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 <[email protected]>
   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/pagemap.h>
  20 #include <linux/debugfs.h>
  21 #include <linux/idr.h>
  22 #include <linux/kthread.h>
  23 #include <linux/workqueue.h>
  24 #include <linux/module.h>
  25 #include <linux/random.h>
  26 #include <linux/reboot.h>
  27 #include <linux/sysfs.h>
  28
  29 unsigned int bch_cutoff_writeback;
  30 unsigned int bch_cutoff_writeback_sync;
  31
  32 static const char bcache_magic[] = {
  33         0xc6, 0x85, 0x73, 0xf6, 0x4e, 0x1a, 0x45, 0xca,
  34         0x82, 0x65, 0xf5, 0x7f, 0x48, 0xba, 0x6d, 0x81
  35 };
  36
  37 static const char invalid_uuid[] = {
  38         0xa0, 0x3e, 0xf8, 0xed, 0x3e, 0xe1, 0xb8, 0x78,
  39         0xc8, 0x50, 0xfc, 0x5e, 0xcb, 0x16, 0xcd, 0x99
  40 };
  41
  42 static struct kobject *bcache_kobj;
  43 struct mutex bch_register_lock;
  44 bool bcache_is_reboot;
  45 LIST_HEAD(bch_cache_sets);
  46 static LIST_HEAD(uncached_devices);
  47
  48 static int bcache_major;
  49 static DEFINE_IDA(bcache_device_idx);
  50 static wait_queue_head_t unregister_wait;
  51 struct workqueue_struct *bcache_wq;
  52 struct workqueue_struct *bch_flush_wq;
  53 struct workqueue_struct *bch_journal_wq;
  54
  55
  56 #define BTREE_MAX_PAGES         (256 * 1024 / PAGE_SIZE)
  57 /* limitation of partitions number on single bcache device */
  58 #define BCACHE_MINORS           128
  59 /* limitation of bcache devices number on single system */
  60 #define BCACHE_DEVICE_IDX_MAX   ((1U << MINORBITS)/BCACHE_MINORS)
  61
  62 /* Superblock */
  63
  64 static unsigned int get_bucket_size(struct cache_sb *sb, struct cache_sb_disk *s)
  65 {
  66         unsigned int bucket_size = le16_to_cpu(s->bucket_size);
  67
  68         if (sb->version >= BCACHE_SB_VERSION_CDEV_WITH_FEATURES) {
  69                 if (bch_has_feature_large_bucket(sb)) {
  70                         unsigned int max, order;
  71
  72                         max = sizeof(unsigned int) * BITS_PER_BYTE - 1;
  73                         order = le16_to_cpu(s->bucket_size);
  74                         /*
  75                          * bcache tool will make sure the overflow won't
  76                          * happen, an error message here is enough.
  77                          */
  78                         if (order > max)
  79                                 pr_err("Bucket size (1 << %u) overflows\n",
  80                                         order);
  81                         bucket_size = 1 << order;
  82                 } else if (bch_has_feature_obso_large_bucket(sb)) {
  83                         bucket_size +=
  84                                 le16_to_cpu(s->obso_bucket_size_hi) << 16;
  85                 }
  86         }
  87
  88         return bucket_size;
  89 }
  90
  91 static const char *read_super_common(struct cache_sb *sb,  struct block_device *bdev,
  92                                      struct cache_sb_disk *s)
  93 {
  94         const char *err;
  95         unsigned int i;
  96
  97         sb->first_bucket= le16_to_cpu(s->first_bucket);
  98         sb->nbuckets    = le64_to_cpu(s->nbuckets);
  99         sb->bucket_size = get_bucket_size(sb, s);
 100
 101         sb->nr_in_set   = le16_to_cpu(s->nr_in_set);
 102         sb->nr_this_dev = le16_to_cpu(s->nr_this_dev);
 103
 104         err = "Too many journal buckets";
 105         if (sb->keys > SB_JOURNAL_BUCKETS)
 106                 goto err;
 107
 108         err = "Too many buckets";
 109         if (sb->nbuckets > LONG_MAX)
 110                 goto err;
 111
 112         err = "Not enough buckets";
 113         if (sb->nbuckets < 1 << 7)
 114                 goto err;
 115
 116         err = "Bad block size (not power of 2)";
 117         if (!is_power_of_2(sb->block_size))
 118                 goto err;
 119
 120         err = "Bad block size (larger than page size)";
 121         if (sb->block_size > PAGE_SECTORS)
 122                 goto err;
 123
 124         err = "Bad bucket size (not power of 2)";
 125         if (!is_power_of_2(sb->bucket_size))
 126                 goto err;
 127
 128         err = "Bad bucket size (smaller than page size)";
 129         if (sb->bucket_size < PAGE_SECTORS)
 130                 goto err;
 131
 132         err = "Invalid superblock: device too small";
 133         if (get_capacity(bdev->bd_disk) <
 134             sb->bucket_size * sb->nbuckets)
 135                 goto err;
 136
 137         err = "Bad UUID";
 138         if (bch_is_zero(sb->set_uuid, 16))
 139                 goto err;
 140
 141         err = "Bad cache device number in set";
 142         if (!sb->nr_in_set ||
 143             sb->nr_in_set <= sb->nr_this_dev ||
 144             sb->nr_in_set > MAX_CACHES_PER_SET)
 145                 goto err;
 146
 147         err = "Journal buckets not sequential";
 148         for (i = 0; i < sb->keys; i++)
 149                 if (sb->d[i] != sb->first_bucket + i)
 150                         goto err;
 151
 152         err = "Too many journal buckets";
 153         if (sb->first_bucket + sb->keys > sb->nbuckets)
 154                 goto err;
 155
 156         err = "Invalid superblock: first bucket comes before end of super";
 157         if (sb->first_bucket * sb->bucket_size < 16)
 158                 goto err;
 159
 160         err = NULL;
 161 err:
 162         return err;
 163 }
 164
 165
 166 static const char *read_super(struct cache_sb *sb, struct block_device *bdev,
 167                               struct cache_sb_disk **res)
 168 {
 169         const char *err;
 170         struct cache_sb_disk *s;
 171         struct page *page;
 172         unsigned int i;
 173
 174         page = read_cache_page_gfp(bdev->bd_mapping,
 175                                    SB_OFFSET >> PAGE_SHIFT, GFP_KERNEL);
 176         if (IS_ERR(page))
 177                 return "IO error";
 178         s = page_address(page) + offset_in_page(SB_OFFSET);
 179
 180         sb->offset              = le64_to_cpu(s->offset);
 181         sb->version             = le64_to_cpu(s->version);
 182
 183         memcpy(sb->magic,       s->magic, 16);
 184         memcpy(sb->uuid,        s->uuid, 16);
 185         memcpy(sb->set_uuid,    s->set_uuid, 16);
 186         memcpy(sb->label,       s->label, SB_LABEL_SIZE);
 187
 188         sb->flags               = le64_to_cpu(s->flags);
 189         sb->seq                 = le64_to_cpu(s->seq);
 190         sb->last_mount          = le32_to_cpu(s->last_mount);
 191         sb->keys                = le16_to_cpu(s->keys);
 192
 193         for (i = 0; i < SB_JOURNAL_BUCKETS; i++)
 194                 sb->d[i] = le64_to_cpu(s->d[i]);
 195
 196         pr_debug("read sb version %llu, flags %llu, seq %llu, journal size %u\n",
 197                  sb->version, sb->flags, sb->seq, sb->keys);
 198
 199         err = "Not a bcache superblock (bad offset)";
 200         if (sb->offset != SB_SECTOR)
 201                 goto err;
 202
 203         err = "Not a bcache superblock (bad magic)";
 204         if (memcmp(sb->magic, bcache_magic, 16))
 205                 goto err;
 206
 207         err = "Bad checksum";
 208         if (s->csum != csum_set(s))
 209                 goto err;
 210
 211         err = "Bad UUID";
 212         if (bch_is_zero(sb->uuid, 16))
 213                 goto err;
 214
 215         sb->block_size  = le16_to_cpu(s->block_size);
 216
 217         err = "Superblock block size smaller than device block size";
 218         if (sb->block_size << 9 < bdev_logical_block_size(bdev))
 219                 goto err;
 220
 221         switch (sb->version) {
 222         case BCACHE_SB_VERSION_BDEV:
 223                 sb->data_offset = BDEV_DATA_START_DEFAULT;
 224                 break;
 225         case BCACHE_SB_VERSION_BDEV_WITH_OFFSET:
 226         case BCACHE_SB_VERSION_BDEV_WITH_FEATURES:
 227                 sb->data_offset = le64_to_cpu(s->data_offset);
 228
 229                 err = "Bad data offset";
 230                 if (sb->data_offset < BDEV_DATA_START_DEFAULT)
 231                         goto err;
 232
 233                 break;
 234         case BCACHE_SB_VERSION_CDEV:
 235         case BCACHE_SB_VERSION_CDEV_WITH_UUID:
 236                 err = read_super_common(sb, bdev, s);
 237                 if (err)
 238                         goto err;
 239                 break;
 240         case BCACHE_SB_VERSION_CDEV_WITH_FEATURES:
 241                 /*
 242                  * Feature bits are needed in read_super_common(),
 243                  * convert them firstly.
 244                  */
 245                 sb->feature_compat = le64_to_cpu(s->feature_compat);
 246                 sb->feature_incompat = le64_to_cpu(s->feature_incompat);
 247                 sb->feature_ro_compat = le64_to_cpu(s->feature_ro_compat);
 248
 249                 /* Check incompatible features */
 250                 err = "Unsupported compatible feature found";
 251                 if (bch_has_unknown_compat_features(sb))
 252                         goto err;
 253
 254                 err = "Unsupported read-only compatible feature found";
 255                 if (bch_has_unknown_ro_compat_features(sb))
 256                         goto err;
 257
 258                 err = "Unsupported incompatible feature found";
 259                 if (bch_has_unknown_incompat_features(sb))
 260                         goto err;
 261
 262                 err = read_super_common(sb, bdev, s);
 263                 if (err)
 264                         goto err;
 265                 break;
 266         default:
 267                 err = "Unsupported superblock version";
 268                 goto err;
 269         }
 270
 271         sb->last_mount = (u32)ktime_get_real_seconds();
 272         *res = s;
 273         return NULL;
 274 err:
 275         put_page(page);
 276         return err;
 277 }
 278
 279 static void write_bdev_super_endio(struct bio *bio)
 280 {
 281         struct cached_dev *dc = bio->bi_private;
 282
 283         if (bio->bi_status)
 284                 bch_count_backing_io_errors(dc, bio);
 285
 286         closure_put(&dc->sb_write);
 287 }
 288
 289 static void __write_super(struct cache_sb *sb, struct cache_sb_disk *out,
 290                 struct bio *bio)
 291 {
 292         unsigned int i;
 293
 294         bio->bi_opf = REQ_OP_WRITE | REQ_SYNC | REQ_META;
 295         bio->bi_iter.bi_sector  = SB_SECTOR;
 296         __bio_add_page(bio, virt_to_page(out), SB_SIZE,
 297                         offset_in_page(out));
 298
 299         out->offset             = cpu_to_le64(sb->offset);
 300
 301         memcpy(out->uuid,       sb->uuid, 16);
 302         memcpy(out->set_uuid,   sb->set_uuid, 16);
 303         memcpy(out->label,      sb->label, SB_LABEL_SIZE);
 304
 305         out->flags              = cpu_to_le64(sb->flags);
 306         out->seq                = cpu_to_le64(sb->seq);
 307
 308         out->last_mount         = cpu_to_le32(sb->last_mount);
 309         out->first_bucket       = cpu_to_le16(sb->first_bucket);
 310         out->keys               = cpu_to_le16(sb->keys);
 311
 312         for (i = 0; i < sb->keys; i++)
 313                 out->d[i] = cpu_to_le64(sb->d[i]);
 314
 315         if (sb->version >= BCACHE_SB_VERSION_CDEV_WITH_FEATURES) {
 316                 out->feature_compat    = cpu_to_le64(sb->feature_compat);
 317                 out->feature_incompat  = cpu_to_le64(sb->feature_incompat);
 318                 out->feature_ro_compat = cpu_to_le64(sb->feature_ro_compat);
 319         }
 320
 321         out->version            = cpu_to_le64(sb->version);
 322         out->csum = csum_set(out);
 323
 324         pr_debug("ver %llu, flags %llu, seq %llu\n",
 325                  sb->version, sb->flags, sb->seq);
 326
 327         submit_bio(bio);
 328 }
 329
 330 static CLOSURE_CALLBACK(bch_write_bdev_super_unlock)
 331 {
 332         closure_type(dc, struct cached_dev, sb_write);
 333
 334         up(&dc->sb_write_mutex);
 335 }
 336
 337 void bch_write_bdev_super(struct cached_dev *dc, struct closure *parent)
 338 {
 339         struct closure *cl = &dc->sb_write;
 340         struct bio *bio = &dc->sb_bio;
 341
 342         down(&dc->sb_write_mutex);
 343         closure_init(cl, parent);
 344
 345         bio_init(bio, dc->bdev, dc->sb_bv, 1, 0);
 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 CLOSURE_CALLBACK(bcache_write_super_unlock)
 367 {
 368         closure_type(c, 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->bdev, ca->sb_bv, 1, 0);
 389         bio->bi_end_io  = write_super_endio;
 390         bio->bi_private = ca;
 391
 392         closure_get(cl);
 393         __write_super(&ca->sb, ca->sb_disk, bio);
 394
 395         closure_return_with_destructor(cl, bcache_write_super_unlock);
 396 }
 397
 398 /* UUID io */
 399
 400 static void uuid_endio(struct bio *bio)
 401 {
 402         struct closure *cl = bio->bi_private;
 403         struct cache_set *c = container_of(cl, struct cache_set, uuid_write);
 404
 405         cache_set_err_on(bio->bi_status, c, "accessing uuids");
 406         bch_bbio_free(bio, c);
 407         closure_put(cl);
 408 }
 409
 410 static CLOSURE_CALLBACK(uuid_io_unlock)
 411 {
 412         closure_type(c, struct cache_set, uuid_write);
 413
 414         up(&c->uuid_write_mutex);
 415 }
 416
 417 static void uuid_io(struct cache_set *c, blk_opf_t opf, struct bkey *k,
 418                     struct closure *parent)
 419 {
 420         struct closure *cl = &c->uuid_write;
 421         struct uuid_entry *u;
 422         unsigned int i;
 423         char buf[80];
 424
 425         BUG_ON(!parent);
 426         down(&c->uuid_write_mutex);
 427         closure_init(cl, parent);
 428
 429         for (i = 0; i < KEY_PTRS(k); i++) {
 430                 struct bio *bio = bch_bbio_alloc(c);
 431
 432                 bio->bi_opf = opf | REQ_SYNC | REQ_META;
 433                 bio->bi_iter.bi_size = KEY_SIZE(k) << 9;
 434
 435                 bio->bi_end_io  = uuid_endio;
 436                 bio->bi_private = cl;
 437                 bch_bio_map(bio, c->uuids);
 438
 439                 bch_submit_bbio(bio, c, k, i);
 440
 441                 if ((opf & REQ_OP_MASK) != REQ_OP_WRITE)
 442                         break;
 443         }
 444
 445         bch_extent_to_text(buf, sizeof(buf), k);
 446         pr_debug("%s UUIDs at %s\n", (opf & REQ_OP_MASK) == REQ_OP_WRITE ?
 447                  "wrote" : "read", buf);
 448
 449         for (u = c->uuids; u < c->uuids + c->nr_uuids; u++)
 450                 if (!bch_is_zero(u->uuid, 16))
 451                         pr_debug("Slot %zi: %pU: %s: 1st: %u last: %u inv: %u\n",
 452                                  u - c->uuids, u->uuid, u->label,
 453                                  u->first_reg, u->last_reg, u->invalidated);
 454
 455         closure_return_with_destructor(cl, uuid_io_unlock);
 456 }
 457
 458 static char *uuid_read(struct cache_set *c, struct jset *j, struct closure *cl)
 459 {
 460         struct bkey *k = &j->uuid_bucket;
 461
 462         if (__bch_btree_ptr_invalid(c, k))
 463                 return "bad uuid pointer";
 464
 465         bkey_copy(&c->uuid_bucket, k);
 466         uuid_io(c, REQ_OP_READ, k, cl);
 467
 468         if (j->version < BCACHE_JSET_VERSION_UUIDv1) {
 469                 struct uuid_entry_v0    *u0 = (void *) c->uuids;
 470                 struct uuid_entry       *u1 = (void *) c->uuids;
 471                 int i;
 472
 473                 closure_sync(cl);
 474
 475                 /*
 476                  * Since the new uuid entry is bigger than the old, we have to
 477                  * convert starting at the highest memory address and work down
 478                  * in order to do it in place
 479                  */
 480
 481                 for (i = c->nr_uuids - 1;
 482                      i >= 0;
 483                      --i) {
 484                         memcpy(u1[i].uuid,      u0[i].uuid, 16);
 485                         memcpy(u1[i].label,     u0[i].label, 32);
 486
 487                         u1[i].first_reg         = u0[i].first_reg;
 488                         u1[i].last_reg          = u0[i].last_reg;
 489                         u1[i].invalidated       = u0[i].invalidated;
 490
 491                         u1[i].flags     = 0;
 492                         u1[i].sectors   = 0;
 493                 }
 494         }
 495
 496         return NULL;
 497 }
 498
 499 static int __uuid_write(struct cache_set *c)
 500 {
 501         BKEY_PADDED(key) k;
 502         struct closure cl;
 503         struct cache *ca = c->cache;
 504         unsigned int size;
 505
 506         closure_init_stack(&cl);
 507         lockdep_assert_held(&bch_register_lock);
 508
 509         if (bch_bucket_alloc_set(c, RESERVE_BTREE, &k.key, true))
 510                 return 1;
 511
 512         size =  meta_bucket_pages(&ca->sb) * PAGE_SECTORS;
 513         SET_KEY_SIZE(&k.key, size);
 514         uuid_io(c, REQ_OP_WRITE, &k.key, &cl);
 515         closure_sync(&cl);
 516
 517         /* Only one bucket used for uuid write */
 518         atomic_long_add(ca->sb.bucket_size, &ca->meta_sectors_written);
 519
 520         bkey_copy(&c->uuid_bucket, &k.key);
 521         bkey_put(c, &k.key);
 522         return 0;
 523 }
 524
 525 int bch_uuid_write(struct cache_set *c)
 526 {
 527         int ret = __uuid_write(c);
 528
 529         if (!ret)
 530                 bch_journal_meta(c, NULL);
 531
 532         return ret;
 533 }
 534
 535 static struct uuid_entry *uuid_find(struct cache_set *c, const char *uuid)
 536 {
 537         struct uuid_entry *u;
 538
 539         for (u = c->uuids;
 540              u < c->uuids + c->nr_uuids; u++)
 541                 if (!memcmp(u->uuid, uuid, 16))
 542                         return u;
 543
 544         return NULL;
 545 }
 546
 547 static struct uuid_entry *uuid_find_empty(struct cache_set *c)
 548 {
 549         static const char zero_uuid[16] = "\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0";
 550
 551         return uuid_find(c, zero_uuid);
 552 }
 553
 554 /*
 555  * Bucket priorities/gens:
 556  *
 557  * For each bucket, we store on disk its
 558  *   8 bit gen
 559  *  16 bit priority
 560  *
 561  * See alloc.c for an explanation of the gen. The priority is used to implement
 562  * lru (and in the future other) cache replacement policies; for most purposes
 563  * it's just an opaque integer.
 564  *
 565  * The gens and the priorities don't have a whole lot to do with each other, and
 566  * it's actually the gens that must be written out at specific times - it's no
 567  * big deal if the priorities don't get written, if we lose them we just reuse
 568  * buckets in suboptimal order.
 569  *
 570  * On disk they're stored in a packed array, and in as many buckets are required
 571  * to fit them all. The buckets we use to store them form a list; the journal
 572  * header points to the first bucket, the first bucket points to the second
 573  * bucket, et cetera.
 574  *
 575  * This code is used by the allocation code; periodically (whenever it runs out
 576  * of buckets to allocate from) the allocation code will invalidate some
 577  * buckets, but it can't use those buckets until their new gens are safely on
 578  * disk.
 579  */
 580
 581 static void prio_endio(struct bio *bio)
 582 {
 583         struct cache *ca = bio->bi_private;
 584
 585         cache_set_err_on(bio->bi_status, ca->set, "accessing priorities");
 586         bch_bbio_free(bio, ca->set);
 587         closure_put(&ca->prio);
 588 }
 589
 590 static void prio_io(struct cache *ca, uint64_t bucket, blk_opf_t opf)
 591 {
 592         struct closure *cl = &ca->prio;
 593         struct bio *bio = bch_bbio_alloc(ca->set);
 594
 595         closure_init_stack(cl);
 596
 597         bio->bi_iter.bi_sector  = bucket * ca->sb.bucket_size;
 598         bio_set_dev(bio, ca->bdev);
 599         bio->bi_iter.bi_size    = meta_bucket_bytes(&ca->sb);
 600
 601         bio->bi_end_io  = prio_endio;
 602         bio->bi_private = ca;
 603         bio->bi_opf = opf | REQ_SYNC | REQ_META;
 604         bch_bio_map(bio, ca->disk_buckets);
 605
 606         closure_bio_submit(ca->set, bio, &ca->prio);
 607         closure_sync(cl);
 608 }
 609
 610 int bch_prio_write(struct cache *ca, bool wait)
 611 {
 612         int i;
 613         struct bucket *b;
 614         struct closure cl;
 615
 616         pr_debug("free_prio=%zu, free_none=%zu, free_inc=%zu\n",
 617                  fifo_used(&ca->free[RESERVE_PRIO]),
 618                  fifo_used(&ca->free[RESERVE_NONE]),
 619                  fifo_used(&ca->free_inc));
 620
 621         /*
 622          * Pre-check if there are enough free buckets. In the non-blocking
 623          * scenario it's better to fail early rather than starting to allocate
 624          * buckets and do a cleanup later in case of failure.
 625          */
 626         if (!wait) {
 627                 size_t avail = fifo_used(&ca->free[RESERVE_PRIO]) +
 628                                fifo_used(&ca->free[RESERVE_NONE]);
 629                 if (prio_buckets(ca) > avail)
 630                         return -ENOMEM;
 631         }
 632
 633         closure_init_stack(&cl);
 634
 635         lockdep_assert_held(&ca->set->bucket_lock);
 636
 637         ca->disk_buckets->seq++;
 638
 639         atomic_long_add(ca->sb.bucket_size * prio_buckets(ca),
 640                         &ca->meta_sectors_written);
 641
 642         for (i = prio_buckets(ca) - 1; i >= 0; --i) {
 643                 long bucket;
 644                 struct prio_set *p = ca->disk_buckets;
 645                 struct bucket_disk *d = p->data;
 646                 struct bucket_disk *end = d + prios_per_bucket(ca);
 647
 648                 for (b = ca->buckets + i * prios_per_bucket(ca);
 649                      b < ca->buckets + ca->sb.nbuckets && d < end;
 650                      b++, d++) {
 651                         d->prio = cpu_to_le16(b->prio);
 652                         d->gen = b->gen;
 653                 }
 654
 655                 p->next_bucket  = ca->prio_buckets[i + 1];
 656                 p->magic        = pset_magic(&ca->sb);
 657                 p->csum         = bch_crc64(&p->magic, meta_bucket_bytes(&ca->sb) - 8);
 658
 659                 bucket = bch_bucket_alloc(ca, RESERVE_PRIO, wait);
 660                 BUG_ON(bucket == -1);
 661
 662                 mutex_unlock(&ca->set->bucket_lock);
 663                 prio_io(ca, bucket, REQ_OP_WRITE);
 664                 mutex_lock(&ca->set->bucket_lock);
 665
 666                 ca->prio_buckets[i] = bucket;
 667                 atomic_dec_bug(&ca->buckets[bucket].pin);
 668         }
 669
 670         mutex_unlock(&ca->set->bucket_lock);
 671
 672         bch_journal_meta(ca->set, &cl);
 673         closure_sync(&cl);
 674
 675         mutex_lock(&ca->set->bucket_lock);
 676
 677         /*
 678          * Don't want the old priorities to get garbage collected until after we
 679          * finish writing the new ones, and they're journalled
 680          */
 681         for (i = 0; i < prio_buckets(ca); i++) {
 682                 if (ca->prio_last_buckets[i])
 683                         __bch_bucket_free(ca,
 684                                 &ca->buckets[ca->prio_last_buckets[i]]);
 685
 686                 ca->prio_last_buckets[i] = ca->prio_buckets[i];
 687         }
 688         return 0;
 689 }
 690
 691 static int prio_read(struct cache *ca, uint64_t bucket)
 692 {
 693         struct prio_set *p = ca->disk_buckets;
 694         struct bucket_disk *d = p->data + prios_per_bucket(ca), *end = d;
 695         struct bucket *b;
 696         unsigned int bucket_nr = 0;
 697         int ret = -EIO;
 698
 699         for (b = ca->buckets;
 700              b < ca->buckets + ca->sb.nbuckets;
 701              b++, d++) {
 702                 if (d == end) {
 703                         ca->prio_buckets[bucket_nr] = bucket;
 704                         ca->prio_last_buckets[bucket_nr] = bucket;
 705                         bucket_nr++;
 706
 707                         prio_io(ca, bucket, REQ_OP_READ);
 708
 709                         if (p->csum !=
 710                             bch_crc64(&p->magic, meta_bucket_bytes(&ca->sb) - 8)) {
 711                                 pr_warn("bad csum reading priorities\n");
 712                                 goto out;
 713                         }
 714
 715                         if (p->magic != pset_magic(&ca->sb)) {
 716                                 pr_warn("bad magic reading priorities\n");
 717                                 goto out;
 718                         }
 719
 720                         bucket = p->next_bucket;
 721                         d = p->data;
 722                 }
 723
 724                 b->prio = le16_to_cpu(d->prio);
 725                 b->gen = b->last_gc = d->gen;
 726         }
 727
 728         ret = 0;
 729 out:
 730         return ret;
 731 }
 732
 733 /* Bcache device */
 734
 735 static int open_dev(struct gendisk *disk, blk_mode_t mode)
 736 {
 737         struct bcache_device *d = disk->private_data;
 738
 739         if (test_bit(BCACHE_DEV_CLOSING, &d->flags))
 740                 return -ENXIO;
 741
 742         closure_get(&d->cl);
 743         return 0;
 744 }
 745
 746 static void release_dev(struct gendisk *b)
 747 {
 748         struct bcache_device *d = b->private_data;
 749
 750         closure_put(&d->cl);
 751 }
 752
 753 static int ioctl_dev(struct block_device *b, blk_mode_t mode,
 754                      unsigned int cmd, unsigned long arg)
 755 {
 756         struct bcache_device *d = b->bd_disk->private_data;
 757
 758         return d->ioctl(d, mode, cmd, arg);
 759 }
 760
 761 static const struct block_device_operations bcache_cached_ops = {
 762         .submit_bio     = cached_dev_submit_bio,
 763         .open           = open_dev,
 764         .release        = release_dev,
 765         .ioctl          = ioctl_dev,
 766         .owner          = THIS_MODULE,
 767 };
 768
 769 static const struct block_device_operations bcache_flash_ops = {
 770         .submit_bio     = flash_dev_submit_bio,
 771         .open           = open_dev,
 772         .release        = release_dev,
 773         .ioctl          = ioctl_dev,
 774         .owner          = THIS_MODULE,
 775 };
 776
 777 void bcache_device_stop(struct bcache_device *d)
 778 {
 779         if (!test_and_set_bit(BCACHE_DEV_CLOSING, &d->flags))
 780                 /*
 781                  * closure_fn set to
 782                  * - cached device: cached_dev_flush()
 783                  * - flash dev: flash_dev_flush()
 784                  */
 785                 closure_queue(&d->cl);
 786 }
 787
 788 static void bcache_device_unlink(struct bcache_device *d)
 789 {
 790         lockdep_assert_held(&bch_register_lock);
 791
 792         if (d->c && !test_and_set_bit(BCACHE_DEV_UNLINK_DONE, &d->flags)) {
 793                 struct cache *ca = d->c->cache;
 794
 795                 sysfs_remove_link(&d->c->kobj, d->name);
 796                 sysfs_remove_link(&d->kobj, "cache");
 797
 798                 bd_unlink_disk_holder(ca->bdev, d->disk);
 799         }
 800 }
 801
 802 static void bcache_device_link(struct bcache_device *d, struct cache_set *c,
 803                                const char *name)
 804 {
 805         struct cache *ca = c->cache;
 806         int ret;
 807
 808         bd_link_disk_holder(ca->bdev, d->disk);
 809
 810         snprintf(d->name, BCACHEDEVNAME_SIZE,
 811                  "%s%u", name, d->id);
 812
 813         ret = sysfs_create_link(&d->kobj, &c->kobj, "cache");
 814         if (ret < 0)
 815                 pr_err("Couldn't create device -> cache set symlink\n");
 816
 817         ret = sysfs_create_link(&c->kobj, &d->kobj, d->name);
 818         if (ret < 0)
 819                 pr_err("Couldn't create cache set -> device symlink\n");
 820
 821         clear_bit(BCACHE_DEV_UNLINK_DONE, &d->flags);
 822 }
 823
 824 static void bcache_device_detach(struct bcache_device *d)
 825 {
 826         lockdep_assert_held(&bch_register_lock);
 827
 828         atomic_dec(&d->c->attached_dev_nr);
 829
 830         if (test_bit(BCACHE_DEV_DETACHING, &d->flags)) {
 831                 struct uuid_entry *u = d->c->uuids + d->id;
 832
 833                 SET_UUID_FLASH_ONLY(u, 0);
 834                 memcpy(u->uuid, invalid_uuid, 16);
 835                 u->invalidated = cpu_to_le32((u32)ktime_get_real_seconds());
 836                 bch_uuid_write(d->c);
 837         }
 838
 839         bcache_device_unlink(d);
 840
 841         d->c->devices[d->id] = NULL;
 842         closure_put(&d->c->caching);
 843         d->c = NULL;
 844 }
 845
 846 static void bcache_device_attach(struct bcache_device *d, struct cache_set *c,
 847                                  unsigned int id)
 848 {
 849         d->id = id;
 850         d->c = c;
 851         c->devices[id] = d;
 852
 853         if (id >= c->devices_max_used)
 854                 c->devices_max_used = id + 1;
 855
 856         closure_get(&c->caching);
 857 }
 858
 859 static inline int first_minor_to_idx(int first_minor)
 860 {
 861         return (first_minor/BCACHE_MINORS);
 862 }
 863
 864 static inline int idx_to_first_minor(int idx)
 865 {
 866         return (idx * BCACHE_MINORS);
 867 }
 868
 869 static void bcache_device_free(struct bcache_device *d)
 870 {
 871         struct gendisk *disk = d->disk;
 872
 873         lockdep_assert_held(&bch_register_lock);
 874
 875         if (disk)
 876                 pr_info("%s stopped\n", disk->disk_name);
 877         else
 878                 pr_err("bcache device (NULL gendisk) stopped\n");
 879
 880         if (d->c)
 881                 bcache_device_detach(d);
 882
 883         if (disk) {
 884                 ida_free(&bcache_device_idx,
 885                          first_minor_to_idx(disk->first_minor));
 886                 put_disk(disk);
 887         }
 888
 889         bioset_exit(&d->bio_split);
 890         kvfree(d->full_dirty_stripes);
 891         kvfree(d->stripe_sectors_dirty);
 892
 893         closure_debug_destroy(&d->cl);
 894 }
 895
 896 static int bcache_device_init(struct bcache_device *d, unsigned int block_size,
 897                 sector_t sectors, struct block_device *cached_bdev,
 898                 const struct block_device_operations *ops)
 899 {
 900         const size_t max_stripes = min_t(size_t, INT_MAX,
 901                                          SIZE_MAX / sizeof(atomic_t));
 902         struct queue_limits lim = {
 903                 .max_hw_sectors         = UINT_MAX,
 904                 .max_sectors            = UINT_MAX,
 905                 .max_segment_size       = UINT_MAX,
 906                 .max_segments           = BIO_MAX_VECS,
 907                 .max_hw_discard_sectors = UINT_MAX,
 908                 .io_min                 = block_size,
 909                 .logical_block_size     = block_size,
 910                 .physical_block_size    = block_size,
 911                 .features               = BLK_FEAT_WRITE_CACHE | BLK_FEAT_FUA,
 912         };
 913         uint64_t n;
 914         int idx;
 915
 916         if (cached_bdev) {
 917                 d->stripe_size = bdev_io_opt(cached_bdev) >> SECTOR_SHIFT;
 918                 lim.io_opt = umax(block_size, bdev_io_opt(cached_bdev));
 919         }
 920         if (!d->stripe_size)
 921                 d->stripe_size = 1 << 31;
 922         else if (d->stripe_size < BCH_MIN_STRIPE_SZ)
 923                 d->stripe_size = roundup(BCH_MIN_STRIPE_SZ, d->stripe_size);
 924
 925         n = DIV_ROUND_UP_ULL(sectors, d->stripe_size);
 926         if (!n || n > max_stripes) {
 927                 pr_err("nr_stripes too large or invalid: %llu (start sector beyond end of disk?)\n",
 928                         n);
 929                 return -ENOMEM;
 930         }
 931         d->nr_stripes = n;
 932
 933         n = d->nr_stripes * sizeof(atomic_t);
 934         d->stripe_sectors_dirty = kvzalloc(n, GFP_KERNEL);
 935         if (!d->stripe_sectors_dirty)
 936                 return -ENOMEM;
 937
 938         n = BITS_TO_LONGS(d->nr_stripes) * sizeof(unsigned long);
 939         d->full_dirty_stripes = kvzalloc(n, GFP_KERNEL);
 940         if (!d->full_dirty_stripes)
 941                 goto out_free_stripe_sectors_dirty;
 942
 943         idx = ida_alloc_max(&bcache_device_idx, BCACHE_DEVICE_IDX_MAX - 1,
 944                             GFP_KERNEL);
 945         if (idx < 0)
 946                 goto out_free_full_dirty_stripes;
 947
 948         if (bioset_init(&d->bio_split, 4, offsetof(struct bbio, bio),
 949                         BIOSET_NEED_BVECS|BIOSET_NEED_RESCUER))
 950                 goto out_ida_remove;
 951
 952         if (lim.logical_block_size > PAGE_SIZE && cached_bdev) {
 953                 /*
 954                  * This should only happen with BCACHE_SB_VERSION_BDEV.
 955                  * Block/page size is checked for BCACHE_SB_VERSION_CDEV.
 956                  */
 957                 pr_info("bcache%i: sb/logical block size (%u) greater than page size (%lu) falling back to device logical block size (%u)\n",
 958                         idx, lim.logical_block_size,
 959                         PAGE_SIZE, bdev_logical_block_size(cached_bdev));
 960
 961                 /* This also adjusts physical block size/min io size if needed */
 962                 lim.logical_block_size = bdev_logical_block_size(cached_bdev);
 963         }
 964
 965         d->disk = blk_alloc_disk(&lim, NUMA_NO_NODE);
 966         if (IS_ERR(d->disk))
 967                 goto out_bioset_exit;
 968
 969         set_capacity(d->disk, sectors);
 970         snprintf(d->disk->disk_name, DISK_NAME_LEN, "bcache%i", idx);
 971
 972         d->disk->major          = bcache_major;
 973         d->disk->first_minor    = idx_to_first_minor(idx);
 974         d->disk->minors         = BCACHE_MINORS;
 975         d->disk->fops           = ops;
 976         d->disk->private_data   = d;
 977         return 0;
 978
 979 out_bioset_exit:
 980         bioset_exit(&d->bio_split);
 981 out_ida_remove:
 982         ida_free(&bcache_device_idx, idx);
 983 out_free_full_dirty_stripes:
 984         kvfree(d->full_dirty_stripes);
 985 out_free_stripe_sectors_dirty:
 986         kvfree(d->stripe_sectors_dirty);
 987         return -ENOMEM;
 988
 989 }
 990
 991 /* Cached device */
 992
 993 static void calc_cached_dev_sectors(struct cache_set *c)
 994 {
 995         uint64_t sectors = 0;
 996         struct cached_dev *dc;
 997
 998         list_for_each_entry(dc, &c->cached_devs, list)
 999                 sectors += bdev_nr_sectors(dc->bdev);
1000
1001         c->cached_dev_sectors = sectors;
1002 }
1003
1004 #define BACKING_DEV_OFFLINE_TIMEOUT 5
1005 static int cached_dev_status_update(void *arg)
1006 {
1007         struct cached_dev *dc = arg;
1008         struct request_queue *q;
1009
1010         /*
1011          * If this delayed worker is stopping outside, directly quit here.
1012          * dc->io_disable might be set via sysfs interface, so check it
1013          * here too.
1014          */
1015         while (!kthread_should_stop() && !dc->io_disable) {
1016                 q = bdev_get_queue(dc->bdev);
1017                 if (blk_queue_dying(q))
1018                         dc->offline_seconds++;
1019                 else
1020                         dc->offline_seconds = 0;
1021
1022                 if (dc->offline_seconds >= BACKING_DEV_OFFLINE_TIMEOUT) {
1023                         pr_err("%pg: device offline for %d seconds\n",
1024                                dc->bdev,
1025                                BACKING_DEV_OFFLINE_TIMEOUT);
1026                         pr_err("%s: disable I/O request due to backing device offline\n",
1027                                dc->disk.name);
1028                         dc->io_disable = true;
1029                         /* let others know earlier that io_disable is true */
1030                         smp_mb();
1031                         bcache_device_stop(&dc->disk);
1032                         break;
1033                 }
1034                 schedule_timeout_interruptible(HZ);
1035         }
1036
1037         wait_for_kthread_stop();
1038         return 0;
1039 }
1040
1041
1042 int bch_cached_dev_run(struct cached_dev *dc)
1043 {
1044         int ret = 0;
1045         struct bcache_device *d = &dc->disk;
1046         char *buf = kmemdup_nul(dc->sb.label, SB_LABEL_SIZE, GFP_KERNEL);
1047         char *env[] = {
1048                 "DRIVER=bcache",
1049                 kasprintf(GFP_KERNEL, "CACHED_UUID=%pU", dc->sb.uuid),
1050                 kasprintf(GFP_KERNEL, "CACHED_LABEL=%s", buf ? : ""),
1051                 NULL,
1052         };
1053
1054         if (dc->io_disable) {
1055                 pr_err("I/O disabled on cached dev %pg\n", dc->bdev);
1056                 ret = -EIO;
1057                 goto out;
1058         }
1059
1060         if (atomic_xchg(&dc->running, 1)) {
1061                 pr_info("cached dev %pg is running already\n", dc->bdev);
1062                 ret = -EBUSY;
1063                 goto out;
1064         }
1065
1066         if (!d->c &&
1067             BDEV_STATE(&dc->sb) != BDEV_STATE_NONE) {
1068                 struct closure cl;
1069
1070                 closure_init_stack(&cl);
1071
1072                 SET_BDEV_STATE(&dc->sb, BDEV_STATE_STALE);
1073                 bch_write_bdev_super(dc, &cl);
1074                 closure_sync(&cl);
1075         }
1076
1077         ret = add_disk(d->disk);
1078         if (ret)
1079                 goto out;
1080         bd_link_disk_holder(dc->bdev, dc->disk.disk);
1081         /*
1082          * won't show up in the uevent file, use udevadm monitor -e instead
1083          * only class / kset properties are persistent
1084          */
1085         kobject_uevent_env(&disk_to_dev(d->disk)->kobj, KOBJ_CHANGE, env);
1086
1087         if (sysfs_create_link(&d->kobj, &disk_to_dev(d->disk)->kobj, "dev") ||
1088             sysfs_create_link(&disk_to_dev(d->disk)->kobj,
1089                               &d->kobj, "bcache")) {
1090                 pr_err("Couldn't create bcache dev <-> disk sysfs symlinks\n");
1091                 ret = -ENOMEM;
1092                 goto out;
1093         }
1094
1095         dc->status_update_thread = kthread_run(cached_dev_status_update,
1096                                                dc, "bcache_status_update");
1097         if (IS_ERR(dc->status_update_thread)) {
1098                 pr_warn("failed to create bcache_status_update kthread, continue to run without monitoring backing device status\n");
1099         }
1100
1101 out:
1102         kfree(env[1]);
1103         kfree(env[2]);
1104         kfree(buf);
1105         return ret;
1106 }
1107
1108 /*
1109  * If BCACHE_DEV_RATE_DW_RUNNING is set, it means routine of the delayed
1110  * work dc->writeback_rate_update is running. Wait until the routine
1111  * quits (BCACHE_DEV_RATE_DW_RUNNING is clear), then continue to
1112  * cancel it. If BCACHE_DEV_RATE_DW_RUNNING is not clear after time_out
1113  * seconds, give up waiting here and continue to cancel it too.
1114  */
1115 static void cancel_writeback_rate_update_dwork(struct cached_dev *dc)
1116 {
1117         int time_out = WRITEBACK_RATE_UPDATE_SECS_MAX * HZ;
1118
1119         do {
1120                 if (!test_bit(BCACHE_DEV_RATE_DW_RUNNING,
1121                               &dc->disk.flags))
1122                         break;
1123                 time_out--;
1124                 schedule_timeout_interruptible(1);
1125         } while (time_out > 0);
1126
1127         if (time_out == 0)
1128                 pr_warn("give up waiting for dc->writeback_write_update to quit\n");
1129
1130         cancel_delayed_work_sync(&dc->writeback_rate_update);
1131 }
1132
1133 static void cached_dev_detach_finish(struct work_struct *w)
1134 {
1135         struct cached_dev *dc = container_of(w, struct cached_dev, detach);
1136         struct cache_set *c = dc->disk.c;
1137
1138         BUG_ON(!test_bit(BCACHE_DEV_DETACHING, &dc->disk.flags));
1139         BUG_ON(refcount_read(&dc->count));
1140
1141
1142         if (test_and_clear_bit(BCACHE_DEV_WB_RUNNING, &dc->disk.flags))
1143                 cancel_writeback_rate_update_dwork(dc);
1144
1145         if (!IS_ERR_OR_NULL(dc->writeback_thread)) {
1146                 kthread_stop(dc->writeback_thread);
1147                 dc->writeback_thread = NULL;
1148         }
1149
1150         mutex_lock(&bch_register_lock);
1151
1152         bcache_device_detach(&dc->disk);
1153         list_move(&dc->list, &uncached_devices);
1154         calc_cached_dev_sectors(c);
1155
1156         clear_bit(BCACHE_DEV_DETACHING, &dc->disk.flags);
1157         clear_bit(BCACHE_DEV_UNLINK_DONE, &dc->disk.flags);
1158
1159         mutex_unlock(&bch_register_lock);
1160
1161         pr_info("Caching disabled for %pg\n", dc->bdev);
1162
1163         /* Drop ref we took in cached_dev_detach() */
1164         closure_put(&dc->disk.cl);
1165 }
1166
1167 void bch_cached_dev_detach(struct cached_dev *dc)
1168 {
1169         lockdep_assert_held(&bch_register_lock);
1170
1171         if (test_bit(BCACHE_DEV_CLOSING, &dc->disk.flags))
1172                 return;
1173
1174         if (test_and_set_bit(BCACHE_DEV_DETACHING, &dc->disk.flags))
1175                 return;
1176
1177         /*
1178          * Block the device from being closed and freed until we're finished
1179          * detaching
1180          */
1181         closure_get(&dc->disk.cl);
1182
1183         bch_writeback_queue(dc);
1184
1185         cached_dev_put(dc);
1186 }
1187
1188 int bch_cached_dev_attach(struct cached_dev *dc, struct cache_set *c,
1189                           uint8_t *set_uuid)
1190 {
1191         uint32_t rtime = cpu_to_le32((u32)ktime_get_real_seconds());
1192         struct uuid_entry *u;
1193         struct cached_dev *exist_dc, *t;
1194         int ret = 0;
1195
1196         if ((set_uuid && memcmp(set_uuid, c->set_uuid, 16)) ||
1197             (!set_uuid && memcmp(dc->sb.set_uuid, c->set_uuid, 16)))
1198                 return -ENOENT;
1199
1200         if (dc->disk.c) {
1201                 pr_err("Can't attach %pg: already attached\n", dc->bdev);
1202                 return -EINVAL;
1203         }
1204
1205         if (test_bit(CACHE_SET_STOPPING, &c->flags)) {
1206                 pr_err("Can't attach %pg: shutting down\n", dc->bdev);
1207                 return -EINVAL;
1208         }
1209
1210         if (dc->sb.block_size < c->cache->sb.block_size) {
1211                 /* Will die */
1212                 pr_err("Couldn't attach %pg: block size less than set's block size\n",
1213                        dc->bdev);
1214                 return -EINVAL;
1215         }
1216
1217         /* Check whether already attached */
1218         list_for_each_entry_safe(exist_dc, t, &c->cached_devs, list) {
1219                 if (!memcmp(dc->sb.uuid, exist_dc->sb.uuid, 16)) {
1220                         pr_err("Tried to attach %pg but duplicate UUID already attached\n",
1221                                 dc->bdev);
1222
1223                         return -EINVAL;
1224                 }
1225         }
1226
1227         u = uuid_find(c, dc->sb.uuid);
1228
1229         if (u &&
1230             (BDEV_STATE(&dc->sb) == BDEV_STATE_STALE ||
1231              BDEV_STATE(&dc->sb) == BDEV_STATE_NONE)) {
1232                 memcpy(u->uuid, invalid_uuid, 16);
1233                 u->invalidated = cpu_to_le32((u32)ktime_get_real_seconds());
1234                 u = NULL;
1235         }
1236
1237         if (!u) {
1238                 if (BDEV_STATE(&dc->sb) == BDEV_STATE_DIRTY) {
1239                         pr_err("Couldn't find uuid for %pg in set\n", dc->bdev);
1240                         return -ENOENT;
1241                 }
1242
1243                 u = uuid_find_empty(c);
1244                 if (!u) {
1245                         pr_err("Not caching %pg, no room for UUID\n", dc->bdev);
1246                         return -EINVAL;
1247                 }
1248         }
1249
1250         /*
1251          * Deadlocks since we're called via sysfs...
1252          * sysfs_remove_file(&dc->kobj, &sysfs_attach);
1253          */
1254
1255         if (bch_is_zero(u->uuid, 16)) {
1256                 struct closure cl;
1257
1258                 closure_init_stack(&cl);
1259
1260                 memcpy(u->uuid, dc->sb.uuid, 16);
1261                 memcpy(u->label, dc->sb.label, SB_LABEL_SIZE);
1262                 u->first_reg = u->last_reg = rtime;
1263                 bch_uuid_write(c);
1264
1265                 memcpy(dc->sb.set_uuid, c->set_uuid, 16);
1266                 SET_BDEV_STATE(&dc->sb, BDEV_STATE_CLEAN);
1267
1268                 bch_write_bdev_super(dc, &cl);
1269                 closure_sync(&cl);
1270         } else {
1271                 u->last_reg = rtime;
1272                 bch_uuid_write(c);
1273         }
1274
1275         bcache_device_attach(&dc->disk, c, u - c->uuids);
1276         list_move(&dc->list, &c->cached_devs);
1277         calc_cached_dev_sectors(c);
1278
1279         /*
1280          * dc->c must be set before dc->count != 0 - paired with the mb in
1281          * cached_dev_get()
1282          */
1283         smp_wmb();
1284         refcount_set(&dc->count, 1);
1285
1286         /* Block writeback thread, but spawn it */
1287         down_write(&dc->writeback_lock);
1288         if (bch_cached_dev_writeback_start(dc)) {
1289                 up_write(&dc->writeback_lock);
1290                 pr_err("Couldn't start writeback facilities for %s\n",
1291                        dc->disk.disk->disk_name);
1292                 return -ENOMEM;
1293         }
1294
1295         if (BDEV_STATE(&dc->sb) == BDEV_STATE_DIRTY) {
1296                 atomic_set(&dc->has_dirty, 1);
1297                 bch_writeback_queue(dc);
1298         }
1299
1300         bch_sectors_dirty_init(&dc->disk);
1301
1302         ret = bch_cached_dev_run(dc);
1303         if (ret && (ret != -EBUSY)) {
1304                 up_write(&dc->writeback_lock);
1305                 /*
1306                  * bch_register_lock is held, bcache_device_stop() is not
1307                  * able to be directly called. The kthread and kworker
1308                  * created previously in bch_cached_dev_writeback_start()
1309                  * have to be stopped manually here.
1310                  */
1311                 kthread_stop(dc->writeback_thread);
1312                 cancel_writeback_rate_update_dwork(dc);
1313                 pr_err("Couldn't run cached device %pg\n", dc->bdev);
1314                 return ret;
1315         }
1316
1317         bcache_device_link(&dc->disk, c, "bdev");
1318         atomic_inc(&c->attached_dev_nr);
1319
1320         if (bch_has_feature_obso_large_bucket(&(c->cache->sb))) {
1321                 pr_err("The obsoleted large bucket layout is unsupported, set the bcache device into read-only\n");
1322                 pr_err("Please update to the latest bcache-tools to create the cache device\n");
1323                 set_disk_ro(dc->disk.disk, 1);
1324         }
1325
1326         /* Allow the writeback thread to proceed */
1327         up_write(&dc->writeback_lock);
1328
1329         pr_info("Caching %pg as %s on set %pU\n",
1330                 dc->bdev,
1331                 dc->disk.disk->disk_name,
1332                 dc->disk.c->set_uuid);
1333         return 0;
1334 }
1335
1336 /* when dc->disk.kobj released */
1337 void bch_cached_dev_release(struct kobject *kobj)
1338 {
1339         struct cached_dev *dc = container_of(kobj, struct cached_dev,
1340                                              disk.kobj);
1341         kfree(dc);
1342         module_put(THIS_MODULE);
1343 }
1344
1345 static CLOSURE_CALLBACK(cached_dev_free)
1346 {
1347         closure_type(dc, struct cached_dev, disk.cl);
1348
1349         if (test_and_clear_bit(BCACHE_DEV_WB_RUNNING, &dc->disk.flags))
1350                 cancel_writeback_rate_update_dwork(dc);
1351
1352         if (!IS_ERR_OR_NULL(dc->writeback_thread))
1353                 kthread_stop(dc->writeback_thread);
1354         if (!IS_ERR_OR_NULL(dc->status_update_thread))
1355                 kthread_stop(dc->status_update_thread);
1356
1357         mutex_lock(&bch_register_lock);
1358
1359         if (atomic_read(&dc->running)) {
1360                 bd_unlink_disk_holder(dc->bdev, dc->disk.disk);
1361                 del_gendisk(dc->disk.disk);
1362         }
1363         bcache_device_free(&dc->disk);
1364         list_del(&dc->list);
1365
1366         mutex_unlock(&bch_register_lock);
1367
1368         if (dc->sb_disk)
1369                 put_page(virt_to_page(dc->sb_disk));
1370
1371         if (dc->bdev_file)
1372                 fput(dc->bdev_file);
1373
1374         wake_up(&unregister_wait);
1375
1376         kobject_put(&dc->disk.kobj);
1377 }
1378
1379 static CLOSURE_CALLBACK(cached_dev_flush)
1380 {
1381         closure_type(dc, struct cached_dev, disk.cl);
1382         struct bcache_device *d = &dc->disk;
1383
1384         mutex_lock(&bch_register_lock);
1385         bcache_device_unlink(d);
1386         mutex_unlock(&bch_register_lock);
1387
1388         bch_cache_accounting_destroy(&dc->accounting);
1389         kobject_del(&d->kobj);
1390
1391         continue_at(cl, cached_dev_free, system_wq);
1392 }
1393
1394 static int cached_dev_init(struct cached_dev *dc, unsigned int block_size)
1395 {
1396         int ret;
1397         struct io *io;
1398         struct request_queue *q = bdev_get_queue(dc->bdev);
1399
1400         __module_get(THIS_MODULE);
1401         INIT_LIST_HEAD(&dc->list);
1402         closure_init(&dc->disk.cl, NULL);
1403         set_closure_fn(&dc->disk.cl, cached_dev_flush, system_wq);
1404         kobject_init(&dc->disk.kobj, &bch_cached_dev_ktype);
1405         INIT_WORK(&dc->detach, cached_dev_detach_finish);
1406         sema_init(&dc->sb_write_mutex, 1);
1407         INIT_LIST_HEAD(&dc->io_lru);
1408         spin_lock_init(&dc->io_lock);
1409         bch_cache_accounting_init(&dc->accounting, &dc->disk.cl);
1410
1411         dc->sequential_cutoff           = 4 << 20;
1412
1413         for (io = dc->io; io < dc->io + RECENT_IO; io++) {
1414                 list_add(&io->lru, &dc->io_lru);
1415                 hlist_add_head(&io->hash, dc->io_hash + RECENT_IO);
1416         }
1417
1418         if (bdev_io_opt(dc->bdev))
1419                 dc->partial_stripes_expensive = !!(q->limits.features &
1420                         BLK_FEAT_RAID_PARTIAL_STRIPES_EXPENSIVE);
1421
1422         ret = bcache_device_init(&dc->disk, block_size,
1423                          bdev_nr_sectors(dc->bdev) - dc->sb.data_offset,
1424                          dc->bdev, &bcache_cached_ops);
1425         if (ret)
1426                 return ret;
1427
1428         atomic_set(&dc->io_errors, 0);
1429         dc->io_disable = false;
1430         dc->error_limit = DEFAULT_CACHED_DEV_ERROR_LIMIT;
1431         /* default to auto */
1432         dc->stop_when_cache_set_failed = BCH_CACHED_DEV_STOP_AUTO;
1433
1434         bch_cached_dev_request_init(dc);
1435         bch_cached_dev_writeback_init(dc);
1436         return 0;
1437 }
1438
1439 /* Cached device - bcache superblock */
1440
1441 static int register_bdev(struct cache_sb *sb, struct cache_sb_disk *sb_disk,
1442                                  struct file *bdev_file,
1443                                  struct cached_dev *dc)
1444 {
1445         const char *err = "cannot allocate memory";
1446         struct cache_set *c;
1447         int ret = -ENOMEM;
1448
1449         memcpy(&dc->sb, sb, sizeof(struct cache_sb));
1450         dc->bdev_file = bdev_file;
1451         dc->bdev = file_bdev(bdev_file);
1452         dc->sb_disk = sb_disk;
1453
1454         if (cached_dev_init(dc, sb->block_size << 9))
1455                 goto err;
1456
1457         err = "error creating kobject";
1458         if (kobject_add(&dc->disk.kobj, bdev_kobj(dc->bdev), "bcache"))
1459                 goto err;
1460         if (bch_cache_accounting_add_kobjs(&dc->accounting, &dc->disk.kobj))
1461                 goto err;
1462
1463         pr_info("registered backing device %pg\n", dc->bdev);
1464
1465         list_add(&dc->list, &uncached_devices);
1466         /* attach to a matched cache set if it exists */
1467         list_for_each_entry(c, &bch_cache_sets, list)
1468                 bch_cached_dev_attach(dc, c, NULL);
1469
1470         if (BDEV_STATE(&dc->sb) == BDEV_STATE_NONE ||
1471             BDEV_STATE(&dc->sb) == BDEV_STATE_STALE) {
1472                 err = "failed to run cached device";
1473                 ret = bch_cached_dev_run(dc);
1474                 if (ret)
1475                         goto err;
1476         }
1477
1478         return 0;
1479 err:
1480         pr_notice("error %pg: %s\n", dc->bdev, err);
1481         bcache_device_stop(&dc->disk);
1482         return ret;
1483 }
1484
1485 /* Flash only volumes */
1486
1487 /* When d->kobj released */
1488 void bch_flash_dev_release(struct kobject *kobj)
1489 {
1490         struct bcache_device *d = container_of(kobj, struct bcache_device,
1491                                                kobj);
1492         kfree(d);
1493 }
1494
1495 static CLOSURE_CALLBACK(flash_dev_free)
1496 {
1497         closure_type(d, struct bcache_device, cl);
1498
1499         mutex_lock(&bch_register_lock);
1500         atomic_long_sub(bcache_dev_sectors_dirty(d),
1501                         &d->c->flash_dev_dirty_sectors);
1502         del_gendisk(d->disk);
1503         bcache_device_free(d);
1504         mutex_unlock(&bch_register_lock);
1505         kobject_put(&d->kobj);
1506 }
1507
1508 static CLOSURE_CALLBACK(flash_dev_flush)
1509 {
1510         closure_type(d, struct bcache_device, cl);
1511
1512         mutex_lock(&bch_register_lock);
1513         bcache_device_unlink(d);
1514         mutex_unlock(&bch_register_lock);
1515         kobject_del(&d->kobj);
1516         continue_at(cl, flash_dev_free, system_wq);
1517 }
1518
1519 static int flash_dev_run(struct cache_set *c, struct uuid_entry *u)
1520 {
1521         int err = -ENOMEM;
1522         struct bcache_device *d = kzalloc(sizeof(struct bcache_device),
1523                                           GFP_KERNEL);
1524         if (!d)
1525                 goto err_ret;
1526
1527         closure_init(&d->cl, NULL);
1528         set_closure_fn(&d->cl, flash_dev_flush, system_wq);
1529
1530         kobject_init(&d->kobj, &bch_flash_dev_ktype);
1531
1532         if (bcache_device_init(d, block_bytes(c->cache), u->sectors,
1533                         NULL, &bcache_flash_ops))
1534                 goto err;
1535
1536         bcache_device_attach(d, c, u - c->uuids);
1537         bch_sectors_dirty_init(d);
1538         bch_flash_dev_request_init(d);
1539         err = add_disk(d->disk);
1540         if (err)
1541                 goto err;
1542
1543         err = kobject_add(&d->kobj, &disk_to_dev(d->disk)->kobj, "bcache");
1544         if (err)
1545                 goto err;
1546
1547         bcache_device_link(d, c, "volume");
1548
1549         if (bch_has_feature_obso_large_bucket(&c->cache->sb)) {
1550                 pr_err("The obsoleted large bucket layout is unsupported, set the bcache device into read-only\n");
1551                 pr_err("Please update to the latest bcache-tools to create the cache device\n");
1552                 set_disk_ro(d->disk, 1);
1553         }
1554
1555         return 0;
1556 err:
1557         kobject_put(&d->kobj);
1558 err_ret:
1559         return err;
1560 }
1561
1562 static int flash_devs_run(struct cache_set *c)
1563 {
1564         int ret = 0;
1565         struct uuid_entry *u;
1566
1567         for (u = c->uuids;
1568              u < c->uuids + c->nr_uuids && !ret;
1569              u++)
1570                 if (UUID_FLASH_ONLY(u))
1571                         ret = flash_dev_run(c, u);
1572
1573         return ret;
1574 }
1575
1576 int bch_flash_dev_create(struct cache_set *c, uint64_t size)
1577 {
1578         struct uuid_entry *u;
1579
1580         if (test_bit(CACHE_SET_STOPPING, &c->flags))
1581                 return -EINTR;
1582
1583         if (!test_bit(CACHE_SET_RUNNING, &c->flags))
1584                 return -EPERM;
1585
1586         u = uuid_find_empty(c);
1587         if (!u) {
1588                 pr_err("Can't create volume, no room for UUID\n");
1589                 return -EINVAL;
1590         }
1591
1592         get_random_bytes(u->uuid, 16);
1593         memset(u->label, 0, 32);
1594         u->first_reg = u->last_reg = cpu_to_le32((u32)ktime_get_real_seconds());
1595
1596         SET_UUID_FLASH_ONLY(u, 1);
1597         u->sectors = size >> 9;
1598
1599         bch_uuid_write(c);
1600
1601         return flash_dev_run(c, u);
1602 }
1603
1604 bool bch_cached_dev_error(struct cached_dev *dc)
1605 {
1606         if (!dc || test_bit(BCACHE_DEV_CLOSING, &dc->disk.flags))
1607                 return false;
1608
1609         dc->io_disable = true;
1610         /* make others know io_disable is true earlier */
1611         smp_mb();
1612
1613         pr_err("stop %s: too many IO errors on backing device %pg\n",
1614                dc->disk.disk->disk_name, dc->bdev);
1615
1616         bcache_device_stop(&dc->disk);
1617         return true;
1618 }
1619
1620 /* Cache set */
1621
1622 __printf(2, 3)
1623 bool bch_cache_set_error(struct cache_set *c, const char *fmt, ...)
1624 {
1625         struct va_format vaf;
1626         va_list args;
1627
1628         if (c->on_error != ON_ERROR_PANIC &&
1629             test_bit(CACHE_SET_STOPPING, &c->flags))
1630                 return false;
1631
1632         if (test_and_set_bit(CACHE_SET_IO_DISABLE, &c->flags))
1633                 pr_info("CACHE_SET_IO_DISABLE already set\n");
1634
1635         /*
1636          * XXX: we can be called from atomic context
1637          * acquire_console_sem();
1638          */
1639
1640         va_start(args, fmt);
1641
1642         vaf.fmt = fmt;
1643         vaf.va = &args;
1644
1645         pr_err("error on %pU: %pV, disabling caching\n",
1646                c->set_uuid, &vaf);
1647
1648         va_end(args);
1649
1650         if (c->on_error == ON_ERROR_PANIC)
1651                 panic("panic forced after error\n");
1652
1653         bch_cache_set_unregister(c);
1654         return true;
1655 }
1656
1657 /* When c->kobj released */
1658 void bch_cache_set_release(struct kobject *kobj)
1659 {
1660         struct cache_set *c = container_of(kobj, struct cache_set, kobj);
1661
1662         kfree(c);
1663         module_put(THIS_MODULE);
1664 }
1665
1666 static CLOSURE_CALLBACK(cache_set_free)
1667 {
1668         closure_type(c, struct cache_set, cl);
1669         struct cache *ca;
1670
1671         debugfs_remove(c->debug);
1672
1673         bch_open_buckets_free(c);
1674         bch_btree_cache_free(c);
1675         bch_journal_free(c);
1676
1677         mutex_lock(&bch_register_lock);
1678         bch_bset_sort_state_free(&c->sort);
1679         free_pages((unsigned long) c->uuids, ilog2(meta_bucket_pages(&c->cache->sb)));
1680
1681         ca = c->cache;
1682         if (ca) {
1683                 ca->set = NULL;
1684                 c->cache = NULL;
1685                 kobject_put(&ca->kobj);
1686         }
1687
1688
1689         if (c->moving_gc_wq)
1690                 destroy_workqueue(c->moving_gc_wq);
1691         bioset_exit(&c->bio_split);
1692         mempool_exit(&c->fill_iter);
1693         mempool_exit(&c->bio_meta);
1694         mempool_exit(&c->search);
1695         kfree(c->devices);
1696
1697         list_del(&c->list);
1698         mutex_unlock(&bch_register_lock);
1699
1700         pr_info("Cache set %pU unregistered\n", c->set_uuid);
1701         wake_up(&unregister_wait);
1702
1703         closure_debug_destroy(&c->cl);
1704         kobject_put(&c->kobj);
1705 }
1706
1707 static CLOSURE_CALLBACK(cache_set_flush)
1708 {
1709         closure_type(c, struct cache_set, caching);
1710         struct cache *ca = c->cache;
1711         struct btree *b;
1712
1713         bch_cache_accounting_destroy(&c->accounting);
1714
1715         kobject_put(&c->internal);
1716         kobject_del(&c->kobj);
1717
1718         if (!IS_ERR_OR_NULL(c->gc_thread))
1719                 kthread_stop(c->gc_thread);
1720
1721         if (!IS_ERR_OR_NULL(c->root))
1722                 list_add(&c->root->list, &c->btree_cache);
1723
1724         /*
1725          * Avoid flushing cached nodes if cache set is retiring
1726          * due to too many I/O errors detected.
1727          */
1728         if (!test_bit(CACHE_SET_IO_DISABLE, &c->flags))
1729                 list_for_each_entry(b, &c->btree_cache, list) {
1730                         mutex_lock(&b->write_lock);
1731                         if (btree_node_dirty(b))
1732                                 __bch_btree_node_write(b, NULL);
1733                         mutex_unlock(&b->write_lock);
1734                 }
1735
1736         if (ca->alloc_thread)
1737                 kthread_stop(ca->alloc_thread);
1738
1739         if (c->journal.cur) {
1740                 cancel_delayed_work_sync(&c->journal.work);
1741                 /* flush last journal entry if needed */
1742                 c->journal.work.work.func(&c->journal.work.work);
1743         }
1744
1745         closure_return(cl);
1746 }
1747
1748 /*
1749  * This function is only called when CACHE_SET_IO_DISABLE is set, which means
1750  * cache set is unregistering due to too many I/O errors. In this condition,
1751  * the bcache device might be stopped, it depends on stop_when_cache_set_failed
1752  * value and whether the broken cache has dirty data:
1753  *
1754  * dc->stop_when_cache_set_failed    dc->has_dirty   stop bcache device
1755  *  BCH_CACHED_STOP_AUTO               0               NO
1756  *  BCH_CACHED_STOP_AUTO               1               YES
1757  *  BCH_CACHED_DEV_STOP_ALWAYS         0               YES
1758  *  BCH_CACHED_DEV_STOP_ALWAYS         1               YES
1759  *
1760  * The expected behavior is, if stop_when_cache_set_failed is configured to
1761  * "auto" via sysfs interface, the bcache device will not be stopped if the
1762  * backing device is clean on the broken cache device.
1763  */
1764 static void conditional_stop_bcache_device(struct cache_set *c,
1765                                            struct bcache_device *d,
1766                                            struct cached_dev *dc)
1767 {
1768         if (dc->stop_when_cache_set_failed == BCH_CACHED_DEV_STOP_ALWAYS) {
1769                 pr_warn("stop_when_cache_set_failed of %s is \"always\", stop it for failed cache set %pU.\n",
1770                         d->disk->disk_name, c->set_uuid);
1771                 bcache_device_stop(d);
1772         } else if (atomic_read(&dc->has_dirty)) {
1773                 /*
1774                  * dc->stop_when_cache_set_failed == BCH_CACHED_STOP_AUTO
1775                  * and dc->has_dirty == 1
1776                  */
1777                 pr_warn("stop_when_cache_set_failed of %s is \"auto\" and cache is dirty, stop it to avoid potential data corruption.\n",
1778                         d->disk->disk_name);
1779                 /*
1780                  * There might be a small time gap that cache set is
1781                  * released but bcache device is not. Inside this time
1782                  * gap, regular I/O requests will directly go into
1783                  * backing device as no cache set attached to. This
1784                  * behavior may also introduce potential inconsistence
1785                  * data in writeback mode while cache is dirty.
1786                  * Therefore before calling bcache_device_stop() due
1787                  * to a broken cache device, dc->io_disable should be
1788                  * explicitly set to true.
1789                  */
1790                 dc->io_disable = true;
1791                 /* make others know io_disable is true earlier */
1792                 smp_mb();
1793                 bcache_device_stop(d);
1794         } else {
1795                 /*
1796                  * dc->stop_when_cache_set_failed == BCH_CACHED_STOP_AUTO
1797                  * and dc->has_dirty == 0
1798                  */
1799                 pr_warn("stop_when_cache_set_failed of %s is \"auto\" and cache is clean, keep it alive.\n",
1800                         d->disk->disk_name);
1801         }
1802 }
1803
1804 static CLOSURE_CALLBACK(__cache_set_unregister)
1805 {
1806         closure_type(c, struct cache_set, caching);
1807         struct cached_dev *dc;
1808         struct bcache_device *d;
1809         size_t i;
1810
1811         mutex_lock(&bch_register_lock);
1812
1813         for (i = 0; i < c->devices_max_used; i++) {
1814                 d = c->devices[i];
1815                 if (!d)
1816                         continue;
1817
1818                 if (!UUID_FLASH_ONLY(&c->uuids[i]) &&
1819                     test_bit(CACHE_SET_UNREGISTERING, &c->flags)) {
1820                         dc = container_of(d, struct cached_dev, disk);
1821                         bch_cached_dev_detach(dc);
1822                         if (test_bit(CACHE_SET_IO_DISABLE, &c->flags))
1823                                 conditional_stop_bcache_device(c, d, dc);
1824                 } else {
1825                         bcache_device_stop(d);
1826                 }
1827         }
1828
1829         mutex_unlock(&bch_register_lock);
1830
1831         continue_at(cl, cache_set_flush, system_wq);
1832 }
1833
1834 void bch_cache_set_stop(struct cache_set *c)
1835 {
1836         if (!test_and_set_bit(CACHE_SET_STOPPING, &c->flags))
1837                 /* closure_fn set to __cache_set_unregister() */
1838                 closure_queue(&c->caching);
1839 }
1840
1841 void bch_cache_set_unregister(struct cache_set *c)
1842 {
1843         set_bit(CACHE_SET_UNREGISTERING, &c->flags);
1844         bch_cache_set_stop(c);
1845 }
1846
1847 #define alloc_meta_bucket_pages(gfp, sb)                \
1848         ((void *) __get_free_pages(__GFP_ZERO|__GFP_COMP|gfp, ilog2(meta_bucket_pages(sb))))
1849
1850 struct cache_set *bch_cache_set_alloc(struct cache_sb *sb)
1851 {
1852         int iter_size;
1853         struct cache *ca = container_of(sb, struct cache, sb);
1854         struct cache_set *c = kzalloc(sizeof(struct cache_set), GFP_KERNEL);
1855
1856         if (!c)
1857                 return NULL;
1858
1859         __module_get(THIS_MODULE);
1860         closure_init(&c->cl, NULL);
1861         set_closure_fn(&c->cl, cache_set_free, system_wq);
1862
1863         closure_init(&c->caching, &c->cl);
1864         set_closure_fn(&c->caching, __cache_set_unregister, system_wq);
1865
1866         /* Maybe create continue_at_noreturn() and use it here? */
1867         closure_set_stopped(&c->cl);
1868         closure_put(&c->cl);
1869
1870         kobject_init(&c->kobj, &bch_cache_set_ktype);
1871         kobject_init(&c->internal, &bch_cache_set_internal_ktype);
1872
1873         bch_cache_accounting_init(&c->accounting, &c->cl);
1874
1875         memcpy(c->set_uuid, sb->set_uuid, 16);
1876
1877         c->cache                = ca;
1878         c->cache->set           = c;
1879         c->bucket_bits          = ilog2(sb->bucket_size);
1880         c->block_bits           = ilog2(sb->block_size);
1881         c->nr_uuids             = meta_bucket_bytes(sb) / sizeof(struct uuid_entry);
1882         c->devices_max_used     = 0;
1883         atomic_set(&c->attached_dev_nr, 0);
1884         c->btree_pages          = meta_bucket_pages(sb);
1885         if (c->btree_pages > BTREE_MAX_PAGES)
1886                 c->btree_pages = max_t(int, c->btree_pages / 4,
1887                                        BTREE_MAX_PAGES);
1888
1889         sema_init(&c->sb_write_mutex, 1);
1890         mutex_init(&c->bucket_lock);
1891         init_waitqueue_head(&c->btree_cache_wait);
1892         spin_lock_init(&c->btree_cannibalize_lock);
1893         init_waitqueue_head(&c->bucket_wait);
1894         init_waitqueue_head(&c->gc_wait);
1895         sema_init(&c->uuid_write_mutex, 1);
1896
1897         spin_lock_init(&c->btree_gc_time.lock);
1898         spin_lock_init(&c->btree_split_time.lock);
1899         spin_lock_init(&c->btree_read_time.lock);
1900
1901         bch_moving_init_cache_set(c);
1902
1903         INIT_LIST_HEAD(&c->list);
1904         INIT_LIST_HEAD(&c->cached_devs);
1905         INIT_LIST_HEAD(&c->btree_cache);
1906         INIT_LIST_HEAD(&c->btree_cache_freeable);
1907         INIT_LIST_HEAD(&c->btree_cache_freed);
1908         INIT_LIST_HEAD(&c->data_buckets);
1909
1910         iter_size = ((meta_bucket_pages(sb) * PAGE_SECTORS) / sb->block_size) *
1911                             sizeof(struct btree_iter_set);
1912
1913         c->devices = kcalloc(c->nr_uuids, sizeof(void *), GFP_KERNEL);
1914         if (!c->devices)
1915                 goto err;
1916
1917         if (mempool_init_slab_pool(&c->search, 32, bch_search_cache))
1918                 goto err;
1919
1920         if (mempool_init_kmalloc_pool(&c->bio_meta, 2,
1921                         sizeof(struct bbio) +
1922                         sizeof(struct bio_vec) * meta_bucket_pages(sb)))
1923                 goto err;
1924
1925         if (mempool_init_kmalloc_pool(&c->fill_iter, 1, iter_size))
1926                 goto err;
1927
1928         if (bioset_init(&c->bio_split, 4, offsetof(struct bbio, bio),
1929                         BIOSET_NEED_RESCUER))
1930                 goto err;
1931
1932         c->uuids = alloc_meta_bucket_pages(GFP_KERNEL, sb);
1933         if (!c->uuids)
1934                 goto err;
1935
1936         c->moving_gc_wq = alloc_workqueue("bcache_gc", WQ_MEM_RECLAIM, 0);
1937         if (!c->moving_gc_wq)
1938                 goto err;
1939
1940         if (bch_journal_alloc(c))
1941                 goto err;
1942
1943         if (bch_btree_cache_alloc(c))
1944                 goto err;
1945
1946         if (bch_open_buckets_alloc(c))
1947                 goto err;
1948
1949         if (bch_bset_sort_state_init(&c->sort, ilog2(c->btree_pages)))
1950                 goto err;
1951
1952         c->congested_read_threshold_us  = 2000;
1953         c->congested_write_threshold_us = 20000;
1954         c->error_limit  = DEFAULT_IO_ERROR_LIMIT;
1955         c->idle_max_writeback_rate_enabled = 1;
1956         WARN_ON(test_and_clear_bit(CACHE_SET_IO_DISABLE, &c->flags));
1957
1958         return c;
1959 err:
1960         bch_cache_set_unregister(c);
1961         return NULL;
1962 }
1963
1964 static int run_cache_set(struct cache_set *c)
1965 {
1966         const char *err = "cannot allocate memory";
1967         struct cached_dev *dc, *t;
1968         struct cache *ca = c->cache;
1969         struct closure cl;
1970         LIST_HEAD(journal);
1971         struct journal_replay *l;
1972
1973         closure_init_stack(&cl);
1974
1975         c->nbuckets = ca->sb.nbuckets;
1976         set_gc_sectors(c);
1977
1978         if (CACHE_SYNC(&c->cache->sb)) {
1979                 struct bkey *k;
1980                 struct jset *j;
1981
1982                 err = "cannot allocate memory for journal";
1983                 if (bch_journal_read(c, &journal))
1984                         goto err;
1985
1986                 pr_debug("btree_journal_read() done\n");
1987
1988                 err = "no journal entries found";
1989                 if (list_empty(&journal))
1990                         goto err;
1991
1992                 j = &list_entry(journal.prev, struct journal_replay, list)->j;
1993
1994                 err = "IO error reading priorities";
1995                 if (prio_read(ca, j->prio_bucket[ca->sb.nr_this_dev]))
1996                         goto err;
1997
1998                 /*
1999                  * If prio_read() fails it'll call cache_set_error and we'll
2000                  * tear everything down right away, but if we perhaps checked
2001                  * sooner we could avoid journal replay.
2002                  */
2003
2004                 k = &j->btree_root;
2005
2006                 err = "bad btree root";
2007                 if (__bch_btree_ptr_invalid(c, k))
2008                         goto err;
2009
2010                 err = "error reading btree root";
2011                 c->root = bch_btree_node_get(c, NULL, k,
2012                                              j->btree_level,
2013                                              true, NULL);
2014                 if (IS_ERR(c->root))
2015                         goto err;
2016
2017                 list_del_init(&c->root->list);
2018                 rw_unlock(true, c->root);
2019
2020                 err = uuid_read(c, j, &cl);
2021                 if (err)
2022                         goto err;
2023
2024                 err = "error in recovery";
2025                 if (bch_btree_check(c))
2026                         goto err;
2027
2028                 bch_journal_mark(c, &journal);
2029                 bch_initial_gc_finish(c);
2030                 pr_debug("btree_check() done\n");
2031
2032                 /*
2033                  * bcache_journal_next() can't happen sooner, or
2034                  * btree_gc_finish() will give spurious errors about last_gc >
2035                  * gc_gen - this is a hack but oh well.
2036                  */
2037                 bch_journal_next(&c->journal);
2038
2039                 err = "error starting allocator thread";
2040                 if (bch_cache_allocator_start(ca))
2041                         goto err;
2042
2043                 /*
2044                  * First place it's safe to allocate: btree_check() and
2045                  * btree_gc_finish() have to run before we have buckets to
2046                  * allocate, and bch_bucket_alloc_set() might cause a journal
2047                  * entry to be written so bcache_journal_next() has to be called
2048                  * first.
2049                  *
2050                  * If the uuids were in the old format we have to rewrite them
2051                  * before the next journal entry is written:
2052                  */
2053                 if (j->version < BCACHE_JSET_VERSION_UUID)
2054                         __uuid_write(c);
2055
2056                 err = "bcache: replay journal failed";
2057                 if (bch_journal_replay(c, &journal))
2058                         goto err;
2059         } else {
2060                 unsigned int j;
2061
2062                 pr_notice("invalidating existing data\n");
2063                 ca->sb.keys = clamp_t(int, ca->sb.nbuckets >> 7,
2064                                         2, SB_JOURNAL_BUCKETS);
2065
2066                 for (j = 0; j < ca->sb.keys; j++)
2067                         ca->sb.d[j] = ca->sb.first_bucket + j;
2068
2069                 bch_initial_gc_finish(c);
2070
2071                 err = "error starting allocator thread";
2072                 if (bch_cache_allocator_start(ca))
2073                         goto err;
2074
2075                 mutex_lock(&c->bucket_lock);
2076                 bch_prio_write(ca, true);
2077                 mutex_unlock(&c->bucket_lock);
2078
2079                 err = "cannot allocate new UUID bucket";
2080                 if (__uuid_write(c))
2081                         goto err;
2082
2083                 err = "cannot allocate new btree root";
2084                 c->root = __bch_btree_node_alloc(c, NULL, 0, true, NULL);
2085                 if (IS_ERR(c->root))
2086                         goto err;
2087
2088                 mutex_lock(&c->root->write_lock);
2089                 bkey_copy_key(&c->root->key, &MAX_KEY);
2090                 bch_btree_node_write(c->root, &cl);
2091                 mutex_unlock(&c->root->write_lock);
2092
2093                 bch_btree_set_root(c->root);
2094                 rw_unlock(true, c->root);
2095
2096                 /*
2097                  * We don't want to write the first journal entry until
2098                  * everything is set up - fortunately journal entries won't be
2099                  * written until the SET_CACHE_SYNC() here:
2100                  */
2101                 SET_CACHE_SYNC(&c->cache->sb, true);
2102
2103                 bch_journal_next(&c->journal);
2104                 bch_journal_meta(c, &cl);
2105         }
2106
2107         err = "error starting gc thread";
2108         if (bch_gc_thread_start(c))
2109                 goto err;
2110
2111         closure_sync(&cl);
2112         c->cache->sb.last_mount = (u32)ktime_get_real_seconds();
2113         bcache_write_super(c);
2114
2115         if (bch_has_feature_obso_large_bucket(&c->cache->sb))
2116                 pr_err("Detect obsoleted large bucket layout, all attached bcache device will be read-only\n");
2117
2118         list_for_each_entry_safe(dc, t, &uncached_devices, list)
2119                 bch_cached_dev_attach(dc, c, NULL);
2120
2121         flash_devs_run(c);
2122
2123         bch_journal_space_reserve(&c->journal);
2124         set_bit(CACHE_SET_RUNNING, &c->flags);
2125         return 0;
2126 err:
2127         while (!list_empty(&journal)) {
2128                 l = list_first_entry(&journal, struct journal_replay, list);
2129                 list_del(&l->list);
2130                 kfree(l);
2131         }
2132
2133         closure_sync(&cl);
2134
2135         bch_cache_set_error(c, "%s", err);
2136
2137         return -EIO;
2138 }
2139
2140 static const char *register_cache_set(struct cache *ca)
2141 {
2142         char buf[12];
2143         const char *err = "cannot allocate memory";
2144         struct cache_set *c;
2145
2146         list_for_each_entry(c, &bch_cache_sets, list)
2147                 if (!memcmp(c->set_uuid, ca->sb.set_uuid, 16)) {
2148                         if (c->cache)
2149                                 return "duplicate cache set member";
2150
2151                         goto found;
2152                 }
2153
2154         c = bch_cache_set_alloc(&ca->sb);
2155         if (!c)
2156                 return err;
2157
2158         err = "error creating kobject";
2159         if (kobject_add(&c->kobj, bcache_kobj, "%pU", c->set_uuid) ||
2160             kobject_add(&c->internal, &c->kobj, "internal"))
2161                 goto err;
2162
2163         if (bch_cache_accounting_add_kobjs(&c->accounting, &c->kobj))
2164                 goto err;
2165
2166         bch_debug_init_cache_set(c);
2167
2168         list_add(&c->list, &bch_cache_sets);
2169 found:
2170         sprintf(buf, "cache%i", ca->sb.nr_this_dev);
2171         if (sysfs_create_link(&ca->kobj, &c->kobj, "set") ||
2172             sysfs_create_link(&c->kobj, &ca->kobj, buf))
2173                 goto err;
2174
2175         kobject_get(&ca->kobj);
2176         ca->set = c;
2177         ca->set->cache = ca;
2178
2179         err = "failed to run cache set";
2180         if (run_cache_set(c) < 0)
2181                 goto err;
2182
2183         return NULL;
2184 err:
2185         bch_cache_set_unregister(c);
2186         return err;
2187 }
2188
2189 /* Cache device */
2190
2191 /* When ca->kobj released */
2192 void bch_cache_release(struct kobject *kobj)
2193 {
2194         struct cache *ca = container_of(kobj, struct cache, kobj);
2195         unsigned int i;
2196
2197         if (ca->set) {
2198                 BUG_ON(ca->set->cache != ca);
2199                 ca->set->cache = NULL;
2200         }
2201
2202         free_pages((unsigned long) ca->disk_buckets, ilog2(meta_bucket_pages(&ca->sb)));
2203         kfree(ca->prio_buckets);
2204         vfree(ca->buckets);
2205
2206         free_heap(&ca->heap);
2207         free_fifo(&ca->free_inc);
2208
2209         for (i = 0; i < RESERVE_NR; i++)
2210                 free_fifo(&ca->free[i]);
2211
2212         if (ca->sb_disk)
2213                 put_page(virt_to_page(ca->sb_disk));
2214
2215         if (ca->bdev_file)
2216                 fput(ca->bdev_file);
2217
2218         kfree(ca);
2219         module_put(THIS_MODULE);
2220 }
2221
2222 static int cache_alloc(struct cache *ca)
2223 {
2224         size_t free;
2225         size_t btree_buckets;
2226         struct bucket *b;
2227         int ret = -ENOMEM;
2228         const char *err = NULL;
2229
2230         __module_get(THIS_MODULE);
2231         kobject_init(&ca->kobj, &bch_cache_ktype);
2232
2233         bio_init(&ca->journal.bio, NULL, ca->journal.bio.bi_inline_vecs, 8, 0);
2234
2235         /*
2236          * when ca->sb.njournal_buckets is not zero, journal exists,
2237          * and in bch_journal_replay(), tree node may split,
2238          * so bucket of RESERVE_BTREE type is needed,
2239          * the worst situation is all journal buckets are valid journal,
2240          * and all the keys need to replay,
2241          * so the number of  RESERVE_BTREE type buckets should be as much
2242          * as journal buckets
2243          */
2244         btree_buckets = ca->sb.njournal_buckets ?: 8;
2245         free = roundup_pow_of_two(ca->sb.nbuckets) >> 10;
2246         if (!free) {
2247                 ret = -EPERM;
2248                 err = "ca->sb.nbuckets is too small";
2249                 goto err_free;
2250         }
2251
2252         if (!init_fifo(&ca->free[RESERVE_BTREE], btree_buckets,
2253                                                 GFP_KERNEL)) {
2254                 err = "ca->free[RESERVE_BTREE] alloc failed";
2255                 goto err_btree_alloc;
2256         }
2257
2258         if (!init_fifo_exact(&ca->free[RESERVE_PRIO], prio_buckets(ca),
2259                                                         GFP_KERNEL)) {
2260                 err = "ca->free[RESERVE_PRIO] alloc failed";
2261                 goto err_prio_alloc;
2262         }
2263
2264         if (!init_fifo(&ca->free[RESERVE_MOVINGGC], free, GFP_KERNEL)) {
2265                 err = "ca->free[RESERVE_MOVINGGC] alloc failed";
2266                 goto err_movinggc_alloc;
2267         }
2268
2269         if (!init_fifo(&ca->free[RESERVE_NONE], free, GFP_KERNEL)) {
2270                 err = "ca->free[RESERVE_NONE] alloc failed";
2271                 goto err_none_alloc;
2272         }
2273
2274         if (!init_fifo(&ca->free_inc, free << 2, GFP_KERNEL)) {
2275                 err = "ca->free_inc alloc failed";
2276                 goto err_free_inc_alloc;
2277         }
2278
2279         if (!init_heap(&ca->heap, free << 3, GFP_KERNEL)) {
2280                 err = "ca->heap alloc failed";
2281                 goto err_heap_alloc;
2282         }
2283
2284         ca->buckets = vzalloc(array_size(sizeof(struct bucket),
2285                               ca->sb.nbuckets));
2286         if (!ca->buckets) {
2287                 err = "ca->buckets alloc failed";
2288                 goto err_buckets_alloc;
2289         }
2290
2291         ca->prio_buckets = kzalloc(array3_size(sizeof(uint64_t),
2292                                    prio_buckets(ca), 2),
2293                                    GFP_KERNEL);
2294         if (!ca->prio_buckets) {
2295                 err = "ca->prio_buckets alloc failed";
2296                 goto err_prio_buckets_alloc;
2297         }
2298
2299         ca->disk_buckets = alloc_meta_bucket_pages(GFP_KERNEL, &ca->sb);
2300         if (!ca->disk_buckets) {
2301                 err = "ca->disk_buckets alloc failed";
2302                 goto err_disk_buckets_alloc;
2303         }
2304
2305         ca->prio_last_buckets = ca->prio_buckets + prio_buckets(ca);
2306
2307         for_each_bucket(b, ca)
2308                 atomic_set(&b->pin, 0);
2309         return 0;
2310
2311 err_disk_buckets_alloc:
2312         kfree(ca->prio_buckets);
2313 err_prio_buckets_alloc:
2314         vfree(ca->buckets);
2315 err_buckets_alloc:
2316         free_heap(&ca->heap);
2317 err_heap_alloc:
2318         free_fifo(&ca->free_inc);
2319 err_free_inc_alloc:
2320         free_fifo(&ca->free[RESERVE_NONE]);
2321 err_none_alloc:
2322         free_fifo(&ca->free[RESERVE_MOVINGGC]);
2323 err_movinggc_alloc:
2324         free_fifo(&ca->free[RESERVE_PRIO]);
2325 err_prio_alloc:
2326         free_fifo(&ca->free[RESERVE_BTREE]);
2327 err_btree_alloc:
2328 err_free:
2329         module_put(THIS_MODULE);
2330         if (err)
2331                 pr_notice("error %pg: %s\n", ca->bdev, err);
2332         return ret;
2333 }
2334
2335 static int register_cache(struct cache_sb *sb, struct cache_sb_disk *sb_disk,
2336                                 struct file *bdev_file,
2337                                 struct cache *ca)
2338 {
2339         const char *err = NULL; /* must be set for any error case */
2340         int ret = 0;
2341
2342         memcpy(&ca->sb, sb, sizeof(struct cache_sb));
2343         ca->bdev_file = bdev_file;
2344         ca->bdev = file_bdev(bdev_file);
2345         ca->sb_disk = sb_disk;
2346
2347         if (bdev_max_discard_sectors(file_bdev(bdev_file)))
2348                 ca->discard = CACHE_DISCARD(&ca->sb);
2349
2350         ret = cache_alloc(ca);
2351         if (ret != 0) {
2352                 if (ret == -ENOMEM)
2353                         err = "cache_alloc(): -ENOMEM";
2354                 else if (ret == -EPERM)
2355                         err = "cache_alloc(): cache device is too small";
2356                 else
2357                         err = "cache_alloc(): unknown error";
2358                 pr_notice("error %pg: %s\n", file_bdev(bdev_file), err);
2359                 /*
2360                  * If we failed here, it means ca->kobj is not initialized yet,
2361                  * kobject_put() won't be called and there is no chance to
2362                  * call fput() to bdev in bch_cache_release(). So
2363                  * we explicitly call fput() on the block device here.
2364                  */
2365                 fput(bdev_file);
2366                 return ret;
2367         }
2368
2369         if (kobject_add(&ca->kobj, bdev_kobj(file_bdev(bdev_file)), "bcache")) {
2370                 pr_notice("error %pg: error calling kobject_add\n",
2371                           file_bdev(bdev_file));
2372                 ret = -ENOMEM;
2373                 goto out;
2374         }
2375
2376         mutex_lock(&bch_register_lock);
2377         err = register_cache_set(ca);
2378         mutex_unlock(&bch_register_lock);
2379
2380         if (err) {
2381                 ret = -ENODEV;
2382                 goto out;
2383         }
2384
2385         pr_info("registered cache device %pg\n", file_bdev(ca->bdev_file));
2386
2387 out:
2388         kobject_put(&ca->kobj);
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 file *bdev_file;
2444         void *holder;
2445 };
2446
2447 static void register_bdev_worker(struct work_struct *work)
2448 {
2449         int fail = false;
2450         struct async_reg_args *args =
2451                 container_of(work, struct async_reg_args, reg_work.work);
2452
2453         mutex_lock(&bch_register_lock);
2454         if (register_bdev(args->sb, args->sb_disk, args->bdev_file,
2455                           args->holder) < 0)
2456                 fail = true;
2457         mutex_unlock(&bch_register_lock);
2458
2459         if (fail)
2460                 pr_info("error %s: fail to register backing device\n",
2461                         args->path);
2462         kfree(args->sb);
2463         kfree(args->path);
2464         kfree(args);
2465         module_put(THIS_MODULE);
2466 }
2467
2468 static void register_cache_worker(struct work_struct *work)
2469 {
2470         int fail = false;
2471         struct async_reg_args *args =
2472                 container_of(work, struct async_reg_args, reg_work.work);
2473
2474         /* blkdev_put() will be called in bch_cache_release() */
2475         if (register_cache(args->sb, args->sb_disk, args->bdev_file,
2476                            args->holder))
2477                 fail = true;
2478
2479         if (fail)
2480                 pr_info("error %s: fail to register cache device\n",
2481                         args->path);
2482         kfree(args->sb);
2483         kfree(args->path);
2484         kfree(args);
2485         module_put(THIS_MODULE);
2486 }
2487
2488 static void register_device_async(struct async_reg_args *args)
2489 {
2490         if (SB_IS_BDEV(args->sb))
2491                 INIT_DELAYED_WORK(&args->reg_work, register_bdev_worker);
2492         else
2493                 INIT_DELAYED_WORK(&args->reg_work, register_cache_worker);
2494
2495         /* 10 jiffies is enough for a delay */
2496         queue_delayed_work(system_wq, &args->reg_work, 10);
2497 }
2498
2499 static void *alloc_holder_object(struct cache_sb *sb)
2500 {
2501         if (SB_IS_BDEV(sb))
2502                 return kzalloc(sizeof(struct cached_dev), GFP_KERNEL);
2503         return kzalloc(sizeof(struct cache), GFP_KERNEL);
2504 }
2505
2506 static ssize_t register_bcache(struct kobject *k, struct kobj_attribute *attr,
2507                                const char *buffer, size_t size)
2508 {
2509         const char *err;
2510         char *path = NULL;
2511         struct cache_sb *sb;
2512         struct cache_sb_disk *sb_disk;
2513         struct file *bdev_file, *bdev_file2;
2514         void *holder = NULL;
2515         ssize_t ret;
2516         bool async_registration = false;
2517         bool quiet = false;
2518
2519 #ifdef CONFIG_BCACHE_ASYNC_REGISTRATION
2520         async_registration = true;
2521 #endif
2522
2523         ret = -EBUSY;
2524         err = "failed to reference bcache module";
2525         if (!try_module_get(THIS_MODULE))
2526                 goto out;
2527
2528         /* For latest state of bcache_is_reboot */
2529         smp_mb();
2530         err = "bcache is in reboot";
2531         if (bcache_is_reboot)
2532                 goto out_module_put;
2533
2534         ret = -ENOMEM;
2535         err = "cannot allocate memory";
2536         path = kstrndup(buffer, size, GFP_KERNEL);
2537         if (!path)
2538                 goto out_module_put;
2539
2540         sb = kmalloc(sizeof(struct cache_sb), GFP_KERNEL);
2541         if (!sb)
2542                 goto out_free_path;
2543
2544         ret = -EINVAL;
2545         err = "failed to open device";
2546         bdev_file = bdev_file_open_by_path(strim(path), BLK_OPEN_READ, NULL, NULL);
2547         if (IS_ERR(bdev_file))
2548                 goto out_free_sb;
2549
2550         err = read_super(sb, file_bdev(bdev_file), &sb_disk);
2551         if (err)
2552                 goto out_blkdev_put;
2553
2554         holder = alloc_holder_object(sb);
2555         if (!holder) {
2556                 ret = -ENOMEM;
2557                 err = "cannot allocate memory";
2558                 goto out_put_sb_page;
2559         }
2560
2561         /* Now reopen in exclusive mode with proper holder */
2562         bdev_file2 = bdev_file_open_by_dev(file_bdev(bdev_file)->bd_dev,
2563                         BLK_OPEN_READ | BLK_OPEN_WRITE, holder, NULL);
2564         fput(bdev_file);
2565         bdev_file = bdev_file2;
2566         if (IS_ERR(bdev_file)) {
2567                 ret = PTR_ERR(bdev_file);
2568                 bdev_file = NULL;
2569                 if (ret == -EBUSY) {
2570                         dev_t dev;
2571
2572                         mutex_lock(&bch_register_lock);
2573                         if (lookup_bdev(strim(path), &dev) == 0 &&
2574                             bch_is_open(dev))
2575                                 err = "device already registered";
2576                         else
2577                                 err = "device busy";
2578                         mutex_unlock(&bch_register_lock);
2579                         if (attr == &ksysfs_register_quiet) {
2580                                 quiet = true;
2581                                 ret = size;
2582                         }
2583                 }
2584                 goto out_free_holder;
2585         }
2586
2587         err = "failed to register device";
2588
2589         if (async_registration) {
2590                 /* register in asynchronous way */
2591                 struct async_reg_args *args =
2592                         kzalloc(sizeof(struct async_reg_args), GFP_KERNEL);
2593
2594                 if (!args) {
2595                         ret = -ENOMEM;
2596                         err = "cannot allocate memory";
2597                         goto out_free_holder;
2598                 }
2599
2600                 args->path      = path;
2601                 args->sb        = sb;
2602                 args->sb_disk   = sb_disk;
2603                 args->bdev_file = bdev_file;
2604                 args->holder    = holder;
2605                 register_device_async(args);
2606                 /* No wait and returns to user space */
2607                 goto async_done;
2608         }
2609
2610         if (SB_IS_BDEV(sb)) {
2611                 mutex_lock(&bch_register_lock);
2612                 ret = register_bdev(sb, sb_disk, bdev_file, holder);
2613                 mutex_unlock(&bch_register_lock);
2614                 /* blkdev_put() will be called in cached_dev_free() */
2615                 if (ret < 0)
2616                         goto out_free_sb;
2617         } else {
2618                 /* blkdev_put() will be called in bch_cache_release() */
2619                 ret = register_cache(sb, sb_disk, bdev_file, holder);
2620                 if (ret)
2621                         goto out_free_sb;
2622         }
2623
2624         kfree(sb);
2625         kfree(path);
2626         module_put(THIS_MODULE);
2627 async_done:
2628         return size;
2629
2630 out_free_holder:
2631         kfree(holder);
2632 out_put_sb_page:
2633         put_page(virt_to_page(sb_disk));
2634 out_blkdev_put:
2635         if (bdev_file)
2636                 fput(bdev_file);
2637 out_free_sb:
2638         kfree(sb);
2639 out_free_path:
2640         kfree(path);
2641         path = NULL;
2642 out_module_put:
2643         module_put(THIS_MODULE);
2644 out:
2645         if (!quiet)
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 acquire
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         if (bch_flush_wq)
2810                 destroy_workqueue(bch_flush_wq);
2811         bch_btree_exit();
2812
2813         if (bcache_major)
2814                 unregister_blkdev(bcache_major, "bcache");
2815         unregister_reboot_notifier(&reboot);
2816         mutex_destroy(&bch_register_lock);
2817 }
2818
2819 /* Check and fixup module parameters */
2820 static void check_module_parameters(void)
2821 {
2822         if (bch_cutoff_writeback_sync == 0)
2823                 bch_cutoff_writeback_sync = CUTOFF_WRITEBACK_SYNC;
2824         else if (bch_cutoff_writeback_sync > CUTOFF_WRITEBACK_SYNC_MAX) {
2825                 pr_warn("set bch_cutoff_writeback_sync (%u) to max value %u\n",
2826                         bch_cutoff_writeback_sync, CUTOFF_WRITEBACK_SYNC_MAX);
2827                 bch_cutoff_writeback_sync = CUTOFF_WRITEBACK_SYNC_MAX;
2828         }
2829
2830         if (bch_cutoff_writeback == 0)
2831                 bch_cutoff_writeback = CUTOFF_WRITEBACK;
2832         else if (bch_cutoff_writeback > CUTOFF_WRITEBACK_MAX) {
2833                 pr_warn("set bch_cutoff_writeback (%u) to max value %u\n",
2834                         bch_cutoff_writeback, CUTOFF_WRITEBACK_MAX);
2835                 bch_cutoff_writeback = CUTOFF_WRITEBACK_MAX;
2836         }
2837
2838         if (bch_cutoff_writeback > bch_cutoff_writeback_sync) {
2839                 pr_warn("set bch_cutoff_writeback (%u) to %u\n",
2840                         bch_cutoff_writeback, bch_cutoff_writeback_sync);
2841                 bch_cutoff_writeback = bch_cutoff_writeback_sync;
2842         }
2843 }
2844
2845 static int __init bcache_init(void)
2846 {
2847         static const struct attribute *files[] = {
2848                 &ksysfs_register.attr,
2849                 &ksysfs_register_quiet.attr,
2850                 &ksysfs_pendings_cleanup.attr,
2851                 NULL
2852         };
2853
2854         check_module_parameters();
2855
2856         mutex_init(&bch_register_lock);
2857         init_waitqueue_head(&unregister_wait);
2858         register_reboot_notifier(&reboot);
2859
2860         bcache_major = register_blkdev(0, "bcache");
2861         if (bcache_major < 0) {
2862                 unregister_reboot_notifier(&reboot);
2863                 mutex_destroy(&bch_register_lock);
2864                 return bcache_major;
2865         }
2866
2867         if (bch_btree_init())
2868                 goto err;
2869
2870         bcache_wq = alloc_workqueue("bcache", WQ_MEM_RECLAIM, 0);
2871         if (!bcache_wq)
2872                 goto err;
2873
2874         /*
2875          * Let's not make this `WQ_MEM_RECLAIM` for the following reasons:
2876          *
2877          * 1. It used `system_wq` before which also does no memory reclaim.
2878          * 2. With `WQ_MEM_RECLAIM` desktop stalls, increased boot times, and
2879          *    reduced throughput can be observed.
2880          *
2881          * We still want to user our own queue to not congest the `system_wq`.
2882          */
2883         bch_flush_wq = alloc_workqueue("bch_flush", 0, 0);
2884         if (!bch_flush_wq)
2885                 goto err;
2886
2887         bch_journal_wq = alloc_workqueue("bch_journal", WQ_MEM_RECLAIM, 0);
2888         if (!bch_journal_wq)
2889                 goto err;
2890
2891         bcache_kobj = kobject_create_and_add("bcache", fs_kobj);
2892         if (!bcache_kobj)
2893                 goto err;
2894
2895         if (bch_request_init() ||
2896             sysfs_create_files(bcache_kobj, files))
2897                 goto err;
2898
2899         bch_debug_init();
2900
2901         bcache_is_reboot = false;
2902
2903         return 0;
2904 err:
2905         bcache_exit();
2906         return -ENOMEM;
2907 }
2908
2909 /*
2910  * Module hooks
2911  */
2912 module_exit(bcache_exit);
2913 module_init(bcache_init);
2914
2915 module_param(bch_cutoff_writeback, uint, 0);
2916 MODULE_PARM_DESC(bch_cutoff_writeback, "threshold to cutoff writeback");
2917
2918 module_param(bch_cutoff_writeback_sync, uint, 0);
2919 MODULE_PARM_DESC(bch_cutoff_writeback_sync, "hard threshold to cutoff writeback");
2920
2921 MODULE_DESCRIPTION("Bcache: a Linux block layer cache");
2922 MODULE_AUTHOR("Kent Overstreet <[email protected]>");
2923 MODULE_LICENSE("GPL");