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