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