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