Commit | Line | Data |
---|---|---|
cafe5635 KO |
1 | /* |
2 | * Copyright (C) 2010 Kent Overstreet <kent.overstreet@gmail.com> | |
3 | * | |
4 | * Uses a block device as cache for other block devices; optimized for SSDs. | |
5 | * All allocation is done in buckets, which should match the erase block size | |
6 | * of the device. | |
7 | * | |
8 | * Buckets containing cached data are kept on a heap sorted by priority; | |
9 | * bucket priority is increased on cache hit, and periodically all the buckets | |
10 | * on the heap have their priority scaled down. This currently is just used as | |
11 | * an LRU but in the future should allow for more intelligent heuristics. | |
12 | * | |
13 | * Buckets have an 8 bit counter; freeing is accomplished by incrementing the | |
14 | * counter. Garbage collection is used to remove stale pointers. | |
15 | * | |
16 | * Indexing is done via a btree; nodes are not necessarily fully sorted, rather | |
17 | * as keys are inserted we only sort the pages that have not yet been written. | |
18 | * When garbage collection is run, we resort the entire node. | |
19 | * | |
20 | * All configuration is done via sysfs; see Documentation/bcache.txt. | |
21 | */ | |
22 | ||
23 | #include "bcache.h" | |
24 | #include "btree.h" | |
25 | #include "debug.h" | |
26 | #include "request.h" | |
279afbad | 27 | #include "writeback.h" |
cafe5635 KO |
28 | |
29 | #include <linux/slab.h> | |
30 | #include <linux/bitops.h> | |
31 | #include <linux/hash.h> | |
cd953ed0 | 32 | #include <linux/prefetch.h> |
cafe5635 KO |
33 | #include <linux/random.h> |
34 | #include <linux/rcupdate.h> | |
35 | #include <trace/events/bcache.h> | |
36 | ||
37 | /* | |
38 | * Todo: | |
39 | * register_bcache: Return errors out to userspace correctly | |
40 | * | |
41 | * Writeback: don't undirty key until after a cache flush | |
42 | * | |
43 | * Create an iterator for key pointers | |
44 | * | |
45 | * On btree write error, mark bucket such that it won't be freed from the cache | |
46 | * | |
47 | * Journalling: | |
48 | * Check for bad keys in replay | |
49 | * Propagate barriers | |
50 | * Refcount journal entries in journal_replay | |
51 | * | |
52 | * Garbage collection: | |
53 | * Finish incremental gc | |
54 | * Gc should free old UUIDs, data for invalid UUIDs | |
55 | * | |
56 | * Provide a way to list backing device UUIDs we have data cached for, and | |
57 | * probably how long it's been since we've seen them, and a way to invalidate | |
58 | * dirty data for devices that will never be attached again | |
59 | * | |
60 | * Keep 1 min/5 min/15 min statistics of how busy a block device has been, so | |
61 | * that based on that and how much dirty data we have we can keep writeback | |
62 | * from being starved | |
63 | * | |
64 | * Add a tracepoint or somesuch to watch for writeback starvation | |
65 | * | |
66 | * When btree depth > 1 and splitting an interior node, we have to make sure | |
67 | * alloc_bucket() cannot fail. This should be true but is not completely | |
68 | * obvious. | |
69 | * | |
70 | * Make sure all allocations get charged to the root cgroup | |
71 | * | |
72 | * Plugging? | |
73 | * | |
74 | * If data write is less than hard sector size of ssd, round up offset in open | |
75 | * bucket to the next whole sector | |
76 | * | |
77 | * Also lookup by cgroup in get_open_bucket() | |
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 | static const char * const op_types[] = { | |
92 | "insert", "replace" | |
93 | }; | |
94 | ||
95 | static const char *op_type(struct btree_op *op) | |
96 | { | |
97 | return op_types[op->type]; | |
98 | } | |
99 | ||
100 | #define MAX_NEED_GC 64 | |
101 | #define MAX_SAVE_PRIO 72 | |
102 | ||
103 | #define PTR_DIRTY_BIT (((uint64_t) 1 << 36)) | |
104 | ||
105 | #define PTR_HASH(c, k) \ | |
106 | (((k)->ptr[0] >> c->bucket_bits) | PTR_GEN(k, 0)) | |
107 | ||
108 | struct workqueue_struct *bch_gc_wq; | |
109 | static struct workqueue_struct *btree_io_wq; | |
110 | ||
111 | void bch_btree_op_init_stack(struct btree_op *op) | |
112 | { | |
113 | memset(op, 0, sizeof(struct btree_op)); | |
114 | closure_init_stack(&op->cl); | |
115 | op->lock = -1; | |
116 | bch_keylist_init(&op->keys); | |
117 | } | |
118 | ||
119 | /* Btree key manipulation */ | |
120 | ||
121 | static void bkey_put(struct cache_set *c, struct bkey *k, int level) | |
122 | { | |
123 | if ((level && KEY_OFFSET(k)) || !level) | |
124 | __bkey_put(c, k); | |
125 | } | |
126 | ||
127 | /* Btree IO */ | |
128 | ||
129 | static uint64_t btree_csum_set(struct btree *b, struct bset *i) | |
130 | { | |
131 | uint64_t crc = b->key.ptr[0]; | |
132 | void *data = (void *) i + 8, *end = end(i); | |
133 | ||
169ef1cf | 134 | crc = bch_crc64_update(crc, data, end - data); |
c19ed23a | 135 | return crc ^ 0xffffffffffffffffULL; |
cafe5635 KO |
136 | } |
137 | ||
57943511 | 138 | void bch_btree_node_read_done(struct btree *b) |
cafe5635 | 139 | { |
cafe5635 | 140 | const char *err = "bad btree header"; |
57943511 KO |
141 | struct bset *i = b->sets[0].data; |
142 | struct btree_iter *iter; | |
cafe5635 | 143 | |
57943511 KO |
144 | iter = mempool_alloc(b->c->fill_iter, GFP_NOWAIT); |
145 | iter->size = b->c->sb.bucket_size / b->c->sb.block_size; | |
cafe5635 KO |
146 | iter->used = 0; |
147 | ||
57943511 | 148 | if (!i->seq) |
cafe5635 KO |
149 | goto err; |
150 | ||
151 | for (; | |
152 | b->written < btree_blocks(b) && i->seq == b->sets[0].data->seq; | |
153 | i = write_block(b)) { | |
154 | err = "unsupported bset version"; | |
155 | if (i->version > BCACHE_BSET_VERSION) | |
156 | goto err; | |
157 | ||
158 | err = "bad btree header"; | |
159 | if (b->written + set_blocks(i, b->c) > btree_blocks(b)) | |
160 | goto err; | |
161 | ||
162 | err = "bad magic"; | |
163 | if (i->magic != bset_magic(b->c)) | |
164 | goto err; | |
165 | ||
166 | err = "bad checksum"; | |
167 | switch (i->version) { | |
168 | case 0: | |
169 | if (i->csum != csum_set(i)) | |
170 | goto err; | |
171 | break; | |
172 | case BCACHE_BSET_VERSION: | |
173 | if (i->csum != btree_csum_set(b, i)) | |
174 | goto err; | |
175 | break; | |
176 | } | |
177 | ||
178 | err = "empty set"; | |
179 | if (i != b->sets[0].data && !i->keys) | |
180 | goto err; | |
181 | ||
182 | bch_btree_iter_push(iter, i->start, end(i)); | |
183 | ||
184 | b->written += set_blocks(i, b->c); | |
185 | } | |
186 | ||
187 | err = "corrupted btree"; | |
188 | for (i = write_block(b); | |
189 | index(i, b) < btree_blocks(b); | |
190 | i = ((void *) i) + block_bytes(b->c)) | |
191 | if (i->seq == b->sets[0].data->seq) | |
192 | goto err; | |
193 | ||
194 | bch_btree_sort_and_fix_extents(b, iter); | |
195 | ||
196 | i = b->sets[0].data; | |
197 | err = "short btree key"; | |
198 | if (b->sets[0].size && | |
199 | bkey_cmp(&b->key, &b->sets[0].end) < 0) | |
200 | goto err; | |
201 | ||
202 | if (b->written < btree_blocks(b)) | |
203 | bch_bset_init_next(b); | |
204 | out: | |
57943511 KO |
205 | mempool_free(iter, b->c->fill_iter); |
206 | return; | |
cafe5635 KO |
207 | err: |
208 | set_btree_node_io_error(b); | |
07e86ccb | 209 | bch_cache_set_error(b->c, "%s at bucket %zu, block %zu, %u keys", |
cafe5635 KO |
210 | err, PTR_BUCKET_NR(b->c, &b->key, 0), |
211 | index(i, b), i->keys); | |
212 | goto out; | |
213 | } | |
214 | ||
57943511 | 215 | static void btree_node_read_endio(struct bio *bio, int error) |
cafe5635 | 216 | { |
57943511 KO |
217 | struct closure *cl = bio->bi_private; |
218 | closure_put(cl); | |
219 | } | |
cafe5635 | 220 | |
57943511 KO |
221 | void bch_btree_node_read(struct btree *b) |
222 | { | |
223 | uint64_t start_time = local_clock(); | |
224 | struct closure cl; | |
225 | struct bio *bio; | |
cafe5635 | 226 | |
c37511b8 KO |
227 | trace_bcache_btree_read(b); |
228 | ||
57943511 | 229 | closure_init_stack(&cl); |
cafe5635 | 230 | |
57943511 KO |
231 | bio = bch_bbio_alloc(b->c); |
232 | bio->bi_rw = REQ_META|READ_SYNC; | |
233 | bio->bi_size = KEY_SIZE(&b->key) << 9; | |
234 | bio->bi_end_io = btree_node_read_endio; | |
235 | bio->bi_private = &cl; | |
cafe5635 | 236 | |
57943511 | 237 | bch_bio_map(bio, b->sets[0].data); |
cafe5635 | 238 | |
57943511 KO |
239 | bch_submit_bbio(bio, b->c, &b->key, 0); |
240 | closure_sync(&cl); | |
cafe5635 | 241 | |
57943511 KO |
242 | if (!test_bit(BIO_UPTODATE, &bio->bi_flags)) |
243 | set_btree_node_io_error(b); | |
244 | ||
245 | bch_bbio_free(bio, b->c); | |
246 | ||
247 | if (btree_node_io_error(b)) | |
248 | goto err; | |
249 | ||
250 | bch_btree_node_read_done(b); | |
251 | ||
252 | spin_lock(&b->c->btree_read_time_lock); | |
253 | bch_time_stats_update(&b->c->btree_read_time, start_time); | |
254 | spin_unlock(&b->c->btree_read_time_lock); | |
255 | ||
256 | return; | |
257 | err: | |
258 | bch_cache_set_error(b->c, "io error reading bucket %lu", | |
259 | PTR_BUCKET_NR(b->c, &b->key, 0)); | |
cafe5635 KO |
260 | } |
261 | ||
262 | static void btree_complete_write(struct btree *b, struct btree_write *w) | |
263 | { | |
264 | if (w->prio_blocked && | |
265 | !atomic_sub_return(w->prio_blocked, &b->c->prio_blocked)) | |
119ba0f8 | 266 | wake_up_allocators(b->c); |
cafe5635 KO |
267 | |
268 | if (w->journal) { | |
269 | atomic_dec_bug(w->journal); | |
270 | __closure_wake_up(&b->c->journal.wait); | |
271 | } | |
272 | ||
cafe5635 KO |
273 | w->prio_blocked = 0; |
274 | w->journal = NULL; | |
cafe5635 KO |
275 | } |
276 | ||
57943511 | 277 | static void __btree_node_write_done(struct closure *cl) |
cafe5635 KO |
278 | { |
279 | struct btree *b = container_of(cl, struct btree, io.cl); | |
280 | struct btree_write *w = btree_prev_write(b); | |
281 | ||
282 | bch_bbio_free(b->bio, b->c); | |
283 | b->bio = NULL; | |
284 | btree_complete_write(b, w); | |
285 | ||
286 | if (btree_node_dirty(b)) | |
287 | queue_delayed_work(btree_io_wq, &b->work, | |
288 | msecs_to_jiffies(30000)); | |
289 | ||
290 | closure_return(cl); | |
291 | } | |
292 | ||
57943511 | 293 | static void btree_node_write_done(struct closure *cl) |
cafe5635 KO |
294 | { |
295 | struct btree *b = container_of(cl, struct btree, io.cl); | |
296 | struct bio_vec *bv; | |
297 | int n; | |
298 | ||
299 | __bio_for_each_segment(bv, b->bio, n, 0) | |
300 | __free_page(bv->bv_page); | |
301 | ||
57943511 | 302 | __btree_node_write_done(cl); |
cafe5635 KO |
303 | } |
304 | ||
57943511 KO |
305 | static void btree_node_write_endio(struct bio *bio, int error) |
306 | { | |
307 | struct closure *cl = bio->bi_private; | |
308 | struct btree *b = container_of(cl, struct btree, io.cl); | |
309 | ||
310 | if (error) | |
311 | set_btree_node_io_error(b); | |
312 | ||
313 | bch_bbio_count_io_errors(b->c, bio, error, "writing btree"); | |
314 | closure_put(cl); | |
315 | } | |
316 | ||
317 | static void do_btree_node_write(struct btree *b) | |
cafe5635 KO |
318 | { |
319 | struct closure *cl = &b->io.cl; | |
320 | struct bset *i = b->sets[b->nsets].data; | |
321 | BKEY_PADDED(key) k; | |
322 | ||
323 | i->version = BCACHE_BSET_VERSION; | |
324 | i->csum = btree_csum_set(b, i); | |
325 | ||
57943511 KO |
326 | BUG_ON(b->bio); |
327 | b->bio = bch_bbio_alloc(b->c); | |
328 | ||
329 | b->bio->bi_end_io = btree_node_write_endio; | |
330 | b->bio->bi_private = &b->io.cl; | |
cafe5635 KO |
331 | b->bio->bi_rw = REQ_META|WRITE_SYNC; |
332 | b->bio->bi_size = set_blocks(i, b->c) * block_bytes(b->c); | |
169ef1cf | 333 | bch_bio_map(b->bio, i); |
cafe5635 KO |
334 | |
335 | bkey_copy(&k.key, &b->key); | |
336 | SET_PTR_OFFSET(&k.key, 0, PTR_OFFSET(&k.key, 0) + bset_offset(b, i)); | |
337 | ||
169ef1cf | 338 | if (!bch_bio_alloc_pages(b->bio, GFP_NOIO)) { |
cafe5635 KO |
339 | int j; |
340 | struct bio_vec *bv; | |
341 | void *base = (void *) ((unsigned long) i & ~(PAGE_SIZE - 1)); | |
342 | ||
343 | bio_for_each_segment(bv, b->bio, j) | |
344 | memcpy(page_address(bv->bv_page), | |
345 | base + j * PAGE_SIZE, PAGE_SIZE); | |
346 | ||
cafe5635 KO |
347 | bch_submit_bbio(b->bio, b->c, &k.key, 0); |
348 | ||
57943511 | 349 | continue_at(cl, btree_node_write_done, NULL); |
cafe5635 KO |
350 | } else { |
351 | b->bio->bi_vcnt = 0; | |
169ef1cf | 352 | bch_bio_map(b->bio, i); |
cafe5635 | 353 | |
cafe5635 KO |
354 | bch_submit_bbio(b->bio, b->c, &k.key, 0); |
355 | ||
356 | closure_sync(cl); | |
57943511 | 357 | __btree_node_write_done(cl); |
cafe5635 KO |
358 | } |
359 | } | |
360 | ||
57943511 | 361 | void bch_btree_node_write(struct btree *b, struct closure *parent) |
cafe5635 KO |
362 | { |
363 | struct bset *i = b->sets[b->nsets].data; | |
364 | ||
c37511b8 KO |
365 | trace_bcache_btree_write(b); |
366 | ||
cafe5635 | 367 | BUG_ON(current->bio_list); |
57943511 KO |
368 | BUG_ON(b->written >= btree_blocks(b)); |
369 | BUG_ON(b->written && !i->keys); | |
370 | BUG_ON(b->sets->data->seq != i->seq); | |
c37511b8 | 371 | bch_check_key_order(b, i); |
cafe5635 | 372 | |
cafe5635 KO |
373 | cancel_delayed_work(&b->work); |
374 | ||
57943511 KO |
375 | /* If caller isn't waiting for write, parent refcount is cache set */ |
376 | closure_lock(&b->io, parent ?: &b->c->cl); | |
377 | ||
cafe5635 KO |
378 | clear_bit(BTREE_NODE_dirty, &b->flags); |
379 | change_bit(BTREE_NODE_write_idx, &b->flags); | |
380 | ||
57943511 | 381 | do_btree_node_write(b); |
cafe5635 | 382 | |
cafe5635 KO |
383 | b->written += set_blocks(i, b->c); |
384 | atomic_long_add(set_blocks(i, b->c) * b->c->sb.block_size, | |
385 | &PTR_CACHE(b->c, &b->key, 0)->btree_sectors_written); | |
386 | ||
387 | bch_btree_sort_lazy(b); | |
388 | ||
389 | if (b->written < btree_blocks(b)) | |
390 | bch_bset_init_next(b); | |
391 | } | |
392 | ||
57943511 | 393 | static void btree_node_write_work(struct work_struct *w) |
cafe5635 KO |
394 | { |
395 | struct btree *b = container_of(to_delayed_work(w), struct btree, work); | |
396 | ||
57943511 | 397 | rw_lock(true, b, b->level); |
cafe5635 KO |
398 | |
399 | if (btree_node_dirty(b)) | |
57943511 KO |
400 | bch_btree_node_write(b, NULL); |
401 | rw_unlock(true, b); | |
cafe5635 KO |
402 | } |
403 | ||
57943511 | 404 | static void bch_btree_leaf_dirty(struct btree *b, struct btree_op *op) |
cafe5635 KO |
405 | { |
406 | struct bset *i = b->sets[b->nsets].data; | |
407 | struct btree_write *w = btree_current_write(b); | |
408 | ||
57943511 KO |
409 | BUG_ON(!b->written); |
410 | BUG_ON(!i->keys); | |
cafe5635 | 411 | |
57943511 KO |
412 | if (!btree_node_dirty(b)) |
413 | queue_delayed_work(btree_io_wq, &b->work, 30 * HZ); | |
cafe5635 | 414 | |
57943511 | 415 | set_btree_node_dirty(b); |
cafe5635 | 416 | |
57943511 | 417 | if (op && op->journal) { |
cafe5635 KO |
418 | if (w->journal && |
419 | journal_pin_cmp(b->c, w, op)) { | |
420 | atomic_dec_bug(w->journal); | |
421 | w->journal = NULL; | |
422 | } | |
423 | ||
424 | if (!w->journal) { | |
425 | w->journal = op->journal; | |
426 | atomic_inc(w->journal); | |
427 | } | |
428 | } | |
429 | ||
cafe5635 | 430 | /* Force write if set is too big */ |
57943511 KO |
431 | if (set_bytes(i) > PAGE_SIZE - 48 && |
432 | !current->bio_list) | |
433 | bch_btree_node_write(b, NULL); | |
cafe5635 KO |
434 | } |
435 | ||
436 | /* | |
437 | * Btree in memory cache - allocation/freeing | |
438 | * mca -> memory cache | |
439 | */ | |
440 | ||
441 | static void mca_reinit(struct btree *b) | |
442 | { | |
443 | unsigned i; | |
444 | ||
445 | b->flags = 0; | |
446 | b->written = 0; | |
447 | b->nsets = 0; | |
448 | ||
449 | for (i = 0; i < MAX_BSETS; i++) | |
450 | b->sets[i].size = 0; | |
451 | /* | |
452 | * Second loop starts at 1 because b->sets[0]->data is the memory we | |
453 | * allocated | |
454 | */ | |
455 | for (i = 1; i < MAX_BSETS; i++) | |
456 | b->sets[i].data = NULL; | |
457 | } | |
458 | ||
459 | #define mca_reserve(c) (((c->root && c->root->level) \ | |
460 | ? c->root->level : 1) * 8 + 16) | |
461 | #define mca_can_free(c) \ | |
462 | max_t(int, 0, c->bucket_cache_used - mca_reserve(c)) | |
463 | ||
464 | static void mca_data_free(struct btree *b) | |
465 | { | |
466 | struct bset_tree *t = b->sets; | |
467 | BUG_ON(!closure_is_unlocked(&b->io.cl)); | |
468 | ||
469 | if (bset_prev_bytes(b) < PAGE_SIZE) | |
470 | kfree(t->prev); | |
471 | else | |
472 | free_pages((unsigned long) t->prev, | |
473 | get_order(bset_prev_bytes(b))); | |
474 | ||
475 | if (bset_tree_bytes(b) < PAGE_SIZE) | |
476 | kfree(t->tree); | |
477 | else | |
478 | free_pages((unsigned long) t->tree, | |
479 | get_order(bset_tree_bytes(b))); | |
480 | ||
481 | free_pages((unsigned long) t->data, b->page_order); | |
482 | ||
483 | t->prev = NULL; | |
484 | t->tree = NULL; | |
485 | t->data = NULL; | |
486 | list_move(&b->list, &b->c->btree_cache_freed); | |
487 | b->c->bucket_cache_used--; | |
488 | } | |
489 | ||
490 | static void mca_bucket_free(struct btree *b) | |
491 | { | |
492 | BUG_ON(btree_node_dirty(b)); | |
493 | ||
494 | b->key.ptr[0] = 0; | |
495 | hlist_del_init_rcu(&b->hash); | |
496 | list_move(&b->list, &b->c->btree_cache_freeable); | |
497 | } | |
498 | ||
499 | static unsigned btree_order(struct bkey *k) | |
500 | { | |
501 | return ilog2(KEY_SIZE(k) / PAGE_SECTORS ?: 1); | |
502 | } | |
503 | ||
504 | static void mca_data_alloc(struct btree *b, struct bkey *k, gfp_t gfp) | |
505 | { | |
506 | struct bset_tree *t = b->sets; | |
507 | BUG_ON(t->data); | |
508 | ||
509 | b->page_order = max_t(unsigned, | |
510 | ilog2(b->c->btree_pages), | |
511 | btree_order(k)); | |
512 | ||
513 | t->data = (void *) __get_free_pages(gfp, b->page_order); | |
514 | if (!t->data) | |
515 | goto err; | |
516 | ||
517 | t->tree = bset_tree_bytes(b) < PAGE_SIZE | |
518 | ? kmalloc(bset_tree_bytes(b), gfp) | |
519 | : (void *) __get_free_pages(gfp, get_order(bset_tree_bytes(b))); | |
520 | if (!t->tree) | |
521 | goto err; | |
522 | ||
523 | t->prev = bset_prev_bytes(b) < PAGE_SIZE | |
524 | ? kmalloc(bset_prev_bytes(b), gfp) | |
525 | : (void *) __get_free_pages(gfp, get_order(bset_prev_bytes(b))); | |
526 | if (!t->prev) | |
527 | goto err; | |
528 | ||
529 | list_move(&b->list, &b->c->btree_cache); | |
530 | b->c->bucket_cache_used++; | |
531 | return; | |
532 | err: | |
533 | mca_data_free(b); | |
534 | } | |
535 | ||
536 | static struct btree *mca_bucket_alloc(struct cache_set *c, | |
537 | struct bkey *k, gfp_t gfp) | |
538 | { | |
539 | struct btree *b = kzalloc(sizeof(struct btree), gfp); | |
540 | if (!b) | |
541 | return NULL; | |
542 | ||
543 | init_rwsem(&b->lock); | |
544 | lockdep_set_novalidate_class(&b->lock); | |
545 | INIT_LIST_HEAD(&b->list); | |
57943511 | 546 | INIT_DELAYED_WORK(&b->work, btree_node_write_work); |
cafe5635 KO |
547 | b->c = c; |
548 | closure_init_unlocked(&b->io); | |
549 | ||
550 | mca_data_alloc(b, k, gfp); | |
551 | return b; | |
552 | } | |
553 | ||
554 | static int mca_reap(struct btree *b, struct closure *cl, unsigned min_order) | |
555 | { | |
556 | lockdep_assert_held(&b->c->bucket_lock); | |
557 | ||
558 | if (!down_write_trylock(&b->lock)) | |
559 | return -ENOMEM; | |
560 | ||
561 | if (b->page_order < min_order) { | |
562 | rw_unlock(true, b); | |
563 | return -ENOMEM; | |
564 | } | |
565 | ||
566 | BUG_ON(btree_node_dirty(b) && !b->sets[0].data); | |
567 | ||
568 | if (cl && btree_node_dirty(b)) | |
57943511 | 569 | bch_btree_node_write(b, NULL); |
cafe5635 KO |
570 | |
571 | if (cl) | |
572 | closure_wait_event_async(&b->io.wait, cl, | |
573 | atomic_read(&b->io.cl.remaining) == -1); | |
574 | ||
575 | if (btree_node_dirty(b) || | |
576 | !closure_is_unlocked(&b->io.cl) || | |
577 | work_pending(&b->work.work)) { | |
578 | rw_unlock(true, b); | |
579 | return -EAGAIN; | |
580 | } | |
581 | ||
582 | return 0; | |
583 | } | |
584 | ||
585 | static int bch_mca_shrink(struct shrinker *shrink, struct shrink_control *sc) | |
586 | { | |
587 | struct cache_set *c = container_of(shrink, struct cache_set, shrink); | |
588 | struct btree *b, *t; | |
589 | unsigned long i, nr = sc->nr_to_scan; | |
590 | ||
591 | if (c->shrinker_disabled) | |
592 | return 0; | |
593 | ||
594 | if (c->try_harder) | |
595 | return 0; | |
596 | ||
597 | /* | |
598 | * If nr == 0, we're supposed to return the number of items we have | |
599 | * cached. Not allowed to return -1. | |
600 | */ | |
601 | if (!nr) | |
602 | return mca_can_free(c) * c->btree_pages; | |
603 | ||
604 | /* Return -1 if we can't do anything right now */ | |
605 | if (sc->gfp_mask & __GFP_WAIT) | |
606 | mutex_lock(&c->bucket_lock); | |
607 | else if (!mutex_trylock(&c->bucket_lock)) | |
608 | return -1; | |
609 | ||
610 | nr /= c->btree_pages; | |
611 | nr = min_t(unsigned long, nr, mca_can_free(c)); | |
612 | ||
613 | i = 0; | |
614 | list_for_each_entry_safe(b, t, &c->btree_cache_freeable, list) { | |
615 | if (!nr) | |
616 | break; | |
617 | ||
618 | if (++i > 3 && | |
619 | !mca_reap(b, NULL, 0)) { | |
620 | mca_data_free(b); | |
621 | rw_unlock(true, b); | |
622 | --nr; | |
623 | } | |
624 | } | |
625 | ||
626 | /* | |
627 | * Can happen right when we first start up, before we've read in any | |
628 | * btree nodes | |
629 | */ | |
630 | if (list_empty(&c->btree_cache)) | |
631 | goto out; | |
632 | ||
633 | for (i = 0; nr && i < c->bucket_cache_used; i++) { | |
634 | b = list_first_entry(&c->btree_cache, struct btree, list); | |
635 | list_rotate_left(&c->btree_cache); | |
636 | ||
637 | if (!b->accessed && | |
638 | !mca_reap(b, NULL, 0)) { | |
639 | mca_bucket_free(b); | |
640 | mca_data_free(b); | |
641 | rw_unlock(true, b); | |
642 | --nr; | |
643 | } else | |
644 | b->accessed = 0; | |
645 | } | |
646 | out: | |
647 | nr = mca_can_free(c) * c->btree_pages; | |
648 | mutex_unlock(&c->bucket_lock); | |
649 | return nr; | |
650 | } | |
651 | ||
652 | void bch_btree_cache_free(struct cache_set *c) | |
653 | { | |
654 | struct btree *b; | |
655 | struct closure cl; | |
656 | closure_init_stack(&cl); | |
657 | ||
658 | if (c->shrink.list.next) | |
659 | unregister_shrinker(&c->shrink); | |
660 | ||
661 | mutex_lock(&c->bucket_lock); | |
662 | ||
663 | #ifdef CONFIG_BCACHE_DEBUG | |
664 | if (c->verify_data) | |
665 | list_move(&c->verify_data->list, &c->btree_cache); | |
666 | #endif | |
667 | ||
668 | list_splice(&c->btree_cache_freeable, | |
669 | &c->btree_cache); | |
670 | ||
671 | while (!list_empty(&c->btree_cache)) { | |
672 | b = list_first_entry(&c->btree_cache, struct btree, list); | |
673 | ||
674 | if (btree_node_dirty(b)) | |
675 | btree_complete_write(b, btree_current_write(b)); | |
676 | clear_bit(BTREE_NODE_dirty, &b->flags); | |
677 | ||
678 | mca_data_free(b); | |
679 | } | |
680 | ||
681 | while (!list_empty(&c->btree_cache_freed)) { | |
682 | b = list_first_entry(&c->btree_cache_freed, | |
683 | struct btree, list); | |
684 | list_del(&b->list); | |
685 | cancel_delayed_work_sync(&b->work); | |
686 | kfree(b); | |
687 | } | |
688 | ||
689 | mutex_unlock(&c->bucket_lock); | |
690 | } | |
691 | ||
692 | int bch_btree_cache_alloc(struct cache_set *c) | |
693 | { | |
694 | unsigned i; | |
695 | ||
696 | /* XXX: doesn't check for errors */ | |
697 | ||
698 | closure_init_unlocked(&c->gc); | |
699 | ||
700 | for (i = 0; i < mca_reserve(c); i++) | |
701 | mca_bucket_alloc(c, &ZERO_KEY, GFP_KERNEL); | |
702 | ||
703 | list_splice_init(&c->btree_cache, | |
704 | &c->btree_cache_freeable); | |
705 | ||
706 | #ifdef CONFIG_BCACHE_DEBUG | |
707 | mutex_init(&c->verify_lock); | |
708 | ||
709 | c->verify_data = mca_bucket_alloc(c, &ZERO_KEY, GFP_KERNEL); | |
710 | ||
711 | if (c->verify_data && | |
712 | c->verify_data->sets[0].data) | |
713 | list_del_init(&c->verify_data->list); | |
714 | else | |
715 | c->verify_data = NULL; | |
716 | #endif | |
717 | ||
718 | c->shrink.shrink = bch_mca_shrink; | |
719 | c->shrink.seeks = 4; | |
720 | c->shrink.batch = c->btree_pages * 2; | |
721 | register_shrinker(&c->shrink); | |
722 | ||
723 | return 0; | |
724 | } | |
725 | ||
726 | /* Btree in memory cache - hash table */ | |
727 | ||
728 | static struct hlist_head *mca_hash(struct cache_set *c, struct bkey *k) | |
729 | { | |
730 | return &c->bucket_hash[hash_32(PTR_HASH(c, k), BUCKET_HASH_BITS)]; | |
731 | } | |
732 | ||
733 | static struct btree *mca_find(struct cache_set *c, struct bkey *k) | |
734 | { | |
735 | struct btree *b; | |
736 | ||
737 | rcu_read_lock(); | |
738 | hlist_for_each_entry_rcu(b, mca_hash(c, k), hash) | |
739 | if (PTR_HASH(c, &b->key) == PTR_HASH(c, k)) | |
740 | goto out; | |
741 | b = NULL; | |
742 | out: | |
743 | rcu_read_unlock(); | |
744 | return b; | |
745 | } | |
746 | ||
747 | static struct btree *mca_cannibalize(struct cache_set *c, struct bkey *k, | |
748 | int level, struct closure *cl) | |
749 | { | |
750 | int ret = -ENOMEM; | |
751 | struct btree *i; | |
752 | ||
c37511b8 KO |
753 | trace_bcache_btree_cache_cannibalize(c); |
754 | ||
cafe5635 KO |
755 | if (!cl) |
756 | return ERR_PTR(-ENOMEM); | |
757 | ||
758 | /* | |
759 | * Trying to free up some memory - i.e. reuse some btree nodes - may | |
760 | * require initiating IO to flush the dirty part of the node. If we're | |
761 | * running under generic_make_request(), that IO will never finish and | |
762 | * we would deadlock. Returning -EAGAIN causes the cache lookup code to | |
763 | * punt to workqueue and retry. | |
764 | */ | |
765 | if (current->bio_list) | |
766 | return ERR_PTR(-EAGAIN); | |
767 | ||
768 | if (c->try_harder && c->try_harder != cl) { | |
769 | closure_wait_event_async(&c->try_wait, cl, !c->try_harder); | |
770 | return ERR_PTR(-EAGAIN); | |
771 | } | |
772 | ||
cafe5635 KO |
773 | c->try_harder = cl; |
774 | c->try_harder_start = local_clock(); | |
775 | retry: | |
776 | list_for_each_entry_reverse(i, &c->btree_cache, list) { | |
777 | int r = mca_reap(i, cl, btree_order(k)); | |
778 | if (!r) | |
779 | return i; | |
780 | if (r != -ENOMEM) | |
781 | ret = r; | |
782 | } | |
783 | ||
784 | if (ret == -EAGAIN && | |
785 | closure_blocking(cl)) { | |
786 | mutex_unlock(&c->bucket_lock); | |
787 | closure_sync(cl); | |
788 | mutex_lock(&c->bucket_lock); | |
789 | goto retry; | |
790 | } | |
791 | ||
792 | return ERR_PTR(ret); | |
793 | } | |
794 | ||
795 | /* | |
796 | * We can only have one thread cannibalizing other cached btree nodes at a time, | |
797 | * or we'll deadlock. We use an open coded mutex to ensure that, which a | |
798 | * cannibalize_bucket() will take. This means every time we unlock the root of | |
799 | * the btree, we need to release this lock if we have it held. | |
800 | */ | |
801 | void bch_cannibalize_unlock(struct cache_set *c, struct closure *cl) | |
802 | { | |
803 | if (c->try_harder == cl) { | |
169ef1cf | 804 | bch_time_stats_update(&c->try_harder_time, c->try_harder_start); |
cafe5635 KO |
805 | c->try_harder = NULL; |
806 | __closure_wake_up(&c->try_wait); | |
807 | } | |
808 | } | |
809 | ||
810 | static struct btree *mca_alloc(struct cache_set *c, struct bkey *k, | |
811 | int level, struct closure *cl) | |
812 | { | |
813 | struct btree *b; | |
814 | ||
815 | lockdep_assert_held(&c->bucket_lock); | |
816 | ||
817 | if (mca_find(c, k)) | |
818 | return NULL; | |
819 | ||
820 | /* btree_free() doesn't free memory; it sticks the node on the end of | |
821 | * the list. Check if there's any freed nodes there: | |
822 | */ | |
823 | list_for_each_entry(b, &c->btree_cache_freeable, list) | |
824 | if (!mca_reap(b, NULL, btree_order(k))) | |
825 | goto out; | |
826 | ||
827 | /* We never free struct btree itself, just the memory that holds the on | |
828 | * disk node. Check the freed list before allocating a new one: | |
829 | */ | |
830 | list_for_each_entry(b, &c->btree_cache_freed, list) | |
831 | if (!mca_reap(b, NULL, 0)) { | |
832 | mca_data_alloc(b, k, __GFP_NOWARN|GFP_NOIO); | |
833 | if (!b->sets[0].data) | |
834 | goto err; | |
835 | else | |
836 | goto out; | |
837 | } | |
838 | ||
839 | b = mca_bucket_alloc(c, k, __GFP_NOWARN|GFP_NOIO); | |
840 | if (!b) | |
841 | goto err; | |
842 | ||
843 | BUG_ON(!down_write_trylock(&b->lock)); | |
844 | if (!b->sets->data) | |
845 | goto err; | |
846 | out: | |
847 | BUG_ON(!closure_is_unlocked(&b->io.cl)); | |
848 | ||
849 | bkey_copy(&b->key, k); | |
850 | list_move(&b->list, &c->btree_cache); | |
851 | hlist_del_init_rcu(&b->hash); | |
852 | hlist_add_head_rcu(&b->hash, mca_hash(c, k)); | |
853 | ||
854 | lock_set_subclass(&b->lock.dep_map, level + 1, _THIS_IP_); | |
855 | b->level = level; | |
856 | ||
857 | mca_reinit(b); | |
858 | ||
859 | return b; | |
860 | err: | |
861 | if (b) | |
862 | rw_unlock(true, b); | |
863 | ||
864 | b = mca_cannibalize(c, k, level, cl); | |
865 | if (!IS_ERR(b)) | |
866 | goto out; | |
867 | ||
868 | return b; | |
869 | } | |
870 | ||
871 | /** | |
872 | * bch_btree_node_get - find a btree node in the cache and lock it, reading it | |
873 | * in from disk if necessary. | |
874 | * | |
875 | * If IO is necessary, it uses the closure embedded in struct btree_op to wait; | |
876 | * if that closure is in non blocking mode, will return -EAGAIN. | |
877 | * | |
878 | * The btree node will have either a read or a write lock held, depending on | |
879 | * level and op->lock. | |
880 | */ | |
881 | struct btree *bch_btree_node_get(struct cache_set *c, struct bkey *k, | |
882 | int level, struct btree_op *op) | |
883 | { | |
884 | int i = 0; | |
885 | bool write = level <= op->lock; | |
886 | struct btree *b; | |
887 | ||
888 | BUG_ON(level < 0); | |
889 | retry: | |
890 | b = mca_find(c, k); | |
891 | ||
892 | if (!b) { | |
57943511 KO |
893 | if (current->bio_list) |
894 | return ERR_PTR(-EAGAIN); | |
895 | ||
cafe5635 KO |
896 | mutex_lock(&c->bucket_lock); |
897 | b = mca_alloc(c, k, level, &op->cl); | |
898 | mutex_unlock(&c->bucket_lock); | |
899 | ||
900 | if (!b) | |
901 | goto retry; | |
902 | if (IS_ERR(b)) | |
903 | return b; | |
904 | ||
57943511 | 905 | bch_btree_node_read(b); |
cafe5635 KO |
906 | |
907 | if (!write) | |
908 | downgrade_write(&b->lock); | |
909 | } else { | |
910 | rw_lock(write, b, level); | |
911 | if (PTR_HASH(c, &b->key) != PTR_HASH(c, k)) { | |
912 | rw_unlock(write, b); | |
913 | goto retry; | |
914 | } | |
915 | BUG_ON(b->level != level); | |
916 | } | |
917 | ||
918 | b->accessed = 1; | |
919 | ||
920 | for (; i <= b->nsets && b->sets[i].size; i++) { | |
921 | prefetch(b->sets[i].tree); | |
922 | prefetch(b->sets[i].data); | |
923 | } | |
924 | ||
925 | for (; i <= b->nsets; i++) | |
926 | prefetch(b->sets[i].data); | |
927 | ||
57943511 | 928 | if (btree_node_io_error(b)) { |
cafe5635 | 929 | rw_unlock(write, b); |
57943511 KO |
930 | return ERR_PTR(-EIO); |
931 | } | |
932 | ||
933 | BUG_ON(!b->written); | |
cafe5635 KO |
934 | |
935 | return b; | |
936 | } | |
937 | ||
938 | static void btree_node_prefetch(struct cache_set *c, struct bkey *k, int level) | |
939 | { | |
940 | struct btree *b; | |
941 | ||
942 | mutex_lock(&c->bucket_lock); | |
943 | b = mca_alloc(c, k, level, NULL); | |
944 | mutex_unlock(&c->bucket_lock); | |
945 | ||
946 | if (!IS_ERR_OR_NULL(b)) { | |
57943511 | 947 | bch_btree_node_read(b); |
cafe5635 KO |
948 | rw_unlock(true, b); |
949 | } | |
950 | } | |
951 | ||
952 | /* Btree alloc */ | |
953 | ||
954 | static void btree_node_free(struct btree *b, struct btree_op *op) | |
955 | { | |
956 | unsigned i; | |
957 | ||
c37511b8 KO |
958 | trace_bcache_btree_node_free(b); |
959 | ||
cafe5635 KO |
960 | /* |
961 | * The BUG_ON() in btree_node_get() implies that we must have a write | |
962 | * lock on parent to free or even invalidate a node | |
963 | */ | |
964 | BUG_ON(op->lock <= b->level); | |
965 | BUG_ON(b == b->c->root); | |
cafe5635 KO |
966 | |
967 | if (btree_node_dirty(b)) | |
968 | btree_complete_write(b, btree_current_write(b)); | |
969 | clear_bit(BTREE_NODE_dirty, &b->flags); | |
970 | ||
cafe5635 KO |
971 | cancel_delayed_work(&b->work); |
972 | ||
973 | mutex_lock(&b->c->bucket_lock); | |
974 | ||
975 | for (i = 0; i < KEY_PTRS(&b->key); i++) { | |
976 | BUG_ON(atomic_read(&PTR_BUCKET(b->c, &b->key, i)->pin)); | |
977 | ||
978 | bch_inc_gen(PTR_CACHE(b->c, &b->key, i), | |
979 | PTR_BUCKET(b->c, &b->key, i)); | |
980 | } | |
981 | ||
982 | bch_bucket_free(b->c, &b->key); | |
983 | mca_bucket_free(b); | |
984 | mutex_unlock(&b->c->bucket_lock); | |
985 | } | |
986 | ||
987 | struct btree *bch_btree_node_alloc(struct cache_set *c, int level, | |
988 | struct closure *cl) | |
989 | { | |
990 | BKEY_PADDED(key) k; | |
991 | struct btree *b = ERR_PTR(-EAGAIN); | |
992 | ||
993 | mutex_lock(&c->bucket_lock); | |
994 | retry: | |
995 | if (__bch_bucket_alloc_set(c, WATERMARK_METADATA, &k.key, 1, cl)) | |
996 | goto err; | |
997 | ||
998 | SET_KEY_SIZE(&k.key, c->btree_pages * PAGE_SECTORS); | |
999 | ||
1000 | b = mca_alloc(c, &k.key, level, cl); | |
1001 | if (IS_ERR(b)) | |
1002 | goto err_free; | |
1003 | ||
1004 | if (!b) { | |
b1a67b0f KO |
1005 | cache_bug(c, |
1006 | "Tried to allocate bucket that was in btree cache"); | |
cafe5635 KO |
1007 | __bkey_put(c, &k.key); |
1008 | goto retry; | |
1009 | } | |
1010 | ||
cafe5635 KO |
1011 | b->accessed = 1; |
1012 | bch_bset_init_next(b); | |
1013 | ||
1014 | mutex_unlock(&c->bucket_lock); | |
c37511b8 KO |
1015 | |
1016 | trace_bcache_btree_node_alloc(b); | |
cafe5635 KO |
1017 | return b; |
1018 | err_free: | |
1019 | bch_bucket_free(c, &k.key); | |
1020 | __bkey_put(c, &k.key); | |
1021 | err: | |
1022 | mutex_unlock(&c->bucket_lock); | |
c37511b8 KO |
1023 | |
1024 | trace_bcache_btree_node_alloc_fail(b); | |
cafe5635 KO |
1025 | return b; |
1026 | } | |
1027 | ||
1028 | static struct btree *btree_node_alloc_replacement(struct btree *b, | |
1029 | struct closure *cl) | |
1030 | { | |
1031 | struct btree *n = bch_btree_node_alloc(b->c, b->level, cl); | |
1032 | if (!IS_ERR_OR_NULL(n)) | |
1033 | bch_btree_sort_into(b, n); | |
1034 | ||
1035 | return n; | |
1036 | } | |
1037 | ||
1038 | /* Garbage collection */ | |
1039 | ||
1040 | uint8_t __bch_btree_mark_key(struct cache_set *c, int level, struct bkey *k) | |
1041 | { | |
1042 | uint8_t stale = 0; | |
1043 | unsigned i; | |
1044 | struct bucket *g; | |
1045 | ||
1046 | /* | |
1047 | * ptr_invalid() can't return true for the keys that mark btree nodes as | |
1048 | * freed, but since ptr_bad() returns true we'll never actually use them | |
1049 | * for anything and thus we don't want mark their pointers here | |
1050 | */ | |
1051 | if (!bkey_cmp(k, &ZERO_KEY)) | |
1052 | return stale; | |
1053 | ||
1054 | for (i = 0; i < KEY_PTRS(k); i++) { | |
1055 | if (!ptr_available(c, k, i)) | |
1056 | continue; | |
1057 | ||
1058 | g = PTR_BUCKET(c, k, i); | |
1059 | ||
1060 | if (gen_after(g->gc_gen, PTR_GEN(k, i))) | |
1061 | g->gc_gen = PTR_GEN(k, i); | |
1062 | ||
1063 | if (ptr_stale(c, k, i)) { | |
1064 | stale = max(stale, ptr_stale(c, k, i)); | |
1065 | continue; | |
1066 | } | |
1067 | ||
1068 | cache_bug_on(GC_MARK(g) && | |
1069 | (GC_MARK(g) == GC_MARK_METADATA) != (level != 0), | |
1070 | c, "inconsistent ptrs: mark = %llu, level = %i", | |
1071 | GC_MARK(g), level); | |
1072 | ||
1073 | if (level) | |
1074 | SET_GC_MARK(g, GC_MARK_METADATA); | |
1075 | else if (KEY_DIRTY(k)) | |
1076 | SET_GC_MARK(g, GC_MARK_DIRTY); | |
1077 | ||
1078 | /* guard against overflow */ | |
1079 | SET_GC_SECTORS_USED(g, min_t(unsigned, | |
1080 | GC_SECTORS_USED(g) + KEY_SIZE(k), | |
1081 | (1 << 14) - 1)); | |
1082 | ||
1083 | BUG_ON(!GC_SECTORS_USED(g)); | |
1084 | } | |
1085 | ||
1086 | return stale; | |
1087 | } | |
1088 | ||
1089 | #define btree_mark_key(b, k) __bch_btree_mark_key(b->c, b->level, k) | |
1090 | ||
1091 | static int btree_gc_mark_node(struct btree *b, unsigned *keys, | |
1092 | struct gc_stat *gc) | |
1093 | { | |
1094 | uint8_t stale = 0; | |
1095 | unsigned last_dev = -1; | |
1096 | struct bcache_device *d = NULL; | |
1097 | struct bkey *k; | |
1098 | struct btree_iter iter; | |
1099 | struct bset_tree *t; | |
1100 | ||
1101 | gc->nodes++; | |
1102 | ||
1103 | for_each_key_filter(b, k, &iter, bch_ptr_invalid) { | |
1104 | if (last_dev != KEY_INODE(k)) { | |
1105 | last_dev = KEY_INODE(k); | |
1106 | ||
1107 | d = KEY_INODE(k) < b->c->nr_uuids | |
1108 | ? b->c->devices[last_dev] | |
1109 | : NULL; | |
1110 | } | |
1111 | ||
1112 | stale = max(stale, btree_mark_key(b, k)); | |
1113 | ||
1114 | if (bch_ptr_bad(b, k)) | |
1115 | continue; | |
1116 | ||
1117 | *keys += bkey_u64s(k); | |
1118 | ||
1119 | gc->key_bytes += bkey_u64s(k); | |
1120 | gc->nkeys++; | |
1121 | ||
1122 | gc->data += KEY_SIZE(k); | |
444fc0b6 | 1123 | if (KEY_DIRTY(k)) |
cafe5635 | 1124 | gc->dirty += KEY_SIZE(k); |
cafe5635 KO |
1125 | } |
1126 | ||
1127 | for (t = b->sets; t <= &b->sets[b->nsets]; t++) | |
1128 | btree_bug_on(t->size && | |
1129 | bset_written(b, t) && | |
1130 | bkey_cmp(&b->key, &t->end) < 0, | |
1131 | b, "found short btree key in gc"); | |
1132 | ||
1133 | return stale; | |
1134 | } | |
1135 | ||
1136 | static struct btree *btree_gc_alloc(struct btree *b, struct bkey *k, | |
1137 | struct btree_op *op) | |
1138 | { | |
1139 | /* | |
1140 | * We block priorities from being written for the duration of garbage | |
1141 | * collection, so we can't sleep in btree_alloc() -> | |
1142 | * bch_bucket_alloc_set(), or we'd risk deadlock - so we don't pass it | |
1143 | * our closure. | |
1144 | */ | |
1145 | struct btree *n = btree_node_alloc_replacement(b, NULL); | |
1146 | ||
1147 | if (!IS_ERR_OR_NULL(n)) { | |
1148 | swap(b, n); | |
57943511 | 1149 | __bkey_put(b->c, &b->key); |
cafe5635 KO |
1150 | |
1151 | memcpy(k->ptr, b->key.ptr, | |
1152 | sizeof(uint64_t) * KEY_PTRS(&b->key)); | |
1153 | ||
cafe5635 KO |
1154 | btree_node_free(n, op); |
1155 | up_write(&n->lock); | |
1156 | } | |
1157 | ||
1158 | return b; | |
1159 | } | |
1160 | ||
1161 | /* | |
1162 | * Leaving this at 2 until we've got incremental garbage collection done; it | |
1163 | * could be higher (and has been tested with 4) except that garbage collection | |
1164 | * could take much longer, adversely affecting latency. | |
1165 | */ | |
1166 | #define GC_MERGE_NODES 2U | |
1167 | ||
1168 | struct gc_merge_info { | |
1169 | struct btree *b; | |
1170 | struct bkey *k; | |
1171 | unsigned keys; | |
1172 | }; | |
1173 | ||
1174 | static void btree_gc_coalesce(struct btree *b, struct btree_op *op, | |
1175 | struct gc_stat *gc, struct gc_merge_info *r) | |
1176 | { | |
1177 | unsigned nodes = 0, keys = 0, blocks; | |
1178 | int i; | |
1179 | ||
1180 | while (nodes < GC_MERGE_NODES && r[nodes].b) | |
1181 | keys += r[nodes++].keys; | |
1182 | ||
1183 | blocks = btree_default_blocks(b->c) * 2 / 3; | |
1184 | ||
1185 | if (nodes < 2 || | |
1186 | __set_blocks(b->sets[0].data, keys, b->c) > blocks * (nodes - 1)) | |
1187 | return; | |
1188 | ||
1189 | for (i = nodes - 1; i >= 0; --i) { | |
1190 | if (r[i].b->written) | |
1191 | r[i].b = btree_gc_alloc(r[i].b, r[i].k, op); | |
1192 | ||
1193 | if (r[i].b->written) | |
1194 | return; | |
1195 | } | |
1196 | ||
1197 | for (i = nodes - 1; i > 0; --i) { | |
1198 | struct bset *n1 = r[i].b->sets->data; | |
1199 | struct bset *n2 = r[i - 1].b->sets->data; | |
1200 | struct bkey *k, *last = NULL; | |
1201 | ||
1202 | keys = 0; | |
1203 | ||
1204 | if (i == 1) { | |
1205 | /* | |
1206 | * Last node we're not getting rid of - we're getting | |
1207 | * rid of the node at r[0]. Have to try and fit all of | |
1208 | * the remaining keys into this node; we can't ensure | |
1209 | * they will always fit due to rounding and variable | |
1210 | * length keys (shouldn't be possible in practice, | |
1211 | * though) | |
1212 | */ | |
1213 | if (__set_blocks(n1, n1->keys + r->keys, | |
1214 | b->c) > btree_blocks(r[i].b)) | |
1215 | return; | |
1216 | ||
1217 | keys = n2->keys; | |
1218 | last = &r->b->key; | |
1219 | } else | |
1220 | for (k = n2->start; | |
1221 | k < end(n2); | |
1222 | k = bkey_next(k)) { | |
1223 | if (__set_blocks(n1, n1->keys + keys + | |
1224 | bkey_u64s(k), b->c) > blocks) | |
1225 | break; | |
1226 | ||
1227 | last = k; | |
1228 | keys += bkey_u64s(k); | |
1229 | } | |
1230 | ||
1231 | BUG_ON(__set_blocks(n1, n1->keys + keys, | |
1232 | b->c) > btree_blocks(r[i].b)); | |
1233 | ||
1234 | if (last) { | |
1235 | bkey_copy_key(&r[i].b->key, last); | |
1236 | bkey_copy_key(r[i].k, last); | |
1237 | } | |
1238 | ||
1239 | memcpy(end(n1), | |
1240 | n2->start, | |
1241 | (void *) node(n2, keys) - (void *) n2->start); | |
1242 | ||
1243 | n1->keys += keys; | |
1244 | ||
1245 | memmove(n2->start, | |
1246 | node(n2, keys), | |
1247 | (void *) end(n2) - (void *) node(n2, keys)); | |
1248 | ||
1249 | n2->keys -= keys; | |
1250 | ||
1251 | r[i].keys = n1->keys; | |
1252 | r[i - 1].keys = n2->keys; | |
1253 | } | |
1254 | ||
1255 | btree_node_free(r->b, op); | |
1256 | up_write(&r->b->lock); | |
1257 | ||
c37511b8 | 1258 | trace_bcache_btree_gc_coalesce(nodes); |
cafe5635 KO |
1259 | |
1260 | gc->nodes--; | |
1261 | nodes--; | |
1262 | ||
1263 | memmove(&r[0], &r[1], sizeof(struct gc_merge_info) * nodes); | |
1264 | memset(&r[nodes], 0, sizeof(struct gc_merge_info)); | |
1265 | } | |
1266 | ||
1267 | static int btree_gc_recurse(struct btree *b, struct btree_op *op, | |
1268 | struct closure *writes, struct gc_stat *gc) | |
1269 | { | |
1270 | void write(struct btree *r) | |
1271 | { | |
1272 | if (!r->written) | |
57943511 KO |
1273 | bch_btree_node_write(r, &op->cl); |
1274 | else if (btree_node_dirty(r)) | |
1275 | bch_btree_node_write(r, writes); | |
cafe5635 KO |
1276 | |
1277 | up_write(&r->lock); | |
1278 | } | |
1279 | ||
1280 | int ret = 0, stale; | |
1281 | unsigned i; | |
1282 | struct gc_merge_info r[GC_MERGE_NODES]; | |
1283 | ||
1284 | memset(r, 0, sizeof(r)); | |
1285 | ||
1286 | while ((r->k = bch_next_recurse_key(b, &b->c->gc_done))) { | |
1287 | r->b = bch_btree_node_get(b->c, r->k, b->level - 1, op); | |
1288 | ||
1289 | if (IS_ERR(r->b)) { | |
1290 | ret = PTR_ERR(r->b); | |
1291 | break; | |
1292 | } | |
1293 | ||
1294 | r->keys = 0; | |
1295 | stale = btree_gc_mark_node(r->b, &r->keys, gc); | |
1296 | ||
1297 | if (!b->written && | |
1298 | (r->b->level || stale > 10 || | |
1299 | b->c->gc_always_rewrite)) | |
1300 | r->b = btree_gc_alloc(r->b, r->k, op); | |
1301 | ||
1302 | if (r->b->level) | |
1303 | ret = btree_gc_recurse(r->b, op, writes, gc); | |
1304 | ||
1305 | if (ret) { | |
1306 | write(r->b); | |
1307 | break; | |
1308 | } | |
1309 | ||
1310 | bkey_copy_key(&b->c->gc_done, r->k); | |
1311 | ||
1312 | if (!b->written) | |
1313 | btree_gc_coalesce(b, op, gc, r); | |
1314 | ||
1315 | if (r[GC_MERGE_NODES - 1].b) | |
1316 | write(r[GC_MERGE_NODES - 1].b); | |
1317 | ||
1318 | memmove(&r[1], &r[0], | |
1319 | sizeof(struct gc_merge_info) * (GC_MERGE_NODES - 1)); | |
1320 | ||
1321 | /* When we've got incremental GC working, we'll want to do | |
1322 | * if (should_resched()) | |
1323 | * return -EAGAIN; | |
1324 | */ | |
1325 | cond_resched(); | |
1326 | #if 0 | |
1327 | if (need_resched()) { | |
1328 | ret = -EAGAIN; | |
1329 | break; | |
1330 | } | |
1331 | #endif | |
1332 | } | |
1333 | ||
1334 | for (i = 1; i < GC_MERGE_NODES && r[i].b; i++) | |
1335 | write(r[i].b); | |
1336 | ||
1337 | /* Might have freed some children, must remove their keys */ | |
1338 | if (!b->written) | |
1339 | bch_btree_sort(b); | |
1340 | ||
1341 | return ret; | |
1342 | } | |
1343 | ||
1344 | static int bch_btree_gc_root(struct btree *b, struct btree_op *op, | |
1345 | struct closure *writes, struct gc_stat *gc) | |
1346 | { | |
1347 | struct btree *n = NULL; | |
1348 | unsigned keys = 0; | |
1349 | int ret = 0, stale = btree_gc_mark_node(b, &keys, gc); | |
1350 | ||
1351 | if (b->level || stale > 10) | |
1352 | n = btree_node_alloc_replacement(b, NULL); | |
1353 | ||
1354 | if (!IS_ERR_OR_NULL(n)) | |
1355 | swap(b, n); | |
1356 | ||
1357 | if (b->level) | |
1358 | ret = btree_gc_recurse(b, op, writes, gc); | |
1359 | ||
1360 | if (!b->written || btree_node_dirty(b)) { | |
57943511 | 1361 | bch_btree_node_write(b, n ? &op->cl : NULL); |
cafe5635 KO |
1362 | } |
1363 | ||
1364 | if (!IS_ERR_OR_NULL(n)) { | |
1365 | closure_sync(&op->cl); | |
1366 | bch_btree_set_root(b); | |
1367 | btree_node_free(n, op); | |
1368 | rw_unlock(true, b); | |
1369 | } | |
1370 | ||
1371 | return ret; | |
1372 | } | |
1373 | ||
1374 | static void btree_gc_start(struct cache_set *c) | |
1375 | { | |
1376 | struct cache *ca; | |
1377 | struct bucket *b; | |
cafe5635 KO |
1378 | unsigned i; |
1379 | ||
1380 | if (!c->gc_mark_valid) | |
1381 | return; | |
1382 | ||
1383 | mutex_lock(&c->bucket_lock); | |
1384 | ||
1385 | c->gc_mark_valid = 0; | |
1386 | c->gc_done = ZERO_KEY; | |
1387 | ||
1388 | for_each_cache(ca, c, i) | |
1389 | for_each_bucket(b, ca) { | |
1390 | b->gc_gen = b->gen; | |
1391 | if (!atomic_read(&b->pin)) | |
1392 | SET_GC_MARK(b, GC_MARK_RECLAIMABLE); | |
1393 | } | |
1394 | ||
cafe5635 KO |
1395 | mutex_unlock(&c->bucket_lock); |
1396 | } | |
1397 | ||
1398 | size_t bch_btree_gc_finish(struct cache_set *c) | |
1399 | { | |
1400 | size_t available = 0; | |
1401 | struct bucket *b; | |
1402 | struct cache *ca; | |
cafe5635 KO |
1403 | unsigned i; |
1404 | ||
1405 | mutex_lock(&c->bucket_lock); | |
1406 | ||
1407 | set_gc_sectors(c); | |
1408 | c->gc_mark_valid = 1; | |
1409 | c->need_gc = 0; | |
1410 | ||
1411 | if (c->root) | |
1412 | for (i = 0; i < KEY_PTRS(&c->root->key); i++) | |
1413 | SET_GC_MARK(PTR_BUCKET(c, &c->root->key, i), | |
1414 | GC_MARK_METADATA); | |
1415 | ||
1416 | for (i = 0; i < KEY_PTRS(&c->uuid_bucket); i++) | |
1417 | SET_GC_MARK(PTR_BUCKET(c, &c->uuid_bucket, i), | |
1418 | GC_MARK_METADATA); | |
1419 | ||
1420 | for_each_cache(ca, c, i) { | |
1421 | uint64_t *i; | |
1422 | ||
1423 | ca->invalidate_needs_gc = 0; | |
1424 | ||
1425 | for (i = ca->sb.d; i < ca->sb.d + ca->sb.keys; i++) | |
1426 | SET_GC_MARK(ca->buckets + *i, GC_MARK_METADATA); | |
1427 | ||
1428 | for (i = ca->prio_buckets; | |
1429 | i < ca->prio_buckets + prio_buckets(ca) * 2; i++) | |
1430 | SET_GC_MARK(ca->buckets + *i, GC_MARK_METADATA); | |
1431 | ||
1432 | for_each_bucket(b, ca) { | |
1433 | b->last_gc = b->gc_gen; | |
1434 | c->need_gc = max(c->need_gc, bucket_gc_gen(b)); | |
1435 | ||
1436 | if (!atomic_read(&b->pin) && | |
1437 | GC_MARK(b) == GC_MARK_RECLAIMABLE) { | |
1438 | available++; | |
1439 | if (!GC_SECTORS_USED(b)) | |
1440 | bch_bucket_add_unused(ca, b); | |
1441 | } | |
1442 | } | |
1443 | } | |
1444 | ||
cafe5635 KO |
1445 | mutex_unlock(&c->bucket_lock); |
1446 | return available; | |
1447 | } | |
1448 | ||
1449 | static void bch_btree_gc(struct closure *cl) | |
1450 | { | |
1451 | struct cache_set *c = container_of(cl, struct cache_set, gc.cl); | |
1452 | int ret; | |
1453 | unsigned long available; | |
1454 | struct gc_stat stats; | |
1455 | struct closure writes; | |
1456 | struct btree_op op; | |
cafe5635 | 1457 | uint64_t start_time = local_clock(); |
57943511 | 1458 | |
c37511b8 | 1459 | trace_bcache_gc_start(c); |
cafe5635 KO |
1460 | |
1461 | memset(&stats, 0, sizeof(struct gc_stat)); | |
1462 | closure_init_stack(&writes); | |
1463 | bch_btree_op_init_stack(&op); | |
1464 | op.lock = SHRT_MAX; | |
1465 | ||
1466 | btree_gc_start(c); | |
1467 | ||
57943511 KO |
1468 | atomic_inc(&c->prio_blocked); |
1469 | ||
cafe5635 KO |
1470 | ret = btree_root(gc_root, c, &op, &writes, &stats); |
1471 | closure_sync(&op.cl); | |
1472 | closure_sync(&writes); | |
1473 | ||
1474 | if (ret) { | |
cafe5635 | 1475 | pr_warn("gc failed!"); |
cafe5635 KO |
1476 | continue_at(cl, bch_btree_gc, bch_gc_wq); |
1477 | } | |
1478 | ||
1479 | /* Possibly wait for new UUIDs or whatever to hit disk */ | |
1480 | bch_journal_meta(c, &op.cl); | |
1481 | closure_sync(&op.cl); | |
1482 | ||
1483 | available = bch_btree_gc_finish(c); | |
1484 | ||
57943511 KO |
1485 | atomic_dec(&c->prio_blocked); |
1486 | wake_up_allocators(c); | |
1487 | ||
169ef1cf | 1488 | bch_time_stats_update(&c->btree_gc_time, start_time); |
cafe5635 KO |
1489 | |
1490 | stats.key_bytes *= sizeof(uint64_t); | |
1491 | stats.dirty <<= 9; | |
1492 | stats.data <<= 9; | |
1493 | stats.in_use = (c->nbuckets - available) * 100 / c->nbuckets; | |
1494 | memcpy(&c->gc_stats, &stats, sizeof(struct gc_stat)); | |
cafe5635 | 1495 | |
c37511b8 | 1496 | trace_bcache_gc_end(c); |
cafe5635 KO |
1497 | |
1498 | continue_at(cl, bch_moving_gc, bch_gc_wq); | |
1499 | } | |
1500 | ||
1501 | void bch_queue_gc(struct cache_set *c) | |
1502 | { | |
1503 | closure_trylock_call(&c->gc.cl, bch_btree_gc, bch_gc_wq, &c->cl); | |
1504 | } | |
1505 | ||
1506 | /* Initial partial gc */ | |
1507 | ||
1508 | static int bch_btree_check_recurse(struct btree *b, struct btree_op *op, | |
1509 | unsigned long **seen) | |
1510 | { | |
1511 | int ret; | |
1512 | unsigned i; | |
1513 | struct bkey *k; | |
1514 | struct bucket *g; | |
1515 | struct btree_iter iter; | |
1516 | ||
1517 | for_each_key_filter(b, k, &iter, bch_ptr_invalid) { | |
1518 | for (i = 0; i < KEY_PTRS(k); i++) { | |
1519 | if (!ptr_available(b->c, k, i)) | |
1520 | continue; | |
1521 | ||
1522 | g = PTR_BUCKET(b->c, k, i); | |
1523 | ||
1524 | if (!__test_and_set_bit(PTR_BUCKET_NR(b->c, k, i), | |
1525 | seen[PTR_DEV(k, i)]) || | |
1526 | !ptr_stale(b->c, k, i)) { | |
1527 | g->gen = PTR_GEN(k, i); | |
1528 | ||
1529 | if (b->level) | |
1530 | g->prio = BTREE_PRIO; | |
1531 | else if (g->prio == BTREE_PRIO) | |
1532 | g->prio = INITIAL_PRIO; | |
1533 | } | |
1534 | } | |
1535 | ||
1536 | btree_mark_key(b, k); | |
1537 | } | |
1538 | ||
1539 | if (b->level) { | |
1540 | k = bch_next_recurse_key(b, &ZERO_KEY); | |
1541 | ||
1542 | while (k) { | |
1543 | struct bkey *p = bch_next_recurse_key(b, k); | |
1544 | if (p) | |
1545 | btree_node_prefetch(b->c, p, b->level - 1); | |
1546 | ||
1547 | ret = btree(check_recurse, k, b, op, seen); | |
1548 | if (ret) | |
1549 | return ret; | |
1550 | ||
1551 | k = p; | |
1552 | } | |
1553 | } | |
1554 | ||
1555 | return 0; | |
1556 | } | |
1557 | ||
1558 | int bch_btree_check(struct cache_set *c, struct btree_op *op) | |
1559 | { | |
1560 | int ret = -ENOMEM; | |
1561 | unsigned i; | |
1562 | unsigned long *seen[MAX_CACHES_PER_SET]; | |
1563 | ||
1564 | memset(seen, 0, sizeof(seen)); | |
1565 | ||
1566 | for (i = 0; c->cache[i]; i++) { | |
1567 | size_t n = DIV_ROUND_UP(c->cache[i]->sb.nbuckets, 8); | |
1568 | seen[i] = kmalloc(n, GFP_KERNEL); | |
1569 | if (!seen[i]) | |
1570 | goto err; | |
1571 | ||
1572 | /* Disables the seen array until prio_read() uses it too */ | |
1573 | memset(seen[i], 0xFF, n); | |
1574 | } | |
1575 | ||
1576 | ret = btree_root(check_recurse, c, op, seen); | |
1577 | err: | |
1578 | for (i = 0; i < MAX_CACHES_PER_SET; i++) | |
1579 | kfree(seen[i]); | |
1580 | return ret; | |
1581 | } | |
1582 | ||
1583 | /* Btree insertion */ | |
1584 | ||
1585 | static void shift_keys(struct btree *b, struct bkey *where, struct bkey *insert) | |
1586 | { | |
1587 | struct bset *i = b->sets[b->nsets].data; | |
1588 | ||
1589 | memmove((uint64_t *) where + bkey_u64s(insert), | |
1590 | where, | |
1591 | (void *) end(i) - (void *) where); | |
1592 | ||
1593 | i->keys += bkey_u64s(insert); | |
1594 | bkey_copy(where, insert); | |
1595 | bch_bset_fix_lookup_table(b, where); | |
1596 | } | |
1597 | ||
1598 | static bool fix_overlapping_extents(struct btree *b, | |
1599 | struct bkey *insert, | |
1600 | struct btree_iter *iter, | |
1601 | struct btree_op *op) | |
1602 | { | |
279afbad | 1603 | void subtract_dirty(struct bkey *k, uint64_t offset, int sectors) |
cafe5635 | 1604 | { |
279afbad KO |
1605 | if (KEY_DIRTY(k)) |
1606 | bcache_dev_sectors_dirty_add(b->c, KEY_INODE(k), | |
1607 | offset, -sectors); | |
cafe5635 KO |
1608 | } |
1609 | ||
279afbad | 1610 | uint64_t old_offset; |
cafe5635 KO |
1611 | unsigned old_size, sectors_found = 0; |
1612 | ||
1613 | while (1) { | |
1614 | struct bkey *k = bch_btree_iter_next(iter); | |
1615 | if (!k || | |
1616 | bkey_cmp(&START_KEY(k), insert) >= 0) | |
1617 | break; | |
1618 | ||
1619 | if (bkey_cmp(k, &START_KEY(insert)) <= 0) | |
1620 | continue; | |
1621 | ||
279afbad | 1622 | old_offset = KEY_START(k); |
cafe5635 KO |
1623 | old_size = KEY_SIZE(k); |
1624 | ||
1625 | /* | |
1626 | * We might overlap with 0 size extents; we can't skip these | |
1627 | * because if they're in the set we're inserting to we have to | |
1628 | * adjust them so they don't overlap with the key we're | |
1629 | * inserting. But we don't want to check them for BTREE_REPLACE | |
1630 | * operations. | |
1631 | */ | |
1632 | ||
1633 | if (op->type == BTREE_REPLACE && | |
1634 | KEY_SIZE(k)) { | |
1635 | /* | |
1636 | * k might have been split since we inserted/found the | |
1637 | * key we're replacing | |
1638 | */ | |
1639 | unsigned i; | |
1640 | uint64_t offset = KEY_START(k) - | |
1641 | KEY_START(&op->replace); | |
1642 | ||
1643 | /* But it must be a subset of the replace key */ | |
1644 | if (KEY_START(k) < KEY_START(&op->replace) || | |
1645 | KEY_OFFSET(k) > KEY_OFFSET(&op->replace)) | |
1646 | goto check_failed; | |
1647 | ||
1648 | /* We didn't find a key that we were supposed to */ | |
1649 | if (KEY_START(k) > KEY_START(insert) + sectors_found) | |
1650 | goto check_failed; | |
1651 | ||
1652 | if (KEY_PTRS(&op->replace) != KEY_PTRS(k)) | |
1653 | goto check_failed; | |
1654 | ||
1655 | /* skip past gen */ | |
1656 | offset <<= 8; | |
1657 | ||
1658 | BUG_ON(!KEY_PTRS(&op->replace)); | |
1659 | ||
1660 | for (i = 0; i < KEY_PTRS(&op->replace); i++) | |
1661 | if (k->ptr[i] != op->replace.ptr[i] + offset) | |
1662 | goto check_failed; | |
1663 | ||
1664 | sectors_found = KEY_OFFSET(k) - KEY_START(insert); | |
1665 | } | |
1666 | ||
1667 | if (bkey_cmp(insert, k) < 0 && | |
1668 | bkey_cmp(&START_KEY(insert), &START_KEY(k)) > 0) { | |
1669 | /* | |
1670 | * We overlapped in the middle of an existing key: that | |
1671 | * means we have to split the old key. But we have to do | |
1672 | * slightly different things depending on whether the | |
1673 | * old key has been written out yet. | |
1674 | */ | |
1675 | ||
1676 | struct bkey *top; | |
1677 | ||
279afbad | 1678 | subtract_dirty(k, KEY_START(insert), KEY_SIZE(insert)); |
cafe5635 KO |
1679 | |
1680 | if (bkey_written(b, k)) { | |
1681 | /* | |
1682 | * We insert a new key to cover the top of the | |
1683 | * old key, and the old key is modified in place | |
1684 | * to represent the bottom split. | |
1685 | * | |
1686 | * It's completely arbitrary whether the new key | |
1687 | * is the top or the bottom, but it has to match | |
1688 | * up with what btree_sort_fixup() does - it | |
1689 | * doesn't check for this kind of overlap, it | |
1690 | * depends on us inserting a new key for the top | |
1691 | * here. | |
1692 | */ | |
1693 | top = bch_bset_search(b, &b->sets[b->nsets], | |
1694 | insert); | |
1695 | shift_keys(b, top, k); | |
1696 | } else { | |
1697 | BKEY_PADDED(key) temp; | |
1698 | bkey_copy(&temp.key, k); | |
1699 | shift_keys(b, k, &temp.key); | |
1700 | top = bkey_next(k); | |
1701 | } | |
1702 | ||
1703 | bch_cut_front(insert, top); | |
1704 | bch_cut_back(&START_KEY(insert), k); | |
1705 | bch_bset_fix_invalidated_key(b, k); | |
1706 | return false; | |
1707 | } | |
1708 | ||
1709 | if (bkey_cmp(insert, k) < 0) { | |
1710 | bch_cut_front(insert, k); | |
1711 | } else { | |
1712 | if (bkey_written(b, k) && | |
1713 | bkey_cmp(&START_KEY(insert), &START_KEY(k)) <= 0) { | |
1714 | /* | |
1715 | * Completely overwrote, so we don't have to | |
1716 | * invalidate the binary search tree | |
1717 | */ | |
1718 | bch_cut_front(k, k); | |
1719 | } else { | |
1720 | __bch_cut_back(&START_KEY(insert), k); | |
1721 | bch_bset_fix_invalidated_key(b, k); | |
1722 | } | |
1723 | } | |
1724 | ||
279afbad | 1725 | subtract_dirty(k, old_offset, old_size - KEY_SIZE(k)); |
cafe5635 KO |
1726 | } |
1727 | ||
1728 | check_failed: | |
1729 | if (op->type == BTREE_REPLACE) { | |
1730 | if (!sectors_found) { | |
1731 | op->insert_collision = true; | |
1732 | return true; | |
1733 | } else if (sectors_found < KEY_SIZE(insert)) { | |
1734 | SET_KEY_OFFSET(insert, KEY_OFFSET(insert) - | |
1735 | (KEY_SIZE(insert) - sectors_found)); | |
1736 | SET_KEY_SIZE(insert, sectors_found); | |
1737 | } | |
1738 | } | |
1739 | ||
1740 | return false; | |
1741 | } | |
1742 | ||
1743 | static bool btree_insert_key(struct btree *b, struct btree_op *op, | |
1744 | struct bkey *k) | |
1745 | { | |
1746 | struct bset *i = b->sets[b->nsets].data; | |
1747 | struct bkey *m, *prev; | |
85b1492e | 1748 | unsigned status = BTREE_INSERT_STATUS_INSERT; |
cafe5635 KO |
1749 | |
1750 | BUG_ON(bkey_cmp(k, &b->key) > 0); | |
1751 | BUG_ON(b->level && !KEY_PTRS(k)); | |
1752 | BUG_ON(!b->level && !KEY_OFFSET(k)); | |
1753 | ||
1754 | if (!b->level) { | |
1755 | struct btree_iter iter; | |
1756 | struct bkey search = KEY(KEY_INODE(k), KEY_START(k), 0); | |
1757 | ||
1758 | /* | |
1759 | * bset_search() returns the first key that is strictly greater | |
1760 | * than the search key - but for back merging, we want to find | |
1761 | * the first key that is greater than or equal to KEY_START(k) - | |
1762 | * unless KEY_START(k) is 0. | |
1763 | */ | |
1764 | if (KEY_OFFSET(&search)) | |
1765 | SET_KEY_OFFSET(&search, KEY_OFFSET(&search) - 1); | |
1766 | ||
1767 | prev = NULL; | |
1768 | m = bch_btree_iter_init(b, &iter, &search); | |
1769 | ||
1770 | if (fix_overlapping_extents(b, k, &iter, op)) | |
1771 | return false; | |
1772 | ||
1773 | while (m != end(i) && | |
1774 | bkey_cmp(k, &START_KEY(m)) > 0) | |
1775 | prev = m, m = bkey_next(m); | |
1776 | ||
1777 | if (key_merging_disabled(b->c)) | |
1778 | goto insert; | |
1779 | ||
1780 | /* prev is in the tree, if we merge we're done */ | |
85b1492e | 1781 | status = BTREE_INSERT_STATUS_BACK_MERGE; |
cafe5635 KO |
1782 | if (prev && |
1783 | bch_bkey_try_merge(b, prev, k)) | |
1784 | goto merged; | |
1785 | ||
85b1492e | 1786 | status = BTREE_INSERT_STATUS_OVERWROTE; |
cafe5635 KO |
1787 | if (m != end(i) && |
1788 | KEY_PTRS(m) == KEY_PTRS(k) && !KEY_SIZE(m)) | |
1789 | goto copy; | |
1790 | ||
85b1492e | 1791 | status = BTREE_INSERT_STATUS_FRONT_MERGE; |
cafe5635 KO |
1792 | if (m != end(i) && |
1793 | bch_bkey_try_merge(b, k, m)) | |
1794 | goto copy; | |
1795 | } else | |
1796 | m = bch_bset_search(b, &b->sets[b->nsets], k); | |
1797 | ||
1798 | insert: shift_keys(b, m, k); | |
1799 | copy: bkey_copy(m, k); | |
1800 | merged: | |
279afbad KO |
1801 | if (KEY_DIRTY(k)) |
1802 | bcache_dev_sectors_dirty_add(b->c, KEY_INODE(k), | |
1803 | KEY_START(k), KEY_SIZE(k)); | |
1804 | ||
85b1492e | 1805 | bch_check_keys(b, "%u for %s", status, op_type(op)); |
cafe5635 KO |
1806 | |
1807 | if (b->level && !KEY_OFFSET(k)) | |
57943511 | 1808 | btree_current_write(b)->prio_blocked++; |
cafe5635 | 1809 | |
85b1492e | 1810 | trace_bcache_btree_insert_key(b, k, op->type, status); |
cafe5635 KO |
1811 | |
1812 | return true; | |
1813 | } | |
1814 | ||
1815 | bool bch_btree_insert_keys(struct btree *b, struct btree_op *op) | |
1816 | { | |
1817 | bool ret = false; | |
1818 | struct bkey *k; | |
1819 | unsigned oldsize = bch_count_data(b); | |
1820 | ||
1821 | while ((k = bch_keylist_pop(&op->keys))) { | |
1822 | bkey_put(b->c, k, b->level); | |
1823 | ret |= btree_insert_key(b, op, k); | |
1824 | } | |
1825 | ||
1826 | BUG_ON(bch_count_data(b) < oldsize); | |
1827 | return ret; | |
1828 | } | |
1829 | ||
1830 | bool bch_btree_insert_check_key(struct btree *b, struct btree_op *op, | |
1831 | struct bio *bio) | |
1832 | { | |
1833 | bool ret = false; | |
1834 | uint64_t btree_ptr = b->key.ptr[0]; | |
1835 | unsigned long seq = b->seq; | |
1836 | BKEY_PADDED(k) tmp; | |
1837 | ||
1838 | rw_unlock(false, b); | |
1839 | rw_lock(true, b, b->level); | |
1840 | ||
1841 | if (b->key.ptr[0] != btree_ptr || | |
1842 | b->seq != seq + 1 || | |
1843 | should_split(b)) | |
1844 | goto out; | |
1845 | ||
1846 | op->replace = KEY(op->inode, bio_end(bio), bio_sectors(bio)); | |
1847 | ||
1848 | SET_KEY_PTRS(&op->replace, 1); | |
1849 | get_random_bytes(&op->replace.ptr[0], sizeof(uint64_t)); | |
1850 | ||
1851 | SET_PTR_DEV(&op->replace, 0, PTR_CHECK_DEV); | |
1852 | ||
1853 | bkey_copy(&tmp.k, &op->replace); | |
1854 | ||
1855 | BUG_ON(op->type != BTREE_INSERT); | |
1856 | BUG_ON(!btree_insert_key(b, op, &tmp.k)); | |
cafe5635 KO |
1857 | ret = true; |
1858 | out: | |
1859 | downgrade_write(&b->lock); | |
1860 | return ret; | |
1861 | } | |
1862 | ||
1863 | static int btree_split(struct btree *b, struct btree_op *op) | |
1864 | { | |
1865 | bool split, root = b == b->c->root; | |
1866 | struct btree *n1, *n2 = NULL, *n3 = NULL; | |
1867 | uint64_t start_time = local_clock(); | |
1868 | ||
1869 | if (b->level) | |
1870 | set_closure_blocking(&op->cl); | |
1871 | ||
1872 | n1 = btree_node_alloc_replacement(b, &op->cl); | |
1873 | if (IS_ERR(n1)) | |
1874 | goto err; | |
1875 | ||
1876 | split = set_blocks(n1->sets[0].data, n1->c) > (btree_blocks(b) * 4) / 5; | |
1877 | ||
cafe5635 KO |
1878 | if (split) { |
1879 | unsigned keys = 0; | |
1880 | ||
c37511b8 KO |
1881 | trace_bcache_btree_node_split(b, n1->sets[0].data->keys); |
1882 | ||
cafe5635 KO |
1883 | n2 = bch_btree_node_alloc(b->c, b->level, &op->cl); |
1884 | if (IS_ERR(n2)) | |
1885 | goto err_free1; | |
1886 | ||
1887 | if (root) { | |
1888 | n3 = bch_btree_node_alloc(b->c, b->level + 1, &op->cl); | |
1889 | if (IS_ERR(n3)) | |
1890 | goto err_free2; | |
1891 | } | |
1892 | ||
1893 | bch_btree_insert_keys(n1, op); | |
1894 | ||
1895 | /* Has to be a linear search because we don't have an auxiliary | |
1896 | * search tree yet | |
1897 | */ | |
1898 | ||
1899 | while (keys < (n1->sets[0].data->keys * 3) / 5) | |
1900 | keys += bkey_u64s(node(n1->sets[0].data, keys)); | |
1901 | ||
1902 | bkey_copy_key(&n1->key, node(n1->sets[0].data, keys)); | |
1903 | keys += bkey_u64s(node(n1->sets[0].data, keys)); | |
1904 | ||
1905 | n2->sets[0].data->keys = n1->sets[0].data->keys - keys; | |
1906 | n1->sets[0].data->keys = keys; | |
1907 | ||
1908 | memcpy(n2->sets[0].data->start, | |
1909 | end(n1->sets[0].data), | |
1910 | n2->sets[0].data->keys * sizeof(uint64_t)); | |
1911 | ||
1912 | bkey_copy_key(&n2->key, &b->key); | |
1913 | ||
1914 | bch_keylist_add(&op->keys, &n2->key); | |
57943511 | 1915 | bch_btree_node_write(n2, &op->cl); |
cafe5635 | 1916 | rw_unlock(true, n2); |
c37511b8 KO |
1917 | } else { |
1918 | trace_bcache_btree_node_compact(b, n1->sets[0].data->keys); | |
1919 | ||
cafe5635 | 1920 | bch_btree_insert_keys(n1, op); |
c37511b8 | 1921 | } |
cafe5635 KO |
1922 | |
1923 | bch_keylist_add(&op->keys, &n1->key); | |
57943511 | 1924 | bch_btree_node_write(n1, &op->cl); |
cafe5635 KO |
1925 | |
1926 | if (n3) { | |
1927 | bkey_copy_key(&n3->key, &MAX_KEY); | |
1928 | bch_btree_insert_keys(n3, op); | |
57943511 | 1929 | bch_btree_node_write(n3, &op->cl); |
cafe5635 KO |
1930 | |
1931 | closure_sync(&op->cl); | |
1932 | bch_btree_set_root(n3); | |
1933 | rw_unlock(true, n3); | |
1934 | } else if (root) { | |
1935 | op->keys.top = op->keys.bottom; | |
1936 | closure_sync(&op->cl); | |
1937 | bch_btree_set_root(n1); | |
1938 | } else { | |
1939 | unsigned i; | |
1940 | ||
1941 | bkey_copy(op->keys.top, &b->key); | |
1942 | bkey_copy_key(op->keys.top, &ZERO_KEY); | |
1943 | ||
1944 | for (i = 0; i < KEY_PTRS(&b->key); i++) { | |
1945 | uint8_t g = PTR_BUCKET(b->c, &b->key, i)->gen + 1; | |
1946 | ||
1947 | SET_PTR_GEN(op->keys.top, i, g); | |
1948 | } | |
1949 | ||
1950 | bch_keylist_push(&op->keys); | |
1951 | closure_sync(&op->cl); | |
1952 | atomic_inc(&b->c->prio_blocked); | |
1953 | } | |
1954 | ||
1955 | rw_unlock(true, n1); | |
1956 | btree_node_free(b, op); | |
1957 | ||
169ef1cf | 1958 | bch_time_stats_update(&b->c->btree_split_time, start_time); |
cafe5635 KO |
1959 | |
1960 | return 0; | |
1961 | err_free2: | |
1962 | __bkey_put(n2->c, &n2->key); | |
1963 | btree_node_free(n2, op); | |
1964 | rw_unlock(true, n2); | |
1965 | err_free1: | |
1966 | __bkey_put(n1->c, &n1->key); | |
1967 | btree_node_free(n1, op); | |
1968 | rw_unlock(true, n1); | |
1969 | err: | |
1970 | if (n3 == ERR_PTR(-EAGAIN) || | |
1971 | n2 == ERR_PTR(-EAGAIN) || | |
1972 | n1 == ERR_PTR(-EAGAIN)) | |
1973 | return -EAGAIN; | |
1974 | ||
1975 | pr_warn("couldn't split"); | |
1976 | return -ENOMEM; | |
1977 | } | |
1978 | ||
1979 | static int bch_btree_insert_recurse(struct btree *b, struct btree_op *op, | |
1980 | struct keylist *stack_keys) | |
1981 | { | |
1982 | if (b->level) { | |
1983 | int ret; | |
1984 | struct bkey *insert = op->keys.bottom; | |
1985 | struct bkey *k = bch_next_recurse_key(b, &START_KEY(insert)); | |
1986 | ||
1987 | if (!k) { | |
1988 | btree_bug(b, "no key to recurse on at level %i/%i", | |
1989 | b->level, b->c->root->level); | |
1990 | ||
1991 | op->keys.top = op->keys.bottom; | |
1992 | return -EIO; | |
1993 | } | |
1994 | ||
1995 | if (bkey_cmp(insert, k) > 0) { | |
1996 | unsigned i; | |
1997 | ||
1998 | if (op->type == BTREE_REPLACE) { | |
1999 | __bkey_put(b->c, insert); | |
2000 | op->keys.top = op->keys.bottom; | |
2001 | op->insert_collision = true; | |
2002 | return 0; | |
2003 | } | |
2004 | ||
2005 | for (i = 0; i < KEY_PTRS(insert); i++) | |
2006 | atomic_inc(&PTR_BUCKET(b->c, insert, i)->pin); | |
2007 | ||
2008 | bkey_copy(stack_keys->top, insert); | |
2009 | ||
2010 | bch_cut_back(k, insert); | |
2011 | bch_cut_front(k, stack_keys->top); | |
2012 | ||
2013 | bch_keylist_push(stack_keys); | |
2014 | } | |
2015 | ||
2016 | ret = btree(insert_recurse, k, b, op, stack_keys); | |
2017 | if (ret) | |
2018 | return ret; | |
2019 | } | |
2020 | ||
2021 | if (!bch_keylist_empty(&op->keys)) { | |
2022 | if (should_split(b)) { | |
2023 | if (op->lock <= b->c->root->level) { | |
2024 | BUG_ON(b->level); | |
2025 | op->lock = b->c->root->level + 1; | |
2026 | return -EINTR; | |
2027 | } | |
2028 | return btree_split(b, op); | |
2029 | } | |
2030 | ||
2031 | BUG_ON(write_block(b) != b->sets[b->nsets].data); | |
2032 | ||
57943511 KO |
2033 | if (bch_btree_insert_keys(b, op)) { |
2034 | if (!b->level) | |
2035 | bch_btree_leaf_dirty(b, op); | |
2036 | else | |
2037 | bch_btree_node_write(b, &op->cl); | |
2038 | } | |
cafe5635 KO |
2039 | } |
2040 | ||
2041 | return 0; | |
2042 | } | |
2043 | ||
2044 | int bch_btree_insert(struct btree_op *op, struct cache_set *c) | |
2045 | { | |
2046 | int ret = 0; | |
2047 | struct keylist stack_keys; | |
2048 | ||
2049 | /* | |
2050 | * Don't want to block with the btree locked unless we have to, | |
2051 | * otherwise we get deadlocks with try_harder and between split/gc | |
2052 | */ | |
2053 | clear_closure_blocking(&op->cl); | |
2054 | ||
2055 | BUG_ON(bch_keylist_empty(&op->keys)); | |
2056 | bch_keylist_copy(&stack_keys, &op->keys); | |
2057 | bch_keylist_init(&op->keys); | |
2058 | ||
2059 | while (!bch_keylist_empty(&stack_keys) || | |
2060 | !bch_keylist_empty(&op->keys)) { | |
2061 | if (bch_keylist_empty(&op->keys)) { | |
2062 | bch_keylist_add(&op->keys, | |
2063 | bch_keylist_pop(&stack_keys)); | |
2064 | op->lock = 0; | |
2065 | } | |
2066 | ||
2067 | ret = btree_root(insert_recurse, c, op, &stack_keys); | |
2068 | ||
2069 | if (ret == -EAGAIN) { | |
2070 | ret = 0; | |
2071 | closure_sync(&op->cl); | |
2072 | } else if (ret) { | |
2073 | struct bkey *k; | |
2074 | ||
2075 | pr_err("error %i trying to insert key for %s", | |
2076 | ret, op_type(op)); | |
2077 | ||
2078 | while ((k = bch_keylist_pop(&stack_keys) ?: | |
2079 | bch_keylist_pop(&op->keys))) | |
2080 | bkey_put(c, k, 0); | |
2081 | } | |
2082 | } | |
2083 | ||
2084 | bch_keylist_free(&stack_keys); | |
2085 | ||
2086 | if (op->journal) | |
2087 | atomic_dec_bug(op->journal); | |
2088 | op->journal = NULL; | |
2089 | return ret; | |
2090 | } | |
2091 | ||
2092 | void bch_btree_set_root(struct btree *b) | |
2093 | { | |
2094 | unsigned i; | |
2095 | ||
c37511b8 KO |
2096 | trace_bcache_btree_set_root(b); |
2097 | ||
cafe5635 KO |
2098 | BUG_ON(!b->written); |
2099 | ||
2100 | for (i = 0; i < KEY_PTRS(&b->key); i++) | |
2101 | BUG_ON(PTR_BUCKET(b->c, &b->key, i)->prio != BTREE_PRIO); | |
2102 | ||
2103 | mutex_lock(&b->c->bucket_lock); | |
2104 | list_del_init(&b->list); | |
2105 | mutex_unlock(&b->c->bucket_lock); | |
2106 | ||
2107 | b->c->root = b; | |
2108 | __bkey_put(b->c, &b->key); | |
2109 | ||
2110 | bch_journal_meta(b->c, NULL); | |
cafe5635 KO |
2111 | } |
2112 | ||
2113 | /* Cache lookup */ | |
2114 | ||
2115 | static int submit_partial_cache_miss(struct btree *b, struct btree_op *op, | |
2116 | struct bkey *k) | |
2117 | { | |
2118 | struct search *s = container_of(op, struct search, op); | |
2119 | struct bio *bio = &s->bio.bio; | |
2120 | int ret = 0; | |
2121 | ||
2122 | while (!ret && | |
2123 | !op->lookup_done) { | |
2124 | unsigned sectors = INT_MAX; | |
2125 | ||
2126 | if (KEY_INODE(k) == op->inode) { | |
2127 | if (KEY_START(k) <= bio->bi_sector) | |
2128 | break; | |
2129 | ||
2130 | sectors = min_t(uint64_t, sectors, | |
2131 | KEY_START(k) - bio->bi_sector); | |
2132 | } | |
2133 | ||
2134 | ret = s->d->cache_miss(b, s, bio, sectors); | |
2135 | } | |
2136 | ||
2137 | return ret; | |
2138 | } | |
2139 | ||
2140 | /* | |
2141 | * Read from a single key, handling the initial cache miss if the key starts in | |
2142 | * the middle of the bio | |
2143 | */ | |
2144 | static int submit_partial_cache_hit(struct btree *b, struct btree_op *op, | |
2145 | struct bkey *k) | |
2146 | { | |
2147 | struct search *s = container_of(op, struct search, op); | |
2148 | struct bio *bio = &s->bio.bio; | |
2149 | unsigned ptr; | |
2150 | struct bio *n; | |
2151 | ||
2152 | int ret = submit_partial_cache_miss(b, op, k); | |
2153 | if (ret || op->lookup_done) | |
2154 | return ret; | |
2155 | ||
2156 | /* XXX: figure out best pointer - for multiple cache devices */ | |
2157 | ptr = 0; | |
2158 | ||
2159 | PTR_BUCKET(b->c, k, ptr)->prio = INITIAL_PRIO; | |
2160 | ||
2161 | while (!op->lookup_done && | |
2162 | KEY_INODE(k) == op->inode && | |
2163 | bio->bi_sector < KEY_OFFSET(k)) { | |
2164 | struct bkey *bio_key; | |
2165 | sector_t sector = PTR_OFFSET(k, ptr) + | |
2166 | (bio->bi_sector - KEY_START(k)); | |
2167 | unsigned sectors = min_t(uint64_t, INT_MAX, | |
2168 | KEY_OFFSET(k) - bio->bi_sector); | |
2169 | ||
2170 | n = bch_bio_split(bio, sectors, GFP_NOIO, s->d->bio_split); | |
2171 | if (!n) | |
2172 | return -EAGAIN; | |
2173 | ||
2174 | if (n == bio) | |
2175 | op->lookup_done = true; | |
2176 | ||
2177 | bio_key = &container_of(n, struct bbio, bio)->key; | |
2178 | ||
2179 | /* | |
2180 | * The bucket we're reading from might be reused while our bio | |
2181 | * is in flight, and we could then end up reading the wrong | |
2182 | * data. | |
2183 | * | |
2184 | * We guard against this by checking (in cache_read_endio()) if | |
2185 | * the pointer is stale again; if so, we treat it as an error | |
2186 | * and reread from the backing device (but we don't pass that | |
2187 | * error up anywhere). | |
2188 | */ | |
2189 | ||
2190 | bch_bkey_copy_single_ptr(bio_key, k, ptr); | |
2191 | SET_PTR_OFFSET(bio_key, 0, sector); | |
2192 | ||
2193 | n->bi_end_io = bch_cache_read_endio; | |
2194 | n->bi_private = &s->cl; | |
2195 | ||
cafe5635 KO |
2196 | __bch_submit_bbio(n, b->c); |
2197 | } | |
2198 | ||
2199 | return 0; | |
2200 | } | |
2201 | ||
2202 | int bch_btree_search_recurse(struct btree *b, struct btree_op *op) | |
2203 | { | |
2204 | struct search *s = container_of(op, struct search, op); | |
2205 | struct bio *bio = &s->bio.bio; | |
2206 | ||
2207 | int ret = 0; | |
2208 | struct bkey *k; | |
2209 | struct btree_iter iter; | |
2210 | bch_btree_iter_init(b, &iter, &KEY(op->inode, bio->bi_sector, 0)); | |
2211 | ||
cafe5635 KO |
2212 | do { |
2213 | k = bch_btree_iter_next_filter(&iter, b, bch_ptr_bad); | |
2214 | if (!k) { | |
2215 | /* | |
2216 | * b->key would be exactly what we want, except that | |
2217 | * pointers to btree nodes have nonzero size - we | |
2218 | * wouldn't go far enough | |
2219 | */ | |
2220 | ||
2221 | ret = submit_partial_cache_miss(b, op, | |
2222 | &KEY(KEY_INODE(&b->key), | |
2223 | KEY_OFFSET(&b->key), 0)); | |
2224 | break; | |
2225 | } | |
2226 | ||
2227 | ret = b->level | |
2228 | ? btree(search_recurse, k, b, op) | |
2229 | : submit_partial_cache_hit(b, op, k); | |
2230 | } while (!ret && | |
2231 | !op->lookup_done); | |
2232 | ||
2233 | return ret; | |
2234 | } | |
2235 | ||
2236 | /* Keybuf code */ | |
2237 | ||
2238 | static inline int keybuf_cmp(struct keybuf_key *l, struct keybuf_key *r) | |
2239 | { | |
2240 | /* Overlapping keys compare equal */ | |
2241 | if (bkey_cmp(&l->key, &START_KEY(&r->key)) <= 0) | |
2242 | return -1; | |
2243 | if (bkey_cmp(&START_KEY(&l->key), &r->key) >= 0) | |
2244 | return 1; | |
2245 | return 0; | |
2246 | } | |
2247 | ||
2248 | static inline int keybuf_nonoverlapping_cmp(struct keybuf_key *l, | |
2249 | struct keybuf_key *r) | |
2250 | { | |
2251 | return clamp_t(int64_t, bkey_cmp(&l->key, &r->key), -1, 1); | |
2252 | } | |
2253 | ||
2254 | static int bch_btree_refill_keybuf(struct btree *b, struct btree_op *op, | |
72c27061 KO |
2255 | struct keybuf *buf, struct bkey *end, |
2256 | keybuf_pred_fn *pred) | |
cafe5635 KO |
2257 | { |
2258 | struct btree_iter iter; | |
2259 | bch_btree_iter_init(b, &iter, &buf->last_scanned); | |
2260 | ||
2261 | while (!array_freelist_empty(&buf->freelist)) { | |
2262 | struct bkey *k = bch_btree_iter_next_filter(&iter, b, | |
2263 | bch_ptr_bad); | |
2264 | ||
2265 | if (!b->level) { | |
2266 | if (!k) { | |
2267 | buf->last_scanned = b->key; | |
2268 | break; | |
2269 | } | |
2270 | ||
2271 | buf->last_scanned = *k; | |
2272 | if (bkey_cmp(&buf->last_scanned, end) >= 0) | |
2273 | break; | |
2274 | ||
72c27061 | 2275 | if (pred(buf, k)) { |
cafe5635 KO |
2276 | struct keybuf_key *w; |
2277 | ||
cafe5635 KO |
2278 | spin_lock(&buf->lock); |
2279 | ||
2280 | w = array_alloc(&buf->freelist); | |
2281 | ||
2282 | w->private = NULL; | |
2283 | bkey_copy(&w->key, k); | |
2284 | ||
2285 | if (RB_INSERT(&buf->keys, w, node, keybuf_cmp)) | |
2286 | array_free(&buf->freelist, w); | |
2287 | ||
2288 | spin_unlock(&buf->lock); | |
2289 | } | |
2290 | } else { | |
2291 | if (!k) | |
2292 | break; | |
2293 | ||
72c27061 | 2294 | btree(refill_keybuf, k, b, op, buf, end, pred); |
cafe5635 KO |
2295 | /* |
2296 | * Might get an error here, but can't really do anything | |
2297 | * and it'll get logged elsewhere. Just read what we | |
2298 | * can. | |
2299 | */ | |
2300 | ||
2301 | if (bkey_cmp(&buf->last_scanned, end) >= 0) | |
2302 | break; | |
2303 | ||
2304 | cond_resched(); | |
2305 | } | |
2306 | } | |
2307 | ||
2308 | return 0; | |
2309 | } | |
2310 | ||
2311 | void bch_refill_keybuf(struct cache_set *c, struct keybuf *buf, | |
72c27061 | 2312 | struct bkey *end, keybuf_pred_fn *pred) |
cafe5635 KO |
2313 | { |
2314 | struct bkey start = buf->last_scanned; | |
2315 | struct btree_op op; | |
2316 | bch_btree_op_init_stack(&op); | |
2317 | ||
2318 | cond_resched(); | |
2319 | ||
72c27061 | 2320 | btree_root(refill_keybuf, c, &op, buf, end, pred); |
cafe5635 KO |
2321 | closure_sync(&op.cl); |
2322 | ||
2323 | pr_debug("found %s keys from %llu:%llu to %llu:%llu", | |
2324 | RB_EMPTY_ROOT(&buf->keys) ? "no" : | |
2325 | array_freelist_empty(&buf->freelist) ? "some" : "a few", | |
2326 | KEY_INODE(&start), KEY_OFFSET(&start), | |
2327 | KEY_INODE(&buf->last_scanned), KEY_OFFSET(&buf->last_scanned)); | |
2328 | ||
2329 | spin_lock(&buf->lock); | |
2330 | ||
2331 | if (!RB_EMPTY_ROOT(&buf->keys)) { | |
2332 | struct keybuf_key *w; | |
2333 | w = RB_FIRST(&buf->keys, struct keybuf_key, node); | |
2334 | buf->start = START_KEY(&w->key); | |
2335 | ||
2336 | w = RB_LAST(&buf->keys, struct keybuf_key, node); | |
2337 | buf->end = w->key; | |
2338 | } else { | |
2339 | buf->start = MAX_KEY; | |
2340 | buf->end = MAX_KEY; | |
2341 | } | |
2342 | ||
2343 | spin_unlock(&buf->lock); | |
2344 | } | |
2345 | ||
2346 | static void __bch_keybuf_del(struct keybuf *buf, struct keybuf_key *w) | |
2347 | { | |
2348 | rb_erase(&w->node, &buf->keys); | |
2349 | array_free(&buf->freelist, w); | |
2350 | } | |
2351 | ||
2352 | void bch_keybuf_del(struct keybuf *buf, struct keybuf_key *w) | |
2353 | { | |
2354 | spin_lock(&buf->lock); | |
2355 | __bch_keybuf_del(buf, w); | |
2356 | spin_unlock(&buf->lock); | |
2357 | } | |
2358 | ||
2359 | bool bch_keybuf_check_overlapping(struct keybuf *buf, struct bkey *start, | |
2360 | struct bkey *end) | |
2361 | { | |
2362 | bool ret = false; | |
2363 | struct keybuf_key *p, *w, s; | |
2364 | s.key = *start; | |
2365 | ||
2366 | if (bkey_cmp(end, &buf->start) <= 0 || | |
2367 | bkey_cmp(start, &buf->end) >= 0) | |
2368 | return false; | |
2369 | ||
2370 | spin_lock(&buf->lock); | |
2371 | w = RB_GREATER(&buf->keys, s, node, keybuf_nonoverlapping_cmp); | |
2372 | ||
2373 | while (w && bkey_cmp(&START_KEY(&w->key), end) < 0) { | |
2374 | p = w; | |
2375 | w = RB_NEXT(w, node); | |
2376 | ||
2377 | if (p->private) | |
2378 | ret = true; | |
2379 | else | |
2380 | __bch_keybuf_del(buf, p); | |
2381 | } | |
2382 | ||
2383 | spin_unlock(&buf->lock); | |
2384 | return ret; | |
2385 | } | |
2386 | ||
2387 | struct keybuf_key *bch_keybuf_next(struct keybuf *buf) | |
2388 | { | |
2389 | struct keybuf_key *w; | |
2390 | spin_lock(&buf->lock); | |
2391 | ||
2392 | w = RB_FIRST(&buf->keys, struct keybuf_key, node); | |
2393 | ||
2394 | while (w && w->private) | |
2395 | w = RB_NEXT(w, node); | |
2396 | ||
2397 | if (w) | |
2398 | w->private = ERR_PTR(-EINTR); | |
2399 | ||
2400 | spin_unlock(&buf->lock); | |
2401 | return w; | |
2402 | } | |
2403 | ||
2404 | struct keybuf_key *bch_keybuf_next_rescan(struct cache_set *c, | |
2405 | struct keybuf *buf, | |
72c27061 KO |
2406 | struct bkey *end, |
2407 | keybuf_pred_fn *pred) | |
cafe5635 KO |
2408 | { |
2409 | struct keybuf_key *ret; | |
2410 | ||
2411 | while (1) { | |
2412 | ret = bch_keybuf_next(buf); | |
2413 | if (ret) | |
2414 | break; | |
2415 | ||
2416 | if (bkey_cmp(&buf->last_scanned, end) >= 0) { | |
2417 | pr_debug("scan finished"); | |
2418 | break; | |
2419 | } | |
2420 | ||
72c27061 | 2421 | bch_refill_keybuf(c, buf, end, pred); |
cafe5635 KO |
2422 | } |
2423 | ||
2424 | return ret; | |
2425 | } | |
2426 | ||
72c27061 | 2427 | void bch_keybuf_init(struct keybuf *buf) |
cafe5635 | 2428 | { |
cafe5635 KO |
2429 | buf->last_scanned = MAX_KEY; |
2430 | buf->keys = RB_ROOT; | |
2431 | ||
2432 | spin_lock_init(&buf->lock); | |
2433 | array_allocator_init(&buf->freelist); | |
2434 | } | |
2435 | ||
2436 | void bch_btree_exit(void) | |
2437 | { | |
2438 | if (btree_io_wq) | |
2439 | destroy_workqueue(btree_io_wq); | |
2440 | if (bch_gc_wq) | |
2441 | destroy_workqueue(bch_gc_wq); | |
2442 | } | |
2443 | ||
2444 | int __init bch_btree_init(void) | |
2445 | { | |
2446 | if (!(bch_gc_wq = create_singlethread_workqueue("bch_btree_gc")) || | |
2447 | !(btree_io_wq = create_singlethread_workqueue("bch_btree_io"))) | |
2448 | return -ENOMEM; | |
2449 | ||
2450 | return 0; | |
2451 | } |