disk->disk_name);
1559 			/*
1560 			 * There might be a small time gap that cache set is
1561 			 * released but bcache device is not. Inside this time
1562 			 * gap, regular I/O requests will directly go into
1563 			 * backing device as no cache set attached to. This
1564 			 * behavior may also introduce potential inconsistence
1565 			 * data in writeback mode while cache is dirty.
1566 			 * Therefore before calling bcache_device_stop() due
1567 			 * to a broken cache device, dc->io_disable should be
1568 			 * explicitly set to true.
1569 			 */
1570 			dc->io_disable = true;
1571 			/* make others know io_disable is true earlier */
1572 			smp_mb();
1573 			bcache_device_stop(d);
1574 	} else {
1575 		/*
1576 		 * dc->stop_when_cache_set_failed == BCH_CACHED_STOP_AUTO
1577 		 * and dc->has_dirty == 0
1578 		 */
1579 		pr_warn("stop_when_cache_set_failed of %s is \"auto\" and cache is clean, keep it alive.",
1580 			d->disk->disk_name);
1581 	}
1582 }
1583 
1584 static void __cache_set_unregister(struct closure *cl)
1585 {
1586 	struct cache_set *c = container_of(cl, struct cache_set, caching);
1587 	struct cached_dev *dc;
1588 	struct bcache_device *d;
1589 	size_t i;
1590 
1591 	mutex_lock(&bch_register_lock);
1592 
1593 	for (i = 0; i < c->devices_max_used; i++) {
1594 		d = c->devices[i];
1595 		if (!d)
1596 			continue;
1597 
1598 		if (!UUID_FLASH_ONLY(&c->uuids[i]) &&
1599 		    test_bit(CACHE_SET_UNREGISTERING, &c->flags)) {
1600 			dc = container_of(d, struct cached_dev, disk);
1601 			bch_cached_dev_detach(dc);
1602 			if (test_bit(CACHE_SET_IO_DISABLE, &c->flags))
1603 				conditional_stop_bcache_device(c, d, dc);
1604 		} else {
1605 			bcache_device_stop(d);
1606 		}
1607 	}
1608 
1609 	mutex_unlock(&bch_register_lock);
1610 
1611 	continue_at(cl, cache_set_flush, system_wq);
1612 }
1613 
1614 void bch_cache_set_stop(struct cache_set *c)
1615 {
1616 	if (!test_and_set_bit(CACHE_SET_STOPPING, &c->flags))
1617 		closure_queue(&c->caching);
1618 }
1619 
1620 void bch_cache_set_unregister(struct cache_set *c)
1621 {
1622 	set_bit(CACHE_SET_UNREGISTERING, &c->flags);
1623 	bch_cache_set_stop(c);
1624 }
1625 
1626 #define alloc_bucket_pages(gfp, c)			\
1627 	((void *) __get_free_pages(__GFP_ZERO|gfp, ilog2(bucket_pages(c))))
1628 
1629 struct cache_set *bch_cache_set_alloc(struct cache_sb *sb)
1630 {
1631 	int iter_size;
1632 	struct cache_set *c = kzalloc(sizeof(struct cache_set), GFP_KERNEL);
1633 	if (!c)
1634 		return NULL;
1635 
1636 	__module_get(THIS_MODULE);
1637 	closure_init(&c->cl, NULL);
1638 	set_closure_fn(&c->cl, cache_set_free, system_wq);
1639 
1640 	closure_init(&c->caching, &c->cl);
1641 	set_closure_fn(&c->caching, __cache_set_unregister, system_wq);
1642 
1643 	/* Maybe create continue_at_noreturn() and use it here? */
1644 	closure_set_stopped(&c->cl);
1645 	closure_put(&c->cl);
1646 
1647 	kobject_init(&c->kobj, &bch_cache_set_ktype);
1648 	kobject_init(&c->internal, &bch_cache_set_internal_ktype);
1649 
1650 	bch_cache_accounting_init(&c->accounting, &c->cl);
1651 
1652 	memcpy(c->sb.set_uuid, sb->set_uuid, 16);
1653 	c->sb.block_size	= sb->block_size;
1654 	c->sb.bucket_size	= sb->bucket_size;
1655 	c->sb.nr_in_set		= sb->nr_in_set;
1656 	c->sb.last_mount	= sb->last_mount;
1657 	c->bucket_bits		= ilog2(sb->bucket_size);
1658 	c->block_bits		= ilog2(sb->block_size);
1659 	c->nr_uuids		= bucket_bytes(c) / sizeof(struct uuid_entry);
1660 	c->devices_max_used	= 0;
1661 	c->btree_pages		= bucket_pages(c);
1662 	if (c->btree_pages > BTREE_MAX_PAGES)
1663 		c->btree_pages = max_t(int, c->btree_pages / 4,
1664 				       BTREE_MAX_PAGES);
1665 
1666 	sema_init(&c->sb_write_mutex, 1);
1667 	mutex_init(&c->bucket_lock);
1668 	init_waitqueue_head(&c->btree_cache_wait);
1669 	init_waitqueue_head(&c->bucket_wait);
1670 	init_waitqueue_head(&c->gc_wait);
1671 	sema_init(&c->uuid_write_mutex, 1);
1672 
1673 	spin_lock_init(&c->btree_gc_time.lock);
1674 	spin_lock_init(&c->btree_split_time.lock);
1675 	spin_lock_init(&c->btree_read_time.lock);
1676 
1677 	bch_moving_init_cache_set(c);
1678 
1679 	INIT_LIST_HEAD(&c->list);
1680 	INIT_LIST_HEAD(&c->cached_devs);
1681 	INIT_LIST_HEAD(&c->btree_cache);
1682 	INIT_LIST_HEAD(&c->btree_cache_freeable);
1683 	INIT_LIST_HEAD(&c->btree_cache_freed);
1684 	INIT_LIST_HEAD(&c->data_buckets);
1685 
1686 	c->search = mempool_create_slab_pool(32, bch_search_cache);
1687 	if (!c->search)
1688 		goto err;
1689 
1690 	iter_size = (sb->bucket_size / sb->block_size + 1) *
1691 		sizeof(struct btree_iter_set);
1692 
1693 	if (!(c->devices = kzalloc(c->nr_uuids * sizeof(void *), GFP_KERNEL)) ||
1694 	    !(c->bio_meta = mempool_create_kmalloc_pool(2,
1695 				sizeof(struct bbio) + sizeof(struct bio_vec) *
1696 				bucket_pages(c))) ||
1697 	    !(c->fill_iter = mempool_create_kmalloc_pool(1, iter_size)) ||
1698 	    !(c->bio_split = bioset_create(4, offsetof(struct bbio, bio),
1699 					   BIOSET_NEED_BVECS |
1700 					   BIOSET_NEED_RESCUER)) ||
1701 	    !(c->uuids = alloc_bucket_pages(GFP_KERNEL, c)) ||
1702 	    !(c->moving_gc_wq = alloc_workqueue("bcache_gc",
1703 						WQ_MEM_RECLAIM, 0)) ||
1704 	    bch_journal_alloc(c) ||
1705 	    bch_btree_cache_alloc(c) ||
1706 	    bch_open_buckets_alloc(c) ||
1707 	    bch_bset_sort_state_init(&c->sort, ilog2(c->btree_pages)))
1708 		goto err;
1709 
1710 	c->congested_read_threshold_us	= 2000;
1711 	c->congested_write_threshold_us	= 20000;
1712 	c->error_limit	= DEFAULT_IO_ERROR_LIMIT;
1713 	WARN_ON(test_and_clear_bit(CACHE_SET_IO_DISABLE, &c->flags));
1714 
1715 	return c;
1716 err:
1717 	bch_cache_set_unregister(c);
1718 	return NULL;
1719 }
1720 
1721 static void run_cache_set(struct cache_set *c)
1722 {
1723 	const char *err = "cannot allocate memory";
1724 	struct cached_dev *dc, *t;
1725 	struct cache *ca;
1726 	struct closure cl;
1727 	unsigned i;
1728 
1729 	closure_init_stack(&cl);
1730 
1731 	for_each_cache(ca, c, i)
1732 		c->nbuckets += ca->sb.nbuckets;
1733 	set_gc_sectors(c);
1734 
1735 	if (CACHE_SYNC(&c->sb)) {
1736 		LIST_HEAD(journal);
1737 		struct bkey *k;
1738 		struct jset *j;
1739 
1740 		err = "cannot allocate memory for journal";
1741 		if (bch_journal_read(c, &journal))
1742 			goto err;
1743 
1744 		pr_debug("btree_journal_read() done");
1745 
1746 		err = "no journal entries found";
1747 		if (list_empty(&journal))
1748 			goto err;
1749 
1750 		j = &list_entry(journal.prev, struct journal_replay, list)->j;
1751 
1752 		err = "IO error reading priorities";
1753 		for_each_cache(ca, c, i)
1754 			prio_read(ca, j->prio_bucket[ca->sb.nr_this_dev]);
1755 
1756 		/*
1757 		 * If prio_read() fails it'll call cache_set_error and we'll
1758 		 * tear everything down right away, but if we perhaps checked
1759 		 * sooner we could avoid journal replay.
1760 		 */
1761 
1762 		k = &j->btree_root;
1763 
1764 		err = "bad btree root";
1765 		if (__bch_btree_ptr_invalid(c, k))
1766 			goto err;
1767 
1768 		err = "error reading btree root";
1769 		c->root = bch_btree_node_get(c, NULL, k, j->btree_level, true, NULL);
1770 		if (IS_ERR_OR_NULL(c->root))
1771 			goto err;
1772 
1773 		list_del_init(&c->root->list);
1774 		rw_unlock(true, c->root);
1775 
1776 		err = uuid_read(c, j, &cl);
1777 		if (err)
1778 			goto err;
1779 
1780 		err = "error in recovery";
1781 		if (bch_btree_check(c))
1782 			goto err;
1783 
1784 		bch_journal_mark(c, &journal);
1785 		bch_initial_gc_finish(c);
1786 		pr_debug("btree_check() done");
1787 
1788 		/*
1789 		 * bcache_journal_next() can't happen sooner, or
1790 		 * btree_gc_finish() will give spurious errors about last_gc >
1791 		 * gc_gen - this is a hack but oh well.
1792 		 */
1793 		bch_journal_next(&c->journal);
1794 
1795 		err = "error starting allocator thread";
1796 		for_each_cache(ca, c, i)
1797 			if (bch_cache_allocator_start(ca))
1798 				goto err;
1799 
1800 		/*
1801 		 * First place it's safe to allocate: btree_check() and
1802 		 * btree_gc_finish() have to run before we have buckets to
1803 		 * allocate, and bch_bucket_alloc_set() might cause a journal
1804 		 * entry to be written so bcache_journal_next() has to be called
1805 		 * first.
1806 		 *
1807 		 * If the uuids were in the old format we have to rewrite them
1808 		 * before the next journal entry is written:
1809 		 */
1810 		if (j->version < BCACHE_JSET_VERSION_UUID)
1811 			__uuid_write(c);
1812 
1813 		bch_journal_replay(c, &journal);
1814 	} else {
1815 		pr_notice("invalidating existing data");
1816 
1817 		for_each_cache(ca, c, i) {
1818 			unsigned j;
1819 
1820 			ca->sb.keys = clamp_t(int, ca->sb.nbuckets >> 7,
1821 					      2, SB_JOURNAL_BUCKETS);
1822 
1823 			for (j = 0; j < ca->sb.keys; j++)
1824 				ca->sb.d[j] = ca->sb.first_bucket + j;
1825 		}
1826 
1827 		bch_initial_gc_finish(c);
1828 
1829 		err = "error starting allocator thread";
1830 		for_each_cache(ca, c, i)
1831 			if (bch_cache_allocator_start(ca))
1832 				goto err;
1833 
1834 		mutex_lock(&c->bucket_lock);
1835 		for_each_cache(ca, c, i)
1836 			bch_prio_write(ca);
1837 		mutex_unlock(&c->bucket_lock);
1838 
1839 		err = "cannot allocate new UUID bucket";
1840 		if (__uuid_write(c))
1841 			goto err;
1842 
1843 		err = "cannot allocate new btree root";
1844 		c->root = __bch_btree_node_alloc(c, NULL, 0, true, NULL);
1845 		if (IS_ERR_OR_NULL(c->root))
1846 			goto err;
1847 
1848 		mutex_lock(&c->root->write_lock);
1849 		bkey_copy_key(&c->root->key, &MAX_KEY);
1850 		bch_btree_node_write(c->root, &cl);
1851 		mutex_unlock(&c->root->write_lock);
1852 
1853 		bch_btree_set_root(c->root);
1854 		rw_unlock(true, c->root);
1855 
1856 		/*
1857 		 * We don't want to write the first journal entry until
1858 		 * everything is set up - fortunately journal entries won't be
1859 		 * written until the SET_CACHE_SYNC() here:
1860 		 */
1861 		SET_CACHE_SYNC(&c->sb, true);
1862 
1863 		bch_journal_next(&c->journal);
1864 		bch_journal_meta(c, &cl);
1865 	}
1866 
1867 	err = "error starting gc thread";
1868 	if (bch_gc_thread_start(c))
1869 		goto err;
1870 
1871 	closure_sync(&cl);
1872 	c->sb.last_mount = get_seconds();
1873 	bcache_write_super(c);
1874 
1875 	list_for_each_entry_safe(dc, t, &uncached_devices, list)
1876 		bch_cached_dev_attach(dc, c, NULL);
1877 
1878 	flash_devs_run(c);
1879 
1880 	set_bit(CACHE_SET_RUNNING, &c->flags);
1881 	return;
1882 err:
1883 	closure_sync(&cl);
1884 	/* XXX: test this, it's broken */
1885 	bch_cache_set_error(c, "%s", err);
1886 }
1887 
1888 static bool can_attach_cache(struct cache *ca, struct cache_set *c)
1889 {
1890 	return ca->sb.block_size	== c->sb.block_size &&
1891 		ca->sb.bucket_size	== c->sb.bucket_size &&
1892 		ca->sb.nr_in_set	== c->sb.nr_in_set;
1893 }
1894 
1895 static const char *register_cache_set(struct cache *ca)
1896 {
1897 	char buf[12];
1898 	const char *err = "cannot allocate memory";
1899 	struct cache_set *c;
1900 
1901 	list_for_each_entry(c, &bch_cache_sets, list)
1902 		if (!memcmp(c->sb.set_uuid, ca->sb.set_uuid, 16)) {
1903 			if (c->cache[ca->sb.nr_this_dev])
1904 				return "duplicate cache set member";
1905 
1906 			if (!can_attach_cache(ca, c))
1907 				return "cache sb does not match set";
1908 
1909 			if (!CACHE_SYNC(&ca->sb))
1910 				SET_CACHE_SYNC(&c->sb, false);
1911 
1912 			goto found;
1913 		}
1914 
1915 	c = bch_cache_set_alloc(&ca->sb);
1916 	if (!c)
1917 		return err;
1918 
1919 	err = "error creating kobject";
1920 	if (kobject_add(&c->kobj, bcache_kobj, "%pU", c->sb.set_uuid) ||
1921 	    kobject_add(&c->internal, &c->kobj, "internal"))
1922 		goto err;
1923 
1924 	if (bch_cache_accounting_add_kobjs(&c->accounting, &c->kobj))
1925 		goto err;
1926 
1927 	bch_debug_init_cache_set(c);
1928 
1929 	list_add(&c->list, &bch_cache_sets);
1930 found:
1931 	sprintf(buf, "cache%i", ca->sb.nr_this_dev);
1932 	if (sysfs_create_link(&ca->kobj, &c->kobj, "set") ||
1933 	    sysfs_create_link(&c->kobj, &ca->kobj, buf))
1934 		goto err;
1935 
1936 	if (ca->sb.seq > c->sb.seq) {
1937 		c->sb.version		= ca->sb.version;
1938 		memcpy(c->sb.set_uuid, ca->sb.set_uuid, 16);
1939 		c->sb.flags             = ca->sb.flags;
1940 		c->sb.seq		= ca->sb.seq;
1941 		pr_debug("set version = %llu", c->sb.version);
1942 	}
1943 
1944 	kobject_get(&ca->kobj);
1945 	ca->set = c;
1946 	ca->set->cache[ca->sb.nr_this_dev] = ca;
1947 	c->cache_by_alloc[c->caches_loaded++] = ca;
1948 
1949 	if (c->caches_loaded == c->sb.nr_in_set)
1950 		run_cache_set(c);
1951 
1952 	return NULL;
1953 err:
1954 	bch_cache_set_unregister(c);
1955 	return err;
1956 }
1957 
1958 /* Cache device */
1959 
1960 void bch_cache_release(struct kobject *kobj)
1961 {
1962 	struct cache *ca = container_of(kobj, struct cache, kobj);
1963 	unsigned i;
1964 
1965 	if (ca->set) {
1966 		BUG_ON(ca->set->cache[ca->sb.nr_this_dev] != ca);
1967 		ca->set->cache[ca->sb.nr_this_dev] = NULL;
1968 	}
1969 
1970 	free_pages((unsigned long) ca->disk_buckets, ilog2(bucket_pages(ca)));
1971 	kfree(ca->prio_buckets);
1972 	vfree(ca->buckets);
1973 
1974 	free_heap(&ca->heap);
1975 	free_fifo(&ca->free_inc);
1976 
1977 	for (i = 0; i < RESERVE_NR; i++)
1978 		free_fifo(&ca->free[i]);
1979 
1980 	if (ca->sb_bio.bi_inline_vecs[0].bv_page)
1981 		put_page(bio_first_page_all(&ca->sb_bio));
1982 
1983 	if (!IS_ERR_OR_NULL(ca->bdev))
1984 		blkdev_put(ca->bdev, FMODE_READ|FMODE_WRITE|FMODE_EXCL);
1985 
1986 	kfree(ca);
1987 	module_put(THIS_MODULE);
1988 }
1989 
1990 static int cache_alloc(struct cache *ca)
1991 {
1992 	size_t free;
1993 	size_t btree_buckets;
1994 	struct bucket *b;
1995 
1996 	__module_get(THIS_MODULE);
1997 	kobject_init(&ca->kobj, &bch_cache_ktype);
1998 
1999 	bio_init(&ca->journal.bio, ca->journal.bio.bi_inline_vecs, 8);
2000 
2001 	/*
2002 	 * when ca->sb.njournal_buckets is not zero, journal exists,
2003 	 * and in bch_journal_replay(), tree node may split,
2004 	 * so bucket of RESERVE_BTREE type is needed,
2005 	 * the worst situation is all journal buckets are valid journal,
2006 	 * and all the keys need to replay,
2007 	 * so the number of  RESERVE_BTREE type buckets should be as much
2008 	 * as journal buckets
2009 	 */
2010 	btree_buckets = ca->sb.njournal_buckets ?: 8;
2011 	free = roundup_pow_of_two(ca->sb.nbuckets) >> 10;
2012 
2013 	if (!init_fifo(&ca->free[RESERVE_BTREE], btree_buckets, GFP_KERNEL) ||
2014 	    !init_fifo_exact(&ca->free[RESERVE_PRIO], prio_buckets(ca), GFP_KERNEL) ||
2015 	    !init_fifo(&ca->free[RESERVE_MOVINGGC], free, GFP_KERNEL) ||
2016 	    !init_fifo(&ca->free[RESERVE_NONE], free, GFP_KERNEL) ||
2017 	    !init_fifo(&ca->free_inc,	free << 2, GFP_KERNEL) ||
2018 	    !init_heap(&ca->heap,	free << 3, GFP_KERNEL) ||
2019 	    !(ca->buckets	= vzalloc(sizeof(struct bucket) *
2020 					  ca->sb.nbuckets)) ||
2021 	    !(ca->prio_buckets	= kzalloc(sizeof(uint64_t) * prio_buckets(ca) *
2022 					  2, GFP_KERNEL)) ||
2023 	    !(ca->disk_buckets	= alloc_bucket_pages(GFP_KERNEL, ca)))
2024 		return -ENOMEM;
2025 
2026 	ca->prio_last_buckets = ca->prio_buckets + prio_buckets(ca);
2027 
2028 	for_each_bucket(b, ca)
2029 		atomic_set(&b->pin, 0);
2030 
2031 	return 0;
2032 }
2033 
2034 static int register_cache(struct cache_sb *sb, struct page *sb_page,
2035 				struct block_device *bdev, struct cache *ca)
2036 {
2037 	const char *err = NULL; /* must be set for any error case */
2038 	int ret = 0;
2039 
2040 	bdevname(bdev, ca->cache_dev_name);
2041 	memcpy(&ca->sb, sb, sizeof(struct cache_sb));
2042 	ca->bdev = bdev;
2043 	ca->bdev->bd_holder = ca;
2044 
2045 	bio_init(&ca->sb_bio, ca->sb_bio.bi_inline_vecs, 1);
2046 	bio_first_bvec_all(&ca->sb_bio)->bv_page = sb_page;
2047 	get_page(sb_page);
2048 
2049 	if (blk_queue_discard(bdev_get_queue(bdev)))
2050 		ca->discard = CACHE_DISCARD(&ca->sb);
2051 
2052 	ret = cache_alloc(ca);
2053 	if (ret != 0) {
2054 		blkdev_put(bdev, FMODE_READ|FMODE_WRITE|FMODE_EXCL);
2055 		if (ret == -ENOMEM)
2056 			err = "cache_alloc(): -ENOMEM";
2057 		else
2058 			err = "cache_alloc(): unknown error";
2059 		goto err;
2060 	}
2061 
2062 	if (kobject_add(&ca->kobj, &part_to_dev(bdev->bd_part)->kobj, "bcache")) {
2063 		err = "error calling kobject_add";
2064 		ret = -ENOMEM;
2065 		goto out;
2066 	}
2067 
2068 	mutex_lock(&bch_register_lock);
2069 	err = register_cache_set(ca);
2070 	mutex_unlock(&bch_register_lock);
2071 
2072 	if (err) {
2073 		ret = -ENODEV;
2074 		goto out;
2075 	}
2076 
2077 	pr_info("registered cache device %s", ca->cache_dev_name);
2078 
2079 out:
2080 	kobject_put(&ca->kobj);
2081 
2082 err:
2083 	if (err)
2084 		pr_notice("error %s: %s", ca->cache_dev_name, err);
2085 
2086 	return ret;
2087 }
2088 
2089 /* Global interfaces/init */
2090 
2091 static ssize_t register_bcache(struct kobject *, struct kobj_attribute *,
2092 			       const char *, size_t);
2093 
2094 kobj_attribute_write(register,		register_bcache);
2095 kobj_attribute_write(register_quiet,	register_bcache);
2096 
2097 static bool bch_is_open_backing(struct block_device *bdev) {
2098 	struct cache_set *c, *tc;
2099 	struct cached_dev *dc, *t;
2100 
2101 	list_for_each_entry_safe(c, tc, &bch_cache_sets, list)
2102 		list_for_each_entry_safe(dc, t, &c->cached_devs, list)
2103 			if (dc->bdev == bdev)
2104 				return true;
2105 	list_for_each_entry_safe(dc, t, &uncached_devices, list)
2106 		if (dc->bdev == bdev)
2107 			return true;
2108 	return false;
2109 }
2110 
2111 static bool bch_is_open_cache(struct block_device *bdev) {
2112 	struct cache_set *c, *tc;
2113 	struct cache *ca;
2114 	unsigned i;
2115 
2116 	list_for_each_entry_safe(c, tc, &bch_cache_sets, list)
2117 		for_each_cache(ca, c, i)
2118 			if (ca->bdev == bdev)
2119 				return true;
2120 	return false;
2121 }
2122 
2123 static bool bch_is_open(struct block_device *bdev) {
2124 	return bch_is_open_cache(bdev) || bch_is_open_backing(bdev);
2125 }
2126 
2127 static ssize_t register_bcache(struct kobject *k, struct kobj_attribute *attr,
2128 			       const char *buffer, size_t size)
2129 {
2130 	ssize_t ret = size;
2131 	const char *err = "cannot allocate memory";
2132 	char *path = NULL;
2133 	struct cache_sb *sb = NULL;
2134 	struct block_device *bdev = NULL;
2135 	struct page *sb_page = NULL;
2136 
2137 	if (!try_module_get(THIS_MODULE))
2138 		return -EBUSY;
2139 
2140 	if (!(path = kstrndup(buffer, size, GFP_KERNEL)) ||
2141 	    !(sb = kmalloc(sizeof(struct cache_sb), GFP_KERNEL)))
2142 		goto err;
2143 
2144 	err = "failed to open device";
2145 	bdev = blkdev_get_by_path(strim(path),
2146 				  FMODE_READ|FMODE_WRITE|FMODE_EXCL,
2147 				  sb);
2148 	if (IS_ERR(bdev)) {
2149 		if (bdev == ERR_PTR(-EBUSY)) {
2150 			bdev = lookup_bdev(strim(path));
2151 			mutex_lock(&bch_register_lock);
2152 			if (!IS_ERR(bdev) && bch_is_open(bdev))
2153 				err = "device already registered";
2154 			else
2155 				err = "device busy";
2156 			mutex_unlock(&bch_register_lock);
2157 			if (!IS_ERR(bdev))
2158 				bdput(bdev);
2159 			if (attr == &ksysfs_register_quiet)
2160 				goto out;
2161 		}
2162 		goto err;
2163 	}
2164 
2165 	err = "failed to set blocksize";
2166 	if (set_blocksize(bdev, 4096))
2167 		goto err_close;
2168 
2169 	err = read_super(sb, bdev, &sb_page);
2170 	if (err)
2171 		goto err_close;
2172 
2173 	err = "failed to register device";
2174 	if (SB_IS_BDEV(sb)) {
2175 		struct cached_dev *dc = kzalloc(sizeof(*dc), GFP_KERNEL);
2176 		if (!dc)
2177 			goto err_close;
2178 
2179 		mutex_lock(&bch_register_lock);
2180 		register_bdev(sb, sb_page, bdev, dc);
2181 		mutex_unlock(&bch_register_lock);
2182 	} else {
2183 		struct cache *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
2184 		if (!ca)
2185 			goto err_close;
2186 
2187 		if (register_cache(sb, sb_page, bdev, ca) != 0)
2188 			goto err;
2189 	}
2190 out:
2191 	if (sb_page)
2192 		put_page(sb_page);
2193 	kfree(sb);
2194 	kfree(path);
2195 	module_put(THIS_MODULE);
2196 	return ret;
2197 
2198 err_close:
2199 	blkdev_put(bdev, FMODE_READ|FMODE_WRITE|FMODE_EXCL);
2200 err:
2201 	pr_info("error %s: %s", path, err);
2202 	ret = -EINVAL;
2203 	goto out;
2204 }
2205 
2206 static int bcache_reboot(struct notifier_block *n, unsigned long code, void *x)
2207 {
2208 	if (code == SYS_DOWN ||
2209 	    code == SYS_HALT ||
2210 	    code == SYS_POWER_OFF) {
2211 		DEFINE_WAIT(wait);
2212 		unsigned long start = jiffies;
2213 		bool stopped = false;
2214 
2215 		struct cache_set *c, *tc;
2216 		struct cached_dev *dc, *tdc;
2217 
2218 		mutex_lock(&bch_register_lock);
2219 
2220 		if (list_empty(&bch_cache_sets) &&
2221 		    list_empty(&uncached_devices))
2222 			goto out;
2223 
2224 		pr_info("Stopping all devices:");
2225 
2226 		list_for_each_entry_safe(c, tc, &bch_cache_sets, list)
2227 			bch_cache_set_stop(c);
2228 
2229 		list_for_each_entry_safe(dc, tdc, &uncached_devices, list)
2230 			bcache_device_stop(&dc->disk);
2231 
2232 		/* What's a condition variable? */
2233 		while (1) {
2234 			long timeout = start + 2 * HZ - jiffies;
2235 
2236 			stopped = list_empty(&bch_cache_sets) &&
2237 				list_empty(&uncached_devices);
2238 
2239 			if (timeout < 0 || stopped)
2240 				break;
2241 
2242 			prepare_to_wait(&unregister_wait, &wait,
2243 					TASK_UNINTERRUPTIBLE);
2244 
2245 			mutex_unlock(&bch_register_lock);
2246 			schedule_timeout(timeout);
2247 			mutex_lock(&bch_register_lock);
2248 		}
2249 
2250 		finish_wait(&unregister_wait, &wait);
2251 
2252 		if (stopped)
2253 			pr_info("All devices stopped");
2254 		else
2255 			pr_notice("Timeout waiting for devices to be closed");
2256 out:
2257 		mutex_unlock(&bch_register_lock);
2258 	}
2259 
2260 	return NOTIFY_DONE;
2261 }
2262 
2263 static struct notifier_block reboot = {
2264 	.notifier_call	= bcache_reboot,
2265 	.priority	= INT_MAX, /* before any real devices */
2266 };
2267 
2268 static void bcache_exit(void)
2269 {
2270 	bch_debug_exit();
2271 	bch_request_exit();
2272 	if (bcache_kobj)
2273 		kobject_put(bcache_kobj);
2274 	if (bcache_wq)
2275 		destroy_workqueue(bcache_wq);
2276 	if (bcache_major)
2277 		unregister_blkdev(bcache_major, "bcache");
2278 	unregister_reboot_notifier(&reboot);
2279 	mutex_destroy(&bch_register_lock);
2280 }
2281 
2282 static int __init bcache_init(void)
2283 {
2284 	static const struct attribute *files[] = {
2285 		&ksysfs_register.attr,
2286 		&ksysfs_register_quiet.attr,
2287 		NULL
2288 	};
2289 
2290 	mutex_init(&bch_register_lock);
2291 	init_waitqueue_head(&unregister_wait);
2292 	register_reboot_notifier(&reboot);
2293 
2294 	bcache_major = register_blkdev(0, "bcache");
2295 	if (bcache_major < 0) {
2296 		unregister_reboot_notifier(&reboot);
2297 		mutex_destroy(&bch_register_lock);
2298 		return bcache_major;
2299 	}
2300 
2301 	if (!(bcache_wq = alloc_workqueue("bcache", WQ_MEM_RECLAIM, 0)) ||
2302 	    !(bcache_kobj = kobject_create_and_add("bcache", fs_kobj)) ||
2303 	    bch_request_init() ||
2304 	    bch_debug_init(bcache_kobj) || closure_debug_init() ||
2305 	    sysfs_create_files(bcache_kobj, files))
2306 		goto err;
2307 
2308 	return 0;
2309 err:
2310 	bcache_exit();
2311 	return -ENOMEM;
2312 }
2313 
2314 module_exit(bcache_exit);
2315 module_init(bcache_init);
2316