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 | ||
e7c590eb KO |
121 | void __bkey_put(struct cache_set *c, struct bkey *k) |
122 | { | |
123 | unsigned i; | |
124 | ||
125 | for (i = 0; i < KEY_PTRS(k); i++) | |
126 | if (ptr_available(c, k, i)) | |
127 | atomic_dec_bug(&PTR_BUCKET(c, k, i)->pin); | |
128 | } | |
129 | ||
cafe5635 KO |
130 | static void bkey_put(struct cache_set *c, struct bkey *k, int level) |
131 | { | |
132 | if ((level && KEY_OFFSET(k)) || !level) | |
133 | __bkey_put(c, k); | |
134 | } | |
135 | ||
136 | /* Btree IO */ | |
137 | ||
138 | static uint64_t btree_csum_set(struct btree *b, struct bset *i) | |
139 | { | |
140 | uint64_t crc = b->key.ptr[0]; | |
141 | void *data = (void *) i + 8, *end = end(i); | |
142 | ||
169ef1cf | 143 | crc = bch_crc64_update(crc, data, end - data); |
c19ed23a | 144 | return crc ^ 0xffffffffffffffffULL; |
cafe5635 KO |
145 | } |
146 | ||
f3059a54 | 147 | static void bch_btree_node_read_done(struct btree *b) |
cafe5635 | 148 | { |
cafe5635 | 149 | const char *err = "bad btree header"; |
57943511 KO |
150 | struct bset *i = b->sets[0].data; |
151 | struct btree_iter *iter; | |
cafe5635 | 152 | |
57943511 KO |
153 | iter = mempool_alloc(b->c->fill_iter, GFP_NOWAIT); |
154 | iter->size = b->c->sb.bucket_size / b->c->sb.block_size; | |
cafe5635 KO |
155 | iter->used = 0; |
156 | ||
57943511 | 157 | if (!i->seq) |
cafe5635 KO |
158 | goto err; |
159 | ||
160 | for (; | |
161 | b->written < btree_blocks(b) && i->seq == b->sets[0].data->seq; | |
162 | i = write_block(b)) { | |
163 | err = "unsupported bset version"; | |
164 | if (i->version > BCACHE_BSET_VERSION) | |
165 | goto err; | |
166 | ||
167 | err = "bad btree header"; | |
168 | if (b->written + set_blocks(i, b->c) > btree_blocks(b)) | |
169 | goto err; | |
170 | ||
171 | err = "bad magic"; | |
172 | if (i->magic != bset_magic(b->c)) | |
173 | goto err; | |
174 | ||
175 | err = "bad checksum"; | |
176 | switch (i->version) { | |
177 | case 0: | |
178 | if (i->csum != csum_set(i)) | |
179 | goto err; | |
180 | break; | |
181 | case BCACHE_BSET_VERSION: | |
182 | if (i->csum != btree_csum_set(b, i)) | |
183 | goto err; | |
184 | break; | |
185 | } | |
186 | ||
187 | err = "empty set"; | |
188 | if (i != b->sets[0].data && !i->keys) | |
189 | goto err; | |
190 | ||
191 | bch_btree_iter_push(iter, i->start, end(i)); | |
192 | ||
193 | b->written += set_blocks(i, b->c); | |
194 | } | |
195 | ||
196 | err = "corrupted btree"; | |
197 | for (i = write_block(b); | |
198 | index(i, b) < btree_blocks(b); | |
199 | i = ((void *) i) + block_bytes(b->c)) | |
200 | if (i->seq == b->sets[0].data->seq) | |
201 | goto err; | |
202 | ||
203 | bch_btree_sort_and_fix_extents(b, iter); | |
204 | ||
205 | i = b->sets[0].data; | |
206 | err = "short btree key"; | |
207 | if (b->sets[0].size && | |
208 | bkey_cmp(&b->key, &b->sets[0].end) < 0) | |
209 | goto err; | |
210 | ||
211 | if (b->written < btree_blocks(b)) | |
212 | bch_bset_init_next(b); | |
213 | out: | |
57943511 KO |
214 | mempool_free(iter, b->c->fill_iter); |
215 | return; | |
cafe5635 KO |
216 | err: |
217 | set_btree_node_io_error(b); | |
07e86ccb | 218 | bch_cache_set_error(b->c, "%s at bucket %zu, block %zu, %u keys", |
cafe5635 KO |
219 | err, PTR_BUCKET_NR(b->c, &b->key, 0), |
220 | index(i, b), i->keys); | |
221 | goto out; | |
222 | } | |
223 | ||
57943511 | 224 | static void btree_node_read_endio(struct bio *bio, int error) |
cafe5635 | 225 | { |
57943511 KO |
226 | struct closure *cl = bio->bi_private; |
227 | closure_put(cl); | |
228 | } | |
cafe5635 | 229 | |
57943511 KO |
230 | void bch_btree_node_read(struct btree *b) |
231 | { | |
232 | uint64_t start_time = local_clock(); | |
233 | struct closure cl; | |
234 | struct bio *bio; | |
cafe5635 | 235 | |
c37511b8 KO |
236 | trace_bcache_btree_read(b); |
237 | ||
57943511 | 238 | closure_init_stack(&cl); |
cafe5635 | 239 | |
57943511 KO |
240 | bio = bch_bbio_alloc(b->c); |
241 | bio->bi_rw = REQ_META|READ_SYNC; | |
242 | bio->bi_size = KEY_SIZE(&b->key) << 9; | |
243 | bio->bi_end_io = btree_node_read_endio; | |
244 | bio->bi_private = &cl; | |
cafe5635 | 245 | |
57943511 | 246 | bch_bio_map(bio, b->sets[0].data); |
cafe5635 | 247 | |
57943511 KO |
248 | bch_submit_bbio(bio, b->c, &b->key, 0); |
249 | closure_sync(&cl); | |
cafe5635 | 250 | |
57943511 KO |
251 | if (!test_bit(BIO_UPTODATE, &bio->bi_flags)) |
252 | set_btree_node_io_error(b); | |
253 | ||
254 | bch_bbio_free(bio, b->c); | |
255 | ||
256 | if (btree_node_io_error(b)) | |
257 | goto err; | |
258 | ||
259 | bch_btree_node_read_done(b); | |
260 | ||
261 | spin_lock(&b->c->btree_read_time_lock); | |
262 | bch_time_stats_update(&b->c->btree_read_time, start_time); | |
263 | spin_unlock(&b->c->btree_read_time_lock); | |
264 | ||
265 | return; | |
266 | err: | |
61cbd250 | 267 | bch_cache_set_error(b->c, "io error reading bucket %zu", |
57943511 | 268 | PTR_BUCKET_NR(b->c, &b->key, 0)); |
cafe5635 KO |
269 | } |
270 | ||
271 | static void btree_complete_write(struct btree *b, struct btree_write *w) | |
272 | { | |
273 | if (w->prio_blocked && | |
274 | !atomic_sub_return(w->prio_blocked, &b->c->prio_blocked)) | |
119ba0f8 | 275 | wake_up_allocators(b->c); |
cafe5635 KO |
276 | |
277 | if (w->journal) { | |
278 | atomic_dec_bug(w->journal); | |
279 | __closure_wake_up(&b->c->journal.wait); | |
280 | } | |
281 | ||
cafe5635 KO |
282 | w->prio_blocked = 0; |
283 | w->journal = NULL; | |
cafe5635 KO |
284 | } |
285 | ||
57943511 | 286 | static void __btree_node_write_done(struct closure *cl) |
cafe5635 KO |
287 | { |
288 | struct btree *b = container_of(cl, struct btree, io.cl); | |
289 | struct btree_write *w = btree_prev_write(b); | |
290 | ||
291 | bch_bbio_free(b->bio, b->c); | |
292 | b->bio = NULL; | |
293 | btree_complete_write(b, w); | |
294 | ||
295 | if (btree_node_dirty(b)) | |
296 | queue_delayed_work(btree_io_wq, &b->work, | |
297 | msecs_to_jiffies(30000)); | |
298 | ||
299 | closure_return(cl); | |
300 | } | |
301 | ||
57943511 | 302 | static void btree_node_write_done(struct closure *cl) |
cafe5635 KO |
303 | { |
304 | struct btree *b = container_of(cl, struct btree, io.cl); | |
305 | struct bio_vec *bv; | |
306 | int n; | |
307 | ||
308 | __bio_for_each_segment(bv, b->bio, n, 0) | |
309 | __free_page(bv->bv_page); | |
310 | ||
57943511 | 311 | __btree_node_write_done(cl); |
cafe5635 KO |
312 | } |
313 | ||
57943511 KO |
314 | static void btree_node_write_endio(struct bio *bio, int error) |
315 | { | |
316 | struct closure *cl = bio->bi_private; | |
317 | struct btree *b = container_of(cl, struct btree, io.cl); | |
318 | ||
319 | if (error) | |
320 | set_btree_node_io_error(b); | |
321 | ||
322 | bch_bbio_count_io_errors(b->c, bio, error, "writing btree"); | |
323 | closure_put(cl); | |
324 | } | |
325 | ||
326 | static void do_btree_node_write(struct btree *b) | |
cafe5635 KO |
327 | { |
328 | struct closure *cl = &b->io.cl; | |
329 | struct bset *i = b->sets[b->nsets].data; | |
330 | BKEY_PADDED(key) k; | |
331 | ||
332 | i->version = BCACHE_BSET_VERSION; | |
333 | i->csum = btree_csum_set(b, i); | |
334 | ||
57943511 KO |
335 | BUG_ON(b->bio); |
336 | b->bio = bch_bbio_alloc(b->c); | |
337 | ||
338 | b->bio->bi_end_io = btree_node_write_endio; | |
339 | b->bio->bi_private = &b->io.cl; | |
e49c7c37 KO |
340 | b->bio->bi_rw = REQ_META|WRITE_SYNC|REQ_FUA; |
341 | b->bio->bi_size = set_blocks(i, b->c) * block_bytes(b->c); | |
169ef1cf | 342 | bch_bio_map(b->bio, i); |
cafe5635 | 343 | |
e49c7c37 KO |
344 | /* |
345 | * If we're appending to a leaf node, we don't technically need FUA - | |
346 | * this write just needs to be persisted before the next journal write, | |
347 | * which will be marked FLUSH|FUA. | |
348 | * | |
349 | * Similarly if we're writing a new btree root - the pointer is going to | |
350 | * be in the next journal entry. | |
351 | * | |
352 | * But if we're writing a new btree node (that isn't a root) or | |
353 | * appending to a non leaf btree node, we need either FUA or a flush | |
354 | * when we write the parent with the new pointer. FUA is cheaper than a | |
355 | * flush, and writes appending to leaf nodes aren't blocking anything so | |
356 | * just make all btree node writes FUA to keep things sane. | |
357 | */ | |
358 | ||
cafe5635 KO |
359 | bkey_copy(&k.key, &b->key); |
360 | SET_PTR_OFFSET(&k.key, 0, PTR_OFFSET(&k.key, 0) + bset_offset(b, i)); | |
361 | ||
8e51e414 | 362 | if (!bio_alloc_pages(b->bio, GFP_NOIO)) { |
cafe5635 KO |
363 | int j; |
364 | struct bio_vec *bv; | |
365 | void *base = (void *) ((unsigned long) i & ~(PAGE_SIZE - 1)); | |
366 | ||
367 | bio_for_each_segment(bv, b->bio, j) | |
368 | memcpy(page_address(bv->bv_page), | |
369 | base + j * PAGE_SIZE, PAGE_SIZE); | |
370 | ||
cafe5635 KO |
371 | bch_submit_bbio(b->bio, b->c, &k.key, 0); |
372 | ||
57943511 | 373 | continue_at(cl, btree_node_write_done, NULL); |
cafe5635 KO |
374 | } else { |
375 | b->bio->bi_vcnt = 0; | |
169ef1cf | 376 | bch_bio_map(b->bio, i); |
cafe5635 | 377 | |
cafe5635 KO |
378 | bch_submit_bbio(b->bio, b->c, &k.key, 0); |
379 | ||
380 | closure_sync(cl); | |
57943511 | 381 | __btree_node_write_done(cl); |
cafe5635 KO |
382 | } |
383 | } | |
384 | ||
57943511 | 385 | void bch_btree_node_write(struct btree *b, struct closure *parent) |
cafe5635 KO |
386 | { |
387 | struct bset *i = b->sets[b->nsets].data; | |
388 | ||
c37511b8 KO |
389 | trace_bcache_btree_write(b); |
390 | ||
cafe5635 | 391 | BUG_ON(current->bio_list); |
57943511 KO |
392 | BUG_ON(b->written >= btree_blocks(b)); |
393 | BUG_ON(b->written && !i->keys); | |
394 | BUG_ON(b->sets->data->seq != i->seq); | |
c37511b8 | 395 | bch_check_key_order(b, i); |
cafe5635 | 396 | |
cafe5635 KO |
397 | cancel_delayed_work(&b->work); |
398 | ||
57943511 KO |
399 | /* If caller isn't waiting for write, parent refcount is cache set */ |
400 | closure_lock(&b->io, parent ?: &b->c->cl); | |
401 | ||
cafe5635 KO |
402 | clear_bit(BTREE_NODE_dirty, &b->flags); |
403 | change_bit(BTREE_NODE_write_idx, &b->flags); | |
404 | ||
57943511 | 405 | do_btree_node_write(b); |
cafe5635 | 406 | |
cafe5635 KO |
407 | b->written += set_blocks(i, b->c); |
408 | atomic_long_add(set_blocks(i, b->c) * b->c->sb.block_size, | |
409 | &PTR_CACHE(b->c, &b->key, 0)->btree_sectors_written); | |
410 | ||
411 | bch_btree_sort_lazy(b); | |
412 | ||
413 | if (b->written < btree_blocks(b)) | |
414 | bch_bset_init_next(b); | |
415 | } | |
416 | ||
57943511 | 417 | static void btree_node_write_work(struct work_struct *w) |
cafe5635 KO |
418 | { |
419 | struct btree *b = container_of(to_delayed_work(w), struct btree, work); | |
420 | ||
57943511 | 421 | rw_lock(true, b, b->level); |
cafe5635 KO |
422 | |
423 | if (btree_node_dirty(b)) | |
57943511 KO |
424 | bch_btree_node_write(b, NULL); |
425 | rw_unlock(true, b); | |
cafe5635 KO |
426 | } |
427 | ||
57943511 | 428 | static void bch_btree_leaf_dirty(struct btree *b, struct btree_op *op) |
cafe5635 KO |
429 | { |
430 | struct bset *i = b->sets[b->nsets].data; | |
431 | struct btree_write *w = btree_current_write(b); | |
432 | ||
57943511 KO |
433 | BUG_ON(!b->written); |
434 | BUG_ON(!i->keys); | |
cafe5635 | 435 | |
57943511 KO |
436 | if (!btree_node_dirty(b)) |
437 | queue_delayed_work(btree_io_wq, &b->work, 30 * HZ); | |
cafe5635 | 438 | |
57943511 | 439 | set_btree_node_dirty(b); |
cafe5635 | 440 | |
57943511 | 441 | if (op && op->journal) { |
cafe5635 KO |
442 | if (w->journal && |
443 | journal_pin_cmp(b->c, w, op)) { | |
444 | atomic_dec_bug(w->journal); | |
445 | w->journal = NULL; | |
446 | } | |
447 | ||
448 | if (!w->journal) { | |
449 | w->journal = op->journal; | |
450 | atomic_inc(w->journal); | |
451 | } | |
452 | } | |
453 | ||
cafe5635 | 454 | /* Force write if set is too big */ |
57943511 KO |
455 | if (set_bytes(i) > PAGE_SIZE - 48 && |
456 | !current->bio_list) | |
457 | bch_btree_node_write(b, NULL); | |
cafe5635 KO |
458 | } |
459 | ||
460 | /* | |
461 | * Btree in memory cache - allocation/freeing | |
462 | * mca -> memory cache | |
463 | */ | |
464 | ||
465 | static void mca_reinit(struct btree *b) | |
466 | { | |
467 | unsigned i; | |
468 | ||
469 | b->flags = 0; | |
470 | b->written = 0; | |
471 | b->nsets = 0; | |
472 | ||
473 | for (i = 0; i < MAX_BSETS; i++) | |
474 | b->sets[i].size = 0; | |
475 | /* | |
476 | * Second loop starts at 1 because b->sets[0]->data is the memory we | |
477 | * allocated | |
478 | */ | |
479 | for (i = 1; i < MAX_BSETS; i++) | |
480 | b->sets[i].data = NULL; | |
481 | } | |
482 | ||
483 | #define mca_reserve(c) (((c->root && c->root->level) \ | |
484 | ? c->root->level : 1) * 8 + 16) | |
485 | #define mca_can_free(c) \ | |
486 | max_t(int, 0, c->bucket_cache_used - mca_reserve(c)) | |
487 | ||
488 | static void mca_data_free(struct btree *b) | |
489 | { | |
490 | struct bset_tree *t = b->sets; | |
491 | BUG_ON(!closure_is_unlocked(&b->io.cl)); | |
492 | ||
493 | if (bset_prev_bytes(b) < PAGE_SIZE) | |
494 | kfree(t->prev); | |
495 | else | |
496 | free_pages((unsigned long) t->prev, | |
497 | get_order(bset_prev_bytes(b))); | |
498 | ||
499 | if (bset_tree_bytes(b) < PAGE_SIZE) | |
500 | kfree(t->tree); | |
501 | else | |
502 | free_pages((unsigned long) t->tree, | |
503 | get_order(bset_tree_bytes(b))); | |
504 | ||
505 | free_pages((unsigned long) t->data, b->page_order); | |
506 | ||
507 | t->prev = NULL; | |
508 | t->tree = NULL; | |
509 | t->data = NULL; | |
510 | list_move(&b->list, &b->c->btree_cache_freed); | |
511 | b->c->bucket_cache_used--; | |
512 | } | |
513 | ||
514 | static void mca_bucket_free(struct btree *b) | |
515 | { | |
516 | BUG_ON(btree_node_dirty(b)); | |
517 | ||
518 | b->key.ptr[0] = 0; | |
519 | hlist_del_init_rcu(&b->hash); | |
520 | list_move(&b->list, &b->c->btree_cache_freeable); | |
521 | } | |
522 | ||
523 | static unsigned btree_order(struct bkey *k) | |
524 | { | |
525 | return ilog2(KEY_SIZE(k) / PAGE_SECTORS ?: 1); | |
526 | } | |
527 | ||
528 | static void mca_data_alloc(struct btree *b, struct bkey *k, gfp_t gfp) | |
529 | { | |
530 | struct bset_tree *t = b->sets; | |
531 | BUG_ON(t->data); | |
532 | ||
533 | b->page_order = max_t(unsigned, | |
534 | ilog2(b->c->btree_pages), | |
535 | btree_order(k)); | |
536 | ||
537 | t->data = (void *) __get_free_pages(gfp, b->page_order); | |
538 | if (!t->data) | |
539 | goto err; | |
540 | ||
541 | t->tree = bset_tree_bytes(b) < PAGE_SIZE | |
542 | ? kmalloc(bset_tree_bytes(b), gfp) | |
543 | : (void *) __get_free_pages(gfp, get_order(bset_tree_bytes(b))); | |
544 | if (!t->tree) | |
545 | goto err; | |
546 | ||
547 | t->prev = bset_prev_bytes(b) < PAGE_SIZE | |
548 | ? kmalloc(bset_prev_bytes(b), gfp) | |
549 | : (void *) __get_free_pages(gfp, get_order(bset_prev_bytes(b))); | |
550 | if (!t->prev) | |
551 | goto err; | |
552 | ||
553 | list_move(&b->list, &b->c->btree_cache); | |
554 | b->c->bucket_cache_used++; | |
555 | return; | |
556 | err: | |
557 | mca_data_free(b); | |
558 | } | |
559 | ||
560 | static struct btree *mca_bucket_alloc(struct cache_set *c, | |
561 | struct bkey *k, gfp_t gfp) | |
562 | { | |
563 | struct btree *b = kzalloc(sizeof(struct btree), gfp); | |
564 | if (!b) | |
565 | return NULL; | |
566 | ||
567 | init_rwsem(&b->lock); | |
568 | lockdep_set_novalidate_class(&b->lock); | |
569 | INIT_LIST_HEAD(&b->list); | |
57943511 | 570 | INIT_DELAYED_WORK(&b->work, btree_node_write_work); |
cafe5635 KO |
571 | b->c = c; |
572 | closure_init_unlocked(&b->io); | |
573 | ||
574 | mca_data_alloc(b, k, gfp); | |
575 | return b; | |
576 | } | |
577 | ||
578 | static int mca_reap(struct btree *b, struct closure *cl, unsigned min_order) | |
579 | { | |
580 | lockdep_assert_held(&b->c->bucket_lock); | |
581 | ||
582 | if (!down_write_trylock(&b->lock)) | |
583 | return -ENOMEM; | |
584 | ||
585 | if (b->page_order < min_order) { | |
586 | rw_unlock(true, b); | |
587 | return -ENOMEM; | |
588 | } | |
589 | ||
590 | BUG_ON(btree_node_dirty(b) && !b->sets[0].data); | |
591 | ||
592 | if (cl && btree_node_dirty(b)) | |
57943511 | 593 | bch_btree_node_write(b, NULL); |
cafe5635 KO |
594 | |
595 | if (cl) | |
596 | closure_wait_event_async(&b->io.wait, cl, | |
597 | atomic_read(&b->io.cl.remaining) == -1); | |
598 | ||
599 | if (btree_node_dirty(b) || | |
600 | !closure_is_unlocked(&b->io.cl) || | |
601 | work_pending(&b->work.work)) { | |
602 | rw_unlock(true, b); | |
603 | return -EAGAIN; | |
604 | } | |
605 | ||
606 | return 0; | |
607 | } | |
608 | ||
7dc19d5a DC |
609 | static unsigned long bch_mca_scan(struct shrinker *shrink, |
610 | struct shrink_control *sc) | |
cafe5635 KO |
611 | { |
612 | struct cache_set *c = container_of(shrink, struct cache_set, shrink); | |
613 | struct btree *b, *t; | |
614 | unsigned long i, nr = sc->nr_to_scan; | |
7dc19d5a | 615 | unsigned long freed = 0; |
cafe5635 KO |
616 | |
617 | if (c->shrinker_disabled) | |
7dc19d5a | 618 | return SHRINK_STOP; |
cafe5635 KO |
619 | |
620 | if (c->try_harder) | |
7dc19d5a | 621 | return SHRINK_STOP; |
cafe5635 KO |
622 | |
623 | /* Return -1 if we can't do anything right now */ | |
a698e08c | 624 | if (sc->gfp_mask & __GFP_IO) |
cafe5635 KO |
625 | mutex_lock(&c->bucket_lock); |
626 | else if (!mutex_trylock(&c->bucket_lock)) | |
627 | return -1; | |
628 | ||
36c9ea98 KO |
629 | /* |
630 | * It's _really_ critical that we don't free too many btree nodes - we | |
631 | * have to always leave ourselves a reserve. The reserve is how we | |
632 | * guarantee that allocating memory for a new btree node can always | |
633 | * succeed, so that inserting keys into the btree can always succeed and | |
634 | * IO can always make forward progress: | |
635 | */ | |
cafe5635 KO |
636 | nr /= c->btree_pages; |
637 | nr = min_t(unsigned long, nr, mca_can_free(c)); | |
638 | ||
639 | i = 0; | |
640 | list_for_each_entry_safe(b, t, &c->btree_cache_freeable, list) { | |
7dc19d5a | 641 | if (freed >= nr) |
cafe5635 KO |
642 | break; |
643 | ||
644 | if (++i > 3 && | |
645 | !mca_reap(b, NULL, 0)) { | |
646 | mca_data_free(b); | |
647 | rw_unlock(true, b); | |
7dc19d5a | 648 | freed++; |
cafe5635 KO |
649 | } |
650 | } | |
651 | ||
652 | /* | |
653 | * Can happen right when we first start up, before we've read in any | |
654 | * btree nodes | |
655 | */ | |
656 | if (list_empty(&c->btree_cache)) | |
657 | goto out; | |
658 | ||
7dc19d5a | 659 | for (i = 0; (nr--) && i < c->bucket_cache_used; i++) { |
cafe5635 KO |
660 | b = list_first_entry(&c->btree_cache, struct btree, list); |
661 | list_rotate_left(&c->btree_cache); | |
662 | ||
663 | if (!b->accessed && | |
664 | !mca_reap(b, NULL, 0)) { | |
665 | mca_bucket_free(b); | |
666 | mca_data_free(b); | |
667 | rw_unlock(true, b); | |
7dc19d5a | 668 | freed++; |
cafe5635 KO |
669 | } else |
670 | b->accessed = 0; | |
671 | } | |
672 | out: | |
cafe5635 | 673 | mutex_unlock(&c->bucket_lock); |
7dc19d5a DC |
674 | return freed; |
675 | } | |
676 | ||
677 | static unsigned long bch_mca_count(struct shrinker *shrink, | |
678 | struct shrink_control *sc) | |
679 | { | |
680 | struct cache_set *c = container_of(shrink, struct cache_set, shrink); | |
681 | ||
682 | if (c->shrinker_disabled) | |
683 | return 0; | |
684 | ||
685 | if (c->try_harder) | |
686 | return 0; | |
687 | ||
688 | return mca_can_free(c) * c->btree_pages; | |
cafe5635 KO |
689 | } |
690 | ||
691 | void bch_btree_cache_free(struct cache_set *c) | |
692 | { | |
693 | struct btree *b; | |
694 | struct closure cl; | |
695 | closure_init_stack(&cl); | |
696 | ||
697 | if (c->shrink.list.next) | |
698 | unregister_shrinker(&c->shrink); | |
699 | ||
700 | mutex_lock(&c->bucket_lock); | |
701 | ||
702 | #ifdef CONFIG_BCACHE_DEBUG | |
703 | if (c->verify_data) | |
704 | list_move(&c->verify_data->list, &c->btree_cache); | |
705 | #endif | |
706 | ||
707 | list_splice(&c->btree_cache_freeable, | |
708 | &c->btree_cache); | |
709 | ||
710 | while (!list_empty(&c->btree_cache)) { | |
711 | b = list_first_entry(&c->btree_cache, struct btree, list); | |
712 | ||
713 | if (btree_node_dirty(b)) | |
714 | btree_complete_write(b, btree_current_write(b)); | |
715 | clear_bit(BTREE_NODE_dirty, &b->flags); | |
716 | ||
717 | mca_data_free(b); | |
718 | } | |
719 | ||
720 | while (!list_empty(&c->btree_cache_freed)) { | |
721 | b = list_first_entry(&c->btree_cache_freed, | |
722 | struct btree, list); | |
723 | list_del(&b->list); | |
724 | cancel_delayed_work_sync(&b->work); | |
725 | kfree(b); | |
726 | } | |
727 | ||
728 | mutex_unlock(&c->bucket_lock); | |
729 | } | |
730 | ||
731 | int bch_btree_cache_alloc(struct cache_set *c) | |
732 | { | |
733 | unsigned i; | |
734 | ||
735 | /* XXX: doesn't check for errors */ | |
736 | ||
737 | closure_init_unlocked(&c->gc); | |
738 | ||
739 | for (i = 0; i < mca_reserve(c); i++) | |
740 | mca_bucket_alloc(c, &ZERO_KEY, GFP_KERNEL); | |
741 | ||
742 | list_splice_init(&c->btree_cache, | |
743 | &c->btree_cache_freeable); | |
744 | ||
745 | #ifdef CONFIG_BCACHE_DEBUG | |
746 | mutex_init(&c->verify_lock); | |
747 | ||
748 | c->verify_data = mca_bucket_alloc(c, &ZERO_KEY, GFP_KERNEL); | |
749 | ||
750 | if (c->verify_data && | |
751 | c->verify_data->sets[0].data) | |
752 | list_del_init(&c->verify_data->list); | |
753 | else | |
754 | c->verify_data = NULL; | |
755 | #endif | |
756 | ||
7dc19d5a DC |
757 | c->shrink.count_objects = bch_mca_count; |
758 | c->shrink.scan_objects = bch_mca_scan; | |
cafe5635 KO |
759 | c->shrink.seeks = 4; |
760 | c->shrink.batch = c->btree_pages * 2; | |
761 | register_shrinker(&c->shrink); | |
762 | ||
763 | return 0; | |
764 | } | |
765 | ||
766 | /* Btree in memory cache - hash table */ | |
767 | ||
768 | static struct hlist_head *mca_hash(struct cache_set *c, struct bkey *k) | |
769 | { | |
770 | return &c->bucket_hash[hash_32(PTR_HASH(c, k), BUCKET_HASH_BITS)]; | |
771 | } | |
772 | ||
773 | static struct btree *mca_find(struct cache_set *c, struct bkey *k) | |
774 | { | |
775 | struct btree *b; | |
776 | ||
777 | rcu_read_lock(); | |
778 | hlist_for_each_entry_rcu(b, mca_hash(c, k), hash) | |
779 | if (PTR_HASH(c, &b->key) == PTR_HASH(c, k)) | |
780 | goto out; | |
781 | b = NULL; | |
782 | out: | |
783 | rcu_read_unlock(); | |
784 | return b; | |
785 | } | |
786 | ||
787 | static struct btree *mca_cannibalize(struct cache_set *c, struct bkey *k, | |
788 | int level, struct closure *cl) | |
789 | { | |
790 | int ret = -ENOMEM; | |
791 | struct btree *i; | |
792 | ||
c37511b8 KO |
793 | trace_bcache_btree_cache_cannibalize(c); |
794 | ||
cafe5635 KO |
795 | if (!cl) |
796 | return ERR_PTR(-ENOMEM); | |
797 | ||
798 | /* | |
799 | * Trying to free up some memory - i.e. reuse some btree nodes - may | |
800 | * require initiating IO to flush the dirty part of the node. If we're | |
801 | * running under generic_make_request(), that IO will never finish and | |
802 | * we would deadlock. Returning -EAGAIN causes the cache lookup code to | |
803 | * punt to workqueue and retry. | |
804 | */ | |
805 | if (current->bio_list) | |
806 | return ERR_PTR(-EAGAIN); | |
807 | ||
808 | if (c->try_harder && c->try_harder != cl) { | |
809 | closure_wait_event_async(&c->try_wait, cl, !c->try_harder); | |
810 | return ERR_PTR(-EAGAIN); | |
811 | } | |
812 | ||
cafe5635 KO |
813 | c->try_harder = cl; |
814 | c->try_harder_start = local_clock(); | |
815 | retry: | |
816 | list_for_each_entry_reverse(i, &c->btree_cache, list) { | |
817 | int r = mca_reap(i, cl, btree_order(k)); | |
818 | if (!r) | |
819 | return i; | |
820 | if (r != -ENOMEM) | |
821 | ret = r; | |
822 | } | |
823 | ||
824 | if (ret == -EAGAIN && | |
825 | closure_blocking(cl)) { | |
826 | mutex_unlock(&c->bucket_lock); | |
827 | closure_sync(cl); | |
828 | mutex_lock(&c->bucket_lock); | |
829 | goto retry; | |
830 | } | |
831 | ||
832 | return ERR_PTR(ret); | |
833 | } | |
834 | ||
835 | /* | |
836 | * We can only have one thread cannibalizing other cached btree nodes at a time, | |
837 | * or we'll deadlock. We use an open coded mutex to ensure that, which a | |
838 | * cannibalize_bucket() will take. This means every time we unlock the root of | |
839 | * the btree, we need to release this lock if we have it held. | |
840 | */ | |
841 | void bch_cannibalize_unlock(struct cache_set *c, struct closure *cl) | |
842 | { | |
843 | if (c->try_harder == cl) { | |
169ef1cf | 844 | bch_time_stats_update(&c->try_harder_time, c->try_harder_start); |
cafe5635 KO |
845 | c->try_harder = NULL; |
846 | __closure_wake_up(&c->try_wait); | |
847 | } | |
848 | } | |
849 | ||
850 | static struct btree *mca_alloc(struct cache_set *c, struct bkey *k, | |
851 | int level, struct closure *cl) | |
852 | { | |
853 | struct btree *b; | |
854 | ||
855 | lockdep_assert_held(&c->bucket_lock); | |
856 | ||
857 | if (mca_find(c, k)) | |
858 | return NULL; | |
859 | ||
860 | /* btree_free() doesn't free memory; it sticks the node on the end of | |
861 | * the list. Check if there's any freed nodes there: | |
862 | */ | |
863 | list_for_each_entry(b, &c->btree_cache_freeable, list) | |
864 | if (!mca_reap(b, NULL, btree_order(k))) | |
865 | goto out; | |
866 | ||
867 | /* We never free struct btree itself, just the memory that holds the on | |
868 | * disk node. Check the freed list before allocating a new one: | |
869 | */ | |
870 | list_for_each_entry(b, &c->btree_cache_freed, list) | |
871 | if (!mca_reap(b, NULL, 0)) { | |
872 | mca_data_alloc(b, k, __GFP_NOWARN|GFP_NOIO); | |
873 | if (!b->sets[0].data) | |
874 | goto err; | |
875 | else | |
876 | goto out; | |
877 | } | |
878 | ||
879 | b = mca_bucket_alloc(c, k, __GFP_NOWARN|GFP_NOIO); | |
880 | if (!b) | |
881 | goto err; | |
882 | ||
883 | BUG_ON(!down_write_trylock(&b->lock)); | |
884 | if (!b->sets->data) | |
885 | goto err; | |
886 | out: | |
887 | BUG_ON(!closure_is_unlocked(&b->io.cl)); | |
888 | ||
889 | bkey_copy(&b->key, k); | |
890 | list_move(&b->list, &c->btree_cache); | |
891 | hlist_del_init_rcu(&b->hash); | |
892 | hlist_add_head_rcu(&b->hash, mca_hash(c, k)); | |
893 | ||
894 | lock_set_subclass(&b->lock.dep_map, level + 1, _THIS_IP_); | |
895 | b->level = level; | |
d6fd3b11 | 896 | b->parent = (void *) ~0UL; |
cafe5635 KO |
897 | |
898 | mca_reinit(b); | |
899 | ||
900 | return b; | |
901 | err: | |
902 | if (b) | |
903 | rw_unlock(true, b); | |
904 | ||
905 | b = mca_cannibalize(c, k, level, cl); | |
906 | if (!IS_ERR(b)) | |
907 | goto out; | |
908 | ||
909 | return b; | |
910 | } | |
911 | ||
912 | /** | |
913 | * bch_btree_node_get - find a btree node in the cache and lock it, reading it | |
914 | * in from disk if necessary. | |
915 | * | |
916 | * If IO is necessary, it uses the closure embedded in struct btree_op to wait; | |
917 | * if that closure is in non blocking mode, will return -EAGAIN. | |
918 | * | |
919 | * The btree node will have either a read or a write lock held, depending on | |
920 | * level and op->lock. | |
921 | */ | |
922 | struct btree *bch_btree_node_get(struct cache_set *c, struct bkey *k, | |
923 | int level, struct btree_op *op) | |
924 | { | |
925 | int i = 0; | |
926 | bool write = level <= op->lock; | |
927 | struct btree *b; | |
928 | ||
929 | BUG_ON(level < 0); | |
930 | retry: | |
931 | b = mca_find(c, k); | |
932 | ||
933 | if (!b) { | |
57943511 KO |
934 | if (current->bio_list) |
935 | return ERR_PTR(-EAGAIN); | |
936 | ||
cafe5635 KO |
937 | mutex_lock(&c->bucket_lock); |
938 | b = mca_alloc(c, k, level, &op->cl); | |
939 | mutex_unlock(&c->bucket_lock); | |
940 | ||
941 | if (!b) | |
942 | goto retry; | |
943 | if (IS_ERR(b)) | |
944 | return b; | |
945 | ||
57943511 | 946 | bch_btree_node_read(b); |
cafe5635 KO |
947 | |
948 | if (!write) | |
949 | downgrade_write(&b->lock); | |
950 | } else { | |
951 | rw_lock(write, b, level); | |
952 | if (PTR_HASH(c, &b->key) != PTR_HASH(c, k)) { | |
953 | rw_unlock(write, b); | |
954 | goto retry; | |
955 | } | |
956 | BUG_ON(b->level != level); | |
957 | } | |
958 | ||
959 | b->accessed = 1; | |
960 | ||
961 | for (; i <= b->nsets && b->sets[i].size; i++) { | |
962 | prefetch(b->sets[i].tree); | |
963 | prefetch(b->sets[i].data); | |
964 | } | |
965 | ||
966 | for (; i <= b->nsets; i++) | |
967 | prefetch(b->sets[i].data); | |
968 | ||
57943511 | 969 | if (btree_node_io_error(b)) { |
cafe5635 | 970 | rw_unlock(write, b); |
57943511 KO |
971 | return ERR_PTR(-EIO); |
972 | } | |
973 | ||
974 | BUG_ON(!b->written); | |
cafe5635 KO |
975 | |
976 | return b; | |
977 | } | |
978 | ||
979 | static void btree_node_prefetch(struct cache_set *c, struct bkey *k, int level) | |
980 | { | |
981 | struct btree *b; | |
982 | ||
983 | mutex_lock(&c->bucket_lock); | |
984 | b = mca_alloc(c, k, level, NULL); | |
985 | mutex_unlock(&c->bucket_lock); | |
986 | ||
987 | if (!IS_ERR_OR_NULL(b)) { | |
57943511 | 988 | bch_btree_node_read(b); |
cafe5635 KO |
989 | rw_unlock(true, b); |
990 | } | |
991 | } | |
992 | ||
993 | /* Btree alloc */ | |
994 | ||
995 | static void btree_node_free(struct btree *b, struct btree_op *op) | |
996 | { | |
997 | unsigned i; | |
998 | ||
c37511b8 KO |
999 | trace_bcache_btree_node_free(b); |
1000 | ||
cafe5635 KO |
1001 | /* |
1002 | * The BUG_ON() in btree_node_get() implies that we must have a write | |
1003 | * lock on parent to free or even invalidate a node | |
1004 | */ | |
1005 | BUG_ON(op->lock <= b->level); | |
1006 | BUG_ON(b == b->c->root); | |
cafe5635 KO |
1007 | |
1008 | if (btree_node_dirty(b)) | |
1009 | btree_complete_write(b, btree_current_write(b)); | |
1010 | clear_bit(BTREE_NODE_dirty, &b->flags); | |
1011 | ||
cafe5635 KO |
1012 | cancel_delayed_work(&b->work); |
1013 | ||
1014 | mutex_lock(&b->c->bucket_lock); | |
1015 | ||
1016 | for (i = 0; i < KEY_PTRS(&b->key); i++) { | |
1017 | BUG_ON(atomic_read(&PTR_BUCKET(b->c, &b->key, i)->pin)); | |
1018 | ||
1019 | bch_inc_gen(PTR_CACHE(b->c, &b->key, i), | |
1020 | PTR_BUCKET(b->c, &b->key, i)); | |
1021 | } | |
1022 | ||
1023 | bch_bucket_free(b->c, &b->key); | |
1024 | mca_bucket_free(b); | |
1025 | mutex_unlock(&b->c->bucket_lock); | |
1026 | } | |
1027 | ||
1028 | struct btree *bch_btree_node_alloc(struct cache_set *c, int level, | |
1029 | struct closure *cl) | |
1030 | { | |
1031 | BKEY_PADDED(key) k; | |
1032 | struct btree *b = ERR_PTR(-EAGAIN); | |
1033 | ||
1034 | mutex_lock(&c->bucket_lock); | |
1035 | retry: | |
1036 | if (__bch_bucket_alloc_set(c, WATERMARK_METADATA, &k.key, 1, cl)) | |
1037 | goto err; | |
1038 | ||
1039 | SET_KEY_SIZE(&k.key, c->btree_pages * PAGE_SECTORS); | |
1040 | ||
1041 | b = mca_alloc(c, &k.key, level, cl); | |
1042 | if (IS_ERR(b)) | |
1043 | goto err_free; | |
1044 | ||
1045 | if (!b) { | |
b1a67b0f KO |
1046 | cache_bug(c, |
1047 | "Tried to allocate bucket that was in btree cache"); | |
cafe5635 KO |
1048 | __bkey_put(c, &k.key); |
1049 | goto retry; | |
1050 | } | |
1051 | ||
cafe5635 KO |
1052 | b->accessed = 1; |
1053 | bch_bset_init_next(b); | |
1054 | ||
1055 | mutex_unlock(&c->bucket_lock); | |
c37511b8 KO |
1056 | |
1057 | trace_bcache_btree_node_alloc(b); | |
cafe5635 KO |
1058 | return b; |
1059 | err_free: | |
1060 | bch_bucket_free(c, &k.key); | |
1061 | __bkey_put(c, &k.key); | |
1062 | err: | |
1063 | mutex_unlock(&c->bucket_lock); | |
c37511b8 KO |
1064 | |
1065 | trace_bcache_btree_node_alloc_fail(b); | |
cafe5635 KO |
1066 | return b; |
1067 | } | |
1068 | ||
1069 | static struct btree *btree_node_alloc_replacement(struct btree *b, | |
1070 | struct closure *cl) | |
1071 | { | |
1072 | struct btree *n = bch_btree_node_alloc(b->c, b->level, cl); | |
1073 | if (!IS_ERR_OR_NULL(n)) | |
1074 | bch_btree_sort_into(b, n); | |
1075 | ||
1076 | return n; | |
1077 | } | |
1078 | ||
1079 | /* Garbage collection */ | |
1080 | ||
1081 | uint8_t __bch_btree_mark_key(struct cache_set *c, int level, struct bkey *k) | |
1082 | { | |
1083 | uint8_t stale = 0; | |
1084 | unsigned i; | |
1085 | struct bucket *g; | |
1086 | ||
1087 | /* | |
1088 | * ptr_invalid() can't return true for the keys that mark btree nodes as | |
1089 | * freed, but since ptr_bad() returns true we'll never actually use them | |
1090 | * for anything and thus we don't want mark their pointers here | |
1091 | */ | |
1092 | if (!bkey_cmp(k, &ZERO_KEY)) | |
1093 | return stale; | |
1094 | ||
1095 | for (i = 0; i < KEY_PTRS(k); i++) { | |
1096 | if (!ptr_available(c, k, i)) | |
1097 | continue; | |
1098 | ||
1099 | g = PTR_BUCKET(c, k, i); | |
1100 | ||
1101 | if (gen_after(g->gc_gen, PTR_GEN(k, i))) | |
1102 | g->gc_gen = PTR_GEN(k, i); | |
1103 | ||
1104 | if (ptr_stale(c, k, i)) { | |
1105 | stale = max(stale, ptr_stale(c, k, i)); | |
1106 | continue; | |
1107 | } | |
1108 | ||
1109 | cache_bug_on(GC_MARK(g) && | |
1110 | (GC_MARK(g) == GC_MARK_METADATA) != (level != 0), | |
1111 | c, "inconsistent ptrs: mark = %llu, level = %i", | |
1112 | GC_MARK(g), level); | |
1113 | ||
1114 | if (level) | |
1115 | SET_GC_MARK(g, GC_MARK_METADATA); | |
1116 | else if (KEY_DIRTY(k)) | |
1117 | SET_GC_MARK(g, GC_MARK_DIRTY); | |
1118 | ||
1119 | /* guard against overflow */ | |
1120 | SET_GC_SECTORS_USED(g, min_t(unsigned, | |
1121 | GC_SECTORS_USED(g) + KEY_SIZE(k), | |
1122 | (1 << 14) - 1)); | |
1123 | ||
1124 | BUG_ON(!GC_SECTORS_USED(g)); | |
1125 | } | |
1126 | ||
1127 | return stale; | |
1128 | } | |
1129 | ||
1130 | #define btree_mark_key(b, k) __bch_btree_mark_key(b->c, b->level, k) | |
1131 | ||
1132 | static int btree_gc_mark_node(struct btree *b, unsigned *keys, | |
1133 | struct gc_stat *gc) | |
1134 | { | |
1135 | uint8_t stale = 0; | |
1136 | unsigned last_dev = -1; | |
1137 | struct bcache_device *d = NULL; | |
1138 | struct bkey *k; | |
1139 | struct btree_iter iter; | |
1140 | struct bset_tree *t; | |
1141 | ||
1142 | gc->nodes++; | |
1143 | ||
1144 | for_each_key_filter(b, k, &iter, bch_ptr_invalid) { | |
1145 | if (last_dev != KEY_INODE(k)) { | |
1146 | last_dev = KEY_INODE(k); | |
1147 | ||
1148 | d = KEY_INODE(k) < b->c->nr_uuids | |
1149 | ? b->c->devices[last_dev] | |
1150 | : NULL; | |
1151 | } | |
1152 | ||
1153 | stale = max(stale, btree_mark_key(b, k)); | |
1154 | ||
1155 | if (bch_ptr_bad(b, k)) | |
1156 | continue; | |
1157 | ||
1158 | *keys += bkey_u64s(k); | |
1159 | ||
1160 | gc->key_bytes += bkey_u64s(k); | |
1161 | gc->nkeys++; | |
1162 | ||
1163 | gc->data += KEY_SIZE(k); | |
444fc0b6 | 1164 | if (KEY_DIRTY(k)) |
cafe5635 | 1165 | gc->dirty += KEY_SIZE(k); |
cafe5635 KO |
1166 | } |
1167 | ||
1168 | for (t = b->sets; t <= &b->sets[b->nsets]; t++) | |
1169 | btree_bug_on(t->size && | |
1170 | bset_written(b, t) && | |
1171 | bkey_cmp(&b->key, &t->end) < 0, | |
1172 | b, "found short btree key in gc"); | |
1173 | ||
1174 | return stale; | |
1175 | } | |
1176 | ||
1177 | static struct btree *btree_gc_alloc(struct btree *b, struct bkey *k, | |
1178 | struct btree_op *op) | |
1179 | { | |
1180 | /* | |
1181 | * We block priorities from being written for the duration of garbage | |
1182 | * collection, so we can't sleep in btree_alloc() -> | |
1183 | * bch_bucket_alloc_set(), or we'd risk deadlock - so we don't pass it | |
1184 | * our closure. | |
1185 | */ | |
1186 | struct btree *n = btree_node_alloc_replacement(b, NULL); | |
1187 | ||
1188 | if (!IS_ERR_OR_NULL(n)) { | |
1189 | swap(b, n); | |
57943511 | 1190 | __bkey_put(b->c, &b->key); |
cafe5635 KO |
1191 | |
1192 | memcpy(k->ptr, b->key.ptr, | |
1193 | sizeof(uint64_t) * KEY_PTRS(&b->key)); | |
1194 | ||
cafe5635 KO |
1195 | btree_node_free(n, op); |
1196 | up_write(&n->lock); | |
1197 | } | |
1198 | ||
1199 | return b; | |
1200 | } | |
1201 | ||
1202 | /* | |
1203 | * Leaving this at 2 until we've got incremental garbage collection done; it | |
1204 | * could be higher (and has been tested with 4) except that garbage collection | |
1205 | * could take much longer, adversely affecting latency. | |
1206 | */ | |
1207 | #define GC_MERGE_NODES 2U | |
1208 | ||
1209 | struct gc_merge_info { | |
1210 | struct btree *b; | |
1211 | struct bkey *k; | |
1212 | unsigned keys; | |
1213 | }; | |
1214 | ||
1215 | static void btree_gc_coalesce(struct btree *b, struct btree_op *op, | |
1216 | struct gc_stat *gc, struct gc_merge_info *r) | |
1217 | { | |
1218 | unsigned nodes = 0, keys = 0, blocks; | |
1219 | int i; | |
1220 | ||
1221 | while (nodes < GC_MERGE_NODES && r[nodes].b) | |
1222 | keys += r[nodes++].keys; | |
1223 | ||
1224 | blocks = btree_default_blocks(b->c) * 2 / 3; | |
1225 | ||
1226 | if (nodes < 2 || | |
1227 | __set_blocks(b->sets[0].data, keys, b->c) > blocks * (nodes - 1)) | |
1228 | return; | |
1229 | ||
1230 | for (i = nodes - 1; i >= 0; --i) { | |
1231 | if (r[i].b->written) | |
1232 | r[i].b = btree_gc_alloc(r[i].b, r[i].k, op); | |
1233 | ||
1234 | if (r[i].b->written) | |
1235 | return; | |
1236 | } | |
1237 | ||
1238 | for (i = nodes - 1; i > 0; --i) { | |
1239 | struct bset *n1 = r[i].b->sets->data; | |
1240 | struct bset *n2 = r[i - 1].b->sets->data; | |
1241 | struct bkey *k, *last = NULL; | |
1242 | ||
1243 | keys = 0; | |
1244 | ||
1245 | if (i == 1) { | |
1246 | /* | |
1247 | * Last node we're not getting rid of - we're getting | |
1248 | * rid of the node at r[0]. Have to try and fit all of | |
1249 | * the remaining keys into this node; we can't ensure | |
1250 | * they will always fit due to rounding and variable | |
1251 | * length keys (shouldn't be possible in practice, | |
1252 | * though) | |
1253 | */ | |
1254 | if (__set_blocks(n1, n1->keys + r->keys, | |
1255 | b->c) > btree_blocks(r[i].b)) | |
1256 | return; | |
1257 | ||
1258 | keys = n2->keys; | |
1259 | last = &r->b->key; | |
1260 | } else | |
1261 | for (k = n2->start; | |
1262 | k < end(n2); | |
1263 | k = bkey_next(k)) { | |
1264 | if (__set_blocks(n1, n1->keys + keys + | |
1265 | bkey_u64s(k), b->c) > blocks) | |
1266 | break; | |
1267 | ||
1268 | last = k; | |
1269 | keys += bkey_u64s(k); | |
1270 | } | |
1271 | ||
1272 | BUG_ON(__set_blocks(n1, n1->keys + keys, | |
1273 | b->c) > btree_blocks(r[i].b)); | |
1274 | ||
1275 | if (last) { | |
1276 | bkey_copy_key(&r[i].b->key, last); | |
1277 | bkey_copy_key(r[i].k, last); | |
1278 | } | |
1279 | ||
1280 | memcpy(end(n1), | |
1281 | n2->start, | |
1282 | (void *) node(n2, keys) - (void *) n2->start); | |
1283 | ||
1284 | n1->keys += keys; | |
1285 | ||
1286 | memmove(n2->start, | |
1287 | node(n2, keys), | |
1288 | (void *) end(n2) - (void *) node(n2, keys)); | |
1289 | ||
1290 | n2->keys -= keys; | |
1291 | ||
1292 | r[i].keys = n1->keys; | |
1293 | r[i - 1].keys = n2->keys; | |
1294 | } | |
1295 | ||
1296 | btree_node_free(r->b, op); | |
1297 | up_write(&r->b->lock); | |
1298 | ||
c37511b8 | 1299 | trace_bcache_btree_gc_coalesce(nodes); |
cafe5635 KO |
1300 | |
1301 | gc->nodes--; | |
1302 | nodes--; | |
1303 | ||
1304 | memmove(&r[0], &r[1], sizeof(struct gc_merge_info) * nodes); | |
1305 | memset(&r[nodes], 0, sizeof(struct gc_merge_info)); | |
1306 | } | |
1307 | ||
1308 | static int btree_gc_recurse(struct btree *b, struct btree_op *op, | |
1309 | struct closure *writes, struct gc_stat *gc) | |
1310 | { | |
1311 | void write(struct btree *r) | |
1312 | { | |
1313 | if (!r->written) | |
57943511 KO |
1314 | bch_btree_node_write(r, &op->cl); |
1315 | else if (btree_node_dirty(r)) | |
1316 | bch_btree_node_write(r, writes); | |
cafe5635 KO |
1317 | |
1318 | up_write(&r->lock); | |
1319 | } | |
1320 | ||
1321 | int ret = 0, stale; | |
1322 | unsigned i; | |
1323 | struct gc_merge_info r[GC_MERGE_NODES]; | |
1324 | ||
1325 | memset(r, 0, sizeof(r)); | |
1326 | ||
1327 | while ((r->k = bch_next_recurse_key(b, &b->c->gc_done))) { | |
1328 | r->b = bch_btree_node_get(b->c, r->k, b->level - 1, op); | |
1329 | ||
1330 | if (IS_ERR(r->b)) { | |
1331 | ret = PTR_ERR(r->b); | |
1332 | break; | |
1333 | } | |
1334 | ||
1335 | r->keys = 0; | |
1336 | stale = btree_gc_mark_node(r->b, &r->keys, gc); | |
1337 | ||
1338 | if (!b->written && | |
1339 | (r->b->level || stale > 10 || | |
1340 | b->c->gc_always_rewrite)) | |
1341 | r->b = btree_gc_alloc(r->b, r->k, op); | |
1342 | ||
1343 | if (r->b->level) | |
1344 | ret = btree_gc_recurse(r->b, op, writes, gc); | |
1345 | ||
1346 | if (ret) { | |
1347 | write(r->b); | |
1348 | break; | |
1349 | } | |
1350 | ||
1351 | bkey_copy_key(&b->c->gc_done, r->k); | |
1352 | ||
1353 | if (!b->written) | |
1354 | btree_gc_coalesce(b, op, gc, r); | |
1355 | ||
1356 | if (r[GC_MERGE_NODES - 1].b) | |
1357 | write(r[GC_MERGE_NODES - 1].b); | |
1358 | ||
1359 | memmove(&r[1], &r[0], | |
1360 | sizeof(struct gc_merge_info) * (GC_MERGE_NODES - 1)); | |
1361 | ||
1362 | /* When we've got incremental GC working, we'll want to do | |
1363 | * if (should_resched()) | |
1364 | * return -EAGAIN; | |
1365 | */ | |
1366 | cond_resched(); | |
1367 | #if 0 | |
1368 | if (need_resched()) { | |
1369 | ret = -EAGAIN; | |
1370 | break; | |
1371 | } | |
1372 | #endif | |
1373 | } | |
1374 | ||
1375 | for (i = 1; i < GC_MERGE_NODES && r[i].b; i++) | |
1376 | write(r[i].b); | |
1377 | ||
1378 | /* Might have freed some children, must remove their keys */ | |
1379 | if (!b->written) | |
1380 | bch_btree_sort(b); | |
1381 | ||
1382 | return ret; | |
1383 | } | |
1384 | ||
1385 | static int bch_btree_gc_root(struct btree *b, struct btree_op *op, | |
1386 | struct closure *writes, struct gc_stat *gc) | |
1387 | { | |
1388 | struct btree *n = NULL; | |
1389 | unsigned keys = 0; | |
1390 | int ret = 0, stale = btree_gc_mark_node(b, &keys, gc); | |
1391 | ||
1392 | if (b->level || stale > 10) | |
1393 | n = btree_node_alloc_replacement(b, NULL); | |
1394 | ||
1395 | if (!IS_ERR_OR_NULL(n)) | |
1396 | swap(b, n); | |
1397 | ||
1398 | if (b->level) | |
1399 | ret = btree_gc_recurse(b, op, writes, gc); | |
1400 | ||
1401 | if (!b->written || btree_node_dirty(b)) { | |
57943511 | 1402 | bch_btree_node_write(b, n ? &op->cl : NULL); |
cafe5635 KO |
1403 | } |
1404 | ||
1405 | if (!IS_ERR_OR_NULL(n)) { | |
1406 | closure_sync(&op->cl); | |
1407 | bch_btree_set_root(b); | |
1408 | btree_node_free(n, op); | |
1409 | rw_unlock(true, b); | |
1410 | } | |
1411 | ||
1412 | return ret; | |
1413 | } | |
1414 | ||
1415 | static void btree_gc_start(struct cache_set *c) | |
1416 | { | |
1417 | struct cache *ca; | |
1418 | struct bucket *b; | |
cafe5635 KO |
1419 | unsigned i; |
1420 | ||
1421 | if (!c->gc_mark_valid) | |
1422 | return; | |
1423 | ||
1424 | mutex_lock(&c->bucket_lock); | |
1425 | ||
1426 | c->gc_mark_valid = 0; | |
1427 | c->gc_done = ZERO_KEY; | |
1428 | ||
1429 | for_each_cache(ca, c, i) | |
1430 | for_each_bucket(b, ca) { | |
1431 | b->gc_gen = b->gen; | |
29ebf465 | 1432 | if (!atomic_read(&b->pin)) { |
cafe5635 | 1433 | SET_GC_MARK(b, GC_MARK_RECLAIMABLE); |
29ebf465 KO |
1434 | SET_GC_SECTORS_USED(b, 0); |
1435 | } | |
cafe5635 KO |
1436 | } |
1437 | ||
cafe5635 KO |
1438 | mutex_unlock(&c->bucket_lock); |
1439 | } | |
1440 | ||
1441 | size_t bch_btree_gc_finish(struct cache_set *c) | |
1442 | { | |
1443 | size_t available = 0; | |
1444 | struct bucket *b; | |
1445 | struct cache *ca; | |
cafe5635 KO |
1446 | unsigned i; |
1447 | ||
1448 | mutex_lock(&c->bucket_lock); | |
1449 | ||
1450 | set_gc_sectors(c); | |
1451 | c->gc_mark_valid = 1; | |
1452 | c->need_gc = 0; | |
1453 | ||
1454 | if (c->root) | |
1455 | for (i = 0; i < KEY_PTRS(&c->root->key); i++) | |
1456 | SET_GC_MARK(PTR_BUCKET(c, &c->root->key, i), | |
1457 | GC_MARK_METADATA); | |
1458 | ||
1459 | for (i = 0; i < KEY_PTRS(&c->uuid_bucket); i++) | |
1460 | SET_GC_MARK(PTR_BUCKET(c, &c->uuid_bucket, i), | |
1461 | GC_MARK_METADATA); | |
1462 | ||
1463 | for_each_cache(ca, c, i) { | |
1464 | uint64_t *i; | |
1465 | ||
1466 | ca->invalidate_needs_gc = 0; | |
1467 | ||
1468 | for (i = ca->sb.d; i < ca->sb.d + ca->sb.keys; i++) | |
1469 | SET_GC_MARK(ca->buckets + *i, GC_MARK_METADATA); | |
1470 | ||
1471 | for (i = ca->prio_buckets; | |
1472 | i < ca->prio_buckets + prio_buckets(ca) * 2; i++) | |
1473 | SET_GC_MARK(ca->buckets + *i, GC_MARK_METADATA); | |
1474 | ||
1475 | for_each_bucket(b, ca) { | |
1476 | b->last_gc = b->gc_gen; | |
1477 | c->need_gc = max(c->need_gc, bucket_gc_gen(b)); | |
1478 | ||
1479 | if (!atomic_read(&b->pin) && | |
1480 | GC_MARK(b) == GC_MARK_RECLAIMABLE) { | |
1481 | available++; | |
1482 | if (!GC_SECTORS_USED(b)) | |
1483 | bch_bucket_add_unused(ca, b); | |
1484 | } | |
1485 | } | |
1486 | } | |
1487 | ||
cafe5635 KO |
1488 | mutex_unlock(&c->bucket_lock); |
1489 | return available; | |
1490 | } | |
1491 | ||
1492 | static void bch_btree_gc(struct closure *cl) | |
1493 | { | |
1494 | struct cache_set *c = container_of(cl, struct cache_set, gc.cl); | |
1495 | int ret; | |
1496 | unsigned long available; | |
1497 | struct gc_stat stats; | |
1498 | struct closure writes; | |
1499 | struct btree_op op; | |
cafe5635 | 1500 | uint64_t start_time = local_clock(); |
57943511 | 1501 | |
c37511b8 | 1502 | trace_bcache_gc_start(c); |
cafe5635 KO |
1503 | |
1504 | memset(&stats, 0, sizeof(struct gc_stat)); | |
1505 | closure_init_stack(&writes); | |
1506 | bch_btree_op_init_stack(&op); | |
1507 | op.lock = SHRT_MAX; | |
1508 | ||
1509 | btree_gc_start(c); | |
1510 | ||
57943511 KO |
1511 | atomic_inc(&c->prio_blocked); |
1512 | ||
cafe5635 KO |
1513 | ret = btree_root(gc_root, c, &op, &writes, &stats); |
1514 | closure_sync(&op.cl); | |
1515 | closure_sync(&writes); | |
1516 | ||
1517 | if (ret) { | |
cafe5635 | 1518 | pr_warn("gc failed!"); |
cafe5635 KO |
1519 | continue_at(cl, bch_btree_gc, bch_gc_wq); |
1520 | } | |
1521 | ||
1522 | /* Possibly wait for new UUIDs or whatever to hit disk */ | |
1523 | bch_journal_meta(c, &op.cl); | |
1524 | closure_sync(&op.cl); | |
1525 | ||
1526 | available = bch_btree_gc_finish(c); | |
1527 | ||
57943511 KO |
1528 | atomic_dec(&c->prio_blocked); |
1529 | wake_up_allocators(c); | |
1530 | ||
169ef1cf | 1531 | bch_time_stats_update(&c->btree_gc_time, start_time); |
cafe5635 KO |
1532 | |
1533 | stats.key_bytes *= sizeof(uint64_t); | |
1534 | stats.dirty <<= 9; | |
1535 | stats.data <<= 9; | |
1536 | stats.in_use = (c->nbuckets - available) * 100 / c->nbuckets; | |
1537 | memcpy(&c->gc_stats, &stats, sizeof(struct gc_stat)); | |
cafe5635 | 1538 | |
c37511b8 | 1539 | trace_bcache_gc_end(c); |
cafe5635 KO |
1540 | |
1541 | continue_at(cl, bch_moving_gc, bch_gc_wq); | |
1542 | } | |
1543 | ||
1544 | void bch_queue_gc(struct cache_set *c) | |
1545 | { | |
1546 | closure_trylock_call(&c->gc.cl, bch_btree_gc, bch_gc_wq, &c->cl); | |
1547 | } | |
1548 | ||
1549 | /* Initial partial gc */ | |
1550 | ||
1551 | static int bch_btree_check_recurse(struct btree *b, struct btree_op *op, | |
1552 | unsigned long **seen) | |
1553 | { | |
1554 | int ret; | |
1555 | unsigned i; | |
1556 | struct bkey *k; | |
1557 | struct bucket *g; | |
1558 | struct btree_iter iter; | |
1559 | ||
1560 | for_each_key_filter(b, k, &iter, bch_ptr_invalid) { | |
1561 | for (i = 0; i < KEY_PTRS(k); i++) { | |
1562 | if (!ptr_available(b->c, k, i)) | |
1563 | continue; | |
1564 | ||
1565 | g = PTR_BUCKET(b->c, k, i); | |
1566 | ||
1567 | if (!__test_and_set_bit(PTR_BUCKET_NR(b->c, k, i), | |
1568 | seen[PTR_DEV(k, i)]) || | |
1569 | !ptr_stale(b->c, k, i)) { | |
1570 | g->gen = PTR_GEN(k, i); | |
1571 | ||
1572 | if (b->level) | |
1573 | g->prio = BTREE_PRIO; | |
1574 | else if (g->prio == BTREE_PRIO) | |
1575 | g->prio = INITIAL_PRIO; | |
1576 | } | |
1577 | } | |
1578 | ||
1579 | btree_mark_key(b, k); | |
1580 | } | |
1581 | ||
1582 | if (b->level) { | |
1583 | k = bch_next_recurse_key(b, &ZERO_KEY); | |
1584 | ||
1585 | while (k) { | |
1586 | struct bkey *p = bch_next_recurse_key(b, k); | |
1587 | if (p) | |
1588 | btree_node_prefetch(b->c, p, b->level - 1); | |
1589 | ||
1590 | ret = btree(check_recurse, k, b, op, seen); | |
1591 | if (ret) | |
1592 | return ret; | |
1593 | ||
1594 | k = p; | |
1595 | } | |
1596 | } | |
1597 | ||
1598 | return 0; | |
1599 | } | |
1600 | ||
1601 | int bch_btree_check(struct cache_set *c, struct btree_op *op) | |
1602 | { | |
1603 | int ret = -ENOMEM; | |
1604 | unsigned i; | |
1605 | unsigned long *seen[MAX_CACHES_PER_SET]; | |
1606 | ||
1607 | memset(seen, 0, sizeof(seen)); | |
1608 | ||
1609 | for (i = 0; c->cache[i]; i++) { | |
1610 | size_t n = DIV_ROUND_UP(c->cache[i]->sb.nbuckets, 8); | |
1611 | seen[i] = kmalloc(n, GFP_KERNEL); | |
1612 | if (!seen[i]) | |
1613 | goto err; | |
1614 | ||
1615 | /* Disables the seen array until prio_read() uses it too */ | |
1616 | memset(seen[i], 0xFF, n); | |
1617 | } | |
1618 | ||
1619 | ret = btree_root(check_recurse, c, op, seen); | |
1620 | err: | |
1621 | for (i = 0; i < MAX_CACHES_PER_SET; i++) | |
1622 | kfree(seen[i]); | |
1623 | return ret; | |
1624 | } | |
1625 | ||
1626 | /* Btree insertion */ | |
1627 | ||
1628 | static void shift_keys(struct btree *b, struct bkey *where, struct bkey *insert) | |
1629 | { | |
1630 | struct bset *i = b->sets[b->nsets].data; | |
1631 | ||
1632 | memmove((uint64_t *) where + bkey_u64s(insert), | |
1633 | where, | |
1634 | (void *) end(i) - (void *) where); | |
1635 | ||
1636 | i->keys += bkey_u64s(insert); | |
1637 | bkey_copy(where, insert); | |
1638 | bch_bset_fix_lookup_table(b, where); | |
1639 | } | |
1640 | ||
1641 | static bool fix_overlapping_extents(struct btree *b, | |
1642 | struct bkey *insert, | |
1643 | struct btree_iter *iter, | |
1644 | struct btree_op *op) | |
1645 | { | |
279afbad | 1646 | void subtract_dirty(struct bkey *k, uint64_t offset, int sectors) |
cafe5635 | 1647 | { |
279afbad KO |
1648 | if (KEY_DIRTY(k)) |
1649 | bcache_dev_sectors_dirty_add(b->c, KEY_INODE(k), | |
1650 | offset, -sectors); | |
cafe5635 KO |
1651 | } |
1652 | ||
279afbad | 1653 | uint64_t old_offset; |
cafe5635 KO |
1654 | unsigned old_size, sectors_found = 0; |
1655 | ||
1656 | while (1) { | |
1657 | struct bkey *k = bch_btree_iter_next(iter); | |
1658 | if (!k || | |
1659 | bkey_cmp(&START_KEY(k), insert) >= 0) | |
1660 | break; | |
1661 | ||
1662 | if (bkey_cmp(k, &START_KEY(insert)) <= 0) | |
1663 | continue; | |
1664 | ||
279afbad | 1665 | old_offset = KEY_START(k); |
cafe5635 KO |
1666 | old_size = KEY_SIZE(k); |
1667 | ||
1668 | /* | |
1669 | * We might overlap with 0 size extents; we can't skip these | |
1670 | * because if they're in the set we're inserting to we have to | |
1671 | * adjust them so they don't overlap with the key we're | |
1672 | * inserting. But we don't want to check them for BTREE_REPLACE | |
1673 | * operations. | |
1674 | */ | |
1675 | ||
1676 | if (op->type == BTREE_REPLACE && | |
1677 | KEY_SIZE(k)) { | |
1678 | /* | |
1679 | * k might have been split since we inserted/found the | |
1680 | * key we're replacing | |
1681 | */ | |
1682 | unsigned i; | |
1683 | uint64_t offset = KEY_START(k) - | |
1684 | KEY_START(&op->replace); | |
1685 | ||
1686 | /* But it must be a subset of the replace key */ | |
1687 | if (KEY_START(k) < KEY_START(&op->replace) || | |
1688 | KEY_OFFSET(k) > KEY_OFFSET(&op->replace)) | |
1689 | goto check_failed; | |
1690 | ||
1691 | /* We didn't find a key that we were supposed to */ | |
1692 | if (KEY_START(k) > KEY_START(insert) + sectors_found) | |
1693 | goto check_failed; | |
1694 | ||
1695 | if (KEY_PTRS(&op->replace) != KEY_PTRS(k)) | |
1696 | goto check_failed; | |
1697 | ||
1698 | /* skip past gen */ | |
1699 | offset <<= 8; | |
1700 | ||
1701 | BUG_ON(!KEY_PTRS(&op->replace)); | |
1702 | ||
1703 | for (i = 0; i < KEY_PTRS(&op->replace); i++) | |
1704 | if (k->ptr[i] != op->replace.ptr[i] + offset) | |
1705 | goto check_failed; | |
1706 | ||
1707 | sectors_found = KEY_OFFSET(k) - KEY_START(insert); | |
1708 | } | |
1709 | ||
1710 | if (bkey_cmp(insert, k) < 0 && | |
1711 | bkey_cmp(&START_KEY(insert), &START_KEY(k)) > 0) { | |
1712 | /* | |
1713 | * We overlapped in the middle of an existing key: that | |
1714 | * means we have to split the old key. But we have to do | |
1715 | * slightly different things depending on whether the | |
1716 | * old key has been written out yet. | |
1717 | */ | |
1718 | ||
1719 | struct bkey *top; | |
1720 | ||
279afbad | 1721 | subtract_dirty(k, KEY_START(insert), KEY_SIZE(insert)); |
cafe5635 KO |
1722 | |
1723 | if (bkey_written(b, k)) { | |
1724 | /* | |
1725 | * We insert a new key to cover the top of the | |
1726 | * old key, and the old key is modified in place | |
1727 | * to represent the bottom split. | |
1728 | * | |
1729 | * It's completely arbitrary whether the new key | |
1730 | * is the top or the bottom, but it has to match | |
1731 | * up with what btree_sort_fixup() does - it | |
1732 | * doesn't check for this kind of overlap, it | |
1733 | * depends on us inserting a new key for the top | |
1734 | * here. | |
1735 | */ | |
1736 | top = bch_bset_search(b, &b->sets[b->nsets], | |
1737 | insert); | |
1738 | shift_keys(b, top, k); | |
1739 | } else { | |
1740 | BKEY_PADDED(key) temp; | |
1741 | bkey_copy(&temp.key, k); | |
1742 | shift_keys(b, k, &temp.key); | |
1743 | top = bkey_next(k); | |
1744 | } | |
1745 | ||
1746 | bch_cut_front(insert, top); | |
1747 | bch_cut_back(&START_KEY(insert), k); | |
1748 | bch_bset_fix_invalidated_key(b, k); | |
1749 | return false; | |
1750 | } | |
1751 | ||
1752 | if (bkey_cmp(insert, k) < 0) { | |
1753 | bch_cut_front(insert, k); | |
1754 | } else { | |
1fa8455d KO |
1755 | if (bkey_cmp(&START_KEY(insert), &START_KEY(k)) > 0) |
1756 | old_offset = KEY_START(insert); | |
1757 | ||
cafe5635 KO |
1758 | if (bkey_written(b, k) && |
1759 | bkey_cmp(&START_KEY(insert), &START_KEY(k)) <= 0) { | |
1760 | /* | |
1761 | * Completely overwrote, so we don't have to | |
1762 | * invalidate the binary search tree | |
1763 | */ | |
1764 | bch_cut_front(k, k); | |
1765 | } else { | |
1766 | __bch_cut_back(&START_KEY(insert), k); | |
1767 | bch_bset_fix_invalidated_key(b, k); | |
1768 | } | |
1769 | } | |
1770 | ||
279afbad | 1771 | subtract_dirty(k, old_offset, old_size - KEY_SIZE(k)); |
cafe5635 KO |
1772 | } |
1773 | ||
1774 | check_failed: | |
1775 | if (op->type == BTREE_REPLACE) { | |
1776 | if (!sectors_found) { | |
1777 | op->insert_collision = true; | |
1778 | return true; | |
1779 | } else if (sectors_found < KEY_SIZE(insert)) { | |
1780 | SET_KEY_OFFSET(insert, KEY_OFFSET(insert) - | |
1781 | (KEY_SIZE(insert) - sectors_found)); | |
1782 | SET_KEY_SIZE(insert, sectors_found); | |
1783 | } | |
1784 | } | |
1785 | ||
1786 | return false; | |
1787 | } | |
1788 | ||
1789 | static bool btree_insert_key(struct btree *b, struct btree_op *op, | |
1790 | struct bkey *k) | |
1791 | { | |
1792 | struct bset *i = b->sets[b->nsets].data; | |
1793 | struct bkey *m, *prev; | |
85b1492e | 1794 | unsigned status = BTREE_INSERT_STATUS_INSERT; |
cafe5635 KO |
1795 | |
1796 | BUG_ON(bkey_cmp(k, &b->key) > 0); | |
1797 | BUG_ON(b->level && !KEY_PTRS(k)); | |
1798 | BUG_ON(!b->level && !KEY_OFFSET(k)); | |
1799 | ||
1800 | if (!b->level) { | |
1801 | struct btree_iter iter; | |
1802 | struct bkey search = KEY(KEY_INODE(k), KEY_START(k), 0); | |
1803 | ||
1804 | /* | |
1805 | * bset_search() returns the first key that is strictly greater | |
1806 | * than the search key - but for back merging, we want to find | |
1807 | * the first key that is greater than or equal to KEY_START(k) - | |
1808 | * unless KEY_START(k) is 0. | |
1809 | */ | |
1810 | if (KEY_OFFSET(&search)) | |
1811 | SET_KEY_OFFSET(&search, KEY_OFFSET(&search) - 1); | |
1812 | ||
1813 | prev = NULL; | |
1814 | m = bch_btree_iter_init(b, &iter, &search); | |
1815 | ||
1816 | if (fix_overlapping_extents(b, k, &iter, op)) | |
1817 | return false; | |
1818 | ||
1fa8455d KO |
1819 | if (KEY_DIRTY(k)) |
1820 | bcache_dev_sectors_dirty_add(b->c, KEY_INODE(k), | |
1821 | KEY_START(k), KEY_SIZE(k)); | |
1822 | ||
cafe5635 KO |
1823 | while (m != end(i) && |
1824 | bkey_cmp(k, &START_KEY(m)) > 0) | |
1825 | prev = m, m = bkey_next(m); | |
1826 | ||
1827 | if (key_merging_disabled(b->c)) | |
1828 | goto insert; | |
1829 | ||
1830 | /* prev is in the tree, if we merge we're done */ | |
85b1492e | 1831 | status = BTREE_INSERT_STATUS_BACK_MERGE; |
cafe5635 KO |
1832 | if (prev && |
1833 | bch_bkey_try_merge(b, prev, k)) | |
1834 | goto merged; | |
1835 | ||
85b1492e | 1836 | status = BTREE_INSERT_STATUS_OVERWROTE; |
cafe5635 KO |
1837 | if (m != end(i) && |
1838 | KEY_PTRS(m) == KEY_PTRS(k) && !KEY_SIZE(m)) | |
1839 | goto copy; | |
1840 | ||
85b1492e | 1841 | status = BTREE_INSERT_STATUS_FRONT_MERGE; |
cafe5635 KO |
1842 | if (m != end(i) && |
1843 | bch_bkey_try_merge(b, k, m)) | |
1844 | goto copy; | |
1845 | } else | |
1846 | m = bch_bset_search(b, &b->sets[b->nsets], k); | |
1847 | ||
1848 | insert: shift_keys(b, m, k); | |
1849 | copy: bkey_copy(m, k); | |
1850 | merged: | |
85b1492e | 1851 | bch_check_keys(b, "%u for %s", status, op_type(op)); |
cafe5635 KO |
1852 | |
1853 | if (b->level && !KEY_OFFSET(k)) | |
57943511 | 1854 | btree_current_write(b)->prio_blocked++; |
cafe5635 | 1855 | |
85b1492e | 1856 | trace_bcache_btree_insert_key(b, k, op->type, status); |
cafe5635 KO |
1857 | |
1858 | return true; | |
1859 | } | |
1860 | ||
26c949f8 KO |
1861 | static bool bch_btree_insert_keys(struct btree *b, struct btree_op *op, |
1862 | struct keylist *insert_keys) | |
cafe5635 KO |
1863 | { |
1864 | bool ret = false; | |
cafe5635 KO |
1865 | unsigned oldsize = bch_count_data(b); |
1866 | ||
26c949f8 | 1867 | while (!bch_keylist_empty(insert_keys)) { |
403b6cde | 1868 | struct bset *i = write_block(b); |
26c949f8 KO |
1869 | struct bkey *k = insert_keys->bottom; |
1870 | ||
403b6cde KO |
1871 | if (b->written + __set_blocks(i, i->keys + bkey_u64s(k), b->c) |
1872 | > btree_blocks(b)) | |
1873 | break; | |
1874 | ||
1875 | if (bkey_cmp(k, &b->key) <= 0) { | |
26c949f8 KO |
1876 | bkey_put(b->c, k, b->level); |
1877 | ||
1878 | ret |= btree_insert_key(b, op, k); | |
1879 | bch_keylist_pop_front(insert_keys); | |
1880 | } else if (bkey_cmp(&START_KEY(k), &b->key) < 0) { | |
1881 | #if 0 | |
1882 | if (op->type == BTREE_REPLACE) { | |
1883 | bkey_put(b->c, k, b->level); | |
1884 | bch_keylist_pop_front(insert_keys); | |
1885 | op->insert_collision = true; | |
1886 | break; | |
1887 | } | |
1888 | #endif | |
1889 | BKEY_PADDED(key) temp; | |
1890 | bkey_copy(&temp.key, insert_keys->bottom); | |
1891 | ||
1892 | bch_cut_back(&b->key, &temp.key); | |
1893 | bch_cut_front(&b->key, insert_keys->bottom); | |
1894 | ||
1895 | ret |= btree_insert_key(b, op, &temp.key); | |
1896 | break; | |
1897 | } else { | |
1898 | break; | |
1899 | } | |
cafe5635 KO |
1900 | } |
1901 | ||
403b6cde KO |
1902 | BUG_ON(!bch_keylist_empty(insert_keys) && b->level); |
1903 | ||
cafe5635 KO |
1904 | BUG_ON(bch_count_data(b) < oldsize); |
1905 | return ret; | |
1906 | } | |
1907 | ||
26c949f8 KO |
1908 | static int btree_split(struct btree *b, struct btree_op *op, |
1909 | struct keylist *insert_keys, | |
1910 | struct keylist *parent_keys) | |
cafe5635 | 1911 | { |
d6fd3b11 | 1912 | bool split; |
cafe5635 KO |
1913 | struct btree *n1, *n2 = NULL, *n3 = NULL; |
1914 | uint64_t start_time = local_clock(); | |
1915 | ||
1916 | if (b->level) | |
1917 | set_closure_blocking(&op->cl); | |
1918 | ||
1919 | n1 = btree_node_alloc_replacement(b, &op->cl); | |
1920 | if (IS_ERR(n1)) | |
1921 | goto err; | |
1922 | ||
1923 | split = set_blocks(n1->sets[0].data, n1->c) > (btree_blocks(b) * 4) / 5; | |
1924 | ||
cafe5635 KO |
1925 | if (split) { |
1926 | unsigned keys = 0; | |
1927 | ||
c37511b8 KO |
1928 | trace_bcache_btree_node_split(b, n1->sets[0].data->keys); |
1929 | ||
cafe5635 KO |
1930 | n2 = bch_btree_node_alloc(b->c, b->level, &op->cl); |
1931 | if (IS_ERR(n2)) | |
1932 | goto err_free1; | |
1933 | ||
d6fd3b11 | 1934 | if (!b->parent) { |
cafe5635 KO |
1935 | n3 = bch_btree_node_alloc(b->c, b->level + 1, &op->cl); |
1936 | if (IS_ERR(n3)) | |
1937 | goto err_free2; | |
1938 | } | |
1939 | ||
26c949f8 | 1940 | bch_btree_insert_keys(n1, op, insert_keys); |
cafe5635 | 1941 | |
d6fd3b11 KO |
1942 | /* |
1943 | * Has to be a linear search because we don't have an auxiliary | |
cafe5635 KO |
1944 | * search tree yet |
1945 | */ | |
1946 | ||
1947 | while (keys < (n1->sets[0].data->keys * 3) / 5) | |
1948 | keys += bkey_u64s(node(n1->sets[0].data, keys)); | |
1949 | ||
1950 | bkey_copy_key(&n1->key, node(n1->sets[0].data, keys)); | |
1951 | keys += bkey_u64s(node(n1->sets[0].data, keys)); | |
1952 | ||
1953 | n2->sets[0].data->keys = n1->sets[0].data->keys - keys; | |
1954 | n1->sets[0].data->keys = keys; | |
1955 | ||
1956 | memcpy(n2->sets[0].data->start, | |
1957 | end(n1->sets[0].data), | |
1958 | n2->sets[0].data->keys * sizeof(uint64_t)); | |
1959 | ||
1960 | bkey_copy_key(&n2->key, &b->key); | |
1961 | ||
26c949f8 | 1962 | bch_keylist_add(parent_keys, &n2->key); |
57943511 | 1963 | bch_btree_node_write(n2, &op->cl); |
cafe5635 | 1964 | rw_unlock(true, n2); |
c37511b8 KO |
1965 | } else { |
1966 | trace_bcache_btree_node_compact(b, n1->sets[0].data->keys); | |
1967 | ||
26c949f8 | 1968 | bch_btree_insert_keys(n1, op, insert_keys); |
c37511b8 | 1969 | } |
cafe5635 | 1970 | |
26c949f8 | 1971 | bch_keylist_add(parent_keys, &n1->key); |
57943511 | 1972 | bch_btree_node_write(n1, &op->cl); |
cafe5635 KO |
1973 | |
1974 | if (n3) { | |
d6fd3b11 KO |
1975 | /* Depth increases, make a new root */ |
1976 | ||
cafe5635 | 1977 | bkey_copy_key(&n3->key, &MAX_KEY); |
26c949f8 | 1978 | bch_btree_insert_keys(n3, op, parent_keys); |
57943511 | 1979 | bch_btree_node_write(n3, &op->cl); |
cafe5635 KO |
1980 | |
1981 | closure_sync(&op->cl); | |
1982 | bch_btree_set_root(n3); | |
1983 | rw_unlock(true, n3); | |
d6fd3b11 KO |
1984 | } else if (!b->parent) { |
1985 | /* Root filled up but didn't need to be split */ | |
1986 | ||
26c949f8 | 1987 | parent_keys->top = parent_keys->bottom; |
cafe5635 KO |
1988 | closure_sync(&op->cl); |
1989 | bch_btree_set_root(n1); | |
1990 | } else { | |
1991 | unsigned i; | |
1992 | ||
26c949f8 KO |
1993 | bkey_copy(parent_keys->top, &b->key); |
1994 | bkey_copy_key(parent_keys->top, &ZERO_KEY); | |
cafe5635 KO |
1995 | |
1996 | for (i = 0; i < KEY_PTRS(&b->key); i++) { | |
1997 | uint8_t g = PTR_BUCKET(b->c, &b->key, i)->gen + 1; | |
1998 | ||
26c949f8 | 1999 | SET_PTR_GEN(parent_keys->top, i, g); |
cafe5635 KO |
2000 | } |
2001 | ||
26c949f8 | 2002 | bch_keylist_push(parent_keys); |
cafe5635 KO |
2003 | closure_sync(&op->cl); |
2004 | atomic_inc(&b->c->prio_blocked); | |
2005 | } | |
2006 | ||
2007 | rw_unlock(true, n1); | |
2008 | btree_node_free(b, op); | |
2009 | ||
169ef1cf | 2010 | bch_time_stats_update(&b->c->btree_split_time, start_time); |
cafe5635 KO |
2011 | |
2012 | return 0; | |
2013 | err_free2: | |
2014 | __bkey_put(n2->c, &n2->key); | |
2015 | btree_node_free(n2, op); | |
2016 | rw_unlock(true, n2); | |
2017 | err_free1: | |
2018 | __bkey_put(n1->c, &n1->key); | |
2019 | btree_node_free(n1, op); | |
2020 | rw_unlock(true, n1); | |
2021 | err: | |
2022 | if (n3 == ERR_PTR(-EAGAIN) || | |
2023 | n2 == ERR_PTR(-EAGAIN) || | |
2024 | n1 == ERR_PTR(-EAGAIN)) | |
2025 | return -EAGAIN; | |
2026 | ||
2027 | pr_warn("couldn't split"); | |
2028 | return -ENOMEM; | |
2029 | } | |
2030 | ||
26c949f8 KO |
2031 | static int bch_btree_insert_node(struct btree *b, struct btree_op *op, |
2032 | struct keylist *insert_keys) | |
cafe5635 | 2033 | { |
26c949f8 KO |
2034 | int ret = 0; |
2035 | struct keylist split_keys; | |
cafe5635 | 2036 | |
26c949f8 | 2037 | bch_keylist_init(&split_keys); |
cafe5635 | 2038 | |
26c949f8 | 2039 | BUG_ON(b->level); |
cafe5635 | 2040 | |
26c949f8 KO |
2041 | do { |
2042 | if (should_split(b)) { | |
2043 | if (current->bio_list) { | |
2044 | op->lock = b->c->root->level + 1; | |
2045 | ret = -EAGAIN; | |
2046 | } else if (op->lock <= b->c->root->level) { | |
2047 | op->lock = b->c->root->level + 1; | |
2048 | ret = -EINTR; | |
2049 | } else { | |
2050 | struct btree *parent = b->parent; | |
cafe5635 | 2051 | |
26c949f8 KO |
2052 | ret = btree_split(b, op, insert_keys, |
2053 | &split_keys); | |
2054 | insert_keys = &split_keys; | |
2055 | b = parent; | |
403b6cde KO |
2056 | if (!ret) |
2057 | ret = -EINTR; | |
cafe5635 | 2058 | } |
26c949f8 KO |
2059 | } else { |
2060 | BUG_ON(write_block(b) != b->sets[b->nsets].data); | |
cafe5635 | 2061 | |
26c949f8 KO |
2062 | if (bch_btree_insert_keys(b, op, insert_keys)) { |
2063 | if (!b->level) | |
2064 | bch_btree_leaf_dirty(b, op); | |
2065 | else | |
2066 | bch_btree_node_write(b, &op->cl); | |
2067 | } | |
cafe5635 | 2068 | } |
26c949f8 | 2069 | } while (!bch_keylist_empty(&split_keys)); |
cafe5635 | 2070 | |
26c949f8 KO |
2071 | return ret; |
2072 | } | |
cafe5635 | 2073 | |
e7c590eb KO |
2074 | int bch_btree_insert_check_key(struct btree *b, struct btree_op *op, |
2075 | struct bkey *check_key) | |
2076 | { | |
2077 | int ret = -EINTR; | |
2078 | uint64_t btree_ptr = b->key.ptr[0]; | |
2079 | unsigned long seq = b->seq; | |
2080 | struct keylist insert; | |
2081 | bool upgrade = op->lock == -1; | |
2082 | ||
2083 | bch_keylist_init(&insert); | |
2084 | ||
2085 | if (upgrade) { | |
2086 | rw_unlock(false, b); | |
2087 | rw_lock(true, b, b->level); | |
2088 | ||
2089 | if (b->key.ptr[0] != btree_ptr || | |
2090 | b->seq != seq + 1) | |
2091 | goto out; | |
2092 | } | |
2093 | ||
2094 | SET_KEY_PTRS(check_key, 1); | |
2095 | get_random_bytes(&check_key->ptr[0], sizeof(uint64_t)); | |
2096 | ||
2097 | SET_PTR_DEV(check_key, 0, PTR_CHECK_DEV); | |
2098 | ||
2099 | bch_keylist_add(&insert, check_key); | |
2100 | ||
2101 | BUG_ON(op->type != BTREE_INSERT); | |
2102 | ||
2103 | ret = bch_btree_insert_node(b, op, &insert); | |
2104 | ||
2105 | BUG_ON(!ret && !bch_keylist_empty(&insert)); | |
2106 | out: | |
2107 | if (upgrade) | |
2108 | downgrade_write(&b->lock); | |
2109 | return ret; | |
2110 | } | |
2111 | ||
26c949f8 KO |
2112 | static int bch_btree_insert_recurse(struct btree *b, struct btree_op *op) |
2113 | { | |
403b6cde KO |
2114 | if (bch_keylist_empty(&op->keys)) |
2115 | return 0; | |
2116 | ||
26c949f8 KO |
2117 | if (b->level) { |
2118 | struct bkey *insert = op->keys.bottom; | |
2119 | struct bkey *k = bch_next_recurse_key(b, &START_KEY(insert)); | |
cafe5635 | 2120 | |
26c949f8 KO |
2121 | if (!k) { |
2122 | btree_bug(b, "no key to recurse on at level %i/%i", | |
2123 | b->level, b->c->root->level); | |
cafe5635 | 2124 | |
26c949f8 KO |
2125 | op->keys.top = op->keys.bottom; |
2126 | return -EIO; | |
57943511 | 2127 | } |
cafe5635 | 2128 | |
26c949f8 KO |
2129 | return btree(insert_recurse, k, b, op); |
2130 | } else { | |
2131 | return bch_btree_insert_node(b, op, &op->keys); | |
2132 | } | |
cafe5635 KO |
2133 | } |
2134 | ||
2135 | int bch_btree_insert(struct btree_op *op, struct cache_set *c) | |
2136 | { | |
2137 | int ret = 0; | |
cafe5635 KO |
2138 | |
2139 | /* | |
2140 | * Don't want to block with the btree locked unless we have to, | |
2141 | * otherwise we get deadlocks with try_harder and between split/gc | |
2142 | */ | |
2143 | clear_closure_blocking(&op->cl); | |
2144 | ||
2145 | BUG_ON(bch_keylist_empty(&op->keys)); | |
cafe5635 | 2146 | |
403b6cde KO |
2147 | while (!bch_keylist_empty(&op->keys)) { |
2148 | op->lock = 0; | |
26c949f8 | 2149 | ret = btree_root(insert_recurse, c, op); |
cafe5635 KO |
2150 | |
2151 | if (ret == -EAGAIN) { | |
2152 | ret = 0; | |
2153 | closure_sync(&op->cl); | |
2154 | } else if (ret) { | |
2155 | struct bkey *k; | |
2156 | ||
2157 | pr_err("error %i trying to insert key for %s", | |
2158 | ret, op_type(op)); | |
2159 | ||
403b6cde | 2160 | while ((k = bch_keylist_pop(&op->keys))) |
cafe5635 KO |
2161 | bkey_put(c, k, 0); |
2162 | } | |
2163 | } | |
2164 | ||
cafe5635 KO |
2165 | if (op->journal) |
2166 | atomic_dec_bug(op->journal); | |
2167 | op->journal = NULL; | |
2168 | return ret; | |
2169 | } | |
2170 | ||
2171 | void bch_btree_set_root(struct btree *b) | |
2172 | { | |
2173 | unsigned i; | |
e49c7c37 KO |
2174 | struct closure cl; |
2175 | ||
2176 | closure_init_stack(&cl); | |
cafe5635 | 2177 | |
c37511b8 KO |
2178 | trace_bcache_btree_set_root(b); |
2179 | ||
cafe5635 KO |
2180 | BUG_ON(!b->written); |
2181 | ||
2182 | for (i = 0; i < KEY_PTRS(&b->key); i++) | |
2183 | BUG_ON(PTR_BUCKET(b->c, &b->key, i)->prio != BTREE_PRIO); | |
2184 | ||
2185 | mutex_lock(&b->c->bucket_lock); | |
2186 | list_del_init(&b->list); | |
2187 | mutex_unlock(&b->c->bucket_lock); | |
2188 | ||
2189 | b->c->root = b; | |
2190 | __bkey_put(b->c, &b->key); | |
2191 | ||
e49c7c37 KO |
2192 | bch_journal_meta(b->c, &cl); |
2193 | closure_sync(&cl); | |
cafe5635 KO |
2194 | } |
2195 | ||
2196 | /* Cache lookup */ | |
2197 | ||
2198 | static int submit_partial_cache_miss(struct btree *b, struct btree_op *op, | |
2199 | struct bkey *k) | |
2200 | { | |
2201 | struct search *s = container_of(op, struct search, op); | |
2202 | struct bio *bio = &s->bio.bio; | |
2203 | int ret = 0; | |
2204 | ||
2205 | while (!ret && | |
2206 | !op->lookup_done) { | |
2207 | unsigned sectors = INT_MAX; | |
2208 | ||
2209 | if (KEY_INODE(k) == op->inode) { | |
2210 | if (KEY_START(k) <= bio->bi_sector) | |
2211 | break; | |
2212 | ||
2213 | sectors = min_t(uint64_t, sectors, | |
2214 | KEY_START(k) - bio->bi_sector); | |
2215 | } | |
2216 | ||
2217 | ret = s->d->cache_miss(b, s, bio, sectors); | |
2218 | } | |
2219 | ||
2220 | return ret; | |
2221 | } | |
2222 | ||
2223 | /* | |
2224 | * Read from a single key, handling the initial cache miss if the key starts in | |
2225 | * the middle of the bio | |
2226 | */ | |
2227 | static int submit_partial_cache_hit(struct btree *b, struct btree_op *op, | |
2228 | struct bkey *k) | |
2229 | { | |
2230 | struct search *s = container_of(op, struct search, op); | |
2231 | struct bio *bio = &s->bio.bio; | |
2232 | unsigned ptr; | |
2233 | struct bio *n; | |
2234 | ||
2235 | int ret = submit_partial_cache_miss(b, op, k); | |
2236 | if (ret || op->lookup_done) | |
2237 | return ret; | |
2238 | ||
2239 | /* XXX: figure out best pointer - for multiple cache devices */ | |
2240 | ptr = 0; | |
2241 | ||
2242 | PTR_BUCKET(b->c, k, ptr)->prio = INITIAL_PRIO; | |
2243 | ||
2244 | while (!op->lookup_done && | |
2245 | KEY_INODE(k) == op->inode && | |
2246 | bio->bi_sector < KEY_OFFSET(k)) { | |
2247 | struct bkey *bio_key; | |
2248 | sector_t sector = PTR_OFFSET(k, ptr) + | |
2249 | (bio->bi_sector - KEY_START(k)); | |
2250 | unsigned sectors = min_t(uint64_t, INT_MAX, | |
2251 | KEY_OFFSET(k) - bio->bi_sector); | |
2252 | ||
2253 | n = bch_bio_split(bio, sectors, GFP_NOIO, s->d->bio_split); | |
cafe5635 KO |
2254 | if (n == bio) |
2255 | op->lookup_done = true; | |
2256 | ||
2257 | bio_key = &container_of(n, struct bbio, bio)->key; | |
2258 | ||
2259 | /* | |
2260 | * The bucket we're reading from might be reused while our bio | |
2261 | * is in flight, and we could then end up reading the wrong | |
2262 | * data. | |
2263 | * | |
2264 | * We guard against this by checking (in cache_read_endio()) if | |
2265 | * the pointer is stale again; if so, we treat it as an error | |
2266 | * and reread from the backing device (but we don't pass that | |
2267 | * error up anywhere). | |
2268 | */ | |
2269 | ||
2270 | bch_bkey_copy_single_ptr(bio_key, k, ptr); | |
2271 | SET_PTR_OFFSET(bio_key, 0, sector); | |
2272 | ||
2273 | n->bi_end_io = bch_cache_read_endio; | |
2274 | n->bi_private = &s->cl; | |
2275 | ||
cafe5635 KO |
2276 | __bch_submit_bbio(n, b->c); |
2277 | } | |
2278 | ||
2279 | return 0; | |
2280 | } | |
2281 | ||
2282 | int bch_btree_search_recurse(struct btree *b, struct btree_op *op) | |
2283 | { | |
2284 | struct search *s = container_of(op, struct search, op); | |
2285 | struct bio *bio = &s->bio.bio; | |
2286 | ||
2287 | int ret = 0; | |
2288 | struct bkey *k; | |
2289 | struct btree_iter iter; | |
2290 | bch_btree_iter_init(b, &iter, &KEY(op->inode, bio->bi_sector, 0)); | |
2291 | ||
cafe5635 KO |
2292 | do { |
2293 | k = bch_btree_iter_next_filter(&iter, b, bch_ptr_bad); | |
2294 | if (!k) { | |
2295 | /* | |
2296 | * b->key would be exactly what we want, except that | |
2297 | * pointers to btree nodes have nonzero size - we | |
2298 | * wouldn't go far enough | |
2299 | */ | |
2300 | ||
2301 | ret = submit_partial_cache_miss(b, op, | |
2302 | &KEY(KEY_INODE(&b->key), | |
2303 | KEY_OFFSET(&b->key), 0)); | |
2304 | break; | |
2305 | } | |
2306 | ||
2307 | ret = b->level | |
2308 | ? btree(search_recurse, k, b, op) | |
2309 | : submit_partial_cache_hit(b, op, k); | |
2310 | } while (!ret && | |
2311 | !op->lookup_done); | |
2312 | ||
2313 | return ret; | |
2314 | } | |
2315 | ||
2316 | /* Keybuf code */ | |
2317 | ||
2318 | static inline int keybuf_cmp(struct keybuf_key *l, struct keybuf_key *r) | |
2319 | { | |
2320 | /* Overlapping keys compare equal */ | |
2321 | if (bkey_cmp(&l->key, &START_KEY(&r->key)) <= 0) | |
2322 | return -1; | |
2323 | if (bkey_cmp(&START_KEY(&l->key), &r->key) >= 0) | |
2324 | return 1; | |
2325 | return 0; | |
2326 | } | |
2327 | ||
2328 | static inline int keybuf_nonoverlapping_cmp(struct keybuf_key *l, | |
2329 | struct keybuf_key *r) | |
2330 | { | |
2331 | return clamp_t(int64_t, bkey_cmp(&l->key, &r->key), -1, 1); | |
2332 | } | |
2333 | ||
2334 | static int bch_btree_refill_keybuf(struct btree *b, struct btree_op *op, | |
72c27061 KO |
2335 | struct keybuf *buf, struct bkey *end, |
2336 | keybuf_pred_fn *pred) | |
cafe5635 KO |
2337 | { |
2338 | struct btree_iter iter; | |
2339 | bch_btree_iter_init(b, &iter, &buf->last_scanned); | |
2340 | ||
2341 | while (!array_freelist_empty(&buf->freelist)) { | |
2342 | struct bkey *k = bch_btree_iter_next_filter(&iter, b, | |
2343 | bch_ptr_bad); | |
2344 | ||
2345 | if (!b->level) { | |
2346 | if (!k) { | |
2347 | buf->last_scanned = b->key; | |
2348 | break; | |
2349 | } | |
2350 | ||
2351 | buf->last_scanned = *k; | |
2352 | if (bkey_cmp(&buf->last_scanned, end) >= 0) | |
2353 | break; | |
2354 | ||
72c27061 | 2355 | if (pred(buf, k)) { |
cafe5635 KO |
2356 | struct keybuf_key *w; |
2357 | ||
cafe5635 KO |
2358 | spin_lock(&buf->lock); |
2359 | ||
2360 | w = array_alloc(&buf->freelist); | |
2361 | ||
2362 | w->private = NULL; | |
2363 | bkey_copy(&w->key, k); | |
2364 | ||
2365 | if (RB_INSERT(&buf->keys, w, node, keybuf_cmp)) | |
2366 | array_free(&buf->freelist, w); | |
2367 | ||
2368 | spin_unlock(&buf->lock); | |
2369 | } | |
2370 | } else { | |
2371 | if (!k) | |
2372 | break; | |
2373 | ||
72c27061 | 2374 | btree(refill_keybuf, k, b, op, buf, end, pred); |
cafe5635 KO |
2375 | /* |
2376 | * Might get an error here, but can't really do anything | |
2377 | * and it'll get logged elsewhere. Just read what we | |
2378 | * can. | |
2379 | */ | |
2380 | ||
2381 | if (bkey_cmp(&buf->last_scanned, end) >= 0) | |
2382 | break; | |
2383 | ||
2384 | cond_resched(); | |
2385 | } | |
2386 | } | |
2387 | ||
2388 | return 0; | |
2389 | } | |
2390 | ||
2391 | void bch_refill_keybuf(struct cache_set *c, struct keybuf *buf, | |
72c27061 | 2392 | struct bkey *end, keybuf_pred_fn *pred) |
cafe5635 KO |
2393 | { |
2394 | struct bkey start = buf->last_scanned; | |
2395 | struct btree_op op; | |
2396 | bch_btree_op_init_stack(&op); | |
2397 | ||
2398 | cond_resched(); | |
2399 | ||
72c27061 | 2400 | btree_root(refill_keybuf, c, &op, buf, end, pred); |
cafe5635 KO |
2401 | closure_sync(&op.cl); |
2402 | ||
2403 | pr_debug("found %s keys from %llu:%llu to %llu:%llu", | |
2404 | RB_EMPTY_ROOT(&buf->keys) ? "no" : | |
2405 | array_freelist_empty(&buf->freelist) ? "some" : "a few", | |
2406 | KEY_INODE(&start), KEY_OFFSET(&start), | |
2407 | KEY_INODE(&buf->last_scanned), KEY_OFFSET(&buf->last_scanned)); | |
2408 | ||
2409 | spin_lock(&buf->lock); | |
2410 | ||
2411 | if (!RB_EMPTY_ROOT(&buf->keys)) { | |
2412 | struct keybuf_key *w; | |
2413 | w = RB_FIRST(&buf->keys, struct keybuf_key, node); | |
2414 | buf->start = START_KEY(&w->key); | |
2415 | ||
2416 | w = RB_LAST(&buf->keys, struct keybuf_key, node); | |
2417 | buf->end = w->key; | |
2418 | } else { | |
2419 | buf->start = MAX_KEY; | |
2420 | buf->end = MAX_KEY; | |
2421 | } | |
2422 | ||
2423 | spin_unlock(&buf->lock); | |
2424 | } | |
2425 | ||
2426 | static void __bch_keybuf_del(struct keybuf *buf, struct keybuf_key *w) | |
2427 | { | |
2428 | rb_erase(&w->node, &buf->keys); | |
2429 | array_free(&buf->freelist, w); | |
2430 | } | |
2431 | ||
2432 | void bch_keybuf_del(struct keybuf *buf, struct keybuf_key *w) | |
2433 | { | |
2434 | spin_lock(&buf->lock); | |
2435 | __bch_keybuf_del(buf, w); | |
2436 | spin_unlock(&buf->lock); | |
2437 | } | |
2438 | ||
2439 | bool bch_keybuf_check_overlapping(struct keybuf *buf, struct bkey *start, | |
2440 | struct bkey *end) | |
2441 | { | |
2442 | bool ret = false; | |
2443 | struct keybuf_key *p, *w, s; | |
2444 | s.key = *start; | |
2445 | ||
2446 | if (bkey_cmp(end, &buf->start) <= 0 || | |
2447 | bkey_cmp(start, &buf->end) >= 0) | |
2448 | return false; | |
2449 | ||
2450 | spin_lock(&buf->lock); | |
2451 | w = RB_GREATER(&buf->keys, s, node, keybuf_nonoverlapping_cmp); | |
2452 | ||
2453 | while (w && bkey_cmp(&START_KEY(&w->key), end) < 0) { | |
2454 | p = w; | |
2455 | w = RB_NEXT(w, node); | |
2456 | ||
2457 | if (p->private) | |
2458 | ret = true; | |
2459 | else | |
2460 | __bch_keybuf_del(buf, p); | |
2461 | } | |
2462 | ||
2463 | spin_unlock(&buf->lock); | |
2464 | return ret; | |
2465 | } | |
2466 | ||
2467 | struct keybuf_key *bch_keybuf_next(struct keybuf *buf) | |
2468 | { | |
2469 | struct keybuf_key *w; | |
2470 | spin_lock(&buf->lock); | |
2471 | ||
2472 | w = RB_FIRST(&buf->keys, struct keybuf_key, node); | |
2473 | ||
2474 | while (w && w->private) | |
2475 | w = RB_NEXT(w, node); | |
2476 | ||
2477 | if (w) | |
2478 | w->private = ERR_PTR(-EINTR); | |
2479 | ||
2480 | spin_unlock(&buf->lock); | |
2481 | return w; | |
2482 | } | |
2483 | ||
2484 | struct keybuf_key *bch_keybuf_next_rescan(struct cache_set *c, | |
2485 | struct keybuf *buf, | |
72c27061 KO |
2486 | struct bkey *end, |
2487 | keybuf_pred_fn *pred) | |
cafe5635 KO |
2488 | { |
2489 | struct keybuf_key *ret; | |
2490 | ||
2491 | while (1) { | |
2492 | ret = bch_keybuf_next(buf); | |
2493 | if (ret) | |
2494 | break; | |
2495 | ||
2496 | if (bkey_cmp(&buf->last_scanned, end) >= 0) { | |
2497 | pr_debug("scan finished"); | |
2498 | break; | |
2499 | } | |
2500 | ||
72c27061 | 2501 | bch_refill_keybuf(c, buf, end, pred); |
cafe5635 KO |
2502 | } |
2503 | ||
2504 | return ret; | |
2505 | } | |
2506 | ||
72c27061 | 2507 | void bch_keybuf_init(struct keybuf *buf) |
cafe5635 | 2508 | { |
cafe5635 KO |
2509 | buf->last_scanned = MAX_KEY; |
2510 | buf->keys = RB_ROOT; | |
2511 | ||
2512 | spin_lock_init(&buf->lock); | |
2513 | array_allocator_init(&buf->freelist); | |
2514 | } | |
2515 | ||
2516 | void bch_btree_exit(void) | |
2517 | { | |
2518 | if (btree_io_wq) | |
2519 | destroy_workqueue(btree_io_wq); | |
2520 | if (bch_gc_wq) | |
2521 | destroy_workqueue(bch_gc_wq); | |
2522 | } | |
2523 | ||
2524 | int __init bch_btree_init(void) | |
2525 | { | |
2526 | if (!(bch_gc_wq = create_singlethread_workqueue("bch_btree_gc")) || | |
2527 | !(btree_io_wq = create_singlethread_workqueue("bch_btree_io"))) | |
2528 | return -ENOMEM; | |
2529 | ||
2530 | return 0; | |
2531 | } |