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
29 #include <linux/slab.h>
30 #include <linux/bitops.h>
31 #include <linux/hash.h>
32 #include <linux/kthread.h>
33 #include <linux/prefetch.h>
34 #include <linux/random.h>
35 #include <linux/rcupdate.h>
36 #include <linux/sched/clock.h>
37 #include <linux/rculist.h>
38 #include <linux/delay.h>
39 #include <trace/events/bcache.h>
40
41 /*
42 * Todo:
43 * register_bcache: Return errors out to userspace correctly
44 *
45 * Writeback: don't undirty key until after a cache flush
46 *
47 * Create an iterator for key pointers
48 *
49 * On btree write error, mark bucket such that it won't be freed from the cache
50 *
51 * Journalling:
52 * Check for bad keys in replay
53 * Propagate barriers
54 * Refcount journal entries in journal_replay
55 *
56 * Garbage collection:
57 * Finish incremental gc
58 * Gc should free old UUIDs, data for invalid UUIDs
59 *
60 * Provide a way to list backing device UUIDs we have data cached for, and
61 * probably how long it's been since we've seen them, and a way to invalidate
62 * dirty data for devices that will never be attached again
63 *
64 * Keep 1 min/5 min/15 min statistics of how busy a block device has been, so
65 * that based on that and how much dirty data we have we can keep writeback
66 * from being starved
67 *
68 * Add a tracepoint or somesuch to watch for writeback starvation
69 *
70 * When btree depth > 1 and splitting an interior node, we have to make sure
71 * alloc_bucket() cannot fail. This should be true but is not completely
72 * obvious.
73 *
74 * Plugging?
75 *
76 * If data write is less than hard sector size of ssd, round up offset in open
77 * bucket to the next whole sector
78 *
79 * Superblock needs to be fleshed out for multiple cache devices
80 *
81 * Add a sysfs tunable for the number of writeback IOs in flight
82 *
83 * Add a sysfs tunable for the number of open data buckets
84 *
85 * IO tracking: Can we track when one process is doing io on behalf of another?
86 * IO tracking: Don't use just an average, weigh more recent stuff higher
87 *
88 * Test module load/unload
89 */
90
91 #define MAX_NEED_GC 64
92 #define MAX_SAVE_PRIO 72
93 #define MAX_GC_TIMES 100
94 #define MIN_GC_NODES 100
95 #define GC_SLEEP_MS 100
96
97 #define PTR_DIRTY_BIT (((uint64_t) 1 << 36))
98
99 #define PTR_HASH(c, k) \
100 (((k)->ptr[0] >> c->bucket_bits) | PTR_GEN(k, 0))
101
102 static struct workqueue_struct *btree_io_wq;
103
104 #define insert_lock(s, b) ((b)->level <= (s)->lock)
105
106
write_block(struct btree * b)107 static inline struct bset *write_block(struct btree *b)
108 {
109 return ((void *) btree_bset_first(b)) + b->written * block_bytes(b->c->cache);
110 }
111
bch_btree_init_next(struct btree * b)112 static void bch_btree_init_next(struct btree *b)
113 {
114 /* If not a leaf node, always sort */
115 if (b->level && b->keys.nsets)
116 bch_btree_sort(&b->keys, &b->c->sort);
117 else
118 bch_btree_sort_lazy(&b->keys, &b->c->sort);
119
120 if (b->written < btree_blocks(b))
121 bch_bset_init_next(&b->keys, write_block(b),
122 bset_magic(&b->c->cache->sb));
123
124 }
125
126 /* Btree key manipulation */
127
bkey_put(struct cache_set * c,struct bkey * k)128 void bkey_put(struct cache_set *c, struct bkey *k)
129 {
130 unsigned int i;
131
132 for (i = 0; i < KEY_PTRS(k); i++)
133 if (ptr_available(c, k, i))
134 atomic_dec_bug(&PTR_BUCKET(c, k, i)->pin);
135 }
136
137 /* Btree IO */
138
btree_csum_set(struct btree * b,struct bset * i)139 static uint64_t btree_csum_set(struct btree *b, struct bset *i)
140 {
141 uint64_t crc = b->key.ptr[0];
142 void *data = (void *) i + 8, *end = bset_bkey_last(i);
143
144 crc = crc64_be(crc, data, end - data);
145 return crc ^ 0xffffffffffffffffULL;
146 }
147
bch_btree_node_read_done(struct btree * b)148 void bch_btree_node_read_done(struct btree *b)
149 {
150 const char *err = "bad btree header";
151 struct bset *i = btree_bset_first(b);
152 struct btree_iter *iter;
153
154 /*
155 * c->fill_iter can allocate an iterator with more memory space
156 * than static MAX_BSETS.
157 * See the comment arount cache_set->fill_iter.
158 */
159 iter = mempool_alloc(&b->c->fill_iter, GFP_NOIO);
160 iter->size = b->c->cache->sb.bucket_size / b->c->cache->sb.block_size;
161 iter->used = 0;
162
163 #ifdef CONFIG_BCACHE_DEBUG
164 iter->b = &b->keys;
165 #endif
166
167 if (!i->seq)
168 goto err;
169
170 for (;
171 b->written < btree_blocks(b) && i->seq == b->keys.set[0].data->seq;
172 i = write_block(b)) {
173 err = "unsupported bset version";
174 if (i->version > BCACHE_BSET_VERSION)
175 goto err;
176
177 err = "bad btree header";
178 if (b->written + set_blocks(i, block_bytes(b->c->cache)) >
179 btree_blocks(b))
180 goto err;
181
182 err = "bad magic";
183 if (i->magic != bset_magic(&b->c->cache->sb))
184 goto err;
185
186 err = "bad checksum";
187 switch (i->version) {
188 case 0:
189 if (i->csum != csum_set(i))
190 goto err;
191 break;
192 case BCACHE_BSET_VERSION:
193 if (i->csum != btree_csum_set(b, i))
194 goto err;
195 break;
196 }
197
198 err = "empty set";
199 if (i != b->keys.set[0].data && !i->keys)
200 goto err;
201
202 bch_btree_iter_push(iter, i->start, bset_bkey_last(i));
203
204 b->written += set_blocks(i, block_bytes(b->c->cache));
205 }
206
207 err = "corrupted btree";
208 for (i = write_block(b);
209 bset_sector_offset(&b->keys, i) < KEY_SIZE(&b->key);
210 i = ((void *) i) + block_bytes(b->c->cache))
211 if (i->seq == b->keys.set[0].data->seq)
212 goto err;
213
214 bch_btree_sort_and_fix_extents(&b->keys, iter, &b->c->sort);
215
216 i = b->keys.set[0].data;
217 err = "short btree key";
218 if (b->keys.set[0].size &&
219 bkey_cmp(&b->key, &b->keys.set[0].end) < 0)
220 goto err;
221
222 if (b->written < btree_blocks(b))
223 bch_bset_init_next(&b->keys, write_block(b),
224 bset_magic(&b->c->cache->sb));
225 out:
226 mempool_free(iter, &b->c->fill_iter);
227 return;
228 err:
229 set_btree_node_io_error(b);
230 bch_cache_set_error(b->c, "%s at bucket %zu, block %u, %u keys",
231 err, PTR_BUCKET_NR(b->c, &b->key, 0),
232 bset_block_offset(b, i), i->keys);
233 goto out;
234 }
235
btree_node_read_endio(struct bio * bio)236 static void btree_node_read_endio(struct bio *bio)
237 {
238 struct closure *cl = bio->bi_private;
239
240 closure_put(cl);
241 }
242
bch_btree_node_read(struct btree * b)243 static void bch_btree_node_read(struct btree *b)
244 {
245 uint64_t start_time = local_clock();
246 struct closure cl;
247 struct bio *bio;
248
249 trace_bcache_btree_read(b);
250
251 closure_init_stack(&cl);
252
253 bio = bch_bbio_alloc(b->c);
254 bio->bi_iter.bi_size = KEY_SIZE(&b->key) << 9;
255 bio->bi_end_io = btree_node_read_endio;
256 bio->bi_private = &cl;
257 bio->bi_opf = REQ_OP_READ | REQ_META;
258
259 bch_bio_map(bio, b->keys.set[0].data);
260
261 bch_submit_bbio(bio, b->c, &b->key, 0);
262 closure_sync(&cl);
263
264 if (bio->bi_status)
265 set_btree_node_io_error(b);
266
267 bch_bbio_free(bio, b->c);
268
269 if (btree_node_io_error(b))
270 goto err;
271
272 bch_btree_node_read_done(b);
273 bch_time_stats_update(&b->c->btree_read_time, start_time);
274
275 return;
276 err:
277 bch_cache_set_error(b->c, "io error reading bucket %zu",
278 PTR_BUCKET_NR(b->c, &b->key, 0));
279 }
280
btree_complete_write(struct btree * b,struct btree_write * w)281 static void btree_complete_write(struct btree *b, struct btree_write *w)
282 {
283 if (w->prio_blocked &&
284 !atomic_sub_return(w->prio_blocked, &b->c->prio_blocked))
285 wake_up_allocators(b->c);
286
287 if (w->journal) {
288 atomic_dec_bug(w->journal);
289 __closure_wake_up(&b->c->journal.wait);
290 }
291
292 w->prio_blocked = 0;
293 w->journal = NULL;
294 }
295
CLOSURE_CALLBACK(btree_node_write_unlock)296 static CLOSURE_CALLBACK(btree_node_write_unlock)
297 {
298 closure_type(b, struct btree, io);
299
300 up(&b->io_mutex);
301 }
302
CLOSURE_CALLBACK(__btree_node_write_done)303 static CLOSURE_CALLBACK(__btree_node_write_done)
304 {
305 closure_type(b, struct btree, io);
306 struct btree_write *w = btree_prev_write(b);
307
308 bch_bbio_free(b->bio, b->c);
309 b->bio = NULL;
310 btree_complete_write(b, w);
311
312 if (btree_node_dirty(b))
313 queue_delayed_work(btree_io_wq, &b->work, 30 * HZ);
314
315 closure_return_with_destructor(cl, btree_node_write_unlock);
316 }
317
CLOSURE_CALLBACK(btree_node_write_done)318 static CLOSURE_CALLBACK(btree_node_write_done)
319 {
320 closure_type(b, struct btree, io);
321
322 bio_free_pages(b->bio);
323 __btree_node_write_done(&cl->work);
324 }
325
btree_node_write_endio(struct bio * bio)326 static void btree_node_write_endio(struct bio *bio)
327 {
328 struct closure *cl = bio->bi_private;
329 struct btree *b = container_of(cl, struct btree, io);
330
331 if (bio->bi_status)
332 set_btree_node_io_error(b);
333
334 bch_bbio_count_io_errors(b->c, bio, bio->bi_status, "writing btree");
335 closure_put(cl);
336 }
337
do_btree_node_write(struct btree * b)338 static void do_btree_node_write(struct btree *b)
339 {
340 struct closure *cl = &b->io;
341 struct bset *i = btree_bset_last(b);
342 BKEY_PADDED(key) k;
343
344 i->version = BCACHE_BSET_VERSION;
345 i->csum = btree_csum_set(b, i);
346
347 BUG_ON(b->bio);
348 b->bio = bch_bbio_alloc(b->c);
349
350 b->bio->bi_end_io = btree_node_write_endio;
351 b->bio->bi_private = cl;
352 b->bio->bi_iter.bi_size = roundup(set_bytes(i), block_bytes(b->c->cache));
353 b->bio->bi_opf = REQ_OP_WRITE | REQ_META | REQ_FUA;
354 bch_bio_map(b->bio, i);
355
356 /*
357 * If we're appending to a leaf node, we don't technically need FUA -
358 * this write just needs to be persisted before the next journal write,
359 * which will be marked FLUSH|FUA.
360 *
361 * Similarly if we're writing a new btree root - the pointer is going to
362 * be in the next journal entry.
363 *
364 * But if we're writing a new btree node (that isn't a root) or
365 * appending to a non leaf btree node, we need either FUA or a flush
366 * when we write the parent with the new pointer. FUA is cheaper than a
367 * flush, and writes appending to leaf nodes aren't blocking anything so
368 * just make all btree node writes FUA to keep things sane.
369 */
370
371 bkey_copy(&k.key, &b->key);
372 SET_PTR_OFFSET(&k.key, 0, PTR_OFFSET(&k.key, 0) +
373 bset_sector_offset(&b->keys, i));
374
375 if (!bch_bio_alloc_pages(b->bio, __GFP_NOWARN|GFP_NOWAIT)) {
376 struct bio_vec *bv;
377 void *addr = (void *) ((unsigned long) i & ~(PAGE_SIZE - 1));
378 struct bvec_iter_all iter_all;
379
380 bio_for_each_segment_all(bv, b->bio, iter_all) {
381 memcpy(page_address(bv->bv_page), addr, PAGE_SIZE);
382 addr += PAGE_SIZE;
383 }
384
385 bch_submit_bbio(b->bio, b->c, &k.key, 0);
386
387 continue_at(cl, btree_node_write_done, NULL);
388 } else {
389 /*
390 * No problem for multipage bvec since the bio is
391 * just allocated
392 */
393 b->bio->bi_vcnt = 0;
394 bch_bio_map(b->bio, i);
395
396 bch_submit_bbio(b->bio, b->c, &k.key, 0);
397
398 closure_sync(cl);
399 continue_at_nobarrier(cl, __btree_node_write_done, NULL);
400 }
401 }
402
__bch_btree_node_write(struct btree * b,struct closure * parent)403 void __bch_btree_node_write(struct btree *b, struct closure *parent)
404 {
405 struct bset *i = btree_bset_last(b);
406
407 lockdep_assert_held(&b->write_lock);
408
409 trace_bcache_btree_write(b);
410
411 BUG_ON(current->bio_list);
412 BUG_ON(b->written >= btree_blocks(b));
413 BUG_ON(b->written && !i->keys);
414 BUG_ON(btree_bset_first(b)->seq != i->seq);
415 bch_check_keys(&b->keys, "writing");
416
417 cancel_delayed_work(&b->work);
418
419 /* If caller isn't waiting for write, parent refcount is cache set */
420 down(&b->io_mutex);
421 closure_init(&b->io, parent ?: &b->c->cl);
422
423 clear_bit(BTREE_NODE_dirty, &b->flags);
424 change_bit(BTREE_NODE_write_idx, &b->flags);
425
426 do_btree_node_write(b);
427
428 atomic_long_add(set_blocks(i, block_bytes(b->c->cache)) * b->c->cache->sb.block_size,
429 &b->c->cache->btree_sectors_written);
430
431 b->written += set_blocks(i, block_bytes(b->c->cache));
432 }
433
bch_btree_node_write(struct btree * b,struct closure * parent)434 void bch_btree_node_write(struct btree *b, struct closure *parent)
435 {
436 unsigned int nsets = b->keys.nsets;
437
438 lockdep_assert_held(&b->lock);
439
440 __bch_btree_node_write(b, parent);
441
442 /*
443 * do verify if there was more than one set initially (i.e. we did a
444 * sort) and we sorted down to a single set:
445 */
446 if (nsets && !b->keys.nsets)
447 bch_btree_verify(b);
448
449 bch_btree_init_next(b);
450 }
451
bch_btree_node_write_sync(struct btree * b)452 static void bch_btree_node_write_sync(struct btree *b)
453 {
454 struct closure cl;
455
456 closure_init_stack(&cl);
457
458 mutex_lock(&b->write_lock);
459 bch_btree_node_write(b, &cl);
460 mutex_unlock(&b->write_lock);
461
462 closure_sync(&cl);
463 }
464
btree_node_write_work(struct work_struct * w)465 static void btree_node_write_work(struct work_struct *w)
466 {
467 struct btree *b = container_of(to_delayed_work(w), struct btree, work);
468
469 mutex_lock(&b->write_lock);
470 if (btree_node_dirty(b))
471 __bch_btree_node_write(b, NULL);
472 mutex_unlock(&b->write_lock);
473 }
474
bch_btree_leaf_dirty(struct btree * b,atomic_t * journal_ref)475 static void bch_btree_leaf_dirty(struct btree *b, atomic_t *journal_ref)
476 {
477 struct bset *i = btree_bset_last(b);
478 struct btree_write *w = btree_current_write(b);
479
480 lockdep_assert_held(&b->write_lock);
481
482 BUG_ON(!b->written);
483 BUG_ON(!i->keys);
484
485 if (!btree_node_dirty(b))
486 queue_delayed_work(btree_io_wq, &b->work, 30 * HZ);
487
488 set_btree_node_dirty(b);
489
490 /*
491 * w->journal is always the oldest journal pin of all bkeys
492 * in the leaf node, to make sure the oldest jset seq won't
493 * be increased before this btree node is flushed.
494 */
495 if (journal_ref) {
496 if (w->journal &&
497 journal_pin_cmp(b->c, w->journal, journal_ref)) {
498 atomic_dec_bug(w->journal);
499 w->journal = NULL;
500 }
501
502 if (!w->journal) {
503 w->journal = journal_ref;
504 atomic_inc(w->journal);
505 }
506 }
507
508 /* Force write if set is too big */
509 if (set_bytes(i) > PAGE_SIZE - 48 &&
510 !current->bio_list)
511 bch_btree_node_write(b, NULL);
512 }
513
514 /*
515 * Btree in memory cache - allocation/freeing
516 * mca -> memory cache
517 */
518
519 #define mca_reserve(c) (((!IS_ERR_OR_NULL(c->root) && c->root->level) \
520 ? c->root->level : 1) * 8 + 16)
521 #define mca_can_free(c) \
522 max_t(int, 0, c->btree_cache_used - mca_reserve(c))
523
mca_data_free(struct btree * b)524 static void mca_data_free(struct btree *b)
525 {
526 BUG_ON(b->io_mutex.count != 1);
527
528 bch_btree_keys_free(&b->keys);
529
530 b->c->btree_cache_used--;
531 list_move(&b->list, &b->c->btree_cache_freed);
532 }
533
mca_bucket_free(struct btree * b)534 static void mca_bucket_free(struct btree *b)
535 {
536 BUG_ON(btree_node_dirty(b));
537
538 b->key.ptr[0] = 0;
539 hlist_del_init_rcu(&b->hash);
540 list_move(&b->list, &b->c->btree_cache_freeable);
541 }
542
btree_order(struct bkey * k)543 static unsigned int btree_order(struct bkey *k)
544 {
545 return ilog2(KEY_SIZE(k) / PAGE_SECTORS ?: 1);
546 }
547
mca_data_alloc(struct btree * b,struct bkey * k,gfp_t gfp)548 static void mca_data_alloc(struct btree *b, struct bkey *k, gfp_t gfp)
549 {
550 if (!bch_btree_keys_alloc(&b->keys,
551 max_t(unsigned int,
552 ilog2(b->c->btree_pages),
553 btree_order(k)),
554 gfp)) {
555 b->c->btree_cache_used++;
556 list_move(&b->list, &b->c->btree_cache);
557 } else {
558 list_move(&b->list, &b->c->btree_cache_freed);
559 }
560 }
561
562 #define cmp_int(l, r) ((l > r) - (l < r))
563
564 #ifdef CONFIG_PROVE_LOCKING
btree_lock_cmp_fn(const struct lockdep_map * _a,const struct lockdep_map * _b)565 static int btree_lock_cmp_fn(const struct lockdep_map *_a,
566 const struct lockdep_map *_b)
567 {
568 const struct btree *a = container_of(_a, struct btree, lock.dep_map);
569 const struct btree *b = container_of(_b, struct btree, lock.dep_map);
570
571 return -cmp_int(a->level, b->level) ?: bkey_cmp(&a->key, &b->key);
572 }
573
btree_lock_print_fn(const struct lockdep_map * map)574 static void btree_lock_print_fn(const struct lockdep_map *map)
575 {
576 const struct btree *b = container_of(map, struct btree, lock.dep_map);
577
578 printk(KERN_CONT " l=%u %llu:%llu", b->level,
579 KEY_INODE(&b->key), KEY_OFFSET(&b->key));
580 }
581 #endif
582
mca_bucket_alloc(struct cache_set * c,struct bkey * k,gfp_t gfp)583 static struct btree *mca_bucket_alloc(struct cache_set *c,
584 struct bkey *k, gfp_t gfp)
585 {
586 /*
587 * kzalloc() is necessary here for initialization,
588 * see code comments in bch_btree_keys_init().
589 */
590 struct btree *b = kzalloc(sizeof(struct btree), gfp);
591
592 if (!b)
593 return NULL;
594
595 init_rwsem(&b->lock);
596 lock_set_cmp_fn(&b->lock, btree_lock_cmp_fn, btree_lock_print_fn);
597 mutex_init(&b->write_lock);
598 lockdep_set_novalidate_class(&b->write_lock);
599 INIT_LIST_HEAD(&b->list);
600 INIT_DELAYED_WORK(&b->work, btree_node_write_work);
601 b->c = c;
602 sema_init(&b->io_mutex, 1);
603
604 mca_data_alloc(b, k, gfp);
605 return b;
606 }
607
mca_reap(struct btree * b,unsigned int min_order,bool flush)608 static int mca_reap(struct btree *b, unsigned int min_order, bool flush)
609 {
610 struct closure cl;
611
612 closure_init_stack(&cl);
613 lockdep_assert_held(&b->c->bucket_lock);
614
615 if (!down_write_trylock(&b->lock))
616 return -ENOMEM;
617
618 BUG_ON(btree_node_dirty(b) && !b->keys.set[0].data);
619
620 if (b->keys.page_order < min_order)
621 goto out_unlock;
622
623 if (!flush) {
624 if (btree_node_dirty(b))
625 goto out_unlock;
626
627 if (down_trylock(&b->io_mutex))
628 goto out_unlock;
629 up(&b->io_mutex);
630 }
631
632 retry:
633 /*
634 * BTREE_NODE_dirty might be cleared in btree_flush_btree() by
635 * __bch_btree_node_write(). To avoid an extra flush, acquire
636 * b->write_lock before checking BTREE_NODE_dirty bit.
637 */
638 mutex_lock(&b->write_lock);
639 /*
640 * If this btree node is selected in btree_flush_write() by journal
641 * code, delay and retry until the node is flushed by journal code
642 * and BTREE_NODE_journal_flush bit cleared by btree_flush_write().
643 */
644 if (btree_node_journal_flush(b)) {
645 pr_debug("bnode %p is flushing by journal, retry\n", b);
646 mutex_unlock(&b->write_lock);
647 udelay(1);
648 goto retry;
649 }
650
651 if (btree_node_dirty(b))
652 __bch_btree_node_write(b, &cl);
653 mutex_unlock(&b->write_lock);
654
655 closure_sync(&cl);
656
657 /* wait for any in flight btree write */
658 down(&b->io_mutex);
659 up(&b->io_mutex);
660
661 return 0;
662 out_unlock:
663 rw_unlock(true, b);
664 return -ENOMEM;
665 }
666
bch_mca_scan(struct shrinker * shrink,struct shrink_control * sc)667 static unsigned long bch_mca_scan(struct shrinker *shrink,
668 struct shrink_control *sc)
669 {
670 struct cache_set *c = shrink->private_data;
671 struct btree *b, *t;
672 unsigned long i, nr = sc->nr_to_scan;
673 unsigned long freed = 0;
674 unsigned int btree_cache_used;
675
676 if (c->shrinker_disabled)
677 return SHRINK_STOP;
678
679 if (c->btree_cache_alloc_lock)
680 return SHRINK_STOP;
681
682 /* Return -1 if we can't do anything right now */
683 if (sc->gfp_mask & __GFP_IO)
684 mutex_lock(&c->bucket_lock);
685 else if (!mutex_trylock(&c->bucket_lock))
686 return -1;
687
688 /*
689 * It's _really_ critical that we don't free too many btree nodes - we
690 * have to always leave ourselves a reserve. The reserve is how we
691 * guarantee that allocating memory for a new btree node can always
692 * succeed, so that inserting keys into the btree can always succeed and
693 * IO can always make forward progress:
694 */
695 nr /= c->btree_pages;
696 if (nr == 0)
697 nr = 1;
698 nr = min_t(unsigned long, nr, mca_can_free(c));
699
700 i = 0;
701 btree_cache_used = c->btree_cache_used;
702 list_for_each_entry_safe_reverse(b, t, &c->btree_cache_freeable, list) {
703 if (nr <= 0)
704 goto out;
705
706 if (!mca_reap(b, 0, false)) {
707 mca_data_free(b);
708 rw_unlock(true, b);
709 freed++;
710 }
711 nr--;
712 i++;
713 }
714
715 list_for_each_entry_safe_reverse(b, t, &c->btree_cache, list) {
716 if (nr <= 0 || i >= btree_cache_used)
717 goto out;
718
719 if (!mca_reap(b, 0, false)) {
720 mca_bucket_free(b);
721 mca_data_free(b);
722 rw_unlock(true, b);
723 freed++;
724 }
725
726 nr--;
727 i++;
728 }
729 out:
730 mutex_unlock(&c->bucket_lock);
731 return freed * c->btree_pages;
732 }
733
bch_mca_count(struct shrinker * shrink,struct shrink_control * sc)734 static unsigned long bch_mca_count(struct shrinker *shrink,
735 struct shrink_control *sc)
736 {
737 struct cache_set *c = shrink->private_data;
738
739 if (c->shrinker_disabled)
740 return 0;
741
742 if (c->btree_cache_alloc_lock)
743 return 0;
744
745 return mca_can_free(c) * c->btree_pages;
746 }
747
bch_btree_cache_free(struct cache_set * c)748 void bch_btree_cache_free(struct cache_set *c)
749 {
750 struct btree *b;
751 struct closure cl;
752
753 closure_init_stack(&cl);
754
755 if (c->shrink)
756 shrinker_free(c->shrink);
757
758 mutex_lock(&c->bucket_lock);
759
760 #ifdef CONFIG_BCACHE_DEBUG
761 if (c->verify_data)
762 list_move(&c->verify_data->list, &c->btree_cache);
763
764 free_pages((unsigned long) c->verify_ondisk, ilog2(meta_bucket_pages(&c->cache->sb)));
765 #endif
766
767 list_splice(&c->btree_cache_freeable,
768 &c->btree_cache);
769
770 while (!list_empty(&c->btree_cache)) {
771 b = list_first_entry(&c->btree_cache, struct btree, list);
772
773 /*
774 * This function is called by cache_set_free(), no I/O
775 * request on cache now, it is unnecessary to acquire
776 * b->write_lock before clearing BTREE_NODE_dirty anymore.
777 */
778 if (btree_node_dirty(b)) {
779 btree_complete_write(b, btree_current_write(b));
780 clear_bit(BTREE_NODE_dirty, &b->flags);
781 }
782 mca_data_free(b);
783 }
784
785 while (!list_empty(&c->btree_cache_freed)) {
786 b = list_first_entry(&c->btree_cache_freed,
787 struct btree, list);
788 list_del(&b->list);
789 cancel_delayed_work_sync(&b->work);
790 kfree(b);
791 }
792
793 mutex_unlock(&c->bucket_lock);
794 }
795
bch_btree_cache_alloc(struct cache_set * c)796 int bch_btree_cache_alloc(struct cache_set *c)
797 {
798 unsigned int i;
799
800 for (i = 0; i < mca_reserve(c); i++)
801 if (!mca_bucket_alloc(c, &ZERO_KEY, GFP_KERNEL))
802 return -ENOMEM;
803
804 list_splice_init(&c->btree_cache,
805 &c->btree_cache_freeable);
806
807 #ifdef CONFIG_BCACHE_DEBUG
808 mutex_init(&c->verify_lock);
809
810 c->verify_ondisk = (void *)
811 __get_free_pages(GFP_KERNEL|__GFP_COMP,
812 ilog2(meta_bucket_pages(&c->cache->sb)));
813 if (!c->verify_ondisk) {
814 /*
815 * Don't worry about the mca_rereserve buckets
816 * allocated in previous for-loop, they will be
817 * handled properly in bch_cache_set_unregister().
818 */
819 return -ENOMEM;
820 }
821
822 c->verify_data = mca_bucket_alloc(c, &ZERO_KEY, GFP_KERNEL);
823
824 if (c->verify_data &&
825 c->verify_data->keys.set->data)
826 list_del_init(&c->verify_data->list);
827 else
828 c->verify_data = NULL;
829 #endif
830
831 c->shrink = shrinker_alloc(0, "md-bcache:%pU", c->set_uuid);
832 if (!c->shrink) {
833 pr_warn("bcache: %s: could not allocate shrinker\n", __func__);
834 return 0;
835 }
836
837 c->shrink->count_objects = bch_mca_count;
838 c->shrink->scan_objects = bch_mca_scan;
839 c->shrink->seeks = 4;
840 c->shrink->batch = c->btree_pages * 2;
841 c->shrink->private_data = c;
842
843 shrinker_register(c->shrink);
844
845 return 0;
846 }
847
848 /* Btree in memory cache - hash table */
849
mca_hash(struct cache_set * c,struct bkey * k)850 static struct hlist_head *mca_hash(struct cache_set *c, struct bkey *k)
851 {
852 return &c->bucket_hash[hash_32(PTR_HASH(c, k), BUCKET_HASH_BITS)];
853 }
854
mca_find(struct cache_set * c,struct bkey * k)855 static struct btree *mca_find(struct cache_set *c, struct bkey *k)
856 {
857 struct btree *b;
858
859 rcu_read_lock();
860 hlist_for_each_entry_rcu(b, mca_hash(c, k), hash)
861 if (PTR_HASH(c, &b->key) == PTR_HASH(c, k))
862 goto out;
863 b = NULL;
864 out:
865 rcu_read_unlock();
866 return b;
867 }
868
mca_cannibalize_lock(struct cache_set * c,struct btree_op * op)869 static int mca_cannibalize_lock(struct cache_set *c, struct btree_op *op)
870 {
871 spin_lock(&c->btree_cannibalize_lock);
872 if (likely(c->btree_cache_alloc_lock == NULL)) {
873 c->btree_cache_alloc_lock = current;
874 } else if (c->btree_cache_alloc_lock != current) {
875 if (op)
876 prepare_to_wait(&c->btree_cache_wait, &op->wait,
877 TASK_UNINTERRUPTIBLE);
878 spin_unlock(&c->btree_cannibalize_lock);
879 return -EINTR;
880 }
881 spin_unlock(&c->btree_cannibalize_lock);
882
883 return 0;
884 }
885
mca_cannibalize(struct cache_set * c,struct btree_op * op,struct bkey * k)886 static struct btree *mca_cannibalize(struct cache_set *c, struct btree_op *op,
887 struct bkey *k)
888 {
889 struct btree *b;
890
891 trace_bcache_btree_cache_cannibalize(c);
892
893 if (mca_cannibalize_lock(c, op))
894 return ERR_PTR(-EINTR);
895
896 list_for_each_entry_reverse(b, &c->btree_cache, list)
897 if (!mca_reap(b, btree_order(k), false))
898 return b;
899
900 list_for_each_entry_reverse(b, &c->btree_cache, list)
901 if (!mca_reap(b, btree_order(k), true))
902 return b;
903
904 WARN(1, "btree cache cannibalize failed\n");
905 return ERR_PTR(-ENOMEM);
906 }
907
908 /*
909 * We can only have one thread cannibalizing other cached btree nodes at a time,
910 * or we'll deadlock. We use an open coded mutex to ensure that, which a
911 * cannibalize_bucket() will take. This means every time we unlock the root of
912 * the btree, we need to release this lock if we have it held.
913 */
bch_cannibalize_unlock(struct cache_set * c)914 void bch_cannibalize_unlock(struct cache_set *c)
915 {
916 spin_lock(&c->btree_cannibalize_lock);
917 if (c->btree_cache_alloc_lock == current) {
918 c->btree_cache_alloc_lock = NULL;
919 wake_up(&c->btree_cache_wait);
920 }
921 spin_unlock(&c->btree_cannibalize_lock);
922 }
923
mca_alloc(struct cache_set * c,struct btree_op * op,struct bkey * k,int level)924 static struct btree *mca_alloc(struct cache_set *c, struct btree_op *op,
925 struct bkey *k, int level)
926 {
927 struct btree *b;
928
929 BUG_ON(current->bio_list);
930
931 lockdep_assert_held(&c->bucket_lock);
932
933 if (mca_find(c, k))
934 return NULL;
935
936 /* btree_free() doesn't free memory; it sticks the node on the end of
937 * the list. Check if there's any freed nodes there:
938 */
939 list_for_each_entry(b, &c->btree_cache_freeable, list)
940 if (!mca_reap(b, btree_order(k), false))
941 goto out;
942
943 /* We never free struct btree itself, just the memory that holds the on
944 * disk node. Check the freed list before allocating a new one:
945 */
946 list_for_each_entry(b, &c->btree_cache_freed, list)
947 if (!mca_reap(b, 0, false)) {
948 mca_data_alloc(b, k, __GFP_NOWARN|GFP_NOIO);
949 if (!b->keys.set[0].data)
950 goto err;
951 else
952 goto out;
953 }
954
955 b = mca_bucket_alloc(c, k, __GFP_NOWARN|GFP_NOIO);
956 if (!b)
957 goto err;
958
959 BUG_ON(!down_write_trylock(&b->lock));
960 if (!b->keys.set->data)
961 goto err;
962 out:
963 BUG_ON(b->io_mutex.count != 1);
964
965 bkey_copy(&b->key, k);
966 list_move(&b->list, &c->btree_cache);
967 hlist_del_init_rcu(&b->hash);
968 hlist_add_head_rcu(&b->hash, mca_hash(c, k));
969
970 lock_set_subclass(&b->lock.dep_map, level + 1, _THIS_IP_);
971 b->parent = (void *) ~0UL;
972 b->flags = 0;
973 b->written = 0;
974 b->level = level;
975
976 if (!b->level)
977 bch_btree_keys_init(&b->keys, &bch_extent_keys_ops,
978 &b->c->expensive_debug_checks);
979 else
980 bch_btree_keys_init(&b->keys, &bch_btree_keys_ops,
981 &b->c->expensive_debug_checks);
982
983 return b;
984 err:
985 if (b)
986 rw_unlock(true, b);
987
988 b = mca_cannibalize(c, op, k);
989 if (!IS_ERR(b))
990 goto out;
991
992 return b;
993 }
994
995 /*
996 * bch_btree_node_get - find a btree node in the cache and lock it, reading it
997 * in from disk if necessary.
998 *
999 * If IO is necessary and running under submit_bio_noacct, returns -EAGAIN.
1000 *
1001 * The btree node will have either a read or a write lock held, depending on
1002 * level and op->lock.
1003 *
1004 * Note: Only error code or btree pointer will be returned, it is unncessary
1005 * for callers to check NULL pointer.
1006 */
bch_btree_node_get(struct cache_set * c,struct btree_op * op,struct bkey * k,int level,bool write,struct btree * parent)1007 struct btree *bch_btree_node_get(struct cache_set *c, struct btree_op *op,
1008 struct bkey *k, int level, bool write,
1009 struct btree *parent)
1010 {
1011 int i = 0;
1012 struct btree *b;
1013
1014 BUG_ON(level < 0);
1015 retry:
1016 b = mca_find(c, k);
1017
1018 if (!b) {
1019 if (current->bio_list)
1020 return ERR_PTR(-EAGAIN);
1021
1022 mutex_lock(&c->bucket_lock);
1023 b = mca_alloc(c, op, k, level);
1024 mutex_unlock(&c->bucket_lock);
1025
1026 if (!b)
1027 goto retry;
1028 if (IS_ERR(b))
1029 return b;
1030
1031 bch_btree_node_read(b);
1032
1033 if (!write)
1034 downgrade_write(&b->lock);
1035 } else {
1036 rw_lock(write, b, level);
1037 if (PTR_HASH(c, &b->key) != PTR_HASH(c, k)) {
1038 rw_unlock(write, b);
1039 goto retry;
1040 }
1041 BUG_ON(b->level != level);
1042 }
1043
1044 if (btree_node_io_error(b)) {
1045 rw_unlock(write, b);
1046 return ERR_PTR(-EIO);
1047 }
1048
1049 BUG_ON(!b->written);
1050
1051 b->parent = parent;
1052
1053 for (; i <= b->keys.nsets && b->keys.set[i].size; i++) {
1054 prefetch(b->keys.set[i].tree);
1055 prefetch(b->keys.set[i].data);
1056 }
1057
1058 for (; i <= b->keys.nsets; i++)
1059 prefetch(b->keys.set[i].data);
1060
1061 return b;
1062 }
1063
btree_node_prefetch(struct btree * parent,struct bkey * k)1064 static void btree_node_prefetch(struct btree *parent, struct bkey *k)
1065 {
1066 struct btree *b;
1067
1068 mutex_lock(&parent->c->bucket_lock);
1069 b = mca_alloc(parent->c, NULL, k, parent->level - 1);
1070 mutex_unlock(&parent->c->bucket_lock);
1071
1072 if (!IS_ERR_OR_NULL(b)) {
1073 b->parent = parent;
1074 bch_btree_node_read(b);
1075 rw_unlock(true, b);
1076 }
1077 }
1078
1079 /* Btree alloc */
1080
btree_node_free(struct btree * b)1081 static void btree_node_free(struct btree *b)
1082 {
1083 trace_bcache_btree_node_free(b);
1084
1085 BUG_ON(b == b->c->root);
1086
1087 retry:
1088 mutex_lock(&b->write_lock);
1089 /*
1090 * If the btree node is selected and flushing in btree_flush_write(),
1091 * delay and retry until the BTREE_NODE_journal_flush bit cleared,
1092 * then it is safe to free the btree node here. Otherwise this btree
1093 * node will be in race condition.
1094 */
1095 if (btree_node_journal_flush(b)) {
1096 mutex_unlock(&b->write_lock);
1097 pr_debug("bnode %p journal_flush set, retry\n", b);
1098 udelay(1);
1099 goto retry;
1100 }
1101
1102 if (btree_node_dirty(b)) {
1103 btree_complete_write(b, btree_current_write(b));
1104 clear_bit(BTREE_NODE_dirty, &b->flags);
1105 }
1106
1107 mutex_unlock(&b->write_lock);
1108
1109 cancel_delayed_work(&b->work);
1110
1111 mutex_lock(&b->c->bucket_lock);
1112 bch_bucket_free(b->c, &b->key);
1113 mca_bucket_free(b);
1114 mutex_unlock(&b->c->bucket_lock);
1115 }
1116
1117 /*
1118 * Only error code or btree pointer will be returned, it is unncessary for
1119 * callers to check NULL pointer.
1120 */
__bch_btree_node_alloc(struct cache_set * c,struct btree_op * op,int level,bool wait,struct btree * parent)1121 struct btree *__bch_btree_node_alloc(struct cache_set *c, struct btree_op *op,
1122 int level, bool wait,
1123 struct btree *parent)
1124 {
1125 BKEY_PADDED(key) k;
1126 struct btree *b;
1127
1128 mutex_lock(&c->bucket_lock);
1129 retry:
1130 /* return ERR_PTR(-EAGAIN) when it fails */
1131 b = ERR_PTR(-EAGAIN);
1132 if (__bch_bucket_alloc_set(c, RESERVE_BTREE, &k.key, wait))
1133 goto err;
1134
1135 bkey_put(c, &k.key);
1136 SET_KEY_SIZE(&k.key, c->btree_pages * PAGE_SECTORS);
1137
1138 b = mca_alloc(c, op, &k.key, level);
1139 if (IS_ERR(b))
1140 goto err_free;
1141
1142 if (!b) {
1143 cache_bug(c,
1144 "Tried to allocate bucket that was in btree cache");
1145 goto retry;
1146 }
1147
1148 b->parent = parent;
1149 bch_bset_init_next(&b->keys, b->keys.set->data, bset_magic(&b->c->cache->sb));
1150
1151 mutex_unlock(&c->bucket_lock);
1152
1153 trace_bcache_btree_node_alloc(b);
1154 return b;
1155 err_free:
1156 bch_bucket_free(c, &k.key);
1157 err:
1158 mutex_unlock(&c->bucket_lock);
1159
1160 trace_bcache_btree_node_alloc_fail(c);
1161 return b;
1162 }
1163
bch_btree_node_alloc(struct cache_set * c,struct btree_op * op,int level,struct btree * parent)1164 static struct btree *bch_btree_node_alloc(struct cache_set *c,
1165 struct btree_op *op, int level,
1166 struct btree *parent)
1167 {
1168 return __bch_btree_node_alloc(c, op, level, op != NULL, parent);
1169 }
1170
btree_node_alloc_replacement(struct btree * b,struct btree_op * op)1171 static struct btree *btree_node_alloc_replacement(struct btree *b,
1172 struct btree_op *op)
1173 {
1174 struct btree *n = bch_btree_node_alloc(b->c, op, b->level, b->parent);
1175
1176 if (!IS_ERR(n)) {
1177 mutex_lock(&n->write_lock);
1178 bch_btree_sort_into(&b->keys, &n->keys, &b->c->sort);
1179 bkey_copy_key(&n->key, &b->key);
1180 mutex_unlock(&n->write_lock);
1181 }
1182
1183 return n;
1184 }
1185
make_btree_freeing_key(struct btree * b,struct bkey * k)1186 static void make_btree_freeing_key(struct btree *b, struct bkey *k)
1187 {
1188 unsigned int i;
1189
1190 mutex_lock(&b->c->bucket_lock);
1191
1192 atomic_inc(&b->c->prio_blocked);
1193
1194 bkey_copy(k, &b->key);
1195 bkey_copy_key(k, &ZERO_KEY);
1196
1197 for (i = 0; i < KEY_PTRS(k); i++)
1198 SET_PTR_GEN(k, i,
1199 bch_inc_gen(b->c->cache,
1200 PTR_BUCKET(b->c, &b->key, i)));
1201
1202 mutex_unlock(&b->c->bucket_lock);
1203 }
1204
btree_check_reserve(struct btree * b,struct btree_op * op)1205 static int btree_check_reserve(struct btree *b, struct btree_op *op)
1206 {
1207 struct cache_set *c = b->c;
1208 struct cache *ca = c->cache;
1209 unsigned int reserve = (c->root->level - b->level) * 2 + 1;
1210
1211 mutex_lock(&c->bucket_lock);
1212
1213 if (fifo_used(&ca->free[RESERVE_BTREE]) < reserve) {
1214 if (op)
1215 prepare_to_wait(&c->btree_cache_wait, &op->wait,
1216 TASK_UNINTERRUPTIBLE);
1217 mutex_unlock(&c->bucket_lock);
1218 return -EINTR;
1219 }
1220
1221 mutex_unlock(&c->bucket_lock);
1222
1223 return mca_cannibalize_lock(b->c, op);
1224 }
1225
1226 /* Garbage collection */
1227
__bch_btree_mark_key(struct cache_set * c,int level,struct bkey * k)1228 static uint8_t __bch_btree_mark_key(struct cache_set *c, int level,
1229 struct bkey *k)
1230 {
1231 uint8_t stale = 0;
1232 unsigned int i;
1233 struct bucket *g;
1234
1235 /*
1236 * ptr_invalid() can't return true for the keys that mark btree nodes as
1237 * freed, but since ptr_bad() returns true we'll never actually use them
1238 * for anything and thus we don't want mark their pointers here
1239 */
1240 if (!bkey_cmp(k, &ZERO_KEY))
1241 return stale;
1242
1243 for (i = 0; i < KEY_PTRS(k); i++) {
1244 if (!ptr_available(c, k, i))
1245 continue;
1246
1247 g = PTR_BUCKET(c, k, i);
1248
1249 if (gen_after(g->last_gc, PTR_GEN(k, i)))
1250 g->last_gc = PTR_GEN(k, i);
1251
1252 if (ptr_stale(c, k, i)) {
1253 stale = max(stale, ptr_stale(c, k, i));
1254 continue;
1255 }
1256
1257 cache_bug_on(GC_MARK(g) &&
1258 (GC_MARK(g) == GC_MARK_METADATA) != (level != 0),
1259 c, "inconsistent ptrs: mark = %llu, level = %i",
1260 GC_MARK(g), level);
1261
1262 if (level)
1263 SET_GC_MARK(g, GC_MARK_METADATA);
1264 else if (KEY_DIRTY(k))
1265 SET_GC_MARK(g, GC_MARK_DIRTY);
1266 else if (!GC_MARK(g))
1267 SET_GC_MARK(g, GC_MARK_RECLAIMABLE);
1268
1269 /* guard against overflow */
1270 SET_GC_SECTORS_USED(g, min_t(unsigned int,
1271 GC_SECTORS_USED(g) + KEY_SIZE(k),
1272 MAX_GC_SECTORS_USED));
1273
1274 BUG_ON(!GC_SECTORS_USED(g));
1275 }
1276
1277 return stale;
1278 }
1279
1280 #define btree_mark_key(b, k) __bch_btree_mark_key(b->c, b->level, k)
1281
bch_initial_mark_key(struct cache_set * c,int level,struct bkey * k)1282 void bch_initial_mark_key(struct cache_set *c, int level, struct bkey *k)
1283 {
1284 unsigned int i;
1285
1286 for (i = 0; i < KEY_PTRS(k); i++)
1287 if (ptr_available(c, k, i) &&
1288 !ptr_stale(c, k, i)) {
1289 struct bucket *b = PTR_BUCKET(c, k, i);
1290
1291 b->gen = PTR_GEN(k, i);
1292
1293 if (level && bkey_cmp(k, &ZERO_KEY))
1294 b->prio = BTREE_PRIO;
1295 else if (!level && b->prio == BTREE_PRIO)
1296 b->prio = INITIAL_PRIO;
1297 }
1298
1299 __bch_btree_mark_key(c, level, k);
1300 }
1301
bch_update_bucket_in_use(struct cache_set * c,struct gc_stat * stats)1302 void bch_update_bucket_in_use(struct cache_set *c, struct gc_stat *stats)
1303 {
1304 stats->in_use = (c->nbuckets - c->avail_nbuckets) * 100 / c->nbuckets;
1305 }
1306
btree_gc_mark_node(struct btree * b,struct gc_stat * gc)1307 static bool btree_gc_mark_node(struct btree *b, struct gc_stat *gc)
1308 {
1309 uint8_t stale = 0;
1310 unsigned int keys = 0, good_keys = 0;
1311 struct bkey *k;
1312 struct btree_iter_stack iter;
1313 struct bset_tree *t;
1314
1315 gc->nodes++;
1316
1317 for_each_key_filter(&b->keys, k, &iter, bch_ptr_invalid) {
1318 stale = max(stale, btree_mark_key(b, k));
1319 keys++;
1320
1321 if (bch_ptr_bad(&b->keys, k))
1322 continue;
1323
1324 gc->key_bytes += bkey_u64s(k);
1325 gc->nkeys++;
1326 good_keys++;
1327
1328 gc->data += KEY_SIZE(k);
1329 }
1330
1331 for (t = b->keys.set; t <= &b->keys.set[b->keys.nsets]; t++)
1332 btree_bug_on(t->size &&
1333 bset_written(&b->keys, t) &&
1334 bkey_cmp(&b->key, &t->end) < 0,
1335 b, "found short btree key in gc");
1336
1337 if (b->c->gc_always_rewrite)
1338 return true;
1339
1340 if (stale > 10)
1341 return true;
1342
1343 if ((keys - good_keys) * 2 > keys)
1344 return true;
1345
1346 return false;
1347 }
1348
1349 #define GC_MERGE_NODES 4U
1350
1351 struct gc_merge_info {
1352 struct btree *b;
1353 unsigned int keys;
1354 };
1355
1356 static int bch_btree_insert_node(struct btree *b, struct btree_op *op,
1357 struct keylist *insert_keys,
1358 atomic_t *journal_ref,
1359 struct bkey *replace_key);
1360
btree_gc_coalesce(struct btree * b,struct btree_op * op,struct gc_stat * gc,struct gc_merge_info * r)1361 static int btree_gc_coalesce(struct btree *b, struct btree_op *op,
1362 struct gc_stat *gc, struct gc_merge_info *r)
1363 {
1364 unsigned int i, nodes = 0, keys = 0, blocks;
1365 struct btree *new_nodes[GC_MERGE_NODES];
1366 struct keylist keylist;
1367 struct closure cl;
1368 struct bkey *k;
1369
1370 bch_keylist_init(&keylist);
1371
1372 if (btree_check_reserve(b, NULL))
1373 return 0;
1374
1375 memset(new_nodes, 0, sizeof(new_nodes));
1376 closure_init_stack(&cl);
1377
1378 while (nodes < GC_MERGE_NODES && !IS_ERR_OR_NULL(r[nodes].b))
1379 keys += r[nodes++].keys;
1380
1381 blocks = btree_default_blocks(b->c) * 2 / 3;
1382
1383 if (nodes < 2 ||
1384 __set_blocks(b->keys.set[0].data, keys,
1385 block_bytes(b->c->cache)) > blocks * (nodes - 1))
1386 return 0;
1387
1388 for (i = 0; i < nodes; i++) {
1389 new_nodes[i] = btree_node_alloc_replacement(r[i].b, NULL);
1390 if (IS_ERR(new_nodes[i]))
1391 goto out_nocoalesce;
1392 }
1393
1394 /*
1395 * We have to check the reserve here, after we've allocated our new
1396 * nodes, to make sure the insert below will succeed - we also check
1397 * before as an optimization to potentially avoid a bunch of expensive
1398 * allocs/sorts
1399 */
1400 if (btree_check_reserve(b, NULL))
1401 goto out_nocoalesce;
1402
1403 for (i = 0; i < nodes; i++)
1404 mutex_lock(&new_nodes[i]->write_lock);
1405
1406 for (i = nodes - 1; i > 0; --i) {
1407 struct bset *n1 = btree_bset_first(new_nodes[i]);
1408 struct bset *n2 = btree_bset_first(new_nodes[i - 1]);
1409 struct bkey *k, *last = NULL;
1410
1411 keys = 0;
1412
1413 if (i > 1) {
1414 for (k = n2->start;
1415 k < bset_bkey_last(n2);
1416 k = bkey_next(k)) {
1417 if (__set_blocks(n1, n1->keys + keys +
1418 bkey_u64s(k),
1419 block_bytes(b->c->cache)) > blocks)
1420 break;
1421
1422 last = k;
1423 keys += bkey_u64s(k);
1424 }
1425 } else {
1426 /*
1427 * Last node we're not getting rid of - we're getting
1428 * rid of the node at r[0]. Have to try and fit all of
1429 * the remaining keys into this node; we can't ensure
1430 * they will always fit due to rounding and variable
1431 * length keys (shouldn't be possible in practice,
1432 * though)
1433 */
1434 if (__set_blocks(n1, n1->keys + n2->keys,
1435 block_bytes(b->c->cache)) >
1436 btree_blocks(new_nodes[i]))
1437 goto out_unlock_nocoalesce;
1438
1439 keys = n2->keys;
1440 /* Take the key of the node we're getting rid of */
1441 last = &r->b->key;
1442 }
1443
1444 BUG_ON(__set_blocks(n1, n1->keys + keys, block_bytes(b->c->cache)) >
1445 btree_blocks(new_nodes[i]));
1446
1447 if (last)
1448 bkey_copy_key(&new_nodes[i]->key, last);
1449
1450 memcpy(bset_bkey_last(n1),
1451 n2->start,
1452 (void *) bset_bkey_idx(n2, keys) - (void *) n2->start);
1453
1454 n1->keys += keys;
1455 r[i].keys = n1->keys;
1456
1457 memmove(n2->start,
1458 bset_bkey_idx(n2, keys),
1459 (void *) bset_bkey_last(n2) -
1460 (void *) bset_bkey_idx(n2, keys));
1461
1462 n2->keys -= keys;
1463
1464 if (__bch_keylist_realloc(&keylist,
1465 bkey_u64s(&new_nodes[i]->key)))
1466 goto out_unlock_nocoalesce;
1467
1468 bch_btree_node_write(new_nodes[i], &cl);
1469 bch_keylist_add(&keylist, &new_nodes[i]->key);
1470 }
1471
1472 for (i = 0; i < nodes; i++)
1473 mutex_unlock(&new_nodes[i]->write_lock);
1474
1475 closure_sync(&cl);
1476
1477 /* We emptied out this node */
1478 BUG_ON(btree_bset_first(new_nodes[0])->keys);
1479 btree_node_free(new_nodes[0]);
1480 rw_unlock(true, new_nodes[0]);
1481 new_nodes[0] = NULL;
1482
1483 for (i = 0; i < nodes; i++) {
1484 if (__bch_keylist_realloc(&keylist, bkey_u64s(&r[i].b->key)))
1485 goto out_nocoalesce;
1486
1487 make_btree_freeing_key(r[i].b, keylist.top);
1488 bch_keylist_push(&keylist);
1489 }
1490
1491 bch_btree_insert_node(b, op, &keylist, NULL, NULL);
1492 BUG_ON(!bch_keylist_empty(&keylist));
1493
1494 for (i = 0; i < nodes; i++) {
1495 btree_node_free(r[i].b);
1496 rw_unlock(true, r[i].b);
1497
1498 r[i].b = new_nodes[i];
1499 }
1500
1501 memmove(r, r + 1, sizeof(r[0]) * (nodes - 1));
1502 r[nodes - 1].b = ERR_PTR(-EINTR);
1503
1504 trace_bcache_btree_gc_coalesce(nodes);
1505 gc->nodes--;
1506
1507 bch_keylist_free(&keylist);
1508
1509 /* Invalidated our iterator */
1510 return -EINTR;
1511
1512 out_unlock_nocoalesce:
1513 for (i = 0; i < nodes; i++)
1514 mutex_unlock(&new_nodes[i]->write_lock);
1515
1516 out_nocoalesce:
1517 closure_sync(&cl);
1518
1519 while ((k = bch_keylist_pop(&keylist)))
1520 if (!bkey_cmp(k, &ZERO_KEY))
1521 atomic_dec(&b->c->prio_blocked);
1522 bch_keylist_free(&keylist);
1523
1524 for (i = 0; i < nodes; i++)
1525 if (!IS_ERR_OR_NULL(new_nodes[i])) {
1526 btree_node_free(new_nodes[i]);
1527 rw_unlock(true, new_nodes[i]);
1528 }
1529 return 0;
1530 }
1531
btree_gc_rewrite_node(struct btree * b,struct btree_op * op,struct btree * replace)1532 static int btree_gc_rewrite_node(struct btree *b, struct btree_op *op,
1533 struct btree *replace)
1534 {
1535 struct keylist keys;
1536 struct btree *n;
1537
1538 if (btree_check_reserve(b, NULL))
1539 return 0;
1540
1541 n = btree_node_alloc_replacement(replace, NULL);
1542 if (IS_ERR(n))
1543 return 0;
1544
1545 /* recheck reserve after allocating replacement node */
1546 if (btree_check_reserve(b, NULL)) {
1547 btree_node_free(n);
1548 rw_unlock(true, n);
1549 return 0;
1550 }
1551
1552 bch_btree_node_write_sync(n);
1553
1554 bch_keylist_init(&keys);
1555 bch_keylist_add(&keys, &n->key);
1556
1557 make_btree_freeing_key(replace, keys.top);
1558 bch_keylist_push(&keys);
1559
1560 bch_btree_insert_node(b, op, &keys, NULL, NULL);
1561 BUG_ON(!bch_keylist_empty(&keys));
1562
1563 btree_node_free(replace);
1564 rw_unlock(true, n);
1565
1566 /* Invalidated our iterator */
1567 return -EINTR;
1568 }
1569
btree_gc_count_keys(struct btree * b)1570 static unsigned int btree_gc_count_keys(struct btree *b)
1571 {
1572 struct bkey *k;
1573 struct btree_iter_stack iter;
1574 unsigned int ret = 0;
1575
1576 for_each_key_filter(&b->keys, k, &iter, bch_ptr_bad)
1577 ret += bkey_u64s(k);
1578
1579 return ret;
1580 }
1581
btree_gc_min_nodes(struct cache_set * c)1582 static size_t btree_gc_min_nodes(struct cache_set *c)
1583 {
1584 size_t min_nodes;
1585
1586 /*
1587 * Since incremental GC would stop 100ms when front
1588 * side I/O comes, so when there are many btree nodes,
1589 * if GC only processes constant (100) nodes each time,
1590 * GC would last a long time, and the front side I/Os
1591 * would run out of the buckets (since no new bucket
1592 * can be allocated during GC), and be blocked again.
1593 * So GC should not process constant nodes, but varied
1594 * nodes according to the number of btree nodes, which
1595 * realized by dividing GC into constant(100) times,
1596 * so when there are many btree nodes, GC can process
1597 * more nodes each time, otherwise, GC will process less
1598 * nodes each time (but no less than MIN_GC_NODES)
1599 */
1600 min_nodes = c->gc_stats.nodes / MAX_GC_TIMES;
1601 if (min_nodes < MIN_GC_NODES)
1602 min_nodes = MIN_GC_NODES;
1603
1604 return min_nodes;
1605 }
1606
1607
btree_gc_recurse(struct btree * b,struct btree_op * op,struct closure * writes,struct gc_stat * gc)1608 static int btree_gc_recurse(struct btree *b, struct btree_op *op,
1609 struct closure *writes, struct gc_stat *gc)
1610 {
1611 int ret = 0;
1612 bool should_rewrite;
1613 struct bkey *k;
1614 struct btree_iter_stack iter;
1615 struct gc_merge_info r[GC_MERGE_NODES];
1616 struct gc_merge_info *i, *last = r + ARRAY_SIZE(r) - 1;
1617
1618 bch_btree_iter_stack_init(&b->keys, &iter, &b->c->gc_done);
1619
1620 for (i = r; i < r + ARRAY_SIZE(r); i++)
1621 i->b = ERR_PTR(-EINTR);
1622
1623 while (1) {
1624 k = bch_btree_iter_next_filter(&iter.iter, &b->keys,
1625 bch_ptr_bad);
1626 if (k) {
1627 r->b = bch_btree_node_get(b->c, op, k, b->level - 1,
1628 true, b);
1629 if (IS_ERR(r->b)) {
1630 ret = PTR_ERR(r->b);
1631 break;
1632 }
1633
1634 r->keys = btree_gc_count_keys(r->b);
1635
1636 ret = btree_gc_coalesce(b, op, gc, r);
1637 if (ret)
1638 break;
1639 }
1640
1641 if (!last->b)
1642 break;
1643
1644 if (!IS_ERR(last->b)) {
1645 should_rewrite = btree_gc_mark_node(last->b, gc);
1646 if (should_rewrite) {
1647 ret = btree_gc_rewrite_node(b, op, last->b);
1648 if (ret)
1649 break;
1650 }
1651
1652 if (last->b->level) {
1653 ret = btree_gc_recurse(last->b, op, writes, gc);
1654 if (ret)
1655 break;
1656 }
1657
1658 bkey_copy_key(&b->c->gc_done, &last->b->key);
1659
1660 /*
1661 * Must flush leaf nodes before gc ends, since replace
1662 * operations aren't journalled
1663 */
1664 mutex_lock(&last->b->write_lock);
1665 if (btree_node_dirty(last->b))
1666 bch_btree_node_write(last->b, writes);
1667 mutex_unlock(&last->b->write_lock);
1668 rw_unlock(true, last->b);
1669 }
1670
1671 memmove(r + 1, r, sizeof(r[0]) * (GC_MERGE_NODES - 1));
1672 r->b = NULL;
1673
1674 if (atomic_read(&b->c->search_inflight) &&
1675 gc->nodes >= gc->nodes_pre + btree_gc_min_nodes(b->c)) {
1676 gc->nodes_pre = gc->nodes;
1677 ret = -EAGAIN;
1678 break;
1679 }
1680
1681 if (need_resched()) {
1682 ret = -EAGAIN;
1683 break;
1684 }
1685 }
1686
1687 for (i = r; i < r + ARRAY_SIZE(r); i++)
1688 if (!IS_ERR_OR_NULL(i->b)) {
1689 mutex_lock(&i->b->write_lock);
1690 if (btree_node_dirty(i->b))
1691 bch_btree_node_write(i->b, writes);
1692 mutex_unlock(&i->b->write_lock);
1693 rw_unlock(true, i->b);
1694 }
1695
1696 return ret;
1697 }
1698
bch_btree_gc_root(struct btree * b,struct btree_op * op,struct closure * writes,struct gc_stat * gc)1699 static int bch_btree_gc_root(struct btree *b, struct btree_op *op,
1700 struct closure *writes, struct gc_stat *gc)
1701 {
1702 struct btree *n = NULL;
1703 int ret = 0;
1704 bool should_rewrite;
1705
1706 should_rewrite = btree_gc_mark_node(b, gc);
1707 if (should_rewrite) {
1708 n = btree_node_alloc_replacement(b, NULL);
1709
1710 if (!IS_ERR(n)) {
1711 bch_btree_node_write_sync(n);
1712
1713 bch_btree_set_root(n);
1714 btree_node_free(b);
1715 rw_unlock(true, n);
1716
1717 return -EINTR;
1718 }
1719 }
1720
1721 __bch_btree_mark_key(b->c, b->level + 1, &b->key);
1722
1723 if (b->level) {
1724 ret = btree_gc_recurse(b, op, writes, gc);
1725 if (ret)
1726 return ret;
1727 }
1728
1729 bkey_copy_key(&b->c->gc_done, &b->key);
1730
1731 return ret;
1732 }
1733
btree_gc_start(struct cache_set * c)1734 static void btree_gc_start(struct cache_set *c)
1735 {
1736 struct cache *ca;
1737 struct bucket *b;
1738
1739 if (!c->gc_mark_valid)
1740 return;
1741
1742 mutex_lock(&c->bucket_lock);
1743
1744 c->gc_done = ZERO_KEY;
1745
1746 ca = c->cache;
1747 for_each_bucket(b, ca) {
1748 b->last_gc = b->gen;
1749 if (bch_can_invalidate_bucket(ca, b))
1750 b->reclaimable_in_gc = 1;
1751 if (!atomic_read(&b->pin)) {
1752 SET_GC_MARK(b, 0);
1753 SET_GC_SECTORS_USED(b, 0);
1754 }
1755 }
1756
1757 c->gc_mark_valid = 0;
1758 mutex_unlock(&c->bucket_lock);
1759 }
1760
bch_btree_gc_finish(struct cache_set * c)1761 static void bch_btree_gc_finish(struct cache_set *c)
1762 {
1763 struct bucket *b;
1764 struct cache *ca;
1765 unsigned int i, j;
1766 uint64_t *k;
1767
1768 mutex_lock(&c->bucket_lock);
1769
1770 set_gc_sectors(c);
1771 c->gc_mark_valid = 1;
1772 c->need_gc = 0;
1773
1774 for (i = 0; i < KEY_PTRS(&c->uuid_bucket); i++)
1775 SET_GC_MARK(PTR_BUCKET(c, &c->uuid_bucket, i),
1776 GC_MARK_METADATA);
1777
1778 /* don't reclaim buckets to which writeback keys point */
1779 rcu_read_lock();
1780 for (i = 0; i < c->devices_max_used; i++) {
1781 struct bcache_device *d = c->devices[i];
1782 struct cached_dev *dc;
1783 struct keybuf_key *w, *n;
1784
1785 if (!d || UUID_FLASH_ONLY(&c->uuids[i]))
1786 continue;
1787 dc = container_of(d, struct cached_dev, disk);
1788
1789 spin_lock(&dc->writeback_keys.lock);
1790 rbtree_postorder_for_each_entry_safe(w, n,
1791 &dc->writeback_keys.keys, node)
1792 for (j = 0; j < KEY_PTRS(&w->key); j++)
1793 SET_GC_MARK(PTR_BUCKET(c, &w->key, j),
1794 GC_MARK_DIRTY);
1795 spin_unlock(&dc->writeback_keys.lock);
1796 }
1797 rcu_read_unlock();
1798
1799 c->avail_nbuckets = 0;
1800
1801 ca = c->cache;
1802 ca->invalidate_needs_gc = 0;
1803
1804 for (k = ca->sb.d; k < ca->sb.d + ca->sb.keys; k++)
1805 SET_GC_MARK(ca->buckets + *k, GC_MARK_METADATA);
1806
1807 for (k = ca->prio_buckets;
1808 k < ca->prio_buckets + prio_buckets(ca) * 2; k++)
1809 SET_GC_MARK(ca->buckets + *k, GC_MARK_METADATA);
1810
1811 for_each_bucket(b, ca) {
1812 c->need_gc = max(c->need_gc, bucket_gc_gen(b));
1813
1814 if (b->reclaimable_in_gc)
1815 b->reclaimable_in_gc = 0;
1816
1817 if (atomic_read(&b->pin))
1818 continue;
1819
1820 BUG_ON(!GC_MARK(b) && GC_SECTORS_USED(b));
1821
1822 if (!GC_MARK(b) || GC_MARK(b) == GC_MARK_RECLAIMABLE)
1823 c->avail_nbuckets++;
1824 }
1825
1826 mutex_unlock(&c->bucket_lock);
1827 }
1828
bch_btree_gc(struct cache_set * c)1829 static void bch_btree_gc(struct cache_set *c)
1830 {
1831 int ret;
1832 struct gc_stat stats;
1833 struct closure writes;
1834 struct btree_op op;
1835 uint64_t start_time = local_clock();
1836
1837 trace_bcache_gc_start(c);
1838
1839 memset(&stats, 0, sizeof(struct gc_stat));
1840 closure_init_stack(&writes);
1841 bch_btree_op_init(&op, SHRT_MAX);
1842
1843 btree_gc_start(c);
1844
1845 /* if CACHE_SET_IO_DISABLE set, gc thread should stop too */
1846 do {
1847 ret = bcache_btree_root(gc_root, c, &op, &writes, &stats);
1848 closure_sync(&writes);
1849 cond_resched();
1850
1851 if (ret == -EAGAIN)
1852 schedule_timeout_interruptible(msecs_to_jiffies
1853 (GC_SLEEP_MS));
1854 else if (ret)
1855 pr_warn("gc failed!\n");
1856 } while (ret && !test_bit(CACHE_SET_IO_DISABLE, &c->flags));
1857
1858 bch_btree_gc_finish(c);
1859 wake_up_allocators(c);
1860
1861 bch_time_stats_update(&c->btree_gc_time, start_time);
1862
1863 stats.key_bytes *= sizeof(uint64_t);
1864 stats.data <<= 9;
1865 bch_update_bucket_in_use(c, &stats);
1866 memcpy(&c->gc_stats, &stats, sizeof(struct gc_stat));
1867
1868 trace_bcache_gc_end(c);
1869
1870 bch_moving_gc(c);
1871 }
1872
gc_should_run(struct cache_set * c)1873 static bool gc_should_run(struct cache_set *c)
1874 {
1875 struct cache *ca = c->cache;
1876
1877 if (ca->invalidate_needs_gc)
1878 return true;
1879
1880 if (atomic_read(&c->sectors_to_gc) < 0)
1881 return true;
1882
1883 return false;
1884 }
1885
bch_gc_thread(void * arg)1886 static int bch_gc_thread(void *arg)
1887 {
1888 struct cache_set *c = arg;
1889
1890 while (1) {
1891 wait_event_interruptible(c->gc_wait,
1892 kthread_should_stop() ||
1893 test_bit(CACHE_SET_IO_DISABLE, &c->flags) ||
1894 gc_should_run(c));
1895
1896 if (kthread_should_stop() ||
1897 test_bit(CACHE_SET_IO_DISABLE, &c->flags))
1898 break;
1899
1900 set_gc_sectors(c);
1901 bch_btree_gc(c);
1902 }
1903
1904 wait_for_kthread_stop();
1905 return 0;
1906 }
1907
bch_gc_thread_start(struct cache_set * c)1908 int bch_gc_thread_start(struct cache_set *c)
1909 {
1910 c->gc_thread = kthread_run(bch_gc_thread, c, "bcache_gc");
1911 return PTR_ERR_OR_ZERO(c->gc_thread);
1912 }
1913
1914 /* Initial partial gc */
1915
bch_btree_check_recurse(struct btree * b,struct btree_op * op)1916 static int bch_btree_check_recurse(struct btree *b, struct btree_op *op)
1917 {
1918 int ret = 0;
1919 struct bkey *k, *p = NULL;
1920 struct btree_iter_stack iter;
1921
1922 for_each_key_filter(&b->keys, k, &iter, bch_ptr_invalid)
1923 bch_initial_mark_key(b->c, b->level, k);
1924
1925 bch_initial_mark_key(b->c, b->level + 1, &b->key);
1926
1927 if (b->level) {
1928 bch_btree_iter_stack_init(&b->keys, &iter, NULL);
1929
1930 do {
1931 k = bch_btree_iter_next_filter(&iter.iter, &b->keys,
1932 bch_ptr_bad);
1933 if (k) {
1934 btree_node_prefetch(b, k);
1935 /*
1936 * initiallize c->gc_stats.nodes
1937 * for incremental GC
1938 */
1939 b->c->gc_stats.nodes++;
1940 }
1941
1942 if (p)
1943 ret = bcache_btree(check_recurse, p, b, op);
1944
1945 p = k;
1946 } while (p && !ret);
1947 }
1948
1949 return ret;
1950 }
1951
1952
bch_btree_check_thread(void * arg)1953 static int bch_btree_check_thread(void *arg)
1954 {
1955 int ret;
1956 struct btree_check_info *info = arg;
1957 struct btree_check_state *check_state = info->state;
1958 struct cache_set *c = check_state->c;
1959 struct btree_iter_stack iter;
1960 struct bkey *k, *p;
1961 int cur_idx, prev_idx, skip_nr;
1962
1963 k = p = NULL;
1964 cur_idx = prev_idx = 0;
1965 ret = 0;
1966
1967 /* root node keys are checked before thread created */
1968 bch_btree_iter_stack_init(&c->root->keys, &iter, NULL);
1969 k = bch_btree_iter_next_filter(&iter.iter, &c->root->keys, bch_ptr_bad);
1970 BUG_ON(!k);
1971
1972 p = k;
1973 while (k) {
1974 /*
1975 * Fetch a root node key index, skip the keys which
1976 * should be fetched by other threads, then check the
1977 * sub-tree indexed by the fetched key.
1978 */
1979 spin_lock(&check_state->idx_lock);
1980 cur_idx = check_state->key_idx;
1981 check_state->key_idx++;
1982 spin_unlock(&check_state->idx_lock);
1983
1984 skip_nr = cur_idx - prev_idx;
1985
1986 while (skip_nr) {
1987 k = bch_btree_iter_next_filter(&iter.iter,
1988 &c->root->keys,
1989 bch_ptr_bad);
1990 if (k)
1991 p = k;
1992 else {
1993 /*
1994 * No more keys to check in root node,
1995 * current checking threads are enough,
1996 * stop creating more.
1997 */
1998 atomic_set(&check_state->enough, 1);
1999 /* Update check_state->enough earlier */
2000 smp_mb__after_atomic();
2001 goto out;
2002 }
2003 skip_nr--;
2004 cond_resched();
2005 }
2006
2007 if (p) {
2008 struct btree_op op;
2009
2010 btree_node_prefetch(c->root, p);
2011 c->gc_stats.nodes++;
2012 bch_btree_op_init(&op, 0);
2013 ret = bcache_btree(check_recurse, p, c->root, &op);
2014 /*
2015 * The op may be added to cache_set's btree_cache_wait
2016 * in mca_cannibalize(), must ensure it is removed from
2017 * the list and release btree_cache_alloc_lock before
2018 * free op memory.
2019 * Otherwise, the btree_cache_wait will be damaged.
2020 */
2021 bch_cannibalize_unlock(c);
2022 finish_wait(&c->btree_cache_wait, &(&op)->wait);
2023 if (ret)
2024 goto out;
2025 }
2026 p = NULL;
2027 prev_idx = cur_idx;
2028 cond_resched();
2029 }
2030
2031 out:
2032 info->result = ret;
2033 /* update check_state->started among all CPUs */
2034 smp_mb__before_atomic();
2035 if (atomic_dec_and_test(&check_state->started))
2036 wake_up(&check_state->wait);
2037
2038 return ret;
2039 }
2040
2041
2042
bch_btree_chkthread_nr(void)2043 static int bch_btree_chkthread_nr(void)
2044 {
2045 int n = num_online_cpus()/2;
2046
2047 if (n == 0)
2048 n = 1;
2049 else if (n > BCH_BTR_CHKTHREAD_MAX)
2050 n = BCH_BTR_CHKTHREAD_MAX;
2051
2052 return n;
2053 }
2054
bch_btree_check(struct cache_set * c)2055 int bch_btree_check(struct cache_set *c)
2056 {
2057 int ret = 0;
2058 int i;
2059 struct bkey *k = NULL;
2060 struct btree_iter_stack iter;
2061 struct btree_check_state check_state;
2062
2063 /* check and mark root node keys */
2064 for_each_key_filter(&c->root->keys, k, &iter, bch_ptr_invalid)
2065 bch_initial_mark_key(c, c->root->level, k);
2066
2067 bch_initial_mark_key(c, c->root->level + 1, &c->root->key);
2068
2069 if (c->root->level == 0)
2070 return 0;
2071
2072 memset(&check_state, 0, sizeof(struct btree_check_state));
2073 check_state.c = c;
2074 check_state.total_threads = bch_btree_chkthread_nr();
2075 check_state.key_idx = 0;
2076 spin_lock_init(&check_state.idx_lock);
2077 atomic_set(&check_state.started, 0);
2078 atomic_set(&check_state.enough, 0);
2079 init_waitqueue_head(&check_state.wait);
2080
2081 rw_lock(0, c->root, c->root->level);
2082 /*
2083 * Run multiple threads to check btree nodes in parallel,
2084 * if check_state.enough is non-zero, it means current
2085 * running check threads are enough, unncessary to create
2086 * more.
2087 */
2088 for (i = 0; i < check_state.total_threads; i++) {
2089 /* fetch latest check_state.enough earlier */
2090 smp_mb__before_atomic();
2091 if (atomic_read(&check_state.enough))
2092 break;
2093
2094 check_state.infos[i].result = 0;
2095 check_state.infos[i].state = &check_state;
2096
2097 check_state.infos[i].thread =
2098 kthread_run(bch_btree_check_thread,
2099 &check_state.infos[i],
2100 "bch_btrchk[%d]", i);
2101 if (IS_ERR(check_state.infos[i].thread)) {
2102 pr_err("fails to run thread bch_btrchk[%d]\n", i);
2103 for (--i; i >= 0; i--)
2104 kthread_stop(check_state.infos[i].thread);
2105 ret = -ENOMEM;
2106 goto out;
2107 }
2108 atomic_inc(&check_state.started);
2109 }
2110
2111 /*
2112 * Must wait for all threads to stop.
2113 */
2114 wait_event(check_state.wait, atomic_read(&check_state.started) == 0);
2115
2116 for (i = 0; i < check_state.total_threads; i++) {
2117 if (check_state.infos[i].result) {
2118 ret = check_state.infos[i].result;
2119 goto out;
2120 }
2121 }
2122
2123 out:
2124 rw_unlock(0, c->root);
2125 return ret;
2126 }
2127
bch_initial_gc_finish(struct cache_set * c)2128 void bch_initial_gc_finish(struct cache_set *c)
2129 {
2130 struct cache *ca = c->cache;
2131 struct bucket *b;
2132
2133 bch_btree_gc_finish(c);
2134
2135 mutex_lock(&c->bucket_lock);
2136
2137 /*
2138 * We need to put some unused buckets directly on the prio freelist in
2139 * order to get the allocator thread started - it needs freed buckets in
2140 * order to rewrite the prios and gens, and it needs to rewrite prios
2141 * and gens in order to free buckets.
2142 *
2143 * This is only safe for buckets that have no live data in them, which
2144 * there should always be some of.
2145 */
2146 for_each_bucket(b, ca) {
2147 if (fifo_full(&ca->free[RESERVE_PRIO]) &&
2148 fifo_full(&ca->free[RESERVE_BTREE]))
2149 break;
2150
2151 if (bch_can_invalidate_bucket(ca, b) &&
2152 !GC_MARK(b)) {
2153 __bch_invalidate_one_bucket(ca, b);
2154 if (!fifo_push(&ca->free[RESERVE_PRIO],
2155 b - ca->buckets))
2156 fifo_push(&ca->free[RESERVE_BTREE],
2157 b - ca->buckets);
2158 }
2159 }
2160
2161 mutex_unlock(&c->bucket_lock);
2162 }
2163
2164 /* Btree insertion */
2165
btree_insert_key(struct btree * b,struct bkey * k,struct bkey * replace_key)2166 static bool btree_insert_key(struct btree *b, struct bkey *k,
2167 struct bkey *replace_key)
2168 {
2169 unsigned int status;
2170
2171 BUG_ON(bkey_cmp(k, &b->key) > 0);
2172
2173 status = bch_btree_insert_key(&b->keys, k, replace_key);
2174 if (status != BTREE_INSERT_STATUS_NO_INSERT) {
2175 bch_check_keys(&b->keys, "%u for %s", status,
2176 replace_key ? "replace" : "insert");
2177
2178 trace_bcache_btree_insert_key(b, k, replace_key != NULL,
2179 status);
2180 return true;
2181 } else
2182 return false;
2183 }
2184
insert_u64s_remaining(struct btree * b)2185 static size_t insert_u64s_remaining(struct btree *b)
2186 {
2187 long ret = bch_btree_keys_u64s_remaining(&b->keys);
2188
2189 /*
2190 * Might land in the middle of an existing extent and have to split it
2191 */
2192 if (b->keys.ops->is_extents)
2193 ret -= KEY_MAX_U64S;
2194
2195 return max(ret, 0L);
2196 }
2197
bch_btree_insert_keys(struct btree * b,struct btree_op * op,struct keylist * insert_keys,struct bkey * replace_key)2198 static bool bch_btree_insert_keys(struct btree *b, struct btree_op *op,
2199 struct keylist *insert_keys,
2200 struct bkey *replace_key)
2201 {
2202 bool ret = false;
2203 int oldsize = bch_count_data(&b->keys);
2204
2205 while (!bch_keylist_empty(insert_keys)) {
2206 struct bkey *k = insert_keys->keys;
2207
2208 if (bkey_u64s(k) > insert_u64s_remaining(b))
2209 break;
2210
2211 if (bkey_cmp(k, &b->key) <= 0) {
2212 if (!b->level)
2213 bkey_put(b->c, k);
2214
2215 ret |= btree_insert_key(b, k, replace_key);
2216 bch_keylist_pop_front(insert_keys);
2217 } else if (bkey_cmp(&START_KEY(k), &b->key) < 0) {
2218 BKEY_PADDED(key) temp;
2219 bkey_copy(&temp.key, insert_keys->keys);
2220
2221 bch_cut_back(&b->key, &temp.key);
2222 bch_cut_front(&b->key, insert_keys->keys);
2223
2224 ret |= btree_insert_key(b, &temp.key, replace_key);
2225 break;
2226 } else {
2227 break;
2228 }
2229 }
2230
2231 if (!ret)
2232 op->insert_collision = true;
2233
2234 BUG_ON(!bch_keylist_empty(insert_keys) && b->level);
2235
2236 BUG_ON(bch_count_data(&b->keys) < oldsize);
2237 return ret;
2238 }
2239
btree_split(struct btree * b,struct btree_op * op,struct keylist * insert_keys,struct bkey * replace_key)2240 static int btree_split(struct btree *b, struct btree_op *op,
2241 struct keylist *insert_keys,
2242 struct bkey *replace_key)
2243 {
2244 bool split;
2245 struct btree *n1, *n2 = NULL, *n3 = NULL;
2246 uint64_t start_time = local_clock();
2247 struct closure cl;
2248 struct keylist parent_keys;
2249
2250 closure_init_stack(&cl);
2251 bch_keylist_init(&parent_keys);
2252
2253 if (btree_check_reserve(b, op)) {
2254 if (!b->level)
2255 return -EINTR;
2256 else
2257 WARN(1, "insufficient reserve for split\n");
2258 }
2259
2260 n1 = btree_node_alloc_replacement(b, op);
2261 if (IS_ERR(n1))
2262 goto err;
2263
2264 split = set_blocks(btree_bset_first(n1),
2265 block_bytes(n1->c->cache)) > (btree_blocks(b) * 4) / 5;
2266
2267 if (split) {
2268 unsigned int keys = 0;
2269
2270 trace_bcache_btree_node_split(b, btree_bset_first(n1)->keys);
2271
2272 n2 = bch_btree_node_alloc(b->c, op, b->level, b->parent);
2273 if (IS_ERR(n2))
2274 goto err_free1;
2275
2276 if (!b->parent) {
2277 n3 = bch_btree_node_alloc(b->c, op, b->level + 1, NULL);
2278 if (IS_ERR(n3))
2279 goto err_free2;
2280 }
2281
2282 mutex_lock(&n1->write_lock);
2283 mutex_lock(&n2->write_lock);
2284
2285 bch_btree_insert_keys(n1, op, insert_keys, replace_key);
2286
2287 /*
2288 * Has to be a linear search because we don't have an auxiliary
2289 * search tree yet
2290 */
2291
2292 while (keys < (btree_bset_first(n1)->keys * 3) / 5)
2293 keys += bkey_u64s(bset_bkey_idx(btree_bset_first(n1),
2294 keys));
2295
2296 bkey_copy_key(&n1->key,
2297 bset_bkey_idx(btree_bset_first(n1), keys));
2298 keys += bkey_u64s(bset_bkey_idx(btree_bset_first(n1), keys));
2299
2300 btree_bset_first(n2)->keys = btree_bset_first(n1)->keys - keys;
2301 btree_bset_first(n1)->keys = keys;
2302
2303 memcpy(btree_bset_first(n2)->start,
2304 bset_bkey_last(btree_bset_first(n1)),
2305 btree_bset_first(n2)->keys * sizeof(uint64_t));
2306
2307 bkey_copy_key(&n2->key, &b->key);
2308
2309 bch_keylist_add(&parent_keys, &n2->key);
2310 bch_btree_node_write(n2, &cl);
2311 mutex_unlock(&n2->write_lock);
2312 rw_unlock(true, n2);
2313 } else {
2314 trace_bcache_btree_node_compact(b, btree_bset_first(n1)->keys);
2315
2316 mutex_lock(&n1->write_lock);
2317 bch_btree_insert_keys(n1, op, insert_keys, replace_key);
2318 }
2319
2320 bch_keylist_add(&parent_keys, &n1->key);
2321 bch_btree_node_write(n1, &cl);
2322 mutex_unlock(&n1->write_lock);
2323
2324 if (n3) {
2325 /* Depth increases, make a new root */
2326 mutex_lock(&n3->write_lock);
2327 bkey_copy_key(&n3->key, &MAX_KEY);
2328 bch_btree_insert_keys(n3, op, &parent_keys, NULL);
2329 bch_btree_node_write(n3, &cl);
2330 mutex_unlock(&n3->write_lock);
2331
2332 closure_sync(&cl);
2333 bch_btree_set_root(n3);
2334 rw_unlock(true, n3);
2335 } else if (!b->parent) {
2336 /* Root filled up but didn't need to be split */
2337 closure_sync(&cl);
2338 bch_btree_set_root(n1);
2339 } else {
2340 /* Split a non root node */
2341 closure_sync(&cl);
2342 make_btree_freeing_key(b, parent_keys.top);
2343 bch_keylist_push(&parent_keys);
2344
2345 bch_btree_insert_node(b->parent, op, &parent_keys, NULL, NULL);
2346 BUG_ON(!bch_keylist_empty(&parent_keys));
2347 }
2348
2349 btree_node_free(b);
2350 rw_unlock(true, n1);
2351
2352 bch_time_stats_update(&b->c->btree_split_time, start_time);
2353
2354 return 0;
2355 err_free2:
2356 bkey_put(b->c, &n2->key);
2357 btree_node_free(n2);
2358 rw_unlock(true, n2);
2359 err_free1:
2360 bkey_put(b->c, &n1->key);
2361 btree_node_free(n1);
2362 rw_unlock(true, n1);
2363 err:
2364 WARN(1, "bcache: btree split failed (level %u)", b->level);
2365
2366 if (n3 == ERR_PTR(-EAGAIN) ||
2367 n2 == ERR_PTR(-EAGAIN) ||
2368 n1 == ERR_PTR(-EAGAIN))
2369 return -EAGAIN;
2370
2371 return -ENOMEM;
2372 }
2373
bch_btree_insert_node(struct btree * b,struct btree_op * op,struct keylist * insert_keys,atomic_t * journal_ref,struct bkey * replace_key)2374 static int bch_btree_insert_node(struct btree *b, struct btree_op *op,
2375 struct keylist *insert_keys,
2376 atomic_t *journal_ref,
2377 struct bkey *replace_key)
2378 {
2379 struct closure cl;
2380
2381 BUG_ON(b->level && replace_key);
2382
2383 closure_init_stack(&cl);
2384
2385 mutex_lock(&b->write_lock);
2386
2387 if (write_block(b) != btree_bset_last(b) &&
2388 b->keys.last_set_unwritten)
2389 bch_btree_init_next(b); /* just wrote a set */
2390
2391 if (bch_keylist_nkeys(insert_keys) > insert_u64s_remaining(b)) {
2392 mutex_unlock(&b->write_lock);
2393 goto split;
2394 }
2395
2396 BUG_ON(write_block(b) != btree_bset_last(b));
2397
2398 if (bch_btree_insert_keys(b, op, insert_keys, replace_key)) {
2399 if (!b->level)
2400 bch_btree_leaf_dirty(b, journal_ref);
2401 else
2402 bch_btree_node_write(b, &cl);
2403 }
2404
2405 mutex_unlock(&b->write_lock);
2406
2407 /* wait for btree node write if necessary, after unlock */
2408 closure_sync(&cl);
2409
2410 return 0;
2411 split:
2412 if (current->bio_list) {
2413 op->lock = b->c->root->level + 1;
2414 return -EAGAIN;
2415 } else if (op->lock <= b->c->root->level) {
2416 op->lock = b->c->root->level + 1;
2417 return -EINTR;
2418 } else {
2419 /* Invalidated all iterators */
2420 int ret = btree_split(b, op, insert_keys, replace_key);
2421
2422 if (bch_keylist_empty(insert_keys))
2423 return 0;
2424 else if (!ret)
2425 return -EINTR;
2426 return ret;
2427 }
2428 }
2429
bch_btree_insert_check_key(struct btree * b,struct btree_op * op,struct bkey * check_key)2430 int bch_btree_insert_check_key(struct btree *b, struct btree_op *op,
2431 struct bkey *check_key)
2432 {
2433 int ret = -EINTR;
2434 uint64_t btree_ptr = b->key.ptr[0];
2435 unsigned long seq = b->seq;
2436 struct keylist insert;
2437 bool upgrade = op->lock == -1;
2438
2439 bch_keylist_init(&insert);
2440
2441 if (upgrade) {
2442 rw_unlock(false, b);
2443 rw_lock(true, b, b->level);
2444
2445 if (b->key.ptr[0] != btree_ptr ||
2446 b->seq != seq + 1) {
2447 op->lock = b->level;
2448 goto out;
2449 }
2450 }
2451
2452 SET_KEY_PTRS(check_key, 1);
2453 get_random_bytes(&check_key->ptr[0], sizeof(uint64_t));
2454
2455 SET_PTR_DEV(check_key, 0, PTR_CHECK_DEV);
2456
2457 bch_keylist_add(&insert, check_key);
2458
2459 ret = bch_btree_insert_node(b, op, &insert, NULL, NULL);
2460
2461 BUG_ON(!ret && !bch_keylist_empty(&insert));
2462 out:
2463 if (upgrade)
2464 downgrade_write(&b->lock);
2465 return ret;
2466 }
2467
2468 struct btree_insert_op {
2469 struct btree_op op;
2470 struct keylist *keys;
2471 atomic_t *journal_ref;
2472 struct bkey *replace_key;
2473 };
2474
btree_insert_fn(struct btree_op * b_op,struct btree * b)2475 static int btree_insert_fn(struct btree_op *b_op, struct btree *b)
2476 {
2477 struct btree_insert_op *op = container_of(b_op,
2478 struct btree_insert_op, op);
2479
2480 int ret = bch_btree_insert_node(b, &op->op, op->keys,
2481 op->journal_ref, op->replace_key);
2482 if (ret && !bch_keylist_empty(op->keys))
2483 return ret;
2484 else
2485 return MAP_DONE;
2486 }
2487
bch_btree_insert(struct cache_set * c,struct keylist * keys,atomic_t * journal_ref,struct bkey * replace_key)2488 int bch_btree_insert(struct cache_set *c, struct keylist *keys,
2489 atomic_t *journal_ref, struct bkey *replace_key)
2490 {
2491 struct btree_insert_op op;
2492 int ret = 0;
2493
2494 BUG_ON(current->bio_list);
2495 BUG_ON(bch_keylist_empty(keys));
2496
2497 bch_btree_op_init(&op.op, 0);
2498 op.keys = keys;
2499 op.journal_ref = journal_ref;
2500 op.replace_key = replace_key;
2501
2502 while (!ret && !bch_keylist_empty(keys)) {
2503 op.op.lock = 0;
2504 ret = bch_btree_map_leaf_nodes(&op.op, c,
2505 &START_KEY(keys->keys),
2506 btree_insert_fn);
2507 }
2508
2509 if (ret) {
2510 struct bkey *k;
2511
2512 pr_err("error %i\n", ret);
2513
2514 while ((k = bch_keylist_pop(keys)))
2515 bkey_put(c, k);
2516 } else if (op.op.insert_collision)
2517 ret = -ESRCH;
2518
2519 return ret;
2520 }
2521
bch_btree_set_root(struct btree * b)2522 void bch_btree_set_root(struct btree *b)
2523 {
2524 unsigned int i;
2525 struct closure cl;
2526
2527 closure_init_stack(&cl);
2528
2529 trace_bcache_btree_set_root(b);
2530
2531 BUG_ON(!b->written);
2532
2533 for (i = 0; i < KEY_PTRS(&b->key); i++)
2534 BUG_ON(PTR_BUCKET(b->c, &b->key, i)->prio != BTREE_PRIO);
2535
2536 mutex_lock(&b->c->bucket_lock);
2537 list_del_init(&b->list);
2538 mutex_unlock(&b->c->bucket_lock);
2539
2540 b->c->root = b;
2541
2542 bch_journal_meta(b->c, &cl);
2543 closure_sync(&cl);
2544 }
2545
2546 /* Map across nodes or keys */
2547
bch_btree_map_nodes_recurse(struct btree * b,struct btree_op * op,struct bkey * from,btree_map_nodes_fn * fn,int flags)2548 static int bch_btree_map_nodes_recurse(struct btree *b, struct btree_op *op,
2549 struct bkey *from,
2550 btree_map_nodes_fn *fn, int flags)
2551 {
2552 int ret = MAP_CONTINUE;
2553
2554 if (b->level) {
2555 struct bkey *k;
2556 struct btree_iter_stack iter;
2557
2558 bch_btree_iter_stack_init(&b->keys, &iter, from);
2559
2560 while ((k = bch_btree_iter_next_filter(&iter.iter, &b->keys,
2561 bch_ptr_bad))) {
2562 ret = bcache_btree(map_nodes_recurse, k, b,
2563 op, from, fn, flags);
2564 from = NULL;
2565
2566 if (ret != MAP_CONTINUE)
2567 return ret;
2568 }
2569 }
2570
2571 if (!b->level || flags == MAP_ALL_NODES)
2572 ret = fn(op, b);
2573
2574 return ret;
2575 }
2576
__bch_btree_map_nodes(struct btree_op * op,struct cache_set * c,struct bkey * from,btree_map_nodes_fn * fn,int flags)2577 int __bch_btree_map_nodes(struct btree_op *op, struct cache_set *c,
2578 struct bkey *from, btree_map_nodes_fn *fn, int flags)
2579 {
2580 return bcache_btree_root(map_nodes_recurse, c, op, from, fn, flags);
2581 }
2582
bch_btree_map_keys_recurse(struct btree * b,struct btree_op * op,struct bkey * from,btree_map_keys_fn * fn,int flags)2583 int bch_btree_map_keys_recurse(struct btree *b, struct btree_op *op,
2584 struct bkey *from, btree_map_keys_fn *fn,
2585 int flags)
2586 {
2587 int ret = MAP_CONTINUE;
2588 struct bkey *k;
2589 struct btree_iter_stack iter;
2590
2591 bch_btree_iter_stack_init(&b->keys, &iter, from);
2592
2593 while ((k = bch_btree_iter_next_filter(&iter.iter, &b->keys,
2594 bch_ptr_bad))) {
2595 ret = !b->level
2596 ? fn(op, b, k)
2597 : bcache_btree(map_keys_recurse, k,
2598 b, op, from, fn, flags);
2599 from = NULL;
2600
2601 if (ret != MAP_CONTINUE)
2602 return ret;
2603 }
2604
2605 if (!b->level && (flags & MAP_END_KEY))
2606 ret = fn(op, b, &KEY(KEY_INODE(&b->key),
2607 KEY_OFFSET(&b->key), 0));
2608
2609 return ret;
2610 }
2611
bch_btree_map_keys(struct btree_op * op,struct cache_set * c,struct bkey * from,btree_map_keys_fn * fn,int flags)2612 int bch_btree_map_keys(struct btree_op *op, struct cache_set *c,
2613 struct bkey *from, btree_map_keys_fn *fn, int flags)
2614 {
2615 return bcache_btree_root(map_keys_recurse, c, op, from, fn, flags);
2616 }
2617
2618 /* Keybuf code */
2619
keybuf_cmp(struct keybuf_key * l,struct keybuf_key * r)2620 static inline int keybuf_cmp(struct keybuf_key *l, struct keybuf_key *r)
2621 {
2622 /* Overlapping keys compare equal */
2623 if (bkey_cmp(&l->key, &START_KEY(&r->key)) <= 0)
2624 return -1;
2625 if (bkey_cmp(&START_KEY(&l->key), &r->key) >= 0)
2626 return 1;
2627 return 0;
2628 }
2629
keybuf_nonoverlapping_cmp(struct keybuf_key * l,struct keybuf_key * r)2630 static inline int keybuf_nonoverlapping_cmp(struct keybuf_key *l,
2631 struct keybuf_key *r)
2632 {
2633 return clamp_t(int64_t, bkey_cmp(&l->key, &r->key), -1, 1);
2634 }
2635
2636 struct refill {
2637 struct btree_op op;
2638 unsigned int nr_found;
2639 struct keybuf *buf;
2640 struct bkey *end;
2641 keybuf_pred_fn *pred;
2642 };
2643
refill_keybuf_fn(struct btree_op * op,struct btree * b,struct bkey * k)2644 static int refill_keybuf_fn(struct btree_op *op, struct btree *b,
2645 struct bkey *k)
2646 {
2647 struct refill *refill = container_of(op, struct refill, op);
2648 struct keybuf *buf = refill->buf;
2649 int ret = MAP_CONTINUE;
2650
2651 if (bkey_cmp(k, refill->end) > 0) {
2652 ret = MAP_DONE;
2653 goto out;
2654 }
2655
2656 if (!KEY_SIZE(k)) /* end key */
2657 goto out;
2658
2659 if (refill->pred(buf, k)) {
2660 struct keybuf_key *w;
2661
2662 spin_lock(&buf->lock);
2663
2664 w = array_alloc(&buf->freelist);
2665 if (!w) {
2666 spin_unlock(&buf->lock);
2667 return MAP_DONE;
2668 }
2669
2670 w->private = NULL;
2671 bkey_copy(&w->key, k);
2672
2673 if (RB_INSERT(&buf->keys, w, node, keybuf_cmp))
2674 array_free(&buf->freelist, w);
2675 else
2676 refill->nr_found++;
2677
2678 if (array_freelist_empty(&buf->freelist))
2679 ret = MAP_DONE;
2680
2681 spin_unlock(&buf->lock);
2682 }
2683 out:
2684 buf->last_scanned = *k;
2685 return ret;
2686 }
2687
bch_refill_keybuf(struct cache_set * c,struct keybuf * buf,struct bkey * end,keybuf_pred_fn * pred)2688 void bch_refill_keybuf(struct cache_set *c, struct keybuf *buf,
2689 struct bkey *end, keybuf_pred_fn *pred)
2690 {
2691 struct bkey start = buf->last_scanned;
2692 struct refill refill;
2693
2694 cond_resched();
2695
2696 bch_btree_op_init(&refill.op, -1);
2697 refill.nr_found = 0;
2698 refill.buf = buf;
2699 refill.end = end;
2700 refill.pred = pred;
2701
2702 bch_btree_map_keys(&refill.op, c, &buf->last_scanned,
2703 refill_keybuf_fn, MAP_END_KEY);
2704
2705 trace_bcache_keyscan(refill.nr_found,
2706 KEY_INODE(&start), KEY_OFFSET(&start),
2707 KEY_INODE(&buf->last_scanned),
2708 KEY_OFFSET(&buf->last_scanned));
2709
2710 spin_lock(&buf->lock);
2711
2712 if (!RB_EMPTY_ROOT(&buf->keys)) {
2713 struct keybuf_key *w;
2714
2715 w = RB_FIRST(&buf->keys, struct keybuf_key, node);
2716 buf->start = START_KEY(&w->key);
2717
2718 w = RB_LAST(&buf->keys, struct keybuf_key, node);
2719 buf->end = w->key;
2720 } else {
2721 buf->start = MAX_KEY;
2722 buf->end = MAX_KEY;
2723 }
2724
2725 spin_unlock(&buf->lock);
2726 }
2727
__bch_keybuf_del(struct keybuf * buf,struct keybuf_key * w)2728 static void __bch_keybuf_del(struct keybuf *buf, struct keybuf_key *w)
2729 {
2730 rb_erase(&w->node, &buf->keys);
2731 array_free(&buf->freelist, w);
2732 }
2733
bch_keybuf_del(struct keybuf * buf,struct keybuf_key * w)2734 void bch_keybuf_del(struct keybuf *buf, struct keybuf_key *w)
2735 {
2736 spin_lock(&buf->lock);
2737 __bch_keybuf_del(buf, w);
2738 spin_unlock(&buf->lock);
2739 }
2740
bch_keybuf_check_overlapping(struct keybuf * buf,struct bkey * start,struct bkey * end)2741 bool bch_keybuf_check_overlapping(struct keybuf *buf, struct bkey *start,
2742 struct bkey *end)
2743 {
2744 bool ret = false;
2745 struct keybuf_key *p, *w, s;
2746
2747 s.key = *start;
2748
2749 if (bkey_cmp(end, &buf->start) <= 0 ||
2750 bkey_cmp(start, &buf->end) >= 0)
2751 return false;
2752
2753 spin_lock(&buf->lock);
2754 w = RB_GREATER(&buf->keys, s, node, keybuf_nonoverlapping_cmp);
2755
2756 while (w && bkey_cmp(&START_KEY(&w->key), end) < 0) {
2757 p = w;
2758 w = RB_NEXT(w, node);
2759
2760 if (p->private)
2761 ret = true;
2762 else
2763 __bch_keybuf_del(buf, p);
2764 }
2765
2766 spin_unlock(&buf->lock);
2767 return ret;
2768 }
2769
bch_keybuf_next(struct keybuf * buf)2770 struct keybuf_key *bch_keybuf_next(struct keybuf *buf)
2771 {
2772 struct keybuf_key *w;
2773
2774 spin_lock(&buf->lock);
2775
2776 w = RB_FIRST(&buf->keys, struct keybuf_key, node);
2777
2778 while (w && w->private)
2779 w = RB_NEXT(w, node);
2780
2781 if (w)
2782 w->private = ERR_PTR(-EINTR);
2783
2784 spin_unlock(&buf->lock);
2785 return w;
2786 }
2787
bch_keybuf_next_rescan(struct cache_set * c,struct keybuf * buf,struct bkey * end,keybuf_pred_fn * pred)2788 struct keybuf_key *bch_keybuf_next_rescan(struct cache_set *c,
2789 struct keybuf *buf,
2790 struct bkey *end,
2791 keybuf_pred_fn *pred)
2792 {
2793 struct keybuf_key *ret;
2794
2795 while (1) {
2796 ret = bch_keybuf_next(buf);
2797 if (ret)
2798 break;
2799
2800 if (bkey_cmp(&buf->last_scanned, end) >= 0) {
2801 pr_debug("scan finished\n");
2802 break;
2803 }
2804
2805 bch_refill_keybuf(c, buf, end, pred);
2806 }
2807
2808 return ret;
2809 }
2810
bch_keybuf_init(struct keybuf * buf)2811 void bch_keybuf_init(struct keybuf *buf)
2812 {
2813 buf->last_scanned = MAX_KEY;
2814 buf->keys = RB_ROOT;
2815
2816 spin_lock_init(&buf->lock);
2817 array_allocator_init(&buf->freelist);
2818 }
2819
bch_btree_exit(void)2820 void bch_btree_exit(void)
2821 {
2822 if (btree_io_wq)
2823 destroy_workqueue(btree_io_wq);
2824 }
2825
bch_btree_init(void)2826 int __init bch_btree_init(void)
2827 {
2828 btree_io_wq = alloc_workqueue("bch_btree_io", WQ_MEM_RECLAIM, 0);
2829 if (!btree_io_wq)
2830 return -ENOMEM;
2831
2832 return 0;
2833 }
2834