Commit | Line | Data |
---|---|---|
cafe5635 KO |
1 | /* |
2 | * Main bcache entry point - handle a read or a write request and decide what to | |
3 | * do with it; the make_request functions are called by the block layer. | |
4 | * | |
5 | * Copyright 2010, 2011 Kent Overstreet <kent.overstreet@gmail.com> | |
6 | * Copyright 2012 Google, Inc. | |
7 | */ | |
8 | ||
9 | #include "bcache.h" | |
10 | #include "btree.h" | |
11 | #include "debug.h" | |
12 | #include "request.h" | |
279afbad | 13 | #include "writeback.h" |
cafe5635 KO |
14 | |
15 | #include <linux/cgroup.h> | |
16 | #include <linux/module.h> | |
17 | #include <linux/hash.h> | |
18 | #include <linux/random.h> | |
19 | #include "blk-cgroup.h" | |
20 | ||
21 | #include <trace/events/bcache.h> | |
22 | ||
23 | #define CUTOFF_CACHE_ADD 95 | |
24 | #define CUTOFF_CACHE_READA 90 | |
cafe5635 KO |
25 | |
26 | struct kmem_cache *bch_search_cache; | |
27 | ||
cafe5635 KO |
28 | /* Cgroup interface */ |
29 | ||
30 | #ifdef CONFIG_CGROUP_BCACHE | |
31 | static struct bch_cgroup bcache_default_cgroup = { .cache_mode = -1 }; | |
32 | ||
33 | static struct bch_cgroup *cgroup_to_bcache(struct cgroup *cgroup) | |
34 | { | |
35 | struct cgroup_subsys_state *css; | |
36 | return cgroup && | |
37 | (css = cgroup_subsys_state(cgroup, bcache_subsys_id)) | |
38 | ? container_of(css, struct bch_cgroup, css) | |
39 | : &bcache_default_cgroup; | |
40 | } | |
41 | ||
42 | struct bch_cgroup *bch_bio_to_cgroup(struct bio *bio) | |
43 | { | |
44 | struct cgroup_subsys_state *css = bio->bi_css | |
45 | ? cgroup_subsys_state(bio->bi_css->cgroup, bcache_subsys_id) | |
46 | : task_subsys_state(current, bcache_subsys_id); | |
47 | ||
48 | return css | |
49 | ? container_of(css, struct bch_cgroup, css) | |
50 | : &bcache_default_cgroup; | |
51 | } | |
52 | ||
53 | static ssize_t cache_mode_read(struct cgroup *cgrp, struct cftype *cft, | |
54 | struct file *file, | |
55 | char __user *buf, size_t nbytes, loff_t *ppos) | |
56 | { | |
57 | char tmp[1024]; | |
169ef1cf KO |
58 | int len = bch_snprint_string_list(tmp, PAGE_SIZE, bch_cache_modes, |
59 | cgroup_to_bcache(cgrp)->cache_mode + 1); | |
cafe5635 KO |
60 | |
61 | if (len < 0) | |
62 | return len; | |
63 | ||
64 | return simple_read_from_buffer(buf, nbytes, ppos, tmp, len); | |
65 | } | |
66 | ||
67 | static int cache_mode_write(struct cgroup *cgrp, struct cftype *cft, | |
68 | const char *buf) | |
69 | { | |
169ef1cf | 70 | int v = bch_read_string_list(buf, bch_cache_modes); |
cafe5635 KO |
71 | if (v < 0) |
72 | return v; | |
73 | ||
74 | cgroup_to_bcache(cgrp)->cache_mode = v - 1; | |
75 | return 0; | |
76 | } | |
77 | ||
78 | static u64 bch_verify_read(struct cgroup *cgrp, struct cftype *cft) | |
79 | { | |
80 | return cgroup_to_bcache(cgrp)->verify; | |
81 | } | |
82 | ||
83 | static int bch_verify_write(struct cgroup *cgrp, struct cftype *cft, u64 val) | |
84 | { | |
85 | cgroup_to_bcache(cgrp)->verify = val; | |
86 | return 0; | |
87 | } | |
88 | ||
89 | static u64 bch_cache_hits_read(struct cgroup *cgrp, struct cftype *cft) | |
90 | { | |
91 | struct bch_cgroup *bcachecg = cgroup_to_bcache(cgrp); | |
92 | return atomic_read(&bcachecg->stats.cache_hits); | |
93 | } | |
94 | ||
95 | static u64 bch_cache_misses_read(struct cgroup *cgrp, struct cftype *cft) | |
96 | { | |
97 | struct bch_cgroup *bcachecg = cgroup_to_bcache(cgrp); | |
98 | return atomic_read(&bcachecg->stats.cache_misses); | |
99 | } | |
100 | ||
101 | static u64 bch_cache_bypass_hits_read(struct cgroup *cgrp, | |
102 | struct cftype *cft) | |
103 | { | |
104 | struct bch_cgroup *bcachecg = cgroup_to_bcache(cgrp); | |
105 | return atomic_read(&bcachecg->stats.cache_bypass_hits); | |
106 | } | |
107 | ||
108 | static u64 bch_cache_bypass_misses_read(struct cgroup *cgrp, | |
109 | struct cftype *cft) | |
110 | { | |
111 | struct bch_cgroup *bcachecg = cgroup_to_bcache(cgrp); | |
112 | return atomic_read(&bcachecg->stats.cache_bypass_misses); | |
113 | } | |
114 | ||
115 | static struct cftype bch_files[] = { | |
116 | { | |
117 | .name = "cache_mode", | |
118 | .read = cache_mode_read, | |
119 | .write_string = cache_mode_write, | |
120 | }, | |
121 | { | |
122 | .name = "verify", | |
123 | .read_u64 = bch_verify_read, | |
124 | .write_u64 = bch_verify_write, | |
125 | }, | |
126 | { | |
127 | .name = "cache_hits", | |
128 | .read_u64 = bch_cache_hits_read, | |
129 | }, | |
130 | { | |
131 | .name = "cache_misses", | |
132 | .read_u64 = bch_cache_misses_read, | |
133 | }, | |
134 | { | |
135 | .name = "cache_bypass_hits", | |
136 | .read_u64 = bch_cache_bypass_hits_read, | |
137 | }, | |
138 | { | |
139 | .name = "cache_bypass_misses", | |
140 | .read_u64 = bch_cache_bypass_misses_read, | |
141 | }, | |
142 | { } /* terminate */ | |
143 | }; | |
144 | ||
145 | static void init_bch_cgroup(struct bch_cgroup *cg) | |
146 | { | |
147 | cg->cache_mode = -1; | |
148 | } | |
149 | ||
150 | static struct cgroup_subsys_state *bcachecg_create(struct cgroup *cgroup) | |
151 | { | |
152 | struct bch_cgroup *cg; | |
153 | ||
154 | cg = kzalloc(sizeof(*cg), GFP_KERNEL); | |
155 | if (!cg) | |
156 | return ERR_PTR(-ENOMEM); | |
157 | init_bch_cgroup(cg); | |
158 | return &cg->css; | |
159 | } | |
160 | ||
161 | static void bcachecg_destroy(struct cgroup *cgroup) | |
162 | { | |
163 | struct bch_cgroup *cg = cgroup_to_bcache(cgroup); | |
164 | free_css_id(&bcache_subsys, &cg->css); | |
165 | kfree(cg); | |
166 | } | |
167 | ||
168 | struct cgroup_subsys bcache_subsys = { | |
169 | .create = bcachecg_create, | |
170 | .destroy = bcachecg_destroy, | |
171 | .subsys_id = bcache_subsys_id, | |
172 | .name = "bcache", | |
173 | .module = THIS_MODULE, | |
174 | }; | |
175 | EXPORT_SYMBOL_GPL(bcache_subsys); | |
176 | #endif | |
177 | ||
178 | static unsigned cache_mode(struct cached_dev *dc, struct bio *bio) | |
179 | { | |
180 | #ifdef CONFIG_CGROUP_BCACHE | |
181 | int r = bch_bio_to_cgroup(bio)->cache_mode; | |
182 | if (r >= 0) | |
183 | return r; | |
184 | #endif | |
185 | return BDEV_CACHE_MODE(&dc->sb); | |
186 | } | |
187 | ||
188 | static bool verify(struct cached_dev *dc, struct bio *bio) | |
189 | { | |
190 | #ifdef CONFIG_CGROUP_BCACHE | |
191 | if (bch_bio_to_cgroup(bio)->verify) | |
192 | return true; | |
193 | #endif | |
194 | return dc->verify; | |
195 | } | |
196 | ||
197 | static void bio_csum(struct bio *bio, struct bkey *k) | |
198 | { | |
199 | struct bio_vec *bv; | |
200 | uint64_t csum = 0; | |
201 | int i; | |
202 | ||
203 | bio_for_each_segment(bv, bio, i) { | |
204 | void *d = kmap(bv->bv_page) + bv->bv_offset; | |
169ef1cf | 205 | csum = bch_crc64_update(csum, d, bv->bv_len); |
cafe5635 KO |
206 | kunmap(bv->bv_page); |
207 | } | |
208 | ||
209 | k->ptr[KEY_PTRS(k)] = csum & (~0ULL >> 1); | |
210 | } | |
211 | ||
212 | /* Insert data into cache */ | |
213 | ||
214 | static void bio_invalidate(struct closure *cl) | |
215 | { | |
216 | struct btree_op *op = container_of(cl, struct btree_op, cl); | |
217 | struct bio *bio = op->cache_bio; | |
218 | ||
219 | pr_debug("invalidating %i sectors from %llu", | |
220 | bio_sectors(bio), (uint64_t) bio->bi_sector); | |
221 | ||
222 | while (bio_sectors(bio)) { | |
223 | unsigned len = min(bio_sectors(bio), 1U << 14); | |
224 | ||
225 | if (bch_keylist_realloc(&op->keys, 0, op->c)) | |
226 | goto out; | |
227 | ||
228 | bio->bi_sector += len; | |
229 | bio->bi_size -= len << 9; | |
230 | ||
231 | bch_keylist_add(&op->keys, | |
232 | &KEY(op->inode, bio->bi_sector, len)); | |
233 | } | |
234 | ||
235 | op->insert_data_done = true; | |
236 | bio_put(bio); | |
237 | out: | |
238 | continue_at(cl, bch_journal, bcache_wq); | |
239 | } | |
240 | ||
241 | struct open_bucket { | |
242 | struct list_head list; | |
243 | struct task_struct *last; | |
244 | unsigned sectors_free; | |
245 | BKEY_PADDED(key); | |
246 | }; | |
247 | ||
248 | void bch_open_buckets_free(struct cache_set *c) | |
249 | { | |
250 | struct open_bucket *b; | |
251 | ||
252 | while (!list_empty(&c->data_buckets)) { | |
253 | b = list_first_entry(&c->data_buckets, | |
254 | struct open_bucket, list); | |
255 | list_del(&b->list); | |
256 | kfree(b); | |
257 | } | |
258 | } | |
259 | ||
260 | int bch_open_buckets_alloc(struct cache_set *c) | |
261 | { | |
262 | int i; | |
263 | ||
264 | spin_lock_init(&c->data_bucket_lock); | |
265 | ||
266 | for (i = 0; i < 6; i++) { | |
267 | struct open_bucket *b = kzalloc(sizeof(*b), GFP_KERNEL); | |
268 | if (!b) | |
269 | return -ENOMEM; | |
270 | ||
271 | list_add(&b->list, &c->data_buckets); | |
272 | } | |
273 | ||
274 | return 0; | |
275 | } | |
276 | ||
277 | /* | |
278 | * We keep multiple buckets open for writes, and try to segregate different | |
279 | * write streams for better cache utilization: first we look for a bucket where | |
280 | * the last write to it was sequential with the current write, and failing that | |
281 | * we look for a bucket that was last used by the same task. | |
282 | * | |
283 | * The ideas is if you've got multiple tasks pulling data into the cache at the | |
284 | * same time, you'll get better cache utilization if you try to segregate their | |
285 | * data and preserve locality. | |
286 | * | |
287 | * For example, say you've starting Firefox at the same time you're copying a | |
288 | * bunch of files. Firefox will likely end up being fairly hot and stay in the | |
289 | * cache awhile, but the data you copied might not be; if you wrote all that | |
290 | * data to the same buckets it'd get invalidated at the same time. | |
291 | * | |
292 | * Both of those tasks will be doing fairly random IO so we can't rely on | |
293 | * detecting sequential IO to segregate their data, but going off of the task | |
294 | * should be a sane heuristic. | |
295 | */ | |
296 | static struct open_bucket *pick_data_bucket(struct cache_set *c, | |
297 | const struct bkey *search, | |
298 | struct task_struct *task, | |
299 | struct bkey *alloc) | |
300 | { | |
301 | struct open_bucket *ret, *ret_task = NULL; | |
302 | ||
303 | list_for_each_entry_reverse(ret, &c->data_buckets, list) | |
304 | if (!bkey_cmp(&ret->key, search)) | |
305 | goto found; | |
306 | else if (ret->last == task) | |
307 | ret_task = ret; | |
308 | ||
309 | ret = ret_task ?: list_first_entry(&c->data_buckets, | |
310 | struct open_bucket, list); | |
311 | found: | |
312 | if (!ret->sectors_free && KEY_PTRS(alloc)) { | |
313 | ret->sectors_free = c->sb.bucket_size; | |
314 | bkey_copy(&ret->key, alloc); | |
315 | bkey_init(alloc); | |
316 | } | |
317 | ||
318 | if (!ret->sectors_free) | |
319 | ret = NULL; | |
320 | ||
321 | return ret; | |
322 | } | |
323 | ||
324 | /* | |
325 | * Allocates some space in the cache to write to, and k to point to the newly | |
326 | * allocated space, and updates KEY_SIZE(k) and KEY_OFFSET(k) (to point to the | |
327 | * end of the newly allocated space). | |
328 | * | |
329 | * May allocate fewer sectors than @sectors, KEY_SIZE(k) indicates how many | |
330 | * sectors were actually allocated. | |
331 | * | |
332 | * If s->writeback is true, will not fail. | |
333 | */ | |
334 | static bool bch_alloc_sectors(struct bkey *k, unsigned sectors, | |
335 | struct search *s) | |
336 | { | |
337 | struct cache_set *c = s->op.c; | |
338 | struct open_bucket *b; | |
339 | BKEY_PADDED(key) alloc; | |
340 | struct closure cl, *w = NULL; | |
341 | unsigned i; | |
342 | ||
343 | if (s->writeback) { | |
344 | closure_init_stack(&cl); | |
345 | w = &cl; | |
346 | } | |
347 | ||
348 | /* | |
349 | * We might have to allocate a new bucket, which we can't do with a | |
350 | * spinlock held. So if we have to allocate, we drop the lock, allocate | |
351 | * and then retry. KEY_PTRS() indicates whether alloc points to | |
352 | * allocated bucket(s). | |
353 | */ | |
354 | ||
355 | bkey_init(&alloc.key); | |
356 | spin_lock(&c->data_bucket_lock); | |
357 | ||
358 | while (!(b = pick_data_bucket(c, k, s->task, &alloc.key))) { | |
359 | unsigned watermark = s->op.write_prio | |
360 | ? WATERMARK_MOVINGGC | |
361 | : WATERMARK_NONE; | |
362 | ||
363 | spin_unlock(&c->data_bucket_lock); | |
364 | ||
365 | if (bch_bucket_alloc_set(c, watermark, &alloc.key, 1, w)) | |
366 | return false; | |
367 | ||
368 | spin_lock(&c->data_bucket_lock); | |
369 | } | |
370 | ||
371 | /* | |
372 | * If we had to allocate, we might race and not need to allocate the | |
373 | * second time we call find_data_bucket(). If we allocated a bucket but | |
374 | * didn't use it, drop the refcount bch_bucket_alloc_set() took: | |
375 | */ | |
376 | if (KEY_PTRS(&alloc.key)) | |
377 | __bkey_put(c, &alloc.key); | |
378 | ||
379 | for (i = 0; i < KEY_PTRS(&b->key); i++) | |
380 | EBUG_ON(ptr_stale(c, &b->key, i)); | |
381 | ||
382 | /* Set up the pointer to the space we're allocating: */ | |
383 | ||
384 | for (i = 0; i < KEY_PTRS(&b->key); i++) | |
385 | k->ptr[i] = b->key.ptr[i]; | |
386 | ||
387 | sectors = min(sectors, b->sectors_free); | |
388 | ||
389 | SET_KEY_OFFSET(k, KEY_OFFSET(k) + sectors); | |
390 | SET_KEY_SIZE(k, sectors); | |
391 | SET_KEY_PTRS(k, KEY_PTRS(&b->key)); | |
392 | ||
393 | /* | |
394 | * Move b to the end of the lru, and keep track of what this bucket was | |
395 | * last used for: | |
396 | */ | |
397 | list_move_tail(&b->list, &c->data_buckets); | |
398 | bkey_copy_key(&b->key, k); | |
399 | b->last = s->task; | |
400 | ||
401 | b->sectors_free -= sectors; | |
402 | ||
403 | for (i = 0; i < KEY_PTRS(&b->key); i++) { | |
404 | SET_PTR_OFFSET(&b->key, i, PTR_OFFSET(&b->key, i) + sectors); | |
405 | ||
406 | atomic_long_add(sectors, | |
407 | &PTR_CACHE(c, &b->key, i)->sectors_written); | |
408 | } | |
409 | ||
410 | if (b->sectors_free < c->sb.block_size) | |
411 | b->sectors_free = 0; | |
412 | ||
413 | /* | |
414 | * k takes refcounts on the buckets it points to until it's inserted | |
415 | * into the btree, but if we're done with this bucket we just transfer | |
416 | * get_data_bucket()'s refcount. | |
417 | */ | |
418 | if (b->sectors_free) | |
419 | for (i = 0; i < KEY_PTRS(&b->key); i++) | |
420 | atomic_inc(&PTR_BUCKET(c, &b->key, i)->pin); | |
421 | ||
422 | spin_unlock(&c->data_bucket_lock); | |
423 | return true; | |
424 | } | |
425 | ||
426 | static void bch_insert_data_error(struct closure *cl) | |
427 | { | |
428 | struct btree_op *op = container_of(cl, struct btree_op, cl); | |
429 | ||
430 | /* | |
431 | * Our data write just errored, which means we've got a bunch of keys to | |
432 | * insert that point to data that wasn't succesfully written. | |
433 | * | |
434 | * We don't have to insert those keys but we still have to invalidate | |
435 | * that region of the cache - so, if we just strip off all the pointers | |
436 | * from the keys we'll accomplish just that. | |
437 | */ | |
438 | ||
c2f95ae2 | 439 | struct bkey *src = op->keys.keys, *dst = op->keys.keys; |
cafe5635 KO |
440 | |
441 | while (src != op->keys.top) { | |
442 | struct bkey *n = bkey_next(src); | |
443 | ||
444 | SET_KEY_PTRS(src, 0); | |
c2f95ae2 | 445 | memmove(dst, src, bkey_bytes(src)); |
cafe5635 KO |
446 | |
447 | dst = bkey_next(dst); | |
448 | src = n; | |
449 | } | |
450 | ||
451 | op->keys.top = dst; | |
452 | ||
453 | bch_journal(cl); | |
454 | } | |
455 | ||
456 | static void bch_insert_data_endio(struct bio *bio, int error) | |
457 | { | |
458 | struct closure *cl = bio->bi_private; | |
459 | struct btree_op *op = container_of(cl, struct btree_op, cl); | |
460 | struct search *s = container_of(op, struct search, op); | |
461 | ||
462 | if (error) { | |
463 | /* TODO: We could try to recover from this. */ | |
464 | if (s->writeback) | |
465 | s->error = error; | |
466 | else if (s->write) | |
467 | set_closure_fn(cl, bch_insert_data_error, bcache_wq); | |
468 | else | |
469 | set_closure_fn(cl, NULL, NULL); | |
470 | } | |
471 | ||
472 | bch_bbio_endio(op->c, bio, error, "writing data to cache"); | |
473 | } | |
474 | ||
475 | static void bch_insert_data_loop(struct closure *cl) | |
476 | { | |
477 | struct btree_op *op = container_of(cl, struct btree_op, cl); | |
478 | struct search *s = container_of(op, struct search, op); | |
479 | struct bio *bio = op->cache_bio, *n; | |
480 | ||
84f0db03 | 481 | if (op->bypass) |
cafe5635 KO |
482 | return bio_invalidate(cl); |
483 | ||
484 | if (atomic_sub_return(bio_sectors(bio), &op->c->sectors_to_gc) < 0) { | |
485 | set_gc_sectors(op->c); | |
486 | bch_queue_gc(op->c); | |
487 | } | |
488 | ||
54d12f2b KO |
489 | /* |
490 | * Journal writes are marked REQ_FLUSH; if the original write was a | |
491 | * flush, it'll wait on the journal write. | |
492 | */ | |
493 | bio->bi_rw &= ~(REQ_FLUSH|REQ_FUA); | |
494 | ||
cafe5635 KO |
495 | do { |
496 | unsigned i; | |
497 | struct bkey *k; | |
498 | struct bio_set *split = s->d | |
499 | ? s->d->bio_split : op->c->bio_split; | |
500 | ||
501 | /* 1 for the device pointer and 1 for the chksum */ | |
502 | if (bch_keylist_realloc(&op->keys, | |
503 | 1 + (op->csum ? 1 : 0), | |
504 | op->c)) | |
505 | continue_at(cl, bch_journal, bcache_wq); | |
506 | ||
507 | k = op->keys.top; | |
508 | bkey_init(k); | |
509 | SET_KEY_INODE(k, op->inode); | |
510 | SET_KEY_OFFSET(k, bio->bi_sector); | |
511 | ||
512 | if (!bch_alloc_sectors(k, bio_sectors(bio), s)) | |
513 | goto err; | |
514 | ||
515 | n = bch_bio_split(bio, KEY_SIZE(k), GFP_NOIO, split); | |
cafe5635 KO |
516 | |
517 | n->bi_end_io = bch_insert_data_endio; | |
518 | n->bi_private = cl; | |
519 | ||
520 | if (s->writeback) { | |
521 | SET_KEY_DIRTY(k, true); | |
522 | ||
523 | for (i = 0; i < KEY_PTRS(k); i++) | |
524 | SET_GC_MARK(PTR_BUCKET(op->c, k, i), | |
525 | GC_MARK_DIRTY); | |
526 | } | |
527 | ||
528 | SET_KEY_CSUM(k, op->csum); | |
529 | if (KEY_CSUM(k)) | |
530 | bio_csum(n, k); | |
531 | ||
c37511b8 | 532 | trace_bcache_cache_insert(k); |
cafe5635 KO |
533 | bch_keylist_push(&op->keys); |
534 | ||
cafe5635 KO |
535 | n->bi_rw |= REQ_WRITE; |
536 | bch_submit_bbio(n, op->c, k, 0); | |
537 | } while (n != bio); | |
538 | ||
539 | op->insert_data_done = true; | |
540 | continue_at(cl, bch_journal, bcache_wq); | |
541 | err: | |
542 | /* bch_alloc_sectors() blocks if s->writeback = true */ | |
543 | BUG_ON(s->writeback); | |
544 | ||
545 | /* | |
546 | * But if it's not a writeback write we'd rather just bail out if | |
547 | * there aren't any buckets ready to write to - it might take awhile and | |
548 | * we might be starving btree writes for gc or something. | |
549 | */ | |
550 | ||
551 | if (s->write) { | |
552 | /* | |
553 | * Writethrough write: We can't complete the write until we've | |
554 | * updated the index. But we don't want to delay the write while | |
555 | * we wait for buckets to be freed up, so just invalidate the | |
556 | * rest of the write. | |
557 | */ | |
84f0db03 | 558 | op->bypass = true; |
cafe5635 KO |
559 | return bio_invalidate(cl); |
560 | } else { | |
561 | /* | |
562 | * From a cache miss, we can just insert the keys for the data | |
563 | * we have written or bail out if we didn't do anything. | |
564 | */ | |
565 | op->insert_data_done = true; | |
566 | bio_put(bio); | |
567 | ||
568 | if (!bch_keylist_empty(&op->keys)) | |
569 | continue_at(cl, bch_journal, bcache_wq); | |
570 | else | |
571 | closure_return(cl); | |
572 | } | |
573 | } | |
574 | ||
575 | /** | |
576 | * bch_insert_data - stick some data in the cache | |
577 | * | |
578 | * This is the starting point for any data to end up in a cache device; it could | |
579 | * be from a normal write, or a writeback write, or a write to a flash only | |
580 | * volume - it's also used by the moving garbage collector to compact data in | |
581 | * mostly empty buckets. | |
582 | * | |
583 | * It first writes the data to the cache, creating a list of keys to be inserted | |
584 | * (if the data had to be fragmented there will be multiple keys); after the | |
585 | * data is written it calls bch_journal, and after the keys have been added to | |
586 | * the next journal write they're inserted into the btree. | |
587 | * | |
588 | * It inserts the data in op->cache_bio; bi_sector is used for the key offset, | |
589 | * and op->inode is used for the key inode. | |
590 | * | |
84f0db03 KO |
591 | * If op->bypass is true, instead of inserting the data it invalidates the |
592 | * region of the cache represented by op->cache_bio and op->inode. | |
cafe5635 KO |
593 | */ |
594 | void bch_insert_data(struct closure *cl) | |
595 | { | |
596 | struct btree_op *op = container_of(cl, struct btree_op, cl); | |
597 | ||
598 | bch_keylist_init(&op->keys); | |
599 | bio_get(op->cache_bio); | |
600 | bch_insert_data_loop(cl); | |
601 | } | |
602 | ||
603 | void bch_btree_insert_async(struct closure *cl) | |
604 | { | |
605 | struct btree_op *op = container_of(cl, struct btree_op, cl); | |
606 | struct search *s = container_of(op, struct search, op); | |
607 | ||
4f3d4014 | 608 | if (bch_btree_insert(op, op->c, &op->keys)) { |
cafe5635 KO |
609 | s->error = -ENOMEM; |
610 | op->insert_data_done = true; | |
611 | } | |
612 | ||
613 | if (op->insert_data_done) { | |
614 | bch_keylist_free(&op->keys); | |
615 | closure_return(cl); | |
616 | } else | |
617 | continue_at(cl, bch_insert_data_loop, bcache_wq); | |
618 | } | |
619 | ||
620 | /* Common code for the make_request functions */ | |
621 | ||
622 | static void request_endio(struct bio *bio, int error) | |
623 | { | |
624 | struct closure *cl = bio->bi_private; | |
625 | ||
626 | if (error) { | |
627 | struct search *s = container_of(cl, struct search, cl); | |
628 | s->error = error; | |
629 | /* Only cache read errors are recoverable */ | |
630 | s->recoverable = false; | |
631 | } | |
632 | ||
633 | bio_put(bio); | |
634 | closure_put(cl); | |
635 | } | |
636 | ||
637 | void bch_cache_read_endio(struct bio *bio, int error) | |
638 | { | |
639 | struct bbio *b = container_of(bio, struct bbio, bio); | |
640 | struct closure *cl = bio->bi_private; | |
641 | struct search *s = container_of(cl, struct search, cl); | |
642 | ||
643 | /* | |
644 | * If the bucket was reused while our bio was in flight, we might have | |
645 | * read the wrong data. Set s->error but not error so it doesn't get | |
646 | * counted against the cache device, but we'll still reread the data | |
647 | * from the backing device. | |
648 | */ | |
649 | ||
650 | if (error) | |
651 | s->error = error; | |
652 | else if (ptr_stale(s->op.c, &b->key, 0)) { | |
653 | atomic_long_inc(&s->op.c->cache_read_races); | |
654 | s->error = -EINTR; | |
655 | } | |
656 | ||
657 | bch_bbio_endio(s->op.c, bio, error, "reading from cache"); | |
658 | } | |
659 | ||
660 | static void bio_complete(struct search *s) | |
661 | { | |
662 | if (s->orig_bio) { | |
663 | int cpu, rw = bio_data_dir(s->orig_bio); | |
664 | unsigned long duration = jiffies - s->start_time; | |
665 | ||
666 | cpu = part_stat_lock(); | |
667 | part_round_stats(cpu, &s->d->disk->part0); | |
668 | part_stat_add(cpu, &s->d->disk->part0, ticks[rw], duration); | |
669 | part_stat_unlock(); | |
670 | ||
671 | trace_bcache_request_end(s, s->orig_bio); | |
672 | bio_endio(s->orig_bio, s->error); | |
673 | s->orig_bio = NULL; | |
674 | } | |
675 | } | |
676 | ||
677 | static void do_bio_hook(struct search *s) | |
678 | { | |
679 | struct bio *bio = &s->bio.bio; | |
680 | memcpy(bio, s->orig_bio, sizeof(struct bio)); | |
681 | ||
682 | bio->bi_end_io = request_endio; | |
683 | bio->bi_private = &s->cl; | |
684 | atomic_set(&bio->bi_cnt, 3); | |
685 | } | |
686 | ||
687 | static void search_free(struct closure *cl) | |
688 | { | |
689 | struct search *s = container_of(cl, struct search, cl); | |
690 | bio_complete(s); | |
691 | ||
692 | if (s->op.cache_bio) | |
693 | bio_put(s->op.cache_bio); | |
694 | ||
695 | if (s->unaligned_bvec) | |
696 | mempool_free(s->bio.bio.bi_io_vec, s->d->unaligned_bvec); | |
697 | ||
698 | closure_debug_destroy(cl); | |
699 | mempool_free(s, s->d->c->search); | |
700 | } | |
701 | ||
702 | static struct search *search_alloc(struct bio *bio, struct bcache_device *d) | |
703 | { | |
704 | struct bio_vec *bv; | |
705 | struct search *s = mempool_alloc(d->c->search, GFP_NOIO); | |
706 | memset(s, 0, offsetof(struct search, op.keys)); | |
707 | ||
708 | __closure_init(&s->cl, NULL); | |
709 | ||
710 | s->op.inode = d->id; | |
711 | s->op.c = d->c; | |
712 | s->d = d; | |
713 | s->op.lock = -1; | |
714 | s->task = current; | |
715 | s->orig_bio = bio; | |
716 | s->write = (bio->bi_rw & REQ_WRITE) != 0; | |
54d12f2b | 717 | s->op.flush_journal = (bio->bi_rw & (REQ_FLUSH|REQ_FUA)) != 0; |
cafe5635 KO |
718 | s->recoverable = 1; |
719 | s->start_time = jiffies; | |
720 | do_bio_hook(s); | |
721 | ||
722 | if (bio->bi_size != bio_segments(bio) * PAGE_SIZE) { | |
723 | bv = mempool_alloc(d->unaligned_bvec, GFP_NOIO); | |
724 | memcpy(bv, bio_iovec(bio), | |
725 | sizeof(struct bio_vec) * bio_segments(bio)); | |
726 | ||
727 | s->bio.bio.bi_io_vec = bv; | |
728 | s->unaligned_bvec = 1; | |
729 | } | |
730 | ||
731 | return s; | |
732 | } | |
733 | ||
734 | static void btree_read_async(struct closure *cl) | |
735 | { | |
736 | struct btree_op *op = container_of(cl, struct btree_op, cl); | |
737 | ||
738 | int ret = btree_root(search_recurse, op->c, op); | |
739 | ||
740 | if (ret == -EAGAIN) | |
741 | continue_at(cl, btree_read_async, bcache_wq); | |
742 | ||
743 | closure_return(cl); | |
744 | } | |
745 | ||
746 | /* Cached devices */ | |
747 | ||
748 | static void cached_dev_bio_complete(struct closure *cl) | |
749 | { | |
750 | struct search *s = container_of(cl, struct search, cl); | |
751 | struct cached_dev *dc = container_of(s->d, struct cached_dev, disk); | |
752 | ||
753 | search_free(cl); | |
754 | cached_dev_put(dc); | |
755 | } | |
756 | ||
84f0db03 KO |
757 | unsigned bch_get_congested(struct cache_set *c) |
758 | { | |
759 | int i; | |
760 | long rand; | |
761 | ||
762 | if (!c->congested_read_threshold_us && | |
763 | !c->congested_write_threshold_us) | |
764 | return 0; | |
765 | ||
766 | i = (local_clock_us() - c->congested_last_us) / 1024; | |
767 | if (i < 0) | |
768 | return 0; | |
769 | ||
770 | i += atomic_read(&c->congested); | |
771 | if (i >= 0) | |
772 | return 0; | |
773 | ||
774 | i += CONGESTED_MAX; | |
775 | ||
776 | if (i > 0) | |
777 | i = fract_exp_two(i, 6); | |
778 | ||
779 | rand = get_random_int(); | |
780 | i -= bitmap_weight(&rand, BITS_PER_LONG); | |
781 | ||
782 | return i > 0 ? i : 1; | |
783 | } | |
784 | ||
785 | static void add_sequential(struct task_struct *t) | |
786 | { | |
787 | ewma_add(t->sequential_io_avg, | |
788 | t->sequential_io, 8, 0); | |
789 | ||
790 | t->sequential_io = 0; | |
791 | } | |
792 | ||
793 | static struct hlist_head *iohash(struct cached_dev *dc, uint64_t k) | |
794 | { | |
795 | return &dc->io_hash[hash_64(k, RECENT_IO_BITS)]; | |
796 | } | |
797 | ||
798 | static bool check_should_bypass(struct cached_dev *dc, struct search *s) | |
799 | { | |
800 | struct cache_set *c = s->op.c; | |
801 | struct bio *bio = &s->bio.bio; | |
802 | unsigned mode = cache_mode(dc, bio); | |
803 | unsigned sectors, congested = bch_get_congested(c); | |
804 | ||
805 | if (atomic_read(&dc->disk.detaching) || | |
806 | c->gc_stats.in_use > CUTOFF_CACHE_ADD || | |
807 | (bio->bi_rw & REQ_DISCARD)) | |
808 | goto skip; | |
809 | ||
810 | if (mode == CACHE_MODE_NONE || | |
811 | (mode == CACHE_MODE_WRITEAROUND && | |
812 | (bio->bi_rw & REQ_WRITE))) | |
813 | goto skip; | |
814 | ||
815 | if (bio->bi_sector & (c->sb.block_size - 1) || | |
816 | bio_sectors(bio) & (c->sb.block_size - 1)) { | |
817 | pr_debug("skipping unaligned io"); | |
818 | goto skip; | |
819 | } | |
820 | ||
821 | if (!congested && !dc->sequential_cutoff) | |
822 | goto rescale; | |
823 | ||
824 | if (!congested && | |
825 | mode == CACHE_MODE_WRITEBACK && | |
826 | (bio->bi_rw & REQ_WRITE) && | |
827 | (bio->bi_rw & REQ_SYNC)) | |
828 | goto rescale; | |
829 | ||
830 | if (dc->sequential_merge) { | |
831 | struct io *i; | |
832 | ||
833 | spin_lock(&dc->io_lock); | |
834 | ||
835 | hlist_for_each_entry(i, iohash(dc, bio->bi_sector), hash) | |
836 | if (i->last == bio->bi_sector && | |
837 | time_before(jiffies, i->jiffies)) | |
838 | goto found; | |
839 | ||
840 | i = list_first_entry(&dc->io_lru, struct io, lru); | |
841 | ||
842 | add_sequential(s->task); | |
843 | i->sequential = 0; | |
844 | found: | |
845 | if (i->sequential + bio->bi_size > i->sequential) | |
846 | i->sequential += bio->bi_size; | |
847 | ||
848 | i->last = bio_end_sector(bio); | |
849 | i->jiffies = jiffies + msecs_to_jiffies(5000); | |
850 | s->task->sequential_io = i->sequential; | |
851 | ||
852 | hlist_del(&i->hash); | |
853 | hlist_add_head(&i->hash, iohash(dc, i->last)); | |
854 | list_move_tail(&i->lru, &dc->io_lru); | |
855 | ||
856 | spin_unlock(&dc->io_lock); | |
857 | } else { | |
858 | s->task->sequential_io = bio->bi_size; | |
859 | ||
860 | add_sequential(s->task); | |
861 | } | |
862 | ||
863 | sectors = max(s->task->sequential_io, | |
864 | s->task->sequential_io_avg) >> 9; | |
865 | ||
866 | if (dc->sequential_cutoff && | |
867 | sectors >= dc->sequential_cutoff >> 9) { | |
868 | trace_bcache_bypass_sequential(s->orig_bio); | |
869 | goto skip; | |
870 | } | |
871 | ||
872 | if (congested && sectors >= congested) { | |
873 | trace_bcache_bypass_congested(s->orig_bio); | |
874 | goto skip; | |
875 | } | |
876 | ||
877 | rescale: | |
878 | bch_rescale_priorities(c, bio_sectors(bio)); | |
879 | return false; | |
880 | skip: | |
881 | bch_mark_sectors_bypassed(s, bio_sectors(bio)); | |
882 | return true; | |
883 | } | |
884 | ||
cafe5635 KO |
885 | /* Process reads */ |
886 | ||
887 | static void cached_dev_read_complete(struct closure *cl) | |
888 | { | |
889 | struct search *s = container_of(cl, struct search, cl); | |
890 | ||
891 | if (s->op.insert_collision) | |
892 | bch_mark_cache_miss_collision(s); | |
893 | ||
894 | if (s->op.cache_bio) { | |
895 | int i; | |
896 | struct bio_vec *bv; | |
897 | ||
898 | __bio_for_each_segment(bv, s->op.cache_bio, i, 0) | |
899 | __free_page(bv->bv_page); | |
900 | } | |
901 | ||
902 | cached_dev_bio_complete(cl); | |
903 | } | |
904 | ||
905 | static void request_read_error(struct closure *cl) | |
906 | { | |
907 | struct search *s = container_of(cl, struct search, cl); | |
908 | struct bio_vec *bv; | |
909 | int i; | |
910 | ||
911 | if (s->recoverable) { | |
c37511b8 KO |
912 | /* Retry from the backing device: */ |
913 | trace_bcache_read_retry(s->orig_bio); | |
cafe5635 KO |
914 | |
915 | s->error = 0; | |
916 | bv = s->bio.bio.bi_io_vec; | |
917 | do_bio_hook(s); | |
918 | s->bio.bio.bi_io_vec = bv; | |
919 | ||
920 | if (!s->unaligned_bvec) | |
921 | bio_for_each_segment(bv, s->orig_bio, i) | |
922 | bv->bv_offset = 0, bv->bv_len = PAGE_SIZE; | |
923 | else | |
924 | memcpy(s->bio.bio.bi_io_vec, | |
925 | bio_iovec(s->orig_bio), | |
926 | sizeof(struct bio_vec) * | |
927 | bio_segments(s->orig_bio)); | |
928 | ||
929 | /* XXX: invalidate cache */ | |
930 | ||
cafe5635 KO |
931 | closure_bio_submit(&s->bio.bio, &s->cl, s->d); |
932 | } | |
933 | ||
934 | continue_at(cl, cached_dev_read_complete, NULL); | |
935 | } | |
936 | ||
937 | static void request_read_done(struct closure *cl) | |
938 | { | |
939 | struct search *s = container_of(cl, struct search, cl); | |
940 | struct cached_dev *dc = container_of(s->d, struct cached_dev, disk); | |
941 | ||
942 | /* | |
943 | * s->cache_bio != NULL implies that we had a cache miss; cache_bio now | |
944 | * contains data ready to be inserted into the cache. | |
945 | * | |
946 | * First, we copy the data we just read from cache_bio's bounce buffers | |
947 | * to the buffers the original bio pointed to: | |
948 | */ | |
949 | ||
950 | if (s->op.cache_bio) { | |
cafe5635 KO |
951 | bio_reset(s->op.cache_bio); |
952 | s->op.cache_bio->bi_sector = s->cache_miss->bi_sector; | |
953 | s->op.cache_bio->bi_bdev = s->cache_miss->bi_bdev; | |
954 | s->op.cache_bio->bi_size = s->cache_bio_sectors << 9; | |
169ef1cf | 955 | bch_bio_map(s->op.cache_bio, NULL); |
cafe5635 | 956 | |
8e51e414 | 957 | bio_copy_data(s->cache_miss, s->op.cache_bio); |
cafe5635 KO |
958 | |
959 | bio_put(s->cache_miss); | |
960 | s->cache_miss = NULL; | |
961 | } | |
962 | ||
963 | if (verify(dc, &s->bio.bio) && s->recoverable) | |
964 | bch_data_verify(s); | |
965 | ||
966 | bio_complete(s); | |
967 | ||
968 | if (s->op.cache_bio && | |
969 | !test_bit(CACHE_SET_STOPPING, &s->op.c->flags)) { | |
970 | s->op.type = BTREE_REPLACE; | |
971 | closure_call(&s->op.cl, bch_insert_data, NULL, cl); | |
972 | } | |
973 | ||
974 | continue_at(cl, cached_dev_read_complete, NULL); | |
975 | } | |
976 | ||
977 | static void request_read_done_bh(struct closure *cl) | |
978 | { | |
979 | struct search *s = container_of(cl, struct search, cl); | |
980 | struct cached_dev *dc = container_of(s->d, struct cached_dev, disk); | |
981 | ||
84f0db03 KO |
982 | bch_mark_cache_accounting(s, !s->cache_miss, s->op.bypass); |
983 | trace_bcache_read(s->orig_bio, !s->cache_miss, s->op.bypass); | |
cafe5635 KO |
984 | |
985 | if (s->error) | |
986 | continue_at_nobarrier(cl, request_read_error, bcache_wq); | |
987 | else if (s->op.cache_bio || verify(dc, &s->bio.bio)) | |
988 | continue_at_nobarrier(cl, request_read_done, bcache_wq); | |
989 | else | |
990 | continue_at_nobarrier(cl, cached_dev_read_complete, NULL); | |
991 | } | |
992 | ||
993 | static int cached_dev_cache_miss(struct btree *b, struct search *s, | |
994 | struct bio *bio, unsigned sectors) | |
995 | { | |
996 | int ret = 0; | |
e7c590eb | 997 | unsigned reada = 0; |
cafe5635 KO |
998 | struct cached_dev *dc = container_of(s->d, struct cached_dev, disk); |
999 | struct bio *miss; | |
1000 | ||
84f0db03 | 1001 | if (s->cache_miss || s->op.bypass) { |
e7c590eb KO |
1002 | miss = bch_bio_split(bio, sectors, GFP_NOIO, s->d->bio_split); |
1003 | if (miss == bio) | |
1004 | s->op.lookup_done = true; | |
1005 | goto out_submit; | |
1006 | } | |
cafe5635 | 1007 | |
e7c590eb KO |
1008 | if (!(bio->bi_rw & REQ_RAHEAD) && |
1009 | !(bio->bi_rw & REQ_META) && | |
1010 | s->op.c->gc_stats.in_use < CUTOFF_CACHE_READA) | |
1011 | reada = min_t(sector_t, dc->readahead >> 9, | |
1012 | bdev_sectors(bio->bi_bdev) - bio_end_sector(bio)); | |
cafe5635 | 1013 | |
e7c590eb | 1014 | s->cache_bio_sectors = min(sectors, bio_sectors(bio) + reada); |
cafe5635 | 1015 | |
e7c590eb KO |
1016 | s->op.replace = KEY(s->op.inode, bio->bi_sector + |
1017 | s->cache_bio_sectors, s->cache_bio_sectors); | |
1018 | ||
1019 | ret = bch_btree_insert_check_key(b, &s->op, &s->op.replace); | |
1020 | if (ret) | |
1021 | return ret; | |
1022 | ||
1023 | miss = bch_bio_split(bio, sectors, GFP_NOIO, s->d->bio_split); | |
1024 | if (miss == bio) | |
1025 | s->op.lookup_done = true; | |
1026 | else | |
1027 | /* btree_search_recurse()'s btree iterator is no good anymore */ | |
1028 | ret = -EINTR; | |
cafe5635 | 1029 | |
cafe5635 KO |
1030 | s->op.cache_bio = bio_alloc_bioset(GFP_NOWAIT, |
1031 | DIV_ROUND_UP(s->cache_bio_sectors, PAGE_SECTORS), | |
1032 | dc->disk.bio_split); | |
1033 | ||
1034 | if (!s->op.cache_bio) | |
1035 | goto out_submit; | |
1036 | ||
1037 | s->op.cache_bio->bi_sector = miss->bi_sector; | |
1038 | s->op.cache_bio->bi_bdev = miss->bi_bdev; | |
1039 | s->op.cache_bio->bi_size = s->cache_bio_sectors << 9; | |
1040 | ||
1041 | s->op.cache_bio->bi_end_io = request_endio; | |
1042 | s->op.cache_bio->bi_private = &s->cl; | |
1043 | ||
169ef1cf | 1044 | bch_bio_map(s->op.cache_bio, NULL); |
8e51e414 | 1045 | if (bio_alloc_pages(s->op.cache_bio, __GFP_NOWARN|GFP_NOIO)) |
cafe5635 KO |
1046 | goto out_put; |
1047 | ||
1048 | s->cache_miss = miss; | |
1049 | bio_get(s->op.cache_bio); | |
1050 | ||
cafe5635 KO |
1051 | closure_bio_submit(s->op.cache_bio, &s->cl, s->d); |
1052 | ||
1053 | return ret; | |
1054 | out_put: | |
1055 | bio_put(s->op.cache_bio); | |
1056 | s->op.cache_bio = NULL; | |
1057 | out_submit: | |
e7c590eb KO |
1058 | miss->bi_end_io = request_endio; |
1059 | miss->bi_private = &s->cl; | |
cafe5635 KO |
1060 | closure_bio_submit(miss, &s->cl, s->d); |
1061 | return ret; | |
1062 | } | |
1063 | ||
1064 | static void request_read(struct cached_dev *dc, struct search *s) | |
1065 | { | |
1066 | struct closure *cl = &s->cl; | |
1067 | ||
cafe5635 | 1068 | closure_call(&s->op.cl, btree_read_async, NULL, cl); |
cafe5635 KO |
1069 | continue_at(cl, request_read_done_bh, NULL); |
1070 | } | |
1071 | ||
1072 | /* Process writes */ | |
1073 | ||
1074 | static void cached_dev_write_complete(struct closure *cl) | |
1075 | { | |
1076 | struct search *s = container_of(cl, struct search, cl); | |
1077 | struct cached_dev *dc = container_of(s->d, struct cached_dev, disk); | |
1078 | ||
1079 | up_read_non_owner(&dc->writeback_lock); | |
1080 | cached_dev_bio_complete(cl); | |
1081 | } | |
1082 | ||
cafe5635 KO |
1083 | static void request_write(struct cached_dev *dc, struct search *s) |
1084 | { | |
1085 | struct closure *cl = &s->cl; | |
1086 | struct bio *bio = &s->bio.bio; | |
84f0db03 KO |
1087 | struct bkey start = KEY(dc->disk.id, bio->bi_sector, 0); |
1088 | struct bkey end = KEY(dc->disk.id, bio_end_sector(bio), 0); | |
cafe5635 KO |
1089 | |
1090 | bch_keybuf_check_overlapping(&s->op.c->moving_gc_keys, &start, &end); | |
1091 | ||
cafe5635 | 1092 | down_read_non_owner(&dc->writeback_lock); |
cafe5635 | 1093 | if (bch_keybuf_check_overlapping(&dc->writeback_keys, &start, &end)) { |
84f0db03 KO |
1094 | /* |
1095 | * We overlap with some dirty data undergoing background | |
1096 | * writeback, force this write to writeback | |
1097 | */ | |
1098 | s->op.bypass = false; | |
cafe5635 KO |
1099 | s->writeback = true; |
1100 | } | |
1101 | ||
84f0db03 KO |
1102 | /* |
1103 | * Discards aren't _required_ to do anything, so skipping if | |
1104 | * check_overlapping returned true is ok | |
1105 | * | |
1106 | * But check_overlapping drops dirty keys for which io hasn't started, | |
1107 | * so we still want to call it. | |
1108 | */ | |
cafe5635 | 1109 | if (bio->bi_rw & REQ_DISCARD) |
84f0db03 | 1110 | s->op.bypass = true; |
cafe5635 | 1111 | |
72c27061 KO |
1112 | if (should_writeback(dc, s->orig_bio, |
1113 | cache_mode(dc, bio), | |
84f0db03 KO |
1114 | s->op.bypass)) { |
1115 | s->op.bypass = false; | |
72c27061 KO |
1116 | s->writeback = true; |
1117 | } | |
1118 | ||
84f0db03 | 1119 | trace_bcache_write(s->orig_bio, s->writeback, s->op.bypass); |
c37511b8 | 1120 | |
84f0db03 KO |
1121 | if (s->op.bypass) { |
1122 | s->op.cache_bio = s->orig_bio; | |
1123 | bio_get(s->op.cache_bio); | |
cafe5635 | 1124 | |
84f0db03 KO |
1125 | if (!(bio->bi_rw & REQ_DISCARD) || |
1126 | blk_queue_discard(bdev_get_queue(dc->bdev))) | |
1127 | closure_bio_submit(bio, cl, s->d); | |
1128 | } else if (s->writeback) { | |
279afbad | 1129 | bch_writeback_add(dc); |
2fe80d3b | 1130 | s->op.cache_bio = bio; |
e49c7c37 | 1131 | |
c0f04d88 | 1132 | if (bio->bi_rw & REQ_FLUSH) { |
e49c7c37 | 1133 | /* Also need to send a flush to the backing device */ |
d4eddd42 | 1134 | struct bio *flush = bio_alloc_bioset(GFP_NOIO, 0, |
c0f04d88 | 1135 | dc->disk.bio_split); |
e49c7c37 | 1136 | |
c0f04d88 KO |
1137 | flush->bi_rw = WRITE_FLUSH; |
1138 | flush->bi_bdev = bio->bi_bdev; | |
1139 | flush->bi_end_io = request_endio; | |
1140 | flush->bi_private = cl; | |
1141 | ||
1142 | closure_bio_submit(flush, cl, s->d); | |
e49c7c37 | 1143 | } |
84f0db03 KO |
1144 | } else { |
1145 | s->op.cache_bio = bio_clone_bioset(bio, GFP_NOIO, | |
1146 | dc->disk.bio_split); | |
1147 | ||
1148 | closure_bio_submit(bio, cl, s->d); | |
cafe5635 | 1149 | } |
84f0db03 | 1150 | |
cafe5635 KO |
1151 | closure_call(&s->op.cl, bch_insert_data, NULL, cl); |
1152 | continue_at(cl, cached_dev_write_complete, NULL); | |
cafe5635 KO |
1153 | } |
1154 | ||
1155 | static void request_nodata(struct cached_dev *dc, struct search *s) | |
1156 | { | |
1157 | struct closure *cl = &s->cl; | |
1158 | struct bio *bio = &s->bio.bio; | |
1159 | ||
cafe5635 KO |
1160 | if (s->op.flush_journal) |
1161 | bch_journal_meta(s->op.c, cl); | |
1162 | ||
84f0db03 | 1163 | /* If it's a flush, we send the flush to the backing device too */ |
cafe5635 KO |
1164 | closure_bio_submit(bio, cl, s->d); |
1165 | ||
1166 | continue_at(cl, cached_dev_bio_complete, NULL); | |
1167 | } | |
1168 | ||
1169 | /* Cached devices - read & write stuff */ | |
1170 | ||
cafe5635 KO |
1171 | static void cached_dev_make_request(struct request_queue *q, struct bio *bio) |
1172 | { | |
1173 | struct search *s; | |
1174 | struct bcache_device *d = bio->bi_bdev->bd_disk->private_data; | |
1175 | struct cached_dev *dc = container_of(d, struct cached_dev, disk); | |
1176 | int cpu, rw = bio_data_dir(bio); | |
1177 | ||
1178 | cpu = part_stat_lock(); | |
1179 | part_stat_inc(cpu, &d->disk->part0, ios[rw]); | |
1180 | part_stat_add(cpu, &d->disk->part0, sectors[rw], bio_sectors(bio)); | |
1181 | part_stat_unlock(); | |
1182 | ||
1183 | bio->bi_bdev = dc->bdev; | |
2903381f | 1184 | bio->bi_sector += dc->sb.data_offset; |
cafe5635 KO |
1185 | |
1186 | if (cached_dev_get(dc)) { | |
1187 | s = search_alloc(bio, d); | |
1188 | trace_bcache_request_start(s, bio); | |
1189 | ||
84f0db03 | 1190 | if (!bio->bi_size) |
cafe5635 | 1191 | request_nodata(dc, s); |
84f0db03 KO |
1192 | else { |
1193 | s->op.bypass = check_should_bypass(dc, s); | |
1194 | ||
1195 | if (rw) | |
1196 | request_write(dc, s); | |
1197 | else | |
1198 | request_read(dc, s); | |
1199 | } | |
cafe5635 KO |
1200 | } else { |
1201 | if ((bio->bi_rw & REQ_DISCARD) && | |
1202 | !blk_queue_discard(bdev_get_queue(dc->bdev))) | |
1203 | bio_endio(bio, 0); | |
1204 | else | |
1205 | bch_generic_make_request(bio, &d->bio_split_hook); | |
1206 | } | |
1207 | } | |
1208 | ||
1209 | static int cached_dev_ioctl(struct bcache_device *d, fmode_t mode, | |
1210 | unsigned int cmd, unsigned long arg) | |
1211 | { | |
1212 | struct cached_dev *dc = container_of(d, struct cached_dev, disk); | |
1213 | return __blkdev_driver_ioctl(dc->bdev, mode, cmd, arg); | |
1214 | } | |
1215 | ||
1216 | static int cached_dev_congested(void *data, int bits) | |
1217 | { | |
1218 | struct bcache_device *d = data; | |
1219 | struct cached_dev *dc = container_of(d, struct cached_dev, disk); | |
1220 | struct request_queue *q = bdev_get_queue(dc->bdev); | |
1221 | int ret = 0; | |
1222 | ||
1223 | if (bdi_congested(&q->backing_dev_info, bits)) | |
1224 | return 1; | |
1225 | ||
1226 | if (cached_dev_get(dc)) { | |
1227 | unsigned i; | |
1228 | struct cache *ca; | |
1229 | ||
1230 | for_each_cache(ca, d->c, i) { | |
1231 | q = bdev_get_queue(ca->bdev); | |
1232 | ret |= bdi_congested(&q->backing_dev_info, bits); | |
1233 | } | |
1234 | ||
1235 | cached_dev_put(dc); | |
1236 | } | |
1237 | ||
1238 | return ret; | |
1239 | } | |
1240 | ||
1241 | void bch_cached_dev_request_init(struct cached_dev *dc) | |
1242 | { | |
1243 | struct gendisk *g = dc->disk.disk; | |
1244 | ||
1245 | g->queue->make_request_fn = cached_dev_make_request; | |
1246 | g->queue->backing_dev_info.congested_fn = cached_dev_congested; | |
1247 | dc->disk.cache_miss = cached_dev_cache_miss; | |
1248 | dc->disk.ioctl = cached_dev_ioctl; | |
1249 | } | |
1250 | ||
1251 | /* Flash backed devices */ | |
1252 | ||
1253 | static int flash_dev_cache_miss(struct btree *b, struct search *s, | |
1254 | struct bio *bio, unsigned sectors) | |
1255 | { | |
8e51e414 KO |
1256 | struct bio_vec *bv; |
1257 | int i; | |
1258 | ||
cafe5635 KO |
1259 | /* Zero fill bio */ |
1260 | ||
8e51e414 | 1261 | bio_for_each_segment(bv, bio, i) { |
cafe5635 KO |
1262 | unsigned j = min(bv->bv_len >> 9, sectors); |
1263 | ||
1264 | void *p = kmap(bv->bv_page); | |
1265 | memset(p + bv->bv_offset, 0, j << 9); | |
1266 | kunmap(bv->bv_page); | |
1267 | ||
8e51e414 | 1268 | sectors -= j; |
cafe5635 KO |
1269 | } |
1270 | ||
8e51e414 KO |
1271 | bio_advance(bio, min(sectors << 9, bio->bi_size)); |
1272 | ||
1273 | if (!bio->bi_size) | |
1274 | s->op.lookup_done = true; | |
cafe5635 KO |
1275 | |
1276 | return 0; | |
1277 | } | |
1278 | ||
1279 | static void flash_dev_make_request(struct request_queue *q, struct bio *bio) | |
1280 | { | |
1281 | struct search *s; | |
1282 | struct closure *cl; | |
1283 | struct bcache_device *d = bio->bi_bdev->bd_disk->private_data; | |
1284 | int cpu, rw = bio_data_dir(bio); | |
1285 | ||
1286 | cpu = part_stat_lock(); | |
1287 | part_stat_inc(cpu, &d->disk->part0, ios[rw]); | |
1288 | part_stat_add(cpu, &d->disk->part0, sectors[rw], bio_sectors(bio)); | |
1289 | part_stat_unlock(); | |
1290 | ||
1291 | s = search_alloc(bio, d); | |
1292 | cl = &s->cl; | |
1293 | bio = &s->bio.bio; | |
1294 | ||
1295 | trace_bcache_request_start(s, bio); | |
1296 | ||
84f0db03 KO |
1297 | if (!bio->bi_size) { |
1298 | if (s->op.flush_journal) | |
1299 | bch_journal_meta(s->op.c, cl); | |
1300 | } else if (rw) { | |
cafe5635 | 1301 | bch_keybuf_check_overlapping(&s->op.c->moving_gc_keys, |
8e51e414 KO |
1302 | &KEY(d->id, bio->bi_sector, 0), |
1303 | &KEY(d->id, bio_end_sector(bio), 0)); | |
cafe5635 | 1304 | |
84f0db03 | 1305 | s->op.bypass = (bio->bi_rw & REQ_DISCARD) != 0; |
cafe5635 KO |
1306 | s->writeback = true; |
1307 | s->op.cache_bio = bio; | |
1308 | ||
1309 | closure_call(&s->op.cl, bch_insert_data, NULL, cl); | |
1310 | } else { | |
84f0db03 | 1311 | closure_call(&s->op.cl, btree_read_async, NULL, cl); |
cafe5635 KO |
1312 | } |
1313 | ||
1314 | continue_at(cl, search_free, NULL); | |
1315 | } | |
1316 | ||
1317 | static int flash_dev_ioctl(struct bcache_device *d, fmode_t mode, | |
1318 | unsigned int cmd, unsigned long arg) | |
1319 | { | |
1320 | return -ENOTTY; | |
1321 | } | |
1322 | ||
1323 | static int flash_dev_congested(void *data, int bits) | |
1324 | { | |
1325 | struct bcache_device *d = data; | |
1326 | struct request_queue *q; | |
1327 | struct cache *ca; | |
1328 | unsigned i; | |
1329 | int ret = 0; | |
1330 | ||
1331 | for_each_cache(ca, d->c, i) { | |
1332 | q = bdev_get_queue(ca->bdev); | |
1333 | ret |= bdi_congested(&q->backing_dev_info, bits); | |
1334 | } | |
1335 | ||
1336 | return ret; | |
1337 | } | |
1338 | ||
1339 | void bch_flash_dev_request_init(struct bcache_device *d) | |
1340 | { | |
1341 | struct gendisk *g = d->disk; | |
1342 | ||
1343 | g->queue->make_request_fn = flash_dev_make_request; | |
1344 | g->queue->backing_dev_info.congested_fn = flash_dev_congested; | |
1345 | d->cache_miss = flash_dev_cache_miss; | |
1346 | d->ioctl = flash_dev_ioctl; | |
1347 | } | |
1348 | ||
1349 | void bch_request_exit(void) | |
1350 | { | |
1351 | #ifdef CONFIG_CGROUP_BCACHE | |
1352 | cgroup_unload_subsys(&bcache_subsys); | |
1353 | #endif | |
1354 | if (bch_search_cache) | |
1355 | kmem_cache_destroy(bch_search_cache); | |
1356 | } | |
1357 | ||
1358 | int __init bch_request_init(void) | |
1359 | { | |
1360 | bch_search_cache = KMEM_CACHE(search, 0); | |
1361 | if (!bch_search_cache) | |
1362 | return -ENOMEM; | |
1363 | ||
1364 | #ifdef CONFIG_CGROUP_BCACHE | |
1365 | cgroup_load_subsys(&bcache_subsys); | |
1366 | init_bch_cgroup(&bcache_default_cgroup); | |
1367 | ||
1368 | cgroup_add_cftypes(&bcache_subsys, bch_files); | |
1369 | #endif | |
1370 | return 0; | |
1371 | } |