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" | |
279afbad | 26 | #include "writeback.h" |
cafe5635 KO |
27 | |
28 | #include <linux/slab.h> | |
29 | #include <linux/bitops.h> | |
72a44517 | 30 | #include <linux/freezer.h> |
cafe5635 | 31 | #include <linux/hash.h> |
72a44517 | 32 | #include <linux/kthread.h> |
cd953ed0 | 33 | #include <linux/prefetch.h> |
cafe5635 KO |
34 | #include <linux/random.h> |
35 | #include <linux/rcupdate.h> | |
36 | #include <trace/events/bcache.h> | |
37 | ||
38 | /* | |
39 | * Todo: | |
40 | * register_bcache: Return errors out to userspace correctly | |
41 | * | |
42 | * Writeback: don't undirty key until after a cache flush | |
43 | * | |
44 | * Create an iterator for key pointers | |
45 | * | |
46 | * On btree write error, mark bucket such that it won't be freed from the cache | |
47 | * | |
48 | * Journalling: | |
49 | * Check for bad keys in replay | |
50 | * Propagate barriers | |
51 | * Refcount journal entries in journal_replay | |
52 | * | |
53 | * Garbage collection: | |
54 | * Finish incremental gc | |
55 | * Gc should free old UUIDs, data for invalid UUIDs | |
56 | * | |
57 | * Provide a way to list backing device UUIDs we have data cached for, and | |
58 | * probably how long it's been since we've seen them, and a way to invalidate | |
59 | * dirty data for devices that will never be attached again | |
60 | * | |
61 | * Keep 1 min/5 min/15 min statistics of how busy a block device has been, so | |
62 | * that based on that and how much dirty data we have we can keep writeback | |
63 | * from being starved | |
64 | * | |
65 | * Add a tracepoint or somesuch to watch for writeback starvation | |
66 | * | |
67 | * When btree depth > 1 and splitting an interior node, we have to make sure | |
68 | * alloc_bucket() cannot fail. This should be true but is not completely | |
69 | * obvious. | |
70 | * | |
71 | * Make sure all allocations get charged to the root cgroup | |
72 | * | |
73 | * Plugging? | |
74 | * | |
75 | * If data write is less than hard sector size of ssd, round up offset in open | |
76 | * bucket to the next whole sector | |
77 | * | |
78 | * Also lookup by cgroup in get_open_bucket() | |
79 | * | |
80 | * Superblock needs to be fleshed out for multiple cache devices | |
81 | * | |
82 | * Add a sysfs tunable for the number of writeback IOs in flight | |
83 | * | |
84 | * Add a sysfs tunable for the number of open data buckets | |
85 | * | |
86 | * IO tracking: Can we track when one process is doing io on behalf of another? | |
87 | * IO tracking: Don't use just an average, weigh more recent stuff higher | |
88 | * | |
89 | * Test module load/unload | |
90 | */ | |
91 | ||
df8e8970 KO |
92 | enum { |
93 | BTREE_INSERT_STATUS_INSERT, | |
94 | BTREE_INSERT_STATUS_BACK_MERGE, | |
95 | BTREE_INSERT_STATUS_OVERWROTE, | |
96 | BTREE_INSERT_STATUS_FRONT_MERGE, | |
97 | }; | |
98 | ||
cafe5635 KO |
99 | #define MAX_NEED_GC 64 |
100 | #define MAX_SAVE_PRIO 72 | |
101 | ||
102 | #define PTR_DIRTY_BIT (((uint64_t) 1 << 36)) | |
103 | ||
104 | #define PTR_HASH(c, k) \ | |
105 | (((k)->ptr[0] >> c->bucket_bits) | PTR_GEN(k, 0)) | |
106 | ||
cafe5635 KO |
107 | static struct workqueue_struct *btree_io_wq; |
108 | ||
df8e8970 KO |
109 | static inline bool should_split(struct btree *b) |
110 | { | |
111 | struct bset *i = write_block(b); | |
112 | return b->written >= btree_blocks(b) || | |
113 | (b->written + __set_blocks(i, i->keys + 15, b->c) | |
114 | > btree_blocks(b)); | |
115 | } | |
116 | ||
117 | #define insert_lock(s, b) ((b)->level <= (s)->lock) | |
118 | ||
119 | /* | |
120 | * These macros are for recursing down the btree - they handle the details of | |
121 | * locking and looking up nodes in the cache for you. They're best treated as | |
122 | * mere syntax when reading code that uses them. | |
123 | * | |
124 | * op->lock determines whether we take a read or a write lock at a given depth. | |
125 | * If you've got a read lock and find that you need a write lock (i.e. you're | |
126 | * going to have to split), set op->lock and return -EINTR; btree_root() will | |
127 | * call you again and you'll have the correct lock. | |
128 | */ | |
129 | ||
130 | /** | |
131 | * btree - recurse down the btree on a specified key | |
132 | * @fn: function to call, which will be passed the child node | |
133 | * @key: key to recurse on | |
134 | * @b: parent btree node | |
135 | * @op: pointer to struct btree_op | |
136 | */ | |
137 | #define btree(fn, key, b, op, ...) \ | |
138 | ({ \ | |
139 | int _r, l = (b)->level - 1; \ | |
140 | bool _w = l <= (op)->lock; \ | |
141 | struct btree *_child = bch_btree_node_get((b)->c, key, l, _w); \ | |
142 | if (!IS_ERR(_child)) { \ | |
143 | _child->parent = (b); \ | |
144 | _r = bch_btree_ ## fn(_child, op, ##__VA_ARGS__); \ | |
145 | rw_unlock(_w, _child); \ | |
146 | } else \ | |
147 | _r = PTR_ERR(_child); \ | |
148 | _r; \ | |
149 | }) | |
150 | ||
151 | /** | |
152 | * btree_root - call a function on the root of the btree | |
153 | * @fn: function to call, which will be passed the child node | |
154 | * @c: cache set | |
155 | * @op: pointer to struct btree_op | |
156 | */ | |
157 | #define btree_root(fn, c, op, ...) \ | |
158 | ({ \ | |
159 | int _r = -EINTR; \ | |
160 | do { \ | |
161 | struct btree *_b = (c)->root; \ | |
162 | bool _w = insert_lock(op, _b); \ | |
163 | rw_lock(_w, _b, _b->level); \ | |
164 | if (_b == (c)->root && \ | |
165 | _w == insert_lock(op, _b)) { \ | |
166 | _b->parent = NULL; \ | |
167 | _r = bch_btree_ ## fn(_b, op, ##__VA_ARGS__); \ | |
168 | } \ | |
169 | rw_unlock(_w, _b); \ | |
170 | bch_cannibalize_unlock(c); \ | |
171 | if (_r == -ENOSPC) { \ | |
172 | wait_event((c)->try_wait, \ | |
173 | !(c)->try_harder); \ | |
174 | _r = -EINTR; \ | |
175 | } \ | |
176 | } while (_r == -EINTR); \ | |
177 | \ | |
178 | _r; \ | |
179 | }) | |
180 | ||
cafe5635 KO |
181 | /* Btree key manipulation */ |
182 | ||
e7c590eb KO |
183 | void __bkey_put(struct cache_set *c, struct bkey *k) |
184 | { | |
185 | unsigned i; | |
186 | ||
187 | for (i = 0; i < KEY_PTRS(k); i++) | |
188 | if (ptr_available(c, k, i)) | |
189 | atomic_dec_bug(&PTR_BUCKET(c, k, i)->pin); | |
190 | } | |
191 | ||
cafe5635 KO |
192 | static void bkey_put(struct cache_set *c, struct bkey *k, int level) |
193 | { | |
194 | if ((level && KEY_OFFSET(k)) || !level) | |
195 | __bkey_put(c, k); | |
196 | } | |
197 | ||
198 | /* Btree IO */ | |
199 | ||
200 | static uint64_t btree_csum_set(struct btree *b, struct bset *i) | |
201 | { | |
202 | uint64_t crc = b->key.ptr[0]; | |
203 | void *data = (void *) i + 8, *end = end(i); | |
204 | ||
169ef1cf | 205 | crc = bch_crc64_update(crc, data, end - data); |
c19ed23a | 206 | return crc ^ 0xffffffffffffffffULL; |
cafe5635 KO |
207 | } |
208 | ||
f3059a54 | 209 | static void bch_btree_node_read_done(struct btree *b) |
cafe5635 | 210 | { |
cafe5635 | 211 | const char *err = "bad btree header"; |
57943511 KO |
212 | struct bset *i = b->sets[0].data; |
213 | struct btree_iter *iter; | |
cafe5635 | 214 | |
57943511 KO |
215 | iter = mempool_alloc(b->c->fill_iter, GFP_NOWAIT); |
216 | iter->size = b->c->sb.bucket_size / b->c->sb.block_size; | |
cafe5635 KO |
217 | iter->used = 0; |
218 | ||
57943511 | 219 | if (!i->seq) |
cafe5635 KO |
220 | goto err; |
221 | ||
222 | for (; | |
223 | b->written < btree_blocks(b) && i->seq == b->sets[0].data->seq; | |
224 | i = write_block(b)) { | |
225 | err = "unsupported bset version"; | |
226 | if (i->version > BCACHE_BSET_VERSION) | |
227 | goto err; | |
228 | ||
229 | err = "bad btree header"; | |
230 | if (b->written + set_blocks(i, b->c) > btree_blocks(b)) | |
231 | goto err; | |
232 | ||
233 | err = "bad magic"; | |
234 | if (i->magic != bset_magic(b->c)) | |
235 | goto err; | |
236 | ||
237 | err = "bad checksum"; | |
238 | switch (i->version) { | |
239 | case 0: | |
240 | if (i->csum != csum_set(i)) | |
241 | goto err; | |
242 | break; | |
243 | case BCACHE_BSET_VERSION: | |
244 | if (i->csum != btree_csum_set(b, i)) | |
245 | goto err; | |
246 | break; | |
247 | } | |
248 | ||
249 | err = "empty set"; | |
250 | if (i != b->sets[0].data && !i->keys) | |
251 | goto err; | |
252 | ||
253 | bch_btree_iter_push(iter, i->start, end(i)); | |
254 | ||
255 | b->written += set_blocks(i, b->c); | |
256 | } | |
257 | ||
258 | err = "corrupted btree"; | |
259 | for (i = write_block(b); | |
260 | index(i, b) < btree_blocks(b); | |
261 | i = ((void *) i) + block_bytes(b->c)) | |
262 | if (i->seq == b->sets[0].data->seq) | |
263 | goto err; | |
264 | ||
265 | bch_btree_sort_and_fix_extents(b, iter); | |
266 | ||
267 | i = b->sets[0].data; | |
268 | err = "short btree key"; | |
269 | if (b->sets[0].size && | |
270 | bkey_cmp(&b->key, &b->sets[0].end) < 0) | |
271 | goto err; | |
272 | ||
273 | if (b->written < btree_blocks(b)) | |
274 | bch_bset_init_next(b); | |
275 | out: | |
57943511 KO |
276 | mempool_free(iter, b->c->fill_iter); |
277 | return; | |
cafe5635 KO |
278 | err: |
279 | set_btree_node_io_error(b); | |
07e86ccb | 280 | bch_cache_set_error(b->c, "%s at bucket %zu, block %zu, %u keys", |
cafe5635 KO |
281 | err, PTR_BUCKET_NR(b->c, &b->key, 0), |
282 | index(i, b), i->keys); | |
283 | goto out; | |
284 | } | |
285 | ||
57943511 | 286 | static void btree_node_read_endio(struct bio *bio, int error) |
cafe5635 | 287 | { |
57943511 KO |
288 | struct closure *cl = bio->bi_private; |
289 | closure_put(cl); | |
290 | } | |
cafe5635 | 291 | |
57943511 KO |
292 | void bch_btree_node_read(struct btree *b) |
293 | { | |
294 | uint64_t start_time = local_clock(); | |
295 | struct closure cl; | |
296 | struct bio *bio; | |
cafe5635 | 297 | |
c37511b8 KO |
298 | trace_bcache_btree_read(b); |
299 | ||
57943511 | 300 | closure_init_stack(&cl); |
cafe5635 | 301 | |
57943511 KO |
302 | bio = bch_bbio_alloc(b->c); |
303 | bio->bi_rw = REQ_META|READ_SYNC; | |
304 | bio->bi_size = KEY_SIZE(&b->key) << 9; | |
305 | bio->bi_end_io = btree_node_read_endio; | |
306 | bio->bi_private = &cl; | |
cafe5635 | 307 | |
57943511 | 308 | bch_bio_map(bio, b->sets[0].data); |
cafe5635 | 309 | |
57943511 KO |
310 | bch_submit_bbio(bio, b->c, &b->key, 0); |
311 | closure_sync(&cl); | |
cafe5635 | 312 | |
57943511 KO |
313 | if (!test_bit(BIO_UPTODATE, &bio->bi_flags)) |
314 | set_btree_node_io_error(b); | |
315 | ||
316 | bch_bbio_free(bio, b->c); | |
317 | ||
318 | if (btree_node_io_error(b)) | |
319 | goto err; | |
320 | ||
321 | bch_btree_node_read_done(b); | |
322 | ||
323 | spin_lock(&b->c->btree_read_time_lock); | |
324 | bch_time_stats_update(&b->c->btree_read_time, start_time); | |
325 | spin_unlock(&b->c->btree_read_time_lock); | |
326 | ||
327 | return; | |
328 | err: | |
61cbd250 | 329 | bch_cache_set_error(b->c, "io error reading bucket %zu", |
57943511 | 330 | PTR_BUCKET_NR(b->c, &b->key, 0)); |
cafe5635 KO |
331 | } |
332 | ||
333 | static void btree_complete_write(struct btree *b, struct btree_write *w) | |
334 | { | |
335 | if (w->prio_blocked && | |
336 | !atomic_sub_return(w->prio_blocked, &b->c->prio_blocked)) | |
119ba0f8 | 337 | wake_up_allocators(b->c); |
cafe5635 KO |
338 | |
339 | if (w->journal) { | |
340 | atomic_dec_bug(w->journal); | |
341 | __closure_wake_up(&b->c->journal.wait); | |
342 | } | |
343 | ||
cafe5635 KO |
344 | w->prio_blocked = 0; |
345 | w->journal = NULL; | |
cafe5635 KO |
346 | } |
347 | ||
57943511 | 348 | static void __btree_node_write_done(struct closure *cl) |
cafe5635 KO |
349 | { |
350 | struct btree *b = container_of(cl, struct btree, io.cl); | |
351 | struct btree_write *w = btree_prev_write(b); | |
352 | ||
353 | bch_bbio_free(b->bio, b->c); | |
354 | b->bio = NULL; | |
355 | btree_complete_write(b, w); | |
356 | ||
357 | if (btree_node_dirty(b)) | |
358 | queue_delayed_work(btree_io_wq, &b->work, | |
359 | msecs_to_jiffies(30000)); | |
360 | ||
361 | closure_return(cl); | |
362 | } | |
363 | ||
57943511 | 364 | static void btree_node_write_done(struct closure *cl) |
cafe5635 KO |
365 | { |
366 | struct btree *b = container_of(cl, struct btree, io.cl); | |
367 | struct bio_vec *bv; | |
368 | int n; | |
369 | ||
370 | __bio_for_each_segment(bv, b->bio, n, 0) | |
371 | __free_page(bv->bv_page); | |
372 | ||
57943511 | 373 | __btree_node_write_done(cl); |
cafe5635 KO |
374 | } |
375 | ||
57943511 KO |
376 | static void btree_node_write_endio(struct bio *bio, int error) |
377 | { | |
378 | struct closure *cl = bio->bi_private; | |
379 | struct btree *b = container_of(cl, struct btree, io.cl); | |
380 | ||
381 | if (error) | |
382 | set_btree_node_io_error(b); | |
383 | ||
384 | bch_bbio_count_io_errors(b->c, bio, error, "writing btree"); | |
385 | closure_put(cl); | |
386 | } | |
387 | ||
388 | static void do_btree_node_write(struct btree *b) | |
cafe5635 KO |
389 | { |
390 | struct closure *cl = &b->io.cl; | |
391 | struct bset *i = b->sets[b->nsets].data; | |
392 | BKEY_PADDED(key) k; | |
393 | ||
394 | i->version = BCACHE_BSET_VERSION; | |
395 | i->csum = btree_csum_set(b, i); | |
396 | ||
57943511 KO |
397 | BUG_ON(b->bio); |
398 | b->bio = bch_bbio_alloc(b->c); | |
399 | ||
400 | b->bio->bi_end_io = btree_node_write_endio; | |
faadf0c9 | 401 | b->bio->bi_private = cl; |
e49c7c37 KO |
402 | b->bio->bi_rw = REQ_META|WRITE_SYNC|REQ_FUA; |
403 | b->bio->bi_size = set_blocks(i, b->c) * block_bytes(b->c); | |
169ef1cf | 404 | bch_bio_map(b->bio, i); |
cafe5635 | 405 | |
e49c7c37 KO |
406 | /* |
407 | * If we're appending to a leaf node, we don't technically need FUA - | |
408 | * this write just needs to be persisted before the next journal write, | |
409 | * which will be marked FLUSH|FUA. | |
410 | * | |
411 | * Similarly if we're writing a new btree root - the pointer is going to | |
412 | * be in the next journal entry. | |
413 | * | |
414 | * But if we're writing a new btree node (that isn't a root) or | |
415 | * appending to a non leaf btree node, we need either FUA or a flush | |
416 | * when we write the parent with the new pointer. FUA is cheaper than a | |
417 | * flush, and writes appending to leaf nodes aren't blocking anything so | |
418 | * just make all btree node writes FUA to keep things sane. | |
419 | */ | |
420 | ||
cafe5635 KO |
421 | bkey_copy(&k.key, &b->key); |
422 | SET_PTR_OFFSET(&k.key, 0, PTR_OFFSET(&k.key, 0) + bset_offset(b, i)); | |
423 | ||
8e51e414 | 424 | if (!bio_alloc_pages(b->bio, GFP_NOIO)) { |
cafe5635 KO |
425 | int j; |
426 | struct bio_vec *bv; | |
427 | void *base = (void *) ((unsigned long) i & ~(PAGE_SIZE - 1)); | |
428 | ||
429 | bio_for_each_segment(bv, b->bio, j) | |
430 | memcpy(page_address(bv->bv_page), | |
431 | base + j * PAGE_SIZE, PAGE_SIZE); | |
432 | ||
cafe5635 KO |
433 | bch_submit_bbio(b->bio, b->c, &k.key, 0); |
434 | ||
57943511 | 435 | continue_at(cl, btree_node_write_done, NULL); |
cafe5635 KO |
436 | } else { |
437 | b->bio->bi_vcnt = 0; | |
169ef1cf | 438 | bch_bio_map(b->bio, i); |
cafe5635 | 439 | |
cafe5635 KO |
440 | bch_submit_bbio(b->bio, b->c, &k.key, 0); |
441 | ||
442 | closure_sync(cl); | |
57943511 | 443 | __btree_node_write_done(cl); |
cafe5635 KO |
444 | } |
445 | } | |
446 | ||
57943511 | 447 | void bch_btree_node_write(struct btree *b, struct closure *parent) |
cafe5635 KO |
448 | { |
449 | struct bset *i = b->sets[b->nsets].data; | |
450 | ||
c37511b8 KO |
451 | trace_bcache_btree_write(b); |
452 | ||
cafe5635 | 453 | BUG_ON(current->bio_list); |
57943511 KO |
454 | BUG_ON(b->written >= btree_blocks(b)); |
455 | BUG_ON(b->written && !i->keys); | |
456 | BUG_ON(b->sets->data->seq != i->seq); | |
c37511b8 | 457 | bch_check_key_order(b, i); |
cafe5635 | 458 | |
cafe5635 KO |
459 | cancel_delayed_work(&b->work); |
460 | ||
57943511 KO |
461 | /* If caller isn't waiting for write, parent refcount is cache set */ |
462 | closure_lock(&b->io, parent ?: &b->c->cl); | |
463 | ||
cafe5635 KO |
464 | clear_bit(BTREE_NODE_dirty, &b->flags); |
465 | change_bit(BTREE_NODE_write_idx, &b->flags); | |
466 | ||
57943511 | 467 | do_btree_node_write(b); |
cafe5635 | 468 | |
cafe5635 KO |
469 | b->written += set_blocks(i, b->c); |
470 | atomic_long_add(set_blocks(i, b->c) * b->c->sb.block_size, | |
471 | &PTR_CACHE(b->c, &b->key, 0)->btree_sectors_written); | |
472 | ||
473 | bch_btree_sort_lazy(b); | |
474 | ||
475 | if (b->written < btree_blocks(b)) | |
476 | bch_bset_init_next(b); | |
477 | } | |
478 | ||
57943511 | 479 | static void btree_node_write_work(struct work_struct *w) |
cafe5635 KO |
480 | { |
481 | struct btree *b = container_of(to_delayed_work(w), struct btree, work); | |
482 | ||
57943511 | 483 | rw_lock(true, b, b->level); |
cafe5635 KO |
484 | |
485 | if (btree_node_dirty(b)) | |
57943511 KO |
486 | bch_btree_node_write(b, NULL); |
487 | rw_unlock(true, b); | |
cafe5635 KO |
488 | } |
489 | ||
c18536a7 | 490 | static void bch_btree_leaf_dirty(struct btree *b, atomic_t *journal_ref) |
cafe5635 KO |
491 | { |
492 | struct bset *i = b->sets[b->nsets].data; | |
493 | struct btree_write *w = btree_current_write(b); | |
494 | ||
57943511 KO |
495 | BUG_ON(!b->written); |
496 | BUG_ON(!i->keys); | |
cafe5635 | 497 | |
57943511 KO |
498 | if (!btree_node_dirty(b)) |
499 | queue_delayed_work(btree_io_wq, &b->work, 30 * HZ); | |
cafe5635 | 500 | |
57943511 | 501 | set_btree_node_dirty(b); |
cafe5635 | 502 | |
c18536a7 | 503 | if (journal_ref) { |
cafe5635 | 504 | if (w->journal && |
c18536a7 | 505 | journal_pin_cmp(b->c, w->journal, journal_ref)) { |
cafe5635 KO |
506 | atomic_dec_bug(w->journal); |
507 | w->journal = NULL; | |
508 | } | |
509 | ||
510 | if (!w->journal) { | |
c18536a7 | 511 | w->journal = journal_ref; |
cafe5635 KO |
512 | atomic_inc(w->journal); |
513 | } | |
514 | } | |
515 | ||
cafe5635 | 516 | /* Force write if set is too big */ |
57943511 KO |
517 | if (set_bytes(i) > PAGE_SIZE - 48 && |
518 | !current->bio_list) | |
519 | bch_btree_node_write(b, NULL); | |
cafe5635 KO |
520 | } |
521 | ||
522 | /* | |
523 | * Btree in memory cache - allocation/freeing | |
524 | * mca -> memory cache | |
525 | */ | |
526 | ||
527 | static void mca_reinit(struct btree *b) | |
528 | { | |
529 | unsigned i; | |
530 | ||
531 | b->flags = 0; | |
532 | b->written = 0; | |
533 | b->nsets = 0; | |
534 | ||
535 | for (i = 0; i < MAX_BSETS; i++) | |
536 | b->sets[i].size = 0; | |
537 | /* | |
538 | * Second loop starts at 1 because b->sets[0]->data is the memory we | |
539 | * allocated | |
540 | */ | |
541 | for (i = 1; i < MAX_BSETS; i++) | |
542 | b->sets[i].data = NULL; | |
543 | } | |
544 | ||
545 | #define mca_reserve(c) (((c->root && c->root->level) \ | |
546 | ? c->root->level : 1) * 8 + 16) | |
547 | #define mca_can_free(c) \ | |
548 | max_t(int, 0, c->bucket_cache_used - mca_reserve(c)) | |
549 | ||
550 | static void mca_data_free(struct btree *b) | |
551 | { | |
552 | struct bset_tree *t = b->sets; | |
553 | BUG_ON(!closure_is_unlocked(&b->io.cl)); | |
554 | ||
555 | if (bset_prev_bytes(b) < PAGE_SIZE) | |
556 | kfree(t->prev); | |
557 | else | |
558 | free_pages((unsigned long) t->prev, | |
559 | get_order(bset_prev_bytes(b))); | |
560 | ||
561 | if (bset_tree_bytes(b) < PAGE_SIZE) | |
562 | kfree(t->tree); | |
563 | else | |
564 | free_pages((unsigned long) t->tree, | |
565 | get_order(bset_tree_bytes(b))); | |
566 | ||
567 | free_pages((unsigned long) t->data, b->page_order); | |
568 | ||
569 | t->prev = NULL; | |
570 | t->tree = NULL; | |
571 | t->data = NULL; | |
572 | list_move(&b->list, &b->c->btree_cache_freed); | |
573 | b->c->bucket_cache_used--; | |
574 | } | |
575 | ||
576 | static void mca_bucket_free(struct btree *b) | |
577 | { | |
578 | BUG_ON(btree_node_dirty(b)); | |
579 | ||
580 | b->key.ptr[0] = 0; | |
581 | hlist_del_init_rcu(&b->hash); | |
582 | list_move(&b->list, &b->c->btree_cache_freeable); | |
583 | } | |
584 | ||
585 | static unsigned btree_order(struct bkey *k) | |
586 | { | |
587 | return ilog2(KEY_SIZE(k) / PAGE_SECTORS ?: 1); | |
588 | } | |
589 | ||
590 | static void mca_data_alloc(struct btree *b, struct bkey *k, gfp_t gfp) | |
591 | { | |
592 | struct bset_tree *t = b->sets; | |
593 | BUG_ON(t->data); | |
594 | ||
595 | b->page_order = max_t(unsigned, | |
596 | ilog2(b->c->btree_pages), | |
597 | btree_order(k)); | |
598 | ||
599 | t->data = (void *) __get_free_pages(gfp, b->page_order); | |
600 | if (!t->data) | |
601 | goto err; | |
602 | ||
603 | t->tree = bset_tree_bytes(b) < PAGE_SIZE | |
604 | ? kmalloc(bset_tree_bytes(b), gfp) | |
605 | : (void *) __get_free_pages(gfp, get_order(bset_tree_bytes(b))); | |
606 | if (!t->tree) | |
607 | goto err; | |
608 | ||
609 | t->prev = bset_prev_bytes(b) < PAGE_SIZE | |
610 | ? kmalloc(bset_prev_bytes(b), gfp) | |
611 | : (void *) __get_free_pages(gfp, get_order(bset_prev_bytes(b))); | |
612 | if (!t->prev) | |
613 | goto err; | |
614 | ||
615 | list_move(&b->list, &b->c->btree_cache); | |
616 | b->c->bucket_cache_used++; | |
617 | return; | |
618 | err: | |
619 | mca_data_free(b); | |
620 | } | |
621 | ||
622 | static struct btree *mca_bucket_alloc(struct cache_set *c, | |
623 | struct bkey *k, gfp_t gfp) | |
624 | { | |
625 | struct btree *b = kzalloc(sizeof(struct btree), gfp); | |
626 | if (!b) | |
627 | return NULL; | |
628 | ||
629 | init_rwsem(&b->lock); | |
630 | lockdep_set_novalidate_class(&b->lock); | |
631 | INIT_LIST_HEAD(&b->list); | |
57943511 | 632 | INIT_DELAYED_WORK(&b->work, btree_node_write_work); |
cafe5635 KO |
633 | b->c = c; |
634 | closure_init_unlocked(&b->io); | |
635 | ||
636 | mca_data_alloc(b, k, gfp); | |
637 | return b; | |
638 | } | |
639 | ||
e8e1d468 | 640 | static int mca_reap(struct btree *b, unsigned min_order, bool flush) |
cafe5635 | 641 | { |
e8e1d468 KO |
642 | struct closure cl; |
643 | ||
644 | closure_init_stack(&cl); | |
cafe5635 KO |
645 | lockdep_assert_held(&b->c->bucket_lock); |
646 | ||
647 | if (!down_write_trylock(&b->lock)) | |
648 | return -ENOMEM; | |
649 | ||
e8e1d468 KO |
650 | BUG_ON(btree_node_dirty(b) && !b->sets[0].data); |
651 | ||
652 | if (b->page_order < min_order || | |
653 | (!flush && | |
654 | (btree_node_dirty(b) || | |
655 | atomic_read(&b->io.cl.remaining) != -1))) { | |
cafe5635 KO |
656 | rw_unlock(true, b); |
657 | return -ENOMEM; | |
658 | } | |
659 | ||
e8e1d468 KO |
660 | if (btree_node_dirty(b)) { |
661 | bch_btree_node_write(b, &cl); | |
662 | closure_sync(&cl); | |
cafe5635 KO |
663 | } |
664 | ||
e8e1d468 | 665 | /* wait for any in flight btree write */ |
faadf0c9 KO |
666 | closure_wait_event(&b->io.wait, &cl, |
667 | atomic_read(&b->io.cl.remaining) == -1); | |
e8e1d468 | 668 | |
cafe5635 KO |
669 | return 0; |
670 | } | |
671 | ||
7dc19d5a DC |
672 | static unsigned long bch_mca_scan(struct shrinker *shrink, |
673 | struct shrink_control *sc) | |
cafe5635 KO |
674 | { |
675 | struct cache_set *c = container_of(shrink, struct cache_set, shrink); | |
676 | struct btree *b, *t; | |
677 | unsigned long i, nr = sc->nr_to_scan; | |
7dc19d5a | 678 | unsigned long freed = 0; |
cafe5635 KO |
679 | |
680 | if (c->shrinker_disabled) | |
7dc19d5a | 681 | return SHRINK_STOP; |
cafe5635 KO |
682 | |
683 | if (c->try_harder) | |
7dc19d5a | 684 | return SHRINK_STOP; |
cafe5635 KO |
685 | |
686 | /* Return -1 if we can't do anything right now */ | |
a698e08c | 687 | if (sc->gfp_mask & __GFP_IO) |
cafe5635 KO |
688 | mutex_lock(&c->bucket_lock); |
689 | else if (!mutex_trylock(&c->bucket_lock)) | |
690 | return -1; | |
691 | ||
36c9ea98 KO |
692 | /* |
693 | * It's _really_ critical that we don't free too many btree nodes - we | |
694 | * have to always leave ourselves a reserve. The reserve is how we | |
695 | * guarantee that allocating memory for a new btree node can always | |
696 | * succeed, so that inserting keys into the btree can always succeed and | |
697 | * IO can always make forward progress: | |
698 | */ | |
cafe5635 KO |
699 | nr /= c->btree_pages; |
700 | nr = min_t(unsigned long, nr, mca_can_free(c)); | |
701 | ||
702 | i = 0; | |
703 | list_for_each_entry_safe(b, t, &c->btree_cache_freeable, list) { | |
7dc19d5a | 704 | if (freed >= nr) |
cafe5635 KO |
705 | break; |
706 | ||
707 | if (++i > 3 && | |
e8e1d468 | 708 | !mca_reap(b, 0, false)) { |
cafe5635 KO |
709 | mca_data_free(b); |
710 | rw_unlock(true, b); | |
7dc19d5a | 711 | freed++; |
cafe5635 KO |
712 | } |
713 | } | |
714 | ||
715 | /* | |
716 | * Can happen right when we first start up, before we've read in any | |
717 | * btree nodes | |
718 | */ | |
719 | if (list_empty(&c->btree_cache)) | |
720 | goto out; | |
721 | ||
7dc19d5a | 722 | for (i = 0; (nr--) && i < c->bucket_cache_used; i++) { |
cafe5635 KO |
723 | b = list_first_entry(&c->btree_cache, struct btree, list); |
724 | list_rotate_left(&c->btree_cache); | |
725 | ||
726 | if (!b->accessed && | |
e8e1d468 | 727 | !mca_reap(b, 0, false)) { |
cafe5635 KO |
728 | mca_bucket_free(b); |
729 | mca_data_free(b); | |
730 | rw_unlock(true, b); | |
7dc19d5a | 731 | freed++; |
cafe5635 KO |
732 | } else |
733 | b->accessed = 0; | |
734 | } | |
735 | out: | |
cafe5635 | 736 | mutex_unlock(&c->bucket_lock); |
7dc19d5a DC |
737 | return freed; |
738 | } | |
739 | ||
740 | static unsigned long bch_mca_count(struct shrinker *shrink, | |
741 | struct shrink_control *sc) | |
742 | { | |
743 | struct cache_set *c = container_of(shrink, struct cache_set, shrink); | |
744 | ||
745 | if (c->shrinker_disabled) | |
746 | return 0; | |
747 | ||
748 | if (c->try_harder) | |
749 | return 0; | |
750 | ||
751 | return mca_can_free(c) * c->btree_pages; | |
cafe5635 KO |
752 | } |
753 | ||
754 | void bch_btree_cache_free(struct cache_set *c) | |
755 | { | |
756 | struct btree *b; | |
757 | struct closure cl; | |
758 | closure_init_stack(&cl); | |
759 | ||
760 | if (c->shrink.list.next) | |
761 | unregister_shrinker(&c->shrink); | |
762 | ||
763 | mutex_lock(&c->bucket_lock); | |
764 | ||
765 | #ifdef CONFIG_BCACHE_DEBUG | |
766 | if (c->verify_data) | |
767 | list_move(&c->verify_data->list, &c->btree_cache); | |
768 | #endif | |
769 | ||
770 | list_splice(&c->btree_cache_freeable, | |
771 | &c->btree_cache); | |
772 | ||
773 | while (!list_empty(&c->btree_cache)) { | |
774 | b = list_first_entry(&c->btree_cache, struct btree, list); | |
775 | ||
776 | if (btree_node_dirty(b)) | |
777 | btree_complete_write(b, btree_current_write(b)); | |
778 | clear_bit(BTREE_NODE_dirty, &b->flags); | |
779 | ||
780 | mca_data_free(b); | |
781 | } | |
782 | ||
783 | while (!list_empty(&c->btree_cache_freed)) { | |
784 | b = list_first_entry(&c->btree_cache_freed, | |
785 | struct btree, list); | |
786 | list_del(&b->list); | |
787 | cancel_delayed_work_sync(&b->work); | |
788 | kfree(b); | |
789 | } | |
790 | ||
791 | mutex_unlock(&c->bucket_lock); | |
792 | } | |
793 | ||
794 | int bch_btree_cache_alloc(struct cache_set *c) | |
795 | { | |
796 | unsigned i; | |
797 | ||
cafe5635 | 798 | for (i = 0; i < mca_reserve(c); i++) |
72a44517 KO |
799 | if (!mca_bucket_alloc(c, &ZERO_KEY, GFP_KERNEL)) |
800 | return -ENOMEM; | |
cafe5635 KO |
801 | |
802 | list_splice_init(&c->btree_cache, | |
803 | &c->btree_cache_freeable); | |
804 | ||
805 | #ifdef CONFIG_BCACHE_DEBUG | |
806 | mutex_init(&c->verify_lock); | |
807 | ||
808 | c->verify_data = mca_bucket_alloc(c, &ZERO_KEY, GFP_KERNEL); | |
809 | ||
810 | if (c->verify_data && | |
811 | c->verify_data->sets[0].data) | |
812 | list_del_init(&c->verify_data->list); | |
813 | else | |
814 | c->verify_data = NULL; | |
815 | #endif | |
816 | ||
7dc19d5a DC |
817 | c->shrink.count_objects = bch_mca_count; |
818 | c->shrink.scan_objects = bch_mca_scan; | |
cafe5635 KO |
819 | c->shrink.seeks = 4; |
820 | c->shrink.batch = c->btree_pages * 2; | |
821 | register_shrinker(&c->shrink); | |
822 | ||
823 | return 0; | |
824 | } | |
825 | ||
826 | /* Btree in memory cache - hash table */ | |
827 | ||
828 | static struct hlist_head *mca_hash(struct cache_set *c, struct bkey *k) | |
829 | { | |
830 | return &c->bucket_hash[hash_32(PTR_HASH(c, k), BUCKET_HASH_BITS)]; | |
831 | } | |
832 | ||
833 | static struct btree *mca_find(struct cache_set *c, struct bkey *k) | |
834 | { | |
835 | struct btree *b; | |
836 | ||
837 | rcu_read_lock(); | |
838 | hlist_for_each_entry_rcu(b, mca_hash(c, k), hash) | |
839 | if (PTR_HASH(c, &b->key) == PTR_HASH(c, k)) | |
840 | goto out; | |
841 | b = NULL; | |
842 | out: | |
843 | rcu_read_unlock(); | |
844 | return b; | |
845 | } | |
846 | ||
e8e1d468 | 847 | static struct btree *mca_cannibalize(struct cache_set *c, struct bkey *k) |
cafe5635 | 848 | { |
e8e1d468 | 849 | struct btree *b; |
cafe5635 | 850 | |
c37511b8 KO |
851 | trace_bcache_btree_cache_cannibalize(c); |
852 | ||
e8e1d468 KO |
853 | if (!c->try_harder) { |
854 | c->try_harder = current; | |
855 | c->try_harder_start = local_clock(); | |
856 | } else if (c->try_harder != current) | |
857 | return ERR_PTR(-ENOSPC); | |
cafe5635 | 858 | |
e8e1d468 KO |
859 | list_for_each_entry_reverse(b, &c->btree_cache, list) |
860 | if (!mca_reap(b, btree_order(k), false)) | |
861 | return b; | |
cafe5635 | 862 | |
e8e1d468 KO |
863 | list_for_each_entry_reverse(b, &c->btree_cache, list) |
864 | if (!mca_reap(b, btree_order(k), true)) | |
865 | return b; | |
cafe5635 | 866 | |
e8e1d468 | 867 | return ERR_PTR(-ENOMEM); |
cafe5635 KO |
868 | } |
869 | ||
870 | /* | |
871 | * We can only have one thread cannibalizing other cached btree nodes at a time, | |
872 | * or we'll deadlock. We use an open coded mutex to ensure that, which a | |
873 | * cannibalize_bucket() will take. This means every time we unlock the root of | |
874 | * the btree, we need to release this lock if we have it held. | |
875 | */ | |
df8e8970 | 876 | static void bch_cannibalize_unlock(struct cache_set *c) |
cafe5635 | 877 | { |
e8e1d468 | 878 | if (c->try_harder == current) { |
169ef1cf | 879 | bch_time_stats_update(&c->try_harder_time, c->try_harder_start); |
cafe5635 | 880 | c->try_harder = NULL; |
e8e1d468 | 881 | wake_up(&c->try_wait); |
cafe5635 KO |
882 | } |
883 | } | |
884 | ||
e8e1d468 | 885 | static struct btree *mca_alloc(struct cache_set *c, struct bkey *k, int level) |
cafe5635 KO |
886 | { |
887 | struct btree *b; | |
888 | ||
e8e1d468 KO |
889 | BUG_ON(current->bio_list); |
890 | ||
cafe5635 KO |
891 | lockdep_assert_held(&c->bucket_lock); |
892 | ||
893 | if (mca_find(c, k)) | |
894 | return NULL; | |
895 | ||
896 | /* btree_free() doesn't free memory; it sticks the node on the end of | |
897 | * the list. Check if there's any freed nodes there: | |
898 | */ | |
899 | list_for_each_entry(b, &c->btree_cache_freeable, list) | |
e8e1d468 | 900 | if (!mca_reap(b, btree_order(k), false)) |
cafe5635 KO |
901 | goto out; |
902 | ||
903 | /* We never free struct btree itself, just the memory that holds the on | |
904 | * disk node. Check the freed list before allocating a new one: | |
905 | */ | |
906 | list_for_each_entry(b, &c->btree_cache_freed, list) | |
e8e1d468 | 907 | if (!mca_reap(b, 0, false)) { |
cafe5635 KO |
908 | mca_data_alloc(b, k, __GFP_NOWARN|GFP_NOIO); |
909 | if (!b->sets[0].data) | |
910 | goto err; | |
911 | else | |
912 | goto out; | |
913 | } | |
914 | ||
915 | b = mca_bucket_alloc(c, k, __GFP_NOWARN|GFP_NOIO); | |
916 | if (!b) | |
917 | goto err; | |
918 | ||
919 | BUG_ON(!down_write_trylock(&b->lock)); | |
920 | if (!b->sets->data) | |
921 | goto err; | |
922 | out: | |
923 | BUG_ON(!closure_is_unlocked(&b->io.cl)); | |
924 | ||
925 | bkey_copy(&b->key, k); | |
926 | list_move(&b->list, &c->btree_cache); | |
927 | hlist_del_init_rcu(&b->hash); | |
928 | hlist_add_head_rcu(&b->hash, mca_hash(c, k)); | |
929 | ||
930 | lock_set_subclass(&b->lock.dep_map, level + 1, _THIS_IP_); | |
931 | b->level = level; | |
d6fd3b11 | 932 | b->parent = (void *) ~0UL; |
cafe5635 KO |
933 | |
934 | mca_reinit(b); | |
935 | ||
936 | return b; | |
937 | err: | |
938 | if (b) | |
939 | rw_unlock(true, b); | |
940 | ||
e8e1d468 | 941 | b = mca_cannibalize(c, k); |
cafe5635 KO |
942 | if (!IS_ERR(b)) |
943 | goto out; | |
944 | ||
945 | return b; | |
946 | } | |
947 | ||
948 | /** | |
949 | * bch_btree_node_get - find a btree node in the cache and lock it, reading it | |
950 | * in from disk if necessary. | |
951 | * | |
b54d6934 | 952 | * If IO is necessary and running under generic_make_request, returns -EAGAIN. |
cafe5635 KO |
953 | * |
954 | * The btree node will have either a read or a write lock held, depending on | |
955 | * level and op->lock. | |
956 | */ | |
957 | struct btree *bch_btree_node_get(struct cache_set *c, struct bkey *k, | |
e8e1d468 | 958 | int level, bool write) |
cafe5635 KO |
959 | { |
960 | int i = 0; | |
cafe5635 KO |
961 | struct btree *b; |
962 | ||
963 | BUG_ON(level < 0); | |
964 | retry: | |
965 | b = mca_find(c, k); | |
966 | ||
967 | if (!b) { | |
57943511 KO |
968 | if (current->bio_list) |
969 | return ERR_PTR(-EAGAIN); | |
970 | ||
cafe5635 | 971 | mutex_lock(&c->bucket_lock); |
e8e1d468 | 972 | b = mca_alloc(c, k, level); |
cafe5635 KO |
973 | mutex_unlock(&c->bucket_lock); |
974 | ||
975 | if (!b) | |
976 | goto retry; | |
977 | if (IS_ERR(b)) | |
978 | return b; | |
979 | ||
57943511 | 980 | bch_btree_node_read(b); |
cafe5635 KO |
981 | |
982 | if (!write) | |
983 | downgrade_write(&b->lock); | |
984 | } else { | |
985 | rw_lock(write, b, level); | |
986 | if (PTR_HASH(c, &b->key) != PTR_HASH(c, k)) { | |
987 | rw_unlock(write, b); | |
988 | goto retry; | |
989 | } | |
990 | BUG_ON(b->level != level); | |
991 | } | |
992 | ||
993 | b->accessed = 1; | |
994 | ||
995 | for (; i <= b->nsets && b->sets[i].size; i++) { | |
996 | prefetch(b->sets[i].tree); | |
997 | prefetch(b->sets[i].data); | |
998 | } | |
999 | ||
1000 | for (; i <= b->nsets; i++) | |
1001 | prefetch(b->sets[i].data); | |
1002 | ||
57943511 | 1003 | if (btree_node_io_error(b)) { |
cafe5635 | 1004 | rw_unlock(write, b); |
57943511 KO |
1005 | return ERR_PTR(-EIO); |
1006 | } | |
1007 | ||
1008 | BUG_ON(!b->written); | |
cafe5635 KO |
1009 | |
1010 | return b; | |
1011 | } | |
1012 | ||
1013 | static void btree_node_prefetch(struct cache_set *c, struct bkey *k, int level) | |
1014 | { | |
1015 | struct btree *b; | |
1016 | ||
1017 | mutex_lock(&c->bucket_lock); | |
e8e1d468 | 1018 | b = mca_alloc(c, k, level); |
cafe5635 KO |
1019 | mutex_unlock(&c->bucket_lock); |
1020 | ||
1021 | if (!IS_ERR_OR_NULL(b)) { | |
57943511 | 1022 | bch_btree_node_read(b); |
cafe5635 KO |
1023 | rw_unlock(true, b); |
1024 | } | |
1025 | } | |
1026 | ||
1027 | /* Btree alloc */ | |
1028 | ||
e8e1d468 | 1029 | static void btree_node_free(struct btree *b) |
cafe5635 KO |
1030 | { |
1031 | unsigned i; | |
1032 | ||
c37511b8 KO |
1033 | trace_bcache_btree_node_free(b); |
1034 | ||
cafe5635 | 1035 | BUG_ON(b == b->c->root); |
cafe5635 KO |
1036 | |
1037 | if (btree_node_dirty(b)) | |
1038 | btree_complete_write(b, btree_current_write(b)); | |
1039 | clear_bit(BTREE_NODE_dirty, &b->flags); | |
1040 | ||
cafe5635 KO |
1041 | cancel_delayed_work(&b->work); |
1042 | ||
1043 | mutex_lock(&b->c->bucket_lock); | |
1044 | ||
1045 | for (i = 0; i < KEY_PTRS(&b->key); i++) { | |
1046 | BUG_ON(atomic_read(&PTR_BUCKET(b->c, &b->key, i)->pin)); | |
1047 | ||
1048 | bch_inc_gen(PTR_CACHE(b->c, &b->key, i), | |
1049 | PTR_BUCKET(b->c, &b->key, i)); | |
1050 | } | |
1051 | ||
1052 | bch_bucket_free(b->c, &b->key); | |
1053 | mca_bucket_free(b); | |
1054 | mutex_unlock(&b->c->bucket_lock); | |
1055 | } | |
1056 | ||
35fcd848 | 1057 | struct btree *bch_btree_node_alloc(struct cache_set *c, int level) |
cafe5635 KO |
1058 | { |
1059 | BKEY_PADDED(key) k; | |
1060 | struct btree *b = ERR_PTR(-EAGAIN); | |
1061 | ||
1062 | mutex_lock(&c->bucket_lock); | |
1063 | retry: | |
35fcd848 | 1064 | if (__bch_bucket_alloc_set(c, WATERMARK_METADATA, &k.key, 1, true)) |
cafe5635 KO |
1065 | goto err; |
1066 | ||
1067 | SET_KEY_SIZE(&k.key, c->btree_pages * PAGE_SECTORS); | |
1068 | ||
e8e1d468 | 1069 | b = mca_alloc(c, &k.key, level); |
cafe5635 KO |
1070 | if (IS_ERR(b)) |
1071 | goto err_free; | |
1072 | ||
1073 | if (!b) { | |
b1a67b0f KO |
1074 | cache_bug(c, |
1075 | "Tried to allocate bucket that was in btree cache"); | |
cafe5635 KO |
1076 | __bkey_put(c, &k.key); |
1077 | goto retry; | |
1078 | } | |
1079 | ||
cafe5635 KO |
1080 | b->accessed = 1; |
1081 | bch_bset_init_next(b); | |
1082 | ||
1083 | mutex_unlock(&c->bucket_lock); | |
c37511b8 KO |
1084 | |
1085 | trace_bcache_btree_node_alloc(b); | |
cafe5635 KO |
1086 | return b; |
1087 | err_free: | |
1088 | bch_bucket_free(c, &k.key); | |
1089 | __bkey_put(c, &k.key); | |
1090 | err: | |
1091 | mutex_unlock(&c->bucket_lock); | |
c37511b8 KO |
1092 | |
1093 | trace_bcache_btree_node_alloc_fail(b); | |
cafe5635 KO |
1094 | return b; |
1095 | } | |
1096 | ||
35fcd848 | 1097 | static struct btree *btree_node_alloc_replacement(struct btree *b) |
cafe5635 | 1098 | { |
35fcd848 | 1099 | struct btree *n = bch_btree_node_alloc(b->c, b->level); |
cafe5635 KO |
1100 | if (!IS_ERR_OR_NULL(n)) |
1101 | bch_btree_sort_into(b, n); | |
1102 | ||
1103 | return n; | |
1104 | } | |
1105 | ||
1106 | /* Garbage collection */ | |
1107 | ||
1108 | uint8_t __bch_btree_mark_key(struct cache_set *c, int level, struct bkey *k) | |
1109 | { | |
1110 | uint8_t stale = 0; | |
1111 | unsigned i; | |
1112 | struct bucket *g; | |
1113 | ||
1114 | /* | |
1115 | * ptr_invalid() can't return true for the keys that mark btree nodes as | |
1116 | * freed, but since ptr_bad() returns true we'll never actually use them | |
1117 | * for anything and thus we don't want mark their pointers here | |
1118 | */ | |
1119 | if (!bkey_cmp(k, &ZERO_KEY)) | |
1120 | return stale; | |
1121 | ||
1122 | for (i = 0; i < KEY_PTRS(k); i++) { | |
1123 | if (!ptr_available(c, k, i)) | |
1124 | continue; | |
1125 | ||
1126 | g = PTR_BUCKET(c, k, i); | |
1127 | ||
1128 | if (gen_after(g->gc_gen, PTR_GEN(k, i))) | |
1129 | g->gc_gen = PTR_GEN(k, i); | |
1130 | ||
1131 | if (ptr_stale(c, k, i)) { | |
1132 | stale = max(stale, ptr_stale(c, k, i)); | |
1133 | continue; | |
1134 | } | |
1135 | ||
1136 | cache_bug_on(GC_MARK(g) && | |
1137 | (GC_MARK(g) == GC_MARK_METADATA) != (level != 0), | |
1138 | c, "inconsistent ptrs: mark = %llu, level = %i", | |
1139 | GC_MARK(g), level); | |
1140 | ||
1141 | if (level) | |
1142 | SET_GC_MARK(g, GC_MARK_METADATA); | |
1143 | else if (KEY_DIRTY(k)) | |
1144 | SET_GC_MARK(g, GC_MARK_DIRTY); | |
1145 | ||
1146 | /* guard against overflow */ | |
1147 | SET_GC_SECTORS_USED(g, min_t(unsigned, | |
1148 | GC_SECTORS_USED(g) + KEY_SIZE(k), | |
1149 | (1 << 14) - 1)); | |
1150 | ||
1151 | BUG_ON(!GC_SECTORS_USED(g)); | |
1152 | } | |
1153 | ||
1154 | return stale; | |
1155 | } | |
1156 | ||
1157 | #define btree_mark_key(b, k) __bch_btree_mark_key(b->c, b->level, k) | |
1158 | ||
1159 | static int btree_gc_mark_node(struct btree *b, unsigned *keys, | |
1160 | struct gc_stat *gc) | |
1161 | { | |
1162 | uint8_t stale = 0; | |
1163 | unsigned last_dev = -1; | |
1164 | struct bcache_device *d = NULL; | |
1165 | struct bkey *k; | |
1166 | struct btree_iter iter; | |
1167 | struct bset_tree *t; | |
1168 | ||
1169 | gc->nodes++; | |
1170 | ||
1171 | for_each_key_filter(b, k, &iter, bch_ptr_invalid) { | |
1172 | if (last_dev != KEY_INODE(k)) { | |
1173 | last_dev = KEY_INODE(k); | |
1174 | ||
1175 | d = KEY_INODE(k) < b->c->nr_uuids | |
1176 | ? b->c->devices[last_dev] | |
1177 | : NULL; | |
1178 | } | |
1179 | ||
1180 | stale = max(stale, btree_mark_key(b, k)); | |
1181 | ||
1182 | if (bch_ptr_bad(b, k)) | |
1183 | continue; | |
1184 | ||
1185 | *keys += bkey_u64s(k); | |
1186 | ||
1187 | gc->key_bytes += bkey_u64s(k); | |
1188 | gc->nkeys++; | |
1189 | ||
1190 | gc->data += KEY_SIZE(k); | |
444fc0b6 | 1191 | if (KEY_DIRTY(k)) |
cafe5635 | 1192 | gc->dirty += KEY_SIZE(k); |
cafe5635 KO |
1193 | } |
1194 | ||
1195 | for (t = b->sets; t <= &b->sets[b->nsets]; t++) | |
1196 | btree_bug_on(t->size && | |
1197 | bset_written(b, t) && | |
1198 | bkey_cmp(&b->key, &t->end) < 0, | |
1199 | b, "found short btree key in gc"); | |
1200 | ||
1201 | return stale; | |
1202 | } | |
1203 | ||
e8e1d468 | 1204 | static struct btree *btree_gc_alloc(struct btree *b, struct bkey *k) |
cafe5635 KO |
1205 | { |
1206 | /* | |
1207 | * We block priorities from being written for the duration of garbage | |
1208 | * collection, so we can't sleep in btree_alloc() -> | |
1209 | * bch_bucket_alloc_set(), or we'd risk deadlock - so we don't pass it | |
1210 | * our closure. | |
1211 | */ | |
35fcd848 | 1212 | struct btree *n = btree_node_alloc_replacement(b); |
cafe5635 KO |
1213 | |
1214 | if (!IS_ERR_OR_NULL(n)) { | |
1215 | swap(b, n); | |
57943511 | 1216 | __bkey_put(b->c, &b->key); |
cafe5635 KO |
1217 | |
1218 | memcpy(k->ptr, b->key.ptr, | |
1219 | sizeof(uint64_t) * KEY_PTRS(&b->key)); | |
1220 | ||
e8e1d468 | 1221 | btree_node_free(n); |
cafe5635 KO |
1222 | up_write(&n->lock); |
1223 | } | |
1224 | ||
1225 | return b; | |
1226 | } | |
1227 | ||
1228 | /* | |
1229 | * Leaving this at 2 until we've got incremental garbage collection done; it | |
1230 | * could be higher (and has been tested with 4) except that garbage collection | |
1231 | * could take much longer, adversely affecting latency. | |
1232 | */ | |
1233 | #define GC_MERGE_NODES 2U | |
1234 | ||
1235 | struct gc_merge_info { | |
1236 | struct btree *b; | |
1237 | struct bkey *k; | |
1238 | unsigned keys; | |
1239 | }; | |
1240 | ||
e8e1d468 KO |
1241 | static void btree_gc_coalesce(struct btree *b, struct gc_stat *gc, |
1242 | struct gc_merge_info *r) | |
cafe5635 KO |
1243 | { |
1244 | unsigned nodes = 0, keys = 0, blocks; | |
1245 | int i; | |
b54d6934 KO |
1246 | struct closure cl; |
1247 | ||
1248 | closure_init_stack(&cl); | |
cafe5635 KO |
1249 | |
1250 | while (nodes < GC_MERGE_NODES && r[nodes].b) | |
1251 | keys += r[nodes++].keys; | |
1252 | ||
1253 | blocks = btree_default_blocks(b->c) * 2 / 3; | |
1254 | ||
1255 | if (nodes < 2 || | |
1256 | __set_blocks(b->sets[0].data, keys, b->c) > blocks * (nodes - 1)) | |
1257 | return; | |
1258 | ||
1259 | for (i = nodes - 1; i >= 0; --i) { | |
1260 | if (r[i].b->written) | |
e8e1d468 | 1261 | r[i].b = btree_gc_alloc(r[i].b, r[i].k); |
cafe5635 KO |
1262 | |
1263 | if (r[i].b->written) | |
1264 | return; | |
1265 | } | |
1266 | ||
1267 | for (i = nodes - 1; i > 0; --i) { | |
1268 | struct bset *n1 = r[i].b->sets->data; | |
1269 | struct bset *n2 = r[i - 1].b->sets->data; | |
1270 | struct bkey *k, *last = NULL; | |
1271 | ||
1272 | keys = 0; | |
1273 | ||
1274 | if (i == 1) { | |
1275 | /* | |
1276 | * Last node we're not getting rid of - we're getting | |
1277 | * rid of the node at r[0]. Have to try and fit all of | |
1278 | * the remaining keys into this node; we can't ensure | |
1279 | * they will always fit due to rounding and variable | |
1280 | * length keys (shouldn't be possible in practice, | |
1281 | * though) | |
1282 | */ | |
1283 | if (__set_blocks(n1, n1->keys + r->keys, | |
1284 | b->c) > btree_blocks(r[i].b)) | |
1285 | return; | |
1286 | ||
1287 | keys = n2->keys; | |
1288 | last = &r->b->key; | |
1289 | } else | |
1290 | for (k = n2->start; | |
1291 | k < end(n2); | |
1292 | k = bkey_next(k)) { | |
1293 | if (__set_blocks(n1, n1->keys + keys + | |
1294 | bkey_u64s(k), b->c) > blocks) | |
1295 | break; | |
1296 | ||
1297 | last = k; | |
1298 | keys += bkey_u64s(k); | |
1299 | } | |
1300 | ||
1301 | BUG_ON(__set_blocks(n1, n1->keys + keys, | |
1302 | b->c) > btree_blocks(r[i].b)); | |
1303 | ||
1304 | if (last) { | |
1305 | bkey_copy_key(&r[i].b->key, last); | |
1306 | bkey_copy_key(r[i].k, last); | |
1307 | } | |
1308 | ||
1309 | memcpy(end(n1), | |
1310 | n2->start, | |
1311 | (void *) node(n2, keys) - (void *) n2->start); | |
1312 | ||
1313 | n1->keys += keys; | |
1314 | ||
1315 | memmove(n2->start, | |
1316 | node(n2, keys), | |
1317 | (void *) end(n2) - (void *) node(n2, keys)); | |
1318 | ||
1319 | n2->keys -= keys; | |
1320 | ||
1321 | r[i].keys = n1->keys; | |
1322 | r[i - 1].keys = n2->keys; | |
1323 | } | |
1324 | ||
e8e1d468 | 1325 | btree_node_free(r->b); |
cafe5635 KO |
1326 | up_write(&r->b->lock); |
1327 | ||
c37511b8 | 1328 | trace_bcache_btree_gc_coalesce(nodes); |
cafe5635 KO |
1329 | |
1330 | gc->nodes--; | |
1331 | nodes--; | |
1332 | ||
1333 | memmove(&r[0], &r[1], sizeof(struct gc_merge_info) * nodes); | |
1334 | memset(&r[nodes], 0, sizeof(struct gc_merge_info)); | |
1335 | } | |
1336 | ||
1337 | static int btree_gc_recurse(struct btree *b, struct btree_op *op, | |
1338 | struct closure *writes, struct gc_stat *gc) | |
1339 | { | |
1340 | void write(struct btree *r) | |
1341 | { | |
b54d6934 | 1342 | if (!r->written || btree_node_dirty(r)) |
57943511 | 1343 | bch_btree_node_write(r, writes); |
cafe5635 KO |
1344 | |
1345 | up_write(&r->lock); | |
1346 | } | |
1347 | ||
1348 | int ret = 0, stale; | |
1349 | unsigned i; | |
1350 | struct gc_merge_info r[GC_MERGE_NODES]; | |
1351 | ||
1352 | memset(r, 0, sizeof(r)); | |
1353 | ||
1354 | while ((r->k = bch_next_recurse_key(b, &b->c->gc_done))) { | |
e8e1d468 | 1355 | r->b = bch_btree_node_get(b->c, r->k, b->level - 1, true); |
cafe5635 KO |
1356 | |
1357 | if (IS_ERR(r->b)) { | |
1358 | ret = PTR_ERR(r->b); | |
1359 | break; | |
1360 | } | |
1361 | ||
1362 | r->keys = 0; | |
1363 | stale = btree_gc_mark_node(r->b, &r->keys, gc); | |
1364 | ||
1365 | if (!b->written && | |
1366 | (r->b->level || stale > 10 || | |
1367 | b->c->gc_always_rewrite)) | |
e8e1d468 | 1368 | r->b = btree_gc_alloc(r->b, r->k); |
cafe5635 KO |
1369 | |
1370 | if (r->b->level) | |
1371 | ret = btree_gc_recurse(r->b, op, writes, gc); | |
1372 | ||
1373 | if (ret) { | |
1374 | write(r->b); | |
1375 | break; | |
1376 | } | |
1377 | ||
1378 | bkey_copy_key(&b->c->gc_done, r->k); | |
1379 | ||
1380 | if (!b->written) | |
e8e1d468 | 1381 | btree_gc_coalesce(b, gc, r); |
cafe5635 KO |
1382 | |
1383 | if (r[GC_MERGE_NODES - 1].b) | |
1384 | write(r[GC_MERGE_NODES - 1].b); | |
1385 | ||
1386 | memmove(&r[1], &r[0], | |
1387 | sizeof(struct gc_merge_info) * (GC_MERGE_NODES - 1)); | |
1388 | ||
1389 | /* When we've got incremental GC working, we'll want to do | |
1390 | * if (should_resched()) | |
1391 | * return -EAGAIN; | |
1392 | */ | |
1393 | cond_resched(); | |
1394 | #if 0 | |
1395 | if (need_resched()) { | |
1396 | ret = -EAGAIN; | |
1397 | break; | |
1398 | } | |
1399 | #endif | |
1400 | } | |
1401 | ||
1402 | for (i = 1; i < GC_MERGE_NODES && r[i].b; i++) | |
1403 | write(r[i].b); | |
1404 | ||
1405 | /* Might have freed some children, must remove their keys */ | |
1406 | if (!b->written) | |
1407 | bch_btree_sort(b); | |
1408 | ||
1409 | return ret; | |
1410 | } | |
1411 | ||
1412 | static int bch_btree_gc_root(struct btree *b, struct btree_op *op, | |
1413 | struct closure *writes, struct gc_stat *gc) | |
1414 | { | |
1415 | struct btree *n = NULL; | |
1416 | unsigned keys = 0; | |
1417 | int ret = 0, stale = btree_gc_mark_node(b, &keys, gc); | |
b54d6934 KO |
1418 | struct closure cl; |
1419 | ||
1420 | closure_init_stack(&cl); | |
cafe5635 KO |
1421 | |
1422 | if (b->level || stale > 10) | |
35fcd848 | 1423 | n = btree_node_alloc_replacement(b); |
cafe5635 KO |
1424 | |
1425 | if (!IS_ERR_OR_NULL(n)) | |
1426 | swap(b, n); | |
1427 | ||
1428 | if (b->level) | |
1429 | ret = btree_gc_recurse(b, op, writes, gc); | |
1430 | ||
1431 | if (!b->written || btree_node_dirty(b)) { | |
b54d6934 | 1432 | bch_btree_node_write(b, n ? &cl : NULL); |
cafe5635 KO |
1433 | } |
1434 | ||
1435 | if (!IS_ERR_OR_NULL(n)) { | |
b54d6934 | 1436 | closure_sync(&cl); |
cafe5635 | 1437 | bch_btree_set_root(b); |
e8e1d468 | 1438 | btree_node_free(n); |
cafe5635 KO |
1439 | rw_unlock(true, b); |
1440 | } | |
1441 | ||
1442 | return ret; | |
1443 | } | |
1444 | ||
1445 | static void btree_gc_start(struct cache_set *c) | |
1446 | { | |
1447 | struct cache *ca; | |
1448 | struct bucket *b; | |
cafe5635 KO |
1449 | unsigned i; |
1450 | ||
1451 | if (!c->gc_mark_valid) | |
1452 | return; | |
1453 | ||
1454 | mutex_lock(&c->bucket_lock); | |
1455 | ||
1456 | c->gc_mark_valid = 0; | |
1457 | c->gc_done = ZERO_KEY; | |
1458 | ||
1459 | for_each_cache(ca, c, i) | |
1460 | for_each_bucket(b, ca) { | |
1461 | b->gc_gen = b->gen; | |
29ebf465 | 1462 | if (!atomic_read(&b->pin)) { |
cafe5635 | 1463 | SET_GC_MARK(b, GC_MARK_RECLAIMABLE); |
29ebf465 KO |
1464 | SET_GC_SECTORS_USED(b, 0); |
1465 | } | |
cafe5635 KO |
1466 | } |
1467 | ||
cafe5635 KO |
1468 | mutex_unlock(&c->bucket_lock); |
1469 | } | |
1470 | ||
1471 | size_t bch_btree_gc_finish(struct cache_set *c) | |
1472 | { | |
1473 | size_t available = 0; | |
1474 | struct bucket *b; | |
1475 | struct cache *ca; | |
cafe5635 KO |
1476 | unsigned i; |
1477 | ||
1478 | mutex_lock(&c->bucket_lock); | |
1479 | ||
1480 | set_gc_sectors(c); | |
1481 | c->gc_mark_valid = 1; | |
1482 | c->need_gc = 0; | |
1483 | ||
1484 | if (c->root) | |
1485 | for (i = 0; i < KEY_PTRS(&c->root->key); i++) | |
1486 | SET_GC_MARK(PTR_BUCKET(c, &c->root->key, i), | |
1487 | GC_MARK_METADATA); | |
1488 | ||
1489 | for (i = 0; i < KEY_PTRS(&c->uuid_bucket); i++) | |
1490 | SET_GC_MARK(PTR_BUCKET(c, &c->uuid_bucket, i), | |
1491 | GC_MARK_METADATA); | |
1492 | ||
1493 | for_each_cache(ca, c, i) { | |
1494 | uint64_t *i; | |
1495 | ||
1496 | ca->invalidate_needs_gc = 0; | |
1497 | ||
1498 | for (i = ca->sb.d; i < ca->sb.d + ca->sb.keys; i++) | |
1499 | SET_GC_MARK(ca->buckets + *i, GC_MARK_METADATA); | |
1500 | ||
1501 | for (i = ca->prio_buckets; | |
1502 | i < ca->prio_buckets + prio_buckets(ca) * 2; i++) | |
1503 | SET_GC_MARK(ca->buckets + *i, GC_MARK_METADATA); | |
1504 | ||
1505 | for_each_bucket(b, ca) { | |
1506 | b->last_gc = b->gc_gen; | |
1507 | c->need_gc = max(c->need_gc, bucket_gc_gen(b)); | |
1508 | ||
1509 | if (!atomic_read(&b->pin) && | |
1510 | GC_MARK(b) == GC_MARK_RECLAIMABLE) { | |
1511 | available++; | |
1512 | if (!GC_SECTORS_USED(b)) | |
1513 | bch_bucket_add_unused(ca, b); | |
1514 | } | |
1515 | } | |
1516 | } | |
1517 | ||
cafe5635 KO |
1518 | mutex_unlock(&c->bucket_lock); |
1519 | return available; | |
1520 | } | |
1521 | ||
72a44517 | 1522 | static void bch_btree_gc(struct cache_set *c) |
cafe5635 | 1523 | { |
cafe5635 KO |
1524 | int ret; |
1525 | unsigned long available; | |
1526 | struct gc_stat stats; | |
1527 | struct closure writes; | |
1528 | struct btree_op op; | |
cafe5635 | 1529 | uint64_t start_time = local_clock(); |
57943511 | 1530 | |
c37511b8 | 1531 | trace_bcache_gc_start(c); |
cafe5635 KO |
1532 | |
1533 | memset(&stats, 0, sizeof(struct gc_stat)); | |
1534 | closure_init_stack(&writes); | |
b54d6934 | 1535 | bch_btree_op_init(&op, SHRT_MAX); |
cafe5635 KO |
1536 | |
1537 | btree_gc_start(c); | |
1538 | ||
57943511 KO |
1539 | atomic_inc(&c->prio_blocked); |
1540 | ||
cafe5635 | 1541 | ret = btree_root(gc_root, c, &op, &writes, &stats); |
cafe5635 KO |
1542 | closure_sync(&writes); |
1543 | ||
1544 | if (ret) { | |
cafe5635 | 1545 | pr_warn("gc failed!"); |
72a44517 | 1546 | return; |
cafe5635 KO |
1547 | } |
1548 | ||
1549 | /* Possibly wait for new UUIDs or whatever to hit disk */ | |
b54d6934 KO |
1550 | bch_journal_meta(c, &writes); |
1551 | closure_sync(&writes); | |
cafe5635 KO |
1552 | |
1553 | available = bch_btree_gc_finish(c); | |
1554 | ||
57943511 KO |
1555 | atomic_dec(&c->prio_blocked); |
1556 | wake_up_allocators(c); | |
1557 | ||
169ef1cf | 1558 | bch_time_stats_update(&c->btree_gc_time, start_time); |
cafe5635 KO |
1559 | |
1560 | stats.key_bytes *= sizeof(uint64_t); | |
1561 | stats.dirty <<= 9; | |
1562 | stats.data <<= 9; | |
1563 | stats.in_use = (c->nbuckets - available) * 100 / c->nbuckets; | |
1564 | memcpy(&c->gc_stats, &stats, sizeof(struct gc_stat)); | |
cafe5635 | 1565 | |
c37511b8 | 1566 | trace_bcache_gc_end(c); |
cafe5635 | 1567 | |
72a44517 KO |
1568 | bch_moving_gc(c); |
1569 | } | |
1570 | ||
1571 | static int bch_gc_thread(void *arg) | |
1572 | { | |
1573 | struct cache_set *c = arg; | |
1574 | ||
1575 | while (1) { | |
1576 | bch_btree_gc(c); | |
1577 | ||
1578 | set_current_state(TASK_INTERRUPTIBLE); | |
1579 | if (kthread_should_stop()) | |
1580 | break; | |
1581 | ||
1582 | try_to_freeze(); | |
1583 | schedule(); | |
1584 | } | |
1585 | ||
1586 | return 0; | |
cafe5635 KO |
1587 | } |
1588 | ||
72a44517 | 1589 | int bch_gc_thread_start(struct cache_set *c) |
cafe5635 | 1590 | { |
72a44517 KO |
1591 | c->gc_thread = kthread_create(bch_gc_thread, c, "bcache_gc"); |
1592 | if (IS_ERR(c->gc_thread)) | |
1593 | return PTR_ERR(c->gc_thread); | |
1594 | ||
1595 | set_task_state(c->gc_thread, TASK_INTERRUPTIBLE); | |
1596 | return 0; | |
cafe5635 KO |
1597 | } |
1598 | ||
1599 | /* Initial partial gc */ | |
1600 | ||
1601 | static int bch_btree_check_recurse(struct btree *b, struct btree_op *op, | |
1602 | unsigned long **seen) | |
1603 | { | |
1604 | int ret; | |
1605 | unsigned i; | |
1606 | struct bkey *k; | |
1607 | struct bucket *g; | |
1608 | struct btree_iter iter; | |
1609 | ||
1610 | for_each_key_filter(b, k, &iter, bch_ptr_invalid) { | |
1611 | for (i = 0; i < KEY_PTRS(k); i++) { | |
1612 | if (!ptr_available(b->c, k, i)) | |
1613 | continue; | |
1614 | ||
1615 | g = PTR_BUCKET(b->c, k, i); | |
1616 | ||
1617 | if (!__test_and_set_bit(PTR_BUCKET_NR(b->c, k, i), | |
1618 | seen[PTR_DEV(k, i)]) || | |
1619 | !ptr_stale(b->c, k, i)) { | |
1620 | g->gen = PTR_GEN(k, i); | |
1621 | ||
1622 | if (b->level) | |
1623 | g->prio = BTREE_PRIO; | |
1624 | else if (g->prio == BTREE_PRIO) | |
1625 | g->prio = INITIAL_PRIO; | |
1626 | } | |
1627 | } | |
1628 | ||
1629 | btree_mark_key(b, k); | |
1630 | } | |
1631 | ||
1632 | if (b->level) { | |
1633 | k = bch_next_recurse_key(b, &ZERO_KEY); | |
1634 | ||
1635 | while (k) { | |
1636 | struct bkey *p = bch_next_recurse_key(b, k); | |
1637 | if (p) | |
1638 | btree_node_prefetch(b->c, p, b->level - 1); | |
1639 | ||
1640 | ret = btree(check_recurse, k, b, op, seen); | |
1641 | if (ret) | |
1642 | return ret; | |
1643 | ||
1644 | k = p; | |
1645 | } | |
1646 | } | |
1647 | ||
1648 | return 0; | |
1649 | } | |
1650 | ||
c18536a7 | 1651 | int bch_btree_check(struct cache_set *c) |
cafe5635 KO |
1652 | { |
1653 | int ret = -ENOMEM; | |
1654 | unsigned i; | |
1655 | unsigned long *seen[MAX_CACHES_PER_SET]; | |
c18536a7 | 1656 | struct btree_op op; |
cafe5635 KO |
1657 | |
1658 | memset(seen, 0, sizeof(seen)); | |
b54d6934 | 1659 | bch_btree_op_init(&op, SHRT_MAX); |
cafe5635 KO |
1660 | |
1661 | for (i = 0; c->cache[i]; i++) { | |
1662 | size_t n = DIV_ROUND_UP(c->cache[i]->sb.nbuckets, 8); | |
1663 | seen[i] = kmalloc(n, GFP_KERNEL); | |
1664 | if (!seen[i]) | |
1665 | goto err; | |
1666 | ||
1667 | /* Disables the seen array until prio_read() uses it too */ | |
1668 | memset(seen[i], 0xFF, n); | |
1669 | } | |
1670 | ||
c18536a7 | 1671 | ret = btree_root(check_recurse, c, &op, seen); |
cafe5635 KO |
1672 | err: |
1673 | for (i = 0; i < MAX_CACHES_PER_SET; i++) | |
1674 | kfree(seen[i]); | |
1675 | return ret; | |
1676 | } | |
1677 | ||
1678 | /* Btree insertion */ | |
1679 | ||
1680 | static void shift_keys(struct btree *b, struct bkey *where, struct bkey *insert) | |
1681 | { | |
1682 | struct bset *i = b->sets[b->nsets].data; | |
1683 | ||
1684 | memmove((uint64_t *) where + bkey_u64s(insert), | |
1685 | where, | |
1686 | (void *) end(i) - (void *) where); | |
1687 | ||
1688 | i->keys += bkey_u64s(insert); | |
1689 | bkey_copy(where, insert); | |
1690 | bch_bset_fix_lookup_table(b, where); | |
1691 | } | |
1692 | ||
1b207d80 | 1693 | static bool fix_overlapping_extents(struct btree *b, struct bkey *insert, |
cafe5635 | 1694 | struct btree_iter *iter, |
1b207d80 | 1695 | struct bkey *replace_key) |
cafe5635 | 1696 | { |
279afbad | 1697 | void subtract_dirty(struct bkey *k, uint64_t offset, int sectors) |
cafe5635 | 1698 | { |
279afbad KO |
1699 | if (KEY_DIRTY(k)) |
1700 | bcache_dev_sectors_dirty_add(b->c, KEY_INODE(k), | |
1701 | offset, -sectors); | |
cafe5635 KO |
1702 | } |
1703 | ||
279afbad | 1704 | uint64_t old_offset; |
cafe5635 KO |
1705 | unsigned old_size, sectors_found = 0; |
1706 | ||
1707 | while (1) { | |
1708 | struct bkey *k = bch_btree_iter_next(iter); | |
1709 | if (!k || | |
1710 | bkey_cmp(&START_KEY(k), insert) >= 0) | |
1711 | break; | |
1712 | ||
1713 | if (bkey_cmp(k, &START_KEY(insert)) <= 0) | |
1714 | continue; | |
1715 | ||
279afbad | 1716 | old_offset = KEY_START(k); |
cafe5635 KO |
1717 | old_size = KEY_SIZE(k); |
1718 | ||
1719 | /* | |
1720 | * We might overlap with 0 size extents; we can't skip these | |
1721 | * because if they're in the set we're inserting to we have to | |
1722 | * adjust them so they don't overlap with the key we're | |
1b207d80 | 1723 | * inserting. But we don't want to check them for replace |
cafe5635 KO |
1724 | * operations. |
1725 | */ | |
1726 | ||
1b207d80 | 1727 | if (replace_key && KEY_SIZE(k)) { |
cafe5635 KO |
1728 | /* |
1729 | * k might have been split since we inserted/found the | |
1730 | * key we're replacing | |
1731 | */ | |
1732 | unsigned i; | |
1733 | uint64_t offset = KEY_START(k) - | |
1b207d80 | 1734 | KEY_START(replace_key); |
cafe5635 KO |
1735 | |
1736 | /* But it must be a subset of the replace key */ | |
1b207d80 KO |
1737 | if (KEY_START(k) < KEY_START(replace_key) || |
1738 | KEY_OFFSET(k) > KEY_OFFSET(replace_key)) | |
cafe5635 KO |
1739 | goto check_failed; |
1740 | ||
1741 | /* We didn't find a key that we were supposed to */ | |
1742 | if (KEY_START(k) > KEY_START(insert) + sectors_found) | |
1743 | goto check_failed; | |
1744 | ||
1b207d80 | 1745 | if (KEY_PTRS(replace_key) != KEY_PTRS(k)) |
cafe5635 KO |
1746 | goto check_failed; |
1747 | ||
1748 | /* skip past gen */ | |
1749 | offset <<= 8; | |
1750 | ||
1b207d80 | 1751 | BUG_ON(!KEY_PTRS(replace_key)); |
cafe5635 | 1752 | |
1b207d80 KO |
1753 | for (i = 0; i < KEY_PTRS(replace_key); i++) |
1754 | if (k->ptr[i] != replace_key->ptr[i] + offset) | |
cafe5635 KO |
1755 | goto check_failed; |
1756 | ||
1757 | sectors_found = KEY_OFFSET(k) - KEY_START(insert); | |
1758 | } | |
1759 | ||
1760 | if (bkey_cmp(insert, k) < 0 && | |
1761 | bkey_cmp(&START_KEY(insert), &START_KEY(k)) > 0) { | |
1762 | /* | |
1763 | * We overlapped in the middle of an existing key: that | |
1764 | * means we have to split the old key. But we have to do | |
1765 | * slightly different things depending on whether the | |
1766 | * old key has been written out yet. | |
1767 | */ | |
1768 | ||
1769 | struct bkey *top; | |
1770 | ||
279afbad | 1771 | subtract_dirty(k, KEY_START(insert), KEY_SIZE(insert)); |
cafe5635 KO |
1772 | |
1773 | if (bkey_written(b, k)) { | |
1774 | /* | |
1775 | * We insert a new key to cover the top of the | |
1776 | * old key, and the old key is modified in place | |
1777 | * to represent the bottom split. | |
1778 | * | |
1779 | * It's completely arbitrary whether the new key | |
1780 | * is the top or the bottom, but it has to match | |
1781 | * up with what btree_sort_fixup() does - it | |
1782 | * doesn't check for this kind of overlap, it | |
1783 | * depends on us inserting a new key for the top | |
1784 | * here. | |
1785 | */ | |
1786 | top = bch_bset_search(b, &b->sets[b->nsets], | |
1787 | insert); | |
1788 | shift_keys(b, top, k); | |
1789 | } else { | |
1790 | BKEY_PADDED(key) temp; | |
1791 | bkey_copy(&temp.key, k); | |
1792 | shift_keys(b, k, &temp.key); | |
1793 | top = bkey_next(k); | |
1794 | } | |
1795 | ||
1796 | bch_cut_front(insert, top); | |
1797 | bch_cut_back(&START_KEY(insert), k); | |
1798 | bch_bset_fix_invalidated_key(b, k); | |
1799 | return false; | |
1800 | } | |
1801 | ||
1802 | if (bkey_cmp(insert, k) < 0) { | |
1803 | bch_cut_front(insert, k); | |
1804 | } else { | |
1fa8455d KO |
1805 | if (bkey_cmp(&START_KEY(insert), &START_KEY(k)) > 0) |
1806 | old_offset = KEY_START(insert); | |
1807 | ||
cafe5635 KO |
1808 | if (bkey_written(b, k) && |
1809 | bkey_cmp(&START_KEY(insert), &START_KEY(k)) <= 0) { | |
1810 | /* | |
1811 | * Completely overwrote, so we don't have to | |
1812 | * invalidate the binary search tree | |
1813 | */ | |
1814 | bch_cut_front(k, k); | |
1815 | } else { | |
1816 | __bch_cut_back(&START_KEY(insert), k); | |
1817 | bch_bset_fix_invalidated_key(b, k); | |
1818 | } | |
1819 | } | |
1820 | ||
279afbad | 1821 | subtract_dirty(k, old_offset, old_size - KEY_SIZE(k)); |
cafe5635 KO |
1822 | } |
1823 | ||
1824 | check_failed: | |
1b207d80 | 1825 | if (replace_key) { |
cafe5635 | 1826 | if (!sectors_found) { |
cafe5635 KO |
1827 | return true; |
1828 | } else if (sectors_found < KEY_SIZE(insert)) { | |
1829 | SET_KEY_OFFSET(insert, KEY_OFFSET(insert) - | |
1830 | (KEY_SIZE(insert) - sectors_found)); | |
1831 | SET_KEY_SIZE(insert, sectors_found); | |
1832 | } | |
1833 | } | |
1834 | ||
1835 | return false; | |
1836 | } | |
1837 | ||
1838 | static bool btree_insert_key(struct btree *b, struct btree_op *op, | |
1b207d80 | 1839 | struct bkey *k, struct bkey *replace_key) |
cafe5635 KO |
1840 | { |
1841 | struct bset *i = b->sets[b->nsets].data; | |
1842 | struct bkey *m, *prev; | |
85b1492e | 1843 | unsigned status = BTREE_INSERT_STATUS_INSERT; |
cafe5635 KO |
1844 | |
1845 | BUG_ON(bkey_cmp(k, &b->key) > 0); | |
1846 | BUG_ON(b->level && !KEY_PTRS(k)); | |
1847 | BUG_ON(!b->level && !KEY_OFFSET(k)); | |
1848 | ||
1849 | if (!b->level) { | |
1850 | struct btree_iter iter; | |
1851 | struct bkey search = KEY(KEY_INODE(k), KEY_START(k), 0); | |
1852 | ||
1853 | /* | |
1854 | * bset_search() returns the first key that is strictly greater | |
1855 | * than the search key - but for back merging, we want to find | |
1856 | * the first key that is greater than or equal to KEY_START(k) - | |
1857 | * unless KEY_START(k) is 0. | |
1858 | */ | |
1859 | if (KEY_OFFSET(&search)) | |
1860 | SET_KEY_OFFSET(&search, KEY_OFFSET(&search) - 1); | |
1861 | ||
1862 | prev = NULL; | |
1863 | m = bch_btree_iter_init(b, &iter, &search); | |
1864 | ||
1b207d80 KO |
1865 | if (fix_overlapping_extents(b, k, &iter, replace_key)) { |
1866 | op->insert_collision = true; | |
cafe5635 | 1867 | return false; |
1b207d80 | 1868 | } |
cafe5635 | 1869 | |
1fa8455d KO |
1870 | if (KEY_DIRTY(k)) |
1871 | bcache_dev_sectors_dirty_add(b->c, KEY_INODE(k), | |
1872 | KEY_START(k), KEY_SIZE(k)); | |
1873 | ||
cafe5635 KO |
1874 | while (m != end(i) && |
1875 | bkey_cmp(k, &START_KEY(m)) > 0) | |
1876 | prev = m, m = bkey_next(m); | |
1877 | ||
1878 | if (key_merging_disabled(b->c)) | |
1879 | goto insert; | |
1880 | ||
1881 | /* prev is in the tree, if we merge we're done */ | |
85b1492e | 1882 | status = BTREE_INSERT_STATUS_BACK_MERGE; |
cafe5635 KO |
1883 | if (prev && |
1884 | bch_bkey_try_merge(b, prev, k)) | |
1885 | goto merged; | |
1886 | ||
85b1492e | 1887 | status = BTREE_INSERT_STATUS_OVERWROTE; |
cafe5635 KO |
1888 | if (m != end(i) && |
1889 | KEY_PTRS(m) == KEY_PTRS(k) && !KEY_SIZE(m)) | |
1890 | goto copy; | |
1891 | ||
85b1492e | 1892 | status = BTREE_INSERT_STATUS_FRONT_MERGE; |
cafe5635 KO |
1893 | if (m != end(i) && |
1894 | bch_bkey_try_merge(b, k, m)) | |
1895 | goto copy; | |
1b207d80 KO |
1896 | } else { |
1897 | BUG_ON(replace_key); | |
cafe5635 | 1898 | m = bch_bset_search(b, &b->sets[b->nsets], k); |
1b207d80 | 1899 | } |
cafe5635 KO |
1900 | |
1901 | insert: shift_keys(b, m, k); | |
1902 | copy: bkey_copy(m, k); | |
1903 | merged: | |
1b207d80 KO |
1904 | bch_check_keys(b, "%u for %s", status, |
1905 | replace_key ? "replace" : "insert"); | |
cafe5635 KO |
1906 | |
1907 | if (b->level && !KEY_OFFSET(k)) | |
57943511 | 1908 | btree_current_write(b)->prio_blocked++; |
cafe5635 | 1909 | |
1b207d80 | 1910 | trace_bcache_btree_insert_key(b, k, replace_key != NULL, status); |
cafe5635 KO |
1911 | |
1912 | return true; | |
1913 | } | |
1914 | ||
26c949f8 | 1915 | static bool bch_btree_insert_keys(struct btree *b, struct btree_op *op, |
1b207d80 KO |
1916 | struct keylist *insert_keys, |
1917 | struct bkey *replace_key) | |
cafe5635 KO |
1918 | { |
1919 | bool ret = false; | |
cafe5635 KO |
1920 | unsigned oldsize = bch_count_data(b); |
1921 | ||
26c949f8 | 1922 | while (!bch_keylist_empty(insert_keys)) { |
403b6cde | 1923 | struct bset *i = write_block(b); |
c2f95ae2 | 1924 | struct bkey *k = insert_keys->keys; |
26c949f8 | 1925 | |
403b6cde KO |
1926 | if (b->written + __set_blocks(i, i->keys + bkey_u64s(k), b->c) |
1927 | > btree_blocks(b)) | |
1928 | break; | |
1929 | ||
1930 | if (bkey_cmp(k, &b->key) <= 0) { | |
26c949f8 KO |
1931 | bkey_put(b->c, k, b->level); |
1932 | ||
1b207d80 | 1933 | ret |= btree_insert_key(b, op, k, replace_key); |
26c949f8 KO |
1934 | bch_keylist_pop_front(insert_keys); |
1935 | } else if (bkey_cmp(&START_KEY(k), &b->key) < 0) { | |
1936 | #if 0 | |
1b207d80 | 1937 | if (replace_key) { |
26c949f8 KO |
1938 | bkey_put(b->c, k, b->level); |
1939 | bch_keylist_pop_front(insert_keys); | |
1940 | op->insert_collision = true; | |
1941 | break; | |
1942 | } | |
1943 | #endif | |
1944 | BKEY_PADDED(key) temp; | |
c2f95ae2 | 1945 | bkey_copy(&temp.key, insert_keys->keys); |
26c949f8 KO |
1946 | |
1947 | bch_cut_back(&b->key, &temp.key); | |
c2f95ae2 | 1948 | bch_cut_front(&b->key, insert_keys->keys); |
26c949f8 | 1949 | |
1b207d80 | 1950 | ret |= btree_insert_key(b, op, &temp.key, replace_key); |
26c949f8 KO |
1951 | break; |
1952 | } else { | |
1953 | break; | |
1954 | } | |
cafe5635 KO |
1955 | } |
1956 | ||
403b6cde KO |
1957 | BUG_ON(!bch_keylist_empty(insert_keys) && b->level); |
1958 | ||
cafe5635 KO |
1959 | BUG_ON(bch_count_data(b) < oldsize); |
1960 | return ret; | |
1961 | } | |
1962 | ||
26c949f8 KO |
1963 | static int btree_split(struct btree *b, struct btree_op *op, |
1964 | struct keylist *insert_keys, | |
1b207d80 KO |
1965 | struct keylist *parent_keys, |
1966 | struct bkey *replace_key) | |
cafe5635 | 1967 | { |
d6fd3b11 | 1968 | bool split; |
cafe5635 KO |
1969 | struct btree *n1, *n2 = NULL, *n3 = NULL; |
1970 | uint64_t start_time = local_clock(); | |
b54d6934 KO |
1971 | struct closure cl; |
1972 | ||
1973 | closure_init_stack(&cl); | |
cafe5635 | 1974 | |
35fcd848 | 1975 | n1 = btree_node_alloc_replacement(b); |
cafe5635 KO |
1976 | if (IS_ERR(n1)) |
1977 | goto err; | |
1978 | ||
1979 | split = set_blocks(n1->sets[0].data, n1->c) > (btree_blocks(b) * 4) / 5; | |
1980 | ||
cafe5635 KO |
1981 | if (split) { |
1982 | unsigned keys = 0; | |
1983 | ||
c37511b8 KO |
1984 | trace_bcache_btree_node_split(b, n1->sets[0].data->keys); |
1985 | ||
35fcd848 | 1986 | n2 = bch_btree_node_alloc(b->c, b->level); |
cafe5635 KO |
1987 | if (IS_ERR(n2)) |
1988 | goto err_free1; | |
1989 | ||
d6fd3b11 | 1990 | if (!b->parent) { |
35fcd848 | 1991 | n3 = bch_btree_node_alloc(b->c, b->level + 1); |
cafe5635 KO |
1992 | if (IS_ERR(n3)) |
1993 | goto err_free2; | |
1994 | } | |
1995 | ||
1b207d80 | 1996 | bch_btree_insert_keys(n1, op, insert_keys, replace_key); |
cafe5635 | 1997 | |
d6fd3b11 KO |
1998 | /* |
1999 | * Has to be a linear search because we don't have an auxiliary | |
cafe5635 KO |
2000 | * search tree yet |
2001 | */ | |
2002 | ||
2003 | while (keys < (n1->sets[0].data->keys * 3) / 5) | |
2004 | keys += bkey_u64s(node(n1->sets[0].data, keys)); | |
2005 | ||
2006 | bkey_copy_key(&n1->key, node(n1->sets[0].data, keys)); | |
2007 | keys += bkey_u64s(node(n1->sets[0].data, keys)); | |
2008 | ||
2009 | n2->sets[0].data->keys = n1->sets[0].data->keys - keys; | |
2010 | n1->sets[0].data->keys = keys; | |
2011 | ||
2012 | memcpy(n2->sets[0].data->start, | |
2013 | end(n1->sets[0].data), | |
2014 | n2->sets[0].data->keys * sizeof(uint64_t)); | |
2015 | ||
2016 | bkey_copy_key(&n2->key, &b->key); | |
2017 | ||
26c949f8 | 2018 | bch_keylist_add(parent_keys, &n2->key); |
b54d6934 | 2019 | bch_btree_node_write(n2, &cl); |
cafe5635 | 2020 | rw_unlock(true, n2); |
c37511b8 KO |
2021 | } else { |
2022 | trace_bcache_btree_node_compact(b, n1->sets[0].data->keys); | |
2023 | ||
1b207d80 | 2024 | bch_btree_insert_keys(n1, op, insert_keys, replace_key); |
c37511b8 | 2025 | } |
cafe5635 | 2026 | |
26c949f8 | 2027 | bch_keylist_add(parent_keys, &n1->key); |
b54d6934 | 2028 | bch_btree_node_write(n1, &cl); |
cafe5635 KO |
2029 | |
2030 | if (n3) { | |
d6fd3b11 KO |
2031 | /* Depth increases, make a new root */ |
2032 | ||
cafe5635 | 2033 | bkey_copy_key(&n3->key, &MAX_KEY); |
1b207d80 | 2034 | bch_btree_insert_keys(n3, op, parent_keys, NULL); |
b54d6934 | 2035 | bch_btree_node_write(n3, &cl); |
cafe5635 | 2036 | |
b54d6934 | 2037 | closure_sync(&cl); |
cafe5635 KO |
2038 | bch_btree_set_root(n3); |
2039 | rw_unlock(true, n3); | |
d6fd3b11 KO |
2040 | } else if (!b->parent) { |
2041 | /* Root filled up but didn't need to be split */ | |
2042 | ||
c2f95ae2 | 2043 | bch_keylist_reset(parent_keys); |
b54d6934 | 2044 | closure_sync(&cl); |
cafe5635 KO |
2045 | bch_btree_set_root(n1); |
2046 | } else { | |
2047 | unsigned i; | |
2048 | ||
26c949f8 KO |
2049 | bkey_copy(parent_keys->top, &b->key); |
2050 | bkey_copy_key(parent_keys->top, &ZERO_KEY); | |
cafe5635 KO |
2051 | |
2052 | for (i = 0; i < KEY_PTRS(&b->key); i++) { | |
2053 | uint8_t g = PTR_BUCKET(b->c, &b->key, i)->gen + 1; | |
2054 | ||
26c949f8 | 2055 | SET_PTR_GEN(parent_keys->top, i, g); |
cafe5635 KO |
2056 | } |
2057 | ||
26c949f8 | 2058 | bch_keylist_push(parent_keys); |
b54d6934 | 2059 | closure_sync(&cl); |
cafe5635 KO |
2060 | atomic_inc(&b->c->prio_blocked); |
2061 | } | |
2062 | ||
2063 | rw_unlock(true, n1); | |
e8e1d468 | 2064 | btree_node_free(b); |
cafe5635 | 2065 | |
169ef1cf | 2066 | bch_time_stats_update(&b->c->btree_split_time, start_time); |
cafe5635 KO |
2067 | |
2068 | return 0; | |
2069 | err_free2: | |
2070 | __bkey_put(n2->c, &n2->key); | |
e8e1d468 | 2071 | btree_node_free(n2); |
cafe5635 KO |
2072 | rw_unlock(true, n2); |
2073 | err_free1: | |
2074 | __bkey_put(n1->c, &n1->key); | |
e8e1d468 | 2075 | btree_node_free(n1); |
cafe5635 KO |
2076 | rw_unlock(true, n1); |
2077 | err: | |
2078 | if (n3 == ERR_PTR(-EAGAIN) || | |
2079 | n2 == ERR_PTR(-EAGAIN) || | |
2080 | n1 == ERR_PTR(-EAGAIN)) | |
2081 | return -EAGAIN; | |
2082 | ||
2083 | pr_warn("couldn't split"); | |
2084 | return -ENOMEM; | |
2085 | } | |
2086 | ||
26c949f8 | 2087 | static int bch_btree_insert_node(struct btree *b, struct btree_op *op, |
c18536a7 | 2088 | struct keylist *insert_keys, |
1b207d80 KO |
2089 | atomic_t *journal_ref, |
2090 | struct bkey *replace_key) | |
cafe5635 | 2091 | { |
26c949f8 KO |
2092 | int ret = 0; |
2093 | struct keylist split_keys; | |
cafe5635 | 2094 | |
26c949f8 | 2095 | bch_keylist_init(&split_keys); |
cafe5635 | 2096 | |
26c949f8 | 2097 | BUG_ON(b->level); |
cafe5635 | 2098 | |
26c949f8 | 2099 | do { |
1b207d80 KO |
2100 | BUG_ON(b->level && replace_key); |
2101 | ||
26c949f8 KO |
2102 | if (should_split(b)) { |
2103 | if (current->bio_list) { | |
2104 | op->lock = b->c->root->level + 1; | |
2105 | ret = -EAGAIN; | |
2106 | } else if (op->lock <= b->c->root->level) { | |
2107 | op->lock = b->c->root->level + 1; | |
2108 | ret = -EINTR; | |
2109 | } else { | |
2110 | struct btree *parent = b->parent; | |
cafe5635 | 2111 | |
26c949f8 | 2112 | ret = btree_split(b, op, insert_keys, |
1b207d80 | 2113 | &split_keys, replace_key); |
26c949f8 | 2114 | insert_keys = &split_keys; |
1b207d80 | 2115 | replace_key = NULL; |
26c949f8 | 2116 | b = parent; |
403b6cde KO |
2117 | if (!ret) |
2118 | ret = -EINTR; | |
cafe5635 | 2119 | } |
26c949f8 KO |
2120 | } else { |
2121 | BUG_ON(write_block(b) != b->sets[b->nsets].data); | |
cafe5635 | 2122 | |
1b207d80 KO |
2123 | if (bch_btree_insert_keys(b, op, insert_keys, |
2124 | replace_key)) { | |
b54d6934 | 2125 | if (!b->level) { |
c18536a7 | 2126 | bch_btree_leaf_dirty(b, journal_ref); |
b54d6934 KO |
2127 | } else { |
2128 | struct closure cl; | |
2129 | ||
2130 | closure_init_stack(&cl); | |
2131 | bch_btree_node_write(b, &cl); | |
2132 | closure_sync(&cl); | |
2133 | } | |
26c949f8 | 2134 | } |
cafe5635 | 2135 | } |
26c949f8 | 2136 | } while (!bch_keylist_empty(&split_keys)); |
cafe5635 | 2137 | |
26c949f8 KO |
2138 | return ret; |
2139 | } | |
cafe5635 | 2140 | |
e7c590eb KO |
2141 | int bch_btree_insert_check_key(struct btree *b, struct btree_op *op, |
2142 | struct bkey *check_key) | |
2143 | { | |
2144 | int ret = -EINTR; | |
2145 | uint64_t btree_ptr = b->key.ptr[0]; | |
2146 | unsigned long seq = b->seq; | |
2147 | struct keylist insert; | |
2148 | bool upgrade = op->lock == -1; | |
2149 | ||
2150 | bch_keylist_init(&insert); | |
2151 | ||
2152 | if (upgrade) { | |
2153 | rw_unlock(false, b); | |
2154 | rw_lock(true, b, b->level); | |
2155 | ||
2156 | if (b->key.ptr[0] != btree_ptr || | |
2157 | b->seq != seq + 1) | |
2158 | goto out; | |
2159 | } | |
2160 | ||
2161 | SET_KEY_PTRS(check_key, 1); | |
2162 | get_random_bytes(&check_key->ptr[0], sizeof(uint64_t)); | |
2163 | ||
2164 | SET_PTR_DEV(check_key, 0, PTR_CHECK_DEV); | |
2165 | ||
2166 | bch_keylist_add(&insert, check_key); | |
2167 | ||
1b207d80 | 2168 | ret = bch_btree_insert_node(b, op, &insert, NULL, NULL); |
e7c590eb KO |
2169 | |
2170 | BUG_ON(!ret && !bch_keylist_empty(&insert)); | |
2171 | out: | |
2172 | if (upgrade) | |
2173 | downgrade_write(&b->lock); | |
2174 | return ret; | |
2175 | } | |
2176 | ||
cc7b8819 KO |
2177 | struct btree_insert_op { |
2178 | struct btree_op op; | |
2179 | struct keylist *keys; | |
2180 | atomic_t *journal_ref; | |
2181 | struct bkey *replace_key; | |
2182 | }; | |
cafe5635 | 2183 | |
cc7b8819 KO |
2184 | int btree_insert_fn(struct btree_op *b_op, struct btree *b) |
2185 | { | |
2186 | struct btree_insert_op *op = container_of(b_op, | |
2187 | struct btree_insert_op, op); | |
cafe5635 | 2188 | |
cc7b8819 KO |
2189 | int ret = bch_btree_insert_node(b, &op->op, op->keys, |
2190 | op->journal_ref, op->replace_key); | |
2191 | if (ret && !bch_keylist_empty(op->keys)) | |
2192 | return ret; | |
2193 | else | |
2194 | return MAP_DONE; | |
cafe5635 KO |
2195 | } |
2196 | ||
cc7b8819 KO |
2197 | int bch_btree_insert(struct cache_set *c, struct keylist *keys, |
2198 | atomic_t *journal_ref, struct bkey *replace_key) | |
cafe5635 | 2199 | { |
cc7b8819 | 2200 | struct btree_insert_op op; |
cafe5635 | 2201 | int ret = 0; |
cafe5635 | 2202 | |
cc7b8819 | 2203 | BUG_ON(current->bio_list); |
4f3d4014 | 2204 | BUG_ON(bch_keylist_empty(keys)); |
cafe5635 | 2205 | |
cc7b8819 KO |
2206 | bch_btree_op_init(&op.op, 0); |
2207 | op.keys = keys; | |
2208 | op.journal_ref = journal_ref; | |
2209 | op.replace_key = replace_key; | |
cafe5635 | 2210 | |
cc7b8819 KO |
2211 | while (!ret && !bch_keylist_empty(keys)) { |
2212 | op.op.lock = 0; | |
2213 | ret = bch_btree_map_leaf_nodes(&op.op, c, | |
2214 | &START_KEY(keys->keys), | |
2215 | btree_insert_fn); | |
2216 | } | |
cafe5635 | 2217 | |
cc7b8819 KO |
2218 | if (ret) { |
2219 | struct bkey *k; | |
cafe5635 | 2220 | |
cc7b8819 | 2221 | pr_err("error %i", ret); |
cafe5635 | 2222 | |
cc7b8819 KO |
2223 | while ((k = bch_keylist_pop(keys))) |
2224 | bkey_put(c, k, 0); | |
2225 | } else if (op.op.insert_collision) | |
2226 | ret = -ESRCH; | |
6054c6d4 | 2227 | |
cafe5635 KO |
2228 | return ret; |
2229 | } | |
2230 | ||
2231 | void bch_btree_set_root(struct btree *b) | |
2232 | { | |
2233 | unsigned i; | |
e49c7c37 KO |
2234 | struct closure cl; |
2235 | ||
2236 | closure_init_stack(&cl); | |
cafe5635 | 2237 | |
c37511b8 KO |
2238 | trace_bcache_btree_set_root(b); |
2239 | ||
cafe5635 KO |
2240 | BUG_ON(!b->written); |
2241 | ||
2242 | for (i = 0; i < KEY_PTRS(&b->key); i++) | |
2243 | BUG_ON(PTR_BUCKET(b->c, &b->key, i)->prio != BTREE_PRIO); | |
2244 | ||
2245 | mutex_lock(&b->c->bucket_lock); | |
2246 | list_del_init(&b->list); | |
2247 | mutex_unlock(&b->c->bucket_lock); | |
2248 | ||
2249 | b->c->root = b; | |
2250 | __bkey_put(b->c, &b->key); | |
2251 | ||
e49c7c37 KO |
2252 | bch_journal_meta(b->c, &cl); |
2253 | closure_sync(&cl); | |
cafe5635 KO |
2254 | } |
2255 | ||
48dad8ba KO |
2256 | /* Map across nodes or keys */ |
2257 | ||
2258 | static int bch_btree_map_nodes_recurse(struct btree *b, struct btree_op *op, | |
2259 | struct bkey *from, | |
2260 | btree_map_nodes_fn *fn, int flags) | |
2261 | { | |
2262 | int ret = MAP_CONTINUE; | |
2263 | ||
2264 | if (b->level) { | |
2265 | struct bkey *k; | |
2266 | struct btree_iter iter; | |
2267 | ||
2268 | bch_btree_iter_init(b, &iter, from); | |
2269 | ||
2270 | while ((k = bch_btree_iter_next_filter(&iter, b, | |
2271 | bch_ptr_bad))) { | |
2272 | ret = btree(map_nodes_recurse, k, b, | |
2273 | op, from, fn, flags); | |
2274 | from = NULL; | |
2275 | ||
2276 | if (ret != MAP_CONTINUE) | |
2277 | return ret; | |
2278 | } | |
2279 | } | |
2280 | ||
2281 | if (!b->level || flags == MAP_ALL_NODES) | |
2282 | ret = fn(op, b); | |
2283 | ||
2284 | return ret; | |
2285 | } | |
2286 | ||
2287 | int __bch_btree_map_nodes(struct btree_op *op, struct cache_set *c, | |
2288 | struct bkey *from, btree_map_nodes_fn *fn, int flags) | |
2289 | { | |
b54d6934 | 2290 | return btree_root(map_nodes_recurse, c, op, from, fn, flags); |
48dad8ba KO |
2291 | } |
2292 | ||
2293 | static int bch_btree_map_keys_recurse(struct btree *b, struct btree_op *op, | |
2294 | struct bkey *from, btree_map_keys_fn *fn, | |
2295 | int flags) | |
2296 | { | |
2297 | int ret = MAP_CONTINUE; | |
2298 | struct bkey *k; | |
2299 | struct btree_iter iter; | |
2300 | ||
2301 | bch_btree_iter_init(b, &iter, from); | |
2302 | ||
2303 | while ((k = bch_btree_iter_next_filter(&iter, b, bch_ptr_bad))) { | |
2304 | ret = !b->level | |
2305 | ? fn(op, b, k) | |
2306 | : btree(map_keys_recurse, k, b, op, from, fn, flags); | |
2307 | from = NULL; | |
2308 | ||
2309 | if (ret != MAP_CONTINUE) | |
2310 | return ret; | |
2311 | } | |
2312 | ||
2313 | if (!b->level && (flags & MAP_END_KEY)) | |
2314 | ret = fn(op, b, &KEY(KEY_INODE(&b->key), | |
2315 | KEY_OFFSET(&b->key), 0)); | |
2316 | ||
2317 | return ret; | |
2318 | } | |
2319 | ||
2320 | int bch_btree_map_keys(struct btree_op *op, struct cache_set *c, | |
2321 | struct bkey *from, btree_map_keys_fn *fn, int flags) | |
2322 | { | |
b54d6934 | 2323 | return btree_root(map_keys_recurse, c, op, from, fn, flags); |
48dad8ba KO |
2324 | } |
2325 | ||
cafe5635 KO |
2326 | /* Keybuf code */ |
2327 | ||
2328 | static inline int keybuf_cmp(struct keybuf_key *l, struct keybuf_key *r) | |
2329 | { | |
2330 | /* Overlapping keys compare equal */ | |
2331 | if (bkey_cmp(&l->key, &START_KEY(&r->key)) <= 0) | |
2332 | return -1; | |
2333 | if (bkey_cmp(&START_KEY(&l->key), &r->key) >= 0) | |
2334 | return 1; | |
2335 | return 0; | |
2336 | } | |
2337 | ||
2338 | static inline int keybuf_nonoverlapping_cmp(struct keybuf_key *l, | |
2339 | struct keybuf_key *r) | |
2340 | { | |
2341 | return clamp_t(int64_t, bkey_cmp(&l->key, &r->key), -1, 1); | |
2342 | } | |
2343 | ||
48dad8ba KO |
2344 | struct refill { |
2345 | struct btree_op op; | |
2346 | struct keybuf *buf; | |
2347 | struct bkey *end; | |
2348 | keybuf_pred_fn *pred; | |
2349 | }; | |
cafe5635 | 2350 | |
48dad8ba KO |
2351 | static int refill_keybuf_fn(struct btree_op *op, struct btree *b, |
2352 | struct bkey *k) | |
2353 | { | |
2354 | struct refill *refill = container_of(op, struct refill, op); | |
2355 | struct keybuf *buf = refill->buf; | |
2356 | int ret = MAP_CONTINUE; | |
cafe5635 | 2357 | |
48dad8ba KO |
2358 | if (bkey_cmp(k, refill->end) >= 0) { |
2359 | ret = MAP_DONE; | |
2360 | goto out; | |
2361 | } | |
cafe5635 | 2362 | |
48dad8ba KO |
2363 | if (!KEY_SIZE(k)) /* end key */ |
2364 | goto out; | |
cafe5635 | 2365 | |
48dad8ba KO |
2366 | if (refill->pred(buf, k)) { |
2367 | struct keybuf_key *w; | |
cafe5635 | 2368 | |
48dad8ba | 2369 | spin_lock(&buf->lock); |
cafe5635 | 2370 | |
48dad8ba KO |
2371 | w = array_alloc(&buf->freelist); |
2372 | if (!w) { | |
2373 | spin_unlock(&buf->lock); | |
2374 | return MAP_DONE; | |
2375 | } | |
cafe5635 | 2376 | |
48dad8ba KO |
2377 | w->private = NULL; |
2378 | bkey_copy(&w->key, k); | |
cafe5635 | 2379 | |
48dad8ba KO |
2380 | if (RB_INSERT(&buf->keys, w, node, keybuf_cmp)) |
2381 | array_free(&buf->freelist, w); | |
cafe5635 | 2382 | |
48dad8ba KO |
2383 | if (array_freelist_empty(&buf->freelist)) |
2384 | ret = MAP_DONE; | |
cafe5635 | 2385 | |
48dad8ba | 2386 | spin_unlock(&buf->lock); |
cafe5635 | 2387 | } |
48dad8ba KO |
2388 | out: |
2389 | buf->last_scanned = *k; | |
2390 | return ret; | |
cafe5635 KO |
2391 | } |
2392 | ||
2393 | void bch_refill_keybuf(struct cache_set *c, struct keybuf *buf, | |
72c27061 | 2394 | struct bkey *end, keybuf_pred_fn *pred) |
cafe5635 KO |
2395 | { |
2396 | struct bkey start = buf->last_scanned; | |
48dad8ba | 2397 | struct refill refill; |
cafe5635 KO |
2398 | |
2399 | cond_resched(); | |
2400 | ||
b54d6934 | 2401 | bch_btree_op_init(&refill.op, -1); |
48dad8ba KO |
2402 | refill.buf = buf; |
2403 | refill.end = end; | |
2404 | refill.pred = pred; | |
2405 | ||
2406 | bch_btree_map_keys(&refill.op, c, &buf->last_scanned, | |
2407 | refill_keybuf_fn, MAP_END_KEY); | |
cafe5635 KO |
2408 | |
2409 | pr_debug("found %s keys from %llu:%llu to %llu:%llu", | |
2410 | RB_EMPTY_ROOT(&buf->keys) ? "no" : | |
2411 | array_freelist_empty(&buf->freelist) ? "some" : "a few", | |
2412 | KEY_INODE(&start), KEY_OFFSET(&start), | |
2413 | KEY_INODE(&buf->last_scanned), KEY_OFFSET(&buf->last_scanned)); | |
2414 | ||
2415 | spin_lock(&buf->lock); | |
2416 | ||
2417 | if (!RB_EMPTY_ROOT(&buf->keys)) { | |
2418 | struct keybuf_key *w; | |
2419 | w = RB_FIRST(&buf->keys, struct keybuf_key, node); | |
2420 | buf->start = START_KEY(&w->key); | |
2421 | ||
2422 | w = RB_LAST(&buf->keys, struct keybuf_key, node); | |
2423 | buf->end = w->key; | |
2424 | } else { | |
2425 | buf->start = MAX_KEY; | |
2426 | buf->end = MAX_KEY; | |
2427 | } | |
2428 | ||
2429 | spin_unlock(&buf->lock); | |
2430 | } | |
2431 | ||
2432 | static void __bch_keybuf_del(struct keybuf *buf, struct keybuf_key *w) | |
2433 | { | |
2434 | rb_erase(&w->node, &buf->keys); | |
2435 | array_free(&buf->freelist, w); | |
2436 | } | |
2437 | ||
2438 | void bch_keybuf_del(struct keybuf *buf, struct keybuf_key *w) | |
2439 | { | |
2440 | spin_lock(&buf->lock); | |
2441 | __bch_keybuf_del(buf, w); | |
2442 | spin_unlock(&buf->lock); | |
2443 | } | |
2444 | ||
2445 | bool bch_keybuf_check_overlapping(struct keybuf *buf, struct bkey *start, | |
2446 | struct bkey *end) | |
2447 | { | |
2448 | bool ret = false; | |
2449 | struct keybuf_key *p, *w, s; | |
2450 | s.key = *start; | |
2451 | ||
2452 | if (bkey_cmp(end, &buf->start) <= 0 || | |
2453 | bkey_cmp(start, &buf->end) >= 0) | |
2454 | return false; | |
2455 | ||
2456 | spin_lock(&buf->lock); | |
2457 | w = RB_GREATER(&buf->keys, s, node, keybuf_nonoverlapping_cmp); | |
2458 | ||
2459 | while (w && bkey_cmp(&START_KEY(&w->key), end) < 0) { | |
2460 | p = w; | |
2461 | w = RB_NEXT(w, node); | |
2462 | ||
2463 | if (p->private) | |
2464 | ret = true; | |
2465 | else | |
2466 | __bch_keybuf_del(buf, p); | |
2467 | } | |
2468 | ||
2469 | spin_unlock(&buf->lock); | |
2470 | return ret; | |
2471 | } | |
2472 | ||
2473 | struct keybuf_key *bch_keybuf_next(struct keybuf *buf) | |
2474 | { | |
2475 | struct keybuf_key *w; | |
2476 | spin_lock(&buf->lock); | |
2477 | ||
2478 | w = RB_FIRST(&buf->keys, struct keybuf_key, node); | |
2479 | ||
2480 | while (w && w->private) | |
2481 | w = RB_NEXT(w, node); | |
2482 | ||
2483 | if (w) | |
2484 | w->private = ERR_PTR(-EINTR); | |
2485 | ||
2486 | spin_unlock(&buf->lock); | |
2487 | return w; | |
2488 | } | |
2489 | ||
2490 | struct keybuf_key *bch_keybuf_next_rescan(struct cache_set *c, | |
48dad8ba KO |
2491 | struct keybuf *buf, |
2492 | struct bkey *end, | |
2493 | keybuf_pred_fn *pred) | |
cafe5635 KO |
2494 | { |
2495 | struct keybuf_key *ret; | |
2496 | ||
2497 | while (1) { | |
2498 | ret = bch_keybuf_next(buf); | |
2499 | if (ret) | |
2500 | break; | |
2501 | ||
2502 | if (bkey_cmp(&buf->last_scanned, end) >= 0) { | |
2503 | pr_debug("scan finished"); | |
2504 | break; | |
2505 | } | |
2506 | ||
72c27061 | 2507 | bch_refill_keybuf(c, buf, end, pred); |
cafe5635 KO |
2508 | } |
2509 | ||
2510 | return ret; | |
2511 | } | |
2512 | ||
72c27061 | 2513 | void bch_keybuf_init(struct keybuf *buf) |
cafe5635 | 2514 | { |
cafe5635 KO |
2515 | buf->last_scanned = MAX_KEY; |
2516 | buf->keys = RB_ROOT; | |
2517 | ||
2518 | spin_lock_init(&buf->lock); | |
2519 | array_allocator_init(&buf->freelist); | |
2520 | } | |
2521 | ||
2522 | void bch_btree_exit(void) | |
2523 | { | |
2524 | if (btree_io_wq) | |
2525 | destroy_workqueue(btree_io_wq); | |
cafe5635 KO |
2526 | } |
2527 | ||
2528 | int __init bch_btree_init(void) | |
2529 | { | |
72a44517 KO |
2530 | btree_io_wq = create_singlethread_workqueue("bch_btree_io"); |
2531 | if (!btree_io_wq) | |
cafe5635 KO |
2532 | return -ENOMEM; |
2533 | ||
2534 | return 0; | |
2535 | } |