5 * SOME HIGH LEVEL CODE DOCUMENTATION:
7 * Bcache mostly works with cache sets, cache devices, and backing devices.
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.
14 * Multiple cache devices is intended to give us the ability to mirror dirty
15 * cached data and metadata, without mirroring clean cached data.
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.
23 * A cache set can have multiple (many) backing devices attached to it.
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.
31 * Flash only volumes work but they're not production ready because the moving
32 * garbage collector needs more work. More on that later.
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
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+
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).
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.
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
61 * Bcache is entirely COW - we never write twice to a bucket, even buckets that
62 * contain metadata (including btree nodes).
66 * Bcache is in large part design around the btree.
68 * At a high level, the btree is just an index of key -> ptr tuples.
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).
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.
79 * Pointers contain the cache device id, the offset on that device, and an 8 bit
80 * generation number. More on the gen later.
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
86 * Updates are fairly well abstracted, though. There are two different ways of
87 * updating the btree; insert and replace.
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.
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.
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.
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.
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.
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.)
117 * Anyways, btree nodes are big - big enough to be inefficient with a textbook
118 * btree implementation.
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.
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
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.
136 * GARBAGE COLLECTION:
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.
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.
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.
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
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.
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
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()).
179 #define pr_fmt(fmt) "bcachefs: %s() " fmt "\n", __func__
181 #include <linux/bug.h>
182 #include <linux/bio.h>
183 #include <linux/closure.h>
184 #include <linux/kobject.h>
185 #include <linux/lglock.h>
186 #include <linux/list.h>
187 #include <linux/mutex.h>
188 #include <linux/percpu-refcount.h>
189 #include <linux/radix-tree.h>
190 #include <linux/rbtree.h>
191 #include <linux/rhashtable.h>
192 #include <linux/rwsem.h>
193 #include <linux/seqlock.h>
194 #include <linux/shrinker.h>
195 #include <linux/types.h>
196 #include <linux/workqueue.h>
197 #include <linux/zstd.h>
199 #include "bcachefs_format.h"
204 #include <linux/dynamic_fault.h>
206 #define bch2_fs_init_fault(name) \
207 dynamic_fault("bcachefs:bch_fs_init:" name)
208 #define bch2_meta_read_fault(name) \
209 dynamic_fault("bcachefs:meta:read:" name)
210 #define bch2_meta_write_fault(name) \
211 dynamic_fault("bcachefs:meta:write:" name)
214 #define bch2_fmt(_c, fmt) "bcachefs (%s): " fmt "\n", ((_c)->name)
216 #define bch2_fmt(_c, fmt) fmt "\n"
219 #define bch_info(c, fmt, ...) \
220 printk(KERN_INFO bch2_fmt(c, fmt), ##__VA_ARGS__)
221 #define bch_notice(c, fmt, ...) \
222 printk(KERN_NOTICE bch2_fmt(c, fmt), ##__VA_ARGS__)
223 #define bch_warn(c, fmt, ...) \
224 printk(KERN_WARNING bch2_fmt(c, fmt), ##__VA_ARGS__)
225 #define bch_err(c, fmt, ...) \
226 printk(KERN_ERR bch2_fmt(c, fmt), ##__VA_ARGS__)
228 #define bch_verbose(c, fmt, ...) \
230 if ((c)->opts.verbose_recovery) \
231 bch_info(c, fmt, ##__VA_ARGS__); \
234 #define pr_verbose_init(opts, fmt, ...) \
236 if (opt_get(opts, verbose_init)) \
237 pr_info(fmt, ##__VA_ARGS__); \
240 /* Parameters that are useful for debugging, but should always be compiled in: */
241 #define BCH_DEBUG_PARAMS_ALWAYS() \
242 BCH_DEBUG_PARAM(key_merging_disabled, \
243 "Disables merging of extents") \
244 BCH_DEBUG_PARAM(btree_gc_always_rewrite, \
245 "Causes mark and sweep to compact and rewrite every " \
246 "btree node it traverses") \
247 BCH_DEBUG_PARAM(btree_gc_rewrite_disabled, \
248 "Disables rewriting of btree nodes during mark and sweep")\
249 BCH_DEBUG_PARAM(btree_shrinker_disabled, \
250 "Disables the shrinker callback for the btree node cache")
252 /* Parameters that should only be compiled in in debug mode: */
253 #define BCH_DEBUG_PARAMS_DEBUG() \
254 BCH_DEBUG_PARAM(expensive_debug_checks, \
255 "Enables various runtime debugging checks that " \
256 "significantly affect performance") \
257 BCH_DEBUG_PARAM(debug_check_bkeys, \
258 "Run bkey_debugcheck (primarily checking GC/allocation "\
259 "information) when iterating over keys") \
260 BCH_DEBUG_PARAM(verify_btree_ondisk, \
261 "Reread btree nodes at various points to verify the " \
262 "mergesort in the read path against modifications " \
265 #define BCH_DEBUG_PARAMS_ALL() BCH_DEBUG_PARAMS_ALWAYS() BCH_DEBUG_PARAMS_DEBUG()
267 #ifdef CONFIG_BCACHEFS_DEBUG
268 #define BCH_DEBUG_PARAMS() BCH_DEBUG_PARAMS_ALL()
270 #define BCH_DEBUG_PARAMS() BCH_DEBUG_PARAMS_ALWAYS()
273 #define BCH_TIME_STATS() \
274 x(btree_node_mem_alloc) \
279 x(btree_lock_contended_read) \
280 x(btree_lock_contended_intent) \
281 x(btree_lock_contended_write) \
290 enum bch_time_stats {
291 #define x(name) BCH_TIME_##name,
297 #include "alloc_types.h"
298 #include "btree_types.h"
299 #include "buckets_types.h"
300 #include "clock_types.h"
301 #include "journal_types.h"
302 #include "keylist_types.h"
303 #include "quota_types.h"
304 #include "rebalance_types.h"
305 #include "super_types.h"
308 * Number of nodes we might have to allocate in a worst case btree split
309 * operation - we split all the way up to the root, then allocate a new root.
311 #define btree_reserve_required_nodes(depth) (((depth) + 1) * 2 + 1)
313 /* Number of nodes btree coalesce will try to coalesce at once */
314 #define GC_MERGE_NODES 4U
316 /* Maximum number of nodes we might need to allocate atomically: */
317 #define BTREE_RESERVE_MAX \
318 (btree_reserve_required_nodes(BTREE_MAX_DEPTH) + GC_MERGE_NODES)
320 /* Size of the freelist we allocate btree nodes from: */
321 #define BTREE_NODE_RESERVE (BTREE_RESERVE_MAX * 4)
324 struct crypto_blkcipher;
328 GC_PHASE_SB = BTREE_ID_NR + 1,
329 GC_PHASE_PENDING_DELETE,
341 u64 sectors[2][BCH_DATA_NR];
346 struct percpu_ref ref;
347 struct completion ref_completion;
348 struct percpu_ref io_ref;
349 struct completion io_ref_completion;
355 * Cached version of this device's member info from superblock
356 * Committed by bch2_write_super() -> bch_fs_mi_update()
358 struct bch_member_cpu mi;
360 char name[BDEVNAME_SIZE];
362 struct bch_sb_handle disk_sb;
365 struct bch_devs_mask self;
367 /* biosets used in cloned bios for writing multiple replicas */
368 struct bio_set replica_set;
372 * Per-bucket arrays are protected by c->usage_lock, bucket_lock and
373 * gc_lock, for device resize - holding any is sufficient for access:
374 * Or rcu_read_lock(), but only for ptr_stale():
376 struct bucket_array __rcu *buckets;
377 unsigned long *buckets_dirty;
378 /* most out of date gen in the btree */
380 struct rw_semaphore bucket_lock;
382 struct bch_dev_usage __percpu *usage_percpu;
383 struct bch_dev_usage usage_cached;
386 struct task_struct __rcu *alloc_thread;
389 * free: Buckets that are ready to be used
391 * free_inc: Incoming buckets - these are buckets that currently have
392 * cached data in them, and we can't reuse them until after we write
393 * their new gen to disk. After prio_write() finishes writing the new
394 * gens/prios, they'll be moved to the free list (and possibly discarded
397 alloc_fifo free[RESERVE_NR];
399 spinlock_t freelist_lock;
400 size_t nr_invalidated;
402 u8 open_buckets_partial[OPEN_BUCKETS_COUNT];
403 unsigned open_buckets_partial_nr;
405 size_t fifo_last_bucket;
407 /* last calculated minimum prio */
408 u16 max_last_bucket_io[2];
410 atomic_long_t saturated_count;
411 size_t inc_gen_needs_gc;
412 size_t inc_gen_really_needs_gc;
413 u64 allocator_journal_seq_flush;
414 bool allocator_invalidating_data;
415 bool allocator_blocked;
417 alloc_heap alloc_heap;
420 struct task_struct *copygc_thread;
421 copygc_heap copygc_heap;
422 struct bch_pd_controller copygc_pd;
423 struct write_point copygc_write_point;
425 atomic64_t rebalance_work;
427 struct journal_device journal;
429 struct work_struct io_error_work;
431 /* The rest of this all shows up in sysfs */
432 atomic64_t cur_latency[2];
433 struct time_stats io_latency[2];
435 #define CONGESTED_MAX 1024
439 struct io_count __percpu *io_done;
443 * Flag bits for what phase of startup/shutdown the cache set is at, how we're
444 * shutting down, etc.:
446 * BCH_FS_UNREGISTERING means we're not just shutting down, we're detaching
447 * all the backing devices first (their cached data gets invalidated, and they
448 * won't automatically reattach).
452 BCH_FS_ALLOC_READ_DONE,
453 BCH_FS_ALLOCATOR_STARTED,
454 BCH_FS_INITIAL_GC_DONE,
460 BCH_FS_WRITE_DISABLE_COMPLETE,
468 BCH_FS_FSCK_FIXED_ERRORS,
470 BCH_FS_REBUILD_REPLICAS,
471 BCH_FS_HOLD_BTREE_WRITES,
476 struct dentry *btree;
477 struct dentry *btree_format;
478 struct dentry *failed;
491 struct list_head list;
493 struct kobject internal;
494 struct kobject opts_dir;
495 struct kobject time_stats;
499 struct device *chardev;
500 struct super_block *vfs_sb;
503 /* ro/rw, add/remove devices: */
504 struct mutex state_lock;
505 enum bch_fs_state state;
507 /* Counts outstanding writes, for clean transition to read-only */
508 struct percpu_ref writes;
509 struct work_struct read_only_work;
511 struct bch_dev __rcu *devs[BCH_SB_MEMBERS_MAX];
513 struct bch_replicas_cpu __rcu *replicas;
514 struct bch_replicas_cpu __rcu *replicas_gc;
515 struct mutex replicas_gc_lock;
517 struct bch_disk_groups_cpu __rcu *disk_groups;
519 struct bch_opts opts;
521 /* Updated by bch2_sb_update():*/
526 u16 encoded_extent_max;
539 struct bch_sb_handle disk_sb;
541 unsigned short block_bits; /* ilog2(block_size) */
543 u16 btree_foreground_merge_threshold;
545 struct closure sb_write;
546 struct mutex sb_lock;
549 struct bio_set btree_bio;
551 struct btree_root btree_roots[BTREE_ID_NR];
552 bool btree_roots_dirty;
553 struct mutex btree_root_lock;
555 struct btree_cache btree_cache;
557 mempool_t btree_reserve_pool;
560 * Cache of allocated btree nodes - if we allocate a btree node and
561 * don't use it, if we free it that space can't be reused until going
562 * _all_ the way through the allocator (which exposes us to a livelock
563 * when allocating btree reserves fail halfway through) - instead, we
564 * can stick them here:
566 struct btree_alloc btree_reserve_cache[BTREE_NODE_RESERVE * 2];
567 unsigned btree_reserve_cache_nr;
568 struct mutex btree_reserve_cache_lock;
570 mempool_t btree_interior_update_pool;
571 struct list_head btree_interior_update_list;
572 struct mutex btree_interior_update_lock;
573 struct closure_waitlist btree_interior_update_wait;
575 struct workqueue_struct *wq;
576 /* copygc needs its own workqueue for index updates.. */
577 struct workqueue_struct *copygc_wq;
580 struct delayed_work pd_controllers_update;
581 unsigned pd_controllers_update_seconds;
583 struct bch_devs_mask rw_devs[BCH_DATA_NR];
585 u64 capacity; /* sectors */
588 * When capacity _decreases_ (due to a disk being removed), we
589 * increment capacity_gen - this invalidates outstanding reservations
590 * and forces them to be revalidated
594 atomic64_t sectors_available;
596 struct bch_fs_usage __percpu *usage_percpu;
597 struct bch_fs_usage usage_cached;
598 struct lglock usage_lock;
600 struct closure_waitlist freelist_wait;
603 * When we invalidate buckets, we use both the priority and the amount
604 * of good data to determine which buckets to reuse first - to weight
605 * those together consistently we keep track of the smallest nonzero
606 * priority of any bucket.
608 struct bucket_clock bucket_clock[2];
610 struct io_clock io_clock[2];
613 spinlock_t freelist_lock;
614 u8 open_buckets_freelist;
615 u8 open_buckets_nr_free;
616 struct closure_waitlist open_buckets_wait;
617 struct open_bucket open_buckets[OPEN_BUCKETS_COUNT];
619 struct write_point btree_write_point;
620 struct write_point rebalance_write_point;
622 struct write_point write_points[WRITE_POINT_COUNT];
623 struct hlist_head write_points_hash[WRITE_POINT_COUNT];
624 struct mutex write_points_hash_lock;
626 /* GARBAGE COLLECTION */
627 struct task_struct *gc_thread;
629 unsigned long gc_count;
632 * Tracks GC's progress - everything in the range [ZERO_KEY..gc_cur_pos]
633 * has been marked by GC.
635 * gc_cur_phase is a superset of btree_ids (BTREE_ID_EXTENTS etc.)
637 * gc_cur_phase == GC_PHASE_DONE indicates that gc is finished/not
638 * currently running, and gc marks are currently valid
640 * Protected by gc_pos_lock. Only written to by GC thread, so GC thread
641 * can read without a lock.
643 seqcount_t gc_pos_lock;
644 struct gc_pos gc_pos;
647 * The allocation code needs gc_mark in struct bucket to be correct, but
648 * it's not while a gc is in progress.
650 struct rw_semaphore gc_lock;
653 struct bio_set bio_read;
654 struct bio_set bio_read_split;
655 struct bio_set bio_write;
656 struct mutex bio_bounce_pages_lock;
657 mempool_t bio_bounce_pages;
658 struct rhashtable promote_table;
660 mempool_t compression_bounce[2];
661 mempool_t compress_workspace[BCH_COMPRESSION_NR];
662 mempool_t decompress_workspace;
663 ZSTD_parameters zstd_params;
665 struct crypto_shash *sha256;
666 struct crypto_skcipher *chacha20;
667 struct crypto_shash *poly1305;
669 atomic64_t key_version;
672 struct bch_fs_rebalance rebalance;
674 /* VFS IO PATH - fs-io.c */
675 struct bio_set writepage_bioset;
676 struct bio_set dio_write_bioset;
677 struct bio_set dio_read_bioset;
679 struct bio_list btree_write_error_list;
680 struct work_struct btree_write_error_work;
681 spinlock_t btree_write_error_lock;
684 struct list_head fsck_errors;
685 struct mutex fsck_error_lock;
689 atomic_long_t nr_inodes;
692 struct bch_memquota_type quotas[QTYP_NR];
695 struct dentry *debug;
696 struct btree_debug btree_debug[BTREE_ID_NR];
697 #ifdef CONFIG_BCACHEFS_DEBUG
698 struct btree *verify_data;
699 struct btree_node *verify_ondisk;
700 struct mutex verify_lock;
703 u64 unused_inode_hint;
706 * A btree node on disk could have too many bsets for an iterator to fit
707 * on the stack - have to dynamically allocate them
711 mempool_t btree_bounce_pool;
713 struct journal journal;
715 unsigned bucket_journal_seq;
717 /* The rest of this all shows up in sysfs */
718 atomic_long_t read_realloc_races;
719 atomic_long_t extent_migrate_done;
720 atomic_long_t extent_migrate_raced;
722 unsigned btree_gc_periodic:1;
723 unsigned copy_gc_enabled:1;
724 bool promote_whole_extents;
726 #define BCH_DEBUG_PARAM(name, description) bool name;
727 BCH_DEBUG_PARAMS_ALL()
728 #undef BCH_DEBUG_PARAM
730 struct time_stats times[BCH_TIME_STAT_NR];
733 static inline void bch2_set_ra_pages(struct bch_fs *c, unsigned ra_pages)
735 #ifndef NO_BCACHEFS_FS
737 c->vfs_sb->s_bdi->ra_pages = ra_pages;
741 static inline bool bch2_fs_running(struct bch_fs *c)
743 return c->state == BCH_FS_RO || c->state == BCH_FS_RW;
746 static inline unsigned bucket_bytes(const struct bch_dev *ca)
748 return ca->mi.bucket_size << 9;
751 static inline unsigned block_bytes(const struct bch_fs *c)
753 return c->opts.block_size << 9;
756 #endif /* _BCACHEFS_H */