drivers/md/bcache/super.c

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