Loading...
1/* SPDX-License-Identifier: GPL-2.0 */
2#ifndef _BCACHE_H
3#define _BCACHE_H
4
5/*
6 * SOME HIGH LEVEL CODE DOCUMENTATION:
7 *
8 * Bcache mostly works with cache sets, cache devices, and backing devices.
9 *
10 * Support for multiple cache devices hasn't quite been finished off yet, but
11 * it's about 95% plumbed through. A cache set and its cache devices is sort of
12 * like a md raid array and its component devices. Most of the code doesn't care
13 * about individual cache devices, the main abstraction is the cache set.
14 *
15 * Multiple cache devices is intended to give us the ability to mirror dirty
16 * cached data and metadata, without mirroring clean cached data.
17 *
18 * Backing devices are different, in that they have a lifetime independent of a
19 * cache set. When you register a newly formatted backing device it'll come up
20 * in passthrough mode, and then you can attach and detach a backing device from
21 * a cache set at runtime - while it's mounted and in use. Detaching implicitly
22 * invalidates any cached data for that backing device.
23 *
24 * A cache set can have multiple (many) backing devices attached to it.
25 *
26 * There's also flash only volumes - this is the reason for the distinction
27 * between struct cached_dev and struct bcache_device. A flash only volume
28 * works much like a bcache device that has a backing device, except the
29 * "cached" data is always dirty. The end result is that we get thin
30 * provisioning with very little additional code.
31 *
32 * Flash only volumes work but they're not production ready because the moving
33 * garbage collector needs more work. More on that later.
34 *
35 * BUCKETS/ALLOCATION:
36 *
37 * Bcache is primarily designed for caching, which means that in normal
38 * operation all of our available space will be allocated. Thus, we need an
39 * efficient way of deleting things from the cache so we can write new things to
40 * it.
41 *
42 * To do this, we first divide the cache device up into buckets. A bucket is the
43 * unit of allocation; they're typically around 1 mb - anywhere from 128k to 2M+
44 * works efficiently.
45 *
46 * Each bucket has a 16 bit priority, and an 8 bit generation associated with
47 * it. The gens and priorities for all the buckets are stored contiguously and
48 * packed on disk (in a linked list of buckets - aside from the superblock, all
49 * of bcache's metadata is stored in buckets).
50 *
51 * The priority is used to implement an LRU. We reset a bucket's priority when
52 * we allocate it or on cache it, and every so often we decrement the priority
53 * of each bucket. It could be used to implement something more sophisticated,
54 * if anyone ever gets around to it.
55 *
56 * The generation is used for invalidating buckets. Each pointer also has an 8
57 * bit generation embedded in it; for a pointer to be considered valid, its gen
58 * must match the gen of the bucket it points into. Thus, to reuse a bucket all
59 * we have to do is increment its gen (and write its new gen to disk; we batch
60 * this up).
61 *
62 * Bcache is entirely COW - we never write twice to a bucket, even buckets that
63 * contain metadata (including btree nodes).
64 *
65 * THE BTREE:
66 *
67 * Bcache is in large part design around the btree.
68 *
69 * At a high level, the btree is just an index of key -> ptr tuples.
70 *
71 * Keys represent extents, and thus have a size field. Keys also have a variable
72 * number of pointers attached to them (potentially zero, which is handy for
73 * invalidating the cache).
74 *
75 * The key itself is an inode:offset pair. The inode number corresponds to a
76 * backing device or a flash only volume. The offset is the ending offset of the
77 * extent within the inode - not the starting offset; this makes lookups
78 * slightly more convenient.
79 *
80 * Pointers contain the cache device id, the offset on that device, and an 8 bit
81 * generation number. More on the gen later.
82 *
83 * Index lookups are not fully abstracted - cache lookups in particular are
84 * still somewhat mixed in with the btree code, but things are headed in that
85 * direction.
86 *
87 * Updates are fairly well abstracted, though. There are two different ways of
88 * updating the btree; insert and replace.
89 *
90 * BTREE_INSERT will just take a list of keys and insert them into the btree -
91 * overwriting (possibly only partially) any extents they overlap with. This is
92 * used to update the index after a write.
93 *
94 * BTREE_REPLACE is really cmpxchg(); it inserts a key into the btree iff it is
95 * overwriting a key that matches another given key. This is used for inserting
96 * data into the cache after a cache miss, and for background writeback, and for
97 * the moving garbage collector.
98 *
99 * There is no "delete" operation; deleting things from the index is
100 * accomplished by either by invalidating pointers (by incrementing a bucket's
101 * gen) or by inserting a key with 0 pointers - which will overwrite anything
102 * previously present at that location in the index.
103 *
104 * This means that there are always stale/invalid keys in the btree. They're
105 * filtered out by the code that iterates through a btree node, and removed when
106 * a btree node is rewritten.
107 *
108 * BTREE NODES:
109 *
110 * Our unit of allocation is a bucket, and we we can't arbitrarily allocate and
111 * free smaller than a bucket - so, that's how big our btree nodes are.
112 *
113 * (If buckets are really big we'll only use part of the bucket for a btree node
114 * - no less than 1/4th - but a bucket still contains no more than a single
115 * btree node. I'd actually like to change this, but for now we rely on the
116 * bucket's gen for deleting btree nodes when we rewrite/split a node.)
117 *
118 * Anyways, btree nodes are big - big enough to be inefficient with a textbook
119 * btree implementation.
120 *
121 * The way this is solved is that btree nodes are internally log structured; we
122 * can append new keys to an existing btree node without rewriting it. This
123 * means each set of keys we write is sorted, but the node is not.
124 *
125 * We maintain this log structure in memory - keeping 1Mb of keys sorted would
126 * be expensive, and we have to distinguish between the keys we have written and
127 * the keys we haven't. So to do a lookup in a btree node, we have to search
128 * each sorted set. But we do merge written sets together lazily, so the cost of
129 * these extra searches is quite low (normally most of the keys in a btree node
130 * will be in one big set, and then there'll be one or two sets that are much
131 * smaller).
132 *
133 * This log structure makes bcache's btree more of a hybrid between a
134 * conventional btree and a compacting data structure, with some of the
135 * advantages of both.
136 *
137 * GARBAGE COLLECTION:
138 *
139 * We can't just invalidate any bucket - it might contain dirty data or
140 * metadata. If it once contained dirty data, other writes might overwrite it
141 * later, leaving no valid pointers into that bucket in the index.
142 *
143 * Thus, the primary purpose of garbage collection is to find buckets to reuse.
144 * It also counts how much valid data it each bucket currently contains, so that
145 * allocation can reuse buckets sooner when they've been mostly overwritten.
146 *
147 * It also does some things that are really internal to the btree
148 * implementation. If a btree node contains pointers that are stale by more than
149 * some threshold, it rewrites the btree node to avoid the bucket's generation
150 * wrapping around. It also merges adjacent btree nodes if they're empty enough.
151 *
152 * THE JOURNAL:
153 *
154 * Bcache's journal is not necessary for consistency; we always strictly
155 * order metadata writes so that the btree and everything else is consistent on
156 * disk in the event of an unclean shutdown, and in fact bcache had writeback
157 * caching (with recovery from unclean shutdown) before journalling was
158 * implemented.
159 *
160 * Rather, the journal is purely a performance optimization; we can't complete a
161 * write until we've updated the index on disk, otherwise the cache would be
162 * inconsistent in the event of an unclean shutdown. This means that without the
163 * journal, on random write workloads we constantly have to update all the leaf
164 * nodes in the btree, and those writes will be mostly empty (appending at most
165 * a few keys each) - highly inefficient in terms of amount of metadata writes,
166 * and it puts more strain on the various btree resorting/compacting code.
167 *
168 * The journal is just a log of keys we've inserted; on startup we just reinsert
169 * all the keys in the open journal entries. That means that when we're updating
170 * a node in the btree, we can wait until a 4k block of keys fills up before
171 * writing them out.
172 *
173 * For simplicity, we only journal updates to leaf nodes; updates to parent
174 * nodes are rare enough (since our leaf nodes are huge) that it wasn't worth
175 * the complexity to deal with journalling them (in particular, journal replay)
176 * - updates to non leaf nodes just happen synchronously (see btree_split()).
177 */
178
179#define pr_fmt(fmt) "bcache: %s() " fmt "\n", __func__
180
181#include <linux/bcache.h>
182#include <linux/bio.h>
183#include <linux/kobject.h>
184#include <linux/list.h>
185#include <linux/mutex.h>
186#include <linux/rbtree.h>
187#include <linux/rwsem.h>
188#include <linux/refcount.h>
189#include <linux/types.h>
190#include <linux/workqueue.h>
191#include <linux/kthread.h>
192
193#include "bset.h"
194#include "util.h"
195#include "closure.h"
196
197struct bucket {
198 atomic_t pin;
199 uint16_t prio;
200 uint8_t gen;
201 uint8_t last_gc; /* Most out of date gen in the btree */
202 uint16_t gc_mark; /* Bitfield used by GC. See below for field */
203};
204
205/*
206 * I'd use bitfields for these, but I don't trust the compiler not to screw me
207 * as multiple threads touch struct bucket without locking
208 */
209
210BITMASK(GC_MARK, struct bucket, gc_mark, 0, 2);
211#define GC_MARK_RECLAIMABLE 1
212#define GC_MARK_DIRTY 2
213#define GC_MARK_METADATA 3
214#define GC_SECTORS_USED_SIZE 13
215#define MAX_GC_SECTORS_USED (~(~0ULL << GC_SECTORS_USED_SIZE))
216BITMASK(GC_SECTORS_USED, struct bucket, gc_mark, 2, GC_SECTORS_USED_SIZE);
217BITMASK(GC_MOVE, struct bucket, gc_mark, 15, 1);
218
219#include "journal.h"
220#include "stats.h"
221struct search;
222struct btree;
223struct keybuf;
224
225struct keybuf_key {
226 struct rb_node node;
227 BKEY_PADDED(key);
228 void *private;
229};
230
231struct keybuf {
232 struct bkey last_scanned;
233 spinlock_t lock;
234
235 /*
236 * Beginning and end of range in rb tree - so that we can skip taking
237 * lock and checking the rb tree when we need to check for overlapping
238 * keys.
239 */
240 struct bkey start;
241 struct bkey end;
242
243 struct rb_root keys;
244
245#define KEYBUF_NR 500
246 DECLARE_ARRAY_ALLOCATOR(struct keybuf_key, freelist, KEYBUF_NR);
247};
248
249struct bcache_device {
250 struct closure cl;
251
252 struct kobject kobj;
253
254 struct cache_set *c;
255 unsigned id;
256#define BCACHEDEVNAME_SIZE 12
257 char name[BCACHEDEVNAME_SIZE];
258
259 struct gendisk *disk;
260
261 unsigned long flags;
262#define BCACHE_DEV_CLOSING 0
263#define BCACHE_DEV_DETACHING 1
264#define BCACHE_DEV_UNLINK_DONE 2
265#define BCACHE_DEV_WB_RUNNING 3
266#define BCACHE_DEV_RATE_DW_RUNNING 4
267 unsigned nr_stripes;
268 unsigned stripe_size;
269 atomic_t *stripe_sectors_dirty;
270 unsigned long *full_dirty_stripes;
271
272 struct bio_set *bio_split;
273
274 unsigned data_csum:1;
275
276 int (*cache_miss)(struct btree *, struct search *,
277 struct bio *, unsigned);
278 int (*ioctl) (struct bcache_device *, fmode_t, unsigned, unsigned long);
279};
280
281struct io {
282 /* Used to track sequential IO so it can be skipped */
283 struct hlist_node hash;
284 struct list_head lru;
285
286 unsigned long jiffies;
287 unsigned sequential;
288 sector_t last;
289};
290
291enum stop_on_failure {
292 BCH_CACHED_DEV_STOP_AUTO = 0,
293 BCH_CACHED_DEV_STOP_ALWAYS,
294 BCH_CACHED_DEV_STOP_MODE_MAX,
295};
296
297struct cached_dev {
298 struct list_head list;
299 struct bcache_device disk;
300 struct block_device *bdev;
301
302 struct cache_sb sb;
303 struct bio sb_bio;
304 struct bio_vec sb_bv[1];
305 struct closure sb_write;
306 struct semaphore sb_write_mutex;
307
308 /* Refcount on the cache set. Always nonzero when we're caching. */
309 refcount_t count;
310 struct work_struct detach;
311
312 /*
313 * Device might not be running if it's dirty and the cache set hasn't
314 * showed up yet.
315 */
316 atomic_t running;
317
318 /*
319 * Writes take a shared lock from start to finish; scanning for dirty
320 * data to refill the rb tree requires an exclusive lock.
321 */
322 struct rw_semaphore writeback_lock;
323
324 /*
325 * Nonzero, and writeback has a refcount (d->count), iff there is dirty
326 * data in the cache. Protected by writeback_lock; must have an
327 * shared lock to set and exclusive lock to clear.
328 */
329 atomic_t has_dirty;
330
331 /*
332 * Set to zero by things that touch the backing volume-- except
333 * writeback. Incremented by writeback. Used to determine when to
334 * accelerate idle writeback.
335 */
336 atomic_t backing_idle;
337
338 struct bch_ratelimit writeback_rate;
339 struct delayed_work writeback_rate_update;
340
341 /* Limit number of writeback bios in flight */
342 struct semaphore in_flight;
343 struct task_struct *writeback_thread;
344 struct workqueue_struct *writeback_write_wq;
345
346 struct keybuf writeback_keys;
347
348 /*
349 * Order the write-half of writeback operations strongly in dispatch
350 * order. (Maintain LBA order; don't allow reads completing out of
351 * order to re-order the writes...)
352 */
353 struct closure_waitlist writeback_ordering_wait;
354 atomic_t writeback_sequence_next;
355
356 /* For tracking sequential IO */
357#define RECENT_IO_BITS 7
358#define RECENT_IO (1 << RECENT_IO_BITS)
359 struct io io[RECENT_IO];
360 struct hlist_head io_hash[RECENT_IO + 1];
361 struct list_head io_lru;
362 spinlock_t io_lock;
363
364 struct cache_accounting accounting;
365
366 /* The rest of this all shows up in sysfs */
367 unsigned sequential_cutoff;
368 unsigned readahead;
369
370 unsigned io_disable:1;
371 unsigned verify:1;
372 unsigned bypass_torture_test:1;
373
374 unsigned partial_stripes_expensive:1;
375 unsigned writeback_metadata:1;
376 unsigned writeback_running:1;
377 unsigned char writeback_percent;
378 unsigned writeback_delay;
379
380 uint64_t writeback_rate_target;
381 int64_t writeback_rate_proportional;
382 int64_t writeback_rate_integral;
383 int64_t writeback_rate_integral_scaled;
384 int32_t writeback_rate_change;
385
386 unsigned writeback_rate_update_seconds;
387 unsigned writeback_rate_i_term_inverse;
388 unsigned writeback_rate_p_term_inverse;
389 unsigned writeback_rate_minimum;
390
391 enum stop_on_failure stop_when_cache_set_failed;
392#define DEFAULT_CACHED_DEV_ERROR_LIMIT 64
393 atomic_t io_errors;
394 unsigned error_limit;
395
396 char backing_dev_name[BDEVNAME_SIZE];
397};
398
399enum alloc_reserve {
400 RESERVE_BTREE,
401 RESERVE_PRIO,
402 RESERVE_MOVINGGC,
403 RESERVE_NONE,
404 RESERVE_NR,
405};
406
407struct cache {
408 struct cache_set *set;
409 struct cache_sb sb;
410 struct bio sb_bio;
411 struct bio_vec sb_bv[1];
412
413 struct kobject kobj;
414 struct block_device *bdev;
415
416 struct task_struct *alloc_thread;
417
418 struct closure prio;
419 struct prio_set *disk_buckets;
420
421 /*
422 * When allocating new buckets, prio_write() gets first dibs - since we
423 * may not be allocate at all without writing priorities and gens.
424 * prio_buckets[] contains the last buckets we wrote priorities to (so
425 * gc can mark them as metadata), prio_next[] contains the buckets
426 * allocated for the next prio write.
427 */
428 uint64_t *prio_buckets;
429 uint64_t *prio_last_buckets;
430
431 /*
432 * free: Buckets that are ready to be used
433 *
434 * free_inc: Incoming buckets - these are buckets that currently have
435 * cached data in them, and we can't reuse them until after we write
436 * their new gen to disk. After prio_write() finishes writing the new
437 * gens/prios, they'll be moved to the free list (and possibly discarded
438 * in the process)
439 */
440 DECLARE_FIFO(long, free)[RESERVE_NR];
441 DECLARE_FIFO(long, free_inc);
442
443 size_t fifo_last_bucket;
444
445 /* Allocation stuff: */
446 struct bucket *buckets;
447
448 DECLARE_HEAP(struct bucket *, heap);
449
450 /*
451 * If nonzero, we know we aren't going to find any buckets to invalidate
452 * until a gc finishes - otherwise we could pointlessly burn a ton of
453 * cpu
454 */
455 unsigned invalidate_needs_gc;
456
457 bool discard; /* Get rid of? */
458
459 struct journal_device journal;
460
461 /* The rest of this all shows up in sysfs */
462#define IO_ERROR_SHIFT 20
463 atomic_t io_errors;
464 atomic_t io_count;
465
466 atomic_long_t meta_sectors_written;
467 atomic_long_t btree_sectors_written;
468 atomic_long_t sectors_written;
469
470 char cache_dev_name[BDEVNAME_SIZE];
471};
472
473struct gc_stat {
474 size_t nodes;
475 size_t key_bytes;
476
477 size_t nkeys;
478 uint64_t data; /* sectors */
479 unsigned in_use; /* percent */
480};
481
482/*
483 * Flag bits, for how the cache set is shutting down, and what phase it's at:
484 *
485 * CACHE_SET_UNREGISTERING means we're not just shutting down, we're detaching
486 * all the backing devices first (their cached data gets invalidated, and they
487 * won't automatically reattach).
488 *
489 * CACHE_SET_STOPPING always gets set first when we're closing down a cache set;
490 * we'll continue to run normally for awhile with CACHE_SET_STOPPING set (i.e.
491 * flushing dirty data).
492 *
493 * CACHE_SET_RUNNING means all cache devices have been registered and journal
494 * replay is complete.
495 *
496 * CACHE_SET_IO_DISABLE is set when bcache is stopping the whold cache set, all
497 * external and internal I/O should be denied when this flag is set.
498 *
499 */
500#define CACHE_SET_UNREGISTERING 0
501#define CACHE_SET_STOPPING 1
502#define CACHE_SET_RUNNING 2
503#define CACHE_SET_IO_DISABLE 3
504
505struct cache_set {
506 struct closure cl;
507
508 struct list_head list;
509 struct kobject kobj;
510 struct kobject internal;
511 struct dentry *debug;
512 struct cache_accounting accounting;
513
514 unsigned long flags;
515
516 struct cache_sb sb;
517
518 struct cache *cache[MAX_CACHES_PER_SET];
519 struct cache *cache_by_alloc[MAX_CACHES_PER_SET];
520 int caches_loaded;
521
522 struct bcache_device **devices;
523 unsigned devices_max_used;
524 struct list_head cached_devs;
525 uint64_t cached_dev_sectors;
526 struct closure caching;
527
528 struct closure sb_write;
529 struct semaphore sb_write_mutex;
530
531 mempool_t *search;
532 mempool_t *bio_meta;
533 struct bio_set *bio_split;
534
535 /* For the btree cache */
536 struct shrinker shrink;
537
538 /* For the btree cache and anything allocation related */
539 struct mutex bucket_lock;
540
541 /* log2(bucket_size), in sectors */
542 unsigned short bucket_bits;
543
544 /* log2(block_size), in sectors */
545 unsigned short block_bits;
546
547 /*
548 * Default number of pages for a new btree node - may be less than a
549 * full bucket
550 */
551 unsigned btree_pages;
552
553 /*
554 * Lists of struct btrees; lru is the list for structs that have memory
555 * allocated for actual btree node, freed is for structs that do not.
556 *
557 * We never free a struct btree, except on shutdown - we just put it on
558 * the btree_cache_freed list and reuse it later. This simplifies the
559 * code, and it doesn't cost us much memory as the memory usage is
560 * dominated by buffers that hold the actual btree node data and those
561 * can be freed - and the number of struct btrees allocated is
562 * effectively bounded.
563 *
564 * btree_cache_freeable effectively is a small cache - we use it because
565 * high order page allocations can be rather expensive, and it's quite
566 * common to delete and allocate btree nodes in quick succession. It
567 * should never grow past ~2-3 nodes in practice.
568 */
569 struct list_head btree_cache;
570 struct list_head btree_cache_freeable;
571 struct list_head btree_cache_freed;
572
573 /* Number of elements in btree_cache + btree_cache_freeable lists */
574 unsigned btree_cache_used;
575
576 /*
577 * If we need to allocate memory for a new btree node and that
578 * allocation fails, we can cannibalize another node in the btree cache
579 * to satisfy the allocation - lock to guarantee only one thread does
580 * this at a time:
581 */
582 wait_queue_head_t btree_cache_wait;
583 struct task_struct *btree_cache_alloc_lock;
584
585 /*
586 * When we free a btree node, we increment the gen of the bucket the
587 * node is in - but we can't rewrite the prios and gens until we
588 * finished whatever it is we were doing, otherwise after a crash the
589 * btree node would be freed but for say a split, we might not have the
590 * pointers to the new nodes inserted into the btree yet.
591 *
592 * This is a refcount that blocks prio_write() until the new keys are
593 * written.
594 */
595 atomic_t prio_blocked;
596 wait_queue_head_t bucket_wait;
597
598 /*
599 * For any bio we don't skip we subtract the number of sectors from
600 * rescale; when it hits 0 we rescale all the bucket priorities.
601 */
602 atomic_t rescale;
603 /*
604 * When we invalidate buckets, we use both the priority and the amount
605 * of good data to determine which buckets to reuse first - to weight
606 * those together consistently we keep track of the smallest nonzero
607 * priority of any bucket.
608 */
609 uint16_t min_prio;
610
611 /*
612 * max(gen - last_gc) for all buckets. When it gets too big we have to gc
613 * to keep gens from wrapping around.
614 */
615 uint8_t need_gc;
616 struct gc_stat gc_stats;
617 size_t nbuckets;
618 size_t avail_nbuckets;
619
620 struct task_struct *gc_thread;
621 /* Where in the btree gc currently is */
622 struct bkey gc_done;
623
624 /*
625 * The allocation code needs gc_mark in struct bucket to be correct, but
626 * it's not while a gc is in progress. Protected by bucket_lock.
627 */
628 int gc_mark_valid;
629
630 /* Counts how many sectors bio_insert has added to the cache */
631 atomic_t sectors_to_gc;
632 wait_queue_head_t gc_wait;
633
634 struct keybuf moving_gc_keys;
635 /* Number of moving GC bios in flight */
636 struct semaphore moving_in_flight;
637
638 struct workqueue_struct *moving_gc_wq;
639
640 struct btree *root;
641
642#ifdef CONFIG_BCACHE_DEBUG
643 struct btree *verify_data;
644 struct bset *verify_ondisk;
645 struct mutex verify_lock;
646#endif
647
648 unsigned nr_uuids;
649 struct uuid_entry *uuids;
650 BKEY_PADDED(uuid_bucket);
651 struct closure uuid_write;
652 struct semaphore uuid_write_mutex;
653
654 /*
655 * A btree node on disk could have too many bsets for an iterator to fit
656 * on the stack - have to dynamically allocate them
657 */
658 mempool_t *fill_iter;
659
660 struct bset_sort_state sort;
661
662 /* List of buckets we're currently writing data to */
663 struct list_head data_buckets;
664 spinlock_t data_bucket_lock;
665
666 struct journal journal;
667
668#define CONGESTED_MAX 1024
669 unsigned congested_last_us;
670 atomic_t congested;
671
672 /* The rest of this all shows up in sysfs */
673 unsigned congested_read_threshold_us;
674 unsigned congested_write_threshold_us;
675
676 struct time_stats btree_gc_time;
677 struct time_stats btree_split_time;
678 struct time_stats btree_read_time;
679
680 atomic_long_t cache_read_races;
681 atomic_long_t writeback_keys_done;
682 atomic_long_t writeback_keys_failed;
683
684 atomic_long_t reclaim;
685 atomic_long_t flush_write;
686 atomic_long_t retry_flush_write;
687
688 enum {
689 ON_ERROR_UNREGISTER,
690 ON_ERROR_PANIC,
691 } on_error;
692#define DEFAULT_IO_ERROR_LIMIT 8
693 unsigned error_limit;
694 unsigned error_decay;
695
696 unsigned short journal_delay_ms;
697 bool expensive_debug_checks;
698 unsigned verify:1;
699 unsigned key_merging_disabled:1;
700 unsigned gc_always_rewrite:1;
701 unsigned shrinker_disabled:1;
702 unsigned copy_gc_enabled:1;
703
704#define BUCKET_HASH_BITS 12
705 struct hlist_head bucket_hash[1 << BUCKET_HASH_BITS];
706
707 DECLARE_HEAP(struct btree *, flush_btree);
708};
709
710struct bbio {
711 unsigned submit_time_us;
712 union {
713 struct bkey key;
714 uint64_t _pad[3];
715 /*
716 * We only need pad = 3 here because we only ever carry around a
717 * single pointer - i.e. the pointer we're doing io to/from.
718 */
719 };
720 struct bio bio;
721};
722
723#define BTREE_PRIO USHRT_MAX
724#define INITIAL_PRIO 32768U
725
726#define btree_bytes(c) ((c)->btree_pages * PAGE_SIZE)
727#define btree_blocks(b) \
728 ((unsigned) (KEY_SIZE(&b->key) >> (b)->c->block_bits))
729
730#define btree_default_blocks(c) \
731 ((unsigned) ((PAGE_SECTORS * (c)->btree_pages) >> (c)->block_bits))
732
733#define bucket_pages(c) ((c)->sb.bucket_size / PAGE_SECTORS)
734#define bucket_bytes(c) ((c)->sb.bucket_size << 9)
735#define block_bytes(c) ((c)->sb.block_size << 9)
736
737#define prios_per_bucket(c) \
738 ((bucket_bytes(c) - sizeof(struct prio_set)) / \
739 sizeof(struct bucket_disk))
740#define prio_buckets(c) \
741 DIV_ROUND_UP((size_t) (c)->sb.nbuckets, prios_per_bucket(c))
742
743static inline size_t sector_to_bucket(struct cache_set *c, sector_t s)
744{
745 return s >> c->bucket_bits;
746}
747
748static inline sector_t bucket_to_sector(struct cache_set *c, size_t b)
749{
750 return ((sector_t) b) << c->bucket_bits;
751}
752
753static inline sector_t bucket_remainder(struct cache_set *c, sector_t s)
754{
755 return s & (c->sb.bucket_size - 1);
756}
757
758static inline struct cache *PTR_CACHE(struct cache_set *c,
759 const struct bkey *k,
760 unsigned ptr)
761{
762 return c->cache[PTR_DEV(k, ptr)];
763}
764
765static inline size_t PTR_BUCKET_NR(struct cache_set *c,
766 const struct bkey *k,
767 unsigned ptr)
768{
769 return sector_to_bucket(c, PTR_OFFSET(k, ptr));
770}
771
772static inline struct bucket *PTR_BUCKET(struct cache_set *c,
773 const struct bkey *k,
774 unsigned ptr)
775{
776 return PTR_CACHE(c, k, ptr)->buckets + PTR_BUCKET_NR(c, k, ptr);
777}
778
779static inline uint8_t gen_after(uint8_t a, uint8_t b)
780{
781 uint8_t r = a - b;
782 return r > 128U ? 0 : r;
783}
784
785static inline uint8_t ptr_stale(struct cache_set *c, const struct bkey *k,
786 unsigned i)
787{
788 return gen_after(PTR_BUCKET(c, k, i)->gen, PTR_GEN(k, i));
789}
790
791static inline bool ptr_available(struct cache_set *c, const struct bkey *k,
792 unsigned i)
793{
794 return (PTR_DEV(k, i) < MAX_CACHES_PER_SET) && PTR_CACHE(c, k, i);
795}
796
797/* Btree key macros */
798
799/*
800 * This is used for various on disk data structures - cache_sb, prio_set, bset,
801 * jset: The checksum is _always_ the first 8 bytes of these structs
802 */
803#define csum_set(i) \
804 bch_crc64(((void *) (i)) + sizeof(uint64_t), \
805 ((void *) bset_bkey_last(i)) - \
806 (((void *) (i)) + sizeof(uint64_t)))
807
808/* Error handling macros */
809
810#define btree_bug(b, ...) \
811do { \
812 if (bch_cache_set_error((b)->c, __VA_ARGS__)) \
813 dump_stack(); \
814} while (0)
815
816#define cache_bug(c, ...) \
817do { \
818 if (bch_cache_set_error(c, __VA_ARGS__)) \
819 dump_stack(); \
820} while (0)
821
822#define btree_bug_on(cond, b, ...) \
823do { \
824 if (cond) \
825 btree_bug(b, __VA_ARGS__); \
826} while (0)
827
828#define cache_bug_on(cond, c, ...) \
829do { \
830 if (cond) \
831 cache_bug(c, __VA_ARGS__); \
832} while (0)
833
834#define cache_set_err_on(cond, c, ...) \
835do { \
836 if (cond) \
837 bch_cache_set_error(c, __VA_ARGS__); \
838} while (0)
839
840/* Looping macros */
841
842#define for_each_cache(ca, cs, iter) \
843 for (iter = 0; ca = cs->cache[iter], iter < (cs)->sb.nr_in_set; iter++)
844
845#define for_each_bucket(b, ca) \
846 for (b = (ca)->buckets + (ca)->sb.first_bucket; \
847 b < (ca)->buckets + (ca)->sb.nbuckets; b++)
848
849static inline void cached_dev_put(struct cached_dev *dc)
850{
851 if (refcount_dec_and_test(&dc->count))
852 schedule_work(&dc->detach);
853}
854
855static inline bool cached_dev_get(struct cached_dev *dc)
856{
857 if (!refcount_inc_not_zero(&dc->count))
858 return false;
859
860 /* Paired with the mb in cached_dev_attach */
861 smp_mb__after_atomic();
862 return true;
863}
864
865/*
866 * bucket_gc_gen() returns the difference between the bucket's current gen and
867 * the oldest gen of any pointer into that bucket in the btree (last_gc).
868 */
869
870static inline uint8_t bucket_gc_gen(struct bucket *b)
871{
872 return b->gen - b->last_gc;
873}
874
875#define BUCKET_GC_GEN_MAX 96U
876
877#define kobj_attribute_write(n, fn) \
878 static struct kobj_attribute ksysfs_##n = __ATTR(n, S_IWUSR, NULL, fn)
879
880#define kobj_attribute_rw(n, show, store) \
881 static struct kobj_attribute ksysfs_##n = \
882 __ATTR(n, S_IWUSR|S_IRUSR, show, store)
883
884static inline void wake_up_allocators(struct cache_set *c)
885{
886 struct cache *ca;
887 unsigned i;
888
889 for_each_cache(ca, c, i)
890 wake_up_process(ca->alloc_thread);
891}
892
893static inline void closure_bio_submit(struct cache_set *c,
894 struct bio *bio,
895 struct closure *cl)
896{
897 closure_get(cl);
898 if (unlikely(test_bit(CACHE_SET_IO_DISABLE, &c->flags))) {
899 bio->bi_status = BLK_STS_IOERR;
900 bio_endio(bio);
901 return;
902 }
903 generic_make_request(bio);
904}
905
906/*
907 * Prevent the kthread exits directly, and make sure when kthread_stop()
908 * is called to stop a kthread, it is still alive. If a kthread might be
909 * stopped by CACHE_SET_IO_DISABLE bit set, wait_for_kthread_stop() is
910 * necessary before the kthread returns.
911 */
912static inline void wait_for_kthread_stop(void)
913{
914 while (!kthread_should_stop()) {
915 set_current_state(TASK_INTERRUPTIBLE);
916 schedule();
917 }
918}
919
920/* Forward declarations */
921
922void bch_count_backing_io_errors(struct cached_dev *dc, struct bio *bio);
923void bch_count_io_errors(struct cache *, blk_status_t, int, const char *);
924void bch_bbio_count_io_errors(struct cache_set *, struct bio *,
925 blk_status_t, const char *);
926void bch_bbio_endio(struct cache_set *, struct bio *, blk_status_t,
927 const char *);
928void bch_bbio_free(struct bio *, struct cache_set *);
929struct bio *bch_bbio_alloc(struct cache_set *);
930
931void __bch_submit_bbio(struct bio *, struct cache_set *);
932void bch_submit_bbio(struct bio *, struct cache_set *, struct bkey *, unsigned);
933
934uint8_t bch_inc_gen(struct cache *, struct bucket *);
935void bch_rescale_priorities(struct cache_set *, int);
936
937bool bch_can_invalidate_bucket(struct cache *, struct bucket *);
938void __bch_invalidate_one_bucket(struct cache *, struct bucket *);
939
940void __bch_bucket_free(struct cache *, struct bucket *);
941void bch_bucket_free(struct cache_set *, struct bkey *);
942
943long bch_bucket_alloc(struct cache *, unsigned, bool);
944int __bch_bucket_alloc_set(struct cache_set *, unsigned,
945 struct bkey *, int, bool);
946int bch_bucket_alloc_set(struct cache_set *, unsigned,
947 struct bkey *, int, bool);
948bool bch_alloc_sectors(struct cache_set *, struct bkey *, unsigned,
949 unsigned, unsigned, bool);
950bool bch_cached_dev_error(struct cached_dev *dc);
951
952__printf(2, 3)
953bool bch_cache_set_error(struct cache_set *, const char *, ...);
954
955void bch_prio_write(struct cache *);
956void bch_write_bdev_super(struct cached_dev *, struct closure *);
957
958extern struct workqueue_struct *bcache_wq;
959extern const char * const bch_cache_modes[];
960extern const char * const bch_stop_on_failure_modes[];
961extern struct mutex bch_register_lock;
962extern struct list_head bch_cache_sets;
963
964extern struct kobj_type bch_cached_dev_ktype;
965extern struct kobj_type bch_flash_dev_ktype;
966extern struct kobj_type bch_cache_set_ktype;
967extern struct kobj_type bch_cache_set_internal_ktype;
968extern struct kobj_type bch_cache_ktype;
969
970void bch_cached_dev_release(struct kobject *);
971void bch_flash_dev_release(struct kobject *);
972void bch_cache_set_release(struct kobject *);
973void bch_cache_release(struct kobject *);
974
975int bch_uuid_write(struct cache_set *);
976void bcache_write_super(struct cache_set *);
977
978int bch_flash_dev_create(struct cache_set *c, uint64_t size);
979
980int bch_cached_dev_attach(struct cached_dev *, struct cache_set *, uint8_t *);
981void bch_cached_dev_detach(struct cached_dev *);
982void bch_cached_dev_run(struct cached_dev *);
983void bcache_device_stop(struct bcache_device *);
984
985void bch_cache_set_unregister(struct cache_set *);
986void bch_cache_set_stop(struct cache_set *);
987
988struct cache_set *bch_cache_set_alloc(struct cache_sb *);
989void bch_btree_cache_free(struct cache_set *);
990int bch_btree_cache_alloc(struct cache_set *);
991void bch_moving_init_cache_set(struct cache_set *);
992int bch_open_buckets_alloc(struct cache_set *);
993void bch_open_buckets_free(struct cache_set *);
994
995int bch_cache_allocator_start(struct cache *ca);
996
997void bch_debug_exit(void);
998int bch_debug_init(struct kobject *);
999void bch_request_exit(void);
1000int bch_request_init(void);
1001
1002#endif /* _BCACHE_H */
1/* SPDX-License-Identifier: GPL-2.0 */
2#ifndef _BCACHE_H
3#define _BCACHE_H
4
5/*
6 * SOME HIGH LEVEL CODE DOCUMENTATION:
7 *
8 * Bcache mostly works with cache sets, cache devices, and backing devices.
9 *
10 * Support for multiple cache devices hasn't quite been finished off yet, but
11 * it's about 95% plumbed through. A cache set and its cache devices is sort of
12 * like a md raid array and its component devices. Most of the code doesn't care
13 * about individual cache devices, the main abstraction is the cache set.
14 *
15 * Multiple cache devices is intended to give us the ability to mirror dirty
16 * cached data and metadata, without mirroring clean cached data.
17 *
18 * Backing devices are different, in that they have a lifetime independent of a
19 * cache set. When you register a newly formatted backing device it'll come up
20 * in passthrough mode, and then you can attach and detach a backing device from
21 * a cache set at runtime - while it's mounted and in use. Detaching implicitly
22 * invalidates any cached data for that backing device.
23 *
24 * A cache set can have multiple (many) backing devices attached to it.
25 *
26 * There's also flash only volumes - this is the reason for the distinction
27 * between struct cached_dev and struct bcache_device. A flash only volume
28 * works much like a bcache device that has a backing device, except the
29 * "cached" data is always dirty. The end result is that we get thin
30 * provisioning with very little additional code.
31 *
32 * Flash only volumes work but they're not production ready because the moving
33 * garbage collector needs more work. More on that later.
34 *
35 * BUCKETS/ALLOCATION:
36 *
37 * Bcache is primarily designed for caching, which means that in normal
38 * operation all of our available space will be allocated. Thus, we need an
39 * efficient way of deleting things from the cache so we can write new things to
40 * it.
41 *
42 * To do this, we first divide the cache device up into buckets. A bucket is the
43 * unit of allocation; they're typically around 1 mb - anywhere from 128k to 2M+
44 * works efficiently.
45 *
46 * Each bucket has a 16 bit priority, and an 8 bit generation associated with
47 * it. The gens and priorities for all the buckets are stored contiguously and
48 * packed on disk (in a linked list of buckets - aside from the superblock, all
49 * of bcache's metadata is stored in buckets).
50 *
51 * The priority is used to implement an LRU. We reset a bucket's priority when
52 * we allocate it or on cache it, and every so often we decrement the priority
53 * of each bucket. It could be used to implement something more sophisticated,
54 * if anyone ever gets around to it.
55 *
56 * The generation is used for invalidating buckets. Each pointer also has an 8
57 * bit generation embedded in it; for a pointer to be considered valid, its gen
58 * must match the gen of the bucket it points into. Thus, to reuse a bucket all
59 * we have to do is increment its gen (and write its new gen to disk; we batch
60 * this up).
61 *
62 * Bcache is entirely COW - we never write twice to a bucket, even buckets that
63 * contain metadata (including btree nodes).
64 *
65 * THE BTREE:
66 *
67 * Bcache is in large part design around the btree.
68 *
69 * At a high level, the btree is just an index of key -> ptr tuples.
70 *
71 * Keys represent extents, and thus have a size field. Keys also have a variable
72 * number of pointers attached to them (potentially zero, which is handy for
73 * invalidating the cache).
74 *
75 * The key itself is an inode:offset pair. The inode number corresponds to a
76 * backing device or a flash only volume. The offset is the ending offset of the
77 * extent within the inode - not the starting offset; this makes lookups
78 * slightly more convenient.
79 *
80 * Pointers contain the cache device id, the offset on that device, and an 8 bit
81 * generation number. More on the gen later.
82 *
83 * Index lookups are not fully abstracted - cache lookups in particular are
84 * still somewhat mixed in with the btree code, but things are headed in that
85 * direction.
86 *
87 * Updates are fairly well abstracted, though. There are two different ways of
88 * updating the btree; insert and replace.
89 *
90 * BTREE_INSERT will just take a list of keys and insert them into the btree -
91 * overwriting (possibly only partially) any extents they overlap with. This is
92 * used to update the index after a write.
93 *
94 * BTREE_REPLACE is really cmpxchg(); it inserts a key into the btree iff it is
95 * overwriting a key that matches another given key. This is used for inserting
96 * data into the cache after a cache miss, and for background writeback, and for
97 * the moving garbage collector.
98 *
99 * There is no "delete" operation; deleting things from the index is
100 * accomplished by either by invalidating pointers (by incrementing a bucket's
101 * gen) or by inserting a key with 0 pointers - which will overwrite anything
102 * previously present at that location in the index.
103 *
104 * This means that there are always stale/invalid keys in the btree. They're
105 * filtered out by the code that iterates through a btree node, and removed when
106 * a btree node is rewritten.
107 *
108 * BTREE NODES:
109 *
110 * Our unit of allocation is a bucket, and we we can't arbitrarily allocate and
111 * free smaller than a bucket - so, that's how big our btree nodes are.
112 *
113 * (If buckets are really big we'll only use part of the bucket for a btree node
114 * - no less than 1/4th - but a bucket still contains no more than a single
115 * btree node. I'd actually like to change this, but for now we rely on the
116 * bucket's gen for deleting btree nodes when we rewrite/split a node.)
117 *
118 * Anyways, btree nodes are big - big enough to be inefficient with a textbook
119 * btree implementation.
120 *
121 * The way this is solved is that btree nodes are internally log structured; we
122 * can append new keys to an existing btree node without rewriting it. This
123 * means each set of keys we write is sorted, but the node is not.
124 *
125 * We maintain this log structure in memory - keeping 1Mb of keys sorted would
126 * be expensive, and we have to distinguish between the keys we have written and
127 * the keys we haven't. So to do a lookup in a btree node, we have to search
128 * each sorted set. But we do merge written sets together lazily, so the cost of
129 * these extra searches is quite low (normally most of the keys in a btree node
130 * will be in one big set, and then there'll be one or two sets that are much
131 * smaller).
132 *
133 * This log structure makes bcache's btree more of a hybrid between a
134 * conventional btree and a compacting data structure, with some of the
135 * advantages of both.
136 *
137 * GARBAGE COLLECTION:
138 *
139 * We can't just invalidate any bucket - it might contain dirty data or
140 * metadata. If it once contained dirty data, other writes might overwrite it
141 * later, leaving no valid pointers into that bucket in the index.
142 *
143 * Thus, the primary purpose of garbage collection is to find buckets to reuse.
144 * It also counts how much valid data it each bucket currently contains, so that
145 * allocation can reuse buckets sooner when they've been mostly overwritten.
146 *
147 * It also does some things that are really internal to the btree
148 * implementation. If a btree node contains pointers that are stale by more than
149 * some threshold, it rewrites the btree node to avoid the bucket's generation
150 * wrapping around. It also merges adjacent btree nodes if they're empty enough.
151 *
152 * THE JOURNAL:
153 *
154 * Bcache's journal is not necessary for consistency; we always strictly
155 * order metadata writes so that the btree and everything else is consistent on
156 * disk in the event of an unclean shutdown, and in fact bcache had writeback
157 * caching (with recovery from unclean shutdown) before journalling was
158 * implemented.
159 *
160 * Rather, the journal is purely a performance optimization; we can't complete a
161 * write until we've updated the index on disk, otherwise the cache would be
162 * inconsistent in the event of an unclean shutdown. This means that without the
163 * journal, on random write workloads we constantly have to update all the leaf
164 * nodes in the btree, and those writes will be mostly empty (appending at most
165 * a few keys each) - highly inefficient in terms of amount of metadata writes,
166 * and it puts more strain on the various btree resorting/compacting code.
167 *
168 * The journal is just a log of keys we've inserted; on startup we just reinsert
169 * all the keys in the open journal entries. That means that when we're updating
170 * a node in the btree, we can wait until a 4k block of keys fills up before
171 * writing them out.
172 *
173 * For simplicity, we only journal updates to leaf nodes; updates to parent
174 * nodes are rare enough (since our leaf nodes are huge) that it wasn't worth
175 * the complexity to deal with journalling them (in particular, journal replay)
176 * - updates to non leaf nodes just happen synchronously (see btree_split()).
177 */
178
179#define pr_fmt(fmt) "bcache: %s() " fmt, __func__
180
181#include <linux/bcache.h>
182#include <linux/bio.h>
183#include <linux/kobject.h>
184#include <linux/list.h>
185#include <linux/mutex.h>
186#include <linux/rbtree.h>
187#include <linux/rwsem.h>
188#include <linux/refcount.h>
189#include <linux/types.h>
190#include <linux/workqueue.h>
191#include <linux/kthread.h>
192
193#include "bset.h"
194#include "util.h"
195#include "closure.h"
196
197struct bucket {
198 atomic_t pin;
199 uint16_t prio;
200 uint8_t gen;
201 uint8_t last_gc; /* Most out of date gen in the btree */
202 uint16_t gc_mark; /* Bitfield used by GC. See below for field */
203};
204
205/*
206 * I'd use bitfields for these, but I don't trust the compiler not to screw me
207 * as multiple threads touch struct bucket without locking
208 */
209
210BITMASK(GC_MARK, struct bucket, gc_mark, 0, 2);
211#define GC_MARK_RECLAIMABLE 1
212#define GC_MARK_DIRTY 2
213#define GC_MARK_METADATA 3
214#define GC_SECTORS_USED_SIZE 13
215#define MAX_GC_SECTORS_USED (~(~0ULL << GC_SECTORS_USED_SIZE))
216BITMASK(GC_SECTORS_USED, struct bucket, gc_mark, 2, GC_SECTORS_USED_SIZE);
217BITMASK(GC_MOVE, struct bucket, gc_mark, 15, 1);
218
219#include "journal.h"
220#include "stats.h"
221struct search;
222struct btree;
223struct keybuf;
224
225struct keybuf_key {
226 struct rb_node node;
227 BKEY_PADDED(key);
228 void *private;
229};
230
231struct keybuf {
232 struct bkey last_scanned;
233 spinlock_t lock;
234
235 /*
236 * Beginning and end of range in rb tree - so that we can skip taking
237 * lock and checking the rb tree when we need to check for overlapping
238 * keys.
239 */
240 struct bkey start;
241 struct bkey end;
242
243 struct rb_root keys;
244
245#define KEYBUF_NR 500
246 DECLARE_ARRAY_ALLOCATOR(struct keybuf_key, freelist, KEYBUF_NR);
247};
248
249struct bcache_device {
250 struct closure cl;
251
252 struct kobject kobj;
253
254 struct cache_set *c;
255 unsigned int id;
256#define BCACHEDEVNAME_SIZE 12
257 char name[BCACHEDEVNAME_SIZE];
258
259 struct gendisk *disk;
260
261 unsigned long flags;
262#define BCACHE_DEV_CLOSING 0
263#define BCACHE_DEV_DETACHING 1
264#define BCACHE_DEV_UNLINK_DONE 2
265#define BCACHE_DEV_WB_RUNNING 3
266#define BCACHE_DEV_RATE_DW_RUNNING 4
267 int nr_stripes;
268 unsigned int stripe_size;
269 atomic_t *stripe_sectors_dirty;
270 unsigned long *full_dirty_stripes;
271
272 struct bio_set bio_split;
273
274 unsigned int data_csum:1;
275
276 int (*cache_miss)(struct btree *b, struct search *s,
277 struct bio *bio, unsigned int sectors);
278 int (*ioctl)(struct bcache_device *d, fmode_t mode,
279 unsigned int cmd, unsigned long arg);
280};
281
282struct io {
283 /* Used to track sequential IO so it can be skipped */
284 struct hlist_node hash;
285 struct list_head lru;
286
287 unsigned long jiffies;
288 unsigned int sequential;
289 sector_t last;
290};
291
292enum stop_on_failure {
293 BCH_CACHED_DEV_STOP_AUTO = 0,
294 BCH_CACHED_DEV_STOP_ALWAYS,
295 BCH_CACHED_DEV_STOP_MODE_MAX,
296};
297
298struct cached_dev {
299 struct list_head list;
300 struct bcache_device disk;
301 struct block_device *bdev;
302
303 struct cache_sb sb;
304 struct cache_sb_disk *sb_disk;
305 struct bio sb_bio;
306 struct bio_vec sb_bv[1];
307 struct closure sb_write;
308 struct semaphore sb_write_mutex;
309
310 /* Refcount on the cache set. Always nonzero when we're caching. */
311 refcount_t count;
312 struct work_struct detach;
313
314 /*
315 * Device might not be running if it's dirty and the cache set hasn't
316 * showed up yet.
317 */
318 atomic_t running;
319
320 /*
321 * Writes take a shared lock from start to finish; scanning for dirty
322 * data to refill the rb tree requires an exclusive lock.
323 */
324 struct rw_semaphore writeback_lock;
325
326 /*
327 * Nonzero, and writeback has a refcount (d->count), iff there is dirty
328 * data in the cache. Protected by writeback_lock; must have an
329 * shared lock to set and exclusive lock to clear.
330 */
331 atomic_t has_dirty;
332
333#define BCH_CACHE_READA_ALL 0
334#define BCH_CACHE_READA_META_ONLY 1
335 unsigned int cache_readahead_policy;
336 struct bch_ratelimit writeback_rate;
337 struct delayed_work writeback_rate_update;
338
339 /* Limit number of writeback bios in flight */
340 struct semaphore in_flight;
341 struct task_struct *writeback_thread;
342 struct workqueue_struct *writeback_write_wq;
343
344 struct keybuf writeback_keys;
345
346 struct task_struct *status_update_thread;
347 /*
348 * Order the write-half of writeback operations strongly in dispatch
349 * order. (Maintain LBA order; don't allow reads completing out of
350 * order to re-order the writes...)
351 */
352 struct closure_waitlist writeback_ordering_wait;
353 atomic_t writeback_sequence_next;
354
355 /* For tracking sequential IO */
356#define RECENT_IO_BITS 7
357#define RECENT_IO (1 << RECENT_IO_BITS)
358 struct io io[RECENT_IO];
359 struct hlist_head io_hash[RECENT_IO + 1];
360 struct list_head io_lru;
361 spinlock_t io_lock;
362
363 struct cache_accounting accounting;
364
365 /* The rest of this all shows up in sysfs */
366 unsigned int sequential_cutoff;
367
368 unsigned int io_disable:1;
369 unsigned int verify:1;
370 unsigned int bypass_torture_test:1;
371
372 unsigned int partial_stripes_expensive:1;
373 unsigned int writeback_metadata:1;
374 unsigned int writeback_running:1;
375 unsigned int writeback_consider_fragment:1;
376 unsigned char writeback_percent;
377 unsigned int writeback_delay;
378
379 uint64_t writeback_rate_target;
380 int64_t writeback_rate_proportional;
381 int64_t writeback_rate_integral;
382 int64_t writeback_rate_integral_scaled;
383 int32_t writeback_rate_change;
384
385 unsigned int writeback_rate_update_seconds;
386 unsigned int writeback_rate_i_term_inverse;
387 unsigned int writeback_rate_p_term_inverse;
388 unsigned int writeback_rate_fp_term_low;
389 unsigned int writeback_rate_fp_term_mid;
390 unsigned int writeback_rate_fp_term_high;
391 unsigned int writeback_rate_minimum;
392
393 enum stop_on_failure stop_when_cache_set_failed;
394#define DEFAULT_CACHED_DEV_ERROR_LIMIT 64
395 atomic_t io_errors;
396 unsigned int error_limit;
397 unsigned int offline_seconds;
398
399 char backing_dev_name[BDEVNAME_SIZE];
400};
401
402enum alloc_reserve {
403 RESERVE_BTREE,
404 RESERVE_PRIO,
405 RESERVE_MOVINGGC,
406 RESERVE_NONE,
407 RESERVE_NR,
408};
409
410struct cache {
411 struct cache_set *set;
412 struct cache_sb sb;
413 struct cache_sb_disk *sb_disk;
414 struct bio sb_bio;
415 struct bio_vec sb_bv[1];
416
417 struct kobject kobj;
418 struct block_device *bdev;
419
420 struct task_struct *alloc_thread;
421
422 struct closure prio;
423 struct prio_set *disk_buckets;
424
425 /*
426 * When allocating new buckets, prio_write() gets first dibs - since we
427 * may not be allocate at all without writing priorities and gens.
428 * prio_last_buckets[] contains the last buckets we wrote priorities to
429 * (so gc can mark them as metadata), prio_buckets[] contains the
430 * buckets allocated for the next prio write.
431 */
432 uint64_t *prio_buckets;
433 uint64_t *prio_last_buckets;
434
435 /*
436 * free: Buckets that are ready to be used
437 *
438 * free_inc: Incoming buckets - these are buckets that currently have
439 * cached data in them, and we can't reuse them until after we write
440 * their new gen to disk. After prio_write() finishes writing the new
441 * gens/prios, they'll be moved to the free list (and possibly discarded
442 * in the process)
443 */
444 DECLARE_FIFO(long, free)[RESERVE_NR];
445 DECLARE_FIFO(long, free_inc);
446
447 size_t fifo_last_bucket;
448
449 /* Allocation stuff: */
450 struct bucket *buckets;
451
452 DECLARE_HEAP(struct bucket *, heap);
453
454 /*
455 * If nonzero, we know we aren't going to find any buckets to invalidate
456 * until a gc finishes - otherwise we could pointlessly burn a ton of
457 * cpu
458 */
459 unsigned int invalidate_needs_gc;
460
461 bool discard; /* Get rid of? */
462
463 struct journal_device journal;
464
465 /* The rest of this all shows up in sysfs */
466#define IO_ERROR_SHIFT 20
467 atomic_t io_errors;
468 atomic_t io_count;
469
470 atomic_long_t meta_sectors_written;
471 atomic_long_t btree_sectors_written;
472 atomic_long_t sectors_written;
473
474 char cache_dev_name[BDEVNAME_SIZE];
475};
476
477struct gc_stat {
478 size_t nodes;
479 size_t nodes_pre;
480 size_t key_bytes;
481
482 size_t nkeys;
483 uint64_t data; /* sectors */
484 unsigned int in_use; /* percent */
485};
486
487/*
488 * Flag bits, for how the cache set is shutting down, and what phase it's at:
489 *
490 * CACHE_SET_UNREGISTERING means we're not just shutting down, we're detaching
491 * all the backing devices first (their cached data gets invalidated, and they
492 * won't automatically reattach).
493 *
494 * CACHE_SET_STOPPING always gets set first when we're closing down a cache set;
495 * we'll continue to run normally for awhile with CACHE_SET_STOPPING set (i.e.
496 * flushing dirty data).
497 *
498 * CACHE_SET_RUNNING means all cache devices have been registered and journal
499 * replay is complete.
500 *
501 * CACHE_SET_IO_DISABLE is set when bcache is stopping the whold cache set, all
502 * external and internal I/O should be denied when this flag is set.
503 *
504 */
505#define CACHE_SET_UNREGISTERING 0
506#define CACHE_SET_STOPPING 1
507#define CACHE_SET_RUNNING 2
508#define CACHE_SET_IO_DISABLE 3
509
510struct cache_set {
511 struct closure cl;
512
513 struct list_head list;
514 struct kobject kobj;
515 struct kobject internal;
516 struct dentry *debug;
517 struct cache_accounting accounting;
518
519 unsigned long flags;
520 atomic_t idle_counter;
521 atomic_t at_max_writeback_rate;
522
523 struct cache *cache;
524
525 struct bcache_device **devices;
526 unsigned int devices_max_used;
527 atomic_t attached_dev_nr;
528 struct list_head cached_devs;
529 uint64_t cached_dev_sectors;
530 atomic_long_t flash_dev_dirty_sectors;
531 struct closure caching;
532
533 struct closure sb_write;
534 struct semaphore sb_write_mutex;
535
536 mempool_t search;
537 mempool_t bio_meta;
538 struct bio_set bio_split;
539
540 /* For the btree cache */
541 struct shrinker shrink;
542
543 /* For the btree cache and anything allocation related */
544 struct mutex bucket_lock;
545
546 /* log2(bucket_size), in sectors */
547 unsigned short bucket_bits;
548
549 /* log2(block_size), in sectors */
550 unsigned short block_bits;
551
552 /*
553 * Default number of pages for a new btree node - may be less than a
554 * full bucket
555 */
556 unsigned int btree_pages;
557
558 /*
559 * Lists of struct btrees; lru is the list for structs that have memory
560 * allocated for actual btree node, freed is for structs that do not.
561 *
562 * We never free a struct btree, except on shutdown - we just put it on
563 * the btree_cache_freed list and reuse it later. This simplifies the
564 * code, and it doesn't cost us much memory as the memory usage is
565 * dominated by buffers that hold the actual btree node data and those
566 * can be freed - and the number of struct btrees allocated is
567 * effectively bounded.
568 *
569 * btree_cache_freeable effectively is a small cache - we use it because
570 * high order page allocations can be rather expensive, and it's quite
571 * common to delete and allocate btree nodes in quick succession. It
572 * should never grow past ~2-3 nodes in practice.
573 */
574 struct list_head btree_cache;
575 struct list_head btree_cache_freeable;
576 struct list_head btree_cache_freed;
577
578 /* Number of elements in btree_cache + btree_cache_freeable lists */
579 unsigned int btree_cache_used;
580
581 /*
582 * If we need to allocate memory for a new btree node and that
583 * allocation fails, we can cannibalize another node in the btree cache
584 * to satisfy the allocation - lock to guarantee only one thread does
585 * this at a time:
586 */
587 wait_queue_head_t btree_cache_wait;
588 struct task_struct *btree_cache_alloc_lock;
589 spinlock_t btree_cannibalize_lock;
590
591 /*
592 * When we free a btree node, we increment the gen of the bucket the
593 * node is in - but we can't rewrite the prios and gens until we
594 * finished whatever it is we were doing, otherwise after a crash the
595 * btree node would be freed but for say a split, we might not have the
596 * pointers to the new nodes inserted into the btree yet.
597 *
598 * This is a refcount that blocks prio_write() until the new keys are
599 * written.
600 */
601 atomic_t prio_blocked;
602 wait_queue_head_t bucket_wait;
603
604 /*
605 * For any bio we don't skip we subtract the number of sectors from
606 * rescale; when it hits 0 we rescale all the bucket priorities.
607 */
608 atomic_t rescale;
609 /*
610 * used for GC, identify if any front side I/Os is inflight
611 */
612 atomic_t search_inflight;
613 /*
614 * When we invalidate buckets, we use both the priority and the amount
615 * of good data to determine which buckets to reuse first - to weight
616 * those together consistently we keep track of the smallest nonzero
617 * priority of any bucket.
618 */
619 uint16_t min_prio;
620
621 /*
622 * max(gen - last_gc) for all buckets. When it gets too big we have to
623 * gc to keep gens from wrapping around.
624 */
625 uint8_t need_gc;
626 struct gc_stat gc_stats;
627 size_t nbuckets;
628 size_t avail_nbuckets;
629
630 struct task_struct *gc_thread;
631 /* Where in the btree gc currently is */
632 struct bkey gc_done;
633
634 /*
635 * For automatical garbage collection after writeback completed, this
636 * varialbe is used as bit fields,
637 * - 0000 0001b (BCH_ENABLE_AUTO_GC): enable gc after writeback
638 * - 0000 0010b (BCH_DO_AUTO_GC): do gc after writeback
639 * This is an optimization for following write request after writeback
640 * finished, but read hit rate dropped due to clean data on cache is
641 * discarded. Unless user explicitly sets it via sysfs, it won't be
642 * enabled.
643 */
644#define BCH_ENABLE_AUTO_GC 1
645#define BCH_DO_AUTO_GC 2
646 uint8_t gc_after_writeback;
647
648 /*
649 * The allocation code needs gc_mark in struct bucket to be correct, but
650 * it's not while a gc is in progress. Protected by bucket_lock.
651 */
652 int gc_mark_valid;
653
654 /* Counts how many sectors bio_insert has added to the cache */
655 atomic_t sectors_to_gc;
656 wait_queue_head_t gc_wait;
657
658 struct keybuf moving_gc_keys;
659 /* Number of moving GC bios in flight */
660 struct semaphore moving_in_flight;
661
662 struct workqueue_struct *moving_gc_wq;
663
664 struct btree *root;
665
666#ifdef CONFIG_BCACHE_DEBUG
667 struct btree *verify_data;
668 struct bset *verify_ondisk;
669 struct mutex verify_lock;
670#endif
671
672 uint8_t set_uuid[16];
673 unsigned int nr_uuids;
674 struct uuid_entry *uuids;
675 BKEY_PADDED(uuid_bucket);
676 struct closure uuid_write;
677 struct semaphore uuid_write_mutex;
678
679 /*
680 * A btree node on disk could have too many bsets for an iterator to fit
681 * on the stack - have to dynamically allocate them.
682 * bch_cache_set_alloc() will make sure the pool can allocate iterators
683 * equipped with enough room that can host
684 * (sb.bucket_size / sb.block_size)
685 * btree_iter_sets, which is more than static MAX_BSETS.
686 */
687 mempool_t fill_iter;
688
689 struct bset_sort_state sort;
690
691 /* List of buckets we're currently writing data to */
692 struct list_head data_buckets;
693 spinlock_t data_bucket_lock;
694
695 struct journal journal;
696
697#define CONGESTED_MAX 1024
698 unsigned int congested_last_us;
699 atomic_t congested;
700
701 /* The rest of this all shows up in sysfs */
702 unsigned int congested_read_threshold_us;
703 unsigned int congested_write_threshold_us;
704
705 struct time_stats btree_gc_time;
706 struct time_stats btree_split_time;
707 struct time_stats btree_read_time;
708
709 atomic_long_t cache_read_races;
710 atomic_long_t writeback_keys_done;
711 atomic_long_t writeback_keys_failed;
712
713 atomic_long_t reclaim;
714 atomic_long_t reclaimed_journal_buckets;
715 atomic_long_t flush_write;
716
717 enum {
718 ON_ERROR_UNREGISTER,
719 ON_ERROR_PANIC,
720 } on_error;
721#define DEFAULT_IO_ERROR_LIMIT 8
722 unsigned int error_limit;
723 unsigned int error_decay;
724
725 unsigned short journal_delay_ms;
726 bool expensive_debug_checks;
727 unsigned int verify:1;
728 unsigned int key_merging_disabled:1;
729 unsigned int gc_always_rewrite:1;
730 unsigned int shrinker_disabled:1;
731 unsigned int copy_gc_enabled:1;
732 unsigned int idle_max_writeback_rate_enabled:1;
733
734#define BUCKET_HASH_BITS 12
735 struct hlist_head bucket_hash[1 << BUCKET_HASH_BITS];
736};
737
738struct bbio {
739 unsigned int submit_time_us;
740 union {
741 struct bkey key;
742 uint64_t _pad[3];
743 /*
744 * We only need pad = 3 here because we only ever carry around a
745 * single pointer - i.e. the pointer we're doing io to/from.
746 */
747 };
748 struct bio bio;
749};
750
751#define BTREE_PRIO USHRT_MAX
752#define INITIAL_PRIO 32768U
753
754#define btree_bytes(c) ((c)->btree_pages * PAGE_SIZE)
755#define btree_blocks(b) \
756 ((unsigned int) (KEY_SIZE(&b->key) >> (b)->c->block_bits))
757
758#define btree_default_blocks(c) \
759 ((unsigned int) ((PAGE_SECTORS * (c)->btree_pages) >> (c)->block_bits))
760
761#define bucket_bytes(ca) ((ca)->sb.bucket_size << 9)
762#define block_bytes(ca) ((ca)->sb.block_size << 9)
763
764static inline unsigned int meta_bucket_pages(struct cache_sb *sb)
765{
766 unsigned int n, max_pages;
767
768 max_pages = min_t(unsigned int,
769 __rounddown_pow_of_two(USHRT_MAX) / PAGE_SECTORS,
770 MAX_ORDER_NR_PAGES);
771
772 n = sb->bucket_size / PAGE_SECTORS;
773 if (n > max_pages)
774 n = max_pages;
775
776 return n;
777}
778
779static inline unsigned int meta_bucket_bytes(struct cache_sb *sb)
780{
781 return meta_bucket_pages(sb) << PAGE_SHIFT;
782}
783
784#define prios_per_bucket(ca) \
785 ((meta_bucket_bytes(&(ca)->sb) - sizeof(struct prio_set)) / \
786 sizeof(struct bucket_disk))
787
788#define prio_buckets(ca) \
789 DIV_ROUND_UP((size_t) (ca)->sb.nbuckets, prios_per_bucket(ca))
790
791static inline size_t sector_to_bucket(struct cache_set *c, sector_t s)
792{
793 return s >> c->bucket_bits;
794}
795
796static inline sector_t bucket_to_sector(struct cache_set *c, size_t b)
797{
798 return ((sector_t) b) << c->bucket_bits;
799}
800
801static inline sector_t bucket_remainder(struct cache_set *c, sector_t s)
802{
803 return s & (c->cache->sb.bucket_size - 1);
804}
805
806static inline size_t PTR_BUCKET_NR(struct cache_set *c,
807 const struct bkey *k,
808 unsigned int ptr)
809{
810 return sector_to_bucket(c, PTR_OFFSET(k, ptr));
811}
812
813static inline struct bucket *PTR_BUCKET(struct cache_set *c,
814 const struct bkey *k,
815 unsigned int ptr)
816{
817 return c->cache->buckets + PTR_BUCKET_NR(c, k, ptr);
818}
819
820static inline uint8_t gen_after(uint8_t a, uint8_t b)
821{
822 uint8_t r = a - b;
823
824 return r > 128U ? 0 : r;
825}
826
827static inline uint8_t ptr_stale(struct cache_set *c, const struct bkey *k,
828 unsigned int i)
829{
830 return gen_after(PTR_BUCKET(c, k, i)->gen, PTR_GEN(k, i));
831}
832
833static inline bool ptr_available(struct cache_set *c, const struct bkey *k,
834 unsigned int i)
835{
836 return (PTR_DEV(k, i) < MAX_CACHES_PER_SET) && c->cache;
837}
838
839/* Btree key macros */
840
841/*
842 * This is used for various on disk data structures - cache_sb, prio_set, bset,
843 * jset: The checksum is _always_ the first 8 bytes of these structs
844 */
845#define csum_set(i) \
846 bch_crc64(((void *) (i)) + sizeof(uint64_t), \
847 ((void *) bset_bkey_last(i)) - \
848 (((void *) (i)) + sizeof(uint64_t)))
849
850/* Error handling macros */
851
852#define btree_bug(b, ...) \
853do { \
854 if (bch_cache_set_error((b)->c, __VA_ARGS__)) \
855 dump_stack(); \
856} while (0)
857
858#define cache_bug(c, ...) \
859do { \
860 if (bch_cache_set_error(c, __VA_ARGS__)) \
861 dump_stack(); \
862} while (0)
863
864#define btree_bug_on(cond, b, ...) \
865do { \
866 if (cond) \
867 btree_bug(b, __VA_ARGS__); \
868} while (0)
869
870#define cache_bug_on(cond, c, ...) \
871do { \
872 if (cond) \
873 cache_bug(c, __VA_ARGS__); \
874} while (0)
875
876#define cache_set_err_on(cond, c, ...) \
877do { \
878 if (cond) \
879 bch_cache_set_error(c, __VA_ARGS__); \
880} while (0)
881
882/* Looping macros */
883
884#define for_each_bucket(b, ca) \
885 for (b = (ca)->buckets + (ca)->sb.first_bucket; \
886 b < (ca)->buckets + (ca)->sb.nbuckets; b++)
887
888static inline void cached_dev_put(struct cached_dev *dc)
889{
890 if (refcount_dec_and_test(&dc->count))
891 schedule_work(&dc->detach);
892}
893
894static inline bool cached_dev_get(struct cached_dev *dc)
895{
896 if (!refcount_inc_not_zero(&dc->count))
897 return false;
898
899 /* Paired with the mb in cached_dev_attach */
900 smp_mb__after_atomic();
901 return true;
902}
903
904/*
905 * bucket_gc_gen() returns the difference between the bucket's current gen and
906 * the oldest gen of any pointer into that bucket in the btree (last_gc).
907 */
908
909static inline uint8_t bucket_gc_gen(struct bucket *b)
910{
911 return b->gen - b->last_gc;
912}
913
914#define BUCKET_GC_GEN_MAX 96U
915
916#define kobj_attribute_write(n, fn) \
917 static struct kobj_attribute ksysfs_##n = __ATTR(n, 0200, NULL, fn)
918
919#define kobj_attribute_rw(n, show, store) \
920 static struct kobj_attribute ksysfs_##n = \
921 __ATTR(n, 0600, show, store)
922
923static inline void wake_up_allocators(struct cache_set *c)
924{
925 struct cache *ca = c->cache;
926
927 wake_up_process(ca->alloc_thread);
928}
929
930static inline void closure_bio_submit(struct cache_set *c,
931 struct bio *bio,
932 struct closure *cl)
933{
934 closure_get(cl);
935 if (unlikely(test_bit(CACHE_SET_IO_DISABLE, &c->flags))) {
936 bio->bi_status = BLK_STS_IOERR;
937 bio_endio(bio);
938 return;
939 }
940 submit_bio_noacct(bio);
941}
942
943/*
944 * Prevent the kthread exits directly, and make sure when kthread_stop()
945 * is called to stop a kthread, it is still alive. If a kthread might be
946 * stopped by CACHE_SET_IO_DISABLE bit set, wait_for_kthread_stop() is
947 * necessary before the kthread returns.
948 */
949static inline void wait_for_kthread_stop(void)
950{
951 while (!kthread_should_stop()) {
952 set_current_state(TASK_INTERRUPTIBLE);
953 schedule();
954 }
955}
956
957/* Forward declarations */
958
959void bch_count_backing_io_errors(struct cached_dev *dc, struct bio *bio);
960void bch_count_io_errors(struct cache *ca, blk_status_t error,
961 int is_read, const char *m);
962void bch_bbio_count_io_errors(struct cache_set *c, struct bio *bio,
963 blk_status_t error, const char *m);
964void bch_bbio_endio(struct cache_set *c, struct bio *bio,
965 blk_status_t error, const char *m);
966void bch_bbio_free(struct bio *bio, struct cache_set *c);
967struct bio *bch_bbio_alloc(struct cache_set *c);
968
969void __bch_submit_bbio(struct bio *bio, struct cache_set *c);
970void bch_submit_bbio(struct bio *bio, struct cache_set *c,
971 struct bkey *k, unsigned int ptr);
972
973uint8_t bch_inc_gen(struct cache *ca, struct bucket *b);
974void bch_rescale_priorities(struct cache_set *c, int sectors);
975
976bool bch_can_invalidate_bucket(struct cache *ca, struct bucket *b);
977void __bch_invalidate_one_bucket(struct cache *ca, struct bucket *b);
978
979void __bch_bucket_free(struct cache *ca, struct bucket *b);
980void bch_bucket_free(struct cache_set *c, struct bkey *k);
981
982long bch_bucket_alloc(struct cache *ca, unsigned int reserve, bool wait);
983int __bch_bucket_alloc_set(struct cache_set *c, unsigned int reserve,
984 struct bkey *k, bool wait);
985int bch_bucket_alloc_set(struct cache_set *c, unsigned int reserve,
986 struct bkey *k, bool wait);
987bool bch_alloc_sectors(struct cache_set *c, struct bkey *k,
988 unsigned int sectors, unsigned int write_point,
989 unsigned int write_prio, bool wait);
990bool bch_cached_dev_error(struct cached_dev *dc);
991
992__printf(2, 3)
993bool bch_cache_set_error(struct cache_set *c, const char *fmt, ...);
994
995int bch_prio_write(struct cache *ca, bool wait);
996void bch_write_bdev_super(struct cached_dev *dc, struct closure *parent);
997
998extern struct workqueue_struct *bcache_wq;
999extern struct workqueue_struct *bch_journal_wq;
1000extern struct workqueue_struct *bch_flush_wq;
1001extern struct mutex bch_register_lock;
1002extern struct list_head bch_cache_sets;
1003
1004extern struct kobj_type bch_cached_dev_ktype;
1005extern struct kobj_type bch_flash_dev_ktype;
1006extern struct kobj_type bch_cache_set_ktype;
1007extern struct kobj_type bch_cache_set_internal_ktype;
1008extern struct kobj_type bch_cache_ktype;
1009
1010void bch_cached_dev_release(struct kobject *kobj);
1011void bch_flash_dev_release(struct kobject *kobj);
1012void bch_cache_set_release(struct kobject *kobj);
1013void bch_cache_release(struct kobject *kobj);
1014
1015int bch_uuid_write(struct cache_set *c);
1016void bcache_write_super(struct cache_set *c);
1017
1018int bch_flash_dev_create(struct cache_set *c, uint64_t size);
1019
1020int bch_cached_dev_attach(struct cached_dev *dc, struct cache_set *c,
1021 uint8_t *set_uuid);
1022void bch_cached_dev_detach(struct cached_dev *dc);
1023int bch_cached_dev_run(struct cached_dev *dc);
1024void bcache_device_stop(struct bcache_device *d);
1025
1026void bch_cache_set_unregister(struct cache_set *c);
1027void bch_cache_set_stop(struct cache_set *c);
1028
1029struct cache_set *bch_cache_set_alloc(struct cache_sb *sb);
1030void bch_btree_cache_free(struct cache_set *c);
1031int bch_btree_cache_alloc(struct cache_set *c);
1032void bch_moving_init_cache_set(struct cache_set *c);
1033int bch_open_buckets_alloc(struct cache_set *c);
1034void bch_open_buckets_free(struct cache_set *c);
1035
1036int bch_cache_allocator_start(struct cache *ca);
1037
1038void bch_debug_exit(void);
1039void bch_debug_init(void);
1040void bch_request_exit(void);
1041int bch_request_init(void);
1042void bch_btree_exit(void);
1043int bch_btree_init(void);
1044
1045#endif /* _BCACHE_H */