xref: /openbmc/linux/drivers/md/bcache/bcache.h (revision 1cac4f26)
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 
197 struct 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 
210 BITMASK(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))
216 BITMASK(GC_SECTORS_USED, struct bucket, gc_mark, 2, GC_SECTORS_USED_SIZE);
217 BITMASK(GC_MOVE, struct bucket, gc_mark, 15, 1);
218 
219 #include "journal.h"
220 #include "stats.h"
221 struct search;
222 struct btree;
223 struct keybuf;
224 
225 struct keybuf_key {
226 	struct rb_node		node;
227 	BKEY_PADDED(key);
228 	void			*private;
229 };
230 
231 struct 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 
249 struct 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 	unsigned 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 
282 struct 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 
292 enum 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 
298 struct 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 bio		sb_bio;
305 	struct bio_vec		sb_bv[1];
306 	struct closure		sb_write;
307 	struct semaphore	sb_write_mutex;
308 
309 	/* Refcount on the cache set. Always nonzero when we're caching. */
310 	refcount_t		count;
311 	struct work_struct	detach;
312 
313 	/*
314 	 * Device might not be running if it's dirty and the cache set hasn't
315 	 * showed up yet.
316 	 */
317 	atomic_t		running;
318 
319 	/*
320 	 * Writes take a shared lock from start to finish; scanning for dirty
321 	 * data to refill the rb tree requires an exclusive lock.
322 	 */
323 	struct rw_semaphore	writeback_lock;
324 
325 	/*
326 	 * Nonzero, and writeback has a refcount (d->count), iff there is dirty
327 	 * data in the cache. Protected by writeback_lock; must have an
328 	 * shared lock to set and exclusive lock to clear.
329 	 */
330 	atomic_t		has_dirty;
331 
332 	struct bch_ratelimit	writeback_rate;
333 	struct delayed_work	writeback_rate_update;
334 
335 	/* Limit number of writeback bios in flight */
336 	struct semaphore	in_flight;
337 	struct task_struct	*writeback_thread;
338 	struct workqueue_struct	*writeback_write_wq;
339 
340 	struct keybuf		writeback_keys;
341 
342 	struct task_struct	*status_update_thread;
343 	/*
344 	 * Order the write-half of writeback operations strongly in dispatch
345 	 * order.  (Maintain LBA order; don't allow reads completing out of
346 	 * order to re-order the writes...)
347 	 */
348 	struct closure_waitlist writeback_ordering_wait;
349 	atomic_t		writeback_sequence_next;
350 
351 	/* For tracking sequential IO */
352 #define RECENT_IO_BITS	7
353 #define RECENT_IO	(1 << RECENT_IO_BITS)
354 	struct io		io[RECENT_IO];
355 	struct hlist_head	io_hash[RECENT_IO + 1];
356 	struct list_head	io_lru;
357 	spinlock_t		io_lock;
358 
359 	struct cache_accounting	accounting;
360 
361 	/* The rest of this all shows up in sysfs */
362 	unsigned int		sequential_cutoff;
363 	unsigned int		readahead;
364 
365 	unsigned int		io_disable:1;
366 	unsigned int		verify:1;
367 	unsigned int		bypass_torture_test:1;
368 
369 	unsigned int		partial_stripes_expensive:1;
370 	unsigned int		writeback_metadata:1;
371 	unsigned int		writeback_running:1;
372 	unsigned char		writeback_percent;
373 	unsigned int		writeback_delay;
374 
375 	uint64_t		writeback_rate_target;
376 	int64_t			writeback_rate_proportional;
377 	int64_t			writeback_rate_integral;
378 	int64_t			writeback_rate_integral_scaled;
379 	int32_t			writeback_rate_change;
380 
381 	unsigned int		writeback_rate_update_seconds;
382 	unsigned int		writeback_rate_i_term_inverse;
383 	unsigned int		writeback_rate_p_term_inverse;
384 	unsigned int		writeback_rate_minimum;
385 
386 	enum stop_on_failure	stop_when_cache_set_failed;
387 #define DEFAULT_CACHED_DEV_ERROR_LIMIT	64
388 	atomic_t		io_errors;
389 	unsigned int		error_limit;
390 	unsigned int		offline_seconds;
391 
392 	char			backing_dev_name[BDEVNAME_SIZE];
393 };
394 
395 enum alloc_reserve {
396 	RESERVE_BTREE,
397 	RESERVE_PRIO,
398 	RESERVE_MOVINGGC,
399 	RESERVE_NONE,
400 	RESERVE_NR,
401 };
402 
403 struct cache {
404 	struct cache_set	*set;
405 	struct cache_sb		sb;
406 	struct bio		sb_bio;
407 	struct bio_vec		sb_bv[1];
408 
409 	struct kobject		kobj;
410 	struct block_device	*bdev;
411 
412 	struct task_struct	*alloc_thread;
413 
414 	struct closure		prio;
415 	struct prio_set		*disk_buckets;
416 
417 	/*
418 	 * When allocating new buckets, prio_write() gets first dibs - since we
419 	 * may not be allocate at all without writing priorities and gens.
420 	 * prio_last_buckets[] contains the last buckets we wrote priorities to
421 	 * (so gc can mark them as metadata), prio_buckets[] contains the
422 	 * buckets allocated for the next prio write.
423 	 */
424 	uint64_t		*prio_buckets;
425 	uint64_t		*prio_last_buckets;
426 
427 	/*
428 	 * free: Buckets that are ready to be used
429 	 *
430 	 * free_inc: Incoming buckets - these are buckets that currently have
431 	 * cached data in them, and we can't reuse them until after we write
432 	 * their new gen to disk. After prio_write() finishes writing the new
433 	 * gens/prios, they'll be moved to the free list (and possibly discarded
434 	 * in the process)
435 	 */
436 	DECLARE_FIFO(long, free)[RESERVE_NR];
437 	DECLARE_FIFO(long, free_inc);
438 
439 	size_t			fifo_last_bucket;
440 
441 	/* Allocation stuff: */
442 	struct bucket		*buckets;
443 
444 	DECLARE_HEAP(struct bucket *, heap);
445 
446 	/*
447 	 * If nonzero, we know we aren't going to find any buckets to invalidate
448 	 * until a gc finishes - otherwise we could pointlessly burn a ton of
449 	 * cpu
450 	 */
451 	unsigned int		invalidate_needs_gc;
452 
453 	bool			discard; /* Get rid of? */
454 
455 	struct journal_device	journal;
456 
457 	/* The rest of this all shows up in sysfs */
458 #define IO_ERROR_SHIFT		20
459 	atomic_t		io_errors;
460 	atomic_t		io_count;
461 
462 	atomic_long_t		meta_sectors_written;
463 	atomic_long_t		btree_sectors_written;
464 	atomic_long_t		sectors_written;
465 
466 	char			cache_dev_name[BDEVNAME_SIZE];
467 };
468 
469 struct gc_stat {
470 	size_t			nodes;
471 	size_t			nodes_pre;
472 	size_t			key_bytes;
473 
474 	size_t			nkeys;
475 	uint64_t		data;	/* sectors */
476 	unsigned int		in_use; /* percent */
477 };
478 
479 /*
480  * Flag bits, for how the cache set is shutting down, and what phase it's at:
481  *
482  * CACHE_SET_UNREGISTERING means we're not just shutting down, we're detaching
483  * all the backing devices first (their cached data gets invalidated, and they
484  * won't automatically reattach).
485  *
486  * CACHE_SET_STOPPING always gets set first when we're closing down a cache set;
487  * we'll continue to run normally for awhile with CACHE_SET_STOPPING set (i.e.
488  * flushing dirty data).
489  *
490  * CACHE_SET_RUNNING means all cache devices have been registered and journal
491  * replay is complete.
492  *
493  * CACHE_SET_IO_DISABLE is set when bcache is stopping the whold cache set, all
494  * external and internal I/O should be denied when this flag is set.
495  *
496  */
497 #define CACHE_SET_UNREGISTERING		0
498 #define	CACHE_SET_STOPPING		1
499 #define	CACHE_SET_RUNNING		2
500 #define CACHE_SET_IO_DISABLE		3
501 
502 struct cache_set {
503 	struct closure		cl;
504 
505 	struct list_head	list;
506 	struct kobject		kobj;
507 	struct kobject		internal;
508 	struct dentry		*debug;
509 	struct cache_accounting accounting;
510 
511 	unsigned long		flags;
512 	atomic_t		idle_counter;
513 	atomic_t		at_max_writeback_rate;
514 
515 	struct cache_sb		sb;
516 
517 	struct cache		*cache[MAX_CACHES_PER_SET];
518 	struct cache		*cache_by_alloc[MAX_CACHES_PER_SET];
519 	int			caches_loaded;
520 
521 	struct bcache_device	**devices;
522 	unsigned int		devices_max_used;
523 	atomic_t		attached_dev_nr;
524 	struct list_head	cached_devs;
525 	uint64_t		cached_dev_sectors;
526 	atomic_long_t		flash_dev_dirty_sectors;
527 	struct closure		caching;
528 
529 	struct closure		sb_write;
530 	struct semaphore	sb_write_mutex;
531 
532 	mempool_t		search;
533 	mempool_t		bio_meta;
534 	struct bio_set		bio_split;
535 
536 	/* For the btree cache */
537 	struct shrinker		shrink;
538 
539 	/* For the btree cache and anything allocation related */
540 	struct mutex		bucket_lock;
541 
542 	/* log2(bucket_size), in sectors */
543 	unsigned short		bucket_bits;
544 
545 	/* log2(block_size), in sectors */
546 	unsigned short		block_bits;
547 
548 	/*
549 	 * Default number of pages for a new btree node - may be less than a
550 	 * full bucket
551 	 */
552 	unsigned int		btree_pages;
553 
554 	/*
555 	 * Lists of struct btrees; lru is the list for structs that have memory
556 	 * allocated for actual btree node, freed is for structs that do not.
557 	 *
558 	 * We never free a struct btree, except on shutdown - we just put it on
559 	 * the btree_cache_freed list and reuse it later. This simplifies the
560 	 * code, and it doesn't cost us much memory as the memory usage is
561 	 * dominated by buffers that hold the actual btree node data and those
562 	 * can be freed - and the number of struct btrees allocated is
563 	 * effectively bounded.
564 	 *
565 	 * btree_cache_freeable effectively is a small cache - we use it because
566 	 * high order page allocations can be rather expensive, and it's quite
567 	 * common to delete and allocate btree nodes in quick succession. It
568 	 * should never grow past ~2-3 nodes in practice.
569 	 */
570 	struct list_head	btree_cache;
571 	struct list_head	btree_cache_freeable;
572 	struct list_head	btree_cache_freed;
573 
574 	/* Number of elements in btree_cache + btree_cache_freeable lists */
575 	unsigned int		btree_cache_used;
576 
577 	/*
578 	 * If we need to allocate memory for a new btree node and that
579 	 * allocation fails, we can cannibalize another node in the btree cache
580 	 * to satisfy the allocation - lock to guarantee only one thread does
581 	 * this at a time:
582 	 */
583 	wait_queue_head_t	btree_cache_wait;
584 	struct task_struct	*btree_cache_alloc_lock;
585 
586 	/*
587 	 * When we free a btree node, we increment the gen of the bucket the
588 	 * node is in - but we can't rewrite the prios and gens until we
589 	 * finished whatever it is we were doing, otherwise after a crash the
590 	 * btree node would be freed but for say a split, we might not have the
591 	 * pointers to the new nodes inserted into the btree yet.
592 	 *
593 	 * This is a refcount that blocks prio_write() until the new keys are
594 	 * written.
595 	 */
596 	atomic_t		prio_blocked;
597 	wait_queue_head_t	bucket_wait;
598 
599 	/*
600 	 * For any bio we don't skip we subtract the number of sectors from
601 	 * rescale; when it hits 0 we rescale all the bucket priorities.
602 	 */
603 	atomic_t		rescale;
604 	/*
605 	 * used for GC, identify if any front side I/Os is inflight
606 	 */
607 	atomic_t		search_inflight;
608 	/*
609 	 * When we invalidate buckets, we use both the priority and the amount
610 	 * of good data to determine which buckets to reuse first - to weight
611 	 * those together consistently we keep track of the smallest nonzero
612 	 * priority of any bucket.
613 	 */
614 	uint16_t		min_prio;
615 
616 	/*
617 	 * max(gen - last_gc) for all buckets. When it gets too big we have to
618 	 * gc to keep gens from wrapping around.
619 	 */
620 	uint8_t			need_gc;
621 	struct gc_stat		gc_stats;
622 	size_t			nbuckets;
623 	size_t			avail_nbuckets;
624 
625 	struct task_struct	*gc_thread;
626 	/* Where in the btree gc currently is */
627 	struct bkey		gc_done;
628 
629 	/*
630 	 * For automatical garbage collection after writeback completed, this
631 	 * varialbe is used as bit fields,
632 	 * - 0000 0001b (BCH_ENABLE_AUTO_GC): enable gc after writeback
633 	 * - 0000 0010b (BCH_DO_AUTO_GC):     do gc after writeback
634 	 * This is an optimization for following write request after writeback
635 	 * finished, but read hit rate dropped due to clean data on cache is
636 	 * discarded. Unless user explicitly sets it via sysfs, it won't be
637 	 * enabled.
638 	 */
639 #define BCH_ENABLE_AUTO_GC	1
640 #define BCH_DO_AUTO_GC		2
641 	uint8_t			gc_after_writeback;
642 
643 	/*
644 	 * The allocation code needs gc_mark in struct bucket to be correct, but
645 	 * it's not while a gc is in progress. Protected by bucket_lock.
646 	 */
647 	int			gc_mark_valid;
648 
649 	/* Counts how many sectors bio_insert has added to the cache */
650 	atomic_t		sectors_to_gc;
651 	wait_queue_head_t	gc_wait;
652 
653 	struct keybuf		moving_gc_keys;
654 	/* Number of moving GC bios in flight */
655 	struct semaphore	moving_in_flight;
656 
657 	struct workqueue_struct	*moving_gc_wq;
658 
659 	struct btree		*root;
660 
661 #ifdef CONFIG_BCACHE_DEBUG
662 	struct btree		*verify_data;
663 	struct bset		*verify_ondisk;
664 	struct mutex		verify_lock;
665 #endif
666 
667 	unsigned int		nr_uuids;
668 	struct uuid_entry	*uuids;
669 	BKEY_PADDED(uuid_bucket);
670 	struct closure		uuid_write;
671 	struct semaphore	uuid_write_mutex;
672 
673 	/*
674 	 * A btree node on disk could have too many bsets for an iterator to fit
675 	 * on the stack - have to dynamically allocate them.
676 	 * bch_cache_set_alloc() will make sure the pool can allocate iterators
677 	 * equipped with enough room that can host
678 	 *     (sb.bucket_size / sb.block_size)
679 	 * btree_iter_sets, which is more than static MAX_BSETS.
680 	 */
681 	mempool_t		fill_iter;
682 
683 	struct bset_sort_state	sort;
684 
685 	/* List of buckets we're currently writing data to */
686 	struct list_head	data_buckets;
687 	spinlock_t		data_bucket_lock;
688 
689 	struct journal		journal;
690 
691 #define CONGESTED_MAX		1024
692 	unsigned int		congested_last_us;
693 	atomic_t		congested;
694 
695 	/* The rest of this all shows up in sysfs */
696 	unsigned int		congested_read_threshold_us;
697 	unsigned int		congested_write_threshold_us;
698 
699 	struct time_stats	btree_gc_time;
700 	struct time_stats	btree_split_time;
701 	struct time_stats	btree_read_time;
702 
703 	atomic_long_t		cache_read_races;
704 	atomic_long_t		writeback_keys_done;
705 	atomic_long_t		writeback_keys_failed;
706 
707 	atomic_long_t		reclaim;
708 	atomic_long_t		flush_write;
709 	atomic_long_t		retry_flush_write;
710 
711 	enum			{
712 		ON_ERROR_UNREGISTER,
713 		ON_ERROR_PANIC,
714 	}			on_error;
715 #define DEFAULT_IO_ERROR_LIMIT 8
716 	unsigned int		error_limit;
717 	unsigned int		error_decay;
718 
719 	unsigned short		journal_delay_ms;
720 	bool			expensive_debug_checks;
721 	unsigned int		verify:1;
722 	unsigned int		key_merging_disabled:1;
723 	unsigned int		gc_always_rewrite:1;
724 	unsigned int		shrinker_disabled:1;
725 	unsigned int		copy_gc_enabled:1;
726 
727 #define BUCKET_HASH_BITS	12
728 	struct hlist_head	bucket_hash[1 << BUCKET_HASH_BITS];
729 
730 	DECLARE_HEAP(struct btree *, flush_btree);
731 };
732 
733 struct bbio {
734 	unsigned int		submit_time_us;
735 	union {
736 		struct bkey	key;
737 		uint64_t	_pad[3];
738 		/*
739 		 * We only need pad = 3 here because we only ever carry around a
740 		 * single pointer - i.e. the pointer we're doing io to/from.
741 		 */
742 	};
743 	struct bio		bio;
744 };
745 
746 #define BTREE_PRIO		USHRT_MAX
747 #define INITIAL_PRIO		32768U
748 
749 #define btree_bytes(c)		((c)->btree_pages * PAGE_SIZE)
750 #define btree_blocks(b)							\
751 	((unsigned int) (KEY_SIZE(&b->key) >> (b)->c->block_bits))
752 
753 #define btree_default_blocks(c)						\
754 	((unsigned int) ((PAGE_SECTORS * (c)->btree_pages) >> (c)->block_bits))
755 
756 #define bucket_pages(c)		((c)->sb.bucket_size / PAGE_SECTORS)
757 #define bucket_bytes(c)		((c)->sb.bucket_size << 9)
758 #define block_bytes(c)		((c)->sb.block_size << 9)
759 
760 #define prios_per_bucket(c)				\
761 	((bucket_bytes(c) - sizeof(struct prio_set)) /	\
762 	 sizeof(struct bucket_disk))
763 #define prio_buckets(c)					\
764 	DIV_ROUND_UP((size_t) (c)->sb.nbuckets, prios_per_bucket(c))
765 
766 static inline size_t sector_to_bucket(struct cache_set *c, sector_t s)
767 {
768 	return s >> c->bucket_bits;
769 }
770 
771 static inline sector_t bucket_to_sector(struct cache_set *c, size_t b)
772 {
773 	return ((sector_t) b) << c->bucket_bits;
774 }
775 
776 static inline sector_t bucket_remainder(struct cache_set *c, sector_t s)
777 {
778 	return s & (c->sb.bucket_size - 1);
779 }
780 
781 static inline struct cache *PTR_CACHE(struct cache_set *c,
782 				      const struct bkey *k,
783 				      unsigned int ptr)
784 {
785 	return c->cache[PTR_DEV(k, ptr)];
786 }
787 
788 static inline size_t PTR_BUCKET_NR(struct cache_set *c,
789 				   const struct bkey *k,
790 				   unsigned int ptr)
791 {
792 	return sector_to_bucket(c, PTR_OFFSET(k, ptr));
793 }
794 
795 static inline struct bucket *PTR_BUCKET(struct cache_set *c,
796 					const struct bkey *k,
797 					unsigned int ptr)
798 {
799 	return PTR_CACHE(c, k, ptr)->buckets + PTR_BUCKET_NR(c, k, ptr);
800 }
801 
802 static inline uint8_t gen_after(uint8_t a, uint8_t b)
803 {
804 	uint8_t r = a - b;
805 
806 	return r > 128U ? 0 : r;
807 }
808 
809 static inline uint8_t ptr_stale(struct cache_set *c, const struct bkey *k,
810 				unsigned int i)
811 {
812 	return gen_after(PTR_BUCKET(c, k, i)->gen, PTR_GEN(k, i));
813 }
814 
815 static inline bool ptr_available(struct cache_set *c, const struct bkey *k,
816 				 unsigned int i)
817 {
818 	return (PTR_DEV(k, i) < MAX_CACHES_PER_SET) && PTR_CACHE(c, k, i);
819 }
820 
821 /* Btree key macros */
822 
823 /*
824  * This is used for various on disk data structures - cache_sb, prio_set, bset,
825  * jset: The checksum is _always_ the first 8 bytes of these structs
826  */
827 #define csum_set(i)							\
828 	bch_crc64(((void *) (i)) + sizeof(uint64_t),			\
829 		  ((void *) bset_bkey_last(i)) -			\
830 		  (((void *) (i)) + sizeof(uint64_t)))
831 
832 /* Error handling macros */
833 
834 #define btree_bug(b, ...)						\
835 do {									\
836 	if (bch_cache_set_error((b)->c, __VA_ARGS__))			\
837 		dump_stack();						\
838 } while (0)
839 
840 #define cache_bug(c, ...)						\
841 do {									\
842 	if (bch_cache_set_error(c, __VA_ARGS__))			\
843 		dump_stack();						\
844 } while (0)
845 
846 #define btree_bug_on(cond, b, ...)					\
847 do {									\
848 	if (cond)							\
849 		btree_bug(b, __VA_ARGS__);				\
850 } while (0)
851 
852 #define cache_bug_on(cond, c, ...)					\
853 do {									\
854 	if (cond)							\
855 		cache_bug(c, __VA_ARGS__);				\
856 } while (0)
857 
858 #define cache_set_err_on(cond, c, ...)					\
859 do {									\
860 	if (cond)							\
861 		bch_cache_set_error(c, __VA_ARGS__);			\
862 } while (0)
863 
864 /* Looping macros */
865 
866 #define for_each_cache(ca, cs, iter)					\
867 	for (iter = 0; ca = cs->cache[iter], iter < (cs)->sb.nr_in_set; iter++)
868 
869 #define for_each_bucket(b, ca)						\
870 	for (b = (ca)->buckets + (ca)->sb.first_bucket;			\
871 	     b < (ca)->buckets + (ca)->sb.nbuckets; b++)
872 
873 static inline void cached_dev_put(struct cached_dev *dc)
874 {
875 	if (refcount_dec_and_test(&dc->count))
876 		schedule_work(&dc->detach);
877 }
878 
879 static inline bool cached_dev_get(struct cached_dev *dc)
880 {
881 	if (!refcount_inc_not_zero(&dc->count))
882 		return false;
883 
884 	/* Paired with the mb in cached_dev_attach */
885 	smp_mb__after_atomic();
886 	return true;
887 }
888 
889 /*
890  * bucket_gc_gen() returns the difference between the bucket's current gen and
891  * the oldest gen of any pointer into that bucket in the btree (last_gc).
892  */
893 
894 static inline uint8_t bucket_gc_gen(struct bucket *b)
895 {
896 	return b->gen - b->last_gc;
897 }
898 
899 #define BUCKET_GC_GEN_MAX	96U
900 
901 #define kobj_attribute_write(n, fn)					\
902 	static struct kobj_attribute ksysfs_##n = __ATTR(n, 0200, NULL, fn)
903 
904 #define kobj_attribute_rw(n, show, store)				\
905 	static struct kobj_attribute ksysfs_##n =			\
906 		__ATTR(n, 0600, show, store)
907 
908 static inline void wake_up_allocators(struct cache_set *c)
909 {
910 	struct cache *ca;
911 	unsigned int i;
912 
913 	for_each_cache(ca, c, i)
914 		wake_up_process(ca->alloc_thread);
915 }
916 
917 static inline void closure_bio_submit(struct cache_set *c,
918 				      struct bio *bio,
919 				      struct closure *cl)
920 {
921 	closure_get(cl);
922 	if (unlikely(test_bit(CACHE_SET_IO_DISABLE, &c->flags))) {
923 		bio->bi_status = BLK_STS_IOERR;
924 		bio_endio(bio);
925 		return;
926 	}
927 	generic_make_request(bio);
928 }
929 
930 /*
931  * Prevent the kthread exits directly, and make sure when kthread_stop()
932  * is called to stop a kthread, it is still alive. If a kthread might be
933  * stopped by CACHE_SET_IO_DISABLE bit set, wait_for_kthread_stop() is
934  * necessary before the kthread returns.
935  */
936 static inline void wait_for_kthread_stop(void)
937 {
938 	while (!kthread_should_stop()) {
939 		set_current_state(TASK_INTERRUPTIBLE);
940 		schedule();
941 	}
942 }
943 
944 /* Forward declarations */
945 
946 void bch_count_backing_io_errors(struct cached_dev *dc, struct bio *bio);
947 void bch_count_io_errors(struct cache *ca, blk_status_t error,
948 			 int is_read, const char *m);
949 void bch_bbio_count_io_errors(struct cache_set *c, struct bio *bio,
950 			      blk_status_t error, const char *m);
951 void bch_bbio_endio(struct cache_set *c, struct bio *bio,
952 		    blk_status_t error, const char *m);
953 void bch_bbio_free(struct bio *bio, struct cache_set *c);
954 struct bio *bch_bbio_alloc(struct cache_set *c);
955 
956 void __bch_submit_bbio(struct bio *bio, struct cache_set *c);
957 void bch_submit_bbio(struct bio *bio, struct cache_set *c,
958 		     struct bkey *k, unsigned int ptr);
959 
960 uint8_t bch_inc_gen(struct cache *ca, struct bucket *b);
961 void bch_rescale_priorities(struct cache_set *c, int sectors);
962 
963 bool bch_can_invalidate_bucket(struct cache *ca, struct bucket *b);
964 void __bch_invalidate_one_bucket(struct cache *ca, struct bucket *b);
965 
966 void __bch_bucket_free(struct cache *ca, struct bucket *b);
967 void bch_bucket_free(struct cache_set *c, struct bkey *k);
968 
969 long bch_bucket_alloc(struct cache *ca, unsigned int reserve, bool wait);
970 int __bch_bucket_alloc_set(struct cache_set *c, unsigned int reserve,
971 			   struct bkey *k, int n, bool wait);
972 int bch_bucket_alloc_set(struct cache_set *c, unsigned int reserve,
973 			 struct bkey *k, int n, bool wait);
974 bool bch_alloc_sectors(struct cache_set *c, struct bkey *k,
975 		       unsigned int sectors, unsigned int write_point,
976 		       unsigned int write_prio, bool wait);
977 bool bch_cached_dev_error(struct cached_dev *dc);
978 
979 __printf(2, 3)
980 bool bch_cache_set_error(struct cache_set *c, const char *fmt, ...);
981 
982 void bch_prio_write(struct cache *ca);
983 void bch_write_bdev_super(struct cached_dev *dc, struct closure *parent);
984 
985 extern struct workqueue_struct *bcache_wq;
986 extern struct workqueue_struct *bch_journal_wq;
987 extern struct mutex bch_register_lock;
988 extern struct list_head bch_cache_sets;
989 
990 extern struct kobj_type bch_cached_dev_ktype;
991 extern struct kobj_type bch_flash_dev_ktype;
992 extern struct kobj_type bch_cache_set_ktype;
993 extern struct kobj_type bch_cache_set_internal_ktype;
994 extern struct kobj_type bch_cache_ktype;
995 
996 void bch_cached_dev_release(struct kobject *kobj);
997 void bch_flash_dev_release(struct kobject *kobj);
998 void bch_cache_set_release(struct kobject *kobj);
999 void bch_cache_release(struct kobject *kobj);
1000 
1001 int bch_uuid_write(struct cache_set *c);
1002 void bcache_write_super(struct cache_set *c);
1003 
1004 int bch_flash_dev_create(struct cache_set *c, uint64_t size);
1005 
1006 int bch_cached_dev_attach(struct cached_dev *dc, struct cache_set *c,
1007 			  uint8_t *set_uuid);
1008 void bch_cached_dev_detach(struct cached_dev *dc);
1009 void bch_cached_dev_run(struct cached_dev *dc);
1010 void bcache_device_stop(struct bcache_device *d);
1011 
1012 void bch_cache_set_unregister(struct cache_set *c);
1013 void bch_cache_set_stop(struct cache_set *c);
1014 
1015 struct cache_set *bch_cache_set_alloc(struct cache_sb *sb);
1016 void bch_btree_cache_free(struct cache_set *c);
1017 int bch_btree_cache_alloc(struct cache_set *c);
1018 void bch_moving_init_cache_set(struct cache_set *c);
1019 int bch_open_buckets_alloc(struct cache_set *c);
1020 void bch_open_buckets_free(struct cache_set *c);
1021 
1022 int bch_cache_allocator_start(struct cache *ca);
1023 
1024 void bch_debug_exit(void);
1025 void bch_debug_init(void);
1026 void bch_request_exit(void);
1027 int bch_request_init(void);
1028 
1029 #endif /* _BCACHE_H */
1030