]> git.ipfire.org Git - thirdparty/kernel/stable.git/blame - drivers/md/bcache/btree.c
bcache: allocate meta data pages as compound pages
[thirdparty/kernel/stable.git] / drivers / md / bcache / btree.c
CommitLineData
b2441318 1// SPDX-License-Identifier: GPL-2.0
cafe5635
KO
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 *
5fb94e9c 21 * All configuration is done via sysfs; see Documentation/admin-guide/bcache.rst.
cafe5635
KO
22 */
23
24#include "bcache.h"
25#include "btree.h"
26#include "debug.h"
65d45231 27#include "extents.h"
cafe5635
KO
28
29#include <linux/slab.h>
30#include <linux/bitops.h>
31#include <linux/hash.h>
72a44517 32#include <linux/kthread.h>
cd953ed0 33#include <linux/prefetch.h>
cafe5635
KO
34#include <linux/random.h>
35#include <linux/rcupdate.h>
e6017571 36#include <linux/sched/clock.h>
b2d09103 37#include <linux/rculist.h>
50a260e8 38#include <linux/delay.h>
cafe5635
KO
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 *
cafe5635
KO
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 *
cafe5635
KO
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
cafe5635
KO
91#define MAX_NEED_GC 64
92#define MAX_SAVE_PRIO 72
7f4a59de 93#define MAX_GC_TIMES 100
5c25c4fc
TJ
94#define MIN_GC_NODES 100
95#define GC_SLEEP_MS 100
cafe5635
KO
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
df8e8970
KO
102#define insert_lock(s, b) ((b)->level <= (s)->lock)
103
df8e8970 104
a85e968e
KO
105static inline struct bset *write_block(struct btree *b)
106{
107 return ((void *) btree_bset_first(b)) + b->written * block_bytes(b->c);
108}
109
2a285686
KO
110static void bch_btree_init_next(struct btree *b)
111{
112 /* If not a leaf node, always sort */
113 if (b->level && b->keys.nsets)
114 bch_btree_sort(&b->keys, &b->c->sort);
115 else
116 bch_btree_sort_lazy(&b->keys, &b->c->sort);
117
118 if (b->written < btree_blocks(b))
119 bch_bset_init_next(&b->keys, write_block(b),
120 bset_magic(&b->c->sb));
121
122}
123
cafe5635
KO
124/* Btree key manipulation */
125
3a3b6a4e 126void bkey_put(struct cache_set *c, struct bkey *k)
e7c590eb 127{
6f10f7d1 128 unsigned int i;
e7c590eb
KO
129
130 for (i = 0; i < KEY_PTRS(k); i++)
131 if (ptr_available(c, k, i))
132 atomic_dec_bug(&PTR_BUCKET(c, k, i)->pin);
133}
134
cafe5635
KO
135/* Btree IO */
136
137static uint64_t btree_csum_set(struct btree *b, struct bset *i)
138{
139 uint64_t crc = b->key.ptr[0];
fafff81c 140 void *data = (void *) i + 8, *end = bset_bkey_last(i);
cafe5635 141
169ef1cf 142 crc = bch_crc64_update(crc, data, end - data);
c19ed23a 143 return crc ^ 0xffffffffffffffffULL;
cafe5635
KO
144}
145
78b77bf8 146void bch_btree_node_read_done(struct btree *b)
cafe5635 147{
cafe5635 148 const char *err = "bad btree header";
ee811287 149 struct bset *i = btree_bset_first(b);
57943511 150 struct btree_iter *iter;
cafe5635 151
d2f96f48
SW
152 /*
153 * c->fill_iter can allocate an iterator with more memory space
154 * than static MAX_BSETS.
155 * See the comment arount cache_set->fill_iter.
156 */
d19936a2 157 iter = mempool_alloc(&b->c->fill_iter, GFP_NOIO);
57943511 158 iter->size = b->c->sb.bucket_size / b->c->sb.block_size;
cafe5635
KO
159 iter->used = 0;
160
280481d0 161#ifdef CONFIG_BCACHE_DEBUG
c052dd9a 162 iter->b = &b->keys;
280481d0
KO
163#endif
164
57943511 165 if (!i->seq)
cafe5635
KO
166 goto err;
167
168 for (;
a85e968e 169 b->written < btree_blocks(b) && i->seq == b->keys.set[0].data->seq;
cafe5635
KO
170 i = write_block(b)) {
171 err = "unsupported bset version";
172 if (i->version > BCACHE_BSET_VERSION)
173 goto err;
174
175 err = "bad btree header";
ee811287
KO
176 if (b->written + set_blocks(i, block_bytes(b->c)) >
177 btree_blocks(b))
cafe5635
KO
178 goto err;
179
180 err = "bad magic";
81ab4190 181 if (i->magic != bset_magic(&b->c->sb))
cafe5635
KO
182 goto err;
183
184 err = "bad checksum";
185 switch (i->version) {
186 case 0:
187 if (i->csum != csum_set(i))
188 goto err;
189 break;
190 case BCACHE_BSET_VERSION:
191 if (i->csum != btree_csum_set(b, i))
192 goto err;
193 break;
194 }
195
196 err = "empty set";
a85e968e 197 if (i != b->keys.set[0].data && !i->keys)
cafe5635
KO
198 goto err;
199
fafff81c 200 bch_btree_iter_push(iter, i->start, bset_bkey_last(i));
cafe5635 201
ee811287 202 b->written += set_blocks(i, block_bytes(b->c));
cafe5635
KO
203 }
204
205 err = "corrupted btree";
206 for (i = write_block(b);
a85e968e 207 bset_sector_offset(&b->keys, i) < KEY_SIZE(&b->key);
cafe5635 208 i = ((void *) i) + block_bytes(b->c))
a85e968e 209 if (i->seq == b->keys.set[0].data->seq)
cafe5635
KO
210 goto err;
211
a85e968e 212 bch_btree_sort_and_fix_extents(&b->keys, iter, &b->c->sort);
cafe5635 213
a85e968e 214 i = b->keys.set[0].data;
cafe5635 215 err = "short btree key";
a85e968e
KO
216 if (b->keys.set[0].size &&
217 bkey_cmp(&b->key, &b->keys.set[0].end) < 0)
cafe5635
KO
218 goto err;
219
220 if (b->written < btree_blocks(b))
a85e968e
KO
221 bch_bset_init_next(&b->keys, write_block(b),
222 bset_magic(&b->c->sb));
cafe5635 223out:
d19936a2 224 mempool_free(iter, &b->c->fill_iter);
57943511 225 return;
cafe5635
KO
226err:
227 set_btree_node_io_error(b);
88b9f8c4 228 bch_cache_set_error(b->c, "%s at bucket %zu, block %u, %u keys",
cafe5635 229 err, PTR_BUCKET_NR(b->c, &b->key, 0),
88b9f8c4 230 bset_block_offset(b, i), i->keys);
cafe5635
KO
231 goto out;
232}
233
4246a0b6 234static void btree_node_read_endio(struct bio *bio)
cafe5635 235{
57943511 236 struct closure *cl = bio->bi_private;
1fae7cf0 237
57943511
KO
238 closure_put(cl);
239}
cafe5635 240
78b77bf8 241static void bch_btree_node_read(struct btree *b)
57943511
KO
242{
243 uint64_t start_time = local_clock();
244 struct closure cl;
245 struct bio *bio;
cafe5635 246
c37511b8
KO
247 trace_bcache_btree_read(b);
248
57943511 249 closure_init_stack(&cl);
cafe5635 250
57943511 251 bio = bch_bbio_alloc(b->c);
4f024f37 252 bio->bi_iter.bi_size = KEY_SIZE(&b->key) << 9;
57943511
KO
253 bio->bi_end_io = btree_node_read_endio;
254 bio->bi_private = &cl;
70fd7614 255 bio->bi_opf = REQ_OP_READ | REQ_META;
cafe5635 256
a85e968e 257 bch_bio_map(bio, b->keys.set[0].data);
cafe5635 258
57943511
KO
259 bch_submit_bbio(bio, b->c, &b->key, 0);
260 closure_sync(&cl);
cafe5635 261
4e4cbee9 262 if (bio->bi_status)
57943511
KO
263 set_btree_node_io_error(b);
264
265 bch_bbio_free(bio, b->c);
266
267 if (btree_node_io_error(b))
268 goto err;
269
270 bch_btree_node_read_done(b);
57943511 271 bch_time_stats_update(&b->c->btree_read_time, start_time);
57943511
KO
272
273 return;
274err:
61cbd250 275 bch_cache_set_error(b->c, "io error reading bucket %zu",
57943511 276 PTR_BUCKET_NR(b->c, &b->key, 0));
cafe5635
KO
277}
278
279static void btree_complete_write(struct btree *b, struct btree_write *w)
280{
281 if (w->prio_blocked &&
282 !atomic_sub_return(w->prio_blocked, &b->c->prio_blocked))
119ba0f8 283 wake_up_allocators(b->c);
cafe5635
KO
284
285 if (w->journal) {
286 atomic_dec_bug(w->journal);
287 __closure_wake_up(&b->c->journal.wait);
288 }
289
cafe5635
KO
290 w->prio_blocked = 0;
291 w->journal = NULL;
cafe5635
KO
292}
293
cb7a583e
KO
294static void btree_node_write_unlock(struct closure *cl)
295{
296 struct btree *b = container_of(cl, struct btree, io);
297
298 up(&b->io_mutex);
299}
300
57943511 301static void __btree_node_write_done(struct closure *cl)
cafe5635 302{
cb7a583e 303 struct btree *b = container_of(cl, struct btree, io);
cafe5635
KO
304 struct btree_write *w = btree_prev_write(b);
305
306 bch_bbio_free(b->bio, b->c);
307 b->bio = NULL;
308 btree_complete_write(b, w);
309
310 if (btree_node_dirty(b))
56b30770 311 schedule_delayed_work(&b->work, 30 * HZ);
cafe5635 312
cb7a583e 313 closure_return_with_destructor(cl, btree_node_write_unlock);
cafe5635
KO
314}
315
57943511 316static void btree_node_write_done(struct closure *cl)
cafe5635 317{
cb7a583e 318 struct btree *b = container_of(cl, struct btree, io);
cafe5635 319
491221f8 320 bio_free_pages(b->bio);
57943511 321 __btree_node_write_done(cl);
cafe5635
KO
322}
323
4246a0b6 324static void btree_node_write_endio(struct bio *bio)
57943511
KO
325{
326 struct closure *cl = bio->bi_private;
cb7a583e 327 struct btree *b = container_of(cl, struct btree, io);
57943511 328
4e4cbee9 329 if (bio->bi_status)
57943511
KO
330 set_btree_node_io_error(b);
331
4e4cbee9 332 bch_bbio_count_io_errors(b->c, bio, bio->bi_status, "writing btree");
57943511
KO
333 closure_put(cl);
334}
335
336static void do_btree_node_write(struct btree *b)
cafe5635 337{
cb7a583e 338 struct closure *cl = &b->io;
ee811287 339 struct bset *i = btree_bset_last(b);
cafe5635
KO
340 BKEY_PADDED(key) k;
341
342 i->version = BCACHE_BSET_VERSION;
343 i->csum = btree_csum_set(b, i);
344
57943511
KO
345 BUG_ON(b->bio);
346 b->bio = bch_bbio_alloc(b->c);
347
348 b->bio->bi_end_io = btree_node_write_endio;
faadf0c9 349 b->bio->bi_private = cl;
ee811287 350 b->bio->bi_iter.bi_size = roundup(set_bytes(i), block_bytes(b->c));
70fd7614 351 b->bio->bi_opf = REQ_OP_WRITE | REQ_META | REQ_FUA;
169ef1cf 352 bch_bio_map(b->bio, i);
cafe5635 353
e49c7c37
KO
354 /*
355 * If we're appending to a leaf node, we don't technically need FUA -
356 * this write just needs to be persisted before the next journal write,
357 * which will be marked FLUSH|FUA.
358 *
359 * Similarly if we're writing a new btree root - the pointer is going to
360 * be in the next journal entry.
361 *
362 * But if we're writing a new btree node (that isn't a root) or
363 * appending to a non leaf btree node, we need either FUA or a flush
364 * when we write the parent with the new pointer. FUA is cheaper than a
365 * flush, and writes appending to leaf nodes aren't blocking anything so
366 * just make all btree node writes FUA to keep things sane.
367 */
368
cafe5635 369 bkey_copy(&k.key, &b->key);
ee811287 370 SET_PTR_OFFSET(&k.key, 0, PTR_OFFSET(&k.key, 0) +
a85e968e 371 bset_sector_offset(&b->keys, i));
cafe5635 372
25d8be77 373 if (!bch_bio_alloc_pages(b->bio, __GFP_NOWARN|GFP_NOWAIT)) {
cafe5635 374 struct bio_vec *bv;
f936b06a 375 void *addr = (void *) ((unsigned long) i & ~(PAGE_SIZE - 1));
6dc4f100 376 struct bvec_iter_all iter_all;
cafe5635 377
2b070cfe 378 bio_for_each_segment_all(bv, b->bio, iter_all) {
f936b06a
CH
379 memcpy(page_address(bv->bv_page), addr, PAGE_SIZE);
380 addr += PAGE_SIZE;
381 }
cafe5635 382
cafe5635
KO
383 bch_submit_bbio(b->bio, b->c, &k.key, 0);
384
57943511 385 continue_at(cl, btree_node_write_done, NULL);
cafe5635 386 } else {
b0d30981
CL
387 /*
388 * No problem for multipage bvec since the bio is
389 * just allocated
390 */
cafe5635 391 b->bio->bi_vcnt = 0;
169ef1cf 392 bch_bio_map(b->bio, i);
cafe5635 393
cafe5635
KO
394 bch_submit_bbio(b->bio, b->c, &k.key, 0);
395
396 closure_sync(cl);
cb7a583e 397 continue_at_nobarrier(cl, __btree_node_write_done, NULL);
cafe5635
KO
398 }
399}
400
2a285686 401void __bch_btree_node_write(struct btree *b, struct closure *parent)
cafe5635 402{
ee811287 403 struct bset *i = btree_bset_last(b);
cafe5635 404
2a285686
KO
405 lockdep_assert_held(&b->write_lock);
406
c37511b8
KO
407 trace_bcache_btree_write(b);
408
cafe5635 409 BUG_ON(current->bio_list);
57943511
KO
410 BUG_ON(b->written >= btree_blocks(b));
411 BUG_ON(b->written && !i->keys);
ee811287 412 BUG_ON(btree_bset_first(b)->seq != i->seq);
dc9d98d6 413 bch_check_keys(&b->keys, "writing");
cafe5635 414
cafe5635
KO
415 cancel_delayed_work(&b->work);
416
57943511 417 /* If caller isn't waiting for write, parent refcount is cache set */
cb7a583e
KO
418 down(&b->io_mutex);
419 closure_init(&b->io, parent ?: &b->c->cl);
57943511 420
cafe5635
KO
421 clear_bit(BTREE_NODE_dirty, &b->flags);
422 change_bit(BTREE_NODE_write_idx, &b->flags);
423
57943511 424 do_btree_node_write(b);
cafe5635 425
ee811287 426 atomic_long_add(set_blocks(i, block_bytes(b->c)) * b->c->sb.block_size,
cafe5635
KO
427 &PTR_CACHE(b->c, &b->key, 0)->btree_sectors_written);
428
a85e968e 429 b->written += set_blocks(i, block_bytes(b->c));
2a285686 430}
a85e968e 431
2a285686
KO
432void bch_btree_node_write(struct btree *b, struct closure *parent)
433{
6f10f7d1 434 unsigned int nsets = b->keys.nsets;
2a285686
KO
435
436 lockdep_assert_held(&b->lock);
437
438 __bch_btree_node_write(b, parent);
cafe5635 439
78b77bf8
KO
440 /*
441 * do verify if there was more than one set initially (i.e. we did a
442 * sort) and we sorted down to a single set:
443 */
2a285686 444 if (nsets && !b->keys.nsets)
78b77bf8
KO
445 bch_btree_verify(b);
446
2a285686 447 bch_btree_init_next(b);
cafe5635
KO
448}
449
f269af5a
KO
450static void bch_btree_node_write_sync(struct btree *b)
451{
452 struct closure cl;
453
454 closure_init_stack(&cl);
2a285686
KO
455
456 mutex_lock(&b->write_lock);
f269af5a 457 bch_btree_node_write(b, &cl);
2a285686
KO
458 mutex_unlock(&b->write_lock);
459
f269af5a
KO
460 closure_sync(&cl);
461}
462
57943511 463static void btree_node_write_work(struct work_struct *w)
cafe5635
KO
464{
465 struct btree *b = container_of(to_delayed_work(w), struct btree, work);
466
2a285686 467 mutex_lock(&b->write_lock);
cafe5635 468 if (btree_node_dirty(b))
2a285686
KO
469 __bch_btree_node_write(b, NULL);
470 mutex_unlock(&b->write_lock);
cafe5635
KO
471}
472
c18536a7 473static void bch_btree_leaf_dirty(struct btree *b, atomic_t *journal_ref)
cafe5635 474{
ee811287 475 struct bset *i = btree_bset_last(b);
cafe5635
KO
476 struct btree_write *w = btree_current_write(b);
477
2a285686
KO
478 lockdep_assert_held(&b->write_lock);
479
57943511
KO
480 BUG_ON(!b->written);
481 BUG_ON(!i->keys);
cafe5635 482
57943511 483 if (!btree_node_dirty(b))
56b30770 484 schedule_delayed_work(&b->work, 30 * HZ);
cafe5635 485
57943511 486 set_btree_node_dirty(b);
cafe5635 487
5dccefd3
CL
488 /*
489 * w->journal is always the oldest journal pin of all bkeys
490 * in the leaf node, to make sure the oldest jset seq won't
491 * be increased before this btree node is flushed.
492 */
c18536a7 493 if (journal_ref) {
cafe5635 494 if (w->journal &&
c18536a7 495 journal_pin_cmp(b->c, w->journal, journal_ref)) {
cafe5635
KO
496 atomic_dec_bug(w->journal);
497 w->journal = NULL;
498 }
499
500 if (!w->journal) {
c18536a7 501 w->journal = journal_ref;
cafe5635
KO
502 atomic_inc(w->journal);
503 }
504 }
505
cafe5635 506 /* Force write if set is too big */
57943511
KO
507 if (set_bytes(i) > PAGE_SIZE - 48 &&
508 !current->bio_list)
509 bch_btree_node_write(b, NULL);
cafe5635
KO
510}
511
512/*
513 * Btree in memory cache - allocation/freeing
514 * mca -> memory cache
515 */
516
cafe5635
KO
517#define mca_reserve(c) (((c->root && c->root->level) \
518 ? c->root->level : 1) * 8 + 16)
519#define mca_can_free(c) \
0a63b66d 520 max_t(int, 0, c->btree_cache_used - mca_reserve(c))
cafe5635
KO
521
522static void mca_data_free(struct btree *b)
523{
cb7a583e 524 BUG_ON(b->io_mutex.count != 1);
cafe5635 525
a85e968e 526 bch_btree_keys_free(&b->keys);
cafe5635 527
0a63b66d 528 b->c->btree_cache_used--;
ee811287 529 list_move(&b->list, &b->c->btree_cache_freed);
cafe5635
KO
530}
531
532static void mca_bucket_free(struct btree *b)
533{
534 BUG_ON(btree_node_dirty(b));
535
536 b->key.ptr[0] = 0;
537 hlist_del_init_rcu(&b->hash);
538 list_move(&b->list, &b->c->btree_cache_freeable);
539}
540
6f10f7d1 541static unsigned int btree_order(struct bkey *k)
cafe5635
KO
542{
543 return ilog2(KEY_SIZE(k) / PAGE_SECTORS ?: 1);
544}
545
546static void mca_data_alloc(struct btree *b, struct bkey *k, gfp_t gfp)
547{
a85e968e 548 if (!bch_btree_keys_alloc(&b->keys,
6f10f7d1 549 max_t(unsigned int,
ee811287
KO
550 ilog2(b->c->btree_pages),
551 btree_order(k)),
552 gfp)) {
0a63b66d 553 b->c->btree_cache_used++;
ee811287
KO
554 list_move(&b->list, &b->c->btree_cache);
555 } else {
556 list_move(&b->list, &b->c->btree_cache_freed);
557 }
cafe5635
KO
558}
559
560static struct btree *mca_bucket_alloc(struct cache_set *c,
561 struct bkey *k, gfp_t gfp)
562{
bd9026c8
CL
563 /*
564 * kzalloc() is necessary here for initialization,
565 * see code comments in bch_btree_keys_init().
566 */
cafe5635 567 struct btree *b = kzalloc(sizeof(struct btree), gfp);
1fae7cf0 568
cafe5635
KO
569 if (!b)
570 return NULL;
571
572 init_rwsem(&b->lock);
573 lockdep_set_novalidate_class(&b->lock);
2a285686
KO
574 mutex_init(&b->write_lock);
575 lockdep_set_novalidate_class(&b->write_lock);
cafe5635 576 INIT_LIST_HEAD(&b->list);
57943511 577 INIT_DELAYED_WORK(&b->work, btree_node_write_work);
cafe5635 578 b->c = c;
cb7a583e 579 sema_init(&b->io_mutex, 1);
cafe5635
KO
580
581 mca_data_alloc(b, k, gfp);
582 return b;
583}
584
6f10f7d1 585static int mca_reap(struct btree *b, unsigned int min_order, bool flush)
cafe5635 586{
e8e1d468
KO
587 struct closure cl;
588
589 closure_init_stack(&cl);
cafe5635
KO
590 lockdep_assert_held(&b->c->bucket_lock);
591
592 if (!down_write_trylock(&b->lock))
593 return -ENOMEM;
594
a85e968e 595 BUG_ON(btree_node_dirty(b) && !b->keys.set[0].data);
e8e1d468 596
a85e968e 597 if (b->keys.page_order < min_order)
cb7a583e
KO
598 goto out_unlock;
599
600 if (!flush) {
601 if (btree_node_dirty(b))
602 goto out_unlock;
603
604 if (down_trylock(&b->io_mutex))
605 goto out_unlock;
606 up(&b->io_mutex);
cafe5635
KO
607 }
608
50a260e8 609retry:
41508bb7
CL
610 /*
611 * BTREE_NODE_dirty might be cleared in btree_flush_btree() by
612 * __bch_btree_node_write(). To avoid an extra flush, acquire
613 * b->write_lock before checking BTREE_NODE_dirty bit.
614 */
2a285686 615 mutex_lock(&b->write_lock);
50a260e8
CL
616 /*
617 * If this btree node is selected in btree_flush_write() by journal
618 * code, delay and retry until the node is flushed by journal code
619 * and BTREE_NODE_journal_flush bit cleared by btree_flush_write().
620 */
621 if (btree_node_journal_flush(b)) {
46f5aa88 622 pr_debug("bnode %p is flushing by journal, retry\n", b);
50a260e8
CL
623 mutex_unlock(&b->write_lock);
624 udelay(1);
625 goto retry;
626 }
627
f269af5a 628 if (btree_node_dirty(b))
2a285686
KO
629 __bch_btree_node_write(b, &cl);
630 mutex_unlock(&b->write_lock);
631
632 closure_sync(&cl);
cafe5635 633
e8e1d468 634 /* wait for any in flight btree write */
cb7a583e
KO
635 down(&b->io_mutex);
636 up(&b->io_mutex);
e8e1d468 637
cafe5635 638 return 0;
cb7a583e
KO
639out_unlock:
640 rw_unlock(true, b);
641 return -ENOMEM;
cafe5635
KO
642}
643
7dc19d5a
DC
644static unsigned long bch_mca_scan(struct shrinker *shrink,
645 struct shrink_control *sc)
cafe5635
KO
646{
647 struct cache_set *c = container_of(shrink, struct cache_set, shrink);
648 struct btree *b, *t;
649 unsigned long i, nr = sc->nr_to_scan;
7dc19d5a 650 unsigned long freed = 0;
ca71df31 651 unsigned int btree_cache_used;
cafe5635
KO
652
653 if (c->shrinker_disabled)
7dc19d5a 654 return SHRINK_STOP;
cafe5635 655
0a63b66d 656 if (c->btree_cache_alloc_lock)
7dc19d5a 657 return SHRINK_STOP;
cafe5635
KO
658
659 /* Return -1 if we can't do anything right now */
a698e08c 660 if (sc->gfp_mask & __GFP_IO)
cafe5635
KO
661 mutex_lock(&c->bucket_lock);
662 else if (!mutex_trylock(&c->bucket_lock))
663 return -1;
664
36c9ea98
KO
665 /*
666 * It's _really_ critical that we don't free too many btree nodes - we
667 * have to always leave ourselves a reserve. The reserve is how we
668 * guarantee that allocating memory for a new btree node can always
669 * succeed, so that inserting keys into the btree can always succeed and
670 * IO can always make forward progress:
671 */
cafe5635 672 nr /= c->btree_pages;
9fcc34b1
CL
673 if (nr == 0)
674 nr = 1;
cafe5635
KO
675 nr = min_t(unsigned long, nr, mca_can_free(c));
676
677 i = 0;
ca71df31 678 btree_cache_used = c->btree_cache_used;
d5c9c470 679 list_for_each_entry_safe_reverse(b, t, &c->btree_cache_freeable, list) {
ca71df31
TJ
680 if (nr <= 0)
681 goto out;
cafe5635 682
d5c9c470 683 if (!mca_reap(b, 0, false)) {
cafe5635
KO
684 mca_data_free(b);
685 rw_unlock(true, b);
7dc19d5a 686 freed++;
cafe5635 687 }
ca71df31 688 nr--;
d5c9c470 689 i++;
cafe5635
KO
690 }
691
e3de0446
CL
692 list_for_each_entry_safe_reverse(b, t, &c->btree_cache, list) {
693 if (nr <= 0 || i >= btree_cache_used)
b0f32a56
KO
694 goto out;
695
125d98ed 696 if (!mca_reap(b, 0, false)) {
cafe5635
KO
697 mca_bucket_free(b);
698 mca_data_free(b);
699 rw_unlock(true, b);
7dc19d5a 700 freed++;
125d98ed 701 }
e3de0446
CL
702
703 nr--;
704 i++;
cafe5635
KO
705 }
706out:
cafe5635 707 mutex_unlock(&c->bucket_lock);
f3641c3a 708 return freed * c->btree_pages;
7dc19d5a
DC
709}
710
711static unsigned long bch_mca_count(struct shrinker *shrink,
712 struct shrink_control *sc)
713{
714 struct cache_set *c = container_of(shrink, struct cache_set, shrink);
715
716 if (c->shrinker_disabled)
717 return 0;
718
0a63b66d 719 if (c->btree_cache_alloc_lock)
7dc19d5a
DC
720 return 0;
721
722 return mca_can_free(c) * c->btree_pages;
cafe5635
KO
723}
724
725void bch_btree_cache_free(struct cache_set *c)
726{
727 struct btree *b;
728 struct closure cl;
1fae7cf0 729
cafe5635
KO
730 closure_init_stack(&cl);
731
732 if (c->shrink.list.next)
733 unregister_shrinker(&c->shrink);
734
735 mutex_lock(&c->bucket_lock);
736
737#ifdef CONFIG_BCACHE_DEBUG
738 if (c->verify_data)
739 list_move(&c->verify_data->list, &c->btree_cache);
78b77bf8
KO
740
741 free_pages((unsigned long) c->verify_ondisk, ilog2(bucket_pages(c)));
cafe5635
KO
742#endif
743
744 list_splice(&c->btree_cache_freeable,
745 &c->btree_cache);
746
747 while (!list_empty(&c->btree_cache)) {
748 b = list_first_entry(&c->btree_cache, struct btree, list);
749
41508bb7
CL
750 /*
751 * This function is called by cache_set_free(), no I/O
752 * request on cache now, it is unnecessary to acquire
753 * b->write_lock before clearing BTREE_NODE_dirty anymore.
754 */
e5ec5f47 755 if (btree_node_dirty(b)) {
cafe5635 756 btree_complete_write(b, btree_current_write(b));
e5ec5f47
CL
757 clear_bit(BTREE_NODE_dirty, &b->flags);
758 }
cafe5635
KO
759 mca_data_free(b);
760 }
761
762 while (!list_empty(&c->btree_cache_freed)) {
763 b = list_first_entry(&c->btree_cache_freed,
764 struct btree, list);
765 list_del(&b->list);
766 cancel_delayed_work_sync(&b->work);
767 kfree(b);
768 }
769
770 mutex_unlock(&c->bucket_lock);
771}
772
773int bch_btree_cache_alloc(struct cache_set *c)
774{
6f10f7d1 775 unsigned int i;
cafe5635 776
cafe5635 777 for (i = 0; i < mca_reserve(c); i++)
72a44517
KO
778 if (!mca_bucket_alloc(c, &ZERO_KEY, GFP_KERNEL))
779 return -ENOMEM;
cafe5635
KO
780
781 list_splice_init(&c->btree_cache,
782 &c->btree_cache_freeable);
783
784#ifdef CONFIG_BCACHE_DEBUG
785 mutex_init(&c->verify_lock);
786
78b77bf8 787 c->verify_ondisk = (void *)
5fe48867 788 __get_free_pages(GFP_KERNEL|__GFP_COMP, ilog2(bucket_pages(c)));
78b77bf8 789
cafe5635
KO
790 c->verify_data = mca_bucket_alloc(c, &ZERO_KEY, GFP_KERNEL);
791
792 if (c->verify_data &&
a85e968e 793 c->verify_data->keys.set->data)
cafe5635
KO
794 list_del_init(&c->verify_data->list);
795 else
796 c->verify_data = NULL;
797#endif
798
7dc19d5a
DC
799 c->shrink.count_objects = bch_mca_count;
800 c->shrink.scan_objects = bch_mca_scan;
cafe5635
KO
801 c->shrink.seeks = 4;
802 c->shrink.batch = c->btree_pages * 2;
6c4ca1e3
ML
803
804 if (register_shrinker(&c->shrink))
46f5aa88 805 pr_warn("bcache: %s: could not register shrinker\n",
6c4ca1e3 806 __func__);
cafe5635
KO
807
808 return 0;
809}
810
811/* Btree in memory cache - hash table */
812
813static struct hlist_head *mca_hash(struct cache_set *c, struct bkey *k)
814{
815 return &c->bucket_hash[hash_32(PTR_HASH(c, k), BUCKET_HASH_BITS)];
816}
817
818static struct btree *mca_find(struct cache_set *c, struct bkey *k)
819{
820 struct btree *b;
821
822 rcu_read_lock();
823 hlist_for_each_entry_rcu(b, mca_hash(c, k), hash)
824 if (PTR_HASH(c, &b->key) == PTR_HASH(c, k))
825 goto out;
826 b = NULL;
827out:
828 rcu_read_unlock();
829 return b;
830}
831
0a63b66d
KO
832static int mca_cannibalize_lock(struct cache_set *c, struct btree_op *op)
833{
34cf78bf
GF
834 spin_lock(&c->btree_cannibalize_lock);
835 if (likely(c->btree_cache_alloc_lock == NULL)) {
836 c->btree_cache_alloc_lock = current;
837 } else if (c->btree_cache_alloc_lock != current) {
0a63b66d
KO
838 if (op)
839 prepare_to_wait(&c->btree_cache_wait, &op->wait,
840 TASK_UNINTERRUPTIBLE);
34cf78bf 841 spin_unlock(&c->btree_cannibalize_lock);
0a63b66d
KO
842 return -EINTR;
843 }
34cf78bf 844 spin_unlock(&c->btree_cannibalize_lock);
0a63b66d
KO
845
846 return 0;
847}
848
849static struct btree *mca_cannibalize(struct cache_set *c, struct btree_op *op,
850 struct bkey *k)
cafe5635 851{
e8e1d468 852 struct btree *b;
cafe5635 853
c37511b8
KO
854 trace_bcache_btree_cache_cannibalize(c);
855
0a63b66d
KO
856 if (mca_cannibalize_lock(c, op))
857 return ERR_PTR(-EINTR);
cafe5635 858
e8e1d468
KO
859 list_for_each_entry_reverse(b, &c->btree_cache, list)
860 if (!mca_reap(b, btree_order(k), false))
861 return b;
cafe5635 862
e8e1d468
KO
863 list_for_each_entry_reverse(b, &c->btree_cache, list)
864 if (!mca_reap(b, btree_order(k), true))
865 return b;
cafe5635 866
0a63b66d 867 WARN(1, "btree cache cannibalize failed\n");
e8e1d468 868 return ERR_PTR(-ENOMEM);
cafe5635
KO
869}
870
871/*
872 * We can only have one thread cannibalizing other cached btree nodes at a time,
873 * or we'll deadlock. We use an open coded mutex to ensure that, which a
874 * cannibalize_bucket() will take. This means every time we unlock the root of
875 * the btree, we need to release this lock if we have it held.
876 */
df8e8970 877static void bch_cannibalize_unlock(struct cache_set *c)
cafe5635 878{
34cf78bf 879 spin_lock(&c->btree_cannibalize_lock);
0a63b66d
KO
880 if (c->btree_cache_alloc_lock == current) {
881 c->btree_cache_alloc_lock = NULL;
882 wake_up(&c->btree_cache_wait);
cafe5635 883 }
34cf78bf 884 spin_unlock(&c->btree_cannibalize_lock);
cafe5635
KO
885}
886
0a63b66d
KO
887static struct btree *mca_alloc(struct cache_set *c, struct btree_op *op,
888 struct bkey *k, int level)
cafe5635
KO
889{
890 struct btree *b;
891
e8e1d468
KO
892 BUG_ON(current->bio_list);
893
cafe5635
KO
894 lockdep_assert_held(&c->bucket_lock);
895
896 if (mca_find(c, k))
897 return NULL;
898
899 /* btree_free() doesn't free memory; it sticks the node on the end of
900 * the list. Check if there's any freed nodes there:
901 */
902 list_for_each_entry(b, &c->btree_cache_freeable, list)
e8e1d468 903 if (!mca_reap(b, btree_order(k), false))
cafe5635
KO
904 goto out;
905
906 /* We never free struct btree itself, just the memory that holds the on
907 * disk node. Check the freed list before allocating a new one:
908 */
909 list_for_each_entry(b, &c->btree_cache_freed, list)
e8e1d468 910 if (!mca_reap(b, 0, false)) {
cafe5635 911 mca_data_alloc(b, k, __GFP_NOWARN|GFP_NOIO);
a85e968e 912 if (!b->keys.set[0].data)
cafe5635
KO
913 goto err;
914 else
915 goto out;
916 }
917
918 b = mca_bucket_alloc(c, k, __GFP_NOWARN|GFP_NOIO);
919 if (!b)
920 goto err;
921
922 BUG_ON(!down_write_trylock(&b->lock));
a85e968e 923 if (!b->keys.set->data)
cafe5635
KO
924 goto err;
925out:
cb7a583e 926 BUG_ON(b->io_mutex.count != 1);
cafe5635
KO
927
928 bkey_copy(&b->key, k);
929 list_move(&b->list, &c->btree_cache);
930 hlist_del_init_rcu(&b->hash);
931 hlist_add_head_rcu(&b->hash, mca_hash(c, k));
932
933 lock_set_subclass(&b->lock.dep_map, level + 1, _THIS_IP_);
d6fd3b11 934 b->parent = (void *) ~0UL;
a85e968e
KO
935 b->flags = 0;
936 b->written = 0;
937 b->level = level;
cafe5635 938
65d45231 939 if (!b->level)
a85e968e
KO
940 bch_btree_keys_init(&b->keys, &bch_extent_keys_ops,
941 &b->c->expensive_debug_checks);
65d45231 942 else
a85e968e
KO
943 bch_btree_keys_init(&b->keys, &bch_btree_keys_ops,
944 &b->c->expensive_debug_checks);
cafe5635
KO
945
946 return b;
947err:
948 if (b)
949 rw_unlock(true, b);
950
0a63b66d 951 b = mca_cannibalize(c, op, k);
cafe5635
KO
952 if (!IS_ERR(b))
953 goto out;
954
955 return b;
956}
957
47344e33 958/*
cafe5635
KO
959 * bch_btree_node_get - find a btree node in the cache and lock it, reading it
960 * in from disk if necessary.
961 *
ed00aabd 962 * If IO is necessary and running under submit_bio_noacct, returns -EAGAIN.
cafe5635
KO
963 *
964 * The btree node will have either a read or a write lock held, depending on
965 * level and op->lock.
966 */
0a63b66d 967struct btree *bch_btree_node_get(struct cache_set *c, struct btree_op *op,
2452cc89
SP
968 struct bkey *k, int level, bool write,
969 struct btree *parent)
cafe5635
KO
970{
971 int i = 0;
cafe5635
KO
972 struct btree *b;
973
974 BUG_ON(level < 0);
975retry:
976 b = mca_find(c, k);
977
978 if (!b) {
57943511
KO
979 if (current->bio_list)
980 return ERR_PTR(-EAGAIN);
981
cafe5635 982 mutex_lock(&c->bucket_lock);
0a63b66d 983 b = mca_alloc(c, op, k, level);
cafe5635
KO
984 mutex_unlock(&c->bucket_lock);
985
986 if (!b)
987 goto retry;
988 if (IS_ERR(b))
989 return b;
990
57943511 991 bch_btree_node_read(b);
cafe5635
KO
992
993 if (!write)
994 downgrade_write(&b->lock);
995 } else {
996 rw_lock(write, b, level);
997 if (PTR_HASH(c, &b->key) != PTR_HASH(c, k)) {
998 rw_unlock(write, b);
999 goto retry;
1000 }
1001 BUG_ON(b->level != level);
1002 }
1003
c2e8dcf7
CL
1004 if (btree_node_io_error(b)) {
1005 rw_unlock(write, b);
1006 return ERR_PTR(-EIO);
1007 }
1008
1009 BUG_ON(!b->written);
1010
2452cc89 1011 b->parent = parent;
cafe5635 1012
a85e968e
KO
1013 for (; i <= b->keys.nsets && b->keys.set[i].size; i++) {
1014 prefetch(b->keys.set[i].tree);
1015 prefetch(b->keys.set[i].data);
cafe5635
KO
1016 }
1017
a85e968e
KO
1018 for (; i <= b->keys.nsets; i++)
1019 prefetch(b->keys.set[i].data);
cafe5635 1020
cafe5635
KO
1021 return b;
1022}
1023
2452cc89 1024static void btree_node_prefetch(struct btree *parent, struct bkey *k)
cafe5635
KO
1025{
1026 struct btree *b;
1027
2452cc89
SP
1028 mutex_lock(&parent->c->bucket_lock);
1029 b = mca_alloc(parent->c, NULL, k, parent->level - 1);
1030 mutex_unlock(&parent->c->bucket_lock);
cafe5635
KO
1031
1032 if (!IS_ERR_OR_NULL(b)) {
2452cc89 1033 b->parent = parent;
57943511 1034 bch_btree_node_read(b);
cafe5635
KO
1035 rw_unlock(true, b);
1036 }
1037}
1038
1039/* Btree alloc */
1040
e8e1d468 1041static void btree_node_free(struct btree *b)
cafe5635 1042{
c37511b8
KO
1043 trace_bcache_btree_node_free(b);
1044
cafe5635 1045 BUG_ON(b == b->c->root);
cafe5635 1046
50a260e8 1047retry:
2a285686 1048 mutex_lock(&b->write_lock);
50a260e8
CL
1049 /*
1050 * If the btree node is selected and flushing in btree_flush_write(),
1051 * delay and retry until the BTREE_NODE_journal_flush bit cleared,
1052 * then it is safe to free the btree node here. Otherwise this btree
1053 * node will be in race condition.
1054 */
1055 if (btree_node_journal_flush(b)) {
1056 mutex_unlock(&b->write_lock);
46f5aa88 1057 pr_debug("bnode %p journal_flush set, retry\n", b);
50a260e8
CL
1058 udelay(1);
1059 goto retry;
1060 }
2a285686 1061
e5ec5f47 1062 if (btree_node_dirty(b)) {
cafe5635 1063 btree_complete_write(b, btree_current_write(b));
e5ec5f47
CL
1064 clear_bit(BTREE_NODE_dirty, &b->flags);
1065 }
cafe5635 1066
2a285686
KO
1067 mutex_unlock(&b->write_lock);
1068
cafe5635
KO
1069 cancel_delayed_work(&b->work);
1070
1071 mutex_lock(&b->c->bucket_lock);
cafe5635
KO
1072 bch_bucket_free(b->c, &b->key);
1073 mca_bucket_free(b);
1074 mutex_unlock(&b->c->bucket_lock);
1075}
1076
c5aa4a31 1077struct btree *__bch_btree_node_alloc(struct cache_set *c, struct btree_op *op,
2452cc89
SP
1078 int level, bool wait,
1079 struct btree *parent)
cafe5635
KO
1080{
1081 BKEY_PADDED(key) k;
1082 struct btree *b = ERR_PTR(-EAGAIN);
1083
1084 mutex_lock(&c->bucket_lock);
1085retry:
c5aa4a31 1086 if (__bch_bucket_alloc_set(c, RESERVE_BTREE, &k.key, 1, wait))
cafe5635
KO
1087 goto err;
1088
3a3b6a4e 1089 bkey_put(c, &k.key);
cafe5635
KO
1090 SET_KEY_SIZE(&k.key, c->btree_pages * PAGE_SECTORS);
1091
0a63b66d 1092 b = mca_alloc(c, op, &k.key, level);
cafe5635
KO
1093 if (IS_ERR(b))
1094 goto err_free;
1095
1096 if (!b) {
b1a67b0f
KO
1097 cache_bug(c,
1098 "Tried to allocate bucket that was in btree cache");
cafe5635
KO
1099 goto retry;
1100 }
1101
2452cc89 1102 b->parent = parent;
a85e968e 1103 bch_bset_init_next(&b->keys, b->keys.set->data, bset_magic(&b->c->sb));
cafe5635
KO
1104
1105 mutex_unlock(&c->bucket_lock);
c37511b8
KO
1106
1107 trace_bcache_btree_node_alloc(b);
cafe5635
KO
1108 return b;
1109err_free:
1110 bch_bucket_free(c, &k.key);
cafe5635
KO
1111err:
1112 mutex_unlock(&c->bucket_lock);
c37511b8 1113
913dc33f 1114 trace_bcache_btree_node_alloc_fail(c);
cafe5635
KO
1115 return b;
1116}
1117
c5aa4a31 1118static struct btree *bch_btree_node_alloc(struct cache_set *c,
2452cc89
SP
1119 struct btree_op *op, int level,
1120 struct btree *parent)
c5aa4a31 1121{
2452cc89 1122 return __bch_btree_node_alloc(c, op, level, op != NULL, parent);
c5aa4a31
SP
1123}
1124
0a63b66d
KO
1125static struct btree *btree_node_alloc_replacement(struct btree *b,
1126 struct btree_op *op)
cafe5635 1127{
2452cc89 1128 struct btree *n = bch_btree_node_alloc(b->c, op, b->level, b->parent);
1fae7cf0 1129
67539e85 1130 if (!IS_ERR_OR_NULL(n)) {
2a285686 1131 mutex_lock(&n->write_lock);
89ebb4a2 1132 bch_btree_sort_into(&b->keys, &n->keys, &b->c->sort);
67539e85 1133 bkey_copy_key(&n->key, &b->key);
2a285686 1134 mutex_unlock(&n->write_lock);
67539e85 1135 }
cafe5635
KO
1136
1137 return n;
1138}
1139
8835c123
KO
1140static void make_btree_freeing_key(struct btree *b, struct bkey *k)
1141{
6f10f7d1 1142 unsigned int i;
8835c123 1143
05335cff
KO
1144 mutex_lock(&b->c->bucket_lock);
1145
1146 atomic_inc(&b->c->prio_blocked);
1147
8835c123
KO
1148 bkey_copy(k, &b->key);
1149 bkey_copy_key(k, &ZERO_KEY);
1150
05335cff
KO
1151 for (i = 0; i < KEY_PTRS(k); i++)
1152 SET_PTR_GEN(k, i,
1153 bch_inc_gen(PTR_CACHE(b->c, &b->key, i),
1154 PTR_BUCKET(b->c, &b->key, i)));
8835c123 1155
05335cff 1156 mutex_unlock(&b->c->bucket_lock);
8835c123
KO
1157}
1158
78365411
KO
1159static int btree_check_reserve(struct btree *b, struct btree_op *op)
1160{
1161 struct cache_set *c = b->c;
1162 struct cache *ca;
6f10f7d1 1163 unsigned int i, reserve = (c->root->level - b->level) * 2 + 1;
78365411
KO
1164
1165 mutex_lock(&c->bucket_lock);
1166
1167 for_each_cache(ca, c, i)
1168 if (fifo_used(&ca->free[RESERVE_BTREE]) < reserve) {
1169 if (op)
0a63b66d 1170 prepare_to_wait(&c->btree_cache_wait, &op->wait,
78365411 1171 TASK_UNINTERRUPTIBLE);
0a63b66d
KO
1172 mutex_unlock(&c->bucket_lock);
1173 return -EINTR;
78365411
KO
1174 }
1175
1176 mutex_unlock(&c->bucket_lock);
0a63b66d
KO
1177
1178 return mca_cannibalize_lock(b->c, op);
78365411
KO
1179}
1180
cafe5635
KO
1181/* Garbage collection */
1182
487dded8
KO
1183static uint8_t __bch_btree_mark_key(struct cache_set *c, int level,
1184 struct bkey *k)
cafe5635
KO
1185{
1186 uint8_t stale = 0;
6f10f7d1 1187 unsigned int i;
cafe5635
KO
1188 struct bucket *g;
1189
1190 /*
1191 * ptr_invalid() can't return true for the keys that mark btree nodes as
1192 * freed, but since ptr_bad() returns true we'll never actually use them
1193 * for anything and thus we don't want mark their pointers here
1194 */
1195 if (!bkey_cmp(k, &ZERO_KEY))
1196 return stale;
1197
1198 for (i = 0; i < KEY_PTRS(k); i++) {
1199 if (!ptr_available(c, k, i))
1200 continue;
1201
1202 g = PTR_BUCKET(c, k, i);
1203
3a2fd9d5
KO
1204 if (gen_after(g->last_gc, PTR_GEN(k, i)))
1205 g->last_gc = PTR_GEN(k, i);
cafe5635
KO
1206
1207 if (ptr_stale(c, k, i)) {
1208 stale = max(stale, ptr_stale(c, k, i));
1209 continue;
1210 }
1211
1212 cache_bug_on(GC_MARK(g) &&
1213 (GC_MARK(g) == GC_MARK_METADATA) != (level != 0),
1214 c, "inconsistent ptrs: mark = %llu, level = %i",
1215 GC_MARK(g), level);
1216
1217 if (level)
1218 SET_GC_MARK(g, GC_MARK_METADATA);
1219 else if (KEY_DIRTY(k))
1220 SET_GC_MARK(g, GC_MARK_DIRTY);
4fe6a816
KO
1221 else if (!GC_MARK(g))
1222 SET_GC_MARK(g, GC_MARK_RECLAIMABLE);
cafe5635
KO
1223
1224 /* guard against overflow */
6f10f7d1 1225 SET_GC_SECTORS_USED(g, min_t(unsigned int,
cafe5635 1226 GC_SECTORS_USED(g) + KEY_SIZE(k),
94717447 1227 MAX_GC_SECTORS_USED));
cafe5635
KO
1228
1229 BUG_ON(!GC_SECTORS_USED(g));
1230 }
1231
1232 return stale;
1233}
1234
1235#define btree_mark_key(b, k) __bch_btree_mark_key(b->c, b->level, k)
1236
487dded8
KO
1237void bch_initial_mark_key(struct cache_set *c, int level, struct bkey *k)
1238{
6f10f7d1 1239 unsigned int i;
487dded8
KO
1240
1241 for (i = 0; i < KEY_PTRS(k); i++)
1242 if (ptr_available(c, k, i) &&
1243 !ptr_stale(c, k, i)) {
1244 struct bucket *b = PTR_BUCKET(c, k, i);
1245
1246 b->gen = PTR_GEN(k, i);
1247
1248 if (level && bkey_cmp(k, &ZERO_KEY))
1249 b->prio = BTREE_PRIO;
1250 else if (!level && b->prio == BTREE_PRIO)
1251 b->prio = INITIAL_PRIO;
1252 }
1253
1254 __bch_btree_mark_key(c, level, k);
1255}
1256
d44c2f9e
TJ
1257void bch_update_bucket_in_use(struct cache_set *c, struct gc_stat *stats)
1258{
1259 stats->in_use = (c->nbuckets - c->avail_nbuckets) * 100 / c->nbuckets;
1260}
1261
a1f0358b 1262static bool btree_gc_mark_node(struct btree *b, struct gc_stat *gc)
cafe5635
KO
1263{
1264 uint8_t stale = 0;
6f10f7d1 1265 unsigned int keys = 0, good_keys = 0;
cafe5635
KO
1266 struct bkey *k;
1267 struct btree_iter iter;
1268 struct bset_tree *t;
1269
1270 gc->nodes++;
1271
c052dd9a 1272 for_each_key_filter(&b->keys, k, &iter, bch_ptr_invalid) {
cafe5635 1273 stale = max(stale, btree_mark_key(b, k));
a1f0358b 1274 keys++;
cafe5635 1275
a85e968e 1276 if (bch_ptr_bad(&b->keys, k))
cafe5635
KO
1277 continue;
1278
cafe5635
KO
1279 gc->key_bytes += bkey_u64s(k);
1280 gc->nkeys++;
a1f0358b 1281 good_keys++;
cafe5635
KO
1282
1283 gc->data += KEY_SIZE(k);
cafe5635
KO
1284 }
1285
a85e968e 1286 for (t = b->keys.set; t <= &b->keys.set[b->keys.nsets]; t++)
cafe5635 1287 btree_bug_on(t->size &&
a85e968e 1288 bset_written(&b->keys, t) &&
cafe5635
KO
1289 bkey_cmp(&b->key, &t->end) < 0,
1290 b, "found short btree key in gc");
1291
a1f0358b
KO
1292 if (b->c->gc_always_rewrite)
1293 return true;
cafe5635 1294
a1f0358b
KO
1295 if (stale > 10)
1296 return true;
cafe5635 1297
a1f0358b
KO
1298 if ((keys - good_keys) * 2 > keys)
1299 return true;
cafe5635 1300
a1f0358b 1301 return false;
cafe5635
KO
1302}
1303
a1f0358b 1304#define GC_MERGE_NODES 4U
cafe5635
KO
1305
1306struct gc_merge_info {
1307 struct btree *b;
6f10f7d1 1308 unsigned int keys;
cafe5635
KO
1309};
1310
fc2d5988
CL
1311static int bch_btree_insert_node(struct btree *b, struct btree_op *op,
1312 struct keylist *insert_keys,
1313 atomic_t *journal_ref,
1314 struct bkey *replace_key);
a1f0358b
KO
1315
1316static int btree_gc_coalesce(struct btree *b, struct btree_op *op,
0a63b66d 1317 struct gc_stat *gc, struct gc_merge_info *r)
cafe5635 1318{
6f10f7d1 1319 unsigned int i, nodes = 0, keys = 0, blocks;
a1f0358b 1320 struct btree *new_nodes[GC_MERGE_NODES];
0a63b66d 1321 struct keylist keylist;
b54d6934 1322 struct closure cl;
a1f0358b 1323 struct bkey *k;
b54d6934 1324
0a63b66d
KO
1325 bch_keylist_init(&keylist);
1326
1327 if (btree_check_reserve(b, NULL))
1328 return 0;
1329
a1f0358b 1330 memset(new_nodes, 0, sizeof(new_nodes));
b54d6934 1331 closure_init_stack(&cl);
cafe5635 1332
a1f0358b 1333 while (nodes < GC_MERGE_NODES && !IS_ERR_OR_NULL(r[nodes].b))
cafe5635
KO
1334 keys += r[nodes++].keys;
1335
1336 blocks = btree_default_blocks(b->c) * 2 / 3;
1337
1338 if (nodes < 2 ||
a85e968e 1339 __set_blocks(b->keys.set[0].data, keys,
ee811287 1340 block_bytes(b->c)) > blocks * (nodes - 1))
a1f0358b 1341 return 0;
cafe5635 1342
a1f0358b 1343 for (i = 0; i < nodes; i++) {
0a63b66d 1344 new_nodes[i] = btree_node_alloc_replacement(r[i].b, NULL);
a1f0358b
KO
1345 if (IS_ERR_OR_NULL(new_nodes[i]))
1346 goto out_nocoalesce;
cafe5635
KO
1347 }
1348
0a63b66d
KO
1349 /*
1350 * We have to check the reserve here, after we've allocated our new
1351 * nodes, to make sure the insert below will succeed - we also check
1352 * before as an optimization to potentially avoid a bunch of expensive
1353 * allocs/sorts
1354 */
1355 if (btree_check_reserve(b, NULL))
1356 goto out_nocoalesce;
1357
2a285686
KO
1358 for (i = 0; i < nodes; i++)
1359 mutex_lock(&new_nodes[i]->write_lock);
1360
cafe5635 1361 for (i = nodes - 1; i > 0; --i) {
ee811287
KO
1362 struct bset *n1 = btree_bset_first(new_nodes[i]);
1363 struct bset *n2 = btree_bset_first(new_nodes[i - 1]);
cafe5635
KO
1364 struct bkey *k, *last = NULL;
1365
1366 keys = 0;
1367
a1f0358b
KO
1368 if (i > 1) {
1369 for (k = n2->start;
fafff81c 1370 k < bset_bkey_last(n2);
a1f0358b
KO
1371 k = bkey_next(k)) {
1372 if (__set_blocks(n1, n1->keys + keys +
ee811287
KO
1373 bkey_u64s(k),
1374 block_bytes(b->c)) > blocks)
a1f0358b
KO
1375 break;
1376
1377 last = k;
1378 keys += bkey_u64s(k);
1379 }
1380 } else {
cafe5635
KO
1381 /*
1382 * Last node we're not getting rid of - we're getting
1383 * rid of the node at r[0]. Have to try and fit all of
1384 * the remaining keys into this node; we can't ensure
1385 * they will always fit due to rounding and variable
1386 * length keys (shouldn't be possible in practice,
1387 * though)
1388 */
a1f0358b 1389 if (__set_blocks(n1, n1->keys + n2->keys,
ee811287
KO
1390 block_bytes(b->c)) >
1391 btree_blocks(new_nodes[i]))
be23e837 1392 goto out_unlock_nocoalesce;
cafe5635
KO
1393
1394 keys = n2->keys;
a1f0358b 1395 /* Take the key of the node we're getting rid of */
cafe5635 1396 last = &r->b->key;
a1f0358b 1397 }
cafe5635 1398
ee811287
KO
1399 BUG_ON(__set_blocks(n1, n1->keys + keys, block_bytes(b->c)) >
1400 btree_blocks(new_nodes[i]));
cafe5635 1401
a1f0358b
KO
1402 if (last)
1403 bkey_copy_key(&new_nodes[i]->key, last);
cafe5635 1404
fafff81c 1405 memcpy(bset_bkey_last(n1),
cafe5635 1406 n2->start,
fafff81c 1407 (void *) bset_bkey_idx(n2, keys) - (void *) n2->start);
cafe5635
KO
1408
1409 n1->keys += keys;
a1f0358b 1410 r[i].keys = n1->keys;
cafe5635
KO
1411
1412 memmove(n2->start,
fafff81c
KO
1413 bset_bkey_idx(n2, keys),
1414 (void *) bset_bkey_last(n2) -
1415 (void *) bset_bkey_idx(n2, keys));
cafe5635
KO
1416
1417 n2->keys -= keys;
1418
0a63b66d 1419 if (__bch_keylist_realloc(&keylist,
085d2a3d 1420 bkey_u64s(&new_nodes[i]->key)))
be23e837 1421 goto out_unlock_nocoalesce;
a1f0358b
KO
1422
1423 bch_btree_node_write(new_nodes[i], &cl);
0a63b66d 1424 bch_keylist_add(&keylist, &new_nodes[i]->key);
cafe5635
KO
1425 }
1426
2a285686
KO
1427 for (i = 0; i < nodes; i++)
1428 mutex_unlock(&new_nodes[i]->write_lock);
1429
05335cff
KO
1430 closure_sync(&cl);
1431
1432 /* We emptied out this node */
1433 BUG_ON(btree_bset_first(new_nodes[0])->keys);
1434 btree_node_free(new_nodes[0]);
1435 rw_unlock(true, new_nodes[0]);
400ffaa2 1436 new_nodes[0] = NULL;
05335cff 1437
a1f0358b 1438 for (i = 0; i < nodes; i++) {
0a63b66d 1439 if (__bch_keylist_realloc(&keylist, bkey_u64s(&r[i].b->key)))
a1f0358b 1440 goto out_nocoalesce;
cafe5635 1441
0a63b66d
KO
1442 make_btree_freeing_key(r[i].b, keylist.top);
1443 bch_keylist_push(&keylist);
a1f0358b 1444 }
cafe5635 1445
0a63b66d
KO
1446 bch_btree_insert_node(b, op, &keylist, NULL, NULL);
1447 BUG_ON(!bch_keylist_empty(&keylist));
a1f0358b
KO
1448
1449 for (i = 0; i < nodes; i++) {
1450 btree_node_free(r[i].b);
1451 rw_unlock(true, r[i].b);
1452
1453 r[i].b = new_nodes[i];
1454 }
1455
a1f0358b
KO
1456 memmove(r, r + 1, sizeof(r[0]) * (nodes - 1));
1457 r[nodes - 1].b = ERR_PTR(-EINTR);
1458
1459 trace_bcache_btree_gc_coalesce(nodes);
cafe5635 1460 gc->nodes--;
cafe5635 1461
0a63b66d
KO
1462 bch_keylist_free(&keylist);
1463
a1f0358b
KO
1464 /* Invalidated our iterator */
1465 return -EINTR;
1466
be23e837
ZL
1467out_unlock_nocoalesce:
1468 for (i = 0; i < nodes; i++)
1469 mutex_unlock(&new_nodes[i]->write_lock);
1470
a1f0358b
KO
1471out_nocoalesce:
1472 closure_sync(&cl);
1473
0a63b66d 1474 while ((k = bch_keylist_pop(&keylist)))
a1f0358b
KO
1475 if (!bkey_cmp(k, &ZERO_KEY))
1476 atomic_dec(&b->c->prio_blocked);
f16277ca 1477 bch_keylist_free(&keylist);
a1f0358b
KO
1478
1479 for (i = 0; i < nodes; i++)
1480 if (!IS_ERR_OR_NULL(new_nodes[i])) {
1481 btree_node_free(new_nodes[i]);
1482 rw_unlock(true, new_nodes[i]);
1483 }
1484 return 0;
cafe5635
KO
1485}
1486
0a63b66d
KO
1487static int btree_gc_rewrite_node(struct btree *b, struct btree_op *op,
1488 struct btree *replace)
1489{
1490 struct keylist keys;
1491 struct btree *n;
1492
1493 if (btree_check_reserve(b, NULL))
1494 return 0;
1495
1496 n = btree_node_alloc_replacement(replace, NULL);
1497
1498 /* recheck reserve after allocating replacement node */
1499 if (btree_check_reserve(b, NULL)) {
1500 btree_node_free(n);
1501 rw_unlock(true, n);
1502 return 0;
1503 }
1504
1505 bch_btree_node_write_sync(n);
1506
1507 bch_keylist_init(&keys);
1508 bch_keylist_add(&keys, &n->key);
1509
1510 make_btree_freeing_key(replace, keys.top);
1511 bch_keylist_push(&keys);
1512
1513 bch_btree_insert_node(b, op, &keys, NULL, NULL);
1514 BUG_ON(!bch_keylist_empty(&keys));
1515
1516 btree_node_free(replace);
1517 rw_unlock(true, n);
1518
1519 /* Invalidated our iterator */
1520 return -EINTR;
1521}
1522
6f10f7d1 1523static unsigned int btree_gc_count_keys(struct btree *b)
cafe5635 1524{
a1f0358b
KO
1525 struct bkey *k;
1526 struct btree_iter iter;
6f10f7d1 1527 unsigned int ret = 0;
cafe5635 1528
c052dd9a 1529 for_each_key_filter(&b->keys, k, &iter, bch_ptr_bad)
a1f0358b
KO
1530 ret += bkey_u64s(k);
1531
1532 return ret;
1533}
cafe5635 1534
7f4a59de
TJ
1535static size_t btree_gc_min_nodes(struct cache_set *c)
1536{
1537 size_t min_nodes;
1538
1539 /*
1540 * Since incremental GC would stop 100ms when front
1541 * side I/O comes, so when there are many btree nodes,
1542 * if GC only processes constant (100) nodes each time,
1543 * GC would last a long time, and the front side I/Os
1544 * would run out of the buckets (since no new bucket
1545 * can be allocated during GC), and be blocked again.
1546 * So GC should not process constant nodes, but varied
1547 * nodes according to the number of btree nodes, which
1548 * realized by dividing GC into constant(100) times,
1549 * so when there are many btree nodes, GC can process
1550 * more nodes each time, otherwise, GC will process less
1551 * nodes each time (but no less than MIN_GC_NODES)
1552 */
1553 min_nodes = c->gc_stats.nodes / MAX_GC_TIMES;
1554 if (min_nodes < MIN_GC_NODES)
1555 min_nodes = MIN_GC_NODES;
1556
1557 return min_nodes;
1558}
1559
1560
a1f0358b
KO
1561static int btree_gc_recurse(struct btree *b, struct btree_op *op,
1562 struct closure *writes, struct gc_stat *gc)
1563{
a1f0358b
KO
1564 int ret = 0;
1565 bool should_rewrite;
a1f0358b 1566 struct bkey *k;
a1f0358b 1567 struct btree_iter iter;
cafe5635 1568 struct gc_merge_info r[GC_MERGE_NODES];
2a285686 1569 struct gc_merge_info *i, *last = r + ARRAY_SIZE(r) - 1;
cafe5635 1570
c052dd9a 1571 bch_btree_iter_init(&b->keys, &iter, &b->c->gc_done);
cafe5635 1572
2a285686
KO
1573 for (i = r; i < r + ARRAY_SIZE(r); i++)
1574 i->b = ERR_PTR(-EINTR);
cafe5635 1575
a1f0358b 1576 while (1) {
a85e968e 1577 k = bch_btree_iter_next_filter(&iter, &b->keys, bch_ptr_bad);
a1f0358b 1578 if (k) {
0a63b66d 1579 r->b = bch_btree_node_get(b->c, op, k, b->level - 1,
2452cc89 1580 true, b);
a1f0358b
KO
1581 if (IS_ERR(r->b)) {
1582 ret = PTR_ERR(r->b);
1583 break;
1584 }
1585
1586 r->keys = btree_gc_count_keys(r->b);
1587
0a63b66d 1588 ret = btree_gc_coalesce(b, op, gc, r);
a1f0358b
KO
1589 if (ret)
1590 break;
cafe5635
KO
1591 }
1592
a1f0358b
KO
1593 if (!last->b)
1594 break;
cafe5635 1595
a1f0358b
KO
1596 if (!IS_ERR(last->b)) {
1597 should_rewrite = btree_gc_mark_node(last->b, gc);
0a63b66d
KO
1598 if (should_rewrite) {
1599 ret = btree_gc_rewrite_node(b, op, last->b);
1600 if (ret)
a1f0358b 1601 break;
a1f0358b
KO
1602 }
1603
1604 if (last->b->level) {
1605 ret = btree_gc_recurse(last->b, op, writes, gc);
1606 if (ret)
1607 break;
1608 }
cafe5635 1609
a1f0358b
KO
1610 bkey_copy_key(&b->c->gc_done, &last->b->key);
1611
1612 /*
1613 * Must flush leaf nodes before gc ends, since replace
1614 * operations aren't journalled
1615 */
2a285686 1616 mutex_lock(&last->b->write_lock);
a1f0358b
KO
1617 if (btree_node_dirty(last->b))
1618 bch_btree_node_write(last->b, writes);
2a285686 1619 mutex_unlock(&last->b->write_lock);
a1f0358b
KO
1620 rw_unlock(true, last->b);
1621 }
1622
1623 memmove(r + 1, r, sizeof(r[0]) * (GC_MERGE_NODES - 1));
1624 r->b = NULL;
cafe5635 1625
5c25c4fc 1626 if (atomic_read(&b->c->search_inflight) &&
7f4a59de 1627 gc->nodes >= gc->nodes_pre + btree_gc_min_nodes(b->c)) {
5c25c4fc
TJ
1628 gc->nodes_pre = gc->nodes;
1629 ret = -EAGAIN;
1630 break;
1631 }
1632
cafe5635
KO
1633 if (need_resched()) {
1634 ret = -EAGAIN;
1635 break;
1636 }
cafe5635
KO
1637 }
1638
2a285686
KO
1639 for (i = r; i < r + ARRAY_SIZE(r); i++)
1640 if (!IS_ERR_OR_NULL(i->b)) {
1641 mutex_lock(&i->b->write_lock);
1642 if (btree_node_dirty(i->b))
1643 bch_btree_node_write(i->b, writes);
1644 mutex_unlock(&i->b->write_lock);
1645 rw_unlock(true, i->b);
a1f0358b 1646 }
cafe5635 1647
cafe5635
KO
1648 return ret;
1649}
1650
1651static int bch_btree_gc_root(struct btree *b, struct btree_op *op,
1652 struct closure *writes, struct gc_stat *gc)
1653{
1654 struct btree *n = NULL;
a1f0358b
KO
1655 int ret = 0;
1656 bool should_rewrite;
cafe5635 1657
a1f0358b
KO
1658 should_rewrite = btree_gc_mark_node(b, gc);
1659 if (should_rewrite) {
0a63b66d 1660 n = btree_node_alloc_replacement(b, NULL);
cafe5635 1661
a1f0358b
KO
1662 if (!IS_ERR_OR_NULL(n)) {
1663 bch_btree_node_write_sync(n);
2a285686 1664
a1f0358b
KO
1665 bch_btree_set_root(n);
1666 btree_node_free(b);
1667 rw_unlock(true, n);
cafe5635 1668
a1f0358b
KO
1669 return -EINTR;
1670 }
1671 }
cafe5635 1672
487dded8
KO
1673 __bch_btree_mark_key(b->c, b->level + 1, &b->key);
1674
a1f0358b
KO
1675 if (b->level) {
1676 ret = btree_gc_recurse(b, op, writes, gc);
1677 if (ret)
1678 return ret;
cafe5635
KO
1679 }
1680
a1f0358b
KO
1681 bkey_copy_key(&b->c->gc_done, &b->key);
1682
cafe5635
KO
1683 return ret;
1684}
1685
1686static void btree_gc_start(struct cache_set *c)
1687{
1688 struct cache *ca;
1689 struct bucket *b;
6f10f7d1 1690 unsigned int i;
cafe5635
KO
1691
1692 if (!c->gc_mark_valid)
1693 return;
1694
1695 mutex_lock(&c->bucket_lock);
1696
1697 c->gc_mark_valid = 0;
1698 c->gc_done = ZERO_KEY;
1699
1700 for_each_cache(ca, c, i)
1701 for_each_bucket(b, ca) {
3a2fd9d5 1702 b->last_gc = b->gen;
29ebf465 1703 if (!atomic_read(&b->pin)) {
4fe6a816 1704 SET_GC_MARK(b, 0);
29ebf465
KO
1705 SET_GC_SECTORS_USED(b, 0);
1706 }
cafe5635
KO
1707 }
1708
cafe5635
KO
1709 mutex_unlock(&c->bucket_lock);
1710}
1711
d44c2f9e 1712static void bch_btree_gc_finish(struct cache_set *c)
cafe5635 1713{
cafe5635
KO
1714 struct bucket *b;
1715 struct cache *ca;
6f10f7d1 1716 unsigned int i;
cafe5635
KO
1717
1718 mutex_lock(&c->bucket_lock);
1719
1720 set_gc_sectors(c);
1721 c->gc_mark_valid = 1;
1722 c->need_gc = 0;
1723
cafe5635
KO
1724 for (i = 0; i < KEY_PTRS(&c->uuid_bucket); i++)
1725 SET_GC_MARK(PTR_BUCKET(c, &c->uuid_bucket, i),
1726 GC_MARK_METADATA);
1727
bf0a628a
NS
1728 /* don't reclaim buckets to which writeback keys point */
1729 rcu_read_lock();
2831231d 1730 for (i = 0; i < c->devices_max_used; i++) {
bf0a628a
NS
1731 struct bcache_device *d = c->devices[i];
1732 struct cached_dev *dc;
1733 struct keybuf_key *w, *n;
6f10f7d1 1734 unsigned int j;
bf0a628a
NS
1735
1736 if (!d || UUID_FLASH_ONLY(&c->uuids[i]))
1737 continue;
1738 dc = container_of(d, struct cached_dev, disk);
1739
1740 spin_lock(&dc->writeback_keys.lock);
1741 rbtree_postorder_for_each_entry_safe(w, n,
1742 &dc->writeback_keys.keys, node)
1743 for (j = 0; j < KEY_PTRS(&w->key); j++)
1744 SET_GC_MARK(PTR_BUCKET(c, &w->key, j),
1745 GC_MARK_DIRTY);
1746 spin_unlock(&dc->writeback_keys.lock);
1747 }
1748 rcu_read_unlock();
1749
d44c2f9e 1750 c->avail_nbuckets = 0;
cafe5635
KO
1751 for_each_cache(ca, c, i) {
1752 uint64_t *i;
1753
1754 ca->invalidate_needs_gc = 0;
1755
1756 for (i = ca->sb.d; i < ca->sb.d + ca->sb.keys; i++)
1757 SET_GC_MARK(ca->buckets + *i, GC_MARK_METADATA);
1758
1759 for (i = ca->prio_buckets;
1760 i < ca->prio_buckets + prio_buckets(ca) * 2; i++)
1761 SET_GC_MARK(ca->buckets + *i, GC_MARK_METADATA);
1762
1763 for_each_bucket(b, ca) {
cafe5635
KO
1764 c->need_gc = max(c->need_gc, bucket_gc_gen(b));
1765
4fe6a816
KO
1766 if (atomic_read(&b->pin))
1767 continue;
1768
1769 BUG_ON(!GC_MARK(b) && GC_SECTORS_USED(b));
1770
1771 if (!GC_MARK(b) || GC_MARK(b) == GC_MARK_RECLAIMABLE)
d44c2f9e 1772 c->avail_nbuckets++;
cafe5635
KO
1773 }
1774 }
1775
cafe5635 1776 mutex_unlock(&c->bucket_lock);
cafe5635
KO
1777}
1778
72a44517 1779static void bch_btree_gc(struct cache_set *c)
cafe5635 1780{
cafe5635 1781 int ret;
cafe5635
KO
1782 struct gc_stat stats;
1783 struct closure writes;
1784 struct btree_op op;
cafe5635 1785 uint64_t start_time = local_clock();
57943511 1786
c37511b8 1787 trace_bcache_gc_start(c);
cafe5635
KO
1788
1789 memset(&stats, 0, sizeof(struct gc_stat));
1790 closure_init_stack(&writes);
b54d6934 1791 bch_btree_op_init(&op, SHRT_MAX);
cafe5635
KO
1792
1793 btree_gc_start(c);
1794
771f393e 1795 /* if CACHE_SET_IO_DISABLE set, gc thread should stop too */
a1f0358b 1796 do {
feac1a70 1797 ret = bcache_btree_root(gc_root, c, &op, &writes, &stats);
a1f0358b 1798 closure_sync(&writes);
c5f1e5ad 1799 cond_resched();
cafe5635 1800
5c25c4fc
TJ
1801 if (ret == -EAGAIN)
1802 schedule_timeout_interruptible(msecs_to_jiffies
1803 (GC_SLEEP_MS));
1804 else if (ret)
46f5aa88 1805 pr_warn("gc failed!\n");
771f393e 1806 } while (ret && !test_bit(CACHE_SET_IO_DISABLE, &c->flags));
cafe5635 1807
d44c2f9e 1808 bch_btree_gc_finish(c);
57943511
KO
1809 wake_up_allocators(c);
1810
169ef1cf 1811 bch_time_stats_update(&c->btree_gc_time, start_time);
cafe5635
KO
1812
1813 stats.key_bytes *= sizeof(uint64_t);
cafe5635 1814 stats.data <<= 9;
d44c2f9e 1815 bch_update_bucket_in_use(c, &stats);
cafe5635 1816 memcpy(&c->gc_stats, &stats, sizeof(struct gc_stat));
cafe5635 1817
c37511b8 1818 trace_bcache_gc_end(c);
cafe5635 1819
72a44517
KO
1820 bch_moving_gc(c);
1821}
1822
be628be0 1823static bool gc_should_run(struct cache_set *c)
72a44517 1824{
a1f0358b 1825 struct cache *ca;
6f10f7d1 1826 unsigned int i;
72a44517 1827
be628be0
KO
1828 for_each_cache(ca, c, i)
1829 if (ca->invalidate_needs_gc)
1830 return true;
72a44517 1831
be628be0
KO
1832 if (atomic_read(&c->sectors_to_gc) < 0)
1833 return true;
72a44517 1834
be628be0
KO
1835 return false;
1836}
a1f0358b 1837
be628be0
KO
1838static int bch_gc_thread(void *arg)
1839{
1840 struct cache_set *c = arg;
a1f0358b 1841
be628be0
KO
1842 while (1) {
1843 wait_event_interruptible(c->gc_wait,
771f393e
CL
1844 kthread_should_stop() ||
1845 test_bit(CACHE_SET_IO_DISABLE, &c->flags) ||
1846 gc_should_run(c));
a1f0358b 1847
771f393e
CL
1848 if (kthread_should_stop() ||
1849 test_bit(CACHE_SET_IO_DISABLE, &c->flags))
be628be0
KO
1850 break;
1851
1852 set_gc_sectors(c);
1853 bch_btree_gc(c);
72a44517
KO
1854 }
1855
771f393e 1856 wait_for_kthread_stop();
72a44517 1857 return 0;
cafe5635
KO
1858}
1859
72a44517 1860int bch_gc_thread_start(struct cache_set *c)
cafe5635 1861{
be628be0 1862 c->gc_thread = kthread_run(bch_gc_thread, c, "bcache_gc");
9d134117 1863 return PTR_ERR_OR_ZERO(c->gc_thread);
cafe5635
KO
1864}
1865
1866/* Initial partial gc */
1867
487dded8 1868static int bch_btree_check_recurse(struct btree *b, struct btree_op *op)
cafe5635 1869{
50310164 1870 int ret = 0;
50310164 1871 struct bkey *k, *p = NULL;
cafe5635
KO
1872 struct btree_iter iter;
1873
487dded8
KO
1874 for_each_key_filter(&b->keys, k, &iter, bch_ptr_invalid)
1875 bch_initial_mark_key(b->c, b->level, k);
cafe5635 1876
487dded8 1877 bch_initial_mark_key(b->c, b->level + 1, &b->key);
cafe5635
KO
1878
1879 if (b->level) {
c052dd9a 1880 bch_btree_iter_init(&b->keys, &iter, NULL);
cafe5635 1881
50310164 1882 do {
a85e968e
KO
1883 k = bch_btree_iter_next_filter(&iter, &b->keys,
1884 bch_ptr_bad);
7f4a59de 1885 if (k) {
2452cc89 1886 btree_node_prefetch(b, k);
7f4a59de
TJ
1887 /*
1888 * initiallize c->gc_stats.nodes
1889 * for incremental GC
1890 */
1891 b->c->gc_stats.nodes++;
1892 }
cafe5635 1893
50310164 1894 if (p)
feac1a70 1895 ret = bcache_btree(check_recurse, p, b, op);
cafe5635 1896
50310164
KO
1897 p = k;
1898 } while (p && !ret);
cafe5635
KO
1899 }
1900
487dded8 1901 return ret;
cafe5635
KO
1902}
1903
8e710227
CL
1904
1905static int bch_btree_check_thread(void *arg)
1906{
1907 int ret;
1908 struct btree_check_info *info = arg;
1909 struct btree_check_state *check_state = info->state;
1910 struct cache_set *c = check_state->c;
1911 struct btree_iter iter;
1912 struct bkey *k, *p;
1913 int cur_idx, prev_idx, skip_nr;
8e710227
CL
1914
1915 k = p = NULL;
8e710227
CL
1916 cur_idx = prev_idx = 0;
1917 ret = 0;
1918
1919 /* root node keys are checked before thread created */
1920 bch_btree_iter_init(&c->root->keys, &iter, NULL);
1921 k = bch_btree_iter_next_filter(&iter, &c->root->keys, bch_ptr_bad);
1922 BUG_ON(!k);
1923
1924 p = k;
1925 while (k) {
1926 /*
1927 * Fetch a root node key index, skip the keys which
1928 * should be fetched by other threads, then check the
1929 * sub-tree indexed by the fetched key.
1930 */
1931 spin_lock(&check_state->idx_lock);
1932 cur_idx = check_state->key_idx;
1933 check_state->key_idx++;
1934 spin_unlock(&check_state->idx_lock);
1935
1936 skip_nr = cur_idx - prev_idx;
1937
1938 while (skip_nr) {
1939 k = bch_btree_iter_next_filter(&iter,
1940 &c->root->keys,
1941 bch_ptr_bad);
1942 if (k)
1943 p = k;
1944 else {
1945 /*
1946 * No more keys to check in root node,
1947 * current checking threads are enough,
1948 * stop creating more.
1949 */
1950 atomic_set(&check_state->enough, 1);
1951 /* Update check_state->enough earlier */
eb9b6666 1952 smp_mb__after_atomic();
8e710227
CL
1953 goto out;
1954 }
1955 skip_nr--;
1956 cond_resched();
1957 }
1958
1959 if (p) {
1960 struct btree_op op;
1961
1962 btree_node_prefetch(c->root, p);
1963 c->gc_stats.nodes++;
1964 bch_btree_op_init(&op, 0);
1965 ret = bcache_btree(check_recurse, p, c->root, &op);
1966 if (ret)
1967 goto out;
1968 }
1969 p = NULL;
1970 prev_idx = cur_idx;
1971 cond_resched();
1972 }
1973
1974out:
1975 info->result = ret;
1976 /* update check_state->started among all CPUs */
eb9b6666 1977 smp_mb__before_atomic();
8e710227
CL
1978 if (atomic_dec_and_test(&check_state->started))
1979 wake_up(&check_state->wait);
1980
1981 return ret;
1982}
1983
1984
1985
1986static int bch_btree_chkthread_nr(void)
1987{
1988 int n = num_online_cpus()/2;
1989
1990 if (n == 0)
1991 n = 1;
1992 else if (n > BCH_BTR_CHKTHREAD_MAX)
1993 n = BCH_BTR_CHKTHREAD_MAX;
1994
1995 return n;
1996}
1997
c18536a7 1998int bch_btree_check(struct cache_set *c)
cafe5635 1999{
8e710227
CL
2000 int ret = 0;
2001 int i;
2002 struct bkey *k = NULL;
2003 struct btree_iter iter;
2004 struct btree_check_state *check_state;
2005 char name[32];
cafe5635 2006
8e710227
CL
2007 /* check and mark root node keys */
2008 for_each_key_filter(&c->root->keys, k, &iter, bch_ptr_invalid)
2009 bch_initial_mark_key(c, c->root->level, k);
2010
2011 bch_initial_mark_key(c, c->root->level + 1, &c->root->key);
2012
2013 if (c->root->level == 0)
2014 return 0;
2015
2016 check_state = kzalloc(sizeof(struct btree_check_state), GFP_KERNEL);
2017 if (!check_state)
2018 return -ENOMEM;
2019
2020 check_state->c = c;
2021 check_state->total_threads = bch_btree_chkthread_nr();
2022 check_state->key_idx = 0;
2023 spin_lock_init(&check_state->idx_lock);
2024 atomic_set(&check_state->started, 0);
2025 atomic_set(&check_state->enough, 0);
2026 init_waitqueue_head(&check_state->wait);
cafe5635 2027
8e710227
CL
2028 /*
2029 * Run multiple threads to check btree nodes in parallel,
2030 * if check_state->enough is non-zero, it means current
2031 * running check threads are enough, unncessary to create
2032 * more.
2033 */
2034 for (i = 0; i < check_state->total_threads; i++) {
2035 /* fetch latest check_state->enough earlier */
eb9b6666 2036 smp_mb__before_atomic();
8e710227
CL
2037 if (atomic_read(&check_state->enough))
2038 break;
2039
2040 check_state->infos[i].result = 0;
2041 check_state->infos[i].state = check_state;
2042 snprintf(name, sizeof(name), "bch_btrchk[%u]", i);
2043 atomic_inc(&check_state->started);
2044
2045 check_state->infos[i].thread =
2046 kthread_run(bch_btree_check_thread,
2047 &check_state->infos[i],
2048 name);
2049 if (IS_ERR(check_state->infos[i].thread)) {
46f5aa88 2050 pr_err("fails to run thread bch_btrchk[%d]\n", i);
8e710227
CL
2051 for (--i; i >= 0; i--)
2052 kthread_stop(check_state->infos[i].thread);
2053 ret = -ENOMEM;
2054 goto out;
2055 }
2056 }
2057
2058 wait_event_interruptible(check_state->wait,
2059 atomic_read(&check_state->started) == 0 ||
2060 test_bit(CACHE_SET_IO_DISABLE, &c->flags));
2061
2062 for (i = 0; i < check_state->total_threads; i++) {
2063 if (check_state->infos[i].result) {
2064 ret = check_state->infos[i].result;
2065 goto out;
2066 }
2067 }
2068
2069out:
2070 kfree(check_state);
2071 return ret;
cafe5635
KO
2072}
2073
2531d9ee
KO
2074void bch_initial_gc_finish(struct cache_set *c)
2075{
2076 struct cache *ca;
2077 struct bucket *b;
6f10f7d1 2078 unsigned int i;
2531d9ee
KO
2079
2080 bch_btree_gc_finish(c);
2081
2082 mutex_lock(&c->bucket_lock);
2083
2084 /*
2085 * We need to put some unused buckets directly on the prio freelist in
2086 * order to get the allocator thread started - it needs freed buckets in
2087 * order to rewrite the prios and gens, and it needs to rewrite prios
2088 * and gens in order to free buckets.
2089 *
2090 * This is only safe for buckets that have no live data in them, which
2091 * there should always be some of.
2092 */
2093 for_each_cache(ca, c, i) {
2094 for_each_bucket(b, ca) {
682811b3
TJ
2095 if (fifo_full(&ca->free[RESERVE_PRIO]) &&
2096 fifo_full(&ca->free[RESERVE_BTREE]))
2531d9ee
KO
2097 break;
2098
2099 if (bch_can_invalidate_bucket(ca, b) &&
2100 !GC_MARK(b)) {
2101 __bch_invalidate_one_bucket(ca, b);
682811b3
TJ
2102 if (!fifo_push(&ca->free[RESERVE_PRIO],
2103 b - ca->buckets))
2104 fifo_push(&ca->free[RESERVE_BTREE],
2105 b - ca->buckets);
2531d9ee
KO
2106 }
2107 }
2108 }
2109
2110 mutex_unlock(&c->bucket_lock);
2111}
2112
cafe5635
KO
2113/* Btree insertion */
2114
829a60b9
KO
2115static bool btree_insert_key(struct btree *b, struct bkey *k,
2116 struct bkey *replace_key)
cafe5635 2117{
6f10f7d1 2118 unsigned int status;
cafe5635
KO
2119
2120 BUG_ON(bkey_cmp(k, &b->key) > 0);
1fa8455d 2121
829a60b9
KO
2122 status = bch_btree_insert_key(&b->keys, k, replace_key);
2123 if (status != BTREE_INSERT_STATUS_NO_INSERT) {
2124 bch_check_keys(&b->keys, "%u for %s", status,
2125 replace_key ? "replace" : "insert");
cafe5635 2126
829a60b9
KO
2127 trace_bcache_btree_insert_key(b, k, replace_key != NULL,
2128 status);
2129 return true;
2130 } else
2131 return false;
cafe5635
KO
2132}
2133
59158fde
KO
2134static size_t insert_u64s_remaining(struct btree *b)
2135{
3572324a 2136 long ret = bch_btree_keys_u64s_remaining(&b->keys);
59158fde
KO
2137
2138 /*
2139 * Might land in the middle of an existing extent and have to split it
2140 */
2141 if (b->keys.ops->is_extents)
2142 ret -= KEY_MAX_U64S;
2143
2144 return max(ret, 0L);
2145}
2146
26c949f8 2147static bool bch_btree_insert_keys(struct btree *b, struct btree_op *op,
1b207d80
KO
2148 struct keylist *insert_keys,
2149 struct bkey *replace_key)
cafe5635
KO
2150{
2151 bool ret = false;
dc9d98d6 2152 int oldsize = bch_count_data(&b->keys);
cafe5635 2153
26c949f8 2154 while (!bch_keylist_empty(insert_keys)) {
c2f95ae2 2155 struct bkey *k = insert_keys->keys;
26c949f8 2156
59158fde 2157 if (bkey_u64s(k) > insert_u64s_remaining(b))
403b6cde
KO
2158 break;
2159
2160 if (bkey_cmp(k, &b->key) <= 0) {
3a3b6a4e
KO
2161 if (!b->level)
2162 bkey_put(b->c, k);
26c949f8 2163
829a60b9 2164 ret |= btree_insert_key(b, k, replace_key);
26c949f8
KO
2165 bch_keylist_pop_front(insert_keys);
2166 } else if (bkey_cmp(&START_KEY(k), &b->key) < 0) {
26c949f8 2167 BKEY_PADDED(key) temp;
c2f95ae2 2168 bkey_copy(&temp.key, insert_keys->keys);
26c949f8
KO
2169
2170 bch_cut_back(&b->key, &temp.key);
c2f95ae2 2171 bch_cut_front(&b->key, insert_keys->keys);
26c949f8 2172
829a60b9 2173 ret |= btree_insert_key(b, &temp.key, replace_key);
26c949f8
KO
2174 break;
2175 } else {
2176 break;
2177 }
cafe5635
KO
2178 }
2179
829a60b9
KO
2180 if (!ret)
2181 op->insert_collision = true;
2182
403b6cde
KO
2183 BUG_ON(!bch_keylist_empty(insert_keys) && b->level);
2184
dc9d98d6 2185 BUG_ON(bch_count_data(&b->keys) < oldsize);
cafe5635
KO
2186 return ret;
2187}
2188
26c949f8
KO
2189static int btree_split(struct btree *b, struct btree_op *op,
2190 struct keylist *insert_keys,
1b207d80 2191 struct bkey *replace_key)
cafe5635 2192{
d6fd3b11 2193 bool split;
cafe5635
KO
2194 struct btree *n1, *n2 = NULL, *n3 = NULL;
2195 uint64_t start_time = local_clock();
b54d6934 2196 struct closure cl;
17e21a9f 2197 struct keylist parent_keys;
b54d6934
KO
2198
2199 closure_init_stack(&cl);
17e21a9f 2200 bch_keylist_init(&parent_keys);
cafe5635 2201
0a63b66d
KO
2202 if (btree_check_reserve(b, op)) {
2203 if (!b->level)
2204 return -EINTR;
2205 else
2206 WARN(1, "insufficient reserve for split\n");
2207 }
78365411 2208
0a63b66d 2209 n1 = btree_node_alloc_replacement(b, op);
cafe5635
KO
2210 if (IS_ERR(n1))
2211 goto err;
2212
ee811287
KO
2213 split = set_blocks(btree_bset_first(n1),
2214 block_bytes(n1->c)) > (btree_blocks(b) * 4) / 5;
cafe5635 2215
cafe5635 2216 if (split) {
6f10f7d1 2217 unsigned int keys = 0;
cafe5635 2218
ee811287 2219 trace_bcache_btree_node_split(b, btree_bset_first(n1)->keys);
c37511b8 2220
2452cc89 2221 n2 = bch_btree_node_alloc(b->c, op, b->level, b->parent);
cafe5635
KO
2222 if (IS_ERR(n2))
2223 goto err_free1;
2224
d6fd3b11 2225 if (!b->parent) {
2452cc89 2226 n3 = bch_btree_node_alloc(b->c, op, b->level + 1, NULL);
cafe5635
KO
2227 if (IS_ERR(n3))
2228 goto err_free2;
2229 }
2230
2a285686
KO
2231 mutex_lock(&n1->write_lock);
2232 mutex_lock(&n2->write_lock);
2233
1b207d80 2234 bch_btree_insert_keys(n1, op, insert_keys, replace_key);
cafe5635 2235
d6fd3b11
KO
2236 /*
2237 * Has to be a linear search because we don't have an auxiliary
cafe5635
KO
2238 * search tree yet
2239 */
2240
ee811287
KO
2241 while (keys < (btree_bset_first(n1)->keys * 3) / 5)
2242 keys += bkey_u64s(bset_bkey_idx(btree_bset_first(n1),
fafff81c 2243 keys));
cafe5635 2244
fafff81c 2245 bkey_copy_key(&n1->key,
ee811287
KO
2246 bset_bkey_idx(btree_bset_first(n1), keys));
2247 keys += bkey_u64s(bset_bkey_idx(btree_bset_first(n1), keys));
cafe5635 2248
ee811287
KO
2249 btree_bset_first(n2)->keys = btree_bset_first(n1)->keys - keys;
2250 btree_bset_first(n1)->keys = keys;
cafe5635 2251
ee811287
KO
2252 memcpy(btree_bset_first(n2)->start,
2253 bset_bkey_last(btree_bset_first(n1)),
2254 btree_bset_first(n2)->keys * sizeof(uint64_t));
cafe5635
KO
2255
2256 bkey_copy_key(&n2->key, &b->key);
2257
17e21a9f 2258 bch_keylist_add(&parent_keys, &n2->key);
b54d6934 2259 bch_btree_node_write(n2, &cl);
2a285686 2260 mutex_unlock(&n2->write_lock);
cafe5635 2261 rw_unlock(true, n2);
c37511b8 2262 } else {
ee811287 2263 trace_bcache_btree_node_compact(b, btree_bset_first(n1)->keys);
c37511b8 2264
2a285686 2265 mutex_lock(&n1->write_lock);
1b207d80 2266 bch_btree_insert_keys(n1, op, insert_keys, replace_key);
c37511b8 2267 }
cafe5635 2268
17e21a9f 2269 bch_keylist_add(&parent_keys, &n1->key);
b54d6934 2270 bch_btree_node_write(n1, &cl);
2a285686 2271 mutex_unlock(&n1->write_lock);
cafe5635
KO
2272
2273 if (n3) {
d6fd3b11 2274 /* Depth increases, make a new root */
2a285686 2275 mutex_lock(&n3->write_lock);
cafe5635 2276 bkey_copy_key(&n3->key, &MAX_KEY);
17e21a9f 2277 bch_btree_insert_keys(n3, op, &parent_keys, NULL);
b54d6934 2278 bch_btree_node_write(n3, &cl);
2a285686 2279 mutex_unlock(&n3->write_lock);
cafe5635 2280
b54d6934 2281 closure_sync(&cl);
cafe5635
KO
2282 bch_btree_set_root(n3);
2283 rw_unlock(true, n3);
d6fd3b11
KO
2284 } else if (!b->parent) {
2285 /* Root filled up but didn't need to be split */
b54d6934 2286 closure_sync(&cl);
cafe5635
KO
2287 bch_btree_set_root(n1);
2288 } else {
17e21a9f 2289 /* Split a non root node */
b54d6934 2290 closure_sync(&cl);
17e21a9f
KO
2291 make_btree_freeing_key(b, parent_keys.top);
2292 bch_keylist_push(&parent_keys);
2293
17e21a9f
KO
2294 bch_btree_insert_node(b->parent, op, &parent_keys, NULL, NULL);
2295 BUG_ON(!bch_keylist_empty(&parent_keys));
cafe5635
KO
2296 }
2297
05335cff 2298 btree_node_free(b);
cafe5635 2299 rw_unlock(true, n1);
cafe5635 2300
169ef1cf 2301 bch_time_stats_update(&b->c->btree_split_time, start_time);
cafe5635
KO
2302
2303 return 0;
2304err_free2:
5f5837d2 2305 bkey_put(b->c, &n2->key);
e8e1d468 2306 btree_node_free(n2);
cafe5635
KO
2307 rw_unlock(true, n2);
2308err_free1:
5f5837d2 2309 bkey_put(b->c, &n1->key);
e8e1d468 2310 btree_node_free(n1);
cafe5635
KO
2311 rw_unlock(true, n1);
2312err:
0a63b66d 2313 WARN(1, "bcache: btree split failed (level %u)", b->level);
5f5837d2 2314
cafe5635
KO
2315 if (n3 == ERR_PTR(-EAGAIN) ||
2316 n2 == ERR_PTR(-EAGAIN) ||
2317 n1 == ERR_PTR(-EAGAIN))
2318 return -EAGAIN;
2319
cafe5635
KO
2320 return -ENOMEM;
2321}
2322
26c949f8 2323static int bch_btree_insert_node(struct btree *b, struct btree_op *op,
c18536a7 2324 struct keylist *insert_keys,
1b207d80
KO
2325 atomic_t *journal_ref,
2326 struct bkey *replace_key)
cafe5635 2327{
2a285686
KO
2328 struct closure cl;
2329
17e21a9f
KO
2330 BUG_ON(b->level && replace_key);
2331
2a285686
KO
2332 closure_init_stack(&cl);
2333
2334 mutex_lock(&b->write_lock);
2335
2336 if (write_block(b) != btree_bset_last(b) &&
2337 b->keys.last_set_unwritten)
2338 bch_btree_init_next(b); /* just wrote a set */
2339
59158fde 2340 if (bch_keylist_nkeys(insert_keys) > insert_u64s_remaining(b)) {
2a285686
KO
2341 mutex_unlock(&b->write_lock);
2342 goto split;
2343 }
3b3e9e50 2344
2a285686 2345 BUG_ON(write_block(b) != btree_bset_last(b));
cafe5635 2346
2a285686
KO
2347 if (bch_btree_insert_keys(b, op, insert_keys, replace_key)) {
2348 if (!b->level)
2349 bch_btree_leaf_dirty(b, journal_ref);
2350 else
2351 bch_btree_node_write(b, &cl);
2352 }
17e21a9f 2353
2a285686
KO
2354 mutex_unlock(&b->write_lock);
2355
2356 /* wait for btree node write if necessary, after unlock */
2357 closure_sync(&cl);
2358
2359 return 0;
2360split:
2361 if (current->bio_list) {
2362 op->lock = b->c->root->level + 1;
2363 return -EAGAIN;
2364 } else if (op->lock <= b->c->root->level) {
2365 op->lock = b->c->root->level + 1;
2366 return -EINTR;
2367 } else {
2368 /* Invalidated all iterators */
2369 int ret = btree_split(b, op, insert_keys, replace_key);
2370
2371 if (bch_keylist_empty(insert_keys))
2372 return 0;
2373 else if (!ret)
2374 return -EINTR;
2375 return ret;
17e21a9f 2376 }
26c949f8 2377}
cafe5635 2378
e7c590eb
KO
2379int bch_btree_insert_check_key(struct btree *b, struct btree_op *op,
2380 struct bkey *check_key)
2381{
2382 int ret = -EINTR;
2383 uint64_t btree_ptr = b->key.ptr[0];
2384 unsigned long seq = b->seq;
2385 struct keylist insert;
2386 bool upgrade = op->lock == -1;
2387
2388 bch_keylist_init(&insert);
2389
2390 if (upgrade) {
2391 rw_unlock(false, b);
2392 rw_lock(true, b, b->level);
2393
2394 if (b->key.ptr[0] != btree_ptr ||
c63ca787 2395 b->seq != seq + 1) {
fd01991d 2396 op->lock = b->level;
e7c590eb 2397 goto out;
c63ca787 2398 }
e7c590eb
KO
2399 }
2400
2401 SET_KEY_PTRS(check_key, 1);
2402 get_random_bytes(&check_key->ptr[0], sizeof(uint64_t));
2403
2404 SET_PTR_DEV(check_key, 0, PTR_CHECK_DEV);
2405
2406 bch_keylist_add(&insert, check_key);
2407
1b207d80 2408 ret = bch_btree_insert_node(b, op, &insert, NULL, NULL);
e7c590eb
KO
2409
2410 BUG_ON(!ret && !bch_keylist_empty(&insert));
2411out:
2412 if (upgrade)
2413 downgrade_write(&b->lock);
2414 return ret;
2415}
2416
cc7b8819
KO
2417struct btree_insert_op {
2418 struct btree_op op;
2419 struct keylist *keys;
2420 atomic_t *journal_ref;
2421 struct bkey *replace_key;
2422};
cafe5635 2423
08239ca2 2424static int btree_insert_fn(struct btree_op *b_op, struct btree *b)
cc7b8819
KO
2425{
2426 struct btree_insert_op *op = container_of(b_op,
2427 struct btree_insert_op, op);
cafe5635 2428
cc7b8819
KO
2429 int ret = bch_btree_insert_node(b, &op->op, op->keys,
2430 op->journal_ref, op->replace_key);
2431 if (ret && !bch_keylist_empty(op->keys))
2432 return ret;
2433 else
2434 return MAP_DONE;
cafe5635
KO
2435}
2436
cc7b8819
KO
2437int bch_btree_insert(struct cache_set *c, struct keylist *keys,
2438 atomic_t *journal_ref, struct bkey *replace_key)
cafe5635 2439{
cc7b8819 2440 struct btree_insert_op op;
cafe5635 2441 int ret = 0;
cafe5635 2442
cc7b8819 2443 BUG_ON(current->bio_list);
4f3d4014 2444 BUG_ON(bch_keylist_empty(keys));
cafe5635 2445
cc7b8819
KO
2446 bch_btree_op_init(&op.op, 0);
2447 op.keys = keys;
2448 op.journal_ref = journal_ref;
2449 op.replace_key = replace_key;
cafe5635 2450
cc7b8819
KO
2451 while (!ret && !bch_keylist_empty(keys)) {
2452 op.op.lock = 0;
2453 ret = bch_btree_map_leaf_nodes(&op.op, c,
2454 &START_KEY(keys->keys),
2455 btree_insert_fn);
2456 }
cafe5635 2457
cc7b8819
KO
2458 if (ret) {
2459 struct bkey *k;
cafe5635 2460
46f5aa88 2461 pr_err("error %i\n", ret);
cafe5635 2462
cc7b8819 2463 while ((k = bch_keylist_pop(keys)))
3a3b6a4e 2464 bkey_put(c, k);
cc7b8819
KO
2465 } else if (op.op.insert_collision)
2466 ret = -ESRCH;
6054c6d4 2467
cafe5635
KO
2468 return ret;
2469}
2470
2471void bch_btree_set_root(struct btree *b)
2472{
6f10f7d1 2473 unsigned int i;
e49c7c37
KO
2474 struct closure cl;
2475
2476 closure_init_stack(&cl);
cafe5635 2477
c37511b8
KO
2478 trace_bcache_btree_set_root(b);
2479
cafe5635
KO
2480 BUG_ON(!b->written);
2481
2482 for (i = 0; i < KEY_PTRS(&b->key); i++)
2483 BUG_ON(PTR_BUCKET(b->c, &b->key, i)->prio != BTREE_PRIO);
2484
2485 mutex_lock(&b->c->bucket_lock);
2486 list_del_init(&b->list);
2487 mutex_unlock(&b->c->bucket_lock);
2488
2489 b->c->root = b;
cafe5635 2490
e49c7c37
KO
2491 bch_journal_meta(b->c, &cl);
2492 closure_sync(&cl);
cafe5635
KO
2493}
2494
48dad8ba
KO
2495/* Map across nodes or keys */
2496
2497static int bch_btree_map_nodes_recurse(struct btree *b, struct btree_op *op,
2498 struct bkey *from,
2499 btree_map_nodes_fn *fn, int flags)
2500{
2501 int ret = MAP_CONTINUE;
2502
2503 if (b->level) {
2504 struct bkey *k;
2505 struct btree_iter iter;
2506
c052dd9a 2507 bch_btree_iter_init(&b->keys, &iter, from);
48dad8ba 2508
a85e968e 2509 while ((k = bch_btree_iter_next_filter(&iter, &b->keys,
48dad8ba 2510 bch_ptr_bad))) {
feac1a70 2511 ret = bcache_btree(map_nodes_recurse, k, b,
48dad8ba
KO
2512 op, from, fn, flags);
2513 from = NULL;
2514
2515 if (ret != MAP_CONTINUE)
2516 return ret;
2517 }
2518 }
2519
2520 if (!b->level || flags == MAP_ALL_NODES)
2521 ret = fn(op, b);
2522
2523 return ret;
2524}
2525
2526int __bch_btree_map_nodes(struct btree_op *op, struct cache_set *c,
2527 struct bkey *from, btree_map_nodes_fn *fn, int flags)
2528{
feac1a70 2529 return bcache_btree_root(map_nodes_recurse, c, op, from, fn, flags);
48dad8ba
KO
2530}
2531
253a99d9 2532int bch_btree_map_keys_recurse(struct btree *b, struct btree_op *op,
48dad8ba
KO
2533 struct bkey *from, btree_map_keys_fn *fn,
2534 int flags)
2535{
2536 int ret = MAP_CONTINUE;
2537 struct bkey *k;
2538 struct btree_iter iter;
2539
c052dd9a 2540 bch_btree_iter_init(&b->keys, &iter, from);
48dad8ba 2541
a85e968e 2542 while ((k = bch_btree_iter_next_filter(&iter, &b->keys, bch_ptr_bad))) {
48dad8ba
KO
2543 ret = !b->level
2544 ? fn(op, b, k)
feac1a70
CL
2545 : bcache_btree(map_keys_recurse, k,
2546 b, op, from, fn, flags);
48dad8ba
KO
2547 from = NULL;
2548
2549 if (ret != MAP_CONTINUE)
2550 return ret;
2551 }
2552
2553 if (!b->level && (flags & MAP_END_KEY))
2554 ret = fn(op, b, &KEY(KEY_INODE(&b->key),
2555 KEY_OFFSET(&b->key), 0));
2556
2557 return ret;
2558}
2559
2560int bch_btree_map_keys(struct btree_op *op, struct cache_set *c,
2561 struct bkey *from, btree_map_keys_fn *fn, int flags)
2562{
feac1a70 2563 return bcache_btree_root(map_keys_recurse, c, op, from, fn, flags);
48dad8ba
KO
2564}
2565
cafe5635
KO
2566/* Keybuf code */
2567
2568static inline int keybuf_cmp(struct keybuf_key *l, struct keybuf_key *r)
2569{
2570 /* Overlapping keys compare equal */
2571 if (bkey_cmp(&l->key, &START_KEY(&r->key)) <= 0)
2572 return -1;
2573 if (bkey_cmp(&START_KEY(&l->key), &r->key) >= 0)
2574 return 1;
2575 return 0;
2576}
2577
2578static inline int keybuf_nonoverlapping_cmp(struct keybuf_key *l,
2579 struct keybuf_key *r)
2580{
2581 return clamp_t(int64_t, bkey_cmp(&l->key, &r->key), -1, 1);
2582}
2583
48dad8ba
KO
2584struct refill {
2585 struct btree_op op;
6f10f7d1 2586 unsigned int nr_found;
48dad8ba
KO
2587 struct keybuf *buf;
2588 struct bkey *end;
2589 keybuf_pred_fn *pred;
2590};
cafe5635 2591
48dad8ba
KO
2592static int refill_keybuf_fn(struct btree_op *op, struct btree *b,
2593 struct bkey *k)
2594{
2595 struct refill *refill = container_of(op, struct refill, op);
2596 struct keybuf *buf = refill->buf;
2597 int ret = MAP_CONTINUE;
cafe5635 2598
2d6cb6ed 2599 if (bkey_cmp(k, refill->end) > 0) {
48dad8ba
KO
2600 ret = MAP_DONE;
2601 goto out;
2602 }
cafe5635 2603
48dad8ba
KO
2604 if (!KEY_SIZE(k)) /* end key */
2605 goto out;
cafe5635 2606
48dad8ba
KO
2607 if (refill->pred(buf, k)) {
2608 struct keybuf_key *w;
cafe5635 2609
48dad8ba 2610 spin_lock(&buf->lock);
cafe5635 2611
48dad8ba
KO
2612 w = array_alloc(&buf->freelist);
2613 if (!w) {
2614 spin_unlock(&buf->lock);
2615 return MAP_DONE;
2616 }
cafe5635 2617
48dad8ba
KO
2618 w->private = NULL;
2619 bkey_copy(&w->key, k);
cafe5635 2620
48dad8ba
KO
2621 if (RB_INSERT(&buf->keys, w, node, keybuf_cmp))
2622 array_free(&buf->freelist, w);
48a915a8
KO
2623 else
2624 refill->nr_found++;
cafe5635 2625
48dad8ba
KO
2626 if (array_freelist_empty(&buf->freelist))
2627 ret = MAP_DONE;
cafe5635 2628
48dad8ba 2629 spin_unlock(&buf->lock);
cafe5635 2630 }
48dad8ba
KO
2631out:
2632 buf->last_scanned = *k;
2633 return ret;
cafe5635
KO
2634}
2635
2636void bch_refill_keybuf(struct cache_set *c, struct keybuf *buf,
72c27061 2637 struct bkey *end, keybuf_pred_fn *pred)
cafe5635
KO
2638{
2639 struct bkey start = buf->last_scanned;
48dad8ba 2640 struct refill refill;
cafe5635
KO
2641
2642 cond_resched();
2643
b54d6934 2644 bch_btree_op_init(&refill.op, -1);
48a915a8
KO
2645 refill.nr_found = 0;
2646 refill.buf = buf;
2647 refill.end = end;
2648 refill.pred = pred;
48dad8ba
KO
2649
2650 bch_btree_map_keys(&refill.op, c, &buf->last_scanned,
2651 refill_keybuf_fn, MAP_END_KEY);
cafe5635 2652
48a915a8
KO
2653 trace_bcache_keyscan(refill.nr_found,
2654 KEY_INODE(&start), KEY_OFFSET(&start),
2655 KEY_INODE(&buf->last_scanned),
2656 KEY_OFFSET(&buf->last_scanned));
cafe5635
KO
2657
2658 spin_lock(&buf->lock);
2659
2660 if (!RB_EMPTY_ROOT(&buf->keys)) {
2661 struct keybuf_key *w;
1fae7cf0 2662
cafe5635
KO
2663 w = RB_FIRST(&buf->keys, struct keybuf_key, node);
2664 buf->start = START_KEY(&w->key);
2665
2666 w = RB_LAST(&buf->keys, struct keybuf_key, node);
2667 buf->end = w->key;
2668 } else {
2669 buf->start = MAX_KEY;
2670 buf->end = MAX_KEY;
2671 }
2672
2673 spin_unlock(&buf->lock);
2674}
2675
2676static void __bch_keybuf_del(struct keybuf *buf, struct keybuf_key *w)
2677{
2678 rb_erase(&w->node, &buf->keys);
2679 array_free(&buf->freelist, w);
2680}
2681
2682void bch_keybuf_del(struct keybuf *buf, struct keybuf_key *w)
2683{
2684 spin_lock(&buf->lock);
2685 __bch_keybuf_del(buf, w);
2686 spin_unlock(&buf->lock);
2687}
2688
2689bool bch_keybuf_check_overlapping(struct keybuf *buf, struct bkey *start,
2690 struct bkey *end)
2691{
2692 bool ret = false;
2693 struct keybuf_key *p, *w, s;
1fae7cf0 2694
cafe5635
KO
2695 s.key = *start;
2696
2697 if (bkey_cmp(end, &buf->start) <= 0 ||
2698 bkey_cmp(start, &buf->end) >= 0)
2699 return false;
2700
2701 spin_lock(&buf->lock);
2702 w = RB_GREATER(&buf->keys, s, node, keybuf_nonoverlapping_cmp);
2703
2704 while (w && bkey_cmp(&START_KEY(&w->key), end) < 0) {
2705 p = w;
2706 w = RB_NEXT(w, node);
2707
2708 if (p->private)
2709 ret = true;
2710 else
2711 __bch_keybuf_del(buf, p);
2712 }
2713
2714 spin_unlock(&buf->lock);
2715 return ret;
2716}
2717
2718struct keybuf_key *bch_keybuf_next(struct keybuf *buf)
2719{
2720 struct keybuf_key *w;
1fae7cf0 2721
cafe5635
KO
2722 spin_lock(&buf->lock);
2723
2724 w = RB_FIRST(&buf->keys, struct keybuf_key, node);
2725
2726 while (w && w->private)
2727 w = RB_NEXT(w, node);
2728
2729 if (w)
2730 w->private = ERR_PTR(-EINTR);
2731
2732 spin_unlock(&buf->lock);
2733 return w;
2734}
2735
2736struct keybuf_key *bch_keybuf_next_rescan(struct cache_set *c,
48dad8ba
KO
2737 struct keybuf *buf,
2738 struct bkey *end,
2739 keybuf_pred_fn *pred)
cafe5635
KO
2740{
2741 struct keybuf_key *ret;
2742
2743 while (1) {
2744 ret = bch_keybuf_next(buf);
2745 if (ret)
2746 break;
2747
2748 if (bkey_cmp(&buf->last_scanned, end) >= 0) {
46f5aa88 2749 pr_debug("scan finished\n");
cafe5635
KO
2750 break;
2751 }
2752
72c27061 2753 bch_refill_keybuf(c, buf, end, pred);
cafe5635
KO
2754 }
2755
2756 return ret;
2757}
2758
72c27061 2759void bch_keybuf_init(struct keybuf *buf)
cafe5635 2760{
cafe5635
KO
2761 buf->last_scanned = MAX_KEY;
2762 buf->keys = RB_ROOT;
2763
2764 spin_lock_init(&buf->lock);
2765 array_allocator_init(&buf->freelist);
2766}