drivers/md/bcache/super.c

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