1 // SPDX-License-Identifier: GPL-2.0
2 /*
3 * bcache setup/teardown code, and some metadata io - read a superblock and
4 * figure out what to do with it.
5 *
6 * Copyright 2010, 2011 Kent Overstreet <kent.overstreet@gmail.com>
7 * Copyright 2012 Google, Inc.
8 */
9
10 #include "bcache.h"
11 #include "btree.h"
12 #include "debug.h"
13 #include "extents.h"
14 #include "request.h"
15 #include "writeback.h"
16 #include "features.h"
17
18 #include <linux/blkdev.h>
19 #include <linux/debugfs.h>
20 #include <linux/genhd.h>
21 #include <linux/idr.h>
22 #include <linux/kthread.h>
23 #include <linux/workqueue.h>
24 #include <linux/module.h>
25 #include <linux/random.h>
26 #include <linux/reboot.h>
27 #include <linux/sysfs.h>
28
29 unsigned int bch_cutoff_writeback;
30 unsigned int bch_cutoff_writeback_sync;
31
32 static const char bcache_magic[] = {
33 0xc6, 0x85, 0x73, 0xf6, 0x4e, 0x1a, 0x45, 0xca,
34 0x82, 0x65, 0xf5, 0x7f, 0x48, 0xba, 0x6d, 0x81
35 };
36
37 static const char invalid_uuid[] = {
38 0xa0, 0x3e, 0xf8, 0xed, 0x3e, 0xe1, 0xb8, 0x78,
39 0xc8, 0x50, 0xfc, 0x5e, 0xcb, 0x16, 0xcd, 0x99
40 };
41
42 static struct kobject *bcache_kobj;
43 struct mutex bch_register_lock;
44 bool bcache_is_reboot;
45 LIST_HEAD(bch_cache_sets);
46 static LIST_HEAD(uncached_devices);
47
48 static int bcache_major;
49 static DEFINE_IDA(bcache_device_idx);
50 static wait_queue_head_t unregister_wait;
51 struct workqueue_struct *bcache_wq;
52 struct workqueue_struct *bch_flush_wq;
53 struct workqueue_struct *bch_journal_wq;
54
55
56 #define BTREE_MAX_PAGES (256 * 1024 / PAGE_SIZE)
57 /* limitation of partitions number on single bcache device */
58 #define BCACHE_MINORS 128
59 /* limitation of bcache devices number on single system */
60 #define BCACHE_DEVICE_IDX_MAX ((1U << MINORBITS)/BCACHE_MINORS)
61
62 /* Superblock */
63
get_bucket_size(struct cache_sb * sb,struct cache_sb_disk * s)64 static unsigned int get_bucket_size(struct cache_sb *sb, struct cache_sb_disk *s)
65 {
66 unsigned int bucket_size = le16_to_cpu(s->bucket_size);
67
68 if (sb->version >= BCACHE_SB_VERSION_CDEV_WITH_FEATURES) {
69 if (bch_has_feature_large_bucket(sb)) {
70 unsigned int max, order;
71
72 max = sizeof(unsigned int) * BITS_PER_BYTE - 1;
73 order = le16_to_cpu(s->bucket_size);
74 /*
75 * bcache tool will make sure the overflow won't
76 * happen, an error message here is enough.
77 */
78 if (order > max)
79 pr_err("Bucket size (1 << %u) overflows\n",
80 order);
81 bucket_size = 1 << order;
82 } else if (bch_has_feature_obso_large_bucket(sb)) {
83 bucket_size +=
84 le16_to_cpu(s->obso_bucket_size_hi) << 16;
85 }
86 }
87
88 return bucket_size;
89 }
90
read_super_common(struct cache_sb * sb,struct block_device * bdev,struct cache_sb_disk * s)91 static const char *read_super_common(struct cache_sb *sb, struct block_device *bdev,
92 struct cache_sb_disk *s)
93 {
94 const char *err;
95 unsigned int i;
96
97 sb->first_bucket= le16_to_cpu(s->first_bucket);
98 sb->nbuckets = le64_to_cpu(s->nbuckets);
99 sb->bucket_size = get_bucket_size(sb, s);
100
101 sb->nr_in_set = le16_to_cpu(s->nr_in_set);
102 sb->nr_this_dev = le16_to_cpu(s->nr_this_dev);
103
104 err = "Too many journal buckets";
105 if (sb->keys > SB_JOURNAL_BUCKETS)
106 goto err;
107
108 err = "Too many buckets";
109 if (sb->nbuckets > LONG_MAX)
110 goto err;
111
112 err = "Not enough buckets";
113 if (sb->nbuckets < 1 << 7)
114 goto err;
115
116 err = "Bad block size (not power of 2)";
117 if (!is_power_of_2(sb->block_size))
118 goto err;
119
120 err = "Bad block size (larger than page size)";
121 if (sb->block_size > PAGE_SECTORS)
122 goto err;
123
124 err = "Bad bucket size (not power of 2)";
125 if (!is_power_of_2(sb->bucket_size))
126 goto err;
127
128 err = "Bad bucket size (smaller than page size)";
129 if (sb->bucket_size < PAGE_SECTORS)
130 goto err;
131
132 err = "Invalid superblock: device too small";
133 if (get_capacity(bdev->bd_disk) <
134 sb->bucket_size * sb->nbuckets)
135 goto err;
136
137 err = "Bad UUID";
138 if (bch_is_zero(sb->set_uuid, 16))
139 goto err;
140
141 err = "Bad cache device number in set";
142 if (!sb->nr_in_set ||
143 sb->nr_in_set <= sb->nr_this_dev ||
144 sb->nr_in_set > MAX_CACHES_PER_SET)
145 goto err;
146
147 err = "Journal buckets not sequential";
148 for (i = 0; i < sb->keys; i++)
149 if (sb->d[i] != sb->first_bucket + i)
150 goto err;
151
152 err = "Too many journal buckets";
153 if (sb->first_bucket + sb->keys > sb->nbuckets)
154 goto err;
155
156 err = "Invalid superblock: first bucket comes before end of super";
157 if (sb->first_bucket * sb->bucket_size < 16)
158 goto err;
159
160 err = NULL;
161 err:
162 return err;
163 }
164
165
read_super(struct cache_sb * sb,struct block_device * bdev,struct cache_sb_disk ** res)166 static const char *read_super(struct cache_sb *sb, struct block_device *bdev,
167 struct cache_sb_disk **res)
168 {
169 const char *err;
170 struct cache_sb_disk *s;
171 struct page *page;
172 unsigned int i;
173
174 page = read_cache_page_gfp(bdev->bd_inode->i_mapping,
175 SB_OFFSET >> PAGE_SHIFT, GFP_KERNEL);
176 if (IS_ERR(page))
177 return "IO error";
178 s = page_address(page) + offset_in_page(SB_OFFSET);
179
180 sb->offset = le64_to_cpu(s->offset);
181 sb->version = le64_to_cpu(s->version);
182
183 memcpy(sb->magic, s->magic, 16);
184 memcpy(sb->uuid, s->uuid, 16);
185 memcpy(sb->set_uuid, s->set_uuid, 16);
186 memcpy(sb->label, s->label, SB_LABEL_SIZE);
187
188 sb->flags = le64_to_cpu(s->flags);
189 sb->seq = le64_to_cpu(s->seq);
190 sb->last_mount = le32_to_cpu(s->last_mount);
191 sb->keys = le16_to_cpu(s->keys);
192
193 for (i = 0; i < SB_JOURNAL_BUCKETS; i++)
194 sb->d[i] = le64_to_cpu(s->d[i]);
195
196 pr_debug("read sb version %llu, flags %llu, seq %llu, journal size %u\n",
197 sb->version, sb->flags, sb->seq, sb->keys);
198
199 err = "Not a bcache superblock (bad offset)";
200 if (sb->offset != SB_SECTOR)
201 goto err;
202
203 err = "Not a bcache superblock (bad magic)";
204 if (memcmp(sb->magic, bcache_magic, 16))
205 goto err;
206
207 err = "Bad checksum";
208 if (s->csum != csum_set(s))
209 goto err;
210
211 err = "Bad UUID";
212 if (bch_is_zero(sb->uuid, 16))
213 goto err;
214
215 sb->block_size = le16_to_cpu(s->block_size);
216
217 err = "Superblock block size smaller than device block size";
218 if (sb->block_size << 9 < bdev_logical_block_size(bdev))
219 goto err;
220
221 switch (sb->version) {
222 case BCACHE_SB_VERSION_BDEV:
223 sb->data_offset = BDEV_DATA_START_DEFAULT;
224 break;
225 case BCACHE_SB_VERSION_BDEV_WITH_OFFSET:
226 case BCACHE_SB_VERSION_BDEV_WITH_FEATURES:
227 sb->data_offset = le64_to_cpu(s->data_offset);
228
229 err = "Bad data offset";
230 if (sb->data_offset < BDEV_DATA_START_DEFAULT)
231 goto err;
232
233 break;
234 case BCACHE_SB_VERSION_CDEV:
235 case BCACHE_SB_VERSION_CDEV_WITH_UUID:
236 err = read_super_common(sb, bdev, s);
237 if (err)
238 goto err;
239 break;
240 case BCACHE_SB_VERSION_CDEV_WITH_FEATURES:
241 /*
242 * Feature bits are needed in read_super_common(),
243 * convert them firstly.
244 */
245 sb->feature_compat = le64_to_cpu(s->feature_compat);
246 sb->feature_incompat = le64_to_cpu(s->feature_incompat);
247 sb->feature_ro_compat = le64_to_cpu(s->feature_ro_compat);
248
249 /* Check incompatible features */
250 err = "Unsupported compatible feature found";
251 if (bch_has_unknown_compat_features(sb))
252 goto err;
253
254 err = "Unsupported read-only compatible feature found";
255 if (bch_has_unknown_ro_compat_features(sb))
256 goto err;
257
258 err = "Unsupported incompatible feature found";
259 if (bch_has_unknown_incompat_features(sb))
260 goto err;
261
262 err = read_super_common(sb, bdev, s);
263 if (err)
264 goto err;
265 break;
266 default:
267 err = "Unsupported superblock version";
268 goto err;
269 }
270
271 sb->last_mount = (u32)ktime_get_real_seconds();
272 *res = s;
273 return NULL;
274 err:
275 put_page(page);
276 return err;
277 }
278
write_bdev_super_endio(struct bio * bio)279 static void write_bdev_super_endio(struct bio *bio)
280 {
281 struct cached_dev *dc = bio->bi_private;
282
283 if (bio->bi_status)
284 bch_count_backing_io_errors(dc, bio);
285
286 closure_put(&dc->sb_write);
287 }
288
__write_super(struct cache_sb * sb,struct cache_sb_disk * out,struct bio * bio)289 static void __write_super(struct cache_sb *sb, struct cache_sb_disk *out,
290 struct bio *bio)
291 {
292 unsigned int i;
293
294 bio->bi_opf = REQ_OP_WRITE | REQ_SYNC | REQ_META;
295 bio->bi_iter.bi_sector = SB_SECTOR;
296 __bio_add_page(bio, virt_to_page(out), SB_SIZE,
297 offset_in_page(out));
298
299 out->offset = cpu_to_le64(sb->offset);
300
301 memcpy(out->uuid, sb->uuid, 16);
302 memcpy(out->set_uuid, sb->set_uuid, 16);
303 memcpy(out->label, sb->label, SB_LABEL_SIZE);
304
305 out->flags = cpu_to_le64(sb->flags);
306 out->seq = cpu_to_le64(sb->seq);
307
308 out->last_mount = cpu_to_le32(sb->last_mount);
309 out->first_bucket = cpu_to_le16(sb->first_bucket);
310 out->keys = cpu_to_le16(sb->keys);
311
312 for (i = 0; i < sb->keys; i++)
313 out->d[i] = cpu_to_le64(sb->d[i]);
314
315 if (sb->version >= BCACHE_SB_VERSION_CDEV_WITH_FEATURES) {
316 out->feature_compat = cpu_to_le64(sb->feature_compat);
317 out->feature_incompat = cpu_to_le64(sb->feature_incompat);
318 out->feature_ro_compat = cpu_to_le64(sb->feature_ro_compat);
319 }
320
321 out->version = cpu_to_le64(sb->version);
322 out->csum = csum_set(out);
323
324 pr_debug("ver %llu, flags %llu, seq %llu\n",
325 sb->version, sb->flags, sb->seq);
326
327 submit_bio(bio);
328 }
329
bch_write_bdev_super_unlock(struct closure * cl)330 static void bch_write_bdev_super_unlock(struct closure *cl)
331 {
332 struct cached_dev *dc = container_of(cl, struct cached_dev, sb_write);
333
334 up(&dc->sb_write_mutex);
335 }
336
bch_write_bdev_super(struct cached_dev * dc,struct closure * parent)337 void bch_write_bdev_super(struct cached_dev *dc, struct closure *parent)
338 {
339 struct closure *cl = &dc->sb_write;
340 struct bio *bio = &dc->sb_bio;
341
342 down(&dc->sb_write_mutex);
343 closure_init(cl, parent);
344
345 bio_init(bio, dc->sb_bv, 1);
346 bio_set_dev(bio, dc->bdev);
347 bio->bi_end_io = write_bdev_super_endio;
348 bio->bi_private = dc;
349
350 closure_get(cl);
351 /* I/O request sent to backing device */
352 __write_super(&dc->sb, dc->sb_disk, bio);
353
354 closure_return_with_destructor(cl, bch_write_bdev_super_unlock);
355 }
356
write_super_endio(struct bio * bio)357 static void write_super_endio(struct bio *bio)
358 {
359 struct cache *ca = bio->bi_private;
360
361 /* is_read = 0 */
362 bch_count_io_errors(ca, bio->bi_status, 0,
363 "writing superblock");
364 closure_put(&ca->set->sb_write);
365 }
366
bcache_write_super_unlock(struct closure * cl)367 static void bcache_write_super_unlock(struct closure *cl)
368 {
369 struct cache_set *c = container_of(cl, struct cache_set, sb_write);
370
371 up(&c->sb_write_mutex);
372 }
373
bcache_write_super(struct cache_set * c)374 void bcache_write_super(struct cache_set *c)
375 {
376 struct closure *cl = &c->sb_write;
377 struct cache *ca = c->cache;
378 struct bio *bio = &ca->sb_bio;
379 unsigned int version = BCACHE_SB_VERSION_CDEV_WITH_UUID;
380
381 down(&c->sb_write_mutex);
382 closure_init(cl, &c->cl);
383
384 ca->sb.seq++;
385
386 if (ca->sb.version < version)
387 ca->sb.version = version;
388
389 bio_init(bio, ca->sb_bv, 1);
390 bio_set_dev(bio, ca->bdev);
391 bio->bi_end_io = write_super_endio;
392 bio->bi_private = ca;
393
394 closure_get(cl);
395 __write_super(&ca->sb, ca->sb_disk, bio);
396
397 closure_return_with_destructor(cl, bcache_write_super_unlock);
398 }
399
400 /* UUID io */
401
uuid_endio(struct bio * bio)402 static void uuid_endio(struct bio *bio)
403 {
404 struct closure *cl = bio->bi_private;
405 struct cache_set *c = container_of(cl, struct cache_set, uuid_write);
406
407 cache_set_err_on(bio->bi_status, c, "accessing uuids");
408 bch_bbio_free(bio, c);
409 closure_put(cl);
410 }
411
uuid_io_unlock(struct closure * cl)412 static void uuid_io_unlock(struct closure *cl)
413 {
414 struct cache_set *c = container_of(cl, struct cache_set, uuid_write);
415
416 up(&c->uuid_write_mutex);
417 }
418
uuid_io(struct cache_set * c,int op,unsigned long op_flags,struct bkey * k,struct closure * parent)419 static void uuid_io(struct cache_set *c, int op, unsigned long op_flags,
420 struct bkey *k, struct closure *parent)
421 {
422 struct closure *cl = &c->uuid_write;
423 struct uuid_entry *u;
424 unsigned int i;
425 char buf[80];
426
427 BUG_ON(!parent);
428 down(&c->uuid_write_mutex);
429 closure_init(cl, parent);
430
431 for (i = 0; i < KEY_PTRS(k); i++) {
432 struct bio *bio = bch_bbio_alloc(c);
433
434 bio->bi_opf = REQ_SYNC | REQ_META | op_flags;
435 bio->bi_iter.bi_size = KEY_SIZE(k) << 9;
436
437 bio->bi_end_io = uuid_endio;
438 bio->bi_private = cl;
439 bio_set_op_attrs(bio, op, REQ_SYNC|REQ_META|op_flags);
440 bch_bio_map(bio, c->uuids);
441
442 bch_submit_bbio(bio, c, k, i);
443
444 if (op != REQ_OP_WRITE)
445 break;
446 }
447
448 bch_extent_to_text(buf, sizeof(buf), k);
449 pr_debug("%s UUIDs at %s\n", op == REQ_OP_WRITE ? "wrote" : "read", buf);
450
451 for (u = c->uuids; u < c->uuids + c->nr_uuids; u++)
452 if (!bch_is_zero(u->uuid, 16))
453 pr_debug("Slot %zi: %pU: %s: 1st: %u last: %u inv: %u\n",
454 u - c->uuids, u->uuid, u->label,
455 u->first_reg, u->last_reg, u->invalidated);
456
457 closure_return_with_destructor(cl, uuid_io_unlock);
458 }
459
uuid_read(struct cache_set * c,struct jset * j,struct closure * cl)460 static char *uuid_read(struct cache_set *c, struct jset *j, struct closure *cl)
461 {
462 struct bkey *k = &j->uuid_bucket;
463
464 if (__bch_btree_ptr_invalid(c, k))
465 return "bad uuid pointer";
466
467 bkey_copy(&c->uuid_bucket, k);
468 uuid_io(c, REQ_OP_READ, 0, k, cl);
469
470 if (j->version < BCACHE_JSET_VERSION_UUIDv1) {
471 struct uuid_entry_v0 *u0 = (void *) c->uuids;
472 struct uuid_entry *u1 = (void *) c->uuids;
473 int i;
474
475 closure_sync(cl);
476
477 /*
478 * Since the new uuid entry is bigger than the old, we have to
479 * convert starting at the highest memory address and work down
480 * in order to do it in place
481 */
482
483 for (i = c->nr_uuids - 1;
484 i >= 0;
485 --i) {
486 memcpy(u1[i].uuid, u0[i].uuid, 16);
487 memcpy(u1[i].label, u0[i].label, 32);
488
489 u1[i].first_reg = u0[i].first_reg;
490 u1[i].last_reg = u0[i].last_reg;
491 u1[i].invalidated = u0[i].invalidated;
492
493 u1[i].flags = 0;
494 u1[i].sectors = 0;
495 }
496 }
497
498 return NULL;
499 }
500
__uuid_write(struct cache_set * c)501 static int __uuid_write(struct cache_set *c)
502 {
503 BKEY_PADDED(key) k;
504 struct closure cl;
505 struct cache *ca = c->cache;
506 unsigned int size;
507
508 closure_init_stack(&cl);
509 lockdep_assert_held(&bch_register_lock);
510
511 if (bch_bucket_alloc_set(c, RESERVE_BTREE, &k.key, true))
512 return 1;
513
514 size = meta_bucket_pages(&ca->sb) * PAGE_SECTORS;
515 SET_KEY_SIZE(&k.key, size);
516 uuid_io(c, REQ_OP_WRITE, 0, &k.key, &cl);
517 closure_sync(&cl);
518
519 /* Only one bucket used for uuid write */
520 atomic_long_add(ca->sb.bucket_size, &ca->meta_sectors_written);
521
522 bkey_copy(&c->uuid_bucket, &k.key);
523 bkey_put(c, &k.key);
524 return 0;
525 }
526
bch_uuid_write(struct cache_set * c)527 int bch_uuid_write(struct cache_set *c)
528 {
529 int ret = __uuid_write(c);
530
531 if (!ret)
532 bch_journal_meta(c, NULL);
533
534 return ret;
535 }
536
uuid_find(struct cache_set * c,const char * uuid)537 static struct uuid_entry *uuid_find(struct cache_set *c, const char *uuid)
538 {
539 struct uuid_entry *u;
540
541 for (u = c->uuids;
542 u < c->uuids + c->nr_uuids; u++)
543 if (!memcmp(u->uuid, uuid, 16))
544 return u;
545
546 return NULL;
547 }
548
uuid_find_empty(struct cache_set * c)549 static struct uuid_entry *uuid_find_empty(struct cache_set *c)
550 {
551 static const char zero_uuid[16] = "\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0";
552
553 return uuid_find(c, zero_uuid);
554 }
555
556 /*
557 * Bucket priorities/gens:
558 *
559 * For each bucket, we store on disk its
560 * 8 bit gen
561 * 16 bit priority
562 *
563 * See alloc.c for an explanation of the gen. The priority is used to implement
564 * lru (and in the future other) cache replacement policies; for most purposes
565 * it's just an opaque integer.
566 *
567 * The gens and the priorities don't have a whole lot to do with each other, and
568 * it's actually the gens that must be written out at specific times - it's no
569 * big deal if the priorities don't get written, if we lose them we just reuse
570 * buckets in suboptimal order.
571 *
572 * On disk they're stored in a packed array, and in as many buckets are required
573 * to fit them all. The buckets we use to store them form a list; the journal
574 * header points to the first bucket, the first bucket points to the second
575 * bucket, et cetera.
576 *
577 * This code is used by the allocation code; periodically (whenever it runs out
578 * of buckets to allocate from) the allocation code will invalidate some
579 * buckets, but it can't use those buckets until their new gens are safely on
580 * disk.
581 */
582
prio_endio(struct bio * bio)583 static void prio_endio(struct bio *bio)
584 {
585 struct cache *ca = bio->bi_private;
586
587 cache_set_err_on(bio->bi_status, ca->set, "accessing priorities");
588 bch_bbio_free(bio, ca->set);
589 closure_put(&ca->prio);
590 }
591
prio_io(struct cache * ca,uint64_t bucket,int op,unsigned long op_flags)592 static void prio_io(struct cache *ca, uint64_t bucket, int op,
593 unsigned long op_flags)
594 {
595 struct closure *cl = &ca->prio;
596 struct bio *bio = bch_bbio_alloc(ca->set);
597
598 closure_init_stack(cl);
599
600 bio->bi_iter.bi_sector = bucket * ca->sb.bucket_size;
601 bio_set_dev(bio, ca->bdev);
602 bio->bi_iter.bi_size = meta_bucket_bytes(&ca->sb);
603
604 bio->bi_end_io = prio_endio;
605 bio->bi_private = ca;
606 bio_set_op_attrs(bio, op, REQ_SYNC|REQ_META|op_flags);
607 bch_bio_map(bio, ca->disk_buckets);
608
609 closure_bio_submit(ca->set, bio, &ca->prio);
610 closure_sync(cl);
611 }
612
bch_prio_write(struct cache * ca,bool wait)613 int bch_prio_write(struct cache *ca, bool wait)
614 {
615 int i;
616 struct bucket *b;
617 struct closure cl;
618
619 pr_debug("free_prio=%zu, free_none=%zu, free_inc=%zu\n",
620 fifo_used(&ca->free[RESERVE_PRIO]),
621 fifo_used(&ca->free[RESERVE_NONE]),
622 fifo_used(&ca->free_inc));
623
624 /*
625 * Pre-check if there are enough free buckets. In the non-blocking
626 * scenario it's better to fail early rather than starting to allocate
627 * buckets and do a cleanup later in case of failure.
628 */
629 if (!wait) {
630 size_t avail = fifo_used(&ca->free[RESERVE_PRIO]) +
631 fifo_used(&ca->free[RESERVE_NONE]);
632 if (prio_buckets(ca) > avail)
633 return -ENOMEM;
634 }
635
636 closure_init_stack(&cl);
637
638 lockdep_assert_held(&ca->set->bucket_lock);
639
640 ca->disk_buckets->seq++;
641
642 atomic_long_add(ca->sb.bucket_size * prio_buckets(ca),
643 &ca->meta_sectors_written);
644
645 for (i = prio_buckets(ca) - 1; i >= 0; --i) {
646 long bucket;
647 struct prio_set *p = ca->disk_buckets;
648 struct bucket_disk *d = p->data;
649 struct bucket_disk *end = d + prios_per_bucket(ca);
650
651 for (b = ca->buckets + i * prios_per_bucket(ca);
652 b < ca->buckets + ca->sb.nbuckets && d < end;
653 b++, d++) {
654 d->prio = cpu_to_le16(b->prio);
655 d->gen = b->gen;
656 }
657
658 p->next_bucket = ca->prio_buckets[i + 1];
659 p->magic = pset_magic(&ca->sb);
660 p->csum = bch_crc64(&p->magic, meta_bucket_bytes(&ca->sb) - 8);
661
662 bucket = bch_bucket_alloc(ca, RESERVE_PRIO, wait);
663 BUG_ON(bucket == -1);
664
665 mutex_unlock(&ca->set->bucket_lock);
666 prio_io(ca, bucket, REQ_OP_WRITE, 0);
667 mutex_lock(&ca->set->bucket_lock);
668
669 ca->prio_buckets[i] = bucket;
670 atomic_dec_bug(&ca->buckets[bucket].pin);
671 }
672
673 mutex_unlock(&ca->set->bucket_lock);
674
675 bch_journal_meta(ca->set, &cl);
676 closure_sync(&cl);
677
678 mutex_lock(&ca->set->bucket_lock);
679
680 /*
681 * Don't want the old priorities to get garbage collected until after we
682 * finish writing the new ones, and they're journalled
683 */
684 for (i = 0; i < prio_buckets(ca); i++) {
685 if (ca->prio_last_buckets[i])
686 __bch_bucket_free(ca,
687 &ca->buckets[ca->prio_last_buckets[i]]);
688
689 ca->prio_last_buckets[i] = ca->prio_buckets[i];
690 }
691 return 0;
692 }
693
prio_read(struct cache * ca,uint64_t bucket)694 static int prio_read(struct cache *ca, uint64_t bucket)
695 {
696 struct prio_set *p = ca->disk_buckets;
697 struct bucket_disk *d = p->data + prios_per_bucket(ca), *end = d;
698 struct bucket *b;
699 unsigned int bucket_nr = 0;
700 int ret = -EIO;
701
702 for (b = ca->buckets;
703 b < ca->buckets + ca->sb.nbuckets;
704 b++, d++) {
705 if (d == end) {
706 ca->prio_buckets[bucket_nr] = bucket;
707 ca->prio_last_buckets[bucket_nr] = bucket;
708 bucket_nr++;
709
710 prio_io(ca, bucket, REQ_OP_READ, 0);
711
712 if (p->csum !=
713 bch_crc64(&p->magic, meta_bucket_bytes(&ca->sb) - 8)) {
714 pr_warn("bad csum reading priorities\n");
715 goto out;
716 }
717
718 if (p->magic != pset_magic(&ca->sb)) {
719 pr_warn("bad magic reading priorities\n");
720 goto out;
721 }
722
723 bucket = p->next_bucket;
724 d = p->data;
725 }
726
727 b->prio = le16_to_cpu(d->prio);
728 b->gen = b->last_gc = d->gen;
729 }
730
731 ret = 0;
732 out:
733 return ret;
734 }
735
736 /* Bcache device */
737
open_dev(struct block_device * b,fmode_t mode)738 static int open_dev(struct block_device *b, fmode_t mode)
739 {
740 struct bcache_device *d = b->bd_disk->private_data;
741
742 if (test_bit(BCACHE_DEV_CLOSING, &d->flags))
743 return -ENXIO;
744
745 closure_get(&d->cl);
746 return 0;
747 }
748
release_dev(struct gendisk * b,fmode_t mode)749 static void release_dev(struct gendisk *b, fmode_t mode)
750 {
751 struct bcache_device *d = b->private_data;
752
753 closure_put(&d->cl);
754 }
755
ioctl_dev(struct block_device * b,fmode_t mode,unsigned int cmd,unsigned long arg)756 static int ioctl_dev(struct block_device *b, fmode_t mode,
757 unsigned int cmd, unsigned long arg)
758 {
759 struct bcache_device *d = b->bd_disk->private_data;
760
761 return d->ioctl(d, mode, cmd, arg);
762 }
763
764 static const struct block_device_operations bcache_cached_ops = {
765 .submit_bio = cached_dev_submit_bio,
766 .open = open_dev,
767 .release = release_dev,
768 .ioctl = ioctl_dev,
769 .owner = THIS_MODULE,
770 };
771
772 static const struct block_device_operations bcache_flash_ops = {
773 .submit_bio = flash_dev_submit_bio,
774 .open = open_dev,
775 .release = release_dev,
776 .ioctl = ioctl_dev,
777 .owner = THIS_MODULE,
778 };
779
bcache_device_stop(struct bcache_device * d)780 void bcache_device_stop(struct bcache_device *d)
781 {
782 if (!test_and_set_bit(BCACHE_DEV_CLOSING, &d->flags))
783 /*
784 * closure_fn set to
785 * - cached device: cached_dev_flush()
786 * - flash dev: flash_dev_flush()
787 */
788 closure_queue(&d->cl);
789 }
790
bcache_device_unlink(struct bcache_device * d)791 static void bcache_device_unlink(struct bcache_device *d)
792 {
793 lockdep_assert_held(&bch_register_lock);
794
795 if (d->c && !test_and_set_bit(BCACHE_DEV_UNLINK_DONE, &d->flags)) {
796 struct cache *ca = d->c->cache;
797
798 sysfs_remove_link(&d->c->kobj, d->name);
799 sysfs_remove_link(&d->kobj, "cache");
800
801 bd_unlink_disk_holder(ca->bdev, d->disk);
802 }
803 }
804
bcache_device_link(struct bcache_device * d,struct cache_set * c,const char * name)805 static void bcache_device_link(struct bcache_device *d, struct cache_set *c,
806 const char *name)
807 {
808 struct cache *ca = c->cache;
809 int ret;
810
811 bd_link_disk_holder(ca->bdev, d->disk);
812
813 snprintf(d->name, BCACHEDEVNAME_SIZE,
814 "%s%u", name, d->id);
815
816 ret = sysfs_create_link(&d->kobj, &c->kobj, "cache");
817 if (ret < 0)
818 pr_err("Couldn't create device -> cache set symlink\n");
819
820 ret = sysfs_create_link(&c->kobj, &d->kobj, d->name);
821 if (ret < 0)
822 pr_err("Couldn't create cache set -> device symlink\n");
823
824 clear_bit(BCACHE_DEV_UNLINK_DONE, &d->flags);
825 }
826
bcache_device_detach(struct bcache_device * d)827 static void bcache_device_detach(struct bcache_device *d)
828 {
829 lockdep_assert_held(&bch_register_lock);
830
831 atomic_dec(&d->c->attached_dev_nr);
832
833 if (test_bit(BCACHE_DEV_DETACHING, &d->flags)) {
834 struct uuid_entry *u = d->c->uuids + d->id;
835
836 SET_UUID_FLASH_ONLY(u, 0);
837 memcpy(u->uuid, invalid_uuid, 16);
838 u->invalidated = cpu_to_le32((u32)ktime_get_real_seconds());
839 bch_uuid_write(d->c);
840 }
841
842 bcache_device_unlink(d);
843
844 d->c->devices[d->id] = NULL;
845 closure_put(&d->c->caching);
846 d->c = NULL;
847 }
848
bcache_device_attach(struct bcache_device * d,struct cache_set * c,unsigned int id)849 static void bcache_device_attach(struct bcache_device *d, struct cache_set *c,
850 unsigned int id)
851 {
852 d->id = id;
853 d->c = c;
854 c->devices[id] = d;
855
856 if (id >= c->devices_max_used)
857 c->devices_max_used = id + 1;
858
859 closure_get(&c->caching);
860 }
861
first_minor_to_idx(int first_minor)862 static inline int first_minor_to_idx(int first_minor)
863 {
864 return (first_minor/BCACHE_MINORS);
865 }
866
idx_to_first_minor(int idx)867 static inline int idx_to_first_minor(int idx)
868 {
869 return (idx * BCACHE_MINORS);
870 }
871
bcache_device_free(struct bcache_device * d)872 static void bcache_device_free(struct bcache_device *d)
873 {
874 struct gendisk *disk = d->disk;
875
876 lockdep_assert_held(&bch_register_lock);
877
878 if (disk)
879 pr_info("%s stopped\n", disk->disk_name);
880 else
881 pr_err("bcache device (NULL gendisk) stopped\n");
882
883 if (d->c)
884 bcache_device_detach(d);
885
886 if (disk) {
887 bool disk_added = (disk->flags & GENHD_FL_UP) != 0;
888
889 if (disk_added)
890 del_gendisk(disk);
891
892 if (disk->queue)
893 blk_cleanup_queue(disk->queue);
894
895 ida_simple_remove(&bcache_device_idx,
896 first_minor_to_idx(disk->first_minor));
897 if (disk_added)
898 put_disk(disk);
899 }
900
901 bioset_exit(&d->bio_split);
902 kvfree(d->full_dirty_stripes);
903 kvfree(d->stripe_sectors_dirty);
904
905 closure_debug_destroy(&d->cl);
906 }
907
bcache_device_init(struct bcache_device * d,unsigned int block_size,sector_t sectors,struct block_device * cached_bdev,const struct block_device_operations * ops)908 static int bcache_device_init(struct bcache_device *d, unsigned int block_size,
909 sector_t sectors, struct block_device *cached_bdev,
910 const struct block_device_operations *ops)
911 {
912 struct request_queue *q;
913 const size_t max_stripes = min_t(size_t, INT_MAX,
914 SIZE_MAX / sizeof(atomic_t));
915 uint64_t n;
916 int idx;
917
918 if (!d->stripe_size)
919 d->stripe_size = 1 << 31;
920
921 n = DIV_ROUND_UP_ULL(sectors, d->stripe_size);
922 if (!n || n > max_stripes) {
923 pr_err("nr_stripes too large or invalid: %llu (start sector beyond end of disk?)\n",
924 n);
925 return -ENOMEM;
926 }
927 d->nr_stripes = n;
928
929 n = d->nr_stripes * sizeof(atomic_t);
930 d->stripe_sectors_dirty = kvzalloc(n, GFP_KERNEL);
931 if (!d->stripe_sectors_dirty)
932 return -ENOMEM;
933
934 n = BITS_TO_LONGS(d->nr_stripes) * sizeof(unsigned long);
935 d->full_dirty_stripes = kvzalloc(n, GFP_KERNEL);
936 if (!d->full_dirty_stripes)
937 goto out_free_stripe_sectors_dirty;
938
939 idx = ida_simple_get(&bcache_device_idx, 0,
940 BCACHE_DEVICE_IDX_MAX, GFP_KERNEL);
941 if (idx < 0)
942 goto out_free_full_dirty_stripes;
943
944 if (bioset_init(&d->bio_split, 4, offsetof(struct bbio, bio),
945 BIOSET_NEED_BVECS|BIOSET_NEED_RESCUER))
946 goto out_ida_remove;
947
948 d->disk = alloc_disk(BCACHE_MINORS);
949 if (!d->disk)
950 goto out_bioset_exit;
951
952 set_capacity(d->disk, sectors);
953 snprintf(d->disk->disk_name, DISK_NAME_LEN, "bcache%i", idx);
954
955 d->disk->major = bcache_major;
956 d->disk->first_minor = idx_to_first_minor(idx);
957 d->disk->fops = ops;
958 d->disk->private_data = d;
959
960 q = blk_alloc_queue(NUMA_NO_NODE);
961 if (!q)
962 return -ENOMEM;
963
964 d->disk->queue = q;
965 q->limits.max_hw_sectors = UINT_MAX;
966 q->limits.max_sectors = UINT_MAX;
967 q->limits.max_segment_size = UINT_MAX;
968 q->limits.max_segments = BIO_MAX_PAGES;
969 blk_queue_max_discard_sectors(q, UINT_MAX);
970 q->limits.discard_granularity = 512;
971 q->limits.io_min = block_size;
972 q->limits.logical_block_size = block_size;
973 q->limits.physical_block_size = block_size;
974
975 if (q->limits.logical_block_size > PAGE_SIZE && cached_bdev) {
976 /*
977 * This should only happen with BCACHE_SB_VERSION_BDEV.
978 * Block/page size is checked for BCACHE_SB_VERSION_CDEV.
979 */
980 pr_info("%s: sb/logical block size (%u) greater than page size (%lu) falling back to device logical block size (%u)\n",
981 d->disk->disk_name, q->limits.logical_block_size,
982 PAGE_SIZE, bdev_logical_block_size(cached_bdev));
983
984 /* This also adjusts physical block size/min io size if needed */
985 blk_queue_logical_block_size(q, bdev_logical_block_size(cached_bdev));
986 }
987
988 blk_queue_flag_set(QUEUE_FLAG_NONROT, d->disk->queue);
989 blk_queue_flag_clear(QUEUE_FLAG_ADD_RANDOM, d->disk->queue);
990 blk_queue_flag_set(QUEUE_FLAG_DISCARD, d->disk->queue);
991
992 blk_queue_write_cache(q, true, true);
993
994 return 0;
995
996 out_bioset_exit:
997 bioset_exit(&d->bio_split);
998 out_ida_remove:
999 ida_simple_remove(&bcache_device_idx, idx);
1000 out_free_full_dirty_stripes:
1001 kvfree(d->full_dirty_stripes);
1002 out_free_stripe_sectors_dirty:
1003 kvfree(d->stripe_sectors_dirty);
1004 return -ENOMEM;
1005
1006 }
1007
1008 /* Cached device */
1009
calc_cached_dev_sectors(struct cache_set * c)1010 static void calc_cached_dev_sectors(struct cache_set *c)
1011 {
1012 uint64_t sectors = 0;
1013 struct cached_dev *dc;
1014
1015 list_for_each_entry(dc, &c->cached_devs, list)
1016 sectors += bdev_sectors(dc->bdev);
1017
1018 c->cached_dev_sectors = sectors;
1019 }
1020
1021 #define BACKING_DEV_OFFLINE_TIMEOUT 5
cached_dev_status_update(void * arg)1022 static int cached_dev_status_update(void *arg)
1023 {
1024 struct cached_dev *dc = arg;
1025 struct request_queue *q;
1026
1027 /*
1028 * If this delayed worker is stopping outside, directly quit here.
1029 * dc->io_disable might be set via sysfs interface, so check it
1030 * here too.
1031 */
1032 while (!kthread_should_stop() && !dc->io_disable) {
1033 q = bdev_get_queue(dc->bdev);
1034 if (blk_queue_dying(q))
1035 dc->offline_seconds++;
1036 else
1037 dc->offline_seconds = 0;
1038
1039 if (dc->offline_seconds >= BACKING_DEV_OFFLINE_TIMEOUT) {
1040 pr_err("%s: device offline for %d seconds\n",
1041 dc->backing_dev_name,
1042 BACKING_DEV_OFFLINE_TIMEOUT);
1043 pr_err("%s: disable I/O request due to backing device offline\n",
1044 dc->disk.name);
1045 dc->io_disable = true;
1046 /* let others know earlier that io_disable is true */
1047 smp_mb();
1048 bcache_device_stop(&dc->disk);
1049 break;
1050 }
1051 schedule_timeout_interruptible(HZ);
1052 }
1053
1054 wait_for_kthread_stop();
1055 return 0;
1056 }
1057
1058
bch_cached_dev_run(struct cached_dev * dc)1059 int bch_cached_dev_run(struct cached_dev *dc)
1060 {
1061 struct bcache_device *d = &dc->disk;
1062 char *buf = kmemdup_nul(dc->sb.label, SB_LABEL_SIZE, GFP_KERNEL);
1063 char *env[] = {
1064 "DRIVER=bcache",
1065 kasprintf(GFP_KERNEL, "CACHED_UUID=%pU", dc->sb.uuid),
1066 kasprintf(GFP_KERNEL, "CACHED_LABEL=%s", buf ? : ""),
1067 NULL,
1068 };
1069
1070 if (dc->io_disable) {
1071 pr_err("I/O disabled on cached dev %s\n",
1072 dc->backing_dev_name);
1073 kfree(env[1]);
1074 kfree(env[2]);
1075 kfree(buf);
1076 return -EIO;
1077 }
1078
1079 if (atomic_xchg(&dc->running, 1)) {
1080 kfree(env[1]);
1081 kfree(env[2]);
1082 kfree(buf);
1083 pr_info("cached dev %s is running already\n",
1084 dc->backing_dev_name);
1085 return -EBUSY;
1086 }
1087
1088 if (!d->c &&
1089 BDEV_STATE(&dc->sb) != BDEV_STATE_NONE) {
1090 struct closure cl;
1091
1092 closure_init_stack(&cl);
1093
1094 SET_BDEV_STATE(&dc->sb, BDEV_STATE_STALE);
1095 bch_write_bdev_super(dc, &cl);
1096 closure_sync(&cl);
1097 }
1098
1099 add_disk(d->disk);
1100 bd_link_disk_holder(dc->bdev, dc->disk.disk);
1101 /*
1102 * won't show up in the uevent file, use udevadm monitor -e instead
1103 * only class / kset properties are persistent
1104 */
1105 kobject_uevent_env(&disk_to_dev(d->disk)->kobj, KOBJ_CHANGE, env);
1106 kfree(env[1]);
1107 kfree(env[2]);
1108 kfree(buf);
1109
1110 if (sysfs_create_link(&d->kobj, &disk_to_dev(d->disk)->kobj, "dev") ||
1111 sysfs_create_link(&disk_to_dev(d->disk)->kobj,
1112 &d->kobj, "bcache")) {
1113 pr_err("Couldn't create bcache dev <-> disk sysfs symlinks\n");
1114 return -ENOMEM;
1115 }
1116
1117 dc->status_update_thread = kthread_run(cached_dev_status_update,
1118 dc, "bcache_status_update");
1119 if (IS_ERR(dc->status_update_thread)) {
1120 pr_warn("failed to create bcache_status_update kthread, continue to run without monitoring backing device status\n");
1121 }
1122
1123 return 0;
1124 }
1125
1126 /*
1127 * If BCACHE_DEV_RATE_DW_RUNNING is set, it means routine of the delayed
1128 * work dc->writeback_rate_update is running. Wait until the routine
1129 * quits (BCACHE_DEV_RATE_DW_RUNNING is clear), then continue to
1130 * cancel it. If BCACHE_DEV_RATE_DW_RUNNING is not clear after time_out
1131 * seconds, give up waiting here and continue to cancel it too.
1132 */
cancel_writeback_rate_update_dwork(struct cached_dev * dc)1133 static void cancel_writeback_rate_update_dwork(struct cached_dev *dc)
1134 {
1135 int time_out = WRITEBACK_RATE_UPDATE_SECS_MAX * HZ;
1136
1137 do {
1138 if (!test_bit(BCACHE_DEV_RATE_DW_RUNNING,
1139 &dc->disk.flags))
1140 break;
1141 time_out--;
1142 schedule_timeout_interruptible(1);
1143 } while (time_out > 0);
1144
1145 if (time_out == 0)
1146 pr_warn("give up waiting for dc->writeback_write_update to quit\n");
1147
1148 cancel_delayed_work_sync(&dc->writeback_rate_update);
1149 }
1150
cached_dev_detach_finish(struct work_struct * w)1151 static void cached_dev_detach_finish(struct work_struct *w)
1152 {
1153 struct cached_dev *dc = container_of(w, struct cached_dev, detach);
1154 struct closure cl;
1155
1156 closure_init_stack(&cl);
1157
1158 BUG_ON(!test_bit(BCACHE_DEV_DETACHING, &dc->disk.flags));
1159 BUG_ON(refcount_read(&dc->count));
1160
1161
1162 if (test_and_clear_bit(BCACHE_DEV_WB_RUNNING, &dc->disk.flags))
1163 cancel_writeback_rate_update_dwork(dc);
1164
1165 if (!IS_ERR_OR_NULL(dc->writeback_thread)) {
1166 kthread_stop(dc->writeback_thread);
1167 dc->writeback_thread = NULL;
1168 }
1169
1170 memset(&dc->sb.set_uuid, 0, 16);
1171 SET_BDEV_STATE(&dc->sb, BDEV_STATE_NONE);
1172
1173 bch_write_bdev_super(dc, &cl);
1174 closure_sync(&cl);
1175
1176 mutex_lock(&bch_register_lock);
1177
1178 calc_cached_dev_sectors(dc->disk.c);
1179 bcache_device_detach(&dc->disk);
1180 list_move(&dc->list, &uncached_devices);
1181
1182 clear_bit(BCACHE_DEV_DETACHING, &dc->disk.flags);
1183 clear_bit(BCACHE_DEV_UNLINK_DONE, &dc->disk.flags);
1184
1185 mutex_unlock(&bch_register_lock);
1186
1187 pr_info("Caching disabled for %s\n", dc->backing_dev_name);
1188
1189 /* Drop ref we took in cached_dev_detach() */
1190 closure_put(&dc->disk.cl);
1191 }
1192
bch_cached_dev_detach(struct cached_dev * dc)1193 void bch_cached_dev_detach(struct cached_dev *dc)
1194 {
1195 lockdep_assert_held(&bch_register_lock);
1196
1197 if (test_bit(BCACHE_DEV_CLOSING, &dc->disk.flags))
1198 return;
1199
1200 if (test_and_set_bit(BCACHE_DEV_DETACHING, &dc->disk.flags))
1201 return;
1202
1203 /*
1204 * Block the device from being closed and freed until we're finished
1205 * detaching
1206 */
1207 closure_get(&dc->disk.cl);
1208
1209 bch_writeback_queue(dc);
1210
1211 cached_dev_put(dc);
1212 }
1213
bch_cached_dev_attach(struct cached_dev * dc,struct cache_set * c,uint8_t * set_uuid)1214 int bch_cached_dev_attach(struct cached_dev *dc, struct cache_set *c,
1215 uint8_t *set_uuid)
1216 {
1217 uint32_t rtime = cpu_to_le32((u32)ktime_get_real_seconds());
1218 struct uuid_entry *u;
1219 struct cached_dev *exist_dc, *t;
1220 int ret = 0;
1221
1222 if ((set_uuid && memcmp(set_uuid, c->set_uuid, 16)) ||
1223 (!set_uuid && memcmp(dc->sb.set_uuid, c->set_uuid, 16)))
1224 return -ENOENT;
1225
1226 if (dc->disk.c) {
1227 pr_err("Can't attach %s: already attached\n",
1228 dc->backing_dev_name);
1229 return -EINVAL;
1230 }
1231
1232 if (test_bit(CACHE_SET_STOPPING, &c->flags)) {
1233 pr_err("Can't attach %s: shutting down\n",
1234 dc->backing_dev_name);
1235 return -EINVAL;
1236 }
1237
1238 if (dc->sb.block_size < c->cache->sb.block_size) {
1239 /* Will die */
1240 pr_err("Couldn't attach %s: block size less than set's block size\n",
1241 dc->backing_dev_name);
1242 return -EINVAL;
1243 }
1244
1245 /* Check whether already attached */
1246 list_for_each_entry_safe(exist_dc, t, &c->cached_devs, list) {
1247 if (!memcmp(dc->sb.uuid, exist_dc->sb.uuid, 16)) {
1248 pr_err("Tried to attach %s but duplicate UUID already attached\n",
1249 dc->backing_dev_name);
1250
1251 return -EINVAL;
1252 }
1253 }
1254
1255 u = uuid_find(c, dc->sb.uuid);
1256
1257 if (u &&
1258 (BDEV_STATE(&dc->sb) == BDEV_STATE_STALE ||
1259 BDEV_STATE(&dc->sb) == BDEV_STATE_NONE)) {
1260 memcpy(u->uuid, invalid_uuid, 16);
1261 u->invalidated = cpu_to_le32((u32)ktime_get_real_seconds());
1262 u = NULL;
1263 }
1264
1265 if (!u) {
1266 if (BDEV_STATE(&dc->sb) == BDEV_STATE_DIRTY) {
1267 pr_err("Couldn't find uuid for %s in set\n",
1268 dc->backing_dev_name);
1269 return -ENOENT;
1270 }
1271
1272 u = uuid_find_empty(c);
1273 if (!u) {
1274 pr_err("Not caching %s, no room for UUID\n",
1275 dc->backing_dev_name);
1276 return -EINVAL;
1277 }
1278 }
1279
1280 /*
1281 * Deadlocks since we're called via sysfs...
1282 * sysfs_remove_file(&dc->kobj, &sysfs_attach);
1283 */
1284
1285 if (bch_is_zero(u->uuid, 16)) {
1286 struct closure cl;
1287
1288 closure_init_stack(&cl);
1289
1290 memcpy(u->uuid, dc->sb.uuid, 16);
1291 memcpy(u->label, dc->sb.label, SB_LABEL_SIZE);
1292 u->first_reg = u->last_reg = rtime;
1293 bch_uuid_write(c);
1294
1295 memcpy(dc->sb.set_uuid, c->set_uuid, 16);
1296 SET_BDEV_STATE(&dc->sb, BDEV_STATE_CLEAN);
1297
1298 bch_write_bdev_super(dc, &cl);
1299 closure_sync(&cl);
1300 } else {
1301 u->last_reg = rtime;
1302 bch_uuid_write(c);
1303 }
1304
1305 bcache_device_attach(&dc->disk, c, u - c->uuids);
1306 list_move(&dc->list, &c->cached_devs);
1307 calc_cached_dev_sectors(c);
1308
1309 /*
1310 * dc->c must be set before dc->count != 0 - paired with the mb in
1311 * cached_dev_get()
1312 */
1313 smp_wmb();
1314 refcount_set(&dc->count, 1);
1315
1316 /* Block writeback thread, but spawn it */
1317 down_write(&dc->writeback_lock);
1318 if (bch_cached_dev_writeback_start(dc)) {
1319 up_write(&dc->writeback_lock);
1320 pr_err("Couldn't start writeback facilities for %s\n",
1321 dc->disk.disk->disk_name);
1322 return -ENOMEM;
1323 }
1324
1325 if (BDEV_STATE(&dc->sb) == BDEV_STATE_DIRTY) {
1326 atomic_set(&dc->has_dirty, 1);
1327 bch_writeback_queue(dc);
1328 }
1329
1330 bch_sectors_dirty_init(&dc->disk);
1331
1332 ret = bch_cached_dev_run(dc);
1333 if (ret && (ret != -EBUSY)) {
1334 up_write(&dc->writeback_lock);
1335 /*
1336 * bch_register_lock is held, bcache_device_stop() is not
1337 * able to be directly called. The kthread and kworker
1338 * created previously in bch_cached_dev_writeback_start()
1339 * have to be stopped manually here.
1340 */
1341 kthread_stop(dc->writeback_thread);
1342 cancel_writeback_rate_update_dwork(dc);
1343 pr_err("Couldn't run cached device %s\n",
1344 dc->backing_dev_name);
1345 return ret;
1346 }
1347
1348 bcache_device_link(&dc->disk, c, "bdev");
1349 atomic_inc(&c->attached_dev_nr);
1350
1351 if (bch_has_feature_obso_large_bucket(&(c->cache->sb))) {
1352 pr_err("The obsoleted large bucket layout is unsupported, set the bcache device into read-only\n");
1353 pr_err("Please update to the latest bcache-tools to create the cache device\n");
1354 set_disk_ro(dc->disk.disk, 1);
1355 }
1356
1357 /* Allow the writeback thread to proceed */
1358 up_write(&dc->writeback_lock);
1359
1360 pr_info("Caching %s as %s on set %pU\n",
1361 dc->backing_dev_name,
1362 dc->disk.disk->disk_name,
1363 dc->disk.c->set_uuid);
1364 return 0;
1365 }
1366
1367 /* when dc->disk.kobj released */
bch_cached_dev_release(struct kobject * kobj)1368 void bch_cached_dev_release(struct kobject *kobj)
1369 {
1370 struct cached_dev *dc = container_of(kobj, struct cached_dev,
1371 disk.kobj);
1372 kfree(dc);
1373 module_put(THIS_MODULE);
1374 }
1375
cached_dev_free(struct closure * cl)1376 static void cached_dev_free(struct closure *cl)
1377 {
1378 struct cached_dev *dc = container_of(cl, struct cached_dev, disk.cl);
1379
1380 if (test_and_clear_bit(BCACHE_DEV_WB_RUNNING, &dc->disk.flags))
1381 cancel_writeback_rate_update_dwork(dc);
1382
1383 if (!IS_ERR_OR_NULL(dc->writeback_thread))
1384 kthread_stop(dc->writeback_thread);
1385 if (!IS_ERR_OR_NULL(dc->status_update_thread))
1386 kthread_stop(dc->status_update_thread);
1387
1388 mutex_lock(&bch_register_lock);
1389
1390 if (atomic_read(&dc->running))
1391 bd_unlink_disk_holder(dc->bdev, dc->disk.disk);
1392 bcache_device_free(&dc->disk);
1393 list_del(&dc->list);
1394
1395 mutex_unlock(&bch_register_lock);
1396
1397 if (dc->sb_disk)
1398 put_page(virt_to_page(dc->sb_disk));
1399
1400 if (!IS_ERR_OR_NULL(dc->bdev))
1401 blkdev_put(dc->bdev, FMODE_READ|FMODE_WRITE|FMODE_EXCL);
1402
1403 wake_up(&unregister_wait);
1404
1405 kobject_put(&dc->disk.kobj);
1406 }
1407
cached_dev_flush(struct closure * cl)1408 static void cached_dev_flush(struct closure *cl)
1409 {
1410 struct cached_dev *dc = container_of(cl, struct cached_dev, disk.cl);
1411 struct bcache_device *d = &dc->disk;
1412
1413 mutex_lock(&bch_register_lock);
1414 bcache_device_unlink(d);
1415 mutex_unlock(&bch_register_lock);
1416
1417 bch_cache_accounting_destroy(&dc->accounting);
1418 kobject_del(&d->kobj);
1419
1420 continue_at(cl, cached_dev_free, system_wq);
1421 }
1422
cached_dev_init(struct cached_dev * dc,unsigned int block_size)1423 static int cached_dev_init(struct cached_dev *dc, unsigned int block_size)
1424 {
1425 int ret;
1426 struct io *io;
1427 struct request_queue *q = bdev_get_queue(dc->bdev);
1428
1429 __module_get(THIS_MODULE);
1430 INIT_LIST_HEAD(&dc->list);
1431 closure_init(&dc->disk.cl, NULL);
1432 set_closure_fn(&dc->disk.cl, cached_dev_flush, system_wq);
1433 kobject_init(&dc->disk.kobj, &bch_cached_dev_ktype);
1434 INIT_WORK(&dc->detach, cached_dev_detach_finish);
1435 sema_init(&dc->sb_write_mutex, 1);
1436 INIT_LIST_HEAD(&dc->io_lru);
1437 spin_lock_init(&dc->io_lock);
1438 bch_cache_accounting_init(&dc->accounting, &dc->disk.cl);
1439
1440 dc->sequential_cutoff = 4 << 20;
1441
1442 for (io = dc->io; io < dc->io + RECENT_IO; io++) {
1443 list_add(&io->lru, &dc->io_lru);
1444 hlist_add_head(&io->hash, dc->io_hash + RECENT_IO);
1445 }
1446
1447 dc->disk.stripe_size = q->limits.io_opt >> 9;
1448
1449 if (dc->disk.stripe_size)
1450 dc->partial_stripes_expensive =
1451 q->limits.raid_partial_stripes_expensive;
1452
1453 ret = bcache_device_init(&dc->disk, block_size,
1454 dc->bdev->bd_part->nr_sects - dc->sb.data_offset,
1455 dc->bdev, &bcache_cached_ops);
1456 if (ret)
1457 return ret;
1458
1459 blk_queue_io_opt(dc->disk.disk->queue,
1460 max(queue_io_opt(dc->disk.disk->queue), queue_io_opt(q)));
1461
1462 atomic_set(&dc->io_errors, 0);
1463 dc->io_disable = false;
1464 dc->error_limit = DEFAULT_CACHED_DEV_ERROR_LIMIT;
1465 /* default to auto */
1466 dc->stop_when_cache_set_failed = BCH_CACHED_DEV_STOP_AUTO;
1467
1468 bch_cached_dev_request_init(dc);
1469 bch_cached_dev_writeback_init(dc);
1470 return 0;
1471 }
1472
1473 /* Cached device - bcache superblock */
1474
register_bdev(struct cache_sb * sb,struct cache_sb_disk * sb_disk,struct block_device * bdev,struct cached_dev * dc)1475 static int register_bdev(struct cache_sb *sb, struct cache_sb_disk *sb_disk,
1476 struct block_device *bdev,
1477 struct cached_dev *dc)
1478 {
1479 const char *err = "cannot allocate memory";
1480 struct cache_set *c;
1481 int ret = -ENOMEM;
1482
1483 bdevname(bdev, dc->backing_dev_name);
1484 memcpy(&dc->sb, sb, sizeof(struct cache_sb));
1485 dc->bdev = bdev;
1486 dc->bdev->bd_holder = dc;
1487 dc->sb_disk = sb_disk;
1488
1489 if (cached_dev_init(dc, sb->block_size << 9))
1490 goto err;
1491
1492 err = "error creating kobject";
1493 if (kobject_add(&dc->disk.kobj, &part_to_dev(bdev->bd_part)->kobj,
1494 "bcache"))
1495 goto err;
1496 if (bch_cache_accounting_add_kobjs(&dc->accounting, &dc->disk.kobj))
1497 goto err;
1498
1499 pr_info("registered backing device %s\n", dc->backing_dev_name);
1500
1501 list_add(&dc->list, &uncached_devices);
1502 /* attach to a matched cache set if it exists */
1503 list_for_each_entry(c, &bch_cache_sets, list)
1504 bch_cached_dev_attach(dc, c, NULL);
1505
1506 if (BDEV_STATE(&dc->sb) == BDEV_STATE_NONE ||
1507 BDEV_STATE(&dc->sb) == BDEV_STATE_STALE) {
1508 err = "failed to run cached device";
1509 ret = bch_cached_dev_run(dc);
1510 if (ret)
1511 goto err;
1512 }
1513
1514 return 0;
1515 err:
1516 pr_notice("error %s: %s\n", dc->backing_dev_name, err);
1517 bcache_device_stop(&dc->disk);
1518 return ret;
1519 }
1520
1521 /* Flash only volumes */
1522
1523 /* When d->kobj released */
bch_flash_dev_release(struct kobject * kobj)1524 void bch_flash_dev_release(struct kobject *kobj)
1525 {
1526 struct bcache_device *d = container_of(kobj, struct bcache_device,
1527 kobj);
1528 kfree(d);
1529 }
1530
flash_dev_free(struct closure * cl)1531 static void flash_dev_free(struct closure *cl)
1532 {
1533 struct bcache_device *d = container_of(cl, struct bcache_device, cl);
1534
1535 mutex_lock(&bch_register_lock);
1536 atomic_long_sub(bcache_dev_sectors_dirty(d),
1537 &d->c->flash_dev_dirty_sectors);
1538 bcache_device_free(d);
1539 mutex_unlock(&bch_register_lock);
1540 kobject_put(&d->kobj);
1541 }
1542
flash_dev_flush(struct closure * cl)1543 static void flash_dev_flush(struct closure *cl)
1544 {
1545 struct bcache_device *d = container_of(cl, struct bcache_device, cl);
1546
1547 mutex_lock(&bch_register_lock);
1548 bcache_device_unlink(d);
1549 mutex_unlock(&bch_register_lock);
1550 kobject_del(&d->kobj);
1551 continue_at(cl, flash_dev_free, system_wq);
1552 }
1553
flash_dev_run(struct cache_set * c,struct uuid_entry * u)1554 static int flash_dev_run(struct cache_set *c, struct uuid_entry *u)
1555 {
1556 struct bcache_device *d = kzalloc(sizeof(struct bcache_device),
1557 GFP_KERNEL);
1558 if (!d)
1559 return -ENOMEM;
1560
1561 closure_init(&d->cl, NULL);
1562 set_closure_fn(&d->cl, flash_dev_flush, system_wq);
1563
1564 kobject_init(&d->kobj, &bch_flash_dev_ktype);
1565
1566 if (bcache_device_init(d, block_bytes(c->cache), u->sectors,
1567 NULL, &bcache_flash_ops))
1568 goto err;
1569
1570 bcache_device_attach(d, c, u - c->uuids);
1571 bch_sectors_dirty_init(d);
1572 bch_flash_dev_request_init(d);
1573 add_disk(d->disk);
1574
1575 if (kobject_add(&d->kobj, &disk_to_dev(d->disk)->kobj, "bcache"))
1576 goto err;
1577
1578 bcache_device_link(d, c, "volume");
1579
1580 if (bch_has_feature_obso_large_bucket(&c->cache->sb)) {
1581 pr_err("The obsoleted large bucket layout is unsupported, set the bcache device into read-only\n");
1582 pr_err("Please update to the latest bcache-tools to create the cache device\n");
1583 set_disk_ro(d->disk, 1);
1584 }
1585
1586 return 0;
1587 err:
1588 kobject_put(&d->kobj);
1589 return -ENOMEM;
1590 }
1591
flash_devs_run(struct cache_set * c)1592 static int flash_devs_run(struct cache_set *c)
1593 {
1594 int ret = 0;
1595 struct uuid_entry *u;
1596
1597 for (u = c->uuids;
1598 u < c->uuids + c->nr_uuids && !ret;
1599 u++)
1600 if (UUID_FLASH_ONLY(u))
1601 ret = flash_dev_run(c, u);
1602
1603 return ret;
1604 }
1605
bch_flash_dev_create(struct cache_set * c,uint64_t size)1606 int bch_flash_dev_create(struct cache_set *c, uint64_t size)
1607 {
1608 struct uuid_entry *u;
1609
1610 if (test_bit(CACHE_SET_STOPPING, &c->flags))
1611 return -EINTR;
1612
1613 if (!test_bit(CACHE_SET_RUNNING, &c->flags))
1614 return -EPERM;
1615
1616 u = uuid_find_empty(c);
1617 if (!u) {
1618 pr_err("Can't create volume, no room for UUID\n");
1619 return -EINVAL;
1620 }
1621
1622 get_random_bytes(u->uuid, 16);
1623 memset(u->label, 0, 32);
1624 u->first_reg = u->last_reg = cpu_to_le32((u32)ktime_get_real_seconds());
1625
1626 SET_UUID_FLASH_ONLY(u, 1);
1627 u->sectors = size >> 9;
1628
1629 bch_uuid_write(c);
1630
1631 return flash_dev_run(c, u);
1632 }
1633
bch_cached_dev_error(struct cached_dev * dc)1634 bool bch_cached_dev_error(struct cached_dev *dc)
1635 {
1636 if (!dc || test_bit(BCACHE_DEV_CLOSING, &dc->disk.flags))
1637 return false;
1638
1639 dc->io_disable = true;
1640 /* make others know io_disable is true earlier */
1641 smp_mb();
1642
1643 pr_err("stop %s: too many IO errors on backing device %s\n",
1644 dc->disk.disk->disk_name, dc->backing_dev_name);
1645
1646 bcache_device_stop(&dc->disk);
1647 return true;
1648 }
1649
1650 /* Cache set */
1651
1652 __printf(2, 3)
bch_cache_set_error(struct cache_set * c,const char * fmt,...)1653 bool bch_cache_set_error(struct cache_set *c, const char *fmt, ...)
1654 {
1655 struct va_format vaf;
1656 va_list args;
1657
1658 if (c->on_error != ON_ERROR_PANIC &&
1659 test_bit(CACHE_SET_STOPPING, &c->flags))
1660 return false;
1661
1662 if (test_and_set_bit(CACHE_SET_IO_DISABLE, &c->flags))
1663 pr_info("CACHE_SET_IO_DISABLE already set\n");
1664
1665 /*
1666 * XXX: we can be called from atomic context
1667 * acquire_console_sem();
1668 */
1669
1670 va_start(args, fmt);
1671
1672 vaf.fmt = fmt;
1673 vaf.va = &args;
1674
1675 pr_err("error on %pU: %pV, disabling caching\n",
1676 c->set_uuid, &vaf);
1677
1678 va_end(args);
1679
1680 if (c->on_error == ON_ERROR_PANIC)
1681 panic("panic forced after error\n");
1682
1683 bch_cache_set_unregister(c);
1684 return true;
1685 }
1686
1687 /* When c->kobj released */
bch_cache_set_release(struct kobject * kobj)1688 void bch_cache_set_release(struct kobject *kobj)
1689 {
1690 struct cache_set *c = container_of(kobj, struct cache_set, kobj);
1691
1692 kfree(c);
1693 module_put(THIS_MODULE);
1694 }
1695
cache_set_free(struct closure * cl)1696 static void cache_set_free(struct closure *cl)
1697 {
1698 struct cache_set *c = container_of(cl, struct cache_set, cl);
1699 struct cache *ca;
1700
1701 debugfs_remove(c->debug);
1702
1703 bch_open_buckets_free(c);
1704 bch_btree_cache_free(c);
1705 bch_journal_free(c);
1706
1707 mutex_lock(&bch_register_lock);
1708 bch_bset_sort_state_free(&c->sort);
1709 free_pages((unsigned long) c->uuids, ilog2(meta_bucket_pages(&c->cache->sb)));
1710
1711 ca = c->cache;
1712 if (ca) {
1713 ca->set = NULL;
1714 c->cache = NULL;
1715 kobject_put(&ca->kobj);
1716 }
1717
1718
1719 if (c->moving_gc_wq)
1720 destroy_workqueue(c->moving_gc_wq);
1721 bioset_exit(&c->bio_split);
1722 mempool_exit(&c->fill_iter);
1723 mempool_exit(&c->bio_meta);
1724 mempool_exit(&c->search);
1725 kfree(c->devices);
1726
1727 list_del(&c->list);
1728 mutex_unlock(&bch_register_lock);
1729
1730 pr_info("Cache set %pU unregistered\n", c->set_uuid);
1731 wake_up(&unregister_wait);
1732
1733 closure_debug_destroy(&c->cl);
1734 kobject_put(&c->kobj);
1735 }
1736
cache_set_flush(struct closure * cl)1737 static void cache_set_flush(struct closure *cl)
1738 {
1739 struct cache_set *c = container_of(cl, struct cache_set, caching);
1740 struct cache *ca = c->cache;
1741 struct btree *b;
1742
1743 bch_cache_accounting_destroy(&c->accounting);
1744
1745 kobject_put(&c->internal);
1746 kobject_del(&c->kobj);
1747
1748 if (!IS_ERR_OR_NULL(c->gc_thread))
1749 kthread_stop(c->gc_thread);
1750
1751 if (!IS_ERR_OR_NULL(c->root))
1752 list_add(&c->root->list, &c->btree_cache);
1753
1754 /*
1755 * Avoid flushing cached nodes if cache set is retiring
1756 * due to too many I/O errors detected.
1757 */
1758 if (!test_bit(CACHE_SET_IO_DISABLE, &c->flags))
1759 list_for_each_entry(b, &c->btree_cache, list) {
1760 mutex_lock(&b->write_lock);
1761 if (btree_node_dirty(b))
1762 __bch_btree_node_write(b, NULL);
1763 mutex_unlock(&b->write_lock);
1764 }
1765
1766 if (ca->alloc_thread)
1767 kthread_stop(ca->alloc_thread);
1768
1769 if (c->journal.cur) {
1770 cancel_delayed_work_sync(&c->journal.work);
1771 /* flush last journal entry if needed */
1772 c->journal.work.work.func(&c->journal.work.work);
1773 }
1774
1775 closure_return(cl);
1776 }
1777
1778 /*
1779 * This function is only called when CACHE_SET_IO_DISABLE is set, which means
1780 * cache set is unregistering due to too many I/O errors. In this condition,
1781 * the bcache device might be stopped, it depends on stop_when_cache_set_failed
1782 * value and whether the broken cache has dirty data:
1783 *
1784 * dc->stop_when_cache_set_failed dc->has_dirty stop bcache device
1785 * BCH_CACHED_STOP_AUTO 0 NO
1786 * BCH_CACHED_STOP_AUTO 1 YES
1787 * BCH_CACHED_DEV_STOP_ALWAYS 0 YES
1788 * BCH_CACHED_DEV_STOP_ALWAYS 1 YES
1789 *
1790 * The expected behavior is, if stop_when_cache_set_failed is configured to
1791 * "auto" via sysfs interface, the bcache device will not be stopped if the
1792 * backing device is clean on the broken cache device.
1793 */
conditional_stop_bcache_device(struct cache_set * c,struct bcache_device * d,struct cached_dev * dc)1794 static void conditional_stop_bcache_device(struct cache_set *c,
1795 struct bcache_device *d,
1796 struct cached_dev *dc)
1797 {
1798 if (dc->stop_when_cache_set_failed == BCH_CACHED_DEV_STOP_ALWAYS) {
1799 pr_warn("stop_when_cache_set_failed of %s is \"always\", stop it for failed cache set %pU.\n",
1800 d->disk->disk_name, c->set_uuid);
1801 bcache_device_stop(d);
1802 } else if (atomic_read(&dc->has_dirty)) {
1803 /*
1804 * dc->stop_when_cache_set_failed == BCH_CACHED_STOP_AUTO
1805 * and dc->has_dirty == 1
1806 */
1807 pr_warn("stop_when_cache_set_failed of %s is \"auto\" and cache is dirty, stop it to avoid potential data corruption.\n",
1808 d->disk->disk_name);
1809 /*
1810 * There might be a small time gap that cache set is
1811 * released but bcache device is not. Inside this time
1812 * gap, regular I/O requests will directly go into
1813 * backing device as no cache set attached to. This
1814 * behavior may also introduce potential inconsistence
1815 * data in writeback mode while cache is dirty.
1816 * Therefore before calling bcache_device_stop() due
1817 * to a broken cache device, dc->io_disable should be
1818 * explicitly set to true.
1819 */
1820 dc->io_disable = true;
1821 /* make others know io_disable is true earlier */
1822 smp_mb();
1823 bcache_device_stop(d);
1824 } else {
1825 /*
1826 * dc->stop_when_cache_set_failed == BCH_CACHED_STOP_AUTO
1827 * and dc->has_dirty == 0
1828 */
1829 pr_warn("stop_when_cache_set_failed of %s is \"auto\" and cache is clean, keep it alive.\n",
1830 d->disk->disk_name);
1831 }
1832 }
1833
__cache_set_unregister(struct closure * cl)1834 static void __cache_set_unregister(struct closure *cl)
1835 {
1836 struct cache_set *c = container_of(cl, struct cache_set, caching);
1837 struct cached_dev *dc;
1838 struct bcache_device *d;
1839 size_t i;
1840
1841 mutex_lock(&bch_register_lock);
1842
1843 for (i = 0; i < c->devices_max_used; i++) {
1844 d = c->devices[i];
1845 if (!d)
1846 continue;
1847
1848 if (!UUID_FLASH_ONLY(&c->uuids[i]) &&
1849 test_bit(CACHE_SET_UNREGISTERING, &c->flags)) {
1850 dc = container_of(d, struct cached_dev, disk);
1851 bch_cached_dev_detach(dc);
1852 if (test_bit(CACHE_SET_IO_DISABLE, &c->flags))
1853 conditional_stop_bcache_device(c, d, dc);
1854 } else {
1855 bcache_device_stop(d);
1856 }
1857 }
1858
1859 mutex_unlock(&bch_register_lock);
1860
1861 continue_at(cl, cache_set_flush, system_wq);
1862 }
1863
bch_cache_set_stop(struct cache_set * c)1864 void bch_cache_set_stop(struct cache_set *c)
1865 {
1866 if (!test_and_set_bit(CACHE_SET_STOPPING, &c->flags))
1867 /* closure_fn set to __cache_set_unregister() */
1868 closure_queue(&c->caching);
1869 }
1870
bch_cache_set_unregister(struct cache_set * c)1871 void bch_cache_set_unregister(struct cache_set *c)
1872 {
1873 set_bit(CACHE_SET_UNREGISTERING, &c->flags);
1874 bch_cache_set_stop(c);
1875 }
1876
1877 #define alloc_meta_bucket_pages(gfp, sb) \
1878 ((void *) __get_free_pages(__GFP_ZERO|__GFP_COMP|gfp, ilog2(meta_bucket_pages(sb))))
1879
bch_cache_set_alloc(struct cache_sb * sb)1880 struct cache_set *bch_cache_set_alloc(struct cache_sb *sb)
1881 {
1882 int iter_size;
1883 struct cache *ca = container_of(sb, struct cache, sb);
1884 struct cache_set *c = kzalloc(sizeof(struct cache_set), GFP_KERNEL);
1885
1886 if (!c)
1887 return NULL;
1888
1889 __module_get(THIS_MODULE);
1890 closure_init(&c->cl, NULL);
1891 set_closure_fn(&c->cl, cache_set_free, system_wq);
1892
1893 closure_init(&c->caching, &c->cl);
1894 set_closure_fn(&c->caching, __cache_set_unregister, system_wq);
1895
1896 /* Maybe create continue_at_noreturn() and use it here? */
1897 closure_set_stopped(&c->cl);
1898 closure_put(&c->cl);
1899
1900 kobject_init(&c->kobj, &bch_cache_set_ktype);
1901 kobject_init(&c->internal, &bch_cache_set_internal_ktype);
1902
1903 bch_cache_accounting_init(&c->accounting, &c->cl);
1904
1905 memcpy(c->set_uuid, sb->set_uuid, 16);
1906
1907 c->cache = ca;
1908 c->cache->set = c;
1909 c->bucket_bits = ilog2(sb->bucket_size);
1910 c->block_bits = ilog2(sb->block_size);
1911 c->nr_uuids = meta_bucket_bytes(sb) / sizeof(struct uuid_entry);
1912 c->devices_max_used = 0;
1913 atomic_set(&c->attached_dev_nr, 0);
1914 c->btree_pages = meta_bucket_pages(sb);
1915 if (c->btree_pages > BTREE_MAX_PAGES)
1916 c->btree_pages = max_t(int, c->btree_pages / 4,
1917 BTREE_MAX_PAGES);
1918
1919 sema_init(&c->sb_write_mutex, 1);
1920 mutex_init(&c->bucket_lock);
1921 init_waitqueue_head(&c->btree_cache_wait);
1922 spin_lock_init(&c->btree_cannibalize_lock);
1923 init_waitqueue_head(&c->bucket_wait);
1924 init_waitqueue_head(&c->gc_wait);
1925 sema_init(&c->uuid_write_mutex, 1);
1926
1927 spin_lock_init(&c->btree_gc_time.lock);
1928 spin_lock_init(&c->btree_split_time.lock);
1929 spin_lock_init(&c->btree_read_time.lock);
1930
1931 bch_moving_init_cache_set(c);
1932
1933 INIT_LIST_HEAD(&c->list);
1934 INIT_LIST_HEAD(&c->cached_devs);
1935 INIT_LIST_HEAD(&c->btree_cache);
1936 INIT_LIST_HEAD(&c->btree_cache_freeable);
1937 INIT_LIST_HEAD(&c->btree_cache_freed);
1938 INIT_LIST_HEAD(&c->data_buckets);
1939
1940 iter_size = ((meta_bucket_pages(sb) * PAGE_SECTORS) / sb->block_size + 1) *
1941 sizeof(struct btree_iter_set);
1942
1943 c->devices = kcalloc(c->nr_uuids, sizeof(void *), GFP_KERNEL);
1944 if (!c->devices)
1945 goto err;
1946
1947 if (mempool_init_slab_pool(&c->search, 32, bch_search_cache))
1948 goto err;
1949
1950 if (mempool_init_kmalloc_pool(&c->bio_meta, 2,
1951 sizeof(struct bbio) +
1952 sizeof(struct bio_vec) * meta_bucket_pages(sb)))
1953 goto err;
1954
1955 if (mempool_init_kmalloc_pool(&c->fill_iter, 1, iter_size))
1956 goto err;
1957
1958 if (bioset_init(&c->bio_split, 4, offsetof(struct bbio, bio),
1959 BIOSET_NEED_BVECS|BIOSET_NEED_RESCUER))
1960 goto err;
1961
1962 c->uuids = alloc_meta_bucket_pages(GFP_KERNEL, sb);
1963 if (!c->uuids)
1964 goto err;
1965
1966 c->moving_gc_wq = alloc_workqueue("bcache_gc", WQ_MEM_RECLAIM, 0);
1967 if (!c->moving_gc_wq)
1968 goto err;
1969
1970 if (bch_journal_alloc(c))
1971 goto err;
1972
1973 if (bch_btree_cache_alloc(c))
1974 goto err;
1975
1976 if (bch_open_buckets_alloc(c))
1977 goto err;
1978
1979 if (bch_bset_sort_state_init(&c->sort, ilog2(c->btree_pages)))
1980 goto err;
1981
1982 c->congested_read_threshold_us = 2000;
1983 c->congested_write_threshold_us = 20000;
1984 c->error_limit = DEFAULT_IO_ERROR_LIMIT;
1985 c->idle_max_writeback_rate_enabled = 1;
1986 WARN_ON(test_and_clear_bit(CACHE_SET_IO_DISABLE, &c->flags));
1987
1988 return c;
1989 err:
1990 bch_cache_set_unregister(c);
1991 return NULL;
1992 }
1993
run_cache_set(struct cache_set * c)1994 static int run_cache_set(struct cache_set *c)
1995 {
1996 const char *err = "cannot allocate memory";
1997 struct cached_dev *dc, *t;
1998 struct cache *ca = c->cache;
1999 struct closure cl;
2000 LIST_HEAD(journal);
2001 struct journal_replay *l;
2002
2003 closure_init_stack(&cl);
2004
2005 c->nbuckets = ca->sb.nbuckets;
2006 set_gc_sectors(c);
2007
2008 if (CACHE_SYNC(&c->cache->sb)) {
2009 struct bkey *k;
2010 struct jset *j;
2011
2012 err = "cannot allocate memory for journal";
2013 if (bch_journal_read(c, &journal))
2014 goto err;
2015
2016 pr_debug("btree_journal_read() done\n");
2017
2018 err = "no journal entries found";
2019 if (list_empty(&journal))
2020 goto err;
2021
2022 j = &list_entry(journal.prev, struct journal_replay, list)->j;
2023
2024 err = "IO error reading priorities";
2025 if (prio_read(ca, j->prio_bucket[ca->sb.nr_this_dev]))
2026 goto err;
2027
2028 /*
2029 * If prio_read() fails it'll call cache_set_error and we'll
2030 * tear everything down right away, but if we perhaps checked
2031 * sooner we could avoid journal replay.
2032 */
2033
2034 k = &j->btree_root;
2035
2036 err = "bad btree root";
2037 if (__bch_btree_ptr_invalid(c, k))
2038 goto err;
2039
2040 err = "error reading btree root";
2041 c->root = bch_btree_node_get(c, NULL, k,
2042 j->btree_level,
2043 true, NULL);
2044 if (IS_ERR_OR_NULL(c->root))
2045 goto err;
2046
2047 list_del_init(&c->root->list);
2048 rw_unlock(true, c->root);
2049
2050 err = uuid_read(c, j, &cl);
2051 if (err)
2052 goto err;
2053
2054 err = "error in recovery";
2055 if (bch_btree_check(c))
2056 goto err;
2057
2058 bch_journal_mark(c, &journal);
2059 bch_initial_gc_finish(c);
2060 pr_debug("btree_check() done\n");
2061
2062 /*
2063 * bcache_journal_next() can't happen sooner, or
2064 * btree_gc_finish() will give spurious errors about last_gc >
2065 * gc_gen - this is a hack but oh well.
2066 */
2067 bch_journal_next(&c->journal);
2068
2069 err = "error starting allocator thread";
2070 if (bch_cache_allocator_start(ca))
2071 goto err;
2072
2073 /*
2074 * First place it's safe to allocate: btree_check() and
2075 * btree_gc_finish() have to run before we have buckets to
2076 * allocate, and bch_bucket_alloc_set() might cause a journal
2077 * entry to be written so bcache_journal_next() has to be called
2078 * first.
2079 *
2080 * If the uuids were in the old format we have to rewrite them
2081 * before the next journal entry is written:
2082 */
2083 if (j->version < BCACHE_JSET_VERSION_UUID)
2084 __uuid_write(c);
2085
2086 err = "bcache: replay journal failed";
2087 if (bch_journal_replay(c, &journal))
2088 goto err;
2089 } else {
2090 unsigned int j;
2091
2092 pr_notice("invalidating existing data\n");
2093 ca->sb.keys = clamp_t(int, ca->sb.nbuckets >> 7,
2094 2, SB_JOURNAL_BUCKETS);
2095
2096 for (j = 0; j < ca->sb.keys; j++)
2097 ca->sb.d[j] = ca->sb.first_bucket + j;
2098
2099 bch_initial_gc_finish(c);
2100
2101 err = "error starting allocator thread";
2102 if (bch_cache_allocator_start(ca))
2103 goto err;
2104
2105 mutex_lock(&c->bucket_lock);
2106 bch_prio_write(ca, true);
2107 mutex_unlock(&c->bucket_lock);
2108
2109 err = "cannot allocate new UUID bucket";
2110 if (__uuid_write(c))
2111 goto err;
2112
2113 err = "cannot allocate new btree root";
2114 c->root = __bch_btree_node_alloc(c, NULL, 0, true, NULL);
2115 if (IS_ERR_OR_NULL(c->root))
2116 goto err;
2117
2118 mutex_lock(&c->root->write_lock);
2119 bkey_copy_key(&c->root->key, &MAX_KEY);
2120 bch_btree_node_write(c->root, &cl);
2121 mutex_unlock(&c->root->write_lock);
2122
2123 bch_btree_set_root(c->root);
2124 rw_unlock(true, c->root);
2125
2126 /*
2127 * We don't want to write the first journal entry until
2128 * everything is set up - fortunately journal entries won't be
2129 * written until the SET_CACHE_SYNC() here:
2130 */
2131 SET_CACHE_SYNC(&c->cache->sb, true);
2132
2133 bch_journal_next(&c->journal);
2134 bch_journal_meta(c, &cl);
2135 }
2136
2137 err = "error starting gc thread";
2138 if (bch_gc_thread_start(c))
2139 goto err;
2140
2141 closure_sync(&cl);
2142 c->cache->sb.last_mount = (u32)ktime_get_real_seconds();
2143 bcache_write_super(c);
2144
2145 if (bch_has_feature_obso_large_bucket(&c->cache->sb))
2146 pr_err("Detect obsoleted large bucket layout, all attached bcache device will be read-only\n");
2147
2148 list_for_each_entry_safe(dc, t, &uncached_devices, list)
2149 bch_cached_dev_attach(dc, c, NULL);
2150
2151 flash_devs_run(c);
2152
2153 set_bit(CACHE_SET_RUNNING, &c->flags);
2154 return 0;
2155 err:
2156 while (!list_empty(&journal)) {
2157 l = list_first_entry(&journal, struct journal_replay, list);
2158 list_del(&l->list);
2159 kfree(l);
2160 }
2161
2162 closure_sync(&cl);
2163
2164 bch_cache_set_error(c, "%s", err);
2165
2166 return -EIO;
2167 }
2168
register_cache_set(struct cache * ca)2169 static const char *register_cache_set(struct cache *ca)
2170 {
2171 char buf[12];
2172 const char *err = "cannot allocate memory";
2173 struct cache_set *c;
2174
2175 list_for_each_entry(c, &bch_cache_sets, list)
2176 if (!memcmp(c->set_uuid, ca->sb.set_uuid, 16)) {
2177 if (c->cache)
2178 return "duplicate cache set member";
2179
2180 goto found;
2181 }
2182
2183 c = bch_cache_set_alloc(&ca->sb);
2184 if (!c)
2185 return err;
2186
2187 err = "error creating kobject";
2188 if (kobject_add(&c->kobj, bcache_kobj, "%pU", c->set_uuid) ||
2189 kobject_add(&c->internal, &c->kobj, "internal"))
2190 goto err;
2191
2192 if (bch_cache_accounting_add_kobjs(&c->accounting, &c->kobj))
2193 goto err;
2194
2195 bch_debug_init_cache_set(c);
2196
2197 list_add(&c->list, &bch_cache_sets);
2198 found:
2199 sprintf(buf, "cache%i", ca->sb.nr_this_dev);
2200 if (sysfs_create_link(&ca->kobj, &c->kobj, "set") ||
2201 sysfs_create_link(&c->kobj, &ca->kobj, buf))
2202 goto err;
2203
2204 kobject_get(&ca->kobj);
2205 ca->set = c;
2206 ca->set->cache = ca;
2207
2208 err = "failed to run cache set";
2209 if (run_cache_set(c) < 0)
2210 goto err;
2211
2212 return NULL;
2213 err:
2214 bch_cache_set_unregister(c);
2215 return err;
2216 }
2217
2218 /* Cache device */
2219
2220 /* When ca->kobj released */
bch_cache_release(struct kobject * kobj)2221 void bch_cache_release(struct kobject *kobj)
2222 {
2223 struct cache *ca = container_of(kobj, struct cache, kobj);
2224 unsigned int i;
2225
2226 if (ca->set) {
2227 BUG_ON(ca->set->cache != ca);
2228 ca->set->cache = NULL;
2229 }
2230
2231 free_pages((unsigned long) ca->disk_buckets, ilog2(meta_bucket_pages(&ca->sb)));
2232 kfree(ca->prio_buckets);
2233 vfree(ca->buckets);
2234
2235 free_heap(&ca->heap);
2236 free_fifo(&ca->free_inc);
2237
2238 for (i = 0; i < RESERVE_NR; i++)
2239 free_fifo(&ca->free[i]);
2240
2241 if (ca->sb_disk)
2242 put_page(virt_to_page(ca->sb_disk));
2243
2244 if (!IS_ERR_OR_NULL(ca->bdev))
2245 blkdev_put(ca->bdev, FMODE_READ|FMODE_WRITE|FMODE_EXCL);
2246
2247 kfree(ca);
2248 module_put(THIS_MODULE);
2249 }
2250
cache_alloc(struct cache * ca)2251 static int cache_alloc(struct cache *ca)
2252 {
2253 size_t free;
2254 size_t btree_buckets;
2255 struct bucket *b;
2256 int ret = -ENOMEM;
2257 const char *err = NULL;
2258
2259 __module_get(THIS_MODULE);
2260 kobject_init(&ca->kobj, &bch_cache_ktype);
2261
2262 bio_init(&ca->journal.bio, ca->journal.bio.bi_inline_vecs, 8);
2263
2264 /*
2265 * when ca->sb.njournal_buckets is not zero, journal exists,
2266 * and in bch_journal_replay(), tree node may split,
2267 * so bucket of RESERVE_BTREE type is needed,
2268 * the worst situation is all journal buckets are valid journal,
2269 * and all the keys need to replay,
2270 * so the number of RESERVE_BTREE type buckets should be as much
2271 * as journal buckets
2272 */
2273 btree_buckets = ca->sb.njournal_buckets ?: 8;
2274 free = roundup_pow_of_two(ca->sb.nbuckets) >> 10;
2275 if (!free) {
2276 ret = -EPERM;
2277 err = "ca->sb.nbuckets is too small";
2278 goto err_free;
2279 }
2280
2281 if (!init_fifo(&ca->free[RESERVE_BTREE], btree_buckets,
2282 GFP_KERNEL)) {
2283 err = "ca->free[RESERVE_BTREE] alloc failed";
2284 goto err_btree_alloc;
2285 }
2286
2287 if (!init_fifo_exact(&ca->free[RESERVE_PRIO], prio_buckets(ca),
2288 GFP_KERNEL)) {
2289 err = "ca->free[RESERVE_PRIO] alloc failed";
2290 goto err_prio_alloc;
2291 }
2292
2293 if (!init_fifo(&ca->free[RESERVE_MOVINGGC], free, GFP_KERNEL)) {
2294 err = "ca->free[RESERVE_MOVINGGC] alloc failed";
2295 goto err_movinggc_alloc;
2296 }
2297
2298 if (!init_fifo(&ca->free[RESERVE_NONE], free, GFP_KERNEL)) {
2299 err = "ca->free[RESERVE_NONE] alloc failed";
2300 goto err_none_alloc;
2301 }
2302
2303 if (!init_fifo(&ca->free_inc, free << 2, GFP_KERNEL)) {
2304 err = "ca->free_inc alloc failed";
2305 goto err_free_inc_alloc;
2306 }
2307
2308 if (!init_heap(&ca->heap, free << 3, GFP_KERNEL)) {
2309 err = "ca->heap alloc failed";
2310 goto err_heap_alloc;
2311 }
2312
2313 ca->buckets = vzalloc(array_size(sizeof(struct bucket),
2314 ca->sb.nbuckets));
2315 if (!ca->buckets) {
2316 err = "ca->buckets alloc failed";
2317 goto err_buckets_alloc;
2318 }
2319
2320 ca->prio_buckets = kzalloc(array3_size(sizeof(uint64_t),
2321 prio_buckets(ca), 2),
2322 GFP_KERNEL);
2323 if (!ca->prio_buckets) {
2324 err = "ca->prio_buckets alloc failed";
2325 goto err_prio_buckets_alloc;
2326 }
2327
2328 ca->disk_buckets = alloc_meta_bucket_pages(GFP_KERNEL, &ca->sb);
2329 if (!ca->disk_buckets) {
2330 err = "ca->disk_buckets alloc failed";
2331 goto err_disk_buckets_alloc;
2332 }
2333
2334 ca->prio_last_buckets = ca->prio_buckets + prio_buckets(ca);
2335
2336 for_each_bucket(b, ca)
2337 atomic_set(&b->pin, 0);
2338 return 0;
2339
2340 err_disk_buckets_alloc:
2341 kfree(ca->prio_buckets);
2342 err_prio_buckets_alloc:
2343 vfree(ca->buckets);
2344 err_buckets_alloc:
2345 free_heap(&ca->heap);
2346 err_heap_alloc:
2347 free_fifo(&ca->free_inc);
2348 err_free_inc_alloc:
2349 free_fifo(&ca->free[RESERVE_NONE]);
2350 err_none_alloc:
2351 free_fifo(&ca->free[RESERVE_MOVINGGC]);
2352 err_movinggc_alloc:
2353 free_fifo(&ca->free[RESERVE_PRIO]);
2354 err_prio_alloc:
2355 free_fifo(&ca->free[RESERVE_BTREE]);
2356 err_btree_alloc:
2357 err_free:
2358 module_put(THIS_MODULE);
2359 if (err)
2360 pr_notice("error %s: %s\n", ca->cache_dev_name, err);
2361 return ret;
2362 }
2363
register_cache(struct cache_sb * sb,struct cache_sb_disk * sb_disk,struct block_device * bdev,struct cache * ca)2364 static int register_cache(struct cache_sb *sb, struct cache_sb_disk *sb_disk,
2365 struct block_device *bdev, struct cache *ca)
2366 {
2367 const char *err = NULL; /* must be set for any error case */
2368 int ret = 0;
2369
2370 bdevname(bdev, ca->cache_dev_name);
2371 memcpy(&ca->sb, sb, sizeof(struct cache_sb));
2372 ca->bdev = bdev;
2373 ca->bdev->bd_holder = ca;
2374 ca->sb_disk = sb_disk;
2375
2376 if (blk_queue_discard(bdev_get_queue(bdev)))
2377 ca->discard = CACHE_DISCARD(&ca->sb);
2378
2379 ret = cache_alloc(ca);
2380 if (ret != 0) {
2381 /*
2382 * If we failed here, it means ca->kobj is not initialized yet,
2383 * kobject_put() won't be called and there is no chance to
2384 * call blkdev_put() to bdev in bch_cache_release(). So we
2385 * explicitly call blkdev_put() here.
2386 */
2387 blkdev_put(bdev, FMODE_READ|FMODE_WRITE|FMODE_EXCL);
2388 if (ret == -ENOMEM)
2389 err = "cache_alloc(): -ENOMEM";
2390 else if (ret == -EPERM)
2391 err = "cache_alloc(): cache device is too small";
2392 else
2393 err = "cache_alloc(): unknown error";
2394 goto err;
2395 }
2396
2397 if (kobject_add(&ca->kobj,
2398 &part_to_dev(bdev->bd_part)->kobj,
2399 "bcache")) {
2400 err = "error calling kobject_add";
2401 ret = -ENOMEM;
2402 goto out;
2403 }
2404
2405 mutex_lock(&bch_register_lock);
2406 err = register_cache_set(ca);
2407 mutex_unlock(&bch_register_lock);
2408
2409 if (err) {
2410 ret = -ENODEV;
2411 goto out;
2412 }
2413
2414 pr_info("registered cache device %s\n", ca->cache_dev_name);
2415
2416 out:
2417 kobject_put(&ca->kobj);
2418
2419 err:
2420 if (err)
2421 pr_notice("error %s: %s\n", ca->cache_dev_name, err);
2422
2423 return ret;
2424 }
2425
2426 /* Global interfaces/init */
2427
2428 static ssize_t register_bcache(struct kobject *k, struct kobj_attribute *attr,
2429 const char *buffer, size_t size);
2430 static ssize_t bch_pending_bdevs_cleanup(struct kobject *k,
2431 struct kobj_attribute *attr,
2432 const char *buffer, size_t size);
2433
2434 kobj_attribute_write(register, register_bcache);
2435 kobj_attribute_write(register_quiet, register_bcache);
2436 kobj_attribute_write(pendings_cleanup, bch_pending_bdevs_cleanup);
2437
bch_is_open_backing(struct block_device * bdev)2438 static bool bch_is_open_backing(struct block_device *bdev)
2439 {
2440 struct cache_set *c, *tc;
2441 struct cached_dev *dc, *t;
2442
2443 list_for_each_entry_safe(c, tc, &bch_cache_sets, list)
2444 list_for_each_entry_safe(dc, t, &c->cached_devs, list)
2445 if (dc->bdev == bdev)
2446 return true;
2447 list_for_each_entry_safe(dc, t, &uncached_devices, list)
2448 if (dc->bdev == bdev)
2449 return true;
2450 return false;
2451 }
2452
bch_is_open_cache(struct block_device * bdev)2453 static bool bch_is_open_cache(struct block_device *bdev)
2454 {
2455 struct cache_set *c, *tc;
2456
2457 list_for_each_entry_safe(c, tc, &bch_cache_sets, list) {
2458 struct cache *ca = c->cache;
2459
2460 if (ca->bdev == bdev)
2461 return true;
2462 }
2463
2464 return false;
2465 }
2466
bch_is_open(struct block_device * bdev)2467 static bool bch_is_open(struct block_device *bdev)
2468 {
2469 return bch_is_open_cache(bdev) || bch_is_open_backing(bdev);
2470 }
2471
2472 struct async_reg_args {
2473 struct delayed_work reg_work;
2474 char *path;
2475 struct cache_sb *sb;
2476 struct cache_sb_disk *sb_disk;
2477 struct block_device *bdev;
2478 };
2479
register_bdev_worker(struct work_struct * work)2480 static void register_bdev_worker(struct work_struct *work)
2481 {
2482 int fail = false;
2483 struct async_reg_args *args =
2484 container_of(work, struct async_reg_args, reg_work.work);
2485 struct cached_dev *dc;
2486
2487 dc = kzalloc(sizeof(*dc), GFP_KERNEL);
2488 if (!dc) {
2489 fail = true;
2490 put_page(virt_to_page(args->sb_disk));
2491 blkdev_put(args->bdev, FMODE_READ | FMODE_WRITE | FMODE_EXCL);
2492 goto out;
2493 }
2494
2495 mutex_lock(&bch_register_lock);
2496 if (register_bdev(args->sb, args->sb_disk, args->bdev, dc) < 0)
2497 fail = true;
2498 mutex_unlock(&bch_register_lock);
2499
2500 out:
2501 if (fail)
2502 pr_info("error %s: fail to register backing device\n",
2503 args->path);
2504 kfree(args->sb);
2505 kfree(args->path);
2506 kfree(args);
2507 module_put(THIS_MODULE);
2508 }
2509
register_cache_worker(struct work_struct * work)2510 static void register_cache_worker(struct work_struct *work)
2511 {
2512 int fail = false;
2513 struct async_reg_args *args =
2514 container_of(work, struct async_reg_args, reg_work.work);
2515 struct cache *ca;
2516
2517 ca = kzalloc(sizeof(*ca), GFP_KERNEL);
2518 if (!ca) {
2519 fail = true;
2520 put_page(virt_to_page(args->sb_disk));
2521 blkdev_put(args->bdev, FMODE_READ | FMODE_WRITE | FMODE_EXCL);
2522 goto out;
2523 }
2524
2525 /* blkdev_put() will be called in bch_cache_release() */
2526 if (register_cache(args->sb, args->sb_disk, args->bdev, ca) != 0)
2527 fail = true;
2528
2529 out:
2530 if (fail)
2531 pr_info("error %s: fail to register cache device\n",
2532 args->path);
2533 kfree(args->sb);
2534 kfree(args->path);
2535 kfree(args);
2536 module_put(THIS_MODULE);
2537 }
2538
register_device_aync(struct async_reg_args * args)2539 static void register_device_aync(struct async_reg_args *args)
2540 {
2541 if (SB_IS_BDEV(args->sb))
2542 INIT_DELAYED_WORK(&args->reg_work, register_bdev_worker);
2543 else
2544 INIT_DELAYED_WORK(&args->reg_work, register_cache_worker);
2545
2546 /* 10 jiffies is enough for a delay */
2547 queue_delayed_work(system_wq, &args->reg_work, 10);
2548 }
2549
register_bcache(struct kobject * k,struct kobj_attribute * attr,const char * buffer,size_t size)2550 static ssize_t register_bcache(struct kobject *k, struct kobj_attribute *attr,
2551 const char *buffer, size_t size)
2552 {
2553 const char *err;
2554 char *path = NULL;
2555 struct cache_sb *sb;
2556 struct cache_sb_disk *sb_disk;
2557 struct block_device *bdev;
2558 ssize_t ret;
2559 bool async_registration = false;
2560
2561 #ifdef CONFIG_BCACHE_ASYNC_REGISTRATION
2562 async_registration = true;
2563 #endif
2564
2565 ret = -EBUSY;
2566 err = "failed to reference bcache module";
2567 if (!try_module_get(THIS_MODULE))
2568 goto out;
2569
2570 /* For latest state of bcache_is_reboot */
2571 smp_mb();
2572 err = "bcache is in reboot";
2573 if (bcache_is_reboot)
2574 goto out_module_put;
2575
2576 ret = -ENOMEM;
2577 err = "cannot allocate memory";
2578 path = kstrndup(buffer, size, GFP_KERNEL);
2579 if (!path)
2580 goto out_module_put;
2581
2582 sb = kmalloc(sizeof(struct cache_sb), GFP_KERNEL);
2583 if (!sb)
2584 goto out_free_path;
2585
2586 ret = -EINVAL;
2587 err = "failed to open device";
2588 bdev = blkdev_get_by_path(strim(path),
2589 FMODE_READ|FMODE_WRITE|FMODE_EXCL,
2590 sb);
2591 if (IS_ERR(bdev)) {
2592 if (bdev == ERR_PTR(-EBUSY)) {
2593 bdev = lookup_bdev(strim(path));
2594 mutex_lock(&bch_register_lock);
2595 if (!IS_ERR(bdev) && bch_is_open(bdev))
2596 err = "device already registered";
2597 else
2598 err = "device busy";
2599 mutex_unlock(&bch_register_lock);
2600 if (!IS_ERR(bdev))
2601 bdput(bdev);
2602 if (attr == &ksysfs_register_quiet)
2603 goto done;
2604 }
2605 goto out_free_sb;
2606 }
2607
2608 err = "failed to set blocksize";
2609 if (set_blocksize(bdev, 4096))
2610 goto out_blkdev_put;
2611
2612 err = read_super(sb, bdev, &sb_disk);
2613 if (err)
2614 goto out_blkdev_put;
2615
2616 err = "failed to register device";
2617
2618 if (async_registration) {
2619 /* register in asynchronous way */
2620 struct async_reg_args *args =
2621 kzalloc(sizeof(struct async_reg_args), GFP_KERNEL);
2622
2623 if (!args) {
2624 ret = -ENOMEM;
2625 err = "cannot allocate memory";
2626 goto out_put_sb_page;
2627 }
2628
2629 args->path = path;
2630 args->sb = sb;
2631 args->sb_disk = sb_disk;
2632 args->bdev = bdev;
2633 register_device_aync(args);
2634 /* No wait and returns to user space */
2635 goto async_done;
2636 }
2637
2638 if (SB_IS_BDEV(sb)) {
2639 struct cached_dev *dc = kzalloc(sizeof(*dc), GFP_KERNEL);
2640
2641 if (!dc)
2642 goto out_put_sb_page;
2643
2644 mutex_lock(&bch_register_lock);
2645 ret = register_bdev(sb, sb_disk, bdev, dc);
2646 mutex_unlock(&bch_register_lock);
2647 /* blkdev_put() will be called in cached_dev_free() */
2648 if (ret < 0)
2649 goto out_free_sb;
2650 } else {
2651 struct cache *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
2652
2653 if (!ca)
2654 goto out_put_sb_page;
2655
2656 /* blkdev_put() will be called in bch_cache_release() */
2657 if (register_cache(sb, sb_disk, bdev, ca) != 0)
2658 goto out_free_sb;
2659 }
2660
2661 done:
2662 kfree(sb);
2663 kfree(path);
2664 module_put(THIS_MODULE);
2665 async_done:
2666 return size;
2667
2668 out_put_sb_page:
2669 put_page(virt_to_page(sb_disk));
2670 out_blkdev_put:
2671 blkdev_put(bdev, FMODE_READ | FMODE_WRITE | FMODE_EXCL);
2672 out_free_sb:
2673 kfree(sb);
2674 out_free_path:
2675 kfree(path);
2676 path = NULL;
2677 out_module_put:
2678 module_put(THIS_MODULE);
2679 out:
2680 pr_info("error %s: %s\n", path?path:"", err);
2681 return ret;
2682 }
2683
2684
2685 struct pdev {
2686 struct list_head list;
2687 struct cached_dev *dc;
2688 };
2689
bch_pending_bdevs_cleanup(struct kobject * k,struct kobj_attribute * attr,const char * buffer,size_t size)2690 static ssize_t bch_pending_bdevs_cleanup(struct kobject *k,
2691 struct kobj_attribute *attr,
2692 const char *buffer,
2693 size_t size)
2694 {
2695 LIST_HEAD(pending_devs);
2696 ssize_t ret = size;
2697 struct cached_dev *dc, *tdc;
2698 struct pdev *pdev, *tpdev;
2699 struct cache_set *c, *tc;
2700
2701 mutex_lock(&bch_register_lock);
2702 list_for_each_entry_safe(dc, tdc, &uncached_devices, list) {
2703 pdev = kmalloc(sizeof(struct pdev), GFP_KERNEL);
2704 if (!pdev)
2705 break;
2706 pdev->dc = dc;
2707 list_add(&pdev->list, &pending_devs);
2708 }
2709
2710 list_for_each_entry_safe(pdev, tpdev, &pending_devs, list) {
2711 list_for_each_entry_safe(c, tc, &bch_cache_sets, list) {
2712 char *pdev_set_uuid = pdev->dc->sb.set_uuid;
2713 char *set_uuid = c->set_uuid;
2714
2715 if (!memcmp(pdev_set_uuid, set_uuid, 16)) {
2716 list_del(&pdev->list);
2717 kfree(pdev);
2718 break;
2719 }
2720 }
2721 }
2722 mutex_unlock(&bch_register_lock);
2723
2724 list_for_each_entry_safe(pdev, tpdev, &pending_devs, list) {
2725 pr_info("delete pdev %p\n", pdev);
2726 list_del(&pdev->list);
2727 bcache_device_stop(&pdev->dc->disk);
2728 kfree(pdev);
2729 }
2730
2731 return ret;
2732 }
2733
bcache_reboot(struct notifier_block * n,unsigned long code,void * x)2734 static int bcache_reboot(struct notifier_block *n, unsigned long code, void *x)
2735 {
2736 if (bcache_is_reboot)
2737 return NOTIFY_DONE;
2738
2739 if (code == SYS_DOWN ||
2740 code == SYS_HALT ||
2741 code == SYS_POWER_OFF) {
2742 DEFINE_WAIT(wait);
2743 unsigned long start = jiffies;
2744 bool stopped = false;
2745
2746 struct cache_set *c, *tc;
2747 struct cached_dev *dc, *tdc;
2748
2749 mutex_lock(&bch_register_lock);
2750
2751 if (bcache_is_reboot)
2752 goto out;
2753
2754 /* New registration is rejected since now */
2755 bcache_is_reboot = true;
2756 /*
2757 * Make registering caller (if there is) on other CPU
2758 * core know bcache_is_reboot set to true earlier
2759 */
2760 smp_mb();
2761
2762 if (list_empty(&bch_cache_sets) &&
2763 list_empty(&uncached_devices))
2764 goto out;
2765
2766 mutex_unlock(&bch_register_lock);
2767
2768 pr_info("Stopping all devices:\n");
2769
2770 /*
2771 * The reason bch_register_lock is not held to call
2772 * bch_cache_set_stop() and bcache_device_stop() is to
2773 * avoid potential deadlock during reboot, because cache
2774 * set or bcache device stopping process will acqurie
2775 * bch_register_lock too.
2776 *
2777 * We are safe here because bcache_is_reboot sets to
2778 * true already, register_bcache() will reject new
2779 * registration now. bcache_is_reboot also makes sure
2780 * bcache_reboot() won't be re-entered on by other thread,
2781 * so there is no race in following list iteration by
2782 * list_for_each_entry_safe().
2783 */
2784 list_for_each_entry_safe(c, tc, &bch_cache_sets, list)
2785 bch_cache_set_stop(c);
2786
2787 list_for_each_entry_safe(dc, tdc, &uncached_devices, list)
2788 bcache_device_stop(&dc->disk);
2789
2790
2791 /*
2792 * Give an early chance for other kthreads and
2793 * kworkers to stop themselves
2794 */
2795 schedule();
2796
2797 /* What's a condition variable? */
2798 while (1) {
2799 long timeout = start + 10 * HZ - jiffies;
2800
2801 mutex_lock(&bch_register_lock);
2802 stopped = list_empty(&bch_cache_sets) &&
2803 list_empty(&uncached_devices);
2804
2805 if (timeout < 0 || stopped)
2806 break;
2807
2808 prepare_to_wait(&unregister_wait, &wait,
2809 TASK_UNINTERRUPTIBLE);
2810
2811 mutex_unlock(&bch_register_lock);
2812 schedule_timeout(timeout);
2813 }
2814
2815 finish_wait(&unregister_wait, &wait);
2816
2817 if (stopped)
2818 pr_info("All devices stopped\n");
2819 else
2820 pr_notice("Timeout waiting for devices to be closed\n");
2821 out:
2822 mutex_unlock(&bch_register_lock);
2823 }
2824
2825 return NOTIFY_DONE;
2826 }
2827
2828 static struct notifier_block reboot = {
2829 .notifier_call = bcache_reboot,
2830 .priority = INT_MAX, /* before any real devices */
2831 };
2832
bcache_exit(void)2833 static void bcache_exit(void)
2834 {
2835 bch_debug_exit();
2836 bch_request_exit();
2837 if (bcache_kobj)
2838 kobject_put(bcache_kobj);
2839 if (bcache_wq)
2840 destroy_workqueue(bcache_wq);
2841 if (bch_journal_wq)
2842 destroy_workqueue(bch_journal_wq);
2843 if (bch_flush_wq)
2844 destroy_workqueue(bch_flush_wq);
2845 bch_btree_exit();
2846
2847 if (bcache_major)
2848 unregister_blkdev(bcache_major, "bcache");
2849 unregister_reboot_notifier(&reboot);
2850 mutex_destroy(&bch_register_lock);
2851 }
2852
2853 /* Check and fixup module parameters */
check_module_parameters(void)2854 static void check_module_parameters(void)
2855 {
2856 if (bch_cutoff_writeback_sync == 0)
2857 bch_cutoff_writeback_sync = CUTOFF_WRITEBACK_SYNC;
2858 else if (bch_cutoff_writeback_sync > CUTOFF_WRITEBACK_SYNC_MAX) {
2859 pr_warn("set bch_cutoff_writeback_sync (%u) to max value %u\n",
2860 bch_cutoff_writeback_sync, CUTOFF_WRITEBACK_SYNC_MAX);
2861 bch_cutoff_writeback_sync = CUTOFF_WRITEBACK_SYNC_MAX;
2862 }
2863
2864 if (bch_cutoff_writeback == 0)
2865 bch_cutoff_writeback = CUTOFF_WRITEBACK;
2866 else if (bch_cutoff_writeback > CUTOFF_WRITEBACK_MAX) {
2867 pr_warn("set bch_cutoff_writeback (%u) to max value %u\n",
2868 bch_cutoff_writeback, CUTOFF_WRITEBACK_MAX);
2869 bch_cutoff_writeback = CUTOFF_WRITEBACK_MAX;
2870 }
2871
2872 if (bch_cutoff_writeback > bch_cutoff_writeback_sync) {
2873 pr_warn("set bch_cutoff_writeback (%u) to %u\n",
2874 bch_cutoff_writeback, bch_cutoff_writeback_sync);
2875 bch_cutoff_writeback = bch_cutoff_writeback_sync;
2876 }
2877 }
2878
bcache_init(void)2879 static int __init bcache_init(void)
2880 {
2881 static const struct attribute *files[] = {
2882 &ksysfs_register.attr,
2883 &ksysfs_register_quiet.attr,
2884 &ksysfs_pendings_cleanup.attr,
2885 NULL
2886 };
2887
2888 check_module_parameters();
2889
2890 mutex_init(&bch_register_lock);
2891 init_waitqueue_head(&unregister_wait);
2892 register_reboot_notifier(&reboot);
2893
2894 bcache_major = register_blkdev(0, "bcache");
2895 if (bcache_major < 0) {
2896 unregister_reboot_notifier(&reboot);
2897 mutex_destroy(&bch_register_lock);
2898 return bcache_major;
2899 }
2900
2901 if (bch_btree_init())
2902 goto err;
2903
2904 bcache_wq = alloc_workqueue("bcache", WQ_MEM_RECLAIM, 0);
2905 if (!bcache_wq)
2906 goto err;
2907
2908 /*
2909 * Let's not make this `WQ_MEM_RECLAIM` for the following reasons:
2910 *
2911 * 1. It used `system_wq` before which also does no memory reclaim.
2912 * 2. With `WQ_MEM_RECLAIM` desktop stalls, increased boot times, and
2913 * reduced throughput can be observed.
2914 *
2915 * We still want to user our own queue to not congest the `system_wq`.
2916 */
2917 bch_flush_wq = alloc_workqueue("bch_flush", 0, 0);
2918 if (!bch_flush_wq)
2919 goto err;
2920
2921 bch_journal_wq = alloc_workqueue("bch_journal", WQ_MEM_RECLAIM, 0);
2922 if (!bch_journal_wq)
2923 goto err;
2924
2925 bcache_kobj = kobject_create_and_add("bcache", fs_kobj);
2926 if (!bcache_kobj)
2927 goto err;
2928
2929 if (bch_request_init() ||
2930 sysfs_create_files(bcache_kobj, files))
2931 goto err;
2932
2933 bch_debug_init();
2934 closure_debug_init();
2935
2936 bcache_is_reboot = false;
2937
2938 return 0;
2939 err:
2940 bcache_exit();
2941 return -ENOMEM;
2942 }
2943
2944 /*
2945 * Module hooks
2946 */
2947 module_exit(bcache_exit);
2948 module_init(bcache_init);
2949
2950 module_param(bch_cutoff_writeback, uint, 0);
2951 MODULE_PARM_DESC(bch_cutoff_writeback, "threshold to cutoff writeback");
2952
2953 module_param(bch_cutoff_writeback_sync, uint, 0);
2954 MODULE_PARM_DESC(bch_cutoff_writeback_sync, "hard threshold to cutoff writeback");
2955
2956 MODULE_DESCRIPTION("Bcache: a Linux block layer cache");
2957 MODULE_AUTHOR("Kent Overstreet <kent.overstreet@gmail.com>");
2958 MODULE_LICENSE("GPL");
2959