xref: /linux/drivers/md/bcache/extents.c (revision cea0f76a483d1270ac6f6513964e3e75193dda48)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Copyright (C) 2010 Kent Overstreet <kent.overstreet@gmail.com>
4  *
5  * Uses a block device as cache for other block devices; optimized for SSDs.
6  * All allocation is done in buckets, which should match the erase block size
7  * of the device.
8  *
9  * Buckets containing cached data are kept on a heap sorted by priority;
10  * bucket priority is increased on cache hit, and periodically all the buckets
11  * on the heap have their priority scaled down. This currently is just used as
12  * an LRU but in the future should allow for more intelligent heuristics.
13  *
14  * Buckets have an 8 bit counter; freeing is accomplished by incrementing the
15  * counter. Garbage collection is used to remove stale pointers.
16  *
17  * Indexing is done via a btree; nodes are not necessarily fully sorted, rather
18  * as keys are inserted we only sort the pages that have not yet been written.
19  * When garbage collection is run, we resort the entire node.
20  *
21  * All configuration is done via sysfs; see Documentation/admin-guide/bcache.rst.
22  */
23 
24 #include "bcache.h"
25 #include "btree.h"
26 #include "debug.h"
27 #include "extents.h"
28 #include "writeback.h"
29 
30 static void sort_key_next(struct btree_iter *iter,
31 			  struct btree_iter_set *i)
32 {
33 	i->k = bkey_next(i->k);
34 
35 	if (i->k == i->end)
36 		*i = iter->data[--iter->used];
37 }
38 
39 static bool bch_key_sort_cmp(struct btree_iter_set l,
40 			     struct btree_iter_set r)
41 {
42 	int64_t c = bkey_cmp(l.k, r.k);
43 
44 	return c ? c > 0 : l.k < r.k;
45 }
46 
47 static bool __ptr_invalid(struct cache_set *c, const struct bkey *k)
48 {
49 	unsigned int i;
50 
51 	for (i = 0; i < KEY_PTRS(k); i++)
52 		if (ptr_available(c, k, i)) {
53 			struct cache *ca = PTR_CACHE(c, k, i);
54 			size_t bucket = PTR_BUCKET_NR(c, k, i);
55 			size_t r = bucket_remainder(c, PTR_OFFSET(k, i));
56 
57 			if (KEY_SIZE(k) + r > c->sb.bucket_size ||
58 			    bucket <  ca->sb.first_bucket ||
59 			    bucket >= ca->sb.nbuckets)
60 				return true;
61 		}
62 
63 	return false;
64 }
65 
66 /* Common among btree and extent ptrs */
67 
68 static const char *bch_ptr_status(struct cache_set *c, const struct bkey *k)
69 {
70 	unsigned int i;
71 
72 	for (i = 0; i < KEY_PTRS(k); i++)
73 		if (ptr_available(c, k, i)) {
74 			struct cache *ca = PTR_CACHE(c, k, i);
75 			size_t bucket = PTR_BUCKET_NR(c, k, i);
76 			size_t r = bucket_remainder(c, PTR_OFFSET(k, i));
77 
78 			if (KEY_SIZE(k) + r > c->sb.bucket_size)
79 				return "bad, length too big";
80 			if (bucket <  ca->sb.first_bucket)
81 				return "bad, short offset";
82 			if (bucket >= ca->sb.nbuckets)
83 				return "bad, offset past end of device";
84 			if (ptr_stale(c, k, i))
85 				return "stale";
86 		}
87 
88 	if (!bkey_cmp(k, &ZERO_KEY))
89 		return "bad, null key";
90 	if (!KEY_PTRS(k))
91 		return "bad, no pointers";
92 	if (!KEY_SIZE(k))
93 		return "zeroed key";
94 	return "";
95 }
96 
97 void bch_extent_to_text(char *buf, size_t size, const struct bkey *k)
98 {
99 	unsigned int i = 0;
100 	char *out = buf, *end = buf + size;
101 
102 #define p(...)	(out += scnprintf(out, end - out, __VA_ARGS__))
103 
104 	p("%llu:%llu len %llu -> [", KEY_INODE(k), KEY_START(k), KEY_SIZE(k));
105 
106 	for (i = 0; i < KEY_PTRS(k); i++) {
107 		if (i)
108 			p(", ");
109 
110 		if (PTR_DEV(k, i) == PTR_CHECK_DEV)
111 			p("check dev");
112 		else
113 			p("%llu:%llu gen %llu", PTR_DEV(k, i),
114 			  PTR_OFFSET(k, i), PTR_GEN(k, i));
115 	}
116 
117 	p("]");
118 
119 	if (KEY_DIRTY(k))
120 		p(" dirty");
121 	if (KEY_CSUM(k))
122 		p(" cs%llu %llx", KEY_CSUM(k), k->ptr[1]);
123 #undef p
124 }
125 
126 static void bch_bkey_dump(struct btree_keys *keys, const struct bkey *k)
127 {
128 	struct btree *b = container_of(keys, struct btree, keys);
129 	unsigned int j;
130 	char buf[80];
131 
132 	bch_extent_to_text(buf, sizeof(buf), k);
133 	pr_cont(" %s", buf);
134 
135 	for (j = 0; j < KEY_PTRS(k); j++) {
136 		size_t n = PTR_BUCKET_NR(b->c, k, j);
137 
138 		pr_cont(" bucket %zu", n);
139 		if (n >= b->c->sb.first_bucket && n < b->c->sb.nbuckets)
140 			pr_cont(" prio %i",
141 				PTR_BUCKET(b->c, k, j)->prio);
142 	}
143 
144 	pr_cont(" %s\n", bch_ptr_status(b->c, k));
145 }
146 
147 /* Btree ptrs */
148 
149 bool __bch_btree_ptr_invalid(struct cache_set *c, const struct bkey *k)
150 {
151 	char buf[80];
152 
153 	if (!KEY_PTRS(k) || !KEY_SIZE(k) || KEY_DIRTY(k))
154 		goto bad;
155 
156 	if (__ptr_invalid(c, k))
157 		goto bad;
158 
159 	return false;
160 bad:
161 	bch_extent_to_text(buf, sizeof(buf), k);
162 	cache_bug(c, "spotted btree ptr %s: %s", buf, bch_ptr_status(c, k));
163 	return true;
164 }
165 
166 static bool bch_btree_ptr_invalid(struct btree_keys *bk, const struct bkey *k)
167 {
168 	struct btree *b = container_of(bk, struct btree, keys);
169 
170 	return __bch_btree_ptr_invalid(b->c, k);
171 }
172 
173 static bool btree_ptr_bad_expensive(struct btree *b, const struct bkey *k)
174 {
175 	unsigned int i;
176 	char buf[80];
177 	struct bucket *g;
178 
179 	if (mutex_trylock(&b->c->bucket_lock)) {
180 		for (i = 0; i < KEY_PTRS(k); i++)
181 			if (ptr_available(b->c, k, i)) {
182 				g = PTR_BUCKET(b->c, k, i);
183 
184 				if (KEY_DIRTY(k) ||
185 				    g->prio != BTREE_PRIO ||
186 				    (b->c->gc_mark_valid &&
187 				     GC_MARK(g) != GC_MARK_METADATA))
188 					goto err;
189 			}
190 
191 		mutex_unlock(&b->c->bucket_lock);
192 	}
193 
194 	return false;
195 err:
196 	mutex_unlock(&b->c->bucket_lock);
197 	bch_extent_to_text(buf, sizeof(buf), k);
198 	btree_bug(b,
199 "inconsistent btree pointer %s: bucket %zi pin %i prio %i gen %i last_gc %i mark %llu",
200 		  buf, PTR_BUCKET_NR(b->c, k, i), atomic_read(&g->pin),
201 		  g->prio, g->gen, g->last_gc, GC_MARK(g));
202 	return true;
203 }
204 
205 static bool bch_btree_ptr_bad(struct btree_keys *bk, const struct bkey *k)
206 {
207 	struct btree *b = container_of(bk, struct btree, keys);
208 	unsigned int i;
209 
210 	if (!bkey_cmp(k, &ZERO_KEY) ||
211 	    !KEY_PTRS(k) ||
212 	    bch_ptr_invalid(bk, k))
213 		return true;
214 
215 	for (i = 0; i < KEY_PTRS(k); i++)
216 		if (!ptr_available(b->c, k, i) ||
217 		    ptr_stale(b->c, k, i))
218 			return true;
219 
220 	if (expensive_debug_checks(b->c) &&
221 	    btree_ptr_bad_expensive(b, k))
222 		return true;
223 
224 	return false;
225 }
226 
227 static bool bch_btree_ptr_insert_fixup(struct btree_keys *bk,
228 				       struct bkey *insert,
229 				       struct btree_iter *iter,
230 				       struct bkey *replace_key)
231 {
232 	struct btree *b = container_of(bk, struct btree, keys);
233 
234 	if (!KEY_OFFSET(insert))
235 		btree_current_write(b)->prio_blocked++;
236 
237 	return false;
238 }
239 
240 const struct btree_keys_ops bch_btree_keys_ops = {
241 	.sort_cmp	= bch_key_sort_cmp,
242 	.insert_fixup	= bch_btree_ptr_insert_fixup,
243 	.key_invalid	= bch_btree_ptr_invalid,
244 	.key_bad	= bch_btree_ptr_bad,
245 	.key_to_text	= bch_extent_to_text,
246 	.key_dump	= bch_bkey_dump,
247 };
248 
249 /* Extents */
250 
251 /*
252  * Returns true if l > r - unless l == r, in which case returns true if l is
253  * older than r.
254  *
255  * Necessary for btree_sort_fixup() - if there are multiple keys that compare
256  * equal in different sets, we have to process them newest to oldest.
257  */
258 static bool bch_extent_sort_cmp(struct btree_iter_set l,
259 				struct btree_iter_set r)
260 {
261 	int64_t c = bkey_cmp(&START_KEY(l.k), &START_KEY(r.k));
262 
263 	return c ? c > 0 : l.k < r.k;
264 }
265 
266 static struct bkey *bch_extent_sort_fixup(struct btree_iter *iter,
267 					  struct bkey *tmp)
268 {
269 	while (iter->used > 1) {
270 		struct btree_iter_set *top = iter->data, *i = top + 1;
271 
272 		if (iter->used > 2 &&
273 		    bch_extent_sort_cmp(i[0], i[1]))
274 			i++;
275 
276 		if (bkey_cmp(top->k, &START_KEY(i->k)) <= 0)
277 			break;
278 
279 		if (!KEY_SIZE(i->k)) {
280 			sort_key_next(iter, i);
281 			heap_sift(iter, i - top, bch_extent_sort_cmp);
282 			continue;
283 		}
284 
285 		if (top->k > i->k) {
286 			if (bkey_cmp(top->k, i->k) >= 0)
287 				sort_key_next(iter, i);
288 			else
289 				bch_cut_front(top->k, i->k);
290 
291 			heap_sift(iter, i - top, bch_extent_sort_cmp);
292 		} else {
293 			/* can't happen because of comparison func */
294 			BUG_ON(!bkey_cmp(&START_KEY(top->k), &START_KEY(i->k)));
295 
296 			if (bkey_cmp(i->k, top->k) < 0) {
297 				bkey_copy(tmp, top->k);
298 
299 				bch_cut_back(&START_KEY(i->k), tmp);
300 				bch_cut_front(i->k, top->k);
301 				heap_sift(iter, 0, bch_extent_sort_cmp);
302 
303 				return tmp;
304 			} else {
305 				bch_cut_back(&START_KEY(i->k), top->k);
306 			}
307 		}
308 	}
309 
310 	return NULL;
311 }
312 
313 static void bch_subtract_dirty(struct bkey *k,
314 			   struct cache_set *c,
315 			   uint64_t offset,
316 			   int sectors)
317 {
318 	if (KEY_DIRTY(k))
319 		bcache_dev_sectors_dirty_add(c, KEY_INODE(k),
320 					     offset, -sectors);
321 }
322 
323 static bool bch_extent_insert_fixup(struct btree_keys *b,
324 				    struct bkey *insert,
325 				    struct btree_iter *iter,
326 				    struct bkey *replace_key)
327 {
328 	struct cache_set *c = container_of(b, struct btree, keys)->c;
329 
330 	uint64_t old_offset;
331 	unsigned int old_size, sectors_found = 0;
332 
333 	BUG_ON(!KEY_OFFSET(insert));
334 	BUG_ON(!KEY_SIZE(insert));
335 
336 	while (1) {
337 		struct bkey *k = bch_btree_iter_next(iter);
338 
339 		if (!k)
340 			break;
341 
342 		if (bkey_cmp(&START_KEY(k), insert) >= 0) {
343 			if (KEY_SIZE(k))
344 				break;
345 			else
346 				continue;
347 		}
348 
349 		if (bkey_cmp(k, &START_KEY(insert)) <= 0)
350 			continue;
351 
352 		old_offset = KEY_START(k);
353 		old_size = KEY_SIZE(k);
354 
355 		/*
356 		 * We might overlap with 0 size extents; we can't skip these
357 		 * because if they're in the set we're inserting to we have to
358 		 * adjust them so they don't overlap with the key we're
359 		 * inserting. But we don't want to check them for replace
360 		 * operations.
361 		 */
362 
363 		if (replace_key && KEY_SIZE(k)) {
364 			/*
365 			 * k might have been split since we inserted/found the
366 			 * key we're replacing
367 			 */
368 			unsigned int i;
369 			uint64_t offset = KEY_START(k) -
370 				KEY_START(replace_key);
371 
372 			/* But it must be a subset of the replace key */
373 			if (KEY_START(k) < KEY_START(replace_key) ||
374 			    KEY_OFFSET(k) > KEY_OFFSET(replace_key))
375 				goto check_failed;
376 
377 			/* We didn't find a key that we were supposed to */
378 			if (KEY_START(k) > KEY_START(insert) + sectors_found)
379 				goto check_failed;
380 
381 			if (!bch_bkey_equal_header(k, replace_key))
382 				goto check_failed;
383 
384 			/* skip past gen */
385 			offset <<= 8;
386 
387 			BUG_ON(!KEY_PTRS(replace_key));
388 
389 			for (i = 0; i < KEY_PTRS(replace_key); i++)
390 				if (k->ptr[i] != replace_key->ptr[i] + offset)
391 					goto check_failed;
392 
393 			sectors_found = KEY_OFFSET(k) - KEY_START(insert);
394 		}
395 
396 		if (bkey_cmp(insert, k) < 0 &&
397 		    bkey_cmp(&START_KEY(insert), &START_KEY(k)) > 0) {
398 			/*
399 			 * We overlapped in the middle of an existing key: that
400 			 * means we have to split the old key. But we have to do
401 			 * slightly different things depending on whether the
402 			 * old key has been written out yet.
403 			 */
404 
405 			struct bkey *top;
406 
407 			bch_subtract_dirty(k, c, KEY_START(insert),
408 				       KEY_SIZE(insert));
409 
410 			if (bkey_written(b, k)) {
411 				/*
412 				 * We insert a new key to cover the top of the
413 				 * old key, and the old key is modified in place
414 				 * to represent the bottom split.
415 				 *
416 				 * It's completely arbitrary whether the new key
417 				 * is the top or the bottom, but it has to match
418 				 * up with what btree_sort_fixup() does - it
419 				 * doesn't check for this kind of overlap, it
420 				 * depends on us inserting a new key for the top
421 				 * here.
422 				 */
423 				top = bch_bset_search(b, bset_tree_last(b),
424 						      insert);
425 				bch_bset_insert(b, top, k);
426 			} else {
427 				BKEY_PADDED(key) temp;
428 				bkey_copy(&temp.key, k);
429 				bch_bset_insert(b, k, &temp.key);
430 				top = bkey_next(k);
431 			}
432 
433 			bch_cut_front(insert, top);
434 			bch_cut_back(&START_KEY(insert), k);
435 			bch_bset_fix_invalidated_key(b, k);
436 			goto out;
437 		}
438 
439 		if (bkey_cmp(insert, k) < 0) {
440 			bch_cut_front(insert, k);
441 		} else {
442 			if (bkey_cmp(&START_KEY(insert), &START_KEY(k)) > 0)
443 				old_offset = KEY_START(insert);
444 
445 			if (bkey_written(b, k) &&
446 			    bkey_cmp(&START_KEY(insert), &START_KEY(k)) <= 0) {
447 				/*
448 				 * Completely overwrote, so we don't have to
449 				 * invalidate the binary search tree
450 				 */
451 				bch_cut_front(k, k);
452 			} else {
453 				__bch_cut_back(&START_KEY(insert), k);
454 				bch_bset_fix_invalidated_key(b, k);
455 			}
456 		}
457 
458 		bch_subtract_dirty(k, c, old_offset, old_size - KEY_SIZE(k));
459 	}
460 
461 check_failed:
462 	if (replace_key) {
463 		if (!sectors_found) {
464 			return true;
465 		} else if (sectors_found < KEY_SIZE(insert)) {
466 			SET_KEY_OFFSET(insert, KEY_OFFSET(insert) -
467 				       (KEY_SIZE(insert) - sectors_found));
468 			SET_KEY_SIZE(insert, sectors_found);
469 		}
470 	}
471 out:
472 	if (KEY_DIRTY(insert))
473 		bcache_dev_sectors_dirty_add(c, KEY_INODE(insert),
474 					     KEY_START(insert),
475 					     KEY_SIZE(insert));
476 
477 	return false;
478 }
479 
480 bool __bch_extent_invalid(struct cache_set *c, const struct bkey *k)
481 {
482 	char buf[80];
483 
484 	if (!KEY_SIZE(k))
485 		return true;
486 
487 	if (KEY_SIZE(k) > KEY_OFFSET(k))
488 		goto bad;
489 
490 	if (__ptr_invalid(c, k))
491 		goto bad;
492 
493 	return false;
494 bad:
495 	bch_extent_to_text(buf, sizeof(buf), k);
496 	cache_bug(c, "spotted extent %s: %s", buf, bch_ptr_status(c, k));
497 	return true;
498 }
499 
500 static bool bch_extent_invalid(struct btree_keys *bk, const struct bkey *k)
501 {
502 	struct btree *b = container_of(bk, struct btree, keys);
503 
504 	return __bch_extent_invalid(b->c, k);
505 }
506 
507 static bool bch_extent_bad_expensive(struct btree *b, const struct bkey *k,
508 				     unsigned int ptr)
509 {
510 	struct bucket *g = PTR_BUCKET(b->c, k, ptr);
511 	char buf[80];
512 
513 	if (mutex_trylock(&b->c->bucket_lock)) {
514 		if (b->c->gc_mark_valid &&
515 		    (!GC_MARK(g) ||
516 		     GC_MARK(g) == GC_MARK_METADATA ||
517 		     (GC_MARK(g) != GC_MARK_DIRTY && KEY_DIRTY(k))))
518 			goto err;
519 
520 		if (g->prio == BTREE_PRIO)
521 			goto err;
522 
523 		mutex_unlock(&b->c->bucket_lock);
524 	}
525 
526 	return false;
527 err:
528 	mutex_unlock(&b->c->bucket_lock);
529 	bch_extent_to_text(buf, sizeof(buf), k);
530 	btree_bug(b,
531 "inconsistent extent pointer %s:\nbucket %zu pin %i prio %i gen %i last_gc %i mark %llu",
532 		  buf, PTR_BUCKET_NR(b->c, k, ptr), atomic_read(&g->pin),
533 		  g->prio, g->gen, g->last_gc, GC_MARK(g));
534 	return true;
535 }
536 
537 static bool bch_extent_bad(struct btree_keys *bk, const struct bkey *k)
538 {
539 	struct btree *b = container_of(bk, struct btree, keys);
540 	unsigned int i, stale;
541 	char buf[80];
542 
543 	if (!KEY_PTRS(k) ||
544 	    bch_extent_invalid(bk, k))
545 		return true;
546 
547 	for (i = 0; i < KEY_PTRS(k); i++)
548 		if (!ptr_available(b->c, k, i))
549 			return true;
550 
551 	for (i = 0; i < KEY_PTRS(k); i++) {
552 		stale = ptr_stale(b->c, k, i);
553 
554 		if (stale && KEY_DIRTY(k)) {
555 			bch_extent_to_text(buf, sizeof(buf), k);
556 			pr_info("stale dirty pointer, stale %u, key: %s\n",
557 				stale, buf);
558 		}
559 
560 		btree_bug_on(stale > BUCKET_GC_GEN_MAX, b,
561 			     "key too stale: %i, need_gc %u",
562 			     stale, b->c->need_gc);
563 
564 		if (stale)
565 			return true;
566 
567 		if (expensive_debug_checks(b->c) &&
568 		    bch_extent_bad_expensive(b, k, i))
569 			return true;
570 	}
571 
572 	return false;
573 }
574 
575 static uint64_t merge_chksums(struct bkey *l, struct bkey *r)
576 {
577 	return (l->ptr[KEY_PTRS(l)] + r->ptr[KEY_PTRS(r)]) &
578 		~((uint64_t)1 << 63);
579 }
580 
581 static bool bch_extent_merge(struct btree_keys *bk,
582 			     struct bkey *l,
583 			     struct bkey *r)
584 {
585 	struct btree *b = container_of(bk, struct btree, keys);
586 	unsigned int i;
587 
588 	if (key_merging_disabled(b->c))
589 		return false;
590 
591 	for (i = 0; i < KEY_PTRS(l); i++)
592 		if (l->ptr[i] + MAKE_PTR(0, KEY_SIZE(l), 0) != r->ptr[i] ||
593 		    PTR_BUCKET_NR(b->c, l, i) != PTR_BUCKET_NR(b->c, r, i))
594 			return false;
595 
596 	/* Keys with no pointers aren't restricted to one bucket and could
597 	 * overflow KEY_SIZE
598 	 */
599 	if (KEY_SIZE(l) + KEY_SIZE(r) > USHRT_MAX) {
600 		SET_KEY_OFFSET(l, KEY_OFFSET(l) + USHRT_MAX - KEY_SIZE(l));
601 		SET_KEY_SIZE(l, USHRT_MAX);
602 
603 		bch_cut_front(l, r);
604 		return false;
605 	}
606 
607 	if (KEY_CSUM(l)) {
608 		if (KEY_CSUM(r))
609 			l->ptr[KEY_PTRS(l)] = merge_chksums(l, r);
610 		else
611 			SET_KEY_CSUM(l, 0);
612 	}
613 
614 	SET_KEY_OFFSET(l, KEY_OFFSET(l) + KEY_SIZE(r));
615 	SET_KEY_SIZE(l, KEY_SIZE(l) + KEY_SIZE(r));
616 
617 	return true;
618 }
619 
620 const struct btree_keys_ops bch_extent_keys_ops = {
621 	.sort_cmp	= bch_extent_sort_cmp,
622 	.sort_fixup	= bch_extent_sort_fixup,
623 	.insert_fixup	= bch_extent_insert_fixup,
624 	.key_invalid	= bch_extent_invalid,
625 	.key_bad	= bch_extent_bad,
626 	.key_merge	= bch_extent_merge,
627 	.key_to_text	= bch_extent_to_text,
628 	.key_dump	= bch_bkey_dump,
629 	.is_extents	= true,
630 };
631