xref: /openbmc/linux/drivers/md/bcache/super.c (revision 0f0709e6bfc3ce4e8e1c0e8573490c45f76cfeee)
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 	struct cached_dev *dc = container_of(d, struct cached_dev, disk);
658 
659 	if (dc->io_disable)
660 		return -EIO;
661 
662 	return d->ioctl(d, mode, cmd, arg);
663 }
664 
665 static const struct block_device_operations bcache_ops = {
666 	.open		= open_dev,
667 	.release	= release_dev,
668 	.ioctl		= ioctl_dev,
669 	.owner		= THIS_MODULE,
670 };
671 
672 void bcache_device_stop(struct bcache_device *d)
673 {
674 	if (!test_and_set_bit(BCACHE_DEV_CLOSING, &d->flags))
675 		closure_queue(&d->cl);
676 }
677 
678 static void bcache_device_unlink(struct bcache_device *d)
679 {
680 	lockdep_assert_held(&bch_register_lock);
681 
682 	if (d->c && !test_and_set_bit(BCACHE_DEV_UNLINK_DONE, &d->flags)) {
683 		unsigned i;
684 		struct cache *ca;
685 
686 		sysfs_remove_link(&d->c->kobj, d->name);
687 		sysfs_remove_link(&d->kobj, "cache");
688 
689 		for_each_cache(ca, d->c, i)
690 			bd_unlink_disk_holder(ca->bdev, d->disk);
691 	}
692 }
693 
694 static void bcache_device_link(struct bcache_device *d, struct cache_set *c,
695 			       const char *name)
696 {
697 	unsigned i;
698 	struct cache *ca;
699 
700 	for_each_cache(ca, d->c, i)
701 		bd_link_disk_holder(ca->bdev, d->disk);
702 
703 	snprintf(d->name, BCACHEDEVNAME_SIZE,
704 		 "%s%u", name, d->id);
705 
706 	WARN(sysfs_create_link(&d->kobj, &c->kobj, "cache") ||
707 	     sysfs_create_link(&c->kobj, &d->kobj, d->name),
708 	     "Couldn't create device <-> cache set symlinks");
709 
710 	clear_bit(BCACHE_DEV_UNLINK_DONE, &d->flags);
711 }
712 
713 static void bcache_device_detach(struct bcache_device *d)
714 {
715 	lockdep_assert_held(&bch_register_lock);
716 
717 	if (test_bit(BCACHE_DEV_DETACHING, &d->flags)) {
718 		struct uuid_entry *u = d->c->uuids + d->id;
719 
720 		SET_UUID_FLASH_ONLY(u, 0);
721 		memcpy(u->uuid, invalid_uuid, 16);
722 		u->invalidated = cpu_to_le32(get_seconds());
723 		bch_uuid_write(d->c);
724 	}
725 
726 	bcache_device_unlink(d);
727 
728 	d->c->devices[d->id] = NULL;
729 	closure_put(&d->c->caching);
730 	d->c = NULL;
731 }
732 
733 static void bcache_device_attach(struct bcache_device *d, struct cache_set *c,
734 				 unsigned id)
735 {
736 	d->id = id;
737 	d->c = c;
738 	c->devices[id] = d;
739 
740 	if (id >= c->devices_max_used)
741 		c->devices_max_used = id + 1;
742 
743 	closure_get(&c->caching);
744 }
745 
746 static inline int first_minor_to_idx(int first_minor)
747 {
748 	return (first_minor/BCACHE_MINORS);
749 }
750 
751 static inline int idx_to_first_minor(int idx)
752 {
753 	return (idx * BCACHE_MINORS);
754 }
755 
756 static void bcache_device_free(struct bcache_device *d)
757 {
758 	lockdep_assert_held(&bch_register_lock);
759 
760 	pr_info("%s stopped", d->disk->disk_name);
761 
762 	if (d->c)
763 		bcache_device_detach(d);
764 	if (d->disk && d->disk->flags & GENHD_FL_UP)
765 		del_gendisk(d->disk);
766 	if (d->disk && d->disk->queue)
767 		blk_cleanup_queue(d->disk->queue);
768 	if (d->disk) {
769 		ida_simple_remove(&bcache_device_idx,
770 				  first_minor_to_idx(d->disk->first_minor));
771 		put_disk(d->disk);
772 	}
773 
774 	if (d->bio_split)
775 		bioset_free(d->bio_split);
776 	kvfree(d->full_dirty_stripes);
777 	kvfree(d->stripe_sectors_dirty);
778 
779 	closure_debug_destroy(&d->cl);
780 }
781 
782 static int bcache_device_init(struct bcache_device *d, unsigned block_size,
783 			      sector_t sectors)
784 {
785 	struct request_queue *q;
786 	const size_t max_stripes = min_t(size_t, INT_MAX,
787 					 SIZE_MAX / sizeof(atomic_t));
788 	size_t n;
789 	int idx;
790 
791 	if (!d->stripe_size)
792 		d->stripe_size = 1 << 31;
793 
794 	d->nr_stripes = DIV_ROUND_UP_ULL(sectors, d->stripe_size);
795 
796 	if (!d->nr_stripes || d->nr_stripes > max_stripes) {
797 		pr_err("nr_stripes too large or invalid: %u (start sector beyond end of disk?)",
798 			(unsigned)d->nr_stripes);
799 		return -ENOMEM;
800 	}
801 
802 	n = d->nr_stripes * sizeof(atomic_t);
803 	d->stripe_sectors_dirty = kvzalloc(n, GFP_KERNEL);
804 	if (!d->stripe_sectors_dirty)
805 		return -ENOMEM;
806 
807 	n = BITS_TO_LONGS(d->nr_stripes) * sizeof(unsigned long);
808 	d->full_dirty_stripes = kvzalloc(n, GFP_KERNEL);
809 	if (!d->full_dirty_stripes)
810 		return -ENOMEM;
811 
812 	idx = ida_simple_get(&bcache_device_idx, 0,
813 				BCACHE_DEVICE_IDX_MAX, GFP_KERNEL);
814 	if (idx < 0)
815 		return idx;
816 
817 	if (!(d->bio_split = bioset_create(4, offsetof(struct bbio, bio),
818 					   BIOSET_NEED_BVECS |
819 					   BIOSET_NEED_RESCUER)) ||
820 	    !(d->disk = alloc_disk(BCACHE_MINORS))) {
821 		ida_simple_remove(&bcache_device_idx, idx);
822 		return -ENOMEM;
823 	}
824 
825 	set_capacity(d->disk, sectors);
826 	snprintf(d->disk->disk_name, DISK_NAME_LEN, "bcache%i", idx);
827 
828 	d->disk->major		= bcache_major;
829 	d->disk->first_minor	= idx_to_first_minor(idx);
830 	d->disk->fops		= &bcache_ops;
831 	d->disk->private_data	= d;
832 
833 	q = blk_alloc_queue(GFP_KERNEL);
834 	if (!q)
835 		return -ENOMEM;
836 
837 	blk_queue_make_request(q, NULL);
838 	d->disk->queue			= q;
839 	q->queuedata			= d;
840 	q->backing_dev_info->congested_data = d;
841 	q->limits.max_hw_sectors	= UINT_MAX;
842 	q->limits.max_sectors		= UINT_MAX;
843 	q->limits.max_segment_size	= UINT_MAX;
844 	q->limits.max_segments		= BIO_MAX_PAGES;
845 	blk_queue_max_discard_sectors(q, UINT_MAX);
846 	q->limits.discard_granularity	= 512;
847 	q->limits.io_min		= block_size;
848 	q->limits.logical_block_size	= block_size;
849 	q->limits.physical_block_size	= block_size;
850 	blk_queue_flag_set(QUEUE_FLAG_NONROT, d->disk->queue);
851 	blk_queue_flag_clear(QUEUE_FLAG_ADD_RANDOM, d->disk->queue);
852 	blk_queue_flag_set(QUEUE_FLAG_DISCARD, d->disk->queue);
853 
854 	blk_queue_write_cache(q, true, true);
855 
856 	return 0;
857 }
858 
859 /* Cached device */
860 
861 static void calc_cached_dev_sectors(struct cache_set *c)
862 {
863 	uint64_t sectors = 0;
864 	struct cached_dev *dc;
865 
866 	list_for_each_entry(dc, &c->cached_devs, list)
867 		sectors += bdev_sectors(dc->bdev);
868 
869 	c->cached_dev_sectors = sectors;
870 }
871 
872 #define BACKING_DEV_OFFLINE_TIMEOUT 5
873 static int cached_dev_status_update(void *arg)
874 {
875 	struct cached_dev *dc = arg;
876 	struct request_queue *q;
877 
878 	/*
879 	 * If this delayed worker is stopping outside, directly quit here.
880 	 * dc->io_disable might be set via sysfs interface, so check it
881 	 * here too.
882 	 */
883 	while (!kthread_should_stop() && !dc->io_disable) {
884 		q = bdev_get_queue(dc->bdev);
885 		if (blk_queue_dying(q))
886 			dc->offline_seconds++;
887 		else
888 			dc->offline_seconds = 0;
889 
890 		if (dc->offline_seconds >= BACKING_DEV_OFFLINE_TIMEOUT) {
891 			pr_err("%s: device offline for %d seconds",
892 			       dc->backing_dev_name,
893 			       BACKING_DEV_OFFLINE_TIMEOUT);
894 			pr_err("%s: disable I/O request due to backing "
895 			       "device offline", dc->disk.name);
896 			dc->io_disable = true;
897 			/* let others know earlier that io_disable is true */
898 			smp_mb();
899 			bcache_device_stop(&dc->disk);
900 			break;
901 		}
902 		schedule_timeout_interruptible(HZ);
903 	}
904 
905 	wait_for_kthread_stop();
906 	return 0;
907 }
908 
909 
910 void bch_cached_dev_run(struct cached_dev *dc)
911 {
912 	struct bcache_device *d = &dc->disk;
913 	char buf[SB_LABEL_SIZE + 1];
914 	char *env[] = {
915 		"DRIVER=bcache",
916 		kasprintf(GFP_KERNEL, "CACHED_UUID=%pU", dc->sb.uuid),
917 		NULL,
918 		NULL,
919 	};
920 
921 	memcpy(buf, dc->sb.label, SB_LABEL_SIZE);
922 	buf[SB_LABEL_SIZE] = '\0';
923 	env[2] = kasprintf(GFP_KERNEL, "CACHED_LABEL=%s", buf);
924 
925 	if (atomic_xchg(&dc->running, 1)) {
926 		kfree(env[1]);
927 		kfree(env[2]);
928 		return;
929 	}
930 
931 	if (!d->c &&
932 	    BDEV_STATE(&dc->sb) != BDEV_STATE_NONE) {
933 		struct closure cl;
934 		closure_init_stack(&cl);
935 
936 		SET_BDEV_STATE(&dc->sb, BDEV_STATE_STALE);
937 		bch_write_bdev_super(dc, &cl);
938 		closure_sync(&cl);
939 	}
940 
941 	add_disk(d->disk);
942 	bd_link_disk_holder(dc->bdev, dc->disk.disk);
943 	/* won't show up in the uevent file, use udevadm monitor -e instead
944 	 * only class / kset properties are persistent */
945 	kobject_uevent_env(&disk_to_dev(d->disk)->kobj, KOBJ_CHANGE, env);
946 	kfree(env[1]);
947 	kfree(env[2]);
948 
949 	if (sysfs_create_link(&d->kobj, &disk_to_dev(d->disk)->kobj, "dev") ||
950 	    sysfs_create_link(&disk_to_dev(d->disk)->kobj, &d->kobj, "bcache"))
951 		pr_debug("error creating sysfs link");
952 
953 	dc->status_update_thread = kthread_run(cached_dev_status_update,
954 					       dc, "bcache_status_update");
955 	if (IS_ERR(dc->status_update_thread)) {
956 		pr_warn("failed to create bcache_status_update kthread, "
957 			"continue to run without monitoring backing "
958 			"device status");
959 	}
960 }
961 
962 /*
963  * If BCACHE_DEV_RATE_DW_RUNNING is set, it means routine of the delayed
964  * work dc->writeback_rate_update is running. Wait until the routine
965  * quits (BCACHE_DEV_RATE_DW_RUNNING is clear), then continue to
966  * cancel it. If BCACHE_DEV_RATE_DW_RUNNING is not clear after time_out
967  * seconds, give up waiting here and continue to cancel it too.
968  */
969 static void cancel_writeback_rate_update_dwork(struct cached_dev *dc)
970 {
971 	int time_out = WRITEBACK_RATE_UPDATE_SECS_MAX * HZ;
972 
973 	do {
974 		if (!test_bit(BCACHE_DEV_RATE_DW_RUNNING,
975 			      &dc->disk.flags))
976 			break;
977 		time_out--;
978 		schedule_timeout_interruptible(1);
979 	} while (time_out > 0);
980 
981 	if (time_out == 0)
982 		pr_warn("give up waiting for dc->writeback_write_update to quit");
983 
984 	cancel_delayed_work_sync(&dc->writeback_rate_update);
985 }
986 
987 static void cached_dev_detach_finish(struct work_struct *w)
988 {
989 	struct cached_dev *dc = container_of(w, struct cached_dev, detach);
990 	struct closure cl;
991 	closure_init_stack(&cl);
992 
993 	BUG_ON(!test_bit(BCACHE_DEV_DETACHING, &dc->disk.flags));
994 	BUG_ON(refcount_read(&dc->count));
995 
996 	mutex_lock(&bch_register_lock);
997 
998 	if (test_and_clear_bit(BCACHE_DEV_WB_RUNNING, &dc->disk.flags))
999 		cancel_writeback_rate_update_dwork(dc);
1000 
1001 	if (!IS_ERR_OR_NULL(dc->writeback_thread)) {
1002 		kthread_stop(dc->writeback_thread);
1003 		dc->writeback_thread = NULL;
1004 	}
1005 
1006 	memset(&dc->sb.set_uuid, 0, 16);
1007 	SET_BDEV_STATE(&dc->sb, BDEV_STATE_NONE);
1008 
1009 	bch_write_bdev_super(dc, &cl);
1010 	closure_sync(&cl);
1011 
1012 	bcache_device_detach(&dc->disk);
1013 	list_move(&dc->list, &uncached_devices);
1014 
1015 	clear_bit(BCACHE_DEV_DETACHING, &dc->disk.flags);
1016 	clear_bit(BCACHE_DEV_UNLINK_DONE, &dc->disk.flags);
1017 
1018 	mutex_unlock(&bch_register_lock);
1019 
1020 	pr_info("Caching disabled for %s", dc->backing_dev_name);
1021 
1022 	/* Drop ref we took in cached_dev_detach() */
1023 	closure_put(&dc->disk.cl);
1024 }
1025 
1026 void bch_cached_dev_detach(struct cached_dev *dc)
1027 {
1028 	lockdep_assert_held(&bch_register_lock);
1029 
1030 	if (test_bit(BCACHE_DEV_CLOSING, &dc->disk.flags))
1031 		return;
1032 
1033 	if (test_and_set_bit(BCACHE_DEV_DETACHING, &dc->disk.flags))
1034 		return;
1035 
1036 	/*
1037 	 * Block the device from being closed and freed until we're finished
1038 	 * detaching
1039 	 */
1040 	closure_get(&dc->disk.cl);
1041 
1042 	bch_writeback_queue(dc);
1043 
1044 	cached_dev_put(dc);
1045 }
1046 
1047 int bch_cached_dev_attach(struct cached_dev *dc, struct cache_set *c,
1048 			  uint8_t *set_uuid)
1049 {
1050 	uint32_t rtime = cpu_to_le32(get_seconds());
1051 	struct uuid_entry *u;
1052 	struct cached_dev *exist_dc, *t;
1053 
1054 	if ((set_uuid && memcmp(set_uuid, c->sb.set_uuid, 16)) ||
1055 	    (!set_uuid && memcmp(dc->sb.set_uuid, c->sb.set_uuid, 16)))
1056 		return -ENOENT;
1057 
1058 	if (dc->disk.c) {
1059 		pr_err("Can't attach %s: already attached",
1060 		       dc->backing_dev_name);
1061 		return -EINVAL;
1062 	}
1063 
1064 	if (test_bit(CACHE_SET_STOPPING, &c->flags)) {
1065 		pr_err("Can't attach %s: shutting down",
1066 		       dc->backing_dev_name);
1067 		return -EINVAL;
1068 	}
1069 
1070 	if (dc->sb.block_size < c->sb.block_size) {
1071 		/* Will die */
1072 		pr_err("Couldn't attach %s: block size less than set's block size",
1073 		       dc->backing_dev_name);
1074 		return -EINVAL;
1075 	}
1076 
1077 	/* Check whether already attached */
1078 	list_for_each_entry_safe(exist_dc, t, &c->cached_devs, list) {
1079 		if (!memcmp(dc->sb.uuid, exist_dc->sb.uuid, 16)) {
1080 			pr_err("Tried to attach %s but duplicate UUID already attached",
1081 				dc->backing_dev_name);
1082 
1083 			return -EINVAL;
1084 		}
1085 	}
1086 
1087 	u = uuid_find(c, dc->sb.uuid);
1088 
1089 	if (u &&
1090 	    (BDEV_STATE(&dc->sb) == BDEV_STATE_STALE ||
1091 	     BDEV_STATE(&dc->sb) == BDEV_STATE_NONE)) {
1092 		memcpy(u->uuid, invalid_uuid, 16);
1093 		u->invalidated = cpu_to_le32(get_seconds());
1094 		u = NULL;
1095 	}
1096 
1097 	if (!u) {
1098 		if (BDEV_STATE(&dc->sb) == BDEV_STATE_DIRTY) {
1099 			pr_err("Couldn't find uuid for %s in set",
1100 			       dc->backing_dev_name);
1101 			return -ENOENT;
1102 		}
1103 
1104 		u = uuid_find_empty(c);
1105 		if (!u) {
1106 			pr_err("Not caching %s, no room for UUID",
1107 			       dc->backing_dev_name);
1108 			return -EINVAL;
1109 		}
1110 	}
1111 
1112 	/* Deadlocks since we're called via sysfs...
1113 	sysfs_remove_file(&dc->kobj, &sysfs_attach);
1114 	 */
1115 
1116 	if (bch_is_zero(u->uuid, 16)) {
1117 		struct closure cl;
1118 		closure_init_stack(&cl);
1119 
1120 		memcpy(u->uuid, dc->sb.uuid, 16);
1121 		memcpy(u->label, dc->sb.label, SB_LABEL_SIZE);
1122 		u->first_reg = u->last_reg = rtime;
1123 		bch_uuid_write(c);
1124 
1125 		memcpy(dc->sb.set_uuid, c->sb.set_uuid, 16);
1126 		SET_BDEV_STATE(&dc->sb, BDEV_STATE_CLEAN);
1127 
1128 		bch_write_bdev_super(dc, &cl);
1129 		closure_sync(&cl);
1130 	} else {
1131 		u->last_reg = rtime;
1132 		bch_uuid_write(c);
1133 	}
1134 
1135 	bcache_device_attach(&dc->disk, c, u - c->uuids);
1136 	list_move(&dc->list, &c->cached_devs);
1137 	calc_cached_dev_sectors(c);
1138 
1139 	smp_wmb();
1140 	/*
1141 	 * dc->c must be set before dc->count != 0 - paired with the mb in
1142 	 * cached_dev_get()
1143 	 */
1144 	refcount_set(&dc->count, 1);
1145 
1146 	/* Block writeback thread, but spawn it */
1147 	down_write(&dc->writeback_lock);
1148 	if (bch_cached_dev_writeback_start(dc)) {
1149 		up_write(&dc->writeback_lock);
1150 		return -ENOMEM;
1151 	}
1152 
1153 	if (BDEV_STATE(&dc->sb) == BDEV_STATE_DIRTY) {
1154 		bch_sectors_dirty_init(&dc->disk);
1155 		atomic_set(&dc->has_dirty, 1);
1156 		bch_writeback_queue(dc);
1157 	}
1158 
1159 	bch_cached_dev_run(dc);
1160 	bcache_device_link(&dc->disk, c, "bdev");
1161 
1162 	/* Allow the writeback thread to proceed */
1163 	up_write(&dc->writeback_lock);
1164 
1165 	pr_info("Caching %s as %s on set %pU",
1166 		dc->backing_dev_name,
1167 		dc->disk.disk->disk_name,
1168 		dc->disk.c->sb.set_uuid);
1169 	return 0;
1170 }
1171 
1172 void bch_cached_dev_release(struct kobject *kobj)
1173 {
1174 	struct cached_dev *dc = container_of(kobj, struct cached_dev,
1175 					     disk.kobj);
1176 	kfree(dc);
1177 	module_put(THIS_MODULE);
1178 }
1179 
1180 static void cached_dev_free(struct closure *cl)
1181 {
1182 	struct cached_dev *dc = container_of(cl, struct cached_dev, disk.cl);
1183 
1184 	mutex_lock(&bch_register_lock);
1185 
1186 	if (test_and_clear_bit(BCACHE_DEV_WB_RUNNING, &dc->disk.flags))
1187 		cancel_writeback_rate_update_dwork(dc);
1188 
1189 	if (!IS_ERR_OR_NULL(dc->writeback_thread))
1190 		kthread_stop(dc->writeback_thread);
1191 	if (dc->writeback_write_wq)
1192 		destroy_workqueue(dc->writeback_write_wq);
1193 	if (!IS_ERR_OR_NULL(dc->status_update_thread))
1194 		kthread_stop(dc->status_update_thread);
1195 
1196 	if (atomic_read(&dc->running))
1197 		bd_unlink_disk_holder(dc->bdev, dc->disk.disk);
1198 	bcache_device_free(&dc->disk);
1199 	list_del(&dc->list);
1200 
1201 	mutex_unlock(&bch_register_lock);
1202 
1203 	if (!IS_ERR_OR_NULL(dc->bdev))
1204 		blkdev_put(dc->bdev, FMODE_READ|FMODE_WRITE|FMODE_EXCL);
1205 
1206 	wake_up(&unregister_wait);
1207 
1208 	kobject_put(&dc->disk.kobj);
1209 }
1210 
1211 static void cached_dev_flush(struct closure *cl)
1212 {
1213 	struct cached_dev *dc = container_of(cl, struct cached_dev, disk.cl);
1214 	struct bcache_device *d = &dc->disk;
1215 
1216 	mutex_lock(&bch_register_lock);
1217 	bcache_device_unlink(d);
1218 	mutex_unlock(&bch_register_lock);
1219 
1220 	bch_cache_accounting_destroy(&dc->accounting);
1221 	kobject_del(&d->kobj);
1222 
1223 	continue_at(cl, cached_dev_free, system_wq);
1224 }
1225 
1226 static int cached_dev_init(struct cached_dev *dc, unsigned block_size)
1227 {
1228 	int ret;
1229 	struct io *io;
1230 	struct request_queue *q = bdev_get_queue(dc->bdev);
1231 
1232 	__module_get(THIS_MODULE);
1233 	INIT_LIST_HEAD(&dc->list);
1234 	closure_init(&dc->disk.cl, NULL);
1235 	set_closure_fn(&dc->disk.cl, cached_dev_flush, system_wq);
1236 	kobject_init(&dc->disk.kobj, &bch_cached_dev_ktype);
1237 	INIT_WORK(&dc->detach, cached_dev_detach_finish);
1238 	sema_init(&dc->sb_write_mutex, 1);
1239 	INIT_LIST_HEAD(&dc->io_lru);
1240 	spin_lock_init(&dc->io_lock);
1241 	bch_cache_accounting_init(&dc->accounting, &dc->disk.cl);
1242 
1243 	dc->sequential_cutoff		= 4 << 20;
1244 
1245 	for (io = dc->io; io < dc->io + RECENT_IO; io++) {
1246 		list_add(&io->lru, &dc->io_lru);
1247 		hlist_add_head(&io->hash, dc->io_hash + RECENT_IO);
1248 	}
1249 
1250 	dc->disk.stripe_size = q->limits.io_opt >> 9;
1251 
1252 	if (dc->disk.stripe_size)
1253 		dc->partial_stripes_expensive =
1254 			q->limits.raid_partial_stripes_expensive;
1255 
1256 	ret = bcache_device_init(&dc->disk, block_size,
1257 			 dc->bdev->bd_part->nr_sects - dc->sb.data_offset);
1258 	if (ret)
1259 		return ret;
1260 
1261 	dc->disk.disk->queue->backing_dev_info->ra_pages =
1262 		max(dc->disk.disk->queue->backing_dev_info->ra_pages,
1263 		    q->backing_dev_info->ra_pages);
1264 
1265 	atomic_set(&dc->io_errors, 0);
1266 	dc->io_disable = false;
1267 	dc->error_limit = DEFAULT_CACHED_DEV_ERROR_LIMIT;
1268 	/* default to auto */
1269 	dc->stop_when_cache_set_failed = BCH_CACHED_DEV_STOP_AUTO;
1270 
1271 	bch_cached_dev_request_init(dc);
1272 	bch_cached_dev_writeback_init(dc);
1273 	return 0;
1274 }
1275 
1276 /* Cached device - bcache superblock */
1277 
1278 static void register_bdev(struct cache_sb *sb, struct page *sb_page,
1279 				 struct block_device *bdev,
1280 				 struct cached_dev *dc)
1281 {
1282 	const char *err = "cannot allocate memory";
1283 	struct cache_set *c;
1284 
1285 	bdevname(bdev, dc->backing_dev_name);
1286 	memcpy(&dc->sb, sb, sizeof(struct cache_sb));
1287 	dc->bdev = bdev;
1288 	dc->bdev->bd_holder = dc;
1289 
1290 	bio_init(&dc->sb_bio, dc->sb_bio.bi_inline_vecs, 1);
1291 	bio_first_bvec_all(&dc->sb_bio)->bv_page = sb_page;
1292 	get_page(sb_page);
1293 
1294 
1295 	if (cached_dev_init(dc, sb->block_size << 9))
1296 		goto err;
1297 
1298 	err = "error creating kobject";
1299 	if (kobject_add(&dc->disk.kobj, &part_to_dev(bdev->bd_part)->kobj,
1300 			"bcache"))
1301 		goto err;
1302 	if (bch_cache_accounting_add_kobjs(&dc->accounting, &dc->disk.kobj))
1303 		goto err;
1304 
1305 	pr_info("registered backing device %s", dc->backing_dev_name);
1306 
1307 	list_add(&dc->list, &uncached_devices);
1308 	list_for_each_entry(c, &bch_cache_sets, list)
1309 		bch_cached_dev_attach(dc, c, NULL);
1310 
1311 	if (BDEV_STATE(&dc->sb) == BDEV_STATE_NONE ||
1312 	    BDEV_STATE(&dc->sb) == BDEV_STATE_STALE)
1313 		bch_cached_dev_run(dc);
1314 
1315 	return;
1316 err:
1317 	pr_notice("error %s: %s", dc->backing_dev_name, err);
1318 	bcache_device_stop(&dc->disk);
1319 }
1320 
1321 /* Flash only volumes */
1322 
1323 void bch_flash_dev_release(struct kobject *kobj)
1324 {
1325 	struct bcache_device *d = container_of(kobj, struct bcache_device,
1326 					       kobj);
1327 	kfree(d);
1328 }
1329 
1330 static void flash_dev_free(struct closure *cl)
1331 {
1332 	struct bcache_device *d = container_of(cl, struct bcache_device, cl);
1333 	mutex_lock(&bch_register_lock);
1334 	bcache_device_free(d);
1335 	mutex_unlock(&bch_register_lock);
1336 	kobject_put(&d->kobj);
1337 }
1338 
1339 static void flash_dev_flush(struct closure *cl)
1340 {
1341 	struct bcache_device *d = container_of(cl, struct bcache_device, cl);
1342 
1343 	mutex_lock(&bch_register_lock);
1344 	bcache_device_unlink(d);
1345 	mutex_unlock(&bch_register_lock);
1346 	kobject_del(&d->kobj);
1347 	continue_at(cl, flash_dev_free, system_wq);
1348 }
1349 
1350 static int flash_dev_run(struct cache_set *c, struct uuid_entry *u)
1351 {
1352 	struct bcache_device *d = kzalloc(sizeof(struct bcache_device),
1353 					  GFP_KERNEL);
1354 	if (!d)
1355 		return -ENOMEM;
1356 
1357 	closure_init(&d->cl, NULL);
1358 	set_closure_fn(&d->cl, flash_dev_flush, system_wq);
1359 
1360 	kobject_init(&d->kobj, &bch_flash_dev_ktype);
1361 
1362 	if (bcache_device_init(d, block_bytes(c), u->sectors))
1363 		goto err;
1364 
1365 	bcache_device_attach(d, c, u - c->uuids);
1366 	bch_sectors_dirty_init(d);
1367 	bch_flash_dev_request_init(d);
1368 	add_disk(d->disk);
1369 
1370 	if (kobject_add(&d->kobj, &disk_to_dev(d->disk)->kobj, "bcache"))
1371 		goto err;
1372 
1373 	bcache_device_link(d, c, "volume");
1374 
1375 	return 0;
1376 err:
1377 	kobject_put(&d->kobj);
1378 	return -ENOMEM;
1379 }
1380 
1381 static int flash_devs_run(struct cache_set *c)
1382 {
1383 	int ret = 0;
1384 	struct uuid_entry *u;
1385 
1386 	for (u = c->uuids;
1387 	     u < c->uuids + c->nr_uuids && !ret;
1388 	     u++)
1389 		if (UUID_FLASH_ONLY(u))
1390 			ret = flash_dev_run(c, u);
1391 
1392 	return ret;
1393 }
1394 
1395 int bch_flash_dev_create(struct cache_set *c, uint64_t size)
1396 {
1397 	struct uuid_entry *u;
1398 
1399 	if (test_bit(CACHE_SET_STOPPING, &c->flags))
1400 		return -EINTR;
1401 
1402 	if (!test_bit(CACHE_SET_RUNNING, &c->flags))
1403 		return -EPERM;
1404 
1405 	u = uuid_find_empty(c);
1406 	if (!u) {
1407 		pr_err("Can't create volume, no room for UUID");
1408 		return -EINVAL;
1409 	}
1410 
1411 	get_random_bytes(u->uuid, 16);
1412 	memset(u->label, 0, 32);
1413 	u->first_reg = u->last_reg = cpu_to_le32(get_seconds());
1414 
1415 	SET_UUID_FLASH_ONLY(u, 1);
1416 	u->sectors = size >> 9;
1417 
1418 	bch_uuid_write(c);
1419 
1420 	return flash_dev_run(c, u);
1421 }
1422 
1423 bool bch_cached_dev_error(struct cached_dev *dc)
1424 {
1425 	struct cache_set *c;
1426 
1427 	if (!dc || test_bit(BCACHE_DEV_CLOSING, &dc->disk.flags))
1428 		return false;
1429 
1430 	dc->io_disable = true;
1431 	/* make others know io_disable is true earlier */
1432 	smp_mb();
1433 
1434 	pr_err("stop %s: too many IO errors on backing device %s\n",
1435 		dc->disk.disk->disk_name, dc->backing_dev_name);
1436 
1437 	/*
1438 	 * If the cached device is still attached to a cache set,
1439 	 * even dc->io_disable is true and no more I/O requests
1440 	 * accepted, cache device internal I/O (writeback scan or
1441 	 * garbage collection) may still prevent bcache device from
1442 	 * being stopped. So here CACHE_SET_IO_DISABLE should be
1443 	 * set to c->flags too, to make the internal I/O to cache
1444 	 * device rejected and stopped immediately.
1445 	 * If c is NULL, that means the bcache device is not attached
1446 	 * to any cache set, then no CACHE_SET_IO_DISABLE bit to set.
1447 	 */
1448 	c = dc->disk.c;
1449 	if (c && test_and_set_bit(CACHE_SET_IO_DISABLE, &c->flags))
1450 		pr_info("CACHE_SET_IO_DISABLE already set");
1451 
1452 	bcache_device_stop(&dc->disk);
1453 	return true;
1454 }
1455 
1456 /* Cache set */
1457 
1458 __printf(2, 3)
1459 bool bch_cache_set_error(struct cache_set *c, const char *fmt, ...)
1460 {
1461 	va_list args;
1462 
1463 	if (c->on_error != ON_ERROR_PANIC &&
1464 	    test_bit(CACHE_SET_STOPPING, &c->flags))
1465 		return false;
1466 
1467 	if (test_and_set_bit(CACHE_SET_IO_DISABLE, &c->flags))
1468 		pr_info("CACHE_SET_IO_DISABLE already set");
1469 
1470 	/* XXX: we can be called from atomic context
1471 	acquire_console_sem();
1472 	*/
1473 
1474 	printk(KERN_ERR "bcache: error on %pU: ", c->sb.set_uuid);
1475 
1476 	va_start(args, fmt);
1477 	vprintk(fmt, args);
1478 	va_end(args);
1479 
1480 	printk(", disabling caching\n");
1481 
1482 	if (c->on_error == ON_ERROR_PANIC)
1483 		panic("panic forced after error\n");
1484 
1485 	bch_cache_set_unregister(c);
1486 	return true;
1487 }
1488 
1489 void bch_cache_set_release(struct kobject *kobj)
1490 {
1491 	struct cache_set *c = container_of(kobj, struct cache_set, kobj);
1492 	kfree(c);
1493 	module_put(THIS_MODULE);
1494 }
1495 
1496 static void cache_set_free(struct closure *cl)
1497 {
1498 	struct cache_set *c = container_of(cl, struct cache_set, cl);
1499 	struct cache *ca;
1500 	unsigned i;
1501 
1502 	if (!IS_ERR_OR_NULL(c->debug))
1503 		debugfs_remove(c->debug);
1504 
1505 	bch_open_buckets_free(c);
1506 	bch_btree_cache_free(c);
1507 	bch_journal_free(c);
1508 
1509 	for_each_cache(ca, c, i)
1510 		if (ca) {
1511 			ca->set = NULL;
1512 			c->cache[ca->sb.nr_this_dev] = NULL;
1513 			kobject_put(&ca->kobj);
1514 		}
1515 
1516 	bch_bset_sort_state_free(&c->sort);
1517 	free_pages((unsigned long) c->uuids, ilog2(bucket_pages(c)));
1518 
1519 	if (c->moving_gc_wq)
1520 		destroy_workqueue(c->moving_gc_wq);
1521 	if (c->bio_split)
1522 		bioset_free(c->bio_split);
1523 	if (c->fill_iter)
1524 		mempool_destroy(c->fill_iter);
1525 	if (c->bio_meta)
1526 		mempool_destroy(c->bio_meta);
1527 	if (c->search)
1528 		mempool_destroy(c->search);
1529 	kfree(c->devices);
1530 
1531 	mutex_lock(&bch_register_lock);
1532 	list_del(&c->list);
1533 	mutex_unlock(&bch_register_lock);
1534 
1535 	pr_info("Cache set %pU unregistered", c->sb.set_uuid);
1536 	wake_up(&unregister_wait);
1537 
1538 	closure_debug_destroy(&c->cl);
1539 	kobject_put(&c->kobj);
1540 }
1541 
1542 static void cache_set_flush(struct closure *cl)
1543 {
1544 	struct cache_set *c = container_of(cl, struct cache_set, caching);
1545 	struct cache *ca;
1546 	struct btree *b;
1547 	unsigned i;
1548 
1549 	bch_cache_accounting_destroy(&c->accounting);
1550 
1551 	kobject_put(&c->internal);
1552 	kobject_del(&c->kobj);
1553 
1554 	if (c->gc_thread)
1555 		kthread_stop(c->gc_thread);
1556 
1557 	if (!IS_ERR_OR_NULL(c->root))
1558 		list_add(&c->root->list, &c->btree_cache);
1559 
1560 	/* Should skip this if we're unregistering because of an error */
1561 	list_for_each_entry(b, &c->btree_cache, list) {
1562 		mutex_lock(&b->write_lock);
1563 		if (btree_node_dirty(b))
1564 			__bch_btree_node_write(b, NULL);
1565 		mutex_unlock(&b->write_lock);
1566 	}
1567 
1568 	for_each_cache(ca, c, i)
1569 		if (ca->alloc_thread)
1570 			kthread_stop(ca->alloc_thread);
1571 
1572 	if (c->journal.cur) {
1573 		cancel_delayed_work_sync(&c->journal.work);
1574 		/* flush last journal entry if needed */
1575 		c->journal.work.work.func(&c->journal.work.work);
1576 	}
1577 
1578 	closure_return(cl);
1579 }
1580 
1581 /*
1582  * This function is only called when CACHE_SET_IO_DISABLE is set, which means
1583  * cache set is unregistering due to too many I/O errors. In this condition,
1584  * the bcache device might be stopped, it depends on stop_when_cache_set_failed
1585  * value and whether the broken cache has dirty data:
1586  *
1587  * dc->stop_when_cache_set_failed    dc->has_dirty   stop bcache device
1588  *  BCH_CACHED_STOP_AUTO               0               NO
1589  *  BCH_CACHED_STOP_AUTO               1               YES
1590  *  BCH_CACHED_DEV_STOP_ALWAYS         0               YES
1591  *  BCH_CACHED_DEV_STOP_ALWAYS         1               YES
1592  *
1593  * The expected behavior is, if stop_when_cache_set_failed is configured to
1594  * "auto" via sysfs interface, the bcache device will not be stopped if the
1595  * backing device is clean on the broken cache device.
1596  */
1597 static void conditional_stop_bcache_device(struct cache_set *c,
1598 					   struct bcache_device *d,
1599 					   struct cached_dev *dc)
1600 {
1601 	if (dc->stop_when_cache_set_failed == BCH_CACHED_DEV_STOP_ALWAYS) {
1602 		pr_warn("stop_when_cache_set_failed of %s is \"always\", stop it for failed cache set %pU.",
1603 			d->disk->disk_name, c->sb.set_uuid);
1604 		bcache_device_stop(d);
1605 	} else if (atomic_read(&dc->has_dirty)) {
1606 		/*
1607 		 * dc->stop_when_cache_set_failed == BCH_CACHED_STOP_AUTO
1608 		 * and dc->has_dirty == 1
1609 		 */
1610 		pr_warn("stop_when_cache_set_failed of %s is \"auto\" and cache is dirty, stop it to avoid potential data corruption.",
1611 			d->disk->disk_name);
1612 			/*
1613 			 * There might be a small time gap that cache set is
1614 			 * released but bcache device is not. Inside this time
1615 			 * gap, regular I/O requests will directly go into
1616 			 * backing device as no cache set attached to. This
1617 			 * behavior may also introduce potential inconsistence
1618 			 * data in writeback mode while cache is dirty.
1619 			 * Therefore before calling bcache_device_stop() due
1620 			 * to a broken cache device, dc->io_disable should be
1621 			 * explicitly set to true.
1622 			 */
1623 			dc->io_disable = true;
1624 			/* make others know io_disable is true earlier */
1625 			smp_mb();
1626 			bcache_device_stop(d);
1627 	} else {
1628 		/*
1629 		 * dc->stop_when_cache_set_failed == BCH_CACHED_STOP_AUTO
1630 		 * and dc->has_dirty == 0
1631 		 */
1632 		pr_warn("stop_when_cache_set_failed of %s is \"auto\" and cache is clean, keep it alive.",
1633 			d->disk->disk_name);
1634 	}
1635 }
1636 
1637 static void __cache_set_unregister(struct closure *cl)
1638 {
1639 	struct cache_set *c = container_of(cl, struct cache_set, caching);
1640 	struct cached_dev *dc;
1641 	struct bcache_device *d;
1642 	size_t i;
1643 
1644 	mutex_lock(&bch_register_lock);
1645 
1646 	for (i = 0; i < c->devices_max_used; i++) {
1647 		d = c->devices[i];
1648 		if (!d)
1649 			continue;
1650 
1651 		if (!UUID_FLASH_ONLY(&c->uuids[i]) &&
1652 		    test_bit(CACHE_SET_UNREGISTERING, &c->flags)) {
1653 			dc = container_of(d, struct cached_dev, disk);
1654 			bch_cached_dev_detach(dc);
1655 			if (test_bit(CACHE_SET_IO_DISABLE, &c->flags))
1656 				conditional_stop_bcache_device(c, d, dc);
1657 		} else {
1658 			bcache_device_stop(d);
1659 		}
1660 	}
1661 
1662 	mutex_unlock(&bch_register_lock);
1663 
1664 	continue_at(cl, cache_set_flush, system_wq);
1665 }
1666 
1667 void bch_cache_set_stop(struct cache_set *c)
1668 {
1669 	if (!test_and_set_bit(CACHE_SET_STOPPING, &c->flags))
1670 		closure_queue(&c->caching);
1671 }
1672 
1673 void bch_cache_set_unregister(struct cache_set *c)
1674 {
1675 	set_bit(CACHE_SET_UNREGISTERING, &c->flags);
1676 	bch_cache_set_stop(c);
1677 }
1678 
1679 #define alloc_bucket_pages(gfp, c)			\
1680 	((void *) __get_free_pages(__GFP_ZERO|gfp, ilog2(bucket_pages(c))))
1681 
1682 struct cache_set *bch_cache_set_alloc(struct cache_sb *sb)
1683 {
1684 	int iter_size;
1685 	struct cache_set *c = kzalloc(sizeof(struct cache_set), GFP_KERNEL);
1686 	if (!c)
1687 		return NULL;
1688 
1689 	__module_get(THIS_MODULE);
1690 	closure_init(&c->cl, NULL);
1691 	set_closure_fn(&c->cl, cache_set_free, system_wq);
1692 
1693 	closure_init(&c->caching, &c->cl);
1694 	set_closure_fn(&c->caching, __cache_set_unregister, system_wq);
1695 
1696 	/* Maybe create continue_at_noreturn() and use it here? */
1697 	closure_set_stopped(&c->cl);
1698 	closure_put(&c->cl);
1699 
1700 	kobject_init(&c->kobj, &bch_cache_set_ktype);
1701 	kobject_init(&c->internal, &bch_cache_set_internal_ktype);
1702 
1703 	bch_cache_accounting_init(&c->accounting, &c->cl);
1704 
1705 	memcpy(c->sb.set_uuid, sb->set_uuid, 16);
1706 	c->sb.block_size	= sb->block_size;
1707 	c->sb.bucket_size	= sb->bucket_size;
1708 	c->sb.nr_in_set		= sb->nr_in_set;
1709 	c->sb.last_mount	= sb->last_mount;
1710 	c->bucket_bits		= ilog2(sb->bucket_size);
1711 	c->block_bits		= ilog2(sb->block_size);
1712 	c->nr_uuids		= bucket_bytes(c) / sizeof(struct uuid_entry);
1713 	c->devices_max_used	= 0;
1714 	c->btree_pages		= bucket_pages(c);
1715 	if (c->btree_pages > BTREE_MAX_PAGES)
1716 		c->btree_pages = max_t(int, c->btree_pages / 4,
1717 				       BTREE_MAX_PAGES);
1718 
1719 	sema_init(&c->sb_write_mutex, 1);
1720 	mutex_init(&c->bucket_lock);
1721 	init_waitqueue_head(&c->btree_cache_wait);
1722 	init_waitqueue_head(&c->bucket_wait);
1723 	init_waitqueue_head(&c->gc_wait);
1724 	sema_init(&c->uuid_write_mutex, 1);
1725 
1726 	spin_lock_init(&c->btree_gc_time.lock);
1727 	spin_lock_init(&c->btree_split_time.lock);
1728 	spin_lock_init(&c->btree_read_time.lock);
1729 
1730 	bch_moving_init_cache_set(c);
1731 
1732 	INIT_LIST_HEAD(&c->list);
1733 	INIT_LIST_HEAD(&c->cached_devs);
1734 	INIT_LIST_HEAD(&c->btree_cache);
1735 	INIT_LIST_HEAD(&c->btree_cache_freeable);
1736 	INIT_LIST_HEAD(&c->btree_cache_freed);
1737 	INIT_LIST_HEAD(&c->data_buckets);
1738 
1739 	c->search = mempool_create_slab_pool(32, bch_search_cache);
1740 	if (!c->search)
1741 		goto err;
1742 
1743 	iter_size = (sb->bucket_size / sb->block_size + 1) *
1744 		sizeof(struct btree_iter_set);
1745 
1746 	if (!(c->devices = kzalloc(c->nr_uuids * sizeof(void *), GFP_KERNEL)) ||
1747 	    !(c->bio_meta = mempool_create_kmalloc_pool(2,
1748 				sizeof(struct bbio) + sizeof(struct bio_vec) *
1749 				bucket_pages(c))) ||
1750 	    !(c->fill_iter = mempool_create_kmalloc_pool(1, iter_size)) ||
1751 	    !(c->bio_split = bioset_create(4, offsetof(struct bbio, bio),
1752 					   BIOSET_NEED_BVECS |
1753 					   BIOSET_NEED_RESCUER)) ||
1754 	    !(c->uuids = alloc_bucket_pages(GFP_KERNEL, c)) ||
1755 	    !(c->moving_gc_wq = alloc_workqueue("bcache_gc",
1756 						WQ_MEM_RECLAIM, 0)) ||
1757 	    bch_journal_alloc(c) ||
1758 	    bch_btree_cache_alloc(c) ||
1759 	    bch_open_buckets_alloc(c) ||
1760 	    bch_bset_sort_state_init(&c->sort, ilog2(c->btree_pages)))
1761 		goto err;
1762 
1763 	c->congested_read_threshold_us	= 2000;
1764 	c->congested_write_threshold_us	= 20000;
1765 	c->error_limit	= DEFAULT_IO_ERROR_LIMIT;
1766 	WARN_ON(test_and_clear_bit(CACHE_SET_IO_DISABLE, &c->flags));
1767 
1768 	return c;
1769 err:
1770 	bch_cache_set_unregister(c);
1771 	return NULL;
1772 }
1773 
1774 static void run_cache_set(struct cache_set *c)
1775 {
1776 	const char *err = "cannot allocate memory";
1777 	struct cached_dev *dc, *t;
1778 	struct cache *ca;
1779 	struct closure cl;
1780 	unsigned i;
1781 
1782 	closure_init_stack(&cl);
1783 
1784 	for_each_cache(ca, c, i)
1785 		c->nbuckets += ca->sb.nbuckets;
1786 	set_gc_sectors(c);
1787 
1788 	if (CACHE_SYNC(&c->sb)) {
1789 		LIST_HEAD(journal);
1790 		struct bkey *k;
1791 		struct jset *j;
1792 
1793 		err = "cannot allocate memory for journal";
1794 		if (bch_journal_read(c, &journal))
1795 			goto err;
1796 
1797 		pr_debug("btree_journal_read() done");
1798 
1799 		err = "no journal entries found";
1800 		if (list_empty(&journal))
1801 			goto err;
1802 
1803 		j = &list_entry(journal.prev, struct journal_replay, list)->j;
1804 
1805 		err = "IO error reading priorities";
1806 		for_each_cache(ca, c, i)
1807 			prio_read(ca, j->prio_bucket[ca->sb.nr_this_dev]);
1808 
1809 		/*
1810 		 * If prio_read() fails it'll call cache_set_error and we'll
1811 		 * tear everything down right away, but if we perhaps checked
1812 		 * sooner we could avoid journal replay.
1813 		 */
1814 
1815 		k = &j->btree_root;
1816 
1817 		err = "bad btree root";
1818 		if (__bch_btree_ptr_invalid(c, k))
1819 			goto err;
1820 
1821 		err = "error reading btree root";
1822 		c->root = bch_btree_node_get(c, NULL, k, j->btree_level, true, NULL);
1823 		if (IS_ERR_OR_NULL(c->root))
1824 			goto err;
1825 
1826 		list_del_init(&c->root->list);
1827 		rw_unlock(true, c->root);
1828 
1829 		err = uuid_read(c, j, &cl);
1830 		if (err)
1831 			goto err;
1832 
1833 		err = "error in recovery";
1834 		if (bch_btree_check(c))
1835 			goto err;
1836 
1837 		bch_journal_mark(c, &journal);
1838 		bch_initial_gc_finish(c);
1839 		pr_debug("btree_check() done");
1840 
1841 		/*
1842 		 * bcache_journal_next() can't happen sooner, or
1843 		 * btree_gc_finish() will give spurious errors about last_gc >
1844 		 * gc_gen - this is a hack but oh well.
1845 		 */
1846 		bch_journal_next(&c->journal);
1847 
1848 		err = "error starting allocator thread";
1849 		for_each_cache(ca, c, i)
1850 			if (bch_cache_allocator_start(ca))
1851 				goto err;
1852 
1853 		/*
1854 		 * First place it's safe to allocate: btree_check() and
1855 		 * btree_gc_finish() have to run before we have buckets to
1856 		 * allocate, and bch_bucket_alloc_set() might cause a journal
1857 		 * entry to be written so bcache_journal_next() has to be called
1858 		 * first.
1859 		 *
1860 		 * If the uuids were in the old format we have to rewrite them
1861 		 * before the next journal entry is written:
1862 		 */
1863 		if (j->version < BCACHE_JSET_VERSION_UUID)
1864 			__uuid_write(c);
1865 
1866 		bch_journal_replay(c, &journal);
1867 	} else {
1868 		pr_notice("invalidating existing data");
1869 
1870 		for_each_cache(ca, c, i) {
1871 			unsigned j;
1872 
1873 			ca->sb.keys = clamp_t(int, ca->sb.nbuckets >> 7,
1874 					      2, SB_JOURNAL_BUCKETS);
1875 
1876 			for (j = 0; j < ca->sb.keys; j++)
1877 				ca->sb.d[j] = ca->sb.first_bucket + j;
1878 		}
1879 
1880 		bch_initial_gc_finish(c);
1881 
1882 		err = "error starting allocator thread";
1883 		for_each_cache(ca, c, i)
1884 			if (bch_cache_allocator_start(ca))
1885 				goto err;
1886 
1887 		mutex_lock(&c->bucket_lock);
1888 		for_each_cache(ca, c, i)
1889 			bch_prio_write(ca);
1890 		mutex_unlock(&c->bucket_lock);
1891 
1892 		err = "cannot allocate new UUID bucket";
1893 		if (__uuid_write(c))
1894 			goto err;
1895 
1896 		err = "cannot allocate new btree root";
1897 		c->root = __bch_btree_node_alloc(c, NULL, 0, true, NULL);
1898 		if (IS_ERR_OR_NULL(c->root))
1899 			goto err;
1900 
1901 		mutex_lock(&c->root->write_lock);
1902 		bkey_copy_key(&c->root->key, &MAX_KEY);
1903 		bch_btree_node_write(c->root, &cl);
1904 		mutex_unlock(&c->root->write_lock);
1905 
1906 		bch_btree_set_root(c->root);
1907 		rw_unlock(true, c->root);
1908 
1909 		/*
1910 		 * We don't want to write the first journal entry until
1911 		 * everything is set up - fortunately journal entries won't be
1912 		 * written until the SET_CACHE_SYNC() here:
1913 		 */
1914 		SET_CACHE_SYNC(&c->sb, true);
1915 
1916 		bch_journal_next(&c->journal);
1917 		bch_journal_meta(c, &cl);
1918 	}
1919 
1920 	err = "error starting gc thread";
1921 	if (bch_gc_thread_start(c))
1922 		goto err;
1923 
1924 	closure_sync(&cl);
1925 	c->sb.last_mount = get_seconds();
1926 	bcache_write_super(c);
1927 
1928 	list_for_each_entry_safe(dc, t, &uncached_devices, list)
1929 		bch_cached_dev_attach(dc, c, NULL);
1930 
1931 	flash_devs_run(c);
1932 
1933 	set_bit(CACHE_SET_RUNNING, &c->flags);
1934 	return;
1935 err:
1936 	closure_sync(&cl);
1937 	/* XXX: test this, it's broken */
1938 	bch_cache_set_error(c, "%s", err);
1939 }
1940 
1941 static bool can_attach_cache(struct cache *ca, struct cache_set *c)
1942 {
1943 	return ca->sb.block_size	== c->sb.block_size &&
1944 		ca->sb.bucket_size	== c->sb.bucket_size &&
1945 		ca->sb.nr_in_set	== c->sb.nr_in_set;
1946 }
1947 
1948 static const char *register_cache_set(struct cache *ca)
1949 {
1950 	char buf[12];
1951 	const char *err = "cannot allocate memory";
1952 	struct cache_set *c;
1953 
1954 	list_for_each_entry(c, &bch_cache_sets, list)
1955 		if (!memcmp(c->sb.set_uuid, ca->sb.set_uuid, 16)) {
1956 			if (c->cache[ca->sb.nr_this_dev])
1957 				return "duplicate cache set member";
1958 
1959 			if (!can_attach_cache(ca, c))
1960 				return "cache sb does not match set";
1961 
1962 			if (!CACHE_SYNC(&ca->sb))
1963 				SET_CACHE_SYNC(&c->sb, false);
1964 
1965 			goto found;
1966 		}
1967 
1968 	c = bch_cache_set_alloc(&ca->sb);
1969 	if (!c)
1970 		return err;
1971 
1972 	err = "error creating kobject";
1973 	if (kobject_add(&c->kobj, bcache_kobj, "%pU", c->sb.set_uuid) ||
1974 	    kobject_add(&c->internal, &c->kobj, "internal"))
1975 		goto err;
1976 
1977 	if (bch_cache_accounting_add_kobjs(&c->accounting, &c->kobj))
1978 		goto err;
1979 
1980 	bch_debug_init_cache_set(c);
1981 
1982 	list_add(&c->list, &bch_cache_sets);
1983 found:
1984 	sprintf(buf, "cache%i", ca->sb.nr_this_dev);
1985 	if (sysfs_create_link(&ca->kobj, &c->kobj, "set") ||
1986 	    sysfs_create_link(&c->kobj, &ca->kobj, buf))
1987 		goto err;
1988 
1989 	if (ca->sb.seq > c->sb.seq) {
1990 		c->sb.version		= ca->sb.version;
1991 		memcpy(c->sb.set_uuid, ca->sb.set_uuid, 16);
1992 		c->sb.flags             = ca->sb.flags;
1993 		c->sb.seq		= ca->sb.seq;
1994 		pr_debug("set version = %llu", c->sb.version);
1995 	}
1996 
1997 	kobject_get(&ca->kobj);
1998 	ca->set = c;
1999 	ca->set->cache[ca->sb.nr_this_dev] = ca;
2000 	c->cache_by_alloc[c->caches_loaded++] = ca;
2001 
2002 	if (c->caches_loaded == c->sb.nr_in_set)
2003 		run_cache_set(c);
2004 
2005 	return NULL;
2006 err:
2007 	bch_cache_set_unregister(c);
2008 	return err;
2009 }
2010 
2011 /* Cache device */
2012 
2013 void bch_cache_release(struct kobject *kobj)
2014 {
2015 	struct cache *ca = container_of(kobj, struct cache, kobj);
2016 	unsigned i;
2017 
2018 	if (ca->set) {
2019 		BUG_ON(ca->set->cache[ca->sb.nr_this_dev] != ca);
2020 		ca->set->cache[ca->sb.nr_this_dev] = NULL;
2021 	}
2022 
2023 	free_pages((unsigned long) ca->disk_buckets, ilog2(bucket_pages(ca)));
2024 	kfree(ca->prio_buckets);
2025 	vfree(ca->buckets);
2026 
2027 	free_heap(&ca->heap);
2028 	free_fifo(&ca->free_inc);
2029 
2030 	for (i = 0; i < RESERVE_NR; i++)
2031 		free_fifo(&ca->free[i]);
2032 
2033 	if (ca->sb_bio.bi_inline_vecs[0].bv_page)
2034 		put_page(bio_first_page_all(&ca->sb_bio));
2035 
2036 	if (!IS_ERR_OR_NULL(ca->bdev))
2037 		blkdev_put(ca->bdev, FMODE_READ|FMODE_WRITE|FMODE_EXCL);
2038 
2039 	kfree(ca);
2040 	module_put(THIS_MODULE);
2041 }
2042 
2043 static int cache_alloc(struct cache *ca)
2044 {
2045 	size_t free;
2046 	size_t btree_buckets;
2047 	struct bucket *b;
2048 
2049 	__module_get(THIS_MODULE);
2050 	kobject_init(&ca->kobj, &bch_cache_ktype);
2051 
2052 	bio_init(&ca->journal.bio, ca->journal.bio.bi_inline_vecs, 8);
2053 
2054 	/*
2055 	 * when ca->sb.njournal_buckets is not zero, journal exists,
2056 	 * and in bch_journal_replay(), tree node may split,
2057 	 * so bucket of RESERVE_BTREE type is needed,
2058 	 * the worst situation is all journal buckets are valid journal,
2059 	 * and all the keys need to replay,
2060 	 * so the number of  RESERVE_BTREE type buckets should be as much
2061 	 * as journal buckets
2062 	 */
2063 	btree_buckets = ca->sb.njournal_buckets ?: 8;
2064 	free = roundup_pow_of_two(ca->sb.nbuckets) >> 10;
2065 
2066 	if (!init_fifo(&ca->free[RESERVE_BTREE], btree_buckets, GFP_KERNEL) ||
2067 	    !init_fifo_exact(&ca->free[RESERVE_PRIO], prio_buckets(ca), GFP_KERNEL) ||
2068 	    !init_fifo(&ca->free[RESERVE_MOVINGGC], free, GFP_KERNEL) ||
2069 	    !init_fifo(&ca->free[RESERVE_NONE], free, GFP_KERNEL) ||
2070 	    !init_fifo(&ca->free_inc,	free << 2, GFP_KERNEL) ||
2071 	    !init_heap(&ca->heap,	free << 3, GFP_KERNEL) ||
2072 	    !(ca->buckets	= vzalloc(sizeof(struct bucket) *
2073 					  ca->sb.nbuckets)) ||
2074 	    !(ca->prio_buckets	= kzalloc(sizeof(uint64_t) * prio_buckets(ca) *
2075 					  2, GFP_KERNEL)) ||
2076 	    !(ca->disk_buckets	= alloc_bucket_pages(GFP_KERNEL, ca)))
2077 		return -ENOMEM;
2078 
2079 	ca->prio_last_buckets = ca->prio_buckets + prio_buckets(ca);
2080 
2081 	for_each_bucket(b, ca)
2082 		atomic_set(&b->pin, 0);
2083 
2084 	return 0;
2085 }
2086 
2087 static int register_cache(struct cache_sb *sb, struct page *sb_page,
2088 				struct block_device *bdev, struct cache *ca)
2089 {
2090 	const char *err = NULL; /* must be set for any error case */
2091 	int ret = 0;
2092 
2093 	bdevname(bdev, ca->cache_dev_name);
2094 	memcpy(&ca->sb, sb, sizeof(struct cache_sb));
2095 	ca->bdev = bdev;
2096 	ca->bdev->bd_holder = ca;
2097 
2098 	bio_init(&ca->sb_bio, ca->sb_bio.bi_inline_vecs, 1);
2099 	bio_first_bvec_all(&ca->sb_bio)->bv_page = sb_page;
2100 	get_page(sb_page);
2101 
2102 	if (blk_queue_discard(bdev_get_queue(bdev)))
2103 		ca->discard = CACHE_DISCARD(&ca->sb);
2104 
2105 	ret = cache_alloc(ca);
2106 	if (ret != 0) {
2107 		blkdev_put(bdev, FMODE_READ|FMODE_WRITE|FMODE_EXCL);
2108 		if (ret == -ENOMEM)
2109 			err = "cache_alloc(): -ENOMEM";
2110 		else
2111 			err = "cache_alloc(): unknown error";
2112 		goto err;
2113 	}
2114 
2115 	if (kobject_add(&ca->kobj, &part_to_dev(bdev->bd_part)->kobj, "bcache")) {
2116 		err = "error calling kobject_add";
2117 		ret = -ENOMEM;
2118 		goto out;
2119 	}
2120 
2121 	mutex_lock(&bch_register_lock);
2122 	err = register_cache_set(ca);
2123 	mutex_unlock(&bch_register_lock);
2124 
2125 	if (err) {
2126 		ret = -ENODEV;
2127 		goto out;
2128 	}
2129 
2130 	pr_info("registered cache device %s", ca->cache_dev_name);
2131 
2132 out:
2133 	kobject_put(&ca->kobj);
2134 
2135 err:
2136 	if (err)
2137 		pr_notice("error %s: %s", ca->cache_dev_name, err);
2138 
2139 	return ret;
2140 }
2141 
2142 /* Global interfaces/init */
2143 
2144 static ssize_t register_bcache(struct kobject *, struct kobj_attribute *,
2145 			       const char *, size_t);
2146 
2147 kobj_attribute_write(register,		register_bcache);
2148 kobj_attribute_write(register_quiet,	register_bcache);
2149 
2150 static bool bch_is_open_backing(struct block_device *bdev) {
2151 	struct cache_set *c, *tc;
2152 	struct cached_dev *dc, *t;
2153 
2154 	list_for_each_entry_safe(c, tc, &bch_cache_sets, list)
2155 		list_for_each_entry_safe(dc, t, &c->cached_devs, list)
2156 			if (dc->bdev == bdev)
2157 				return true;
2158 	list_for_each_entry_safe(dc, t, &uncached_devices, list)
2159 		if (dc->bdev == bdev)
2160 			return true;
2161 	return false;
2162 }
2163 
2164 static bool bch_is_open_cache(struct block_device *bdev) {
2165 	struct cache_set *c, *tc;
2166 	struct cache *ca;
2167 	unsigned i;
2168 
2169 	list_for_each_entry_safe(c, tc, &bch_cache_sets, list)
2170 		for_each_cache(ca, c, i)
2171 			if (ca->bdev == bdev)
2172 				return true;
2173 	return false;
2174 }
2175 
2176 static bool bch_is_open(struct block_device *bdev) {
2177 	return bch_is_open_cache(bdev) || bch_is_open_backing(bdev);
2178 }
2179 
2180 static ssize_t register_bcache(struct kobject *k, struct kobj_attribute *attr,
2181 			       const char *buffer, size_t size)
2182 {
2183 	ssize_t ret = size;
2184 	const char *err = "cannot allocate memory";
2185 	char *path = NULL;
2186 	struct cache_sb *sb = NULL;
2187 	struct block_device *bdev = NULL;
2188 	struct page *sb_page = NULL;
2189 
2190 	if (!try_module_get(THIS_MODULE))
2191 		return -EBUSY;
2192 
2193 	if (!(path = kstrndup(buffer, size, GFP_KERNEL)) ||
2194 	    !(sb = kmalloc(sizeof(struct cache_sb), GFP_KERNEL)))
2195 		goto err;
2196 
2197 	err = "failed to open device";
2198 	bdev = blkdev_get_by_path(strim(path),
2199 				  FMODE_READ|FMODE_WRITE|FMODE_EXCL,
2200 				  sb);
2201 	if (IS_ERR(bdev)) {
2202 		if (bdev == ERR_PTR(-EBUSY)) {
2203 			bdev = lookup_bdev(strim(path));
2204 			mutex_lock(&bch_register_lock);
2205 			if (!IS_ERR(bdev) && bch_is_open(bdev))
2206 				err = "device already registered";
2207 			else
2208 				err = "device busy";
2209 			mutex_unlock(&bch_register_lock);
2210 			if (!IS_ERR(bdev))
2211 				bdput(bdev);
2212 			if (attr == &ksysfs_register_quiet)
2213 				goto out;
2214 		}
2215 		goto err;
2216 	}
2217 
2218 	err = "failed to set blocksize";
2219 	if (set_blocksize(bdev, 4096))
2220 		goto err_close;
2221 
2222 	err = read_super(sb, bdev, &sb_page);
2223 	if (err)
2224 		goto err_close;
2225 
2226 	err = "failed to register device";
2227 	if (SB_IS_BDEV(sb)) {
2228 		struct cached_dev *dc = kzalloc(sizeof(*dc), GFP_KERNEL);
2229 		if (!dc)
2230 			goto err_close;
2231 
2232 		mutex_lock(&bch_register_lock);
2233 		register_bdev(sb, sb_page, bdev, dc);
2234 		mutex_unlock(&bch_register_lock);
2235 	} else {
2236 		struct cache *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
2237 		if (!ca)
2238 			goto err_close;
2239 
2240 		if (register_cache(sb, sb_page, bdev, ca) != 0)
2241 			goto err;
2242 	}
2243 out:
2244 	if (sb_page)
2245 		put_page(sb_page);
2246 	kfree(sb);
2247 	kfree(path);
2248 	module_put(THIS_MODULE);
2249 	return ret;
2250 
2251 err_close:
2252 	blkdev_put(bdev, FMODE_READ|FMODE_WRITE|FMODE_EXCL);
2253 err:
2254 	pr_info("error %s: %s", path, err);
2255 	ret = -EINVAL;
2256 	goto out;
2257 }
2258 
2259 static int bcache_reboot(struct notifier_block *n, unsigned long code, void *x)
2260 {
2261 	if (code == SYS_DOWN ||
2262 	    code == SYS_HALT ||
2263 	    code == SYS_POWER_OFF) {
2264 		DEFINE_WAIT(wait);
2265 		unsigned long start = jiffies;
2266 		bool stopped = false;
2267 
2268 		struct cache_set *c, *tc;
2269 		struct cached_dev *dc, *tdc;
2270 
2271 		mutex_lock(&bch_register_lock);
2272 
2273 		if (list_empty(&bch_cache_sets) &&
2274 		    list_empty(&uncached_devices))
2275 			goto out;
2276 
2277 		pr_info("Stopping all devices:");
2278 
2279 		list_for_each_entry_safe(c, tc, &bch_cache_sets, list)
2280 			bch_cache_set_stop(c);
2281 
2282 		list_for_each_entry_safe(dc, tdc, &uncached_devices, list)
2283 			bcache_device_stop(&dc->disk);
2284 
2285 		/* What's a condition variable? */
2286 		while (1) {
2287 			long timeout = start + 2 * HZ - jiffies;
2288 
2289 			stopped = list_empty(&bch_cache_sets) &&
2290 				list_empty(&uncached_devices);
2291 
2292 			if (timeout < 0 || stopped)
2293 				break;
2294 
2295 			prepare_to_wait(&unregister_wait, &wait,
2296 					TASK_UNINTERRUPTIBLE);
2297 
2298 			mutex_unlock(&bch_register_lock);
2299 			schedule_timeout(timeout);
2300 			mutex_lock(&bch_register_lock);
2301 		}
2302 
2303 		finish_wait(&unregister_wait, &wait);
2304 
2305 		if (stopped)
2306 			pr_info("All devices stopped");
2307 		else
2308 			pr_notice("Timeout waiting for devices to be closed");
2309 out:
2310 		mutex_unlock(&bch_register_lock);
2311 	}
2312 
2313 	return NOTIFY_DONE;
2314 }
2315 
2316 static struct notifier_block reboot = {
2317 	.notifier_call	= bcache_reboot,
2318 	.priority	= INT_MAX, /* before any real devices */
2319 };
2320 
2321 static void bcache_exit(void)
2322 {
2323 	bch_debug_exit();
2324 	bch_request_exit();
2325 	if (bcache_kobj)
2326 		kobject_put(bcache_kobj);
2327 	if (bcache_wq)
2328 		destroy_workqueue(bcache_wq);
2329 	if (bcache_major)
2330 		unregister_blkdev(bcache_major, "bcache");
2331 	unregister_reboot_notifier(&reboot);
2332 	mutex_destroy(&bch_register_lock);
2333 }
2334 
2335 static int __init bcache_init(void)
2336 {
2337 	static const struct attribute *files[] = {
2338 		&ksysfs_register.attr,
2339 		&ksysfs_register_quiet.attr,
2340 		NULL
2341 	};
2342 
2343 	mutex_init(&bch_register_lock);
2344 	init_waitqueue_head(&unregister_wait);
2345 	register_reboot_notifier(&reboot);
2346 
2347 	bcache_major = register_blkdev(0, "bcache");
2348 	if (bcache_major < 0) {
2349 		unregister_reboot_notifier(&reboot);
2350 		mutex_destroy(&bch_register_lock);
2351 		return bcache_major;
2352 	}
2353 
2354 	if (!(bcache_wq = alloc_workqueue("bcache", WQ_MEM_RECLAIM, 0)) ||
2355 	    !(bcache_kobj = kobject_create_and_add("bcache", fs_kobj)) ||
2356 	    bch_request_init() ||
2357 	    bch_debug_init(bcache_kobj) || closure_debug_init() ||
2358 	    sysfs_create_files(bcache_kobj, files))
2359 		goto err;
2360 
2361 	return 0;
2362 err:
2363 	bcache_exit();
2364 	return -ENOMEM;
2365 }
2366 
2367 module_exit(bcache_exit);
2368 module_init(bcache_init);
2369