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 | ||
a34a8bfd KO |
28 | static void bch_data_insert_start(struct closure *); |
29 | ||
cafe5635 KO |
30 | /* Cgroup interface */ |
31 | ||
32 | #ifdef CONFIG_CGROUP_BCACHE | |
33 | static struct bch_cgroup bcache_default_cgroup = { .cache_mode = -1 }; | |
34 | ||
35 | static struct bch_cgroup *cgroup_to_bcache(struct cgroup *cgroup) | |
36 | { | |
37 | struct cgroup_subsys_state *css; | |
38 | return cgroup && | |
39 | (css = cgroup_subsys_state(cgroup, bcache_subsys_id)) | |
40 | ? container_of(css, struct bch_cgroup, css) | |
41 | : &bcache_default_cgroup; | |
42 | } | |
43 | ||
44 | struct bch_cgroup *bch_bio_to_cgroup(struct bio *bio) | |
45 | { | |
46 | struct cgroup_subsys_state *css = bio->bi_css | |
47 | ? cgroup_subsys_state(bio->bi_css->cgroup, bcache_subsys_id) | |
48 | : task_subsys_state(current, bcache_subsys_id); | |
49 | ||
50 | return css | |
51 | ? container_of(css, struct bch_cgroup, css) | |
52 | : &bcache_default_cgroup; | |
53 | } | |
54 | ||
55 | static ssize_t cache_mode_read(struct cgroup *cgrp, struct cftype *cft, | |
56 | struct file *file, | |
57 | char __user *buf, size_t nbytes, loff_t *ppos) | |
58 | { | |
59 | char tmp[1024]; | |
169ef1cf KO |
60 | int len = bch_snprint_string_list(tmp, PAGE_SIZE, bch_cache_modes, |
61 | cgroup_to_bcache(cgrp)->cache_mode + 1); | |
cafe5635 KO |
62 | |
63 | if (len < 0) | |
64 | return len; | |
65 | ||
66 | return simple_read_from_buffer(buf, nbytes, ppos, tmp, len); | |
67 | } | |
68 | ||
69 | static int cache_mode_write(struct cgroup *cgrp, struct cftype *cft, | |
70 | const char *buf) | |
71 | { | |
169ef1cf | 72 | int v = bch_read_string_list(buf, bch_cache_modes); |
cafe5635 KO |
73 | if (v < 0) |
74 | return v; | |
75 | ||
76 | cgroup_to_bcache(cgrp)->cache_mode = v - 1; | |
77 | return 0; | |
78 | } | |
79 | ||
80 | static u64 bch_verify_read(struct cgroup *cgrp, struct cftype *cft) | |
81 | { | |
82 | return cgroup_to_bcache(cgrp)->verify; | |
83 | } | |
84 | ||
85 | static int bch_verify_write(struct cgroup *cgrp, struct cftype *cft, u64 val) | |
86 | { | |
87 | cgroup_to_bcache(cgrp)->verify = val; | |
88 | return 0; | |
89 | } | |
90 | ||
91 | static u64 bch_cache_hits_read(struct cgroup *cgrp, struct cftype *cft) | |
92 | { | |
93 | struct bch_cgroup *bcachecg = cgroup_to_bcache(cgrp); | |
94 | return atomic_read(&bcachecg->stats.cache_hits); | |
95 | } | |
96 | ||
97 | static u64 bch_cache_misses_read(struct cgroup *cgrp, struct cftype *cft) | |
98 | { | |
99 | struct bch_cgroup *bcachecg = cgroup_to_bcache(cgrp); | |
100 | return atomic_read(&bcachecg->stats.cache_misses); | |
101 | } | |
102 | ||
103 | static u64 bch_cache_bypass_hits_read(struct cgroup *cgrp, | |
104 | struct cftype *cft) | |
105 | { | |
106 | struct bch_cgroup *bcachecg = cgroup_to_bcache(cgrp); | |
107 | return atomic_read(&bcachecg->stats.cache_bypass_hits); | |
108 | } | |
109 | ||
110 | static u64 bch_cache_bypass_misses_read(struct cgroup *cgrp, | |
111 | struct cftype *cft) | |
112 | { | |
113 | struct bch_cgroup *bcachecg = cgroup_to_bcache(cgrp); | |
114 | return atomic_read(&bcachecg->stats.cache_bypass_misses); | |
115 | } | |
116 | ||
117 | static struct cftype bch_files[] = { | |
118 | { | |
119 | .name = "cache_mode", | |
120 | .read = cache_mode_read, | |
121 | .write_string = cache_mode_write, | |
122 | }, | |
123 | { | |
124 | .name = "verify", | |
125 | .read_u64 = bch_verify_read, | |
126 | .write_u64 = bch_verify_write, | |
127 | }, | |
128 | { | |
129 | .name = "cache_hits", | |
130 | .read_u64 = bch_cache_hits_read, | |
131 | }, | |
132 | { | |
133 | .name = "cache_misses", | |
134 | .read_u64 = bch_cache_misses_read, | |
135 | }, | |
136 | { | |
137 | .name = "cache_bypass_hits", | |
138 | .read_u64 = bch_cache_bypass_hits_read, | |
139 | }, | |
140 | { | |
141 | .name = "cache_bypass_misses", | |
142 | .read_u64 = bch_cache_bypass_misses_read, | |
143 | }, | |
144 | { } /* terminate */ | |
145 | }; | |
146 | ||
147 | static void init_bch_cgroup(struct bch_cgroup *cg) | |
148 | { | |
149 | cg->cache_mode = -1; | |
150 | } | |
151 | ||
152 | static struct cgroup_subsys_state *bcachecg_create(struct cgroup *cgroup) | |
153 | { | |
154 | struct bch_cgroup *cg; | |
155 | ||
156 | cg = kzalloc(sizeof(*cg), GFP_KERNEL); | |
157 | if (!cg) | |
158 | return ERR_PTR(-ENOMEM); | |
159 | init_bch_cgroup(cg); | |
160 | return &cg->css; | |
161 | } | |
162 | ||
163 | static void bcachecg_destroy(struct cgroup *cgroup) | |
164 | { | |
165 | struct bch_cgroup *cg = cgroup_to_bcache(cgroup); | |
166 | free_css_id(&bcache_subsys, &cg->css); | |
167 | kfree(cg); | |
168 | } | |
169 | ||
170 | struct cgroup_subsys bcache_subsys = { | |
171 | .create = bcachecg_create, | |
172 | .destroy = bcachecg_destroy, | |
173 | .subsys_id = bcache_subsys_id, | |
174 | .name = "bcache", | |
175 | .module = THIS_MODULE, | |
176 | }; | |
177 | EXPORT_SYMBOL_GPL(bcache_subsys); | |
178 | #endif | |
179 | ||
180 | static unsigned cache_mode(struct cached_dev *dc, struct bio *bio) | |
181 | { | |
182 | #ifdef CONFIG_CGROUP_BCACHE | |
183 | int r = bch_bio_to_cgroup(bio)->cache_mode; | |
184 | if (r >= 0) | |
185 | return r; | |
186 | #endif | |
187 | return BDEV_CACHE_MODE(&dc->sb); | |
188 | } | |
189 | ||
190 | static bool verify(struct cached_dev *dc, struct bio *bio) | |
191 | { | |
192 | #ifdef CONFIG_CGROUP_BCACHE | |
193 | if (bch_bio_to_cgroup(bio)->verify) | |
194 | return true; | |
195 | #endif | |
196 | return dc->verify; | |
197 | } | |
198 | ||
199 | static void bio_csum(struct bio *bio, struct bkey *k) | |
200 | { | |
7988613b KO |
201 | struct bio_vec bv; |
202 | struct bvec_iter iter; | |
cafe5635 | 203 | uint64_t csum = 0; |
cafe5635 | 204 | |
7988613b KO |
205 | bio_for_each_segment(bv, bio, iter) { |
206 | void *d = kmap(bv.bv_page) + bv.bv_offset; | |
207 | csum = bch_crc64_update(csum, d, bv.bv_len); | |
208 | kunmap(bv.bv_page); | |
cafe5635 KO |
209 | } |
210 | ||
211 | k->ptr[KEY_PTRS(k)] = csum & (~0ULL >> 1); | |
212 | } | |
213 | ||
214 | /* Insert data into cache */ | |
215 | ||
a34a8bfd | 216 | static void bch_data_insert_keys(struct closure *cl) |
cafe5635 | 217 | { |
220bb38c | 218 | struct data_insert_op *op = container_of(cl, struct data_insert_op, cl); |
c18536a7 | 219 | atomic_t *journal_ref = NULL; |
220bb38c | 220 | struct bkey *replace_key = op->replace ? &op->replace_key : NULL; |
6054c6d4 | 221 | int ret; |
cafe5635 | 222 | |
a34a8bfd KO |
223 | /* |
224 | * If we're looping, might already be waiting on | |
225 | * another journal write - can't wait on more than one journal write at | |
226 | * a time | |
227 | * | |
228 | * XXX: this looks wrong | |
229 | */ | |
230 | #if 0 | |
231 | while (atomic_read(&s->cl.remaining) & CLOSURE_WAITING) | |
232 | closure_sync(&s->cl); | |
233 | #endif | |
cafe5635 | 234 | |
220bb38c KO |
235 | if (!op->replace) |
236 | journal_ref = bch_journal(op->c, &op->insert_keys, | |
237 | op->flush_journal ? cl : NULL); | |
cafe5635 | 238 | |
220bb38c | 239 | ret = bch_btree_insert(op->c, &op->insert_keys, |
6054c6d4 KO |
240 | journal_ref, replace_key); |
241 | if (ret == -ESRCH) { | |
220bb38c | 242 | op->replace_collision = true; |
6054c6d4 | 243 | } else if (ret) { |
220bb38c KO |
244 | op->error = -ENOMEM; |
245 | op->insert_data_done = true; | |
a34a8bfd | 246 | } |
cafe5635 | 247 | |
c18536a7 KO |
248 | if (journal_ref) |
249 | atomic_dec_bug(journal_ref); | |
cafe5635 | 250 | |
220bb38c | 251 | if (!op->insert_data_done) |
a34a8bfd | 252 | continue_at(cl, bch_data_insert_start, bcache_wq); |
cafe5635 | 253 | |
220bb38c | 254 | bch_keylist_free(&op->insert_keys); |
a34a8bfd | 255 | closure_return(cl); |
cafe5635 KO |
256 | } |
257 | ||
085d2a3d KO |
258 | static int bch_keylist_realloc(struct keylist *l, unsigned u64s, |
259 | struct cache_set *c) | |
260 | { | |
261 | size_t oldsize = bch_keylist_nkeys(l); | |
262 | size_t newsize = oldsize + u64s; | |
263 | ||
264 | /* | |
265 | * The journalling code doesn't handle the case where the keys to insert | |
266 | * is bigger than an empty write: If we just return -ENOMEM here, | |
267 | * bio_insert() and bio_invalidate() will insert the keys created so far | |
268 | * and finish the rest when the keylist is empty. | |
269 | */ | |
270 | if (newsize * sizeof(uint64_t) > block_bytes(c) - sizeof(struct jset)) | |
271 | return -ENOMEM; | |
272 | ||
273 | return __bch_keylist_realloc(l, u64s); | |
274 | } | |
275 | ||
a34a8bfd KO |
276 | static void bch_data_invalidate(struct closure *cl) |
277 | { | |
220bb38c KO |
278 | struct data_insert_op *op = container_of(cl, struct data_insert_op, cl); |
279 | struct bio *bio = op->bio; | |
a34a8bfd KO |
280 | |
281 | pr_debug("invalidating %i sectors from %llu", | |
4f024f37 | 282 | bio_sectors(bio), (uint64_t) bio->bi_iter.bi_sector); |
a34a8bfd KO |
283 | |
284 | while (bio_sectors(bio)) { | |
81ab4190 KO |
285 | unsigned sectors = min(bio_sectors(bio), |
286 | 1U << (KEY_SIZE_BITS - 1)); | |
a34a8bfd | 287 | |
085d2a3d | 288 | if (bch_keylist_realloc(&op->insert_keys, 2, op->c)) |
a34a8bfd KO |
289 | goto out; |
290 | ||
4f024f37 KO |
291 | bio->bi_iter.bi_sector += sectors; |
292 | bio->bi_iter.bi_size -= sectors << 9; | |
a34a8bfd | 293 | |
220bb38c | 294 | bch_keylist_add(&op->insert_keys, |
4f024f37 | 295 | &KEY(op->inode, bio->bi_iter.bi_sector, sectors)); |
a34a8bfd KO |
296 | } |
297 | ||
220bb38c | 298 | op->insert_data_done = true; |
a34a8bfd KO |
299 | bio_put(bio); |
300 | out: | |
301 | continue_at(cl, bch_data_insert_keys, bcache_wq); | |
302 | } | |
303 | ||
304 | static void bch_data_insert_error(struct closure *cl) | |
cafe5635 | 305 | { |
220bb38c | 306 | struct data_insert_op *op = container_of(cl, struct data_insert_op, cl); |
cafe5635 KO |
307 | |
308 | /* | |
309 | * Our data write just errored, which means we've got a bunch of keys to | |
310 | * insert that point to data that wasn't succesfully written. | |
311 | * | |
312 | * We don't have to insert those keys but we still have to invalidate | |
313 | * that region of the cache - so, if we just strip off all the pointers | |
314 | * from the keys we'll accomplish just that. | |
315 | */ | |
316 | ||
220bb38c | 317 | struct bkey *src = op->insert_keys.keys, *dst = op->insert_keys.keys; |
cafe5635 | 318 | |
220bb38c | 319 | while (src != op->insert_keys.top) { |
cafe5635 KO |
320 | struct bkey *n = bkey_next(src); |
321 | ||
322 | SET_KEY_PTRS(src, 0); | |
c2f95ae2 | 323 | memmove(dst, src, bkey_bytes(src)); |
cafe5635 KO |
324 | |
325 | dst = bkey_next(dst); | |
326 | src = n; | |
327 | } | |
328 | ||
220bb38c | 329 | op->insert_keys.top = dst; |
cafe5635 | 330 | |
a34a8bfd | 331 | bch_data_insert_keys(cl); |
cafe5635 KO |
332 | } |
333 | ||
a34a8bfd | 334 | static void bch_data_insert_endio(struct bio *bio, int error) |
cafe5635 KO |
335 | { |
336 | struct closure *cl = bio->bi_private; | |
220bb38c | 337 | struct data_insert_op *op = container_of(cl, struct data_insert_op, cl); |
cafe5635 KO |
338 | |
339 | if (error) { | |
340 | /* TODO: We could try to recover from this. */ | |
220bb38c KO |
341 | if (op->writeback) |
342 | op->error = error; | |
343 | else if (!op->replace) | |
a34a8bfd | 344 | set_closure_fn(cl, bch_data_insert_error, bcache_wq); |
cafe5635 KO |
345 | else |
346 | set_closure_fn(cl, NULL, NULL); | |
347 | } | |
348 | ||
220bb38c | 349 | bch_bbio_endio(op->c, bio, error, "writing data to cache"); |
cafe5635 KO |
350 | } |
351 | ||
a34a8bfd | 352 | static void bch_data_insert_start(struct closure *cl) |
cafe5635 | 353 | { |
220bb38c KO |
354 | struct data_insert_op *op = container_of(cl, struct data_insert_op, cl); |
355 | struct bio *bio = op->bio, *n; | |
cafe5635 | 356 | |
220bb38c | 357 | if (op->bypass) |
a34a8bfd | 358 | return bch_data_invalidate(cl); |
cafe5635 | 359 | |
220bb38c KO |
360 | if (atomic_sub_return(bio_sectors(bio), &op->c->sectors_to_gc) < 0) { |
361 | set_gc_sectors(op->c); | |
362 | wake_up_gc(op->c); | |
cafe5635 KO |
363 | } |
364 | ||
54d12f2b KO |
365 | /* |
366 | * Journal writes are marked REQ_FLUSH; if the original write was a | |
367 | * flush, it'll wait on the journal write. | |
368 | */ | |
369 | bio->bi_rw &= ~(REQ_FLUSH|REQ_FUA); | |
370 | ||
cafe5635 KO |
371 | do { |
372 | unsigned i; | |
373 | struct bkey *k; | |
220bb38c | 374 | struct bio_set *split = op->c->bio_split; |
cafe5635 KO |
375 | |
376 | /* 1 for the device pointer and 1 for the chksum */ | |
220bb38c | 377 | if (bch_keylist_realloc(&op->insert_keys, |
085d2a3d | 378 | 3 + (op->csum ? 1 : 0), |
220bb38c | 379 | op->c)) |
a34a8bfd | 380 | continue_at(cl, bch_data_insert_keys, bcache_wq); |
cafe5635 | 381 | |
220bb38c | 382 | k = op->insert_keys.top; |
cafe5635 | 383 | bkey_init(k); |
220bb38c | 384 | SET_KEY_INODE(k, op->inode); |
4f024f37 | 385 | SET_KEY_OFFSET(k, bio->bi_iter.bi_sector); |
cafe5635 | 386 | |
2599b53b KO |
387 | if (!bch_alloc_sectors(op->c, k, bio_sectors(bio), |
388 | op->write_point, op->write_prio, | |
389 | op->writeback)) | |
cafe5635 KO |
390 | goto err; |
391 | ||
20d0189b | 392 | n = bio_next_split(bio, KEY_SIZE(k), GFP_NOIO, split); |
cafe5635 | 393 | |
a34a8bfd | 394 | n->bi_end_io = bch_data_insert_endio; |
cafe5635 KO |
395 | n->bi_private = cl; |
396 | ||
220bb38c | 397 | if (op->writeback) { |
cafe5635 KO |
398 | SET_KEY_DIRTY(k, true); |
399 | ||
400 | for (i = 0; i < KEY_PTRS(k); i++) | |
220bb38c | 401 | SET_GC_MARK(PTR_BUCKET(op->c, k, i), |
cafe5635 KO |
402 | GC_MARK_DIRTY); |
403 | } | |
404 | ||
220bb38c | 405 | SET_KEY_CSUM(k, op->csum); |
cafe5635 KO |
406 | if (KEY_CSUM(k)) |
407 | bio_csum(n, k); | |
408 | ||
c37511b8 | 409 | trace_bcache_cache_insert(k); |
220bb38c | 410 | bch_keylist_push(&op->insert_keys); |
cafe5635 | 411 | |
cafe5635 | 412 | n->bi_rw |= REQ_WRITE; |
220bb38c | 413 | bch_submit_bbio(n, op->c, k, 0); |
cafe5635 KO |
414 | } while (n != bio); |
415 | ||
220bb38c | 416 | op->insert_data_done = true; |
a34a8bfd | 417 | continue_at(cl, bch_data_insert_keys, bcache_wq); |
cafe5635 KO |
418 | err: |
419 | /* bch_alloc_sectors() blocks if s->writeback = true */ | |
220bb38c | 420 | BUG_ON(op->writeback); |
cafe5635 KO |
421 | |
422 | /* | |
423 | * But if it's not a writeback write we'd rather just bail out if | |
424 | * there aren't any buckets ready to write to - it might take awhile and | |
425 | * we might be starving btree writes for gc or something. | |
426 | */ | |
427 | ||
220bb38c | 428 | if (!op->replace) { |
cafe5635 KO |
429 | /* |
430 | * Writethrough write: We can't complete the write until we've | |
431 | * updated the index. But we don't want to delay the write while | |
432 | * we wait for buckets to be freed up, so just invalidate the | |
433 | * rest of the write. | |
434 | */ | |
220bb38c | 435 | op->bypass = true; |
a34a8bfd | 436 | return bch_data_invalidate(cl); |
cafe5635 KO |
437 | } else { |
438 | /* | |
439 | * From a cache miss, we can just insert the keys for the data | |
440 | * we have written or bail out if we didn't do anything. | |
441 | */ | |
220bb38c | 442 | op->insert_data_done = true; |
cafe5635 KO |
443 | bio_put(bio); |
444 | ||
220bb38c | 445 | if (!bch_keylist_empty(&op->insert_keys)) |
a34a8bfd | 446 | continue_at(cl, bch_data_insert_keys, bcache_wq); |
cafe5635 KO |
447 | else |
448 | closure_return(cl); | |
449 | } | |
450 | } | |
451 | ||
452 | /** | |
a34a8bfd | 453 | * bch_data_insert - stick some data in the cache |
cafe5635 KO |
454 | * |
455 | * This is the starting point for any data to end up in a cache device; it could | |
456 | * be from a normal write, or a writeback write, or a write to a flash only | |
457 | * volume - it's also used by the moving garbage collector to compact data in | |
458 | * mostly empty buckets. | |
459 | * | |
460 | * It first writes the data to the cache, creating a list of keys to be inserted | |
461 | * (if the data had to be fragmented there will be multiple keys); after the | |
462 | * data is written it calls bch_journal, and after the keys have been added to | |
463 | * the next journal write they're inserted into the btree. | |
464 | * | |
c18536a7 | 465 | * It inserts the data in s->cache_bio; bi_sector is used for the key offset, |
cafe5635 KO |
466 | * and op->inode is used for the key inode. |
467 | * | |
c18536a7 KO |
468 | * If s->bypass is true, instead of inserting the data it invalidates the |
469 | * region of the cache represented by s->cache_bio and op->inode. | |
cafe5635 | 470 | */ |
a34a8bfd | 471 | void bch_data_insert(struct closure *cl) |
cafe5635 | 472 | { |
220bb38c | 473 | struct data_insert_op *op = container_of(cl, struct data_insert_op, cl); |
cafe5635 | 474 | |
220bb38c KO |
475 | trace_bcache_write(op->bio, op->writeback, op->bypass); |
476 | ||
477 | bch_keylist_init(&op->insert_keys); | |
478 | bio_get(op->bio); | |
a34a8bfd | 479 | bch_data_insert_start(cl); |
cafe5635 KO |
480 | } |
481 | ||
220bb38c KO |
482 | /* Congested? */ |
483 | ||
484 | unsigned bch_get_congested(struct cache_set *c) | |
485 | { | |
486 | int i; | |
487 | long rand; | |
488 | ||
489 | if (!c->congested_read_threshold_us && | |
490 | !c->congested_write_threshold_us) | |
491 | return 0; | |
492 | ||
493 | i = (local_clock_us() - c->congested_last_us) / 1024; | |
494 | if (i < 0) | |
495 | return 0; | |
496 | ||
497 | i += atomic_read(&c->congested); | |
498 | if (i >= 0) | |
499 | return 0; | |
500 | ||
501 | i += CONGESTED_MAX; | |
502 | ||
503 | if (i > 0) | |
504 | i = fract_exp_two(i, 6); | |
505 | ||
506 | rand = get_random_int(); | |
507 | i -= bitmap_weight(&rand, BITS_PER_LONG); | |
508 | ||
509 | return i > 0 ? i : 1; | |
510 | } | |
511 | ||
512 | static void add_sequential(struct task_struct *t) | |
513 | { | |
514 | ewma_add(t->sequential_io_avg, | |
515 | t->sequential_io, 8, 0); | |
516 | ||
517 | t->sequential_io = 0; | |
518 | } | |
519 | ||
520 | static struct hlist_head *iohash(struct cached_dev *dc, uint64_t k) | |
521 | { | |
522 | return &dc->io_hash[hash_64(k, RECENT_IO_BITS)]; | |
523 | } | |
524 | ||
525 | static bool check_should_bypass(struct cached_dev *dc, struct bio *bio) | |
526 | { | |
527 | struct cache_set *c = dc->disk.c; | |
528 | unsigned mode = cache_mode(dc, bio); | |
529 | unsigned sectors, congested = bch_get_congested(c); | |
530 | struct task_struct *task = current; | |
8aee1220 | 531 | struct io *i; |
220bb38c | 532 | |
c4d951dd | 533 | if (test_bit(BCACHE_DEV_DETACHING, &dc->disk.flags) || |
220bb38c KO |
534 | c->gc_stats.in_use > CUTOFF_CACHE_ADD || |
535 | (bio->bi_rw & REQ_DISCARD)) | |
536 | goto skip; | |
537 | ||
538 | if (mode == CACHE_MODE_NONE || | |
539 | (mode == CACHE_MODE_WRITEAROUND && | |
540 | (bio->bi_rw & REQ_WRITE))) | |
541 | goto skip; | |
542 | ||
4f024f37 | 543 | if (bio->bi_iter.bi_sector & (c->sb.block_size - 1) || |
220bb38c KO |
544 | bio_sectors(bio) & (c->sb.block_size - 1)) { |
545 | pr_debug("skipping unaligned io"); | |
546 | goto skip; | |
547 | } | |
548 | ||
5ceaaad7 KO |
549 | if (bypass_torture_test(dc)) { |
550 | if ((get_random_int() & 3) == 3) | |
551 | goto skip; | |
552 | else | |
553 | goto rescale; | |
554 | } | |
555 | ||
220bb38c KO |
556 | if (!congested && !dc->sequential_cutoff) |
557 | goto rescale; | |
558 | ||
559 | if (!congested && | |
560 | mode == CACHE_MODE_WRITEBACK && | |
561 | (bio->bi_rw & REQ_WRITE) && | |
562 | (bio->bi_rw & REQ_SYNC)) | |
563 | goto rescale; | |
564 | ||
8aee1220 | 565 | spin_lock(&dc->io_lock); |
220bb38c | 566 | |
4f024f37 KO |
567 | hlist_for_each_entry(i, iohash(dc, bio->bi_iter.bi_sector), hash) |
568 | if (i->last == bio->bi_iter.bi_sector && | |
8aee1220 KO |
569 | time_before(jiffies, i->jiffies)) |
570 | goto found; | |
220bb38c | 571 | |
8aee1220 | 572 | i = list_first_entry(&dc->io_lru, struct io, lru); |
220bb38c | 573 | |
8aee1220 KO |
574 | add_sequential(task); |
575 | i->sequential = 0; | |
220bb38c | 576 | found: |
4f024f37 KO |
577 | if (i->sequential + bio->bi_iter.bi_size > i->sequential) |
578 | i->sequential += bio->bi_iter.bi_size; | |
220bb38c | 579 | |
8aee1220 KO |
580 | i->last = bio_end_sector(bio); |
581 | i->jiffies = jiffies + msecs_to_jiffies(5000); | |
582 | task->sequential_io = i->sequential; | |
220bb38c | 583 | |
8aee1220 KO |
584 | hlist_del(&i->hash); |
585 | hlist_add_head(&i->hash, iohash(dc, i->last)); | |
586 | list_move_tail(&i->lru, &dc->io_lru); | |
220bb38c | 587 | |
8aee1220 | 588 | spin_unlock(&dc->io_lock); |
220bb38c KO |
589 | |
590 | sectors = max(task->sequential_io, | |
591 | task->sequential_io_avg) >> 9; | |
592 | ||
593 | if (dc->sequential_cutoff && | |
594 | sectors >= dc->sequential_cutoff >> 9) { | |
595 | trace_bcache_bypass_sequential(bio); | |
596 | goto skip; | |
597 | } | |
598 | ||
599 | if (congested && sectors >= congested) { | |
600 | trace_bcache_bypass_congested(bio); | |
601 | goto skip; | |
602 | } | |
603 | ||
604 | rescale: | |
605 | bch_rescale_priorities(c, bio_sectors(bio)); | |
606 | return false; | |
607 | skip: | |
608 | bch_mark_sectors_bypassed(c, dc, bio_sectors(bio)); | |
609 | return true; | |
610 | } | |
611 | ||
2c1953e2 | 612 | /* Cache lookup */ |
cafe5635 | 613 | |
220bb38c KO |
614 | struct search { |
615 | /* Stack frame for bio_complete */ | |
616 | struct closure cl; | |
617 | ||
220bb38c KO |
618 | struct bbio bio; |
619 | struct bio *orig_bio; | |
620 | struct bio *cache_miss; | |
a5ae4300 | 621 | struct bcache_device *d; |
220bb38c KO |
622 | |
623 | unsigned insert_bio_sectors; | |
220bb38c | 624 | unsigned recoverable:1; |
220bb38c | 625 | unsigned write:1; |
5ceaaad7 | 626 | unsigned read_dirty_data:1; |
220bb38c KO |
627 | |
628 | unsigned long start_time; | |
629 | ||
630 | struct btree_op op; | |
631 | struct data_insert_op iop; | |
632 | }; | |
633 | ||
2c1953e2 | 634 | static void bch_cache_read_endio(struct bio *bio, int error) |
cafe5635 KO |
635 | { |
636 | struct bbio *b = container_of(bio, struct bbio, bio); | |
637 | struct closure *cl = bio->bi_private; | |
638 | struct search *s = container_of(cl, struct search, cl); | |
639 | ||
640 | /* | |
641 | * If the bucket was reused while our bio was in flight, we might have | |
642 | * read the wrong data. Set s->error but not error so it doesn't get | |
643 | * counted against the cache device, but we'll still reread the data | |
644 | * from the backing device. | |
645 | */ | |
646 | ||
647 | if (error) | |
220bb38c | 648 | s->iop.error = error; |
d56d000a KO |
649 | else if (!KEY_DIRTY(&b->key) && |
650 | ptr_stale(s->iop.c, &b->key, 0)) { | |
220bb38c KO |
651 | atomic_long_inc(&s->iop.c->cache_read_races); |
652 | s->iop.error = -EINTR; | |
cafe5635 KO |
653 | } |
654 | ||
220bb38c | 655 | bch_bbio_endio(s->iop.c, bio, error, "reading from cache"); |
cafe5635 KO |
656 | } |
657 | ||
2c1953e2 KO |
658 | /* |
659 | * Read from a single key, handling the initial cache miss if the key starts in | |
660 | * the middle of the bio | |
661 | */ | |
cc231966 | 662 | static int cache_lookup_fn(struct btree_op *op, struct btree *b, struct bkey *k) |
2c1953e2 KO |
663 | { |
664 | struct search *s = container_of(op, struct search, op); | |
cc231966 KO |
665 | struct bio *n, *bio = &s->bio.bio; |
666 | struct bkey *bio_key; | |
2c1953e2 | 667 | unsigned ptr; |
2c1953e2 | 668 | |
4f024f37 | 669 | if (bkey_cmp(k, &KEY(s->iop.inode, bio->bi_iter.bi_sector, 0)) <= 0) |
cc231966 KO |
670 | return MAP_CONTINUE; |
671 | ||
220bb38c | 672 | if (KEY_INODE(k) != s->iop.inode || |
4f024f37 | 673 | KEY_START(k) > bio->bi_iter.bi_sector) { |
cc231966 | 674 | unsigned bio_sectors = bio_sectors(bio); |
220bb38c | 675 | unsigned sectors = KEY_INODE(k) == s->iop.inode |
cc231966 | 676 | ? min_t(uint64_t, INT_MAX, |
4f024f37 | 677 | KEY_START(k) - bio->bi_iter.bi_sector) |
cc231966 KO |
678 | : INT_MAX; |
679 | ||
680 | int ret = s->d->cache_miss(b, s, bio, sectors); | |
681 | if (ret != MAP_CONTINUE) | |
682 | return ret; | |
683 | ||
684 | /* if this was a complete miss we shouldn't get here */ | |
685 | BUG_ON(bio_sectors <= sectors); | |
686 | } | |
687 | ||
688 | if (!KEY_SIZE(k)) | |
689 | return MAP_CONTINUE; | |
2c1953e2 KO |
690 | |
691 | /* XXX: figure out best pointer - for multiple cache devices */ | |
692 | ptr = 0; | |
693 | ||
694 | PTR_BUCKET(b->c, k, ptr)->prio = INITIAL_PRIO; | |
695 | ||
5ceaaad7 KO |
696 | if (KEY_DIRTY(k)) |
697 | s->read_dirty_data = true; | |
698 | ||
20d0189b KO |
699 | n = bio_next_split(bio, min_t(uint64_t, INT_MAX, |
700 | KEY_OFFSET(k) - bio->bi_iter.bi_sector), | |
701 | GFP_NOIO, s->d->bio_split); | |
2c1953e2 | 702 | |
cc231966 KO |
703 | bio_key = &container_of(n, struct bbio, bio)->key; |
704 | bch_bkey_copy_single_ptr(bio_key, k, ptr); | |
2c1953e2 | 705 | |
4f024f37 | 706 | bch_cut_front(&KEY(s->iop.inode, n->bi_iter.bi_sector, 0), bio_key); |
220bb38c | 707 | bch_cut_back(&KEY(s->iop.inode, bio_end_sector(n), 0), bio_key); |
2c1953e2 | 708 | |
cc231966 KO |
709 | n->bi_end_io = bch_cache_read_endio; |
710 | n->bi_private = &s->cl; | |
2c1953e2 | 711 | |
cc231966 KO |
712 | /* |
713 | * The bucket we're reading from might be reused while our bio | |
714 | * is in flight, and we could then end up reading the wrong | |
715 | * data. | |
716 | * | |
717 | * We guard against this by checking (in cache_read_endio()) if | |
718 | * the pointer is stale again; if so, we treat it as an error | |
719 | * and reread from the backing device (but we don't pass that | |
720 | * error up anywhere). | |
721 | */ | |
2c1953e2 | 722 | |
cc231966 KO |
723 | __bch_submit_bbio(n, b->c); |
724 | return n == bio ? MAP_DONE : MAP_CONTINUE; | |
2c1953e2 KO |
725 | } |
726 | ||
727 | static void cache_lookup(struct closure *cl) | |
728 | { | |
220bb38c | 729 | struct search *s = container_of(cl, struct search, iop.cl); |
2c1953e2 | 730 | struct bio *bio = &s->bio.bio; |
a5ae4300 | 731 | int ret; |
2c1953e2 | 732 | |
a5ae4300 KO |
733 | bch_btree_op_init(&s->op, -1); |
734 | ||
735 | ret = bch_btree_map_keys(&s->op, s->iop.c, | |
736 | &KEY(s->iop.inode, bio->bi_iter.bi_sector, 0), | |
737 | cache_lookup_fn, MAP_END_KEY); | |
2c1953e2 KO |
738 | if (ret == -EAGAIN) |
739 | continue_at(cl, cache_lookup, bcache_wq); | |
740 | ||
741 | closure_return(cl); | |
742 | } | |
743 | ||
744 | /* Common code for the make_request functions */ | |
745 | ||
746 | static void request_endio(struct bio *bio, int error) | |
747 | { | |
748 | struct closure *cl = bio->bi_private; | |
749 | ||
750 | if (error) { | |
751 | struct search *s = container_of(cl, struct search, cl); | |
220bb38c | 752 | s->iop.error = error; |
2c1953e2 KO |
753 | /* Only cache read errors are recoverable */ |
754 | s->recoverable = false; | |
755 | } | |
756 | ||
757 | bio_put(bio); | |
758 | closure_put(cl); | |
759 | } | |
760 | ||
cafe5635 KO |
761 | static void bio_complete(struct search *s) |
762 | { | |
763 | if (s->orig_bio) { | |
764 | int cpu, rw = bio_data_dir(s->orig_bio); | |
765 | unsigned long duration = jiffies - s->start_time; | |
766 | ||
767 | cpu = part_stat_lock(); | |
768 | part_round_stats(cpu, &s->d->disk->part0); | |
769 | part_stat_add(cpu, &s->d->disk->part0, ticks[rw], duration); | |
770 | part_stat_unlock(); | |
771 | ||
220bb38c KO |
772 | trace_bcache_request_end(s->d, s->orig_bio); |
773 | bio_endio(s->orig_bio, s->iop.error); | |
cafe5635 KO |
774 | s->orig_bio = NULL; |
775 | } | |
776 | } | |
777 | ||
a5ae4300 | 778 | static void do_bio_hook(struct search *s, struct bio *orig_bio) |
cafe5635 KO |
779 | { |
780 | struct bio *bio = &s->bio.bio; | |
cafe5635 | 781 | |
ed9c47be | 782 | bio_init(bio); |
a5ae4300 | 783 | __bio_clone_fast(bio, orig_bio); |
cafe5635 KO |
784 | bio->bi_end_io = request_endio; |
785 | bio->bi_private = &s->cl; | |
ed9c47be | 786 | |
cafe5635 KO |
787 | atomic_set(&bio->bi_cnt, 3); |
788 | } | |
789 | ||
790 | static void search_free(struct closure *cl) | |
791 | { | |
792 | struct search *s = container_of(cl, struct search, cl); | |
793 | bio_complete(s); | |
794 | ||
220bb38c KO |
795 | if (s->iop.bio) |
796 | bio_put(s->iop.bio); | |
cafe5635 | 797 | |
cafe5635 KO |
798 | closure_debug_destroy(cl); |
799 | mempool_free(s, s->d->c->search); | |
800 | } | |
801 | ||
a5ae4300 KO |
802 | static inline struct search *search_alloc(struct bio *bio, |
803 | struct bcache_device *d) | |
cafe5635 | 804 | { |
0b93207a | 805 | struct search *s; |
0b93207a KO |
806 | |
807 | s = mempool_alloc(d->c->search, GFP_NOIO); | |
cafe5635 | 808 | |
a5ae4300 KO |
809 | closure_init(&s->cl, NULL); |
810 | do_bio_hook(s, bio); | |
cafe5635 | 811 | |
cafe5635 | 812 | s->orig_bio = bio; |
a5ae4300 KO |
813 | s->cache_miss = NULL; |
814 | s->d = d; | |
cafe5635 | 815 | s->recoverable = 1; |
a5ae4300 KO |
816 | s->write = (bio->bi_rw & REQ_WRITE) != 0; |
817 | s->read_dirty_data = 0; | |
cafe5635 | 818 | s->start_time = jiffies; |
a5ae4300 KO |
819 | |
820 | s->iop.c = d->c; | |
821 | s->iop.bio = NULL; | |
822 | s->iop.inode = d->id; | |
823 | s->iop.write_point = hash_long((unsigned long) current, 16); | |
824 | s->iop.write_prio = 0; | |
825 | s->iop.error = 0; | |
826 | s->iop.flags = 0; | |
827 | s->iop.flush_journal = (bio->bi_rw & (REQ_FLUSH|REQ_FUA)) != 0; | |
cafe5635 | 828 | |
cafe5635 KO |
829 | return s; |
830 | } | |
831 | ||
cafe5635 KO |
832 | /* Cached devices */ |
833 | ||
834 | static void cached_dev_bio_complete(struct closure *cl) | |
835 | { | |
836 | struct search *s = container_of(cl, struct search, cl); | |
837 | struct cached_dev *dc = container_of(s->d, struct cached_dev, disk); | |
838 | ||
839 | search_free(cl); | |
840 | cached_dev_put(dc); | |
841 | } | |
842 | ||
843 | /* Process reads */ | |
844 | ||
cdd972b1 | 845 | static void cached_dev_cache_miss_done(struct closure *cl) |
cafe5635 KO |
846 | { |
847 | struct search *s = container_of(cl, struct search, cl); | |
848 | ||
220bb38c KO |
849 | if (s->iop.replace_collision) |
850 | bch_mark_cache_miss_collision(s->iop.c, s->d); | |
cafe5635 | 851 | |
220bb38c | 852 | if (s->iop.bio) { |
cafe5635 KO |
853 | int i; |
854 | struct bio_vec *bv; | |
855 | ||
220bb38c | 856 | bio_for_each_segment_all(bv, s->iop.bio, i) |
cafe5635 KO |
857 | __free_page(bv->bv_page); |
858 | } | |
859 | ||
860 | cached_dev_bio_complete(cl); | |
861 | } | |
862 | ||
cdd972b1 | 863 | static void cached_dev_read_error(struct closure *cl) |
cafe5635 KO |
864 | { |
865 | struct search *s = container_of(cl, struct search, cl); | |
cdd972b1 | 866 | struct bio *bio = &s->bio.bio; |
cafe5635 KO |
867 | |
868 | if (s->recoverable) { | |
c37511b8 KO |
869 | /* Retry from the backing device: */ |
870 | trace_bcache_read_retry(s->orig_bio); | |
cafe5635 | 871 | |
220bb38c | 872 | s->iop.error = 0; |
a5ae4300 | 873 | do_bio_hook(s, s->orig_bio); |
cafe5635 KO |
874 | |
875 | /* XXX: invalidate cache */ | |
876 | ||
cdd972b1 | 877 | closure_bio_submit(bio, cl, s->d); |
cafe5635 KO |
878 | } |
879 | ||
cdd972b1 | 880 | continue_at(cl, cached_dev_cache_miss_done, NULL); |
cafe5635 KO |
881 | } |
882 | ||
cdd972b1 | 883 | static void cached_dev_read_done(struct closure *cl) |
cafe5635 KO |
884 | { |
885 | struct search *s = container_of(cl, struct search, cl); | |
886 | struct cached_dev *dc = container_of(s->d, struct cached_dev, disk); | |
887 | ||
888 | /* | |
cdd972b1 KO |
889 | * We had a cache miss; cache_bio now contains data ready to be inserted |
890 | * into the cache. | |
cafe5635 KO |
891 | * |
892 | * First, we copy the data we just read from cache_bio's bounce buffers | |
893 | * to the buffers the original bio pointed to: | |
894 | */ | |
895 | ||
220bb38c KO |
896 | if (s->iop.bio) { |
897 | bio_reset(s->iop.bio); | |
4f024f37 | 898 | s->iop.bio->bi_iter.bi_sector = s->cache_miss->bi_iter.bi_sector; |
220bb38c | 899 | s->iop.bio->bi_bdev = s->cache_miss->bi_bdev; |
4f024f37 | 900 | s->iop.bio->bi_iter.bi_size = s->insert_bio_sectors << 9; |
220bb38c | 901 | bch_bio_map(s->iop.bio, NULL); |
cafe5635 | 902 | |
220bb38c | 903 | bio_copy_data(s->cache_miss, s->iop.bio); |
cafe5635 KO |
904 | |
905 | bio_put(s->cache_miss); | |
906 | s->cache_miss = NULL; | |
907 | } | |
908 | ||
ed9c47be | 909 | if (verify(dc, &s->bio.bio) && s->recoverable && !s->read_dirty_data) |
220bb38c | 910 | bch_data_verify(dc, s->orig_bio); |
cafe5635 KO |
911 | |
912 | bio_complete(s); | |
913 | ||
220bb38c KO |
914 | if (s->iop.bio && |
915 | !test_bit(CACHE_SET_STOPPING, &s->iop.c->flags)) { | |
916 | BUG_ON(!s->iop.replace); | |
917 | closure_call(&s->iop.cl, bch_data_insert, NULL, cl); | |
cafe5635 KO |
918 | } |
919 | ||
cdd972b1 | 920 | continue_at(cl, cached_dev_cache_miss_done, NULL); |
cafe5635 KO |
921 | } |
922 | ||
cdd972b1 | 923 | static void cached_dev_read_done_bh(struct closure *cl) |
cafe5635 KO |
924 | { |
925 | struct search *s = container_of(cl, struct search, cl); | |
926 | struct cached_dev *dc = container_of(s->d, struct cached_dev, disk); | |
927 | ||
220bb38c KO |
928 | bch_mark_cache_accounting(s->iop.c, s->d, |
929 | !s->cache_miss, s->iop.bypass); | |
930 | trace_bcache_read(s->orig_bio, !s->cache_miss, s->iop.bypass); | |
cafe5635 | 931 | |
220bb38c | 932 | if (s->iop.error) |
cdd972b1 | 933 | continue_at_nobarrier(cl, cached_dev_read_error, bcache_wq); |
220bb38c | 934 | else if (s->iop.bio || verify(dc, &s->bio.bio)) |
cdd972b1 | 935 | continue_at_nobarrier(cl, cached_dev_read_done, bcache_wq); |
cafe5635 | 936 | else |
cdd972b1 | 937 | continue_at_nobarrier(cl, cached_dev_bio_complete, NULL); |
cafe5635 KO |
938 | } |
939 | ||
940 | static int cached_dev_cache_miss(struct btree *b, struct search *s, | |
941 | struct bio *bio, unsigned sectors) | |
942 | { | |
2c1953e2 | 943 | int ret = MAP_CONTINUE; |
e7c590eb | 944 | unsigned reada = 0; |
cafe5635 | 945 | struct cached_dev *dc = container_of(s->d, struct cached_dev, disk); |
cdd972b1 | 946 | struct bio *miss, *cache_bio; |
cafe5635 | 947 | |
220bb38c | 948 | if (s->cache_miss || s->iop.bypass) { |
20d0189b | 949 | miss = bio_next_split(bio, sectors, GFP_NOIO, s->d->bio_split); |
2c1953e2 | 950 | ret = miss == bio ? MAP_DONE : MAP_CONTINUE; |
e7c590eb KO |
951 | goto out_submit; |
952 | } | |
cafe5635 | 953 | |
e7c590eb KO |
954 | if (!(bio->bi_rw & REQ_RAHEAD) && |
955 | !(bio->bi_rw & REQ_META) && | |
220bb38c | 956 | s->iop.c->gc_stats.in_use < CUTOFF_CACHE_READA) |
e7c590eb KO |
957 | reada = min_t(sector_t, dc->readahead >> 9, |
958 | bdev_sectors(bio->bi_bdev) - bio_end_sector(bio)); | |
cafe5635 | 959 | |
220bb38c | 960 | s->insert_bio_sectors = min(sectors, bio_sectors(bio) + reada); |
cafe5635 | 961 | |
220bb38c | 962 | s->iop.replace_key = KEY(s->iop.inode, |
4f024f37 | 963 | bio->bi_iter.bi_sector + s->insert_bio_sectors, |
220bb38c | 964 | s->insert_bio_sectors); |
e7c590eb | 965 | |
220bb38c | 966 | ret = bch_btree_insert_check_key(b, &s->op, &s->iop.replace_key); |
e7c590eb KO |
967 | if (ret) |
968 | return ret; | |
969 | ||
220bb38c | 970 | s->iop.replace = true; |
1b207d80 | 971 | |
20d0189b | 972 | miss = bio_next_split(bio, sectors, GFP_NOIO, s->d->bio_split); |
2c1953e2 KO |
973 | |
974 | /* btree_search_recurse()'s btree iterator is no good anymore */ | |
975 | ret = miss == bio ? MAP_DONE : -EINTR; | |
cafe5635 | 976 | |
cdd972b1 | 977 | cache_bio = bio_alloc_bioset(GFP_NOWAIT, |
220bb38c | 978 | DIV_ROUND_UP(s->insert_bio_sectors, PAGE_SECTORS), |
cafe5635 | 979 | dc->disk.bio_split); |
cdd972b1 | 980 | if (!cache_bio) |
cafe5635 KO |
981 | goto out_submit; |
982 | ||
4f024f37 KO |
983 | cache_bio->bi_iter.bi_sector = miss->bi_iter.bi_sector; |
984 | cache_bio->bi_bdev = miss->bi_bdev; | |
985 | cache_bio->bi_iter.bi_size = s->insert_bio_sectors << 9; | |
cafe5635 | 986 | |
cdd972b1 KO |
987 | cache_bio->bi_end_io = request_endio; |
988 | cache_bio->bi_private = &s->cl; | |
cafe5635 | 989 | |
cdd972b1 KO |
990 | bch_bio_map(cache_bio, NULL); |
991 | if (bio_alloc_pages(cache_bio, __GFP_NOWARN|GFP_NOIO)) | |
cafe5635 KO |
992 | goto out_put; |
993 | ||
220bb38c KO |
994 | if (reada) |
995 | bch_mark_cache_readahead(s->iop.c, s->d); | |
996 | ||
cdd972b1 | 997 | s->cache_miss = miss; |
220bb38c | 998 | s->iop.bio = cache_bio; |
cdd972b1 KO |
999 | bio_get(cache_bio); |
1000 | closure_bio_submit(cache_bio, &s->cl, s->d); | |
cafe5635 KO |
1001 | |
1002 | return ret; | |
1003 | out_put: | |
cdd972b1 | 1004 | bio_put(cache_bio); |
cafe5635 | 1005 | out_submit: |
e7c590eb KO |
1006 | miss->bi_end_io = request_endio; |
1007 | miss->bi_private = &s->cl; | |
cafe5635 KO |
1008 | closure_bio_submit(miss, &s->cl, s->d); |
1009 | return ret; | |
1010 | } | |
1011 | ||
cdd972b1 | 1012 | static void cached_dev_read(struct cached_dev *dc, struct search *s) |
cafe5635 KO |
1013 | { |
1014 | struct closure *cl = &s->cl; | |
1015 | ||
220bb38c | 1016 | closure_call(&s->iop.cl, cache_lookup, NULL, cl); |
cdd972b1 | 1017 | continue_at(cl, cached_dev_read_done_bh, NULL); |
cafe5635 KO |
1018 | } |
1019 | ||
1020 | /* Process writes */ | |
1021 | ||
1022 | static void cached_dev_write_complete(struct closure *cl) | |
1023 | { | |
1024 | struct search *s = container_of(cl, struct search, cl); | |
1025 | struct cached_dev *dc = container_of(s->d, struct cached_dev, disk); | |
1026 | ||
1027 | up_read_non_owner(&dc->writeback_lock); | |
1028 | cached_dev_bio_complete(cl); | |
1029 | } | |
1030 | ||
cdd972b1 | 1031 | static void cached_dev_write(struct cached_dev *dc, struct search *s) |
cafe5635 KO |
1032 | { |
1033 | struct closure *cl = &s->cl; | |
1034 | struct bio *bio = &s->bio.bio; | |
4f024f37 | 1035 | struct bkey start = KEY(dc->disk.id, bio->bi_iter.bi_sector, 0); |
84f0db03 | 1036 | struct bkey end = KEY(dc->disk.id, bio_end_sector(bio), 0); |
cafe5635 | 1037 | |
220bb38c | 1038 | bch_keybuf_check_overlapping(&s->iop.c->moving_gc_keys, &start, &end); |
cafe5635 | 1039 | |
cafe5635 | 1040 | down_read_non_owner(&dc->writeback_lock); |
cafe5635 | 1041 | if (bch_keybuf_check_overlapping(&dc->writeback_keys, &start, &end)) { |
84f0db03 KO |
1042 | /* |
1043 | * We overlap with some dirty data undergoing background | |
1044 | * writeback, force this write to writeback | |
1045 | */ | |
220bb38c KO |
1046 | s->iop.bypass = false; |
1047 | s->iop.writeback = true; | |
cafe5635 KO |
1048 | } |
1049 | ||
84f0db03 KO |
1050 | /* |
1051 | * Discards aren't _required_ to do anything, so skipping if | |
1052 | * check_overlapping returned true is ok | |
1053 | * | |
1054 | * But check_overlapping drops dirty keys for which io hasn't started, | |
1055 | * so we still want to call it. | |
1056 | */ | |
cafe5635 | 1057 | if (bio->bi_rw & REQ_DISCARD) |
220bb38c | 1058 | s->iop.bypass = true; |
cafe5635 | 1059 | |
72c27061 KO |
1060 | if (should_writeback(dc, s->orig_bio, |
1061 | cache_mode(dc, bio), | |
220bb38c KO |
1062 | s->iop.bypass)) { |
1063 | s->iop.bypass = false; | |
1064 | s->iop.writeback = true; | |
72c27061 KO |
1065 | } |
1066 | ||
220bb38c KO |
1067 | if (s->iop.bypass) { |
1068 | s->iop.bio = s->orig_bio; | |
1069 | bio_get(s->iop.bio); | |
cafe5635 | 1070 | |
84f0db03 KO |
1071 | if (!(bio->bi_rw & REQ_DISCARD) || |
1072 | blk_queue_discard(bdev_get_queue(dc->bdev))) | |
1073 | closure_bio_submit(bio, cl, s->d); | |
220bb38c | 1074 | } else if (s->iop.writeback) { |
279afbad | 1075 | bch_writeback_add(dc); |
220bb38c | 1076 | s->iop.bio = bio; |
e49c7c37 | 1077 | |
c0f04d88 | 1078 | if (bio->bi_rw & REQ_FLUSH) { |
e49c7c37 | 1079 | /* Also need to send a flush to the backing device */ |
d4eddd42 | 1080 | struct bio *flush = bio_alloc_bioset(GFP_NOIO, 0, |
c0f04d88 | 1081 | dc->disk.bio_split); |
e49c7c37 | 1082 | |
c0f04d88 KO |
1083 | flush->bi_rw = WRITE_FLUSH; |
1084 | flush->bi_bdev = bio->bi_bdev; | |
1085 | flush->bi_end_io = request_endio; | |
1086 | flush->bi_private = cl; | |
1087 | ||
1088 | closure_bio_submit(flush, cl, s->d); | |
e49c7c37 | 1089 | } |
84f0db03 | 1090 | } else { |
59d276fe | 1091 | s->iop.bio = bio_clone_fast(bio, GFP_NOIO, dc->disk.bio_split); |
84f0db03 KO |
1092 | |
1093 | closure_bio_submit(bio, cl, s->d); | |
cafe5635 | 1094 | } |
84f0db03 | 1095 | |
220bb38c | 1096 | closure_call(&s->iop.cl, bch_data_insert, NULL, cl); |
cafe5635 | 1097 | continue_at(cl, cached_dev_write_complete, NULL); |
cafe5635 KO |
1098 | } |
1099 | ||
a34a8bfd | 1100 | static void cached_dev_nodata(struct closure *cl) |
cafe5635 | 1101 | { |
a34a8bfd | 1102 | struct search *s = container_of(cl, struct search, cl); |
cafe5635 KO |
1103 | struct bio *bio = &s->bio.bio; |
1104 | ||
220bb38c KO |
1105 | if (s->iop.flush_journal) |
1106 | bch_journal_meta(s->iop.c, cl); | |
cafe5635 | 1107 | |
84f0db03 | 1108 | /* If it's a flush, we send the flush to the backing device too */ |
cafe5635 KO |
1109 | closure_bio_submit(bio, cl, s->d); |
1110 | ||
1111 | continue_at(cl, cached_dev_bio_complete, NULL); | |
1112 | } | |
1113 | ||
1114 | /* Cached devices - read & write stuff */ | |
1115 | ||
cafe5635 KO |
1116 | static void cached_dev_make_request(struct request_queue *q, struct bio *bio) |
1117 | { | |
1118 | struct search *s; | |
1119 | struct bcache_device *d = bio->bi_bdev->bd_disk->private_data; | |
1120 | struct cached_dev *dc = container_of(d, struct cached_dev, disk); | |
1121 | int cpu, rw = bio_data_dir(bio); | |
1122 | ||
1123 | cpu = part_stat_lock(); | |
1124 | part_stat_inc(cpu, &d->disk->part0, ios[rw]); | |
1125 | part_stat_add(cpu, &d->disk->part0, sectors[rw], bio_sectors(bio)); | |
1126 | part_stat_unlock(); | |
1127 | ||
1128 | bio->bi_bdev = dc->bdev; | |
4f024f37 | 1129 | bio->bi_iter.bi_sector += dc->sb.data_offset; |
cafe5635 KO |
1130 | |
1131 | if (cached_dev_get(dc)) { | |
1132 | s = search_alloc(bio, d); | |
220bb38c | 1133 | trace_bcache_request_start(s->d, bio); |
cafe5635 | 1134 | |
4f024f37 | 1135 | if (!bio->bi_iter.bi_size) { |
a34a8bfd KO |
1136 | /* |
1137 | * can't call bch_journal_meta from under | |
1138 | * generic_make_request | |
1139 | */ | |
1140 | continue_at_nobarrier(&s->cl, | |
1141 | cached_dev_nodata, | |
1142 | bcache_wq); | |
1143 | } else { | |
220bb38c | 1144 | s->iop.bypass = check_should_bypass(dc, bio); |
84f0db03 KO |
1145 | |
1146 | if (rw) | |
cdd972b1 | 1147 | cached_dev_write(dc, s); |
84f0db03 | 1148 | else |
cdd972b1 | 1149 | cached_dev_read(dc, s); |
84f0db03 | 1150 | } |
cafe5635 KO |
1151 | } else { |
1152 | if ((bio->bi_rw & REQ_DISCARD) && | |
1153 | !blk_queue_discard(bdev_get_queue(dc->bdev))) | |
1154 | bio_endio(bio, 0); | |
1155 | else | |
1156 | bch_generic_make_request(bio, &d->bio_split_hook); | |
1157 | } | |
1158 | } | |
1159 | ||
1160 | static int cached_dev_ioctl(struct bcache_device *d, fmode_t mode, | |
1161 | unsigned int cmd, unsigned long arg) | |
1162 | { | |
1163 | struct cached_dev *dc = container_of(d, struct cached_dev, disk); | |
1164 | return __blkdev_driver_ioctl(dc->bdev, mode, cmd, arg); | |
1165 | } | |
1166 | ||
1167 | static int cached_dev_congested(void *data, int bits) | |
1168 | { | |
1169 | struct bcache_device *d = data; | |
1170 | struct cached_dev *dc = container_of(d, struct cached_dev, disk); | |
1171 | struct request_queue *q = bdev_get_queue(dc->bdev); | |
1172 | int ret = 0; | |
1173 | ||
1174 | if (bdi_congested(&q->backing_dev_info, bits)) | |
1175 | return 1; | |
1176 | ||
1177 | if (cached_dev_get(dc)) { | |
1178 | unsigned i; | |
1179 | struct cache *ca; | |
1180 | ||
1181 | for_each_cache(ca, d->c, i) { | |
1182 | q = bdev_get_queue(ca->bdev); | |
1183 | ret |= bdi_congested(&q->backing_dev_info, bits); | |
1184 | } | |
1185 | ||
1186 | cached_dev_put(dc); | |
1187 | } | |
1188 | ||
1189 | return ret; | |
1190 | } | |
1191 | ||
1192 | void bch_cached_dev_request_init(struct cached_dev *dc) | |
1193 | { | |
1194 | struct gendisk *g = dc->disk.disk; | |
1195 | ||
1196 | g->queue->make_request_fn = cached_dev_make_request; | |
1197 | g->queue->backing_dev_info.congested_fn = cached_dev_congested; | |
1198 | dc->disk.cache_miss = cached_dev_cache_miss; | |
1199 | dc->disk.ioctl = cached_dev_ioctl; | |
1200 | } | |
1201 | ||
1202 | /* Flash backed devices */ | |
1203 | ||
1204 | static int flash_dev_cache_miss(struct btree *b, struct search *s, | |
1205 | struct bio *bio, unsigned sectors) | |
1206 | { | |
7988613b KO |
1207 | struct bio_vec bv; |
1208 | struct bvec_iter iter; | |
8e51e414 | 1209 | |
cafe5635 KO |
1210 | /* Zero fill bio */ |
1211 | ||
7988613b KO |
1212 | bio_for_each_segment(bv, bio, iter) { |
1213 | unsigned j = min(bv.bv_len >> 9, sectors); | |
cafe5635 | 1214 | |
7988613b KO |
1215 | void *p = kmap(bv.bv_page); |
1216 | memset(p + bv.bv_offset, 0, j << 9); | |
1217 | kunmap(bv.bv_page); | |
cafe5635 | 1218 | |
8e51e414 | 1219 | sectors -= j; |
cafe5635 KO |
1220 | } |
1221 | ||
4f024f37 | 1222 | bio_advance(bio, min(sectors << 9, bio->bi_iter.bi_size)); |
8e51e414 | 1223 | |
4f024f37 | 1224 | if (!bio->bi_iter.bi_size) |
2c1953e2 | 1225 | return MAP_DONE; |
cafe5635 | 1226 | |
2c1953e2 | 1227 | return MAP_CONTINUE; |
cafe5635 KO |
1228 | } |
1229 | ||
a34a8bfd KO |
1230 | static void flash_dev_nodata(struct closure *cl) |
1231 | { | |
1232 | struct search *s = container_of(cl, struct search, cl); | |
1233 | ||
220bb38c KO |
1234 | if (s->iop.flush_journal) |
1235 | bch_journal_meta(s->iop.c, cl); | |
a34a8bfd KO |
1236 | |
1237 | continue_at(cl, search_free, NULL); | |
1238 | } | |
1239 | ||
cafe5635 KO |
1240 | static void flash_dev_make_request(struct request_queue *q, struct bio *bio) |
1241 | { | |
1242 | struct search *s; | |
1243 | struct closure *cl; | |
1244 | struct bcache_device *d = bio->bi_bdev->bd_disk->private_data; | |
1245 | int cpu, rw = bio_data_dir(bio); | |
1246 | ||
1247 | cpu = part_stat_lock(); | |
1248 | part_stat_inc(cpu, &d->disk->part0, ios[rw]); | |
1249 | part_stat_add(cpu, &d->disk->part0, sectors[rw], bio_sectors(bio)); | |
1250 | part_stat_unlock(); | |
1251 | ||
1252 | s = search_alloc(bio, d); | |
1253 | cl = &s->cl; | |
1254 | bio = &s->bio.bio; | |
1255 | ||
220bb38c | 1256 | trace_bcache_request_start(s->d, bio); |
cafe5635 | 1257 | |
4f024f37 | 1258 | if (!bio->bi_iter.bi_size) { |
a34a8bfd KO |
1259 | /* |
1260 | * can't call bch_journal_meta from under | |
1261 | * generic_make_request | |
1262 | */ | |
1263 | continue_at_nobarrier(&s->cl, | |
1264 | flash_dev_nodata, | |
1265 | bcache_wq); | |
84f0db03 | 1266 | } else if (rw) { |
220bb38c | 1267 | bch_keybuf_check_overlapping(&s->iop.c->moving_gc_keys, |
4f024f37 | 1268 | &KEY(d->id, bio->bi_iter.bi_sector, 0), |
8e51e414 | 1269 | &KEY(d->id, bio_end_sector(bio), 0)); |
cafe5635 | 1270 | |
220bb38c KO |
1271 | s->iop.bypass = (bio->bi_rw & REQ_DISCARD) != 0; |
1272 | s->iop.writeback = true; | |
1273 | s->iop.bio = bio; | |
cafe5635 | 1274 | |
220bb38c | 1275 | closure_call(&s->iop.cl, bch_data_insert, NULL, cl); |
cafe5635 | 1276 | } else { |
220bb38c | 1277 | closure_call(&s->iop.cl, cache_lookup, NULL, cl); |
cafe5635 KO |
1278 | } |
1279 | ||
1280 | continue_at(cl, search_free, NULL); | |
1281 | } | |
1282 | ||
1283 | static int flash_dev_ioctl(struct bcache_device *d, fmode_t mode, | |
1284 | unsigned int cmd, unsigned long arg) | |
1285 | { | |
1286 | return -ENOTTY; | |
1287 | } | |
1288 | ||
1289 | static int flash_dev_congested(void *data, int bits) | |
1290 | { | |
1291 | struct bcache_device *d = data; | |
1292 | struct request_queue *q; | |
1293 | struct cache *ca; | |
1294 | unsigned i; | |
1295 | int ret = 0; | |
1296 | ||
1297 | for_each_cache(ca, d->c, i) { | |
1298 | q = bdev_get_queue(ca->bdev); | |
1299 | ret |= bdi_congested(&q->backing_dev_info, bits); | |
1300 | } | |
1301 | ||
1302 | return ret; | |
1303 | } | |
1304 | ||
1305 | void bch_flash_dev_request_init(struct bcache_device *d) | |
1306 | { | |
1307 | struct gendisk *g = d->disk; | |
1308 | ||
1309 | g->queue->make_request_fn = flash_dev_make_request; | |
1310 | g->queue->backing_dev_info.congested_fn = flash_dev_congested; | |
1311 | d->cache_miss = flash_dev_cache_miss; | |
1312 | d->ioctl = flash_dev_ioctl; | |
1313 | } | |
1314 | ||
1315 | void bch_request_exit(void) | |
1316 | { | |
1317 | #ifdef CONFIG_CGROUP_BCACHE | |
1318 | cgroup_unload_subsys(&bcache_subsys); | |
1319 | #endif | |
1320 | if (bch_search_cache) | |
1321 | kmem_cache_destroy(bch_search_cache); | |
1322 | } | |
1323 | ||
1324 | int __init bch_request_init(void) | |
1325 | { | |
1326 | bch_search_cache = KMEM_CACHE(search, 0); | |
1327 | if (!bch_search_cache) | |
1328 | return -ENOMEM; | |
1329 | ||
1330 | #ifdef CONFIG_CGROUP_BCACHE | |
1331 | cgroup_load_subsys(&bcache_subsys); | |
1332 | init_bch_cgroup(&bcache_default_cgroup); | |
1333 | ||
1334 | cgroup_add_cftypes(&bcache_subsys, bch_files); | |
1335 | #endif | |
1336 | return 0; | |
1337 | } |