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