xref: /linux/drivers/md/bcache/bset.h (revision f86fd32d)
1 /* SPDX-License-Identifier: GPL-2.0 */
2 #ifndef _BCACHE_BSET_H
3 #define _BCACHE_BSET_H
4 
5 #include <linux/bcache.h>
6 #include <linux/kernel.h>
7 #include <linux/types.h>
8 
9 #include "util.h" /* for time_stats */
10 
11 /*
12  * BKEYS:
13  *
14  * A bkey contains a key, a size field, a variable number of pointers, and some
15  * ancillary flag bits.
16  *
17  * We use two different functions for validating bkeys, bch_ptr_invalid and
18  * bch_ptr_bad().
19  *
20  * bch_ptr_invalid() primarily filters out keys and pointers that would be
21  * invalid due to some sort of bug, whereas bch_ptr_bad() filters out keys and
22  * pointer that occur in normal practice but don't point to real data.
23  *
24  * The one exception to the rule that ptr_invalid() filters out invalid keys is
25  * that it also filters out keys of size 0 - these are keys that have been
26  * completely overwritten. It'd be safe to delete these in memory while leaving
27  * them on disk, just unnecessary work - so we filter them out when resorting
28  * instead.
29  *
30  * We can't filter out stale keys when we're resorting, because garbage
31  * collection needs to find them to ensure bucket gens don't wrap around -
32  * unless we're rewriting the btree node those stale keys still exist on disk.
33  *
34  * We also implement functions here for removing some number of sectors from the
35  * front or the back of a bkey - this is mainly used for fixing overlapping
36  * extents, by removing the overlapping sectors from the older key.
37  *
38  * BSETS:
39  *
40  * A bset is an array of bkeys laid out contiguously in memory in sorted order,
41  * along with a header. A btree node is made up of a number of these, written at
42  * different times.
43  *
44  * There could be many of them on disk, but we never allow there to be more than
45  * 4 in memory - we lazily resort as needed.
46  *
47  * We implement code here for creating and maintaining auxiliary search trees
48  * (described below) for searching an individial bset, and on top of that we
49  * implement a btree iterator.
50  *
51  * BTREE ITERATOR:
52  *
53  * Most of the code in bcache doesn't care about an individual bset - it needs
54  * to search entire btree nodes and iterate over them in sorted order.
55  *
56  * The btree iterator code serves both functions; it iterates through the keys
57  * in a btree node in sorted order, starting from either keys after a specific
58  * point (if you pass it a search key) or the start of the btree node.
59  *
60  * AUXILIARY SEARCH TREES:
61  *
62  * Since keys are variable length, we can't use a binary search on a bset - we
63  * wouldn't be able to find the start of the next key. But binary searches are
64  * slow anyways, due to terrible cache behaviour; bcache originally used binary
65  * searches and that code topped out at under 50k lookups/second.
66  *
67  * So we need to construct some sort of lookup table. Since we only insert keys
68  * into the last (unwritten) set, most of the keys within a given btree node are
69  * usually in sets that are mostly constant. We use two different types of
70  * lookup tables to take advantage of this.
71  *
72  * Both lookup tables share in common that they don't index every key in the
73  * set; they index one key every BSET_CACHELINE bytes, and then a linear search
74  * is used for the rest.
75  *
76  * For sets that have been written to disk and are no longer being inserted
77  * into, we construct a binary search tree in an array - traversing a binary
78  * search tree in an array gives excellent locality of reference and is very
79  * fast, since both children of any node are adjacent to each other in memory
80  * (and their grandchildren, and great grandchildren...) - this means
81  * prefetching can be used to great effect.
82  *
83  * It's quite useful performance wise to keep these nodes small - not just
84  * because they're more likely to be in L2, but also because we can prefetch
85  * more nodes on a single cacheline and thus prefetch more iterations in advance
86  * when traversing this tree.
87  *
88  * Nodes in the auxiliary search tree must contain both a key to compare against
89  * (we don't want to fetch the key from the set, that would defeat the purpose),
90  * and a pointer to the key. We use a few tricks to compress both of these.
91  *
92  * To compress the pointer, we take advantage of the fact that one node in the
93  * search tree corresponds to precisely BSET_CACHELINE bytes in the set. We have
94  * a function (to_inorder()) that takes the index of a node in a binary tree and
95  * returns what its index would be in an inorder traversal, so we only have to
96  * store the low bits of the offset.
97  *
98  * The key is 84 bits (KEY_DEV + key->key, the offset on the device). To
99  * compress that,  we take advantage of the fact that when we're traversing the
100  * search tree at every iteration we know that both our search key and the key
101  * we're looking for lie within some range - bounded by our previous
102  * comparisons. (We special case the start of a search so that this is true even
103  * at the root of the tree).
104  *
105  * So we know the key we're looking for is between a and b, and a and b don't
106  * differ higher than bit 50, we don't need to check anything higher than bit
107  * 50.
108  *
109  * We don't usually need the rest of the bits, either; we only need enough bits
110  * to partition the key range we're currently checking.  Consider key n - the
111  * key our auxiliary search tree node corresponds to, and key p, the key
112  * immediately preceding n.  The lowest bit we need to store in the auxiliary
113  * search tree is the highest bit that differs between n and p.
114  *
115  * Note that this could be bit 0 - we might sometimes need all 80 bits to do the
116  * comparison. But we'd really like our nodes in the auxiliary search tree to be
117  * of fixed size.
118  *
119  * The solution is to make them fixed size, and when we're constructing a node
120  * check if p and n differed in the bits we needed them to. If they don't we
121  * flag that node, and when doing lookups we fallback to comparing against the
122  * real key. As long as this doesn't happen to often (and it seems to reliably
123  * happen a bit less than 1% of the time), we win - even on failures, that key
124  * is then more likely to be in cache than if we were doing binary searches all
125  * the way, since we're touching so much less memory.
126  *
127  * The keys in the auxiliary search tree are stored in (software) floating
128  * point, with an exponent and a mantissa. The exponent needs to be big enough
129  * to address all the bits in the original key, but the number of bits in the
130  * mantissa is somewhat arbitrary; more bits just gets us fewer failures.
131  *
132  * We need 7 bits for the exponent and 3 bits for the key's offset (since keys
133  * are 8 byte aligned); using 22 bits for the mantissa means a node is 4 bytes.
134  * We need one node per 128 bytes in the btree node, which means the auxiliary
135  * search trees take up 3% as much memory as the btree itself.
136  *
137  * Constructing these auxiliary search trees is moderately expensive, and we
138  * don't want to be constantly rebuilding the search tree for the last set
139  * whenever we insert another key into it. For the unwritten set, we use a much
140  * simpler lookup table - it's just a flat array, so index i in the lookup table
141  * corresponds to the i range of BSET_CACHELINE bytes in the set. Indexing
142  * within each byte range works the same as with the auxiliary search trees.
143  *
144  * These are much easier to keep up to date when we insert a key - we do it
145  * somewhat lazily; when we shift a key up we usually just increment the pointer
146  * to it, only when it would overflow do we go to the trouble of finding the
147  * first key in that range of bytes again.
148  */
149 
150 struct btree_keys;
151 struct btree_iter;
152 struct btree_iter_set;
153 struct bkey_float;
154 
155 #define MAX_BSETS		4U
156 
157 struct bset_tree {
158 	/*
159 	 * We construct a binary tree in an array as if the array
160 	 * started at 1, so that things line up on the same cachelines
161 	 * better: see comments in bset.c at cacheline_to_bkey() for
162 	 * details
163 	 */
164 
165 	/* size of the binary tree and prev array */
166 	unsigned int		size;
167 
168 	/* function of size - precalculated for to_inorder() */
169 	unsigned int		extra;
170 
171 	/* copy of the last key in the set */
172 	struct bkey		end;
173 	struct bkey_float	*tree;
174 
175 	/*
176 	 * The nodes in the bset tree point to specific keys - this
177 	 * array holds the sizes of the previous key.
178 	 *
179 	 * Conceptually it's a member of struct bkey_float, but we want
180 	 * to keep bkey_float to 4 bytes and prev isn't used in the fast
181 	 * path.
182 	 */
183 	uint8_t			*prev;
184 
185 	/* The actual btree node, with pointers to each sorted set */
186 	struct bset		*data;
187 };
188 
189 struct btree_keys_ops {
190 	bool		(*sort_cmp)(struct btree_iter_set l,
191 				    struct btree_iter_set r);
192 	struct bkey	*(*sort_fixup)(struct btree_iter *iter,
193 				       struct bkey *tmp);
194 	bool		(*insert_fixup)(struct btree_keys *b,
195 					struct bkey *insert,
196 					struct btree_iter *iter,
197 					struct bkey *replace_key);
198 	bool		(*key_invalid)(struct btree_keys *bk,
199 				       const struct bkey *k);
200 	bool		(*key_bad)(struct btree_keys *bk,
201 				   const struct bkey *k);
202 	bool		(*key_merge)(struct btree_keys *bk,
203 				     struct bkey *l, struct bkey *r);
204 	void		(*key_to_text)(char *buf,
205 				       size_t size,
206 				       const struct bkey *k);
207 	void		(*key_dump)(struct btree_keys *keys,
208 				    const struct bkey *k);
209 
210 	/*
211 	 * Only used for deciding whether to use START_KEY(k) or just the key
212 	 * itself in a couple places
213 	 */
214 	bool		is_extents;
215 };
216 
217 struct btree_keys {
218 	const struct btree_keys_ops	*ops;
219 	uint8_t			page_order;
220 	uint8_t			nsets;
221 	unsigned int		last_set_unwritten:1;
222 	bool			*expensive_debug_checks;
223 
224 	/*
225 	 * Sets of sorted keys - the real btree node - plus a binary search tree
226 	 *
227 	 * set[0] is special; set[0]->tree, set[0]->prev and set[0]->data point
228 	 * to the memory we have allocated for this btree node. Additionally,
229 	 * set[0]->data points to the entire btree node as it exists on disk.
230 	 */
231 	struct bset_tree	set[MAX_BSETS];
232 };
233 
234 static inline struct bset_tree *bset_tree_last(struct btree_keys *b)
235 {
236 	return b->set + b->nsets;
237 }
238 
239 static inline bool bset_written(struct btree_keys *b, struct bset_tree *t)
240 {
241 	return t <= b->set + b->nsets - b->last_set_unwritten;
242 }
243 
244 static inline bool bkey_written(struct btree_keys *b, struct bkey *k)
245 {
246 	return !b->last_set_unwritten || k < b->set[b->nsets].data->start;
247 }
248 
249 static inline unsigned int bset_byte_offset(struct btree_keys *b,
250 					    struct bset *i)
251 {
252 	return ((size_t) i) - ((size_t) b->set->data);
253 }
254 
255 static inline unsigned int bset_sector_offset(struct btree_keys *b,
256 					      struct bset *i)
257 {
258 	return bset_byte_offset(b, i) >> 9;
259 }
260 
261 #define __set_bytes(i, k)	(sizeof(*(i)) + (k) * sizeof(uint64_t))
262 #define set_bytes(i)		__set_bytes(i, i->keys)
263 
264 #define __set_blocks(i, k, block_bytes)				\
265 	DIV_ROUND_UP(__set_bytes(i, k), block_bytes)
266 #define set_blocks(i, block_bytes)				\
267 	__set_blocks(i, (i)->keys, block_bytes)
268 
269 static inline size_t bch_btree_keys_u64s_remaining(struct btree_keys *b)
270 {
271 	struct bset_tree *t = bset_tree_last(b);
272 
273 	BUG_ON((PAGE_SIZE << b->page_order) <
274 	       (bset_byte_offset(b, t->data) + set_bytes(t->data)));
275 
276 	if (!b->last_set_unwritten)
277 		return 0;
278 
279 	return ((PAGE_SIZE << b->page_order) -
280 		(bset_byte_offset(b, t->data) + set_bytes(t->data))) /
281 		sizeof(u64);
282 }
283 
284 static inline struct bset *bset_next_set(struct btree_keys *b,
285 					 unsigned int block_bytes)
286 {
287 	struct bset *i = bset_tree_last(b)->data;
288 
289 	return ((void *) i) + roundup(set_bytes(i), block_bytes);
290 }
291 
292 void bch_btree_keys_free(struct btree_keys *b);
293 int bch_btree_keys_alloc(struct btree_keys *b, unsigned int page_order,
294 			 gfp_t gfp);
295 void bch_btree_keys_init(struct btree_keys *b, const struct btree_keys_ops *ops,
296 			 bool *expensive_debug_checks);
297 
298 void bch_bset_init_next(struct btree_keys *b, struct bset *i, uint64_t magic);
299 void bch_bset_build_written_tree(struct btree_keys *b);
300 void bch_bset_fix_invalidated_key(struct btree_keys *b, struct bkey *k);
301 bool bch_bkey_try_merge(struct btree_keys *b, struct bkey *l, struct bkey *r);
302 void bch_bset_insert(struct btree_keys *b, struct bkey *where,
303 		     struct bkey *insert);
304 unsigned int bch_btree_insert_key(struct btree_keys *b, struct bkey *k,
305 			      struct bkey *replace_key);
306 
307 enum {
308 	BTREE_INSERT_STATUS_NO_INSERT = 0,
309 	BTREE_INSERT_STATUS_INSERT,
310 	BTREE_INSERT_STATUS_BACK_MERGE,
311 	BTREE_INSERT_STATUS_OVERWROTE,
312 	BTREE_INSERT_STATUS_FRONT_MERGE,
313 };
314 
315 /* Btree key iteration */
316 
317 struct btree_iter {
318 	size_t size, used;
319 #ifdef CONFIG_BCACHE_DEBUG
320 	struct btree_keys *b;
321 #endif
322 	struct btree_iter_set {
323 		struct bkey *k, *end;
324 	} data[MAX_BSETS];
325 };
326 
327 typedef bool (*ptr_filter_fn)(struct btree_keys *b, const struct bkey *k);
328 
329 struct bkey *bch_btree_iter_next(struct btree_iter *iter);
330 struct bkey *bch_btree_iter_next_filter(struct btree_iter *iter,
331 					struct btree_keys *b,
332 					ptr_filter_fn fn);
333 
334 void bch_btree_iter_push(struct btree_iter *iter, struct bkey *k,
335 			 struct bkey *end);
336 struct bkey *bch_btree_iter_init(struct btree_keys *b,
337 				 struct btree_iter *iter,
338 				 struct bkey *search);
339 
340 struct bkey *__bch_bset_search(struct btree_keys *b, struct bset_tree *t,
341 			       const struct bkey *search);
342 
343 /*
344  * Returns the first key that is strictly greater than search
345  */
346 static inline struct bkey *bch_bset_search(struct btree_keys *b,
347 					   struct bset_tree *t,
348 					   const struct bkey *search)
349 {
350 	return search ? __bch_bset_search(b, t, search) : t->data->start;
351 }
352 
353 #define for_each_key_filter(b, k, iter, filter)				\
354 	for (bch_btree_iter_init((b), (iter), NULL);			\
355 	     ((k) = bch_btree_iter_next_filter((iter), (b), filter));)
356 
357 #define for_each_key(b, k, iter)					\
358 	for (bch_btree_iter_init((b), (iter), NULL);			\
359 	     ((k) = bch_btree_iter_next(iter));)
360 
361 /* Sorting */
362 
363 struct bset_sort_state {
364 	mempool_t		pool;
365 
366 	unsigned int		page_order;
367 	unsigned int		crit_factor;
368 
369 	struct time_stats	time;
370 };
371 
372 void bch_bset_sort_state_free(struct bset_sort_state *state);
373 int bch_bset_sort_state_init(struct bset_sort_state *state,
374 			     unsigned int page_order);
375 void bch_btree_sort_lazy(struct btree_keys *b, struct bset_sort_state *state);
376 void bch_btree_sort_into(struct btree_keys *b, struct btree_keys *new,
377 			 struct bset_sort_state *state);
378 void bch_btree_sort_and_fix_extents(struct btree_keys *b,
379 				    struct btree_iter *iter,
380 				    struct bset_sort_state *state);
381 void bch_btree_sort_partial(struct btree_keys *b, unsigned int start,
382 			    struct bset_sort_state *state);
383 
384 static inline void bch_btree_sort(struct btree_keys *b,
385 				  struct bset_sort_state *state)
386 {
387 	bch_btree_sort_partial(b, 0, state);
388 }
389 
390 struct bset_stats {
391 	size_t sets_written, sets_unwritten;
392 	size_t bytes_written, bytes_unwritten;
393 	size_t floats, failed;
394 };
395 
396 void bch_btree_keys_stats(struct btree_keys *b, struct bset_stats *state);
397 
398 /* Bkey utility code */
399 
400 #define bset_bkey_last(i)	bkey_idx((struct bkey *) (i)->d, \
401 					 (unsigned int)(i)->keys)
402 
403 static inline struct bkey *bset_bkey_idx(struct bset *i, unsigned int idx)
404 {
405 	return bkey_idx(i->start, idx);
406 }
407 
408 static inline void bkey_init(struct bkey *k)
409 {
410 	*k = ZERO_KEY;
411 }
412 
413 static __always_inline int64_t bkey_cmp(const struct bkey *l,
414 					const struct bkey *r)
415 {
416 	return unlikely(KEY_INODE(l) != KEY_INODE(r))
417 		? (int64_t) KEY_INODE(l) - (int64_t) KEY_INODE(r)
418 		: (int64_t) KEY_OFFSET(l) - (int64_t) KEY_OFFSET(r);
419 }
420 
421 void bch_bkey_copy_single_ptr(struct bkey *dest, const struct bkey *src,
422 			      unsigned int i);
423 bool __bch_cut_front(const struct bkey *where, struct bkey *k);
424 bool __bch_cut_back(const struct bkey *where, struct bkey *k);
425 
426 static inline bool bch_cut_front(const struct bkey *where, struct bkey *k)
427 {
428 	BUG_ON(bkey_cmp(where, k) > 0);
429 	return __bch_cut_front(where, k);
430 }
431 
432 static inline bool bch_cut_back(const struct bkey *where, struct bkey *k)
433 {
434 	BUG_ON(bkey_cmp(where, &START_KEY(k)) < 0);
435 	return __bch_cut_back(where, k);
436 }
437 
438 /*
439  * Pointer '*preceding_key_p' points to a memory object to store preceding
440  * key of k. If the preceding key does not exist, set '*preceding_key_p' to
441  * NULL. So the caller of preceding_key() needs to take care of memory
442  * which '*preceding_key_p' pointed to before calling preceding_key().
443  * Currently the only caller of preceding_key() is bch_btree_insert_key(),
444  * and it points to an on-stack variable, so the memory release is handled
445  * by stackframe itself.
446  */
447 static inline void preceding_key(struct bkey *k, struct bkey **preceding_key_p)
448 {
449 	if (KEY_INODE(k) || KEY_OFFSET(k)) {
450 		(**preceding_key_p) = KEY(KEY_INODE(k), KEY_OFFSET(k), 0);
451 		if (!(*preceding_key_p)->low)
452 			(*preceding_key_p)->high--;
453 		(*preceding_key_p)->low--;
454 	} else {
455 		(*preceding_key_p) = NULL;
456 	}
457 }
458 
459 static inline bool bch_ptr_invalid(struct btree_keys *b, const struct bkey *k)
460 {
461 	return b->ops->key_invalid(b, k);
462 }
463 
464 static inline bool bch_ptr_bad(struct btree_keys *b, const struct bkey *k)
465 {
466 	return b->ops->key_bad(b, k);
467 }
468 
469 static inline void bch_bkey_to_text(struct btree_keys *b, char *buf,
470 				    size_t size, const struct bkey *k)
471 {
472 	return b->ops->key_to_text(buf, size, k);
473 }
474 
475 static inline bool bch_bkey_equal_header(const struct bkey *l,
476 					 const struct bkey *r)
477 {
478 	return (KEY_DIRTY(l) == KEY_DIRTY(r) &&
479 		KEY_PTRS(l) == KEY_PTRS(r) &&
480 		KEY_CSUM(l) == KEY_CSUM(r));
481 }
482 
483 /* Keylists */
484 
485 struct keylist {
486 	union {
487 		struct bkey		*keys;
488 		uint64_t		*keys_p;
489 	};
490 	union {
491 		struct bkey		*top;
492 		uint64_t		*top_p;
493 	};
494 
495 	/* Enough room for btree_split's keys without realloc */
496 #define KEYLIST_INLINE		16
497 	uint64_t		inline_keys[KEYLIST_INLINE];
498 };
499 
500 static inline void bch_keylist_init(struct keylist *l)
501 {
502 	l->top_p = l->keys_p = l->inline_keys;
503 }
504 
505 static inline void bch_keylist_init_single(struct keylist *l, struct bkey *k)
506 {
507 	l->keys = k;
508 	l->top = bkey_next(k);
509 }
510 
511 static inline void bch_keylist_push(struct keylist *l)
512 {
513 	l->top = bkey_next(l->top);
514 }
515 
516 static inline void bch_keylist_add(struct keylist *l, struct bkey *k)
517 {
518 	bkey_copy(l->top, k);
519 	bch_keylist_push(l);
520 }
521 
522 static inline bool bch_keylist_empty(struct keylist *l)
523 {
524 	return l->top == l->keys;
525 }
526 
527 static inline void bch_keylist_reset(struct keylist *l)
528 {
529 	l->top = l->keys;
530 }
531 
532 static inline void bch_keylist_free(struct keylist *l)
533 {
534 	if (l->keys_p != l->inline_keys)
535 		kfree(l->keys_p);
536 }
537 
538 static inline size_t bch_keylist_nkeys(struct keylist *l)
539 {
540 	return l->top_p - l->keys_p;
541 }
542 
543 static inline size_t bch_keylist_bytes(struct keylist *l)
544 {
545 	return bch_keylist_nkeys(l) * sizeof(uint64_t);
546 }
547 
548 struct bkey *bch_keylist_pop(struct keylist *l);
549 void bch_keylist_pop_front(struct keylist *l);
550 int __bch_keylist_realloc(struct keylist *l, unsigned int u64s);
551 
552 /* Debug stuff */
553 
554 #ifdef CONFIG_BCACHE_DEBUG
555 
556 int __bch_count_data(struct btree_keys *b);
557 void __printf(2, 3) __bch_check_keys(struct btree_keys *b,
558 				     const char *fmt,
559 				     ...);
560 void bch_dump_bset(struct btree_keys *b, struct bset *i, unsigned int set);
561 void bch_dump_bucket(struct btree_keys *b);
562 
563 #else
564 
565 static inline int __bch_count_data(struct btree_keys *b) { return -1; }
566 static inline void __printf(2, 3)
567 	__bch_check_keys(struct btree_keys *b, const char *fmt, ...) {}
568 static inline void bch_dump_bucket(struct btree_keys *b) {}
569 void bch_dump_bset(struct btree_keys *b, struct bset *i, unsigned int set);
570 
571 #endif
572 
573 static inline bool btree_keys_expensive_checks(struct btree_keys *b)
574 {
575 #ifdef CONFIG_BCACHE_DEBUG
576 	return *b->expensive_debug_checks;
577 #else
578 	return false;
579 #endif
580 }
581 
582 static inline int bch_count_data(struct btree_keys *b)
583 {
584 	return btree_keys_expensive_checks(b) ? __bch_count_data(b) : -1;
585 }
586 
587 #define bch_check_keys(b, ...)						\
588 do {									\
589 	if (btree_keys_expensive_checks(b))				\
590 		__bch_check_keys(b, __VA_ARGS__);			\
591 } while (0)
592 
593 #endif
594