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 | ||
f3059a54 | 138 | static 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: | |
61cbd250 | 258 | bch_cache_set_error(b->c, "io error reading bucket %zu", |
57943511 | 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; | |
e49c7c37 KO |
331 | b->bio->bi_rw = REQ_META|WRITE_SYNC|REQ_FUA; |
332 | b->bio->bi_size = set_blocks(i, b->c) * block_bytes(b->c); | |
169ef1cf | 333 | bch_bio_map(b->bio, i); |
cafe5635 | 334 | |
e49c7c37 KO |
335 | /* |
336 | * If we're appending to a leaf node, we don't technically need FUA - | |
337 | * this write just needs to be persisted before the next journal write, | |
338 | * which will be marked FLUSH|FUA. | |
339 | * | |
340 | * Similarly if we're writing a new btree root - the pointer is going to | |
341 | * be in the next journal entry. | |
342 | * | |
343 | * But if we're writing a new btree node (that isn't a root) or | |
344 | * appending to a non leaf btree node, we need either FUA or a flush | |
345 | * when we write the parent with the new pointer. FUA is cheaper than a | |
346 | * flush, and writes appending to leaf nodes aren't blocking anything so | |
347 | * just make all btree node writes FUA to keep things sane. | |
348 | */ | |
349 | ||
cafe5635 KO |
350 | bkey_copy(&k.key, &b->key); |
351 | SET_PTR_OFFSET(&k.key, 0, PTR_OFFSET(&k.key, 0) + bset_offset(b, i)); | |
352 | ||
8e51e414 | 353 | if (!bio_alloc_pages(b->bio, GFP_NOIO)) { |
cafe5635 KO |
354 | int j; |
355 | struct bio_vec *bv; | |
356 | void *base = (void *) ((unsigned long) i & ~(PAGE_SIZE - 1)); | |
357 | ||
358 | bio_for_each_segment(bv, b->bio, j) | |
359 | memcpy(page_address(bv->bv_page), | |
360 | base + j * PAGE_SIZE, PAGE_SIZE); | |
361 | ||
cafe5635 KO |
362 | bch_submit_bbio(b->bio, b->c, &k.key, 0); |
363 | ||
57943511 | 364 | continue_at(cl, btree_node_write_done, NULL); |
cafe5635 KO |
365 | } else { |
366 | b->bio->bi_vcnt = 0; | |
169ef1cf | 367 | bch_bio_map(b->bio, i); |
cafe5635 | 368 | |
cafe5635 KO |
369 | bch_submit_bbio(b->bio, b->c, &k.key, 0); |
370 | ||
371 | closure_sync(cl); | |
57943511 | 372 | __btree_node_write_done(cl); |
cafe5635 KO |
373 | } |
374 | } | |
375 | ||
57943511 | 376 | void bch_btree_node_write(struct btree *b, struct closure *parent) |
cafe5635 KO |
377 | { |
378 | struct bset *i = b->sets[b->nsets].data; | |
379 | ||
c37511b8 KO |
380 | trace_bcache_btree_write(b); |
381 | ||
cafe5635 | 382 | BUG_ON(current->bio_list); |
57943511 KO |
383 | BUG_ON(b->written >= btree_blocks(b)); |
384 | BUG_ON(b->written && !i->keys); | |
385 | BUG_ON(b->sets->data->seq != i->seq); | |
c37511b8 | 386 | bch_check_key_order(b, i); |
cafe5635 | 387 | |
cafe5635 KO |
388 | cancel_delayed_work(&b->work); |
389 | ||
57943511 KO |
390 | /* If caller isn't waiting for write, parent refcount is cache set */ |
391 | closure_lock(&b->io, parent ?: &b->c->cl); | |
392 | ||
cafe5635 KO |
393 | clear_bit(BTREE_NODE_dirty, &b->flags); |
394 | change_bit(BTREE_NODE_write_idx, &b->flags); | |
395 | ||
57943511 | 396 | do_btree_node_write(b); |
cafe5635 | 397 | |
cafe5635 KO |
398 | b->written += set_blocks(i, b->c); |
399 | atomic_long_add(set_blocks(i, b->c) * b->c->sb.block_size, | |
400 | &PTR_CACHE(b->c, &b->key, 0)->btree_sectors_written); | |
401 | ||
402 | bch_btree_sort_lazy(b); | |
403 | ||
404 | if (b->written < btree_blocks(b)) | |
405 | bch_bset_init_next(b); | |
406 | } | |
407 | ||
57943511 | 408 | static void btree_node_write_work(struct work_struct *w) |
cafe5635 KO |
409 | { |
410 | struct btree *b = container_of(to_delayed_work(w), struct btree, work); | |
411 | ||
57943511 | 412 | rw_lock(true, b, b->level); |
cafe5635 KO |
413 | |
414 | if (btree_node_dirty(b)) | |
57943511 KO |
415 | bch_btree_node_write(b, NULL); |
416 | rw_unlock(true, b); | |
cafe5635 KO |
417 | } |
418 | ||
57943511 | 419 | static void bch_btree_leaf_dirty(struct btree *b, struct btree_op *op) |
cafe5635 KO |
420 | { |
421 | struct bset *i = b->sets[b->nsets].data; | |
422 | struct btree_write *w = btree_current_write(b); | |
423 | ||
57943511 KO |
424 | BUG_ON(!b->written); |
425 | BUG_ON(!i->keys); | |
cafe5635 | 426 | |
57943511 KO |
427 | if (!btree_node_dirty(b)) |
428 | queue_delayed_work(btree_io_wq, &b->work, 30 * HZ); | |
cafe5635 | 429 | |
57943511 | 430 | set_btree_node_dirty(b); |
cafe5635 | 431 | |
57943511 | 432 | if (op && op->journal) { |
cafe5635 KO |
433 | if (w->journal && |
434 | journal_pin_cmp(b->c, w, op)) { | |
435 | atomic_dec_bug(w->journal); | |
436 | w->journal = NULL; | |
437 | } | |
438 | ||
439 | if (!w->journal) { | |
440 | w->journal = op->journal; | |
441 | atomic_inc(w->journal); | |
442 | } | |
443 | } | |
444 | ||
cafe5635 | 445 | /* Force write if set is too big */ |
57943511 KO |
446 | if (set_bytes(i) > PAGE_SIZE - 48 && |
447 | !current->bio_list) | |
448 | bch_btree_node_write(b, NULL); | |
cafe5635 KO |
449 | } |
450 | ||
451 | /* | |
452 | * Btree in memory cache - allocation/freeing | |
453 | * mca -> memory cache | |
454 | */ | |
455 | ||
456 | static void mca_reinit(struct btree *b) | |
457 | { | |
458 | unsigned i; | |
459 | ||
460 | b->flags = 0; | |
461 | b->written = 0; | |
462 | b->nsets = 0; | |
463 | ||
464 | for (i = 0; i < MAX_BSETS; i++) | |
465 | b->sets[i].size = 0; | |
466 | /* | |
467 | * Second loop starts at 1 because b->sets[0]->data is the memory we | |
468 | * allocated | |
469 | */ | |
470 | for (i = 1; i < MAX_BSETS; i++) | |
471 | b->sets[i].data = NULL; | |
472 | } | |
473 | ||
474 | #define mca_reserve(c) (((c->root && c->root->level) \ | |
475 | ? c->root->level : 1) * 8 + 16) | |
476 | #define mca_can_free(c) \ | |
477 | max_t(int, 0, c->bucket_cache_used - mca_reserve(c)) | |
478 | ||
479 | static void mca_data_free(struct btree *b) | |
480 | { | |
481 | struct bset_tree *t = b->sets; | |
482 | BUG_ON(!closure_is_unlocked(&b->io.cl)); | |
483 | ||
484 | if (bset_prev_bytes(b) < PAGE_SIZE) | |
485 | kfree(t->prev); | |
486 | else | |
487 | free_pages((unsigned long) t->prev, | |
488 | get_order(bset_prev_bytes(b))); | |
489 | ||
490 | if (bset_tree_bytes(b) < PAGE_SIZE) | |
491 | kfree(t->tree); | |
492 | else | |
493 | free_pages((unsigned long) t->tree, | |
494 | get_order(bset_tree_bytes(b))); | |
495 | ||
496 | free_pages((unsigned long) t->data, b->page_order); | |
497 | ||
498 | t->prev = NULL; | |
499 | t->tree = NULL; | |
500 | t->data = NULL; | |
501 | list_move(&b->list, &b->c->btree_cache_freed); | |
502 | b->c->bucket_cache_used--; | |
503 | } | |
504 | ||
505 | static void mca_bucket_free(struct btree *b) | |
506 | { | |
507 | BUG_ON(btree_node_dirty(b)); | |
508 | ||
509 | b->key.ptr[0] = 0; | |
510 | hlist_del_init_rcu(&b->hash); | |
511 | list_move(&b->list, &b->c->btree_cache_freeable); | |
512 | } | |
513 | ||
514 | static unsigned btree_order(struct bkey *k) | |
515 | { | |
516 | return ilog2(KEY_SIZE(k) / PAGE_SECTORS ?: 1); | |
517 | } | |
518 | ||
519 | static void mca_data_alloc(struct btree *b, struct bkey *k, gfp_t gfp) | |
520 | { | |
521 | struct bset_tree *t = b->sets; | |
522 | BUG_ON(t->data); | |
523 | ||
524 | b->page_order = max_t(unsigned, | |
525 | ilog2(b->c->btree_pages), | |
526 | btree_order(k)); | |
527 | ||
528 | t->data = (void *) __get_free_pages(gfp, b->page_order); | |
529 | if (!t->data) | |
530 | goto err; | |
531 | ||
532 | t->tree = bset_tree_bytes(b) < PAGE_SIZE | |
533 | ? kmalloc(bset_tree_bytes(b), gfp) | |
534 | : (void *) __get_free_pages(gfp, get_order(bset_tree_bytes(b))); | |
535 | if (!t->tree) | |
536 | goto err; | |
537 | ||
538 | t->prev = bset_prev_bytes(b) < PAGE_SIZE | |
539 | ? kmalloc(bset_prev_bytes(b), gfp) | |
540 | : (void *) __get_free_pages(gfp, get_order(bset_prev_bytes(b))); | |
541 | if (!t->prev) | |
542 | goto err; | |
543 | ||
544 | list_move(&b->list, &b->c->btree_cache); | |
545 | b->c->bucket_cache_used++; | |
546 | return; | |
547 | err: | |
548 | mca_data_free(b); | |
549 | } | |
550 | ||
551 | static struct btree *mca_bucket_alloc(struct cache_set *c, | |
552 | struct bkey *k, gfp_t gfp) | |
553 | { | |
554 | struct btree *b = kzalloc(sizeof(struct btree), gfp); | |
555 | if (!b) | |
556 | return NULL; | |
557 | ||
558 | init_rwsem(&b->lock); | |
559 | lockdep_set_novalidate_class(&b->lock); | |
560 | INIT_LIST_HEAD(&b->list); | |
57943511 | 561 | INIT_DELAYED_WORK(&b->work, btree_node_write_work); |
cafe5635 KO |
562 | b->c = c; |
563 | closure_init_unlocked(&b->io); | |
564 | ||
565 | mca_data_alloc(b, k, gfp); | |
566 | return b; | |
567 | } | |
568 | ||
569 | static int mca_reap(struct btree *b, struct closure *cl, unsigned min_order) | |
570 | { | |
571 | lockdep_assert_held(&b->c->bucket_lock); | |
572 | ||
573 | if (!down_write_trylock(&b->lock)) | |
574 | return -ENOMEM; | |
575 | ||
576 | if (b->page_order < min_order) { | |
577 | rw_unlock(true, b); | |
578 | return -ENOMEM; | |
579 | } | |
580 | ||
581 | BUG_ON(btree_node_dirty(b) && !b->sets[0].data); | |
582 | ||
583 | if (cl && btree_node_dirty(b)) | |
57943511 | 584 | bch_btree_node_write(b, NULL); |
cafe5635 KO |
585 | |
586 | if (cl) | |
587 | closure_wait_event_async(&b->io.wait, cl, | |
588 | atomic_read(&b->io.cl.remaining) == -1); | |
589 | ||
590 | if (btree_node_dirty(b) || | |
591 | !closure_is_unlocked(&b->io.cl) || | |
592 | work_pending(&b->work.work)) { | |
593 | rw_unlock(true, b); | |
594 | return -EAGAIN; | |
595 | } | |
596 | ||
597 | return 0; | |
598 | } | |
599 | ||
7dc19d5a DC |
600 | static unsigned long bch_mca_scan(struct shrinker *shrink, |
601 | struct shrink_control *sc) | |
cafe5635 KO |
602 | { |
603 | struct cache_set *c = container_of(shrink, struct cache_set, shrink); | |
604 | struct btree *b, *t; | |
605 | unsigned long i, nr = sc->nr_to_scan; | |
7dc19d5a | 606 | unsigned long freed = 0; |
cafe5635 KO |
607 | |
608 | if (c->shrinker_disabled) | |
7dc19d5a | 609 | return SHRINK_STOP; |
cafe5635 KO |
610 | |
611 | if (c->try_harder) | |
7dc19d5a | 612 | return SHRINK_STOP; |
cafe5635 KO |
613 | |
614 | /* Return -1 if we can't do anything right now */ | |
a698e08c | 615 | if (sc->gfp_mask & __GFP_IO) |
cafe5635 KO |
616 | mutex_lock(&c->bucket_lock); |
617 | else if (!mutex_trylock(&c->bucket_lock)) | |
618 | return -1; | |
619 | ||
36c9ea98 KO |
620 | /* |
621 | * It's _really_ critical that we don't free too many btree nodes - we | |
622 | * have to always leave ourselves a reserve. The reserve is how we | |
623 | * guarantee that allocating memory for a new btree node can always | |
624 | * succeed, so that inserting keys into the btree can always succeed and | |
625 | * IO can always make forward progress: | |
626 | */ | |
cafe5635 KO |
627 | nr /= c->btree_pages; |
628 | nr = min_t(unsigned long, nr, mca_can_free(c)); | |
629 | ||
630 | i = 0; | |
631 | list_for_each_entry_safe(b, t, &c->btree_cache_freeable, list) { | |
7dc19d5a | 632 | if (freed >= nr) |
cafe5635 KO |
633 | break; |
634 | ||
635 | if (++i > 3 && | |
636 | !mca_reap(b, NULL, 0)) { | |
637 | mca_data_free(b); | |
638 | rw_unlock(true, b); | |
7dc19d5a | 639 | freed++; |
cafe5635 KO |
640 | } |
641 | } | |
642 | ||
643 | /* | |
644 | * Can happen right when we first start up, before we've read in any | |
645 | * btree nodes | |
646 | */ | |
647 | if (list_empty(&c->btree_cache)) | |
648 | goto out; | |
649 | ||
7dc19d5a | 650 | for (i = 0; (nr--) && i < c->bucket_cache_used; i++) { |
cafe5635 KO |
651 | b = list_first_entry(&c->btree_cache, struct btree, list); |
652 | list_rotate_left(&c->btree_cache); | |
653 | ||
654 | if (!b->accessed && | |
655 | !mca_reap(b, NULL, 0)) { | |
656 | mca_bucket_free(b); | |
657 | mca_data_free(b); | |
658 | rw_unlock(true, b); | |
7dc19d5a | 659 | freed++; |
cafe5635 KO |
660 | } else |
661 | b->accessed = 0; | |
662 | } | |
663 | out: | |
cafe5635 | 664 | mutex_unlock(&c->bucket_lock); |
7dc19d5a DC |
665 | return freed; |
666 | } | |
667 | ||
668 | static unsigned long bch_mca_count(struct shrinker *shrink, | |
669 | struct shrink_control *sc) | |
670 | { | |
671 | struct cache_set *c = container_of(shrink, struct cache_set, shrink); | |
672 | ||
673 | if (c->shrinker_disabled) | |
674 | return 0; | |
675 | ||
676 | if (c->try_harder) | |
677 | return 0; | |
678 | ||
679 | return mca_can_free(c) * c->btree_pages; | |
cafe5635 KO |
680 | } |
681 | ||
682 | void bch_btree_cache_free(struct cache_set *c) | |
683 | { | |
684 | struct btree *b; | |
685 | struct closure cl; | |
686 | closure_init_stack(&cl); | |
687 | ||
688 | if (c->shrink.list.next) | |
689 | unregister_shrinker(&c->shrink); | |
690 | ||
691 | mutex_lock(&c->bucket_lock); | |
692 | ||
693 | #ifdef CONFIG_BCACHE_DEBUG | |
694 | if (c->verify_data) | |
695 | list_move(&c->verify_data->list, &c->btree_cache); | |
696 | #endif | |
697 | ||
698 | list_splice(&c->btree_cache_freeable, | |
699 | &c->btree_cache); | |
700 | ||
701 | while (!list_empty(&c->btree_cache)) { | |
702 | b = list_first_entry(&c->btree_cache, struct btree, list); | |
703 | ||
704 | if (btree_node_dirty(b)) | |
705 | btree_complete_write(b, btree_current_write(b)); | |
706 | clear_bit(BTREE_NODE_dirty, &b->flags); | |
707 | ||
708 | mca_data_free(b); | |
709 | } | |
710 | ||
711 | while (!list_empty(&c->btree_cache_freed)) { | |
712 | b = list_first_entry(&c->btree_cache_freed, | |
713 | struct btree, list); | |
714 | list_del(&b->list); | |
715 | cancel_delayed_work_sync(&b->work); | |
716 | kfree(b); | |
717 | } | |
718 | ||
719 | mutex_unlock(&c->bucket_lock); | |
720 | } | |
721 | ||
722 | int bch_btree_cache_alloc(struct cache_set *c) | |
723 | { | |
724 | unsigned i; | |
725 | ||
726 | /* XXX: doesn't check for errors */ | |
727 | ||
728 | closure_init_unlocked(&c->gc); | |
729 | ||
730 | for (i = 0; i < mca_reserve(c); i++) | |
731 | mca_bucket_alloc(c, &ZERO_KEY, GFP_KERNEL); | |
732 | ||
733 | list_splice_init(&c->btree_cache, | |
734 | &c->btree_cache_freeable); | |
735 | ||
736 | #ifdef CONFIG_BCACHE_DEBUG | |
737 | mutex_init(&c->verify_lock); | |
738 | ||
739 | c->verify_data = mca_bucket_alloc(c, &ZERO_KEY, GFP_KERNEL); | |
740 | ||
741 | if (c->verify_data && | |
742 | c->verify_data->sets[0].data) | |
743 | list_del_init(&c->verify_data->list); | |
744 | else | |
745 | c->verify_data = NULL; | |
746 | #endif | |
747 | ||
7dc19d5a DC |
748 | c->shrink.count_objects = bch_mca_count; |
749 | c->shrink.scan_objects = bch_mca_scan; | |
cafe5635 KO |
750 | c->shrink.seeks = 4; |
751 | c->shrink.batch = c->btree_pages * 2; | |
752 | register_shrinker(&c->shrink); | |
753 | ||
754 | return 0; | |
755 | } | |
756 | ||
757 | /* Btree in memory cache - hash table */ | |
758 | ||
759 | static struct hlist_head *mca_hash(struct cache_set *c, struct bkey *k) | |
760 | { | |
761 | return &c->bucket_hash[hash_32(PTR_HASH(c, k), BUCKET_HASH_BITS)]; | |
762 | } | |
763 | ||
764 | static struct btree *mca_find(struct cache_set *c, struct bkey *k) | |
765 | { | |
766 | struct btree *b; | |
767 | ||
768 | rcu_read_lock(); | |
769 | hlist_for_each_entry_rcu(b, mca_hash(c, k), hash) | |
770 | if (PTR_HASH(c, &b->key) == PTR_HASH(c, k)) | |
771 | goto out; | |
772 | b = NULL; | |
773 | out: | |
774 | rcu_read_unlock(); | |
775 | return b; | |
776 | } | |
777 | ||
778 | static struct btree *mca_cannibalize(struct cache_set *c, struct bkey *k, | |
779 | int level, struct closure *cl) | |
780 | { | |
781 | int ret = -ENOMEM; | |
782 | struct btree *i; | |
783 | ||
c37511b8 KO |
784 | trace_bcache_btree_cache_cannibalize(c); |
785 | ||
cafe5635 KO |
786 | if (!cl) |
787 | return ERR_PTR(-ENOMEM); | |
788 | ||
789 | /* | |
790 | * Trying to free up some memory - i.e. reuse some btree nodes - may | |
791 | * require initiating IO to flush the dirty part of the node. If we're | |
792 | * running under generic_make_request(), that IO will never finish and | |
793 | * we would deadlock. Returning -EAGAIN causes the cache lookup code to | |
794 | * punt to workqueue and retry. | |
795 | */ | |
796 | if (current->bio_list) | |
797 | return ERR_PTR(-EAGAIN); | |
798 | ||
799 | if (c->try_harder && c->try_harder != cl) { | |
800 | closure_wait_event_async(&c->try_wait, cl, !c->try_harder); | |
801 | return ERR_PTR(-EAGAIN); | |
802 | } | |
803 | ||
cafe5635 KO |
804 | c->try_harder = cl; |
805 | c->try_harder_start = local_clock(); | |
806 | retry: | |
807 | list_for_each_entry_reverse(i, &c->btree_cache, list) { | |
808 | int r = mca_reap(i, cl, btree_order(k)); | |
809 | if (!r) | |
810 | return i; | |
811 | if (r != -ENOMEM) | |
812 | ret = r; | |
813 | } | |
814 | ||
815 | if (ret == -EAGAIN && | |
816 | closure_blocking(cl)) { | |
817 | mutex_unlock(&c->bucket_lock); | |
818 | closure_sync(cl); | |
819 | mutex_lock(&c->bucket_lock); | |
820 | goto retry; | |
821 | } | |
822 | ||
823 | return ERR_PTR(ret); | |
824 | } | |
825 | ||
826 | /* | |
827 | * We can only have one thread cannibalizing other cached btree nodes at a time, | |
828 | * or we'll deadlock. We use an open coded mutex to ensure that, which a | |
829 | * cannibalize_bucket() will take. This means every time we unlock the root of | |
830 | * the btree, we need to release this lock if we have it held. | |
831 | */ | |
832 | void bch_cannibalize_unlock(struct cache_set *c, struct closure *cl) | |
833 | { | |
834 | if (c->try_harder == cl) { | |
169ef1cf | 835 | bch_time_stats_update(&c->try_harder_time, c->try_harder_start); |
cafe5635 KO |
836 | c->try_harder = NULL; |
837 | __closure_wake_up(&c->try_wait); | |
838 | } | |
839 | } | |
840 | ||
841 | static struct btree *mca_alloc(struct cache_set *c, struct bkey *k, | |
842 | int level, struct closure *cl) | |
843 | { | |
844 | struct btree *b; | |
845 | ||
846 | lockdep_assert_held(&c->bucket_lock); | |
847 | ||
848 | if (mca_find(c, k)) | |
849 | return NULL; | |
850 | ||
851 | /* btree_free() doesn't free memory; it sticks the node on the end of | |
852 | * the list. Check if there's any freed nodes there: | |
853 | */ | |
854 | list_for_each_entry(b, &c->btree_cache_freeable, list) | |
855 | if (!mca_reap(b, NULL, btree_order(k))) | |
856 | goto out; | |
857 | ||
858 | /* We never free struct btree itself, just the memory that holds the on | |
859 | * disk node. Check the freed list before allocating a new one: | |
860 | */ | |
861 | list_for_each_entry(b, &c->btree_cache_freed, list) | |
862 | if (!mca_reap(b, NULL, 0)) { | |
863 | mca_data_alloc(b, k, __GFP_NOWARN|GFP_NOIO); | |
864 | if (!b->sets[0].data) | |
865 | goto err; | |
866 | else | |
867 | goto out; | |
868 | } | |
869 | ||
870 | b = mca_bucket_alloc(c, k, __GFP_NOWARN|GFP_NOIO); | |
871 | if (!b) | |
872 | goto err; | |
873 | ||
874 | BUG_ON(!down_write_trylock(&b->lock)); | |
875 | if (!b->sets->data) | |
876 | goto err; | |
877 | out: | |
878 | BUG_ON(!closure_is_unlocked(&b->io.cl)); | |
879 | ||
880 | bkey_copy(&b->key, k); | |
881 | list_move(&b->list, &c->btree_cache); | |
882 | hlist_del_init_rcu(&b->hash); | |
883 | hlist_add_head_rcu(&b->hash, mca_hash(c, k)); | |
884 | ||
885 | lock_set_subclass(&b->lock.dep_map, level + 1, _THIS_IP_); | |
886 | b->level = level; | |
d6fd3b11 | 887 | b->parent = (void *) ~0UL; |
cafe5635 KO |
888 | |
889 | mca_reinit(b); | |
890 | ||
891 | return b; | |
892 | err: | |
893 | if (b) | |
894 | rw_unlock(true, b); | |
895 | ||
896 | b = mca_cannibalize(c, k, level, cl); | |
897 | if (!IS_ERR(b)) | |
898 | goto out; | |
899 | ||
900 | return b; | |
901 | } | |
902 | ||
903 | /** | |
904 | * bch_btree_node_get - find a btree node in the cache and lock it, reading it | |
905 | * in from disk if necessary. | |
906 | * | |
907 | * If IO is necessary, it uses the closure embedded in struct btree_op to wait; | |
908 | * if that closure is in non blocking mode, will return -EAGAIN. | |
909 | * | |
910 | * The btree node will have either a read or a write lock held, depending on | |
911 | * level and op->lock. | |
912 | */ | |
913 | struct btree *bch_btree_node_get(struct cache_set *c, struct bkey *k, | |
914 | int level, struct btree_op *op) | |
915 | { | |
916 | int i = 0; | |
917 | bool write = level <= op->lock; | |
918 | struct btree *b; | |
919 | ||
920 | BUG_ON(level < 0); | |
921 | retry: | |
922 | b = mca_find(c, k); | |
923 | ||
924 | if (!b) { | |
57943511 KO |
925 | if (current->bio_list) |
926 | return ERR_PTR(-EAGAIN); | |
927 | ||
cafe5635 KO |
928 | mutex_lock(&c->bucket_lock); |
929 | b = mca_alloc(c, k, level, &op->cl); | |
930 | mutex_unlock(&c->bucket_lock); | |
931 | ||
932 | if (!b) | |
933 | goto retry; | |
934 | if (IS_ERR(b)) | |
935 | return b; | |
936 | ||
57943511 | 937 | bch_btree_node_read(b); |
cafe5635 KO |
938 | |
939 | if (!write) | |
940 | downgrade_write(&b->lock); | |
941 | } else { | |
942 | rw_lock(write, b, level); | |
943 | if (PTR_HASH(c, &b->key) != PTR_HASH(c, k)) { | |
944 | rw_unlock(write, b); | |
945 | goto retry; | |
946 | } | |
947 | BUG_ON(b->level != level); | |
948 | } | |
949 | ||
950 | b->accessed = 1; | |
951 | ||
952 | for (; i <= b->nsets && b->sets[i].size; i++) { | |
953 | prefetch(b->sets[i].tree); | |
954 | prefetch(b->sets[i].data); | |
955 | } | |
956 | ||
957 | for (; i <= b->nsets; i++) | |
958 | prefetch(b->sets[i].data); | |
959 | ||
57943511 | 960 | if (btree_node_io_error(b)) { |
cafe5635 | 961 | rw_unlock(write, b); |
57943511 KO |
962 | return ERR_PTR(-EIO); |
963 | } | |
964 | ||
965 | BUG_ON(!b->written); | |
cafe5635 KO |
966 | |
967 | return b; | |
968 | } | |
969 | ||
970 | static void btree_node_prefetch(struct cache_set *c, struct bkey *k, int level) | |
971 | { | |
972 | struct btree *b; | |
973 | ||
974 | mutex_lock(&c->bucket_lock); | |
975 | b = mca_alloc(c, k, level, NULL); | |
976 | mutex_unlock(&c->bucket_lock); | |
977 | ||
978 | if (!IS_ERR_OR_NULL(b)) { | |
57943511 | 979 | bch_btree_node_read(b); |
cafe5635 KO |
980 | rw_unlock(true, b); |
981 | } | |
982 | } | |
983 | ||
984 | /* Btree alloc */ | |
985 | ||
986 | static void btree_node_free(struct btree *b, struct btree_op *op) | |
987 | { | |
988 | unsigned i; | |
989 | ||
c37511b8 KO |
990 | trace_bcache_btree_node_free(b); |
991 | ||
cafe5635 KO |
992 | /* |
993 | * The BUG_ON() in btree_node_get() implies that we must have a write | |
994 | * lock on parent to free or even invalidate a node | |
995 | */ | |
996 | BUG_ON(op->lock <= b->level); | |
997 | BUG_ON(b == b->c->root); | |
cafe5635 KO |
998 | |
999 | if (btree_node_dirty(b)) | |
1000 | btree_complete_write(b, btree_current_write(b)); | |
1001 | clear_bit(BTREE_NODE_dirty, &b->flags); | |
1002 | ||
cafe5635 KO |
1003 | cancel_delayed_work(&b->work); |
1004 | ||
1005 | mutex_lock(&b->c->bucket_lock); | |
1006 | ||
1007 | for (i = 0; i < KEY_PTRS(&b->key); i++) { | |
1008 | BUG_ON(atomic_read(&PTR_BUCKET(b->c, &b->key, i)->pin)); | |
1009 | ||
1010 | bch_inc_gen(PTR_CACHE(b->c, &b->key, i), | |
1011 | PTR_BUCKET(b->c, &b->key, i)); | |
1012 | } | |
1013 | ||
1014 | bch_bucket_free(b->c, &b->key); | |
1015 | mca_bucket_free(b); | |
1016 | mutex_unlock(&b->c->bucket_lock); | |
1017 | } | |
1018 | ||
1019 | struct btree *bch_btree_node_alloc(struct cache_set *c, int level, | |
1020 | struct closure *cl) | |
1021 | { | |
1022 | BKEY_PADDED(key) k; | |
1023 | struct btree *b = ERR_PTR(-EAGAIN); | |
1024 | ||
1025 | mutex_lock(&c->bucket_lock); | |
1026 | retry: | |
1027 | if (__bch_bucket_alloc_set(c, WATERMARK_METADATA, &k.key, 1, cl)) | |
1028 | goto err; | |
1029 | ||
1030 | SET_KEY_SIZE(&k.key, c->btree_pages * PAGE_SECTORS); | |
1031 | ||
1032 | b = mca_alloc(c, &k.key, level, cl); | |
1033 | if (IS_ERR(b)) | |
1034 | goto err_free; | |
1035 | ||
1036 | if (!b) { | |
b1a67b0f KO |
1037 | cache_bug(c, |
1038 | "Tried to allocate bucket that was in btree cache"); | |
cafe5635 KO |
1039 | __bkey_put(c, &k.key); |
1040 | goto retry; | |
1041 | } | |
1042 | ||
cafe5635 KO |
1043 | b->accessed = 1; |
1044 | bch_bset_init_next(b); | |
1045 | ||
1046 | mutex_unlock(&c->bucket_lock); | |
c37511b8 KO |
1047 | |
1048 | trace_bcache_btree_node_alloc(b); | |
cafe5635 KO |
1049 | return b; |
1050 | err_free: | |
1051 | bch_bucket_free(c, &k.key); | |
1052 | __bkey_put(c, &k.key); | |
1053 | err: | |
1054 | mutex_unlock(&c->bucket_lock); | |
c37511b8 KO |
1055 | |
1056 | trace_bcache_btree_node_alloc_fail(b); | |
cafe5635 KO |
1057 | return b; |
1058 | } | |
1059 | ||
1060 | static struct btree *btree_node_alloc_replacement(struct btree *b, | |
1061 | struct closure *cl) | |
1062 | { | |
1063 | struct btree *n = bch_btree_node_alloc(b->c, b->level, cl); | |
1064 | if (!IS_ERR_OR_NULL(n)) | |
1065 | bch_btree_sort_into(b, n); | |
1066 | ||
1067 | return n; | |
1068 | } | |
1069 | ||
1070 | /* Garbage collection */ | |
1071 | ||
1072 | uint8_t __bch_btree_mark_key(struct cache_set *c, int level, struct bkey *k) | |
1073 | { | |
1074 | uint8_t stale = 0; | |
1075 | unsigned i; | |
1076 | struct bucket *g; | |
1077 | ||
1078 | /* | |
1079 | * ptr_invalid() can't return true for the keys that mark btree nodes as | |
1080 | * freed, but since ptr_bad() returns true we'll never actually use them | |
1081 | * for anything and thus we don't want mark their pointers here | |
1082 | */ | |
1083 | if (!bkey_cmp(k, &ZERO_KEY)) | |
1084 | return stale; | |
1085 | ||
1086 | for (i = 0; i < KEY_PTRS(k); i++) { | |
1087 | if (!ptr_available(c, k, i)) | |
1088 | continue; | |
1089 | ||
1090 | g = PTR_BUCKET(c, k, i); | |
1091 | ||
1092 | if (gen_after(g->gc_gen, PTR_GEN(k, i))) | |
1093 | g->gc_gen = PTR_GEN(k, i); | |
1094 | ||
1095 | if (ptr_stale(c, k, i)) { | |
1096 | stale = max(stale, ptr_stale(c, k, i)); | |
1097 | continue; | |
1098 | } | |
1099 | ||
1100 | cache_bug_on(GC_MARK(g) && | |
1101 | (GC_MARK(g) == GC_MARK_METADATA) != (level != 0), | |
1102 | c, "inconsistent ptrs: mark = %llu, level = %i", | |
1103 | GC_MARK(g), level); | |
1104 | ||
1105 | if (level) | |
1106 | SET_GC_MARK(g, GC_MARK_METADATA); | |
1107 | else if (KEY_DIRTY(k)) | |
1108 | SET_GC_MARK(g, GC_MARK_DIRTY); | |
1109 | ||
1110 | /* guard against overflow */ | |
1111 | SET_GC_SECTORS_USED(g, min_t(unsigned, | |
1112 | GC_SECTORS_USED(g) + KEY_SIZE(k), | |
1113 | (1 << 14) - 1)); | |
1114 | ||
1115 | BUG_ON(!GC_SECTORS_USED(g)); | |
1116 | } | |
1117 | ||
1118 | return stale; | |
1119 | } | |
1120 | ||
1121 | #define btree_mark_key(b, k) __bch_btree_mark_key(b->c, b->level, k) | |
1122 | ||
1123 | static int btree_gc_mark_node(struct btree *b, unsigned *keys, | |
1124 | struct gc_stat *gc) | |
1125 | { | |
1126 | uint8_t stale = 0; | |
1127 | unsigned last_dev = -1; | |
1128 | struct bcache_device *d = NULL; | |
1129 | struct bkey *k; | |
1130 | struct btree_iter iter; | |
1131 | struct bset_tree *t; | |
1132 | ||
1133 | gc->nodes++; | |
1134 | ||
1135 | for_each_key_filter(b, k, &iter, bch_ptr_invalid) { | |
1136 | if (last_dev != KEY_INODE(k)) { | |
1137 | last_dev = KEY_INODE(k); | |
1138 | ||
1139 | d = KEY_INODE(k) < b->c->nr_uuids | |
1140 | ? b->c->devices[last_dev] | |
1141 | : NULL; | |
1142 | } | |
1143 | ||
1144 | stale = max(stale, btree_mark_key(b, k)); | |
1145 | ||
1146 | if (bch_ptr_bad(b, k)) | |
1147 | continue; | |
1148 | ||
1149 | *keys += bkey_u64s(k); | |
1150 | ||
1151 | gc->key_bytes += bkey_u64s(k); | |
1152 | gc->nkeys++; | |
1153 | ||
1154 | gc->data += KEY_SIZE(k); | |
444fc0b6 | 1155 | if (KEY_DIRTY(k)) |
cafe5635 | 1156 | gc->dirty += KEY_SIZE(k); |
cafe5635 KO |
1157 | } |
1158 | ||
1159 | for (t = b->sets; t <= &b->sets[b->nsets]; t++) | |
1160 | btree_bug_on(t->size && | |
1161 | bset_written(b, t) && | |
1162 | bkey_cmp(&b->key, &t->end) < 0, | |
1163 | b, "found short btree key in gc"); | |
1164 | ||
1165 | return stale; | |
1166 | } | |
1167 | ||
1168 | static struct btree *btree_gc_alloc(struct btree *b, struct bkey *k, | |
1169 | struct btree_op *op) | |
1170 | { | |
1171 | /* | |
1172 | * We block priorities from being written for the duration of garbage | |
1173 | * collection, so we can't sleep in btree_alloc() -> | |
1174 | * bch_bucket_alloc_set(), or we'd risk deadlock - so we don't pass it | |
1175 | * our closure. | |
1176 | */ | |
1177 | struct btree *n = btree_node_alloc_replacement(b, NULL); | |
1178 | ||
1179 | if (!IS_ERR_OR_NULL(n)) { | |
1180 | swap(b, n); | |
57943511 | 1181 | __bkey_put(b->c, &b->key); |
cafe5635 KO |
1182 | |
1183 | memcpy(k->ptr, b->key.ptr, | |
1184 | sizeof(uint64_t) * KEY_PTRS(&b->key)); | |
1185 | ||
cafe5635 KO |
1186 | btree_node_free(n, op); |
1187 | up_write(&n->lock); | |
1188 | } | |
1189 | ||
1190 | return b; | |
1191 | } | |
1192 | ||
1193 | /* | |
1194 | * Leaving this at 2 until we've got incremental garbage collection done; it | |
1195 | * could be higher (and has been tested with 4) except that garbage collection | |
1196 | * could take much longer, adversely affecting latency. | |
1197 | */ | |
1198 | #define GC_MERGE_NODES 2U | |
1199 | ||
1200 | struct gc_merge_info { | |
1201 | struct btree *b; | |
1202 | struct bkey *k; | |
1203 | unsigned keys; | |
1204 | }; | |
1205 | ||
1206 | static void btree_gc_coalesce(struct btree *b, struct btree_op *op, | |
1207 | struct gc_stat *gc, struct gc_merge_info *r) | |
1208 | { | |
1209 | unsigned nodes = 0, keys = 0, blocks; | |
1210 | int i; | |
1211 | ||
1212 | while (nodes < GC_MERGE_NODES && r[nodes].b) | |
1213 | keys += r[nodes++].keys; | |
1214 | ||
1215 | blocks = btree_default_blocks(b->c) * 2 / 3; | |
1216 | ||
1217 | if (nodes < 2 || | |
1218 | __set_blocks(b->sets[0].data, keys, b->c) > blocks * (nodes - 1)) | |
1219 | return; | |
1220 | ||
1221 | for (i = nodes - 1; i >= 0; --i) { | |
1222 | if (r[i].b->written) | |
1223 | r[i].b = btree_gc_alloc(r[i].b, r[i].k, op); | |
1224 | ||
1225 | if (r[i].b->written) | |
1226 | return; | |
1227 | } | |
1228 | ||
1229 | for (i = nodes - 1; i > 0; --i) { | |
1230 | struct bset *n1 = r[i].b->sets->data; | |
1231 | struct bset *n2 = r[i - 1].b->sets->data; | |
1232 | struct bkey *k, *last = NULL; | |
1233 | ||
1234 | keys = 0; | |
1235 | ||
1236 | if (i == 1) { | |
1237 | /* | |
1238 | * Last node we're not getting rid of - we're getting | |
1239 | * rid of the node at r[0]. Have to try and fit all of | |
1240 | * the remaining keys into this node; we can't ensure | |
1241 | * they will always fit due to rounding and variable | |
1242 | * length keys (shouldn't be possible in practice, | |
1243 | * though) | |
1244 | */ | |
1245 | if (__set_blocks(n1, n1->keys + r->keys, | |
1246 | b->c) > btree_blocks(r[i].b)) | |
1247 | return; | |
1248 | ||
1249 | keys = n2->keys; | |
1250 | last = &r->b->key; | |
1251 | } else | |
1252 | for (k = n2->start; | |
1253 | k < end(n2); | |
1254 | k = bkey_next(k)) { | |
1255 | if (__set_blocks(n1, n1->keys + keys + | |
1256 | bkey_u64s(k), b->c) > blocks) | |
1257 | break; | |
1258 | ||
1259 | last = k; | |
1260 | keys += bkey_u64s(k); | |
1261 | } | |
1262 | ||
1263 | BUG_ON(__set_blocks(n1, n1->keys + keys, | |
1264 | b->c) > btree_blocks(r[i].b)); | |
1265 | ||
1266 | if (last) { | |
1267 | bkey_copy_key(&r[i].b->key, last); | |
1268 | bkey_copy_key(r[i].k, last); | |
1269 | } | |
1270 | ||
1271 | memcpy(end(n1), | |
1272 | n2->start, | |
1273 | (void *) node(n2, keys) - (void *) n2->start); | |
1274 | ||
1275 | n1->keys += keys; | |
1276 | ||
1277 | memmove(n2->start, | |
1278 | node(n2, keys), | |
1279 | (void *) end(n2) - (void *) node(n2, keys)); | |
1280 | ||
1281 | n2->keys -= keys; | |
1282 | ||
1283 | r[i].keys = n1->keys; | |
1284 | r[i - 1].keys = n2->keys; | |
1285 | } | |
1286 | ||
1287 | btree_node_free(r->b, op); | |
1288 | up_write(&r->b->lock); | |
1289 | ||
c37511b8 | 1290 | trace_bcache_btree_gc_coalesce(nodes); |
cafe5635 KO |
1291 | |
1292 | gc->nodes--; | |
1293 | nodes--; | |
1294 | ||
1295 | memmove(&r[0], &r[1], sizeof(struct gc_merge_info) * nodes); | |
1296 | memset(&r[nodes], 0, sizeof(struct gc_merge_info)); | |
1297 | } | |
1298 | ||
1299 | static int btree_gc_recurse(struct btree *b, struct btree_op *op, | |
1300 | struct closure *writes, struct gc_stat *gc) | |
1301 | { | |
1302 | void write(struct btree *r) | |
1303 | { | |
1304 | if (!r->written) | |
57943511 KO |
1305 | bch_btree_node_write(r, &op->cl); |
1306 | else if (btree_node_dirty(r)) | |
1307 | bch_btree_node_write(r, writes); | |
cafe5635 KO |
1308 | |
1309 | up_write(&r->lock); | |
1310 | } | |
1311 | ||
1312 | int ret = 0, stale; | |
1313 | unsigned i; | |
1314 | struct gc_merge_info r[GC_MERGE_NODES]; | |
1315 | ||
1316 | memset(r, 0, sizeof(r)); | |
1317 | ||
1318 | while ((r->k = bch_next_recurse_key(b, &b->c->gc_done))) { | |
1319 | r->b = bch_btree_node_get(b->c, r->k, b->level - 1, op); | |
1320 | ||
1321 | if (IS_ERR(r->b)) { | |
1322 | ret = PTR_ERR(r->b); | |
1323 | break; | |
1324 | } | |
1325 | ||
1326 | r->keys = 0; | |
1327 | stale = btree_gc_mark_node(r->b, &r->keys, gc); | |
1328 | ||
1329 | if (!b->written && | |
1330 | (r->b->level || stale > 10 || | |
1331 | b->c->gc_always_rewrite)) | |
1332 | r->b = btree_gc_alloc(r->b, r->k, op); | |
1333 | ||
1334 | if (r->b->level) | |
1335 | ret = btree_gc_recurse(r->b, op, writes, gc); | |
1336 | ||
1337 | if (ret) { | |
1338 | write(r->b); | |
1339 | break; | |
1340 | } | |
1341 | ||
1342 | bkey_copy_key(&b->c->gc_done, r->k); | |
1343 | ||
1344 | if (!b->written) | |
1345 | btree_gc_coalesce(b, op, gc, r); | |
1346 | ||
1347 | if (r[GC_MERGE_NODES - 1].b) | |
1348 | write(r[GC_MERGE_NODES - 1].b); | |
1349 | ||
1350 | memmove(&r[1], &r[0], | |
1351 | sizeof(struct gc_merge_info) * (GC_MERGE_NODES - 1)); | |
1352 | ||
1353 | /* When we've got incremental GC working, we'll want to do | |
1354 | * if (should_resched()) | |
1355 | * return -EAGAIN; | |
1356 | */ | |
1357 | cond_resched(); | |
1358 | #if 0 | |
1359 | if (need_resched()) { | |
1360 | ret = -EAGAIN; | |
1361 | break; | |
1362 | } | |
1363 | #endif | |
1364 | } | |
1365 | ||
1366 | for (i = 1; i < GC_MERGE_NODES && r[i].b; i++) | |
1367 | write(r[i].b); | |
1368 | ||
1369 | /* Might have freed some children, must remove their keys */ | |
1370 | if (!b->written) | |
1371 | bch_btree_sort(b); | |
1372 | ||
1373 | return ret; | |
1374 | } | |
1375 | ||
1376 | static int bch_btree_gc_root(struct btree *b, struct btree_op *op, | |
1377 | struct closure *writes, struct gc_stat *gc) | |
1378 | { | |
1379 | struct btree *n = NULL; | |
1380 | unsigned keys = 0; | |
1381 | int ret = 0, stale = btree_gc_mark_node(b, &keys, gc); | |
1382 | ||
1383 | if (b->level || stale > 10) | |
1384 | n = btree_node_alloc_replacement(b, NULL); | |
1385 | ||
1386 | if (!IS_ERR_OR_NULL(n)) | |
1387 | swap(b, n); | |
1388 | ||
1389 | if (b->level) | |
1390 | ret = btree_gc_recurse(b, op, writes, gc); | |
1391 | ||
1392 | if (!b->written || btree_node_dirty(b)) { | |
57943511 | 1393 | bch_btree_node_write(b, n ? &op->cl : NULL); |
cafe5635 KO |
1394 | } |
1395 | ||
1396 | if (!IS_ERR_OR_NULL(n)) { | |
1397 | closure_sync(&op->cl); | |
1398 | bch_btree_set_root(b); | |
1399 | btree_node_free(n, op); | |
1400 | rw_unlock(true, b); | |
1401 | } | |
1402 | ||
1403 | return ret; | |
1404 | } | |
1405 | ||
1406 | static void btree_gc_start(struct cache_set *c) | |
1407 | { | |
1408 | struct cache *ca; | |
1409 | struct bucket *b; | |
cafe5635 KO |
1410 | unsigned i; |
1411 | ||
1412 | if (!c->gc_mark_valid) | |
1413 | return; | |
1414 | ||
1415 | mutex_lock(&c->bucket_lock); | |
1416 | ||
1417 | c->gc_mark_valid = 0; | |
1418 | c->gc_done = ZERO_KEY; | |
1419 | ||
1420 | for_each_cache(ca, c, i) | |
1421 | for_each_bucket(b, ca) { | |
1422 | b->gc_gen = b->gen; | |
29ebf465 | 1423 | if (!atomic_read(&b->pin)) { |
cafe5635 | 1424 | SET_GC_MARK(b, GC_MARK_RECLAIMABLE); |
29ebf465 KO |
1425 | SET_GC_SECTORS_USED(b, 0); |
1426 | } | |
cafe5635 KO |
1427 | } |
1428 | ||
cafe5635 KO |
1429 | mutex_unlock(&c->bucket_lock); |
1430 | } | |
1431 | ||
1432 | size_t bch_btree_gc_finish(struct cache_set *c) | |
1433 | { | |
1434 | size_t available = 0; | |
1435 | struct bucket *b; | |
1436 | struct cache *ca; | |
cafe5635 KO |
1437 | unsigned i; |
1438 | ||
1439 | mutex_lock(&c->bucket_lock); | |
1440 | ||
1441 | set_gc_sectors(c); | |
1442 | c->gc_mark_valid = 1; | |
1443 | c->need_gc = 0; | |
1444 | ||
1445 | if (c->root) | |
1446 | for (i = 0; i < KEY_PTRS(&c->root->key); i++) | |
1447 | SET_GC_MARK(PTR_BUCKET(c, &c->root->key, i), | |
1448 | GC_MARK_METADATA); | |
1449 | ||
1450 | for (i = 0; i < KEY_PTRS(&c->uuid_bucket); i++) | |
1451 | SET_GC_MARK(PTR_BUCKET(c, &c->uuid_bucket, i), | |
1452 | GC_MARK_METADATA); | |
1453 | ||
1454 | for_each_cache(ca, c, i) { | |
1455 | uint64_t *i; | |
1456 | ||
1457 | ca->invalidate_needs_gc = 0; | |
1458 | ||
1459 | for (i = ca->sb.d; i < ca->sb.d + ca->sb.keys; i++) | |
1460 | SET_GC_MARK(ca->buckets + *i, GC_MARK_METADATA); | |
1461 | ||
1462 | for (i = ca->prio_buckets; | |
1463 | i < ca->prio_buckets + prio_buckets(ca) * 2; i++) | |
1464 | SET_GC_MARK(ca->buckets + *i, GC_MARK_METADATA); | |
1465 | ||
1466 | for_each_bucket(b, ca) { | |
1467 | b->last_gc = b->gc_gen; | |
1468 | c->need_gc = max(c->need_gc, bucket_gc_gen(b)); | |
1469 | ||
1470 | if (!atomic_read(&b->pin) && | |
1471 | GC_MARK(b) == GC_MARK_RECLAIMABLE) { | |
1472 | available++; | |
1473 | if (!GC_SECTORS_USED(b)) | |
1474 | bch_bucket_add_unused(ca, b); | |
1475 | } | |
1476 | } | |
1477 | } | |
1478 | ||
cafe5635 KO |
1479 | mutex_unlock(&c->bucket_lock); |
1480 | return available; | |
1481 | } | |
1482 | ||
1483 | static void bch_btree_gc(struct closure *cl) | |
1484 | { | |
1485 | struct cache_set *c = container_of(cl, struct cache_set, gc.cl); | |
1486 | int ret; | |
1487 | unsigned long available; | |
1488 | struct gc_stat stats; | |
1489 | struct closure writes; | |
1490 | struct btree_op op; | |
cafe5635 | 1491 | uint64_t start_time = local_clock(); |
57943511 | 1492 | |
c37511b8 | 1493 | trace_bcache_gc_start(c); |
cafe5635 KO |
1494 | |
1495 | memset(&stats, 0, sizeof(struct gc_stat)); | |
1496 | closure_init_stack(&writes); | |
1497 | bch_btree_op_init_stack(&op); | |
1498 | op.lock = SHRT_MAX; | |
1499 | ||
1500 | btree_gc_start(c); | |
1501 | ||
57943511 KO |
1502 | atomic_inc(&c->prio_blocked); |
1503 | ||
cafe5635 KO |
1504 | ret = btree_root(gc_root, c, &op, &writes, &stats); |
1505 | closure_sync(&op.cl); | |
1506 | closure_sync(&writes); | |
1507 | ||
1508 | if (ret) { | |
cafe5635 | 1509 | pr_warn("gc failed!"); |
cafe5635 KO |
1510 | continue_at(cl, bch_btree_gc, bch_gc_wq); |
1511 | } | |
1512 | ||
1513 | /* Possibly wait for new UUIDs or whatever to hit disk */ | |
1514 | bch_journal_meta(c, &op.cl); | |
1515 | closure_sync(&op.cl); | |
1516 | ||
1517 | available = bch_btree_gc_finish(c); | |
1518 | ||
57943511 KO |
1519 | atomic_dec(&c->prio_blocked); |
1520 | wake_up_allocators(c); | |
1521 | ||
169ef1cf | 1522 | bch_time_stats_update(&c->btree_gc_time, start_time); |
cafe5635 KO |
1523 | |
1524 | stats.key_bytes *= sizeof(uint64_t); | |
1525 | stats.dirty <<= 9; | |
1526 | stats.data <<= 9; | |
1527 | stats.in_use = (c->nbuckets - available) * 100 / c->nbuckets; | |
1528 | memcpy(&c->gc_stats, &stats, sizeof(struct gc_stat)); | |
cafe5635 | 1529 | |
c37511b8 | 1530 | trace_bcache_gc_end(c); |
cafe5635 KO |
1531 | |
1532 | continue_at(cl, bch_moving_gc, bch_gc_wq); | |
1533 | } | |
1534 | ||
1535 | void bch_queue_gc(struct cache_set *c) | |
1536 | { | |
1537 | closure_trylock_call(&c->gc.cl, bch_btree_gc, bch_gc_wq, &c->cl); | |
1538 | } | |
1539 | ||
1540 | /* Initial partial gc */ | |
1541 | ||
1542 | static int bch_btree_check_recurse(struct btree *b, struct btree_op *op, | |
1543 | unsigned long **seen) | |
1544 | { | |
1545 | int ret; | |
1546 | unsigned i; | |
1547 | struct bkey *k; | |
1548 | struct bucket *g; | |
1549 | struct btree_iter iter; | |
1550 | ||
1551 | for_each_key_filter(b, k, &iter, bch_ptr_invalid) { | |
1552 | for (i = 0; i < KEY_PTRS(k); i++) { | |
1553 | if (!ptr_available(b->c, k, i)) | |
1554 | continue; | |
1555 | ||
1556 | g = PTR_BUCKET(b->c, k, i); | |
1557 | ||
1558 | if (!__test_and_set_bit(PTR_BUCKET_NR(b->c, k, i), | |
1559 | seen[PTR_DEV(k, i)]) || | |
1560 | !ptr_stale(b->c, k, i)) { | |
1561 | g->gen = PTR_GEN(k, i); | |
1562 | ||
1563 | if (b->level) | |
1564 | g->prio = BTREE_PRIO; | |
1565 | else if (g->prio == BTREE_PRIO) | |
1566 | g->prio = INITIAL_PRIO; | |
1567 | } | |
1568 | } | |
1569 | ||
1570 | btree_mark_key(b, k); | |
1571 | } | |
1572 | ||
1573 | if (b->level) { | |
1574 | k = bch_next_recurse_key(b, &ZERO_KEY); | |
1575 | ||
1576 | while (k) { | |
1577 | struct bkey *p = bch_next_recurse_key(b, k); | |
1578 | if (p) | |
1579 | btree_node_prefetch(b->c, p, b->level - 1); | |
1580 | ||
1581 | ret = btree(check_recurse, k, b, op, seen); | |
1582 | if (ret) | |
1583 | return ret; | |
1584 | ||
1585 | k = p; | |
1586 | } | |
1587 | } | |
1588 | ||
1589 | return 0; | |
1590 | } | |
1591 | ||
1592 | int bch_btree_check(struct cache_set *c, struct btree_op *op) | |
1593 | { | |
1594 | int ret = -ENOMEM; | |
1595 | unsigned i; | |
1596 | unsigned long *seen[MAX_CACHES_PER_SET]; | |
1597 | ||
1598 | memset(seen, 0, sizeof(seen)); | |
1599 | ||
1600 | for (i = 0; c->cache[i]; i++) { | |
1601 | size_t n = DIV_ROUND_UP(c->cache[i]->sb.nbuckets, 8); | |
1602 | seen[i] = kmalloc(n, GFP_KERNEL); | |
1603 | if (!seen[i]) | |
1604 | goto err; | |
1605 | ||
1606 | /* Disables the seen array until prio_read() uses it too */ | |
1607 | memset(seen[i], 0xFF, n); | |
1608 | } | |
1609 | ||
1610 | ret = btree_root(check_recurse, c, op, seen); | |
1611 | err: | |
1612 | for (i = 0; i < MAX_CACHES_PER_SET; i++) | |
1613 | kfree(seen[i]); | |
1614 | return ret; | |
1615 | } | |
1616 | ||
1617 | /* Btree insertion */ | |
1618 | ||
1619 | static void shift_keys(struct btree *b, struct bkey *where, struct bkey *insert) | |
1620 | { | |
1621 | struct bset *i = b->sets[b->nsets].data; | |
1622 | ||
1623 | memmove((uint64_t *) where + bkey_u64s(insert), | |
1624 | where, | |
1625 | (void *) end(i) - (void *) where); | |
1626 | ||
1627 | i->keys += bkey_u64s(insert); | |
1628 | bkey_copy(where, insert); | |
1629 | bch_bset_fix_lookup_table(b, where); | |
1630 | } | |
1631 | ||
1632 | static bool fix_overlapping_extents(struct btree *b, | |
1633 | struct bkey *insert, | |
1634 | struct btree_iter *iter, | |
1635 | struct btree_op *op) | |
1636 | { | |
279afbad | 1637 | void subtract_dirty(struct bkey *k, uint64_t offset, int sectors) |
cafe5635 | 1638 | { |
279afbad KO |
1639 | if (KEY_DIRTY(k)) |
1640 | bcache_dev_sectors_dirty_add(b->c, KEY_INODE(k), | |
1641 | offset, -sectors); | |
cafe5635 KO |
1642 | } |
1643 | ||
279afbad | 1644 | uint64_t old_offset; |
cafe5635 KO |
1645 | unsigned old_size, sectors_found = 0; |
1646 | ||
1647 | while (1) { | |
1648 | struct bkey *k = bch_btree_iter_next(iter); | |
1649 | if (!k || | |
1650 | bkey_cmp(&START_KEY(k), insert) >= 0) | |
1651 | break; | |
1652 | ||
1653 | if (bkey_cmp(k, &START_KEY(insert)) <= 0) | |
1654 | continue; | |
1655 | ||
279afbad | 1656 | old_offset = KEY_START(k); |
cafe5635 KO |
1657 | old_size = KEY_SIZE(k); |
1658 | ||
1659 | /* | |
1660 | * We might overlap with 0 size extents; we can't skip these | |
1661 | * because if they're in the set we're inserting to we have to | |
1662 | * adjust them so they don't overlap with the key we're | |
1663 | * inserting. But we don't want to check them for BTREE_REPLACE | |
1664 | * operations. | |
1665 | */ | |
1666 | ||
1667 | if (op->type == BTREE_REPLACE && | |
1668 | KEY_SIZE(k)) { | |
1669 | /* | |
1670 | * k might have been split since we inserted/found the | |
1671 | * key we're replacing | |
1672 | */ | |
1673 | unsigned i; | |
1674 | uint64_t offset = KEY_START(k) - | |
1675 | KEY_START(&op->replace); | |
1676 | ||
1677 | /* But it must be a subset of the replace key */ | |
1678 | if (KEY_START(k) < KEY_START(&op->replace) || | |
1679 | KEY_OFFSET(k) > KEY_OFFSET(&op->replace)) | |
1680 | goto check_failed; | |
1681 | ||
1682 | /* We didn't find a key that we were supposed to */ | |
1683 | if (KEY_START(k) > KEY_START(insert) + sectors_found) | |
1684 | goto check_failed; | |
1685 | ||
1686 | if (KEY_PTRS(&op->replace) != KEY_PTRS(k)) | |
1687 | goto check_failed; | |
1688 | ||
1689 | /* skip past gen */ | |
1690 | offset <<= 8; | |
1691 | ||
1692 | BUG_ON(!KEY_PTRS(&op->replace)); | |
1693 | ||
1694 | for (i = 0; i < KEY_PTRS(&op->replace); i++) | |
1695 | if (k->ptr[i] != op->replace.ptr[i] + offset) | |
1696 | goto check_failed; | |
1697 | ||
1698 | sectors_found = KEY_OFFSET(k) - KEY_START(insert); | |
1699 | } | |
1700 | ||
1701 | if (bkey_cmp(insert, k) < 0 && | |
1702 | bkey_cmp(&START_KEY(insert), &START_KEY(k)) > 0) { | |
1703 | /* | |
1704 | * We overlapped in the middle of an existing key: that | |
1705 | * means we have to split the old key. But we have to do | |
1706 | * slightly different things depending on whether the | |
1707 | * old key has been written out yet. | |
1708 | */ | |
1709 | ||
1710 | struct bkey *top; | |
1711 | ||
279afbad | 1712 | subtract_dirty(k, KEY_START(insert), KEY_SIZE(insert)); |
cafe5635 KO |
1713 | |
1714 | if (bkey_written(b, k)) { | |
1715 | /* | |
1716 | * We insert a new key to cover the top of the | |
1717 | * old key, and the old key is modified in place | |
1718 | * to represent the bottom split. | |
1719 | * | |
1720 | * It's completely arbitrary whether the new key | |
1721 | * is the top or the bottom, but it has to match | |
1722 | * up with what btree_sort_fixup() does - it | |
1723 | * doesn't check for this kind of overlap, it | |
1724 | * depends on us inserting a new key for the top | |
1725 | * here. | |
1726 | */ | |
1727 | top = bch_bset_search(b, &b->sets[b->nsets], | |
1728 | insert); | |
1729 | shift_keys(b, top, k); | |
1730 | } else { | |
1731 | BKEY_PADDED(key) temp; | |
1732 | bkey_copy(&temp.key, k); | |
1733 | shift_keys(b, k, &temp.key); | |
1734 | top = bkey_next(k); | |
1735 | } | |
1736 | ||
1737 | bch_cut_front(insert, top); | |
1738 | bch_cut_back(&START_KEY(insert), k); | |
1739 | bch_bset_fix_invalidated_key(b, k); | |
1740 | return false; | |
1741 | } | |
1742 | ||
1743 | if (bkey_cmp(insert, k) < 0) { | |
1744 | bch_cut_front(insert, k); | |
1745 | } else { | |
1fa8455d KO |
1746 | if (bkey_cmp(&START_KEY(insert), &START_KEY(k)) > 0) |
1747 | old_offset = KEY_START(insert); | |
1748 | ||
cafe5635 KO |
1749 | if (bkey_written(b, k) && |
1750 | bkey_cmp(&START_KEY(insert), &START_KEY(k)) <= 0) { | |
1751 | /* | |
1752 | * Completely overwrote, so we don't have to | |
1753 | * invalidate the binary search tree | |
1754 | */ | |
1755 | bch_cut_front(k, k); | |
1756 | } else { | |
1757 | __bch_cut_back(&START_KEY(insert), k); | |
1758 | bch_bset_fix_invalidated_key(b, k); | |
1759 | } | |
1760 | } | |
1761 | ||
279afbad | 1762 | subtract_dirty(k, old_offset, old_size - KEY_SIZE(k)); |
cafe5635 KO |
1763 | } |
1764 | ||
1765 | check_failed: | |
1766 | if (op->type == BTREE_REPLACE) { | |
1767 | if (!sectors_found) { | |
1768 | op->insert_collision = true; | |
1769 | return true; | |
1770 | } else if (sectors_found < KEY_SIZE(insert)) { | |
1771 | SET_KEY_OFFSET(insert, KEY_OFFSET(insert) - | |
1772 | (KEY_SIZE(insert) - sectors_found)); | |
1773 | SET_KEY_SIZE(insert, sectors_found); | |
1774 | } | |
1775 | } | |
1776 | ||
1777 | return false; | |
1778 | } | |
1779 | ||
1780 | static bool btree_insert_key(struct btree *b, struct btree_op *op, | |
1781 | struct bkey *k) | |
1782 | { | |
1783 | struct bset *i = b->sets[b->nsets].data; | |
1784 | struct bkey *m, *prev; | |
85b1492e | 1785 | unsigned status = BTREE_INSERT_STATUS_INSERT; |
cafe5635 KO |
1786 | |
1787 | BUG_ON(bkey_cmp(k, &b->key) > 0); | |
1788 | BUG_ON(b->level && !KEY_PTRS(k)); | |
1789 | BUG_ON(!b->level && !KEY_OFFSET(k)); | |
1790 | ||
1791 | if (!b->level) { | |
1792 | struct btree_iter iter; | |
1793 | struct bkey search = KEY(KEY_INODE(k), KEY_START(k), 0); | |
1794 | ||
1795 | /* | |
1796 | * bset_search() returns the first key that is strictly greater | |
1797 | * than the search key - but for back merging, we want to find | |
1798 | * the first key that is greater than or equal to KEY_START(k) - | |
1799 | * unless KEY_START(k) is 0. | |
1800 | */ | |
1801 | if (KEY_OFFSET(&search)) | |
1802 | SET_KEY_OFFSET(&search, KEY_OFFSET(&search) - 1); | |
1803 | ||
1804 | prev = NULL; | |
1805 | m = bch_btree_iter_init(b, &iter, &search); | |
1806 | ||
1807 | if (fix_overlapping_extents(b, k, &iter, op)) | |
1808 | return false; | |
1809 | ||
1fa8455d KO |
1810 | if (KEY_DIRTY(k)) |
1811 | bcache_dev_sectors_dirty_add(b->c, KEY_INODE(k), | |
1812 | KEY_START(k), KEY_SIZE(k)); | |
1813 | ||
cafe5635 KO |
1814 | while (m != end(i) && |
1815 | bkey_cmp(k, &START_KEY(m)) > 0) | |
1816 | prev = m, m = bkey_next(m); | |
1817 | ||
1818 | if (key_merging_disabled(b->c)) | |
1819 | goto insert; | |
1820 | ||
1821 | /* prev is in the tree, if we merge we're done */ | |
85b1492e | 1822 | status = BTREE_INSERT_STATUS_BACK_MERGE; |
cafe5635 KO |
1823 | if (prev && |
1824 | bch_bkey_try_merge(b, prev, k)) | |
1825 | goto merged; | |
1826 | ||
85b1492e | 1827 | status = BTREE_INSERT_STATUS_OVERWROTE; |
cafe5635 KO |
1828 | if (m != end(i) && |
1829 | KEY_PTRS(m) == KEY_PTRS(k) && !KEY_SIZE(m)) | |
1830 | goto copy; | |
1831 | ||
85b1492e | 1832 | status = BTREE_INSERT_STATUS_FRONT_MERGE; |
cafe5635 KO |
1833 | if (m != end(i) && |
1834 | bch_bkey_try_merge(b, k, m)) | |
1835 | goto copy; | |
1836 | } else | |
1837 | m = bch_bset_search(b, &b->sets[b->nsets], k); | |
1838 | ||
1839 | insert: shift_keys(b, m, k); | |
1840 | copy: bkey_copy(m, k); | |
1841 | merged: | |
85b1492e | 1842 | bch_check_keys(b, "%u for %s", status, op_type(op)); |
cafe5635 KO |
1843 | |
1844 | if (b->level && !KEY_OFFSET(k)) | |
57943511 | 1845 | btree_current_write(b)->prio_blocked++; |
cafe5635 | 1846 | |
85b1492e | 1847 | trace_bcache_btree_insert_key(b, k, op->type, status); |
cafe5635 KO |
1848 | |
1849 | return true; | |
1850 | } | |
1851 | ||
f3059a54 | 1852 | static bool bch_btree_insert_keys(struct btree *b, struct btree_op *op) |
cafe5635 KO |
1853 | { |
1854 | bool ret = false; | |
1855 | struct bkey *k; | |
1856 | unsigned oldsize = bch_count_data(b); | |
1857 | ||
1858 | while ((k = bch_keylist_pop(&op->keys))) { | |
1859 | bkey_put(b->c, k, b->level); | |
1860 | ret |= btree_insert_key(b, op, k); | |
1861 | } | |
1862 | ||
1863 | BUG_ON(bch_count_data(b) < oldsize); | |
1864 | return ret; | |
1865 | } | |
1866 | ||
1867 | bool bch_btree_insert_check_key(struct btree *b, struct btree_op *op, | |
1868 | struct bio *bio) | |
1869 | { | |
1870 | bool ret = false; | |
1871 | uint64_t btree_ptr = b->key.ptr[0]; | |
1872 | unsigned long seq = b->seq; | |
1873 | BKEY_PADDED(k) tmp; | |
1874 | ||
1875 | rw_unlock(false, b); | |
1876 | rw_lock(true, b, b->level); | |
1877 | ||
1878 | if (b->key.ptr[0] != btree_ptr || | |
1879 | b->seq != seq + 1 || | |
1880 | should_split(b)) | |
1881 | goto out; | |
1882 | ||
8e51e414 | 1883 | op->replace = KEY(op->inode, bio_end_sector(bio), bio_sectors(bio)); |
cafe5635 KO |
1884 | |
1885 | SET_KEY_PTRS(&op->replace, 1); | |
1886 | get_random_bytes(&op->replace.ptr[0], sizeof(uint64_t)); | |
1887 | ||
1888 | SET_PTR_DEV(&op->replace, 0, PTR_CHECK_DEV); | |
1889 | ||
1890 | bkey_copy(&tmp.k, &op->replace); | |
1891 | ||
1892 | BUG_ON(op->type != BTREE_INSERT); | |
1893 | BUG_ON(!btree_insert_key(b, op, &tmp.k)); | |
cafe5635 KO |
1894 | ret = true; |
1895 | out: | |
1896 | downgrade_write(&b->lock); | |
1897 | return ret; | |
1898 | } | |
1899 | ||
1900 | static int btree_split(struct btree *b, struct btree_op *op) | |
1901 | { | |
d6fd3b11 | 1902 | bool split; |
cafe5635 KO |
1903 | struct btree *n1, *n2 = NULL, *n3 = NULL; |
1904 | uint64_t start_time = local_clock(); | |
1905 | ||
1906 | if (b->level) | |
1907 | set_closure_blocking(&op->cl); | |
1908 | ||
1909 | n1 = btree_node_alloc_replacement(b, &op->cl); | |
1910 | if (IS_ERR(n1)) | |
1911 | goto err; | |
1912 | ||
1913 | split = set_blocks(n1->sets[0].data, n1->c) > (btree_blocks(b) * 4) / 5; | |
1914 | ||
cafe5635 KO |
1915 | if (split) { |
1916 | unsigned keys = 0; | |
1917 | ||
c37511b8 KO |
1918 | trace_bcache_btree_node_split(b, n1->sets[0].data->keys); |
1919 | ||
cafe5635 KO |
1920 | n2 = bch_btree_node_alloc(b->c, b->level, &op->cl); |
1921 | if (IS_ERR(n2)) | |
1922 | goto err_free1; | |
1923 | ||
d6fd3b11 | 1924 | if (!b->parent) { |
cafe5635 KO |
1925 | n3 = bch_btree_node_alloc(b->c, b->level + 1, &op->cl); |
1926 | if (IS_ERR(n3)) | |
1927 | goto err_free2; | |
1928 | } | |
1929 | ||
1930 | bch_btree_insert_keys(n1, op); | |
1931 | ||
d6fd3b11 KO |
1932 | /* |
1933 | * Has to be a linear search because we don't have an auxiliary | |
cafe5635 KO |
1934 | * search tree yet |
1935 | */ | |
1936 | ||
1937 | while (keys < (n1->sets[0].data->keys * 3) / 5) | |
1938 | keys += bkey_u64s(node(n1->sets[0].data, keys)); | |
1939 | ||
1940 | bkey_copy_key(&n1->key, node(n1->sets[0].data, keys)); | |
1941 | keys += bkey_u64s(node(n1->sets[0].data, keys)); | |
1942 | ||
1943 | n2->sets[0].data->keys = n1->sets[0].data->keys - keys; | |
1944 | n1->sets[0].data->keys = keys; | |
1945 | ||
1946 | memcpy(n2->sets[0].data->start, | |
1947 | end(n1->sets[0].data), | |
1948 | n2->sets[0].data->keys * sizeof(uint64_t)); | |
1949 | ||
1950 | bkey_copy_key(&n2->key, &b->key); | |
1951 | ||
1952 | bch_keylist_add(&op->keys, &n2->key); | |
57943511 | 1953 | bch_btree_node_write(n2, &op->cl); |
cafe5635 | 1954 | rw_unlock(true, n2); |
c37511b8 KO |
1955 | } else { |
1956 | trace_bcache_btree_node_compact(b, n1->sets[0].data->keys); | |
1957 | ||
cafe5635 | 1958 | bch_btree_insert_keys(n1, op); |
c37511b8 | 1959 | } |
cafe5635 KO |
1960 | |
1961 | bch_keylist_add(&op->keys, &n1->key); | |
57943511 | 1962 | bch_btree_node_write(n1, &op->cl); |
cafe5635 KO |
1963 | |
1964 | if (n3) { | |
d6fd3b11 KO |
1965 | /* Depth increases, make a new root */ |
1966 | ||
cafe5635 KO |
1967 | bkey_copy_key(&n3->key, &MAX_KEY); |
1968 | bch_btree_insert_keys(n3, op); | |
57943511 | 1969 | bch_btree_node_write(n3, &op->cl); |
cafe5635 KO |
1970 | |
1971 | closure_sync(&op->cl); | |
1972 | bch_btree_set_root(n3); | |
1973 | rw_unlock(true, n3); | |
d6fd3b11 KO |
1974 | } else if (!b->parent) { |
1975 | /* Root filled up but didn't need to be split */ | |
1976 | ||
cafe5635 KO |
1977 | op->keys.top = op->keys.bottom; |
1978 | closure_sync(&op->cl); | |
1979 | bch_btree_set_root(n1); | |
1980 | } else { | |
1981 | unsigned i; | |
1982 | ||
1983 | bkey_copy(op->keys.top, &b->key); | |
1984 | bkey_copy_key(op->keys.top, &ZERO_KEY); | |
1985 | ||
1986 | for (i = 0; i < KEY_PTRS(&b->key); i++) { | |
1987 | uint8_t g = PTR_BUCKET(b->c, &b->key, i)->gen + 1; | |
1988 | ||
1989 | SET_PTR_GEN(op->keys.top, i, g); | |
1990 | } | |
1991 | ||
1992 | bch_keylist_push(&op->keys); | |
1993 | closure_sync(&op->cl); | |
1994 | atomic_inc(&b->c->prio_blocked); | |
1995 | } | |
1996 | ||
1997 | rw_unlock(true, n1); | |
1998 | btree_node_free(b, op); | |
1999 | ||
169ef1cf | 2000 | bch_time_stats_update(&b->c->btree_split_time, start_time); |
cafe5635 KO |
2001 | |
2002 | return 0; | |
2003 | err_free2: | |
2004 | __bkey_put(n2->c, &n2->key); | |
2005 | btree_node_free(n2, op); | |
2006 | rw_unlock(true, n2); | |
2007 | err_free1: | |
2008 | __bkey_put(n1->c, &n1->key); | |
2009 | btree_node_free(n1, op); | |
2010 | rw_unlock(true, n1); | |
2011 | err: | |
2012 | if (n3 == ERR_PTR(-EAGAIN) || | |
2013 | n2 == ERR_PTR(-EAGAIN) || | |
2014 | n1 == ERR_PTR(-EAGAIN)) | |
2015 | return -EAGAIN; | |
2016 | ||
2017 | pr_warn("couldn't split"); | |
2018 | return -ENOMEM; | |
2019 | } | |
2020 | ||
2021 | static int bch_btree_insert_recurse(struct btree *b, struct btree_op *op, | |
2022 | struct keylist *stack_keys) | |
2023 | { | |
2024 | if (b->level) { | |
2025 | int ret; | |
2026 | struct bkey *insert = op->keys.bottom; | |
2027 | struct bkey *k = bch_next_recurse_key(b, &START_KEY(insert)); | |
2028 | ||
2029 | if (!k) { | |
2030 | btree_bug(b, "no key to recurse on at level %i/%i", | |
2031 | b->level, b->c->root->level); | |
2032 | ||
2033 | op->keys.top = op->keys.bottom; | |
2034 | return -EIO; | |
2035 | } | |
2036 | ||
2037 | if (bkey_cmp(insert, k) > 0) { | |
2038 | unsigned i; | |
2039 | ||
2040 | if (op->type == BTREE_REPLACE) { | |
2041 | __bkey_put(b->c, insert); | |
2042 | op->keys.top = op->keys.bottom; | |
2043 | op->insert_collision = true; | |
2044 | return 0; | |
2045 | } | |
2046 | ||
2047 | for (i = 0; i < KEY_PTRS(insert); i++) | |
2048 | atomic_inc(&PTR_BUCKET(b->c, insert, i)->pin); | |
2049 | ||
2050 | bkey_copy(stack_keys->top, insert); | |
2051 | ||
2052 | bch_cut_back(k, insert); | |
2053 | bch_cut_front(k, stack_keys->top); | |
2054 | ||
2055 | bch_keylist_push(stack_keys); | |
2056 | } | |
2057 | ||
2058 | ret = btree(insert_recurse, k, b, op, stack_keys); | |
2059 | if (ret) | |
2060 | return ret; | |
2061 | } | |
2062 | ||
2063 | if (!bch_keylist_empty(&op->keys)) { | |
2064 | if (should_split(b)) { | |
2065 | if (op->lock <= b->c->root->level) { | |
2066 | BUG_ON(b->level); | |
2067 | op->lock = b->c->root->level + 1; | |
2068 | return -EINTR; | |
2069 | } | |
2070 | return btree_split(b, op); | |
2071 | } | |
2072 | ||
2073 | BUG_ON(write_block(b) != b->sets[b->nsets].data); | |
2074 | ||
57943511 KO |
2075 | if (bch_btree_insert_keys(b, op)) { |
2076 | if (!b->level) | |
2077 | bch_btree_leaf_dirty(b, op); | |
2078 | else | |
2079 | bch_btree_node_write(b, &op->cl); | |
2080 | } | |
cafe5635 KO |
2081 | } |
2082 | ||
2083 | return 0; | |
2084 | } | |
2085 | ||
2086 | int bch_btree_insert(struct btree_op *op, struct cache_set *c) | |
2087 | { | |
2088 | int ret = 0; | |
2089 | struct keylist stack_keys; | |
2090 | ||
2091 | /* | |
2092 | * Don't want to block with the btree locked unless we have to, | |
2093 | * otherwise we get deadlocks with try_harder and between split/gc | |
2094 | */ | |
2095 | clear_closure_blocking(&op->cl); | |
2096 | ||
2097 | BUG_ON(bch_keylist_empty(&op->keys)); | |
2098 | bch_keylist_copy(&stack_keys, &op->keys); | |
2099 | bch_keylist_init(&op->keys); | |
2100 | ||
2101 | while (!bch_keylist_empty(&stack_keys) || | |
2102 | !bch_keylist_empty(&op->keys)) { | |
2103 | if (bch_keylist_empty(&op->keys)) { | |
2104 | bch_keylist_add(&op->keys, | |
2105 | bch_keylist_pop(&stack_keys)); | |
2106 | op->lock = 0; | |
2107 | } | |
2108 | ||
2109 | ret = btree_root(insert_recurse, c, op, &stack_keys); | |
2110 | ||
2111 | if (ret == -EAGAIN) { | |
2112 | ret = 0; | |
2113 | closure_sync(&op->cl); | |
2114 | } else if (ret) { | |
2115 | struct bkey *k; | |
2116 | ||
2117 | pr_err("error %i trying to insert key for %s", | |
2118 | ret, op_type(op)); | |
2119 | ||
2120 | while ((k = bch_keylist_pop(&stack_keys) ?: | |
2121 | bch_keylist_pop(&op->keys))) | |
2122 | bkey_put(c, k, 0); | |
2123 | } | |
2124 | } | |
2125 | ||
2126 | bch_keylist_free(&stack_keys); | |
2127 | ||
2128 | if (op->journal) | |
2129 | atomic_dec_bug(op->journal); | |
2130 | op->journal = NULL; | |
2131 | return ret; | |
2132 | } | |
2133 | ||
2134 | void bch_btree_set_root(struct btree *b) | |
2135 | { | |
2136 | unsigned i; | |
e49c7c37 KO |
2137 | struct closure cl; |
2138 | ||
2139 | closure_init_stack(&cl); | |
cafe5635 | 2140 | |
c37511b8 KO |
2141 | trace_bcache_btree_set_root(b); |
2142 | ||
cafe5635 KO |
2143 | BUG_ON(!b->written); |
2144 | ||
2145 | for (i = 0; i < KEY_PTRS(&b->key); i++) | |
2146 | BUG_ON(PTR_BUCKET(b->c, &b->key, i)->prio != BTREE_PRIO); | |
2147 | ||
2148 | mutex_lock(&b->c->bucket_lock); | |
2149 | list_del_init(&b->list); | |
2150 | mutex_unlock(&b->c->bucket_lock); | |
2151 | ||
2152 | b->c->root = b; | |
2153 | __bkey_put(b->c, &b->key); | |
2154 | ||
e49c7c37 KO |
2155 | bch_journal_meta(b->c, &cl); |
2156 | closure_sync(&cl); | |
cafe5635 KO |
2157 | } |
2158 | ||
2159 | /* Cache lookup */ | |
2160 | ||
2161 | static int submit_partial_cache_miss(struct btree *b, struct btree_op *op, | |
2162 | struct bkey *k) | |
2163 | { | |
2164 | struct search *s = container_of(op, struct search, op); | |
2165 | struct bio *bio = &s->bio.bio; | |
2166 | int ret = 0; | |
2167 | ||
2168 | while (!ret && | |
2169 | !op->lookup_done) { | |
2170 | unsigned sectors = INT_MAX; | |
2171 | ||
2172 | if (KEY_INODE(k) == op->inode) { | |
2173 | if (KEY_START(k) <= bio->bi_sector) | |
2174 | break; | |
2175 | ||
2176 | sectors = min_t(uint64_t, sectors, | |
2177 | KEY_START(k) - bio->bi_sector); | |
2178 | } | |
2179 | ||
2180 | ret = s->d->cache_miss(b, s, bio, sectors); | |
2181 | } | |
2182 | ||
2183 | return ret; | |
2184 | } | |
2185 | ||
2186 | /* | |
2187 | * Read from a single key, handling the initial cache miss if the key starts in | |
2188 | * the middle of the bio | |
2189 | */ | |
2190 | static int submit_partial_cache_hit(struct btree *b, struct btree_op *op, | |
2191 | struct bkey *k) | |
2192 | { | |
2193 | struct search *s = container_of(op, struct search, op); | |
2194 | struct bio *bio = &s->bio.bio; | |
2195 | unsigned ptr; | |
2196 | struct bio *n; | |
2197 | ||
2198 | int ret = submit_partial_cache_miss(b, op, k); | |
2199 | if (ret || op->lookup_done) | |
2200 | return ret; | |
2201 | ||
2202 | /* XXX: figure out best pointer - for multiple cache devices */ | |
2203 | ptr = 0; | |
2204 | ||
2205 | PTR_BUCKET(b->c, k, ptr)->prio = INITIAL_PRIO; | |
2206 | ||
2207 | while (!op->lookup_done && | |
2208 | KEY_INODE(k) == op->inode && | |
2209 | bio->bi_sector < KEY_OFFSET(k)) { | |
2210 | struct bkey *bio_key; | |
2211 | sector_t sector = PTR_OFFSET(k, ptr) + | |
2212 | (bio->bi_sector - KEY_START(k)); | |
2213 | unsigned sectors = min_t(uint64_t, INT_MAX, | |
2214 | KEY_OFFSET(k) - bio->bi_sector); | |
2215 | ||
2216 | n = bch_bio_split(bio, sectors, GFP_NOIO, s->d->bio_split); | |
cafe5635 KO |
2217 | if (n == bio) |
2218 | op->lookup_done = true; | |
2219 | ||
2220 | bio_key = &container_of(n, struct bbio, bio)->key; | |
2221 | ||
2222 | /* | |
2223 | * The bucket we're reading from might be reused while our bio | |
2224 | * is in flight, and we could then end up reading the wrong | |
2225 | * data. | |
2226 | * | |
2227 | * We guard against this by checking (in cache_read_endio()) if | |
2228 | * the pointer is stale again; if so, we treat it as an error | |
2229 | * and reread from the backing device (but we don't pass that | |
2230 | * error up anywhere). | |
2231 | */ | |
2232 | ||
2233 | bch_bkey_copy_single_ptr(bio_key, k, ptr); | |
2234 | SET_PTR_OFFSET(bio_key, 0, sector); | |
2235 | ||
2236 | n->bi_end_io = bch_cache_read_endio; | |
2237 | n->bi_private = &s->cl; | |
2238 | ||
cafe5635 KO |
2239 | __bch_submit_bbio(n, b->c); |
2240 | } | |
2241 | ||
2242 | return 0; | |
2243 | } | |
2244 | ||
2245 | int bch_btree_search_recurse(struct btree *b, struct btree_op *op) | |
2246 | { | |
2247 | struct search *s = container_of(op, struct search, op); | |
2248 | struct bio *bio = &s->bio.bio; | |
2249 | ||
2250 | int ret = 0; | |
2251 | struct bkey *k; | |
2252 | struct btree_iter iter; | |
2253 | bch_btree_iter_init(b, &iter, &KEY(op->inode, bio->bi_sector, 0)); | |
2254 | ||
cafe5635 KO |
2255 | do { |
2256 | k = bch_btree_iter_next_filter(&iter, b, bch_ptr_bad); | |
2257 | if (!k) { | |
2258 | /* | |
2259 | * b->key would be exactly what we want, except that | |
2260 | * pointers to btree nodes have nonzero size - we | |
2261 | * wouldn't go far enough | |
2262 | */ | |
2263 | ||
2264 | ret = submit_partial_cache_miss(b, op, | |
2265 | &KEY(KEY_INODE(&b->key), | |
2266 | KEY_OFFSET(&b->key), 0)); | |
2267 | break; | |
2268 | } | |
2269 | ||
2270 | ret = b->level | |
2271 | ? btree(search_recurse, k, b, op) | |
2272 | : submit_partial_cache_hit(b, op, k); | |
2273 | } while (!ret && | |
2274 | !op->lookup_done); | |
2275 | ||
2276 | return ret; | |
2277 | } | |
2278 | ||
2279 | /* Keybuf code */ | |
2280 | ||
2281 | static inline int keybuf_cmp(struct keybuf_key *l, struct keybuf_key *r) | |
2282 | { | |
2283 | /* Overlapping keys compare equal */ | |
2284 | if (bkey_cmp(&l->key, &START_KEY(&r->key)) <= 0) | |
2285 | return -1; | |
2286 | if (bkey_cmp(&START_KEY(&l->key), &r->key) >= 0) | |
2287 | return 1; | |
2288 | return 0; | |
2289 | } | |
2290 | ||
2291 | static inline int keybuf_nonoverlapping_cmp(struct keybuf_key *l, | |
2292 | struct keybuf_key *r) | |
2293 | { | |
2294 | return clamp_t(int64_t, bkey_cmp(&l->key, &r->key), -1, 1); | |
2295 | } | |
2296 | ||
2297 | static int bch_btree_refill_keybuf(struct btree *b, struct btree_op *op, | |
72c27061 KO |
2298 | struct keybuf *buf, struct bkey *end, |
2299 | keybuf_pred_fn *pred) | |
cafe5635 KO |
2300 | { |
2301 | struct btree_iter iter; | |
2302 | bch_btree_iter_init(b, &iter, &buf->last_scanned); | |
2303 | ||
2304 | while (!array_freelist_empty(&buf->freelist)) { | |
2305 | struct bkey *k = bch_btree_iter_next_filter(&iter, b, | |
2306 | bch_ptr_bad); | |
2307 | ||
2308 | if (!b->level) { | |
2309 | if (!k) { | |
2310 | buf->last_scanned = b->key; | |
2311 | break; | |
2312 | } | |
2313 | ||
2314 | buf->last_scanned = *k; | |
2315 | if (bkey_cmp(&buf->last_scanned, end) >= 0) | |
2316 | break; | |
2317 | ||
72c27061 | 2318 | if (pred(buf, k)) { |
cafe5635 KO |
2319 | struct keybuf_key *w; |
2320 | ||
cafe5635 KO |
2321 | spin_lock(&buf->lock); |
2322 | ||
2323 | w = array_alloc(&buf->freelist); | |
2324 | ||
2325 | w->private = NULL; | |
2326 | bkey_copy(&w->key, k); | |
2327 | ||
2328 | if (RB_INSERT(&buf->keys, w, node, keybuf_cmp)) | |
2329 | array_free(&buf->freelist, w); | |
2330 | ||
2331 | spin_unlock(&buf->lock); | |
2332 | } | |
2333 | } else { | |
2334 | if (!k) | |
2335 | break; | |
2336 | ||
72c27061 | 2337 | btree(refill_keybuf, k, b, op, buf, end, pred); |
cafe5635 KO |
2338 | /* |
2339 | * Might get an error here, but can't really do anything | |
2340 | * and it'll get logged elsewhere. Just read what we | |
2341 | * can. | |
2342 | */ | |
2343 | ||
2344 | if (bkey_cmp(&buf->last_scanned, end) >= 0) | |
2345 | break; | |
2346 | ||
2347 | cond_resched(); | |
2348 | } | |
2349 | } | |
2350 | ||
2351 | return 0; | |
2352 | } | |
2353 | ||
2354 | void bch_refill_keybuf(struct cache_set *c, struct keybuf *buf, | |
72c27061 | 2355 | struct bkey *end, keybuf_pred_fn *pred) |
cafe5635 KO |
2356 | { |
2357 | struct bkey start = buf->last_scanned; | |
2358 | struct btree_op op; | |
2359 | bch_btree_op_init_stack(&op); | |
2360 | ||
2361 | cond_resched(); | |
2362 | ||
72c27061 | 2363 | btree_root(refill_keybuf, c, &op, buf, end, pred); |
cafe5635 KO |
2364 | closure_sync(&op.cl); |
2365 | ||
2366 | pr_debug("found %s keys from %llu:%llu to %llu:%llu", | |
2367 | RB_EMPTY_ROOT(&buf->keys) ? "no" : | |
2368 | array_freelist_empty(&buf->freelist) ? "some" : "a few", | |
2369 | KEY_INODE(&start), KEY_OFFSET(&start), | |
2370 | KEY_INODE(&buf->last_scanned), KEY_OFFSET(&buf->last_scanned)); | |
2371 | ||
2372 | spin_lock(&buf->lock); | |
2373 | ||
2374 | if (!RB_EMPTY_ROOT(&buf->keys)) { | |
2375 | struct keybuf_key *w; | |
2376 | w = RB_FIRST(&buf->keys, struct keybuf_key, node); | |
2377 | buf->start = START_KEY(&w->key); | |
2378 | ||
2379 | w = RB_LAST(&buf->keys, struct keybuf_key, node); | |
2380 | buf->end = w->key; | |
2381 | } else { | |
2382 | buf->start = MAX_KEY; | |
2383 | buf->end = MAX_KEY; | |
2384 | } | |
2385 | ||
2386 | spin_unlock(&buf->lock); | |
2387 | } | |
2388 | ||
2389 | static void __bch_keybuf_del(struct keybuf *buf, struct keybuf_key *w) | |
2390 | { | |
2391 | rb_erase(&w->node, &buf->keys); | |
2392 | array_free(&buf->freelist, w); | |
2393 | } | |
2394 | ||
2395 | void bch_keybuf_del(struct keybuf *buf, struct keybuf_key *w) | |
2396 | { | |
2397 | spin_lock(&buf->lock); | |
2398 | __bch_keybuf_del(buf, w); | |
2399 | spin_unlock(&buf->lock); | |
2400 | } | |
2401 | ||
2402 | bool bch_keybuf_check_overlapping(struct keybuf *buf, struct bkey *start, | |
2403 | struct bkey *end) | |
2404 | { | |
2405 | bool ret = false; | |
2406 | struct keybuf_key *p, *w, s; | |
2407 | s.key = *start; | |
2408 | ||
2409 | if (bkey_cmp(end, &buf->start) <= 0 || | |
2410 | bkey_cmp(start, &buf->end) >= 0) | |
2411 | return false; | |
2412 | ||
2413 | spin_lock(&buf->lock); | |
2414 | w = RB_GREATER(&buf->keys, s, node, keybuf_nonoverlapping_cmp); | |
2415 | ||
2416 | while (w && bkey_cmp(&START_KEY(&w->key), end) < 0) { | |
2417 | p = w; | |
2418 | w = RB_NEXT(w, node); | |
2419 | ||
2420 | if (p->private) | |
2421 | ret = true; | |
2422 | else | |
2423 | __bch_keybuf_del(buf, p); | |
2424 | } | |
2425 | ||
2426 | spin_unlock(&buf->lock); | |
2427 | return ret; | |
2428 | } | |
2429 | ||
2430 | struct keybuf_key *bch_keybuf_next(struct keybuf *buf) | |
2431 | { | |
2432 | struct keybuf_key *w; | |
2433 | spin_lock(&buf->lock); | |
2434 | ||
2435 | w = RB_FIRST(&buf->keys, struct keybuf_key, node); | |
2436 | ||
2437 | while (w && w->private) | |
2438 | w = RB_NEXT(w, node); | |
2439 | ||
2440 | if (w) | |
2441 | w->private = ERR_PTR(-EINTR); | |
2442 | ||
2443 | spin_unlock(&buf->lock); | |
2444 | return w; | |
2445 | } | |
2446 | ||
2447 | struct keybuf_key *bch_keybuf_next_rescan(struct cache_set *c, | |
2448 | struct keybuf *buf, | |
72c27061 KO |
2449 | struct bkey *end, |
2450 | keybuf_pred_fn *pred) | |
cafe5635 KO |
2451 | { |
2452 | struct keybuf_key *ret; | |
2453 | ||
2454 | while (1) { | |
2455 | ret = bch_keybuf_next(buf); | |
2456 | if (ret) | |
2457 | break; | |
2458 | ||
2459 | if (bkey_cmp(&buf->last_scanned, end) >= 0) { | |
2460 | pr_debug("scan finished"); | |
2461 | break; | |
2462 | } | |
2463 | ||
72c27061 | 2464 | bch_refill_keybuf(c, buf, end, pred); |
cafe5635 KO |
2465 | } |
2466 | ||
2467 | return ret; | |
2468 | } | |
2469 | ||
72c27061 | 2470 | void bch_keybuf_init(struct keybuf *buf) |
cafe5635 | 2471 | { |
cafe5635 KO |
2472 | buf->last_scanned = MAX_KEY; |
2473 | buf->keys = RB_ROOT; | |
2474 | ||
2475 | spin_lock_init(&buf->lock); | |
2476 | array_allocator_init(&buf->freelist); | |
2477 | } | |
2478 | ||
2479 | void bch_btree_exit(void) | |
2480 | { | |
2481 | if (btree_io_wq) | |
2482 | destroy_workqueue(btree_io_wq); | |
2483 | if (bch_gc_wq) | |
2484 | destroy_workqueue(bch_gc_wq); | |
2485 | } | |
2486 | ||
2487 | int __init bch_btree_init(void) | |
2488 | { | |
2489 | if (!(bch_gc_wq = create_singlethread_workqueue("bch_btree_gc")) || | |
2490 | !(btree_io_wq = create_singlethread_workqueue("bch_btree_io"))) | |
2491 | return -ENOMEM; | |
2492 | ||
2493 | return 0; | |
2494 | } |