]> Git Repo - linux.git/blob - drivers/md/bcache/request.c
nvme: fix visibility of "uuid" ns attribute
[linux.git] / drivers / md / bcache / request.c
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 <[email protected]>
6  * Copyright 2012 Google, Inc.
7  */
8
9 #include "bcache.h"
10 #include "btree.h"
11 #include "debug.h"
12 #include "request.h"
13 #include "writeback.h"
14
15 #include <linux/module.h>
16 #include <linux/hash.h>
17 #include <linux/random.h>
18 #include <linux/backing-dev.h>
19
20 #include <trace/events/bcache.h>
21
22 #define CUTOFF_CACHE_ADD        95
23 #define CUTOFF_CACHE_READA      90
24
25 struct kmem_cache *bch_search_cache;
26
27 static void bch_data_insert_start(struct closure *);
28
29 static unsigned cache_mode(struct cached_dev *dc)
30 {
31         return BDEV_CACHE_MODE(&dc->sb);
32 }
33
34 static bool verify(struct cached_dev *dc)
35 {
36         return dc->verify;
37 }
38
39 static void bio_csum(struct bio *bio, struct bkey *k)
40 {
41         struct bio_vec bv;
42         struct bvec_iter iter;
43         uint64_t csum = 0;
44
45         bio_for_each_segment(bv, bio, iter) {
46                 void *d = kmap(bv.bv_page) + bv.bv_offset;
47                 csum = bch_crc64_update(csum, d, bv.bv_len);
48                 kunmap(bv.bv_page);
49         }
50
51         k->ptr[KEY_PTRS(k)] = csum & (~0ULL >> 1);
52 }
53
54 /* Insert data into cache */
55
56 static void bch_data_insert_keys(struct closure *cl)
57 {
58         struct data_insert_op *op = container_of(cl, struct data_insert_op, cl);
59         atomic_t *journal_ref = NULL;
60         struct bkey *replace_key = op->replace ? &op->replace_key : NULL;
61         int ret;
62
63         /*
64          * If we're looping, might already be waiting on
65          * another journal write - can't wait on more than one journal write at
66          * a time
67          *
68          * XXX: this looks wrong
69          */
70 #if 0
71         while (atomic_read(&s->cl.remaining) & CLOSURE_WAITING)
72                 closure_sync(&s->cl);
73 #endif
74
75         if (!op->replace)
76                 journal_ref = bch_journal(op->c, &op->insert_keys,
77                                           op->flush_journal ? cl : NULL);
78
79         ret = bch_btree_insert(op->c, &op->insert_keys,
80                                journal_ref, replace_key);
81         if (ret == -ESRCH) {
82                 op->replace_collision = true;
83         } else if (ret) {
84                 op->status              = BLK_STS_RESOURCE;
85                 op->insert_data_done    = true;
86         }
87
88         if (journal_ref)
89                 atomic_dec_bug(journal_ref);
90
91         if (!op->insert_data_done) {
92                 continue_at(cl, bch_data_insert_start, op->wq);
93                 return;
94         }
95
96         bch_keylist_free(&op->insert_keys);
97         closure_return(cl);
98 }
99
100 static int bch_keylist_realloc(struct keylist *l, unsigned u64s,
101                                struct cache_set *c)
102 {
103         size_t oldsize = bch_keylist_nkeys(l);
104         size_t newsize = oldsize + u64s;
105
106         /*
107          * The journalling code doesn't handle the case where the keys to insert
108          * is bigger than an empty write: If we just return -ENOMEM here,
109          * bio_insert() and bio_invalidate() will insert the keys created so far
110          * and finish the rest when the keylist is empty.
111          */
112         if (newsize * sizeof(uint64_t) > block_bytes(c) - sizeof(struct jset))
113                 return -ENOMEM;
114
115         return __bch_keylist_realloc(l, u64s);
116 }
117
118 static void bch_data_invalidate(struct closure *cl)
119 {
120         struct data_insert_op *op = container_of(cl, struct data_insert_op, cl);
121         struct bio *bio = op->bio;
122
123         pr_debug("invalidating %i sectors from %llu",
124                  bio_sectors(bio), (uint64_t) bio->bi_iter.bi_sector);
125
126         while (bio_sectors(bio)) {
127                 unsigned sectors = min(bio_sectors(bio),
128                                        1U << (KEY_SIZE_BITS - 1));
129
130                 if (bch_keylist_realloc(&op->insert_keys, 2, op->c))
131                         goto out;
132
133                 bio->bi_iter.bi_sector  += sectors;
134                 bio->bi_iter.bi_size    -= sectors << 9;
135
136                 bch_keylist_add(&op->insert_keys,
137                                 &KEY(op->inode, bio->bi_iter.bi_sector, sectors));
138         }
139
140         op->insert_data_done = true;
141         bio_put(bio);
142 out:
143         continue_at(cl, bch_data_insert_keys, op->wq);
144 }
145
146 static void bch_data_insert_error(struct closure *cl)
147 {
148         struct data_insert_op *op = container_of(cl, struct data_insert_op, cl);
149
150         /*
151          * Our data write just errored, which means we've got a bunch of keys to
152          * insert that point to data that wasn't succesfully written.
153          *
154          * We don't have to insert those keys but we still have to invalidate
155          * that region of the cache - so, if we just strip off all the pointers
156          * from the keys we'll accomplish just that.
157          */
158
159         struct bkey *src = op->insert_keys.keys, *dst = op->insert_keys.keys;
160
161         while (src != op->insert_keys.top) {
162                 struct bkey *n = bkey_next(src);
163
164                 SET_KEY_PTRS(src, 0);
165                 memmove(dst, src, bkey_bytes(src));
166
167                 dst = bkey_next(dst);
168                 src = n;
169         }
170
171         op->insert_keys.top = dst;
172
173         bch_data_insert_keys(cl);
174 }
175
176 static void bch_data_insert_endio(struct bio *bio)
177 {
178         struct closure *cl = bio->bi_private;
179         struct data_insert_op *op = container_of(cl, struct data_insert_op, cl);
180
181         if (bio->bi_status) {
182                 /* TODO: We could try to recover from this. */
183                 if (op->writeback)
184                         op->status = bio->bi_status;
185                 else if (!op->replace)
186                         set_closure_fn(cl, bch_data_insert_error, op->wq);
187                 else
188                         set_closure_fn(cl, NULL, NULL);
189         }
190
191         bch_bbio_endio(op->c, bio, bio->bi_status, "writing data to cache");
192 }
193
194 static void bch_data_insert_start(struct closure *cl)
195 {
196         struct data_insert_op *op = container_of(cl, struct data_insert_op, cl);
197         struct bio *bio = op->bio, *n;
198
199         if (op->bypass)
200                 return bch_data_invalidate(cl);
201
202         if (atomic_sub_return(bio_sectors(bio), &op->c->sectors_to_gc) < 0)
203                 wake_up_gc(op->c);
204
205         /*
206          * Journal writes are marked REQ_PREFLUSH; if the original write was a
207          * flush, it'll wait on the journal write.
208          */
209         bio->bi_opf &= ~(REQ_PREFLUSH|REQ_FUA);
210
211         do {
212                 unsigned i;
213                 struct bkey *k;
214                 struct bio_set *split = op->c->bio_split;
215
216                 /* 1 for the device pointer and 1 for the chksum */
217                 if (bch_keylist_realloc(&op->insert_keys,
218                                         3 + (op->csum ? 1 : 0),
219                                         op->c)) {
220                         continue_at(cl, bch_data_insert_keys, op->wq);
221                         return;
222                 }
223
224                 k = op->insert_keys.top;
225                 bkey_init(k);
226                 SET_KEY_INODE(k, op->inode);
227                 SET_KEY_OFFSET(k, bio->bi_iter.bi_sector);
228
229                 if (!bch_alloc_sectors(op->c, k, bio_sectors(bio),
230                                        op->write_point, op->write_prio,
231                                        op->writeback))
232                         goto err;
233
234                 n = bio_next_split(bio, KEY_SIZE(k), GFP_NOIO, split);
235
236                 n->bi_end_io    = bch_data_insert_endio;
237                 n->bi_private   = cl;
238
239                 if (op->writeback) {
240                         SET_KEY_DIRTY(k, true);
241
242                         for (i = 0; i < KEY_PTRS(k); i++)
243                                 SET_GC_MARK(PTR_BUCKET(op->c, k, i),
244                                             GC_MARK_DIRTY);
245                 }
246
247                 SET_KEY_CSUM(k, op->csum);
248                 if (KEY_CSUM(k))
249                         bio_csum(n, k);
250
251                 trace_bcache_cache_insert(k);
252                 bch_keylist_push(&op->insert_keys);
253
254                 bio_set_op_attrs(n, REQ_OP_WRITE, 0);
255                 bch_submit_bbio(n, op->c, k, 0);
256         } while (n != bio);
257
258         op->insert_data_done = true;
259         continue_at(cl, bch_data_insert_keys, op->wq);
260         return;
261 err:
262         /* bch_alloc_sectors() blocks if s->writeback = true */
263         BUG_ON(op->writeback);
264
265         /*
266          * But if it's not a writeback write we'd rather just bail out if
267          * there aren't any buckets ready to write to - it might take awhile and
268          * we might be starving btree writes for gc or something.
269          */
270
271         if (!op->replace) {
272                 /*
273                  * Writethrough write: We can't complete the write until we've
274                  * updated the index. But we don't want to delay the write while
275                  * we wait for buckets to be freed up, so just invalidate the
276                  * rest of the write.
277                  */
278                 op->bypass = true;
279                 return bch_data_invalidate(cl);
280         } else {
281                 /*
282                  * From a cache miss, we can just insert the keys for the data
283                  * we have written or bail out if we didn't do anything.
284                  */
285                 op->insert_data_done = true;
286                 bio_put(bio);
287
288                 if (!bch_keylist_empty(&op->insert_keys))
289                         continue_at(cl, bch_data_insert_keys, op->wq);
290                 else
291                         closure_return(cl);
292         }
293 }
294
295 /**
296  * bch_data_insert - stick some data in the cache
297  *
298  * This is the starting point for any data to end up in a cache device; it could
299  * be from a normal write, or a writeback write, or a write to a flash only
300  * volume - it's also used by the moving garbage collector to compact data in
301  * mostly empty buckets.
302  *
303  * It first writes the data to the cache, creating a list of keys to be inserted
304  * (if the data had to be fragmented there will be multiple keys); after the
305  * data is written it calls bch_journal, and after the keys have been added to
306  * the next journal write they're inserted into the btree.
307  *
308  * It inserts the data in s->cache_bio; bi_sector is used for the key offset,
309  * and op->inode is used for the key inode.
310  *
311  * If s->bypass is true, instead of inserting the data it invalidates the
312  * region of the cache represented by s->cache_bio and op->inode.
313  */
314 void bch_data_insert(struct closure *cl)
315 {
316         struct data_insert_op *op = container_of(cl, struct data_insert_op, cl);
317
318         trace_bcache_write(op->c, op->inode, op->bio,
319                            op->writeback, op->bypass);
320
321         bch_keylist_init(&op->insert_keys);
322         bio_get(op->bio);
323         bch_data_insert_start(cl);
324 }
325
326 /* Congested? */
327
328 unsigned bch_get_congested(struct cache_set *c)
329 {
330         int i;
331         long rand;
332
333         if (!c->congested_read_threshold_us &&
334             !c->congested_write_threshold_us)
335                 return 0;
336
337         i = (local_clock_us() - c->congested_last_us) / 1024;
338         if (i < 0)
339                 return 0;
340
341         i += atomic_read(&c->congested);
342         if (i >= 0)
343                 return 0;
344
345         i += CONGESTED_MAX;
346
347         if (i > 0)
348                 i = fract_exp_two(i, 6);
349
350         rand = get_random_int();
351         i -= bitmap_weight(&rand, BITS_PER_LONG);
352
353         return i > 0 ? i : 1;
354 }
355
356 static void add_sequential(struct task_struct *t)
357 {
358         ewma_add(t->sequential_io_avg,
359                  t->sequential_io, 8, 0);
360
361         t->sequential_io = 0;
362 }
363
364 static struct hlist_head *iohash(struct cached_dev *dc, uint64_t k)
365 {
366         return &dc->io_hash[hash_64(k, RECENT_IO_BITS)];
367 }
368
369 static bool check_should_bypass(struct cached_dev *dc, struct bio *bio)
370 {
371         struct cache_set *c = dc->disk.c;
372         unsigned mode = cache_mode(dc);
373         unsigned sectors, congested = bch_get_congested(c);
374         struct task_struct *task = current;
375         struct io *i;
376
377         if (test_bit(BCACHE_DEV_DETACHING, &dc->disk.flags) ||
378             c->gc_stats.in_use > CUTOFF_CACHE_ADD ||
379             (bio_op(bio) == REQ_OP_DISCARD))
380                 goto skip;
381
382         if (mode == CACHE_MODE_NONE ||
383             (mode == CACHE_MODE_WRITEAROUND &&
384              op_is_write(bio_op(bio))))
385                 goto skip;
386
387         /*
388          * Flag for bypass if the IO is for read-ahead or background,
389          * unless the read-ahead request is for metadata (eg, for gfs2).
390          */
391         if (bio->bi_opf & (REQ_RAHEAD|REQ_BACKGROUND) &&
392             !(bio->bi_opf & REQ_META))
393                 goto skip;
394
395         if (bio->bi_iter.bi_sector & (c->sb.block_size - 1) ||
396             bio_sectors(bio) & (c->sb.block_size - 1)) {
397                 pr_debug("skipping unaligned io");
398                 goto skip;
399         }
400
401         if (bypass_torture_test(dc)) {
402                 if ((get_random_int() & 3) == 3)
403                         goto skip;
404                 else
405                         goto rescale;
406         }
407
408         if (!congested && !dc->sequential_cutoff)
409                 goto rescale;
410
411         spin_lock(&dc->io_lock);
412
413         hlist_for_each_entry(i, iohash(dc, bio->bi_iter.bi_sector), hash)
414                 if (i->last == bio->bi_iter.bi_sector &&
415                     time_before(jiffies, i->jiffies))
416                         goto found;
417
418         i = list_first_entry(&dc->io_lru, struct io, lru);
419
420         add_sequential(task);
421         i->sequential = 0;
422 found:
423         if (i->sequential + bio->bi_iter.bi_size > i->sequential)
424                 i->sequential   += bio->bi_iter.bi_size;
425
426         i->last                  = bio_end_sector(bio);
427         i->jiffies               = jiffies + msecs_to_jiffies(5000);
428         task->sequential_io      = i->sequential;
429
430         hlist_del(&i->hash);
431         hlist_add_head(&i->hash, iohash(dc, i->last));
432         list_move_tail(&i->lru, &dc->io_lru);
433
434         spin_unlock(&dc->io_lock);
435
436         sectors = max(task->sequential_io,
437                       task->sequential_io_avg) >> 9;
438
439         if (dc->sequential_cutoff &&
440             sectors >= dc->sequential_cutoff >> 9) {
441                 trace_bcache_bypass_sequential(bio);
442                 goto skip;
443         }
444
445         if (congested && sectors >= congested) {
446                 trace_bcache_bypass_congested(bio);
447                 goto skip;
448         }
449
450 rescale:
451         bch_rescale_priorities(c, bio_sectors(bio));
452         return false;
453 skip:
454         bch_mark_sectors_bypassed(c, dc, bio_sectors(bio));
455         return true;
456 }
457
458 /* Cache lookup */
459
460 struct search {
461         /* Stack frame for bio_complete */
462         struct closure          cl;
463
464         struct bbio             bio;
465         struct bio              *orig_bio;
466         struct bio              *cache_miss;
467         struct bcache_device    *d;
468
469         unsigned                insert_bio_sectors;
470         unsigned                recoverable:1;
471         unsigned                write:1;
472         unsigned                read_dirty_data:1;
473         unsigned                cache_missed:1;
474
475         unsigned long           start_time;
476
477         struct btree_op         op;
478         struct data_insert_op   iop;
479 };
480
481 static void bch_cache_read_endio(struct bio *bio)
482 {
483         struct bbio *b = container_of(bio, struct bbio, bio);
484         struct closure *cl = bio->bi_private;
485         struct search *s = container_of(cl, struct search, cl);
486
487         /*
488          * If the bucket was reused while our bio was in flight, we might have
489          * read the wrong data. Set s->error but not error so it doesn't get
490          * counted against the cache device, but we'll still reread the data
491          * from the backing device.
492          */
493
494         if (bio->bi_status)
495                 s->iop.status = bio->bi_status;
496         else if (!KEY_DIRTY(&b->key) &&
497                  ptr_stale(s->iop.c, &b->key, 0)) {
498                 atomic_long_inc(&s->iop.c->cache_read_races);
499                 s->iop.status = BLK_STS_IOERR;
500         }
501
502         bch_bbio_endio(s->iop.c, bio, bio->bi_status, "reading from cache");
503 }
504
505 /*
506  * Read from a single key, handling the initial cache miss if the key starts in
507  * the middle of the bio
508  */
509 static int cache_lookup_fn(struct btree_op *op, struct btree *b, struct bkey *k)
510 {
511         struct search *s = container_of(op, struct search, op);
512         struct bio *n, *bio = &s->bio.bio;
513         struct bkey *bio_key;
514         unsigned ptr;
515
516         if (bkey_cmp(k, &KEY(s->iop.inode, bio->bi_iter.bi_sector, 0)) <= 0)
517                 return MAP_CONTINUE;
518
519         if (KEY_INODE(k) != s->iop.inode ||
520             KEY_START(k) > bio->bi_iter.bi_sector) {
521                 unsigned bio_sectors = bio_sectors(bio);
522                 unsigned sectors = KEY_INODE(k) == s->iop.inode
523                         ? min_t(uint64_t, INT_MAX,
524                                 KEY_START(k) - bio->bi_iter.bi_sector)
525                         : INT_MAX;
526
527                 int ret = s->d->cache_miss(b, s, bio, sectors);
528                 if (ret != MAP_CONTINUE)
529                         return ret;
530
531                 /* if this was a complete miss we shouldn't get here */
532                 BUG_ON(bio_sectors <= sectors);
533         }
534
535         if (!KEY_SIZE(k))
536                 return MAP_CONTINUE;
537
538         /* XXX: figure out best pointer - for multiple cache devices */
539         ptr = 0;
540
541         PTR_BUCKET(b->c, k, ptr)->prio = INITIAL_PRIO;
542
543         if (KEY_DIRTY(k))
544                 s->read_dirty_data = true;
545
546         n = bio_next_split(bio, min_t(uint64_t, INT_MAX,
547                                       KEY_OFFSET(k) - bio->bi_iter.bi_sector),
548                            GFP_NOIO, s->d->bio_split);
549
550         bio_key = &container_of(n, struct bbio, bio)->key;
551         bch_bkey_copy_single_ptr(bio_key, k, ptr);
552
553         bch_cut_front(&KEY(s->iop.inode, n->bi_iter.bi_sector, 0), bio_key);
554         bch_cut_back(&KEY(s->iop.inode, bio_end_sector(n), 0), bio_key);
555
556         n->bi_end_io    = bch_cache_read_endio;
557         n->bi_private   = &s->cl;
558
559         /*
560          * The bucket we're reading from might be reused while our bio
561          * is in flight, and we could then end up reading the wrong
562          * data.
563          *
564          * We guard against this by checking (in cache_read_endio()) if
565          * the pointer is stale again; if so, we treat it as an error
566          * and reread from the backing device (but we don't pass that
567          * error up anywhere).
568          */
569
570         __bch_submit_bbio(n, b->c);
571         return n == bio ? MAP_DONE : MAP_CONTINUE;
572 }
573
574 static void cache_lookup(struct closure *cl)
575 {
576         struct search *s = container_of(cl, struct search, iop.cl);
577         struct bio *bio = &s->bio.bio;
578         int ret;
579
580         bch_btree_op_init(&s->op, -1);
581
582         ret = bch_btree_map_keys(&s->op, s->iop.c,
583                                  &KEY(s->iop.inode, bio->bi_iter.bi_sector, 0),
584                                  cache_lookup_fn, MAP_END_KEY);
585         if (ret == -EAGAIN) {
586                 continue_at(cl, cache_lookup, bcache_wq);
587                 return;
588         }
589
590         closure_return(cl);
591 }
592
593 /* Common code for the make_request functions */
594
595 static void request_endio(struct bio *bio)
596 {
597         struct closure *cl = bio->bi_private;
598
599         if (bio->bi_status) {
600                 struct search *s = container_of(cl, struct search, cl);
601                 s->iop.status = bio->bi_status;
602                 /* Only cache read errors are recoverable */
603                 s->recoverable = false;
604         }
605
606         bio_put(bio);
607         closure_put(cl);
608 }
609
610 static void bio_complete(struct search *s)
611 {
612         if (s->orig_bio) {
613                 struct request_queue *q = s->orig_bio->bi_disk->queue;
614                 generic_end_io_acct(q, bio_data_dir(s->orig_bio),
615                                     &s->d->disk->part0, s->start_time);
616
617                 trace_bcache_request_end(s->d, s->orig_bio);
618                 s->orig_bio->bi_status = s->iop.status;
619                 bio_endio(s->orig_bio);
620                 s->orig_bio = NULL;
621         }
622 }
623
624 static void do_bio_hook(struct search *s, struct bio *orig_bio)
625 {
626         struct bio *bio = &s->bio.bio;
627
628         bio_init(bio, NULL, 0);
629         __bio_clone_fast(bio, orig_bio);
630         bio->bi_end_io          = request_endio;
631         bio->bi_private         = &s->cl;
632
633         bio_cnt_set(bio, 3);
634 }
635
636 static void search_free(struct closure *cl)
637 {
638         struct search *s = container_of(cl, struct search, cl);
639         bio_complete(s);
640
641         if (s->iop.bio)
642                 bio_put(s->iop.bio);
643
644         closure_debug_destroy(cl);
645         mempool_free(s, s->d->c->search);
646 }
647
648 static inline struct search *search_alloc(struct bio *bio,
649                                           struct bcache_device *d)
650 {
651         struct search *s;
652
653         s = mempool_alloc(d->c->search, GFP_NOIO);
654
655         closure_init(&s->cl, NULL);
656         do_bio_hook(s, bio);
657
658         s->orig_bio             = bio;
659         s->cache_miss           = NULL;
660         s->cache_missed         = 0;
661         s->d                    = d;
662         s->recoverable          = 1;
663         s->write                = op_is_write(bio_op(bio));
664         s->read_dirty_data      = 0;
665         s->start_time           = jiffies;
666
667         s->iop.c                = d->c;
668         s->iop.bio              = NULL;
669         s->iop.inode            = d->id;
670         s->iop.write_point      = hash_long((unsigned long) current, 16);
671         s->iop.write_prio       = 0;
672         s->iop.status           = 0;
673         s->iop.flags            = 0;
674         s->iop.flush_journal    = op_is_flush(bio->bi_opf);
675         s->iop.wq               = bcache_wq;
676
677         return s;
678 }
679
680 /* Cached devices */
681
682 static void cached_dev_bio_complete(struct closure *cl)
683 {
684         struct search *s = container_of(cl, struct search, cl);
685         struct cached_dev *dc = container_of(s->d, struct cached_dev, disk);
686
687         search_free(cl);
688         cached_dev_put(dc);
689 }
690
691 /* Process reads */
692
693 static void cached_dev_cache_miss_done(struct closure *cl)
694 {
695         struct search *s = container_of(cl, struct search, cl);
696
697         if (s->iop.replace_collision)
698                 bch_mark_cache_miss_collision(s->iop.c, s->d);
699
700         if (s->iop.bio)
701                 bio_free_pages(s->iop.bio);
702
703         cached_dev_bio_complete(cl);
704 }
705
706 static void cached_dev_read_error(struct closure *cl)
707 {
708         struct search *s = container_of(cl, struct search, cl);
709         struct bio *bio = &s->bio.bio;
710         struct cached_dev *dc = container_of(s->d, struct cached_dev, disk);
711
712         /*
713          * If cache device is dirty (dc->has_dirty is non-zero), then
714          * recovery a failed read request from cached device may get a
715          * stale data back. So read failure recovery is only permitted
716          * when cache device is clean.
717          */
718         if (s->recoverable &&
719             (dc && !atomic_read(&dc->has_dirty))) {
720                 /* Retry from the backing device: */
721                 trace_bcache_read_retry(s->orig_bio);
722
723                 s->iop.status = 0;
724                 do_bio_hook(s, s->orig_bio);
725
726                 /* XXX: invalidate cache */
727
728                 closure_bio_submit(bio, cl);
729         }
730
731         continue_at(cl, cached_dev_cache_miss_done, NULL);
732 }
733
734 static void cached_dev_read_done(struct closure *cl)
735 {
736         struct search *s = container_of(cl, struct search, cl);
737         struct cached_dev *dc = container_of(s->d, struct cached_dev, disk);
738
739         /*
740          * We had a cache miss; cache_bio now contains data ready to be inserted
741          * into the cache.
742          *
743          * First, we copy the data we just read from cache_bio's bounce buffers
744          * to the buffers the original bio pointed to:
745          */
746
747         if (s->iop.bio) {
748                 bio_reset(s->iop.bio);
749                 s->iop.bio->bi_iter.bi_sector = s->cache_miss->bi_iter.bi_sector;
750                 bio_copy_dev(s->iop.bio, s->cache_miss);
751                 s->iop.bio->bi_iter.bi_size = s->insert_bio_sectors << 9;
752                 bch_bio_map(s->iop.bio, NULL);
753
754                 bio_copy_data(s->cache_miss, s->iop.bio);
755
756                 bio_put(s->cache_miss);
757                 s->cache_miss = NULL;
758         }
759
760         if (verify(dc) && s->recoverable && !s->read_dirty_data)
761                 bch_data_verify(dc, s->orig_bio);
762
763         bio_complete(s);
764
765         if (s->iop.bio &&
766             !test_bit(CACHE_SET_STOPPING, &s->iop.c->flags)) {
767                 BUG_ON(!s->iop.replace);
768                 closure_call(&s->iop.cl, bch_data_insert, NULL, cl);
769         }
770
771         continue_at(cl, cached_dev_cache_miss_done, NULL);
772 }
773
774 static void cached_dev_read_done_bh(struct closure *cl)
775 {
776         struct search *s = container_of(cl, struct search, cl);
777         struct cached_dev *dc = container_of(s->d, struct cached_dev, disk);
778
779         bch_mark_cache_accounting(s->iop.c, s->d,
780                                   !s->cache_missed, s->iop.bypass);
781         trace_bcache_read(s->orig_bio, !s->cache_miss, s->iop.bypass);
782
783         if (s->iop.status)
784                 continue_at_nobarrier(cl, cached_dev_read_error, bcache_wq);
785         else if (s->iop.bio || verify(dc))
786                 continue_at_nobarrier(cl, cached_dev_read_done, bcache_wq);
787         else
788                 continue_at_nobarrier(cl, cached_dev_bio_complete, NULL);
789 }
790
791 static int cached_dev_cache_miss(struct btree *b, struct search *s,
792                                  struct bio *bio, unsigned sectors)
793 {
794         int ret = MAP_CONTINUE;
795         unsigned reada = 0;
796         struct cached_dev *dc = container_of(s->d, struct cached_dev, disk);
797         struct bio *miss, *cache_bio;
798
799         s->cache_missed = 1;
800
801         if (s->cache_miss || s->iop.bypass) {
802                 miss = bio_next_split(bio, sectors, GFP_NOIO, s->d->bio_split);
803                 ret = miss == bio ? MAP_DONE : MAP_CONTINUE;
804                 goto out_submit;
805         }
806
807         if (!(bio->bi_opf & REQ_RAHEAD) &&
808             !(bio->bi_opf & REQ_META) &&
809             s->iop.c->gc_stats.in_use < CUTOFF_CACHE_READA)
810                 reada = min_t(sector_t, dc->readahead >> 9,
811                               get_capacity(bio->bi_disk) - bio_end_sector(bio));
812
813         s->insert_bio_sectors = min(sectors, bio_sectors(bio) + reada);
814
815         s->iop.replace_key = KEY(s->iop.inode,
816                                  bio->bi_iter.bi_sector + s->insert_bio_sectors,
817                                  s->insert_bio_sectors);
818
819         ret = bch_btree_insert_check_key(b, &s->op, &s->iop.replace_key);
820         if (ret)
821                 return ret;
822
823         s->iop.replace = true;
824
825         miss = bio_next_split(bio, sectors, GFP_NOIO, s->d->bio_split);
826
827         /* btree_search_recurse()'s btree iterator is no good anymore */
828         ret = miss == bio ? MAP_DONE : -EINTR;
829
830         cache_bio = bio_alloc_bioset(GFP_NOWAIT,
831                         DIV_ROUND_UP(s->insert_bio_sectors, PAGE_SECTORS),
832                         dc->disk.bio_split);
833         if (!cache_bio)
834                 goto out_submit;
835
836         cache_bio->bi_iter.bi_sector    = miss->bi_iter.bi_sector;
837         bio_copy_dev(cache_bio, miss);
838         cache_bio->bi_iter.bi_size      = s->insert_bio_sectors << 9;
839
840         cache_bio->bi_end_io    = request_endio;
841         cache_bio->bi_private   = &s->cl;
842
843         bch_bio_map(cache_bio, NULL);
844         if (bio_alloc_pages(cache_bio, __GFP_NOWARN|GFP_NOIO))
845                 goto out_put;
846
847         if (reada)
848                 bch_mark_cache_readahead(s->iop.c, s->d);
849
850         s->cache_miss   = miss;
851         s->iop.bio      = cache_bio;
852         bio_get(cache_bio);
853         closure_bio_submit(cache_bio, &s->cl);
854
855         return ret;
856 out_put:
857         bio_put(cache_bio);
858 out_submit:
859         miss->bi_end_io         = request_endio;
860         miss->bi_private        = &s->cl;
861         closure_bio_submit(miss, &s->cl);
862         return ret;
863 }
864
865 static void cached_dev_read(struct cached_dev *dc, struct search *s)
866 {
867         struct closure *cl = &s->cl;
868
869         closure_call(&s->iop.cl, cache_lookup, NULL, cl);
870         continue_at(cl, cached_dev_read_done_bh, NULL);
871 }
872
873 /* Process writes */
874
875 static void cached_dev_write_complete(struct closure *cl)
876 {
877         struct search *s = container_of(cl, struct search, cl);
878         struct cached_dev *dc = container_of(s->d, struct cached_dev, disk);
879
880         up_read_non_owner(&dc->writeback_lock);
881         cached_dev_bio_complete(cl);
882 }
883
884 static void cached_dev_write(struct cached_dev *dc, struct search *s)
885 {
886         struct closure *cl = &s->cl;
887         struct bio *bio = &s->bio.bio;
888         struct bkey start = KEY(dc->disk.id, bio->bi_iter.bi_sector, 0);
889         struct bkey end = KEY(dc->disk.id, bio_end_sector(bio), 0);
890
891         bch_keybuf_check_overlapping(&s->iop.c->moving_gc_keys, &start, &end);
892
893         down_read_non_owner(&dc->writeback_lock);
894         if (bch_keybuf_check_overlapping(&dc->writeback_keys, &start, &end)) {
895                 /*
896                  * We overlap with some dirty data undergoing background
897                  * writeback, force this write to writeback
898                  */
899                 s->iop.bypass = false;
900                 s->iop.writeback = true;
901         }
902
903         /*
904          * Discards aren't _required_ to do anything, so skipping if
905          * check_overlapping returned true is ok
906          *
907          * But check_overlapping drops dirty keys for which io hasn't started,
908          * so we still want to call it.
909          */
910         if (bio_op(bio) == REQ_OP_DISCARD)
911                 s->iop.bypass = true;
912
913         if (should_writeback(dc, s->orig_bio,
914                              cache_mode(dc),
915                              s->iop.bypass)) {
916                 s->iop.bypass = false;
917                 s->iop.writeback = true;
918         }
919
920         if (s->iop.bypass) {
921                 s->iop.bio = s->orig_bio;
922                 bio_get(s->iop.bio);
923
924                 if ((bio_op(bio) != REQ_OP_DISCARD) ||
925                     blk_queue_discard(bdev_get_queue(dc->bdev)))
926                         closure_bio_submit(bio, cl);
927         } else if (s->iop.writeback) {
928                 bch_writeback_add(dc);
929                 s->iop.bio = bio;
930
931                 if (bio->bi_opf & REQ_PREFLUSH) {
932                         /* Also need to send a flush to the backing device */
933                         struct bio *flush = bio_alloc_bioset(GFP_NOIO, 0,
934                                                              dc->disk.bio_split);
935
936                         bio_copy_dev(flush, bio);
937                         flush->bi_end_io = request_endio;
938                         flush->bi_private = cl;
939                         flush->bi_opf = REQ_OP_WRITE | REQ_PREFLUSH;
940
941                         closure_bio_submit(flush, cl);
942                 }
943         } else {
944                 s->iop.bio = bio_clone_fast(bio, GFP_NOIO, dc->disk.bio_split);
945
946                 closure_bio_submit(bio, cl);
947         }
948
949         closure_call(&s->iop.cl, bch_data_insert, NULL, cl);
950         continue_at(cl, cached_dev_write_complete, NULL);
951 }
952
953 static void cached_dev_nodata(struct closure *cl)
954 {
955         struct search *s = container_of(cl, struct search, cl);
956         struct bio *bio = &s->bio.bio;
957
958         if (s->iop.flush_journal)
959                 bch_journal_meta(s->iop.c, cl);
960
961         /* If it's a flush, we send the flush to the backing device too */
962         closure_bio_submit(bio, cl);
963
964         continue_at(cl, cached_dev_bio_complete, NULL);
965 }
966
967 /* Cached devices - read & write stuff */
968
969 static blk_qc_t cached_dev_make_request(struct request_queue *q,
970                                         struct bio *bio)
971 {
972         struct search *s;
973         struct bcache_device *d = bio->bi_disk->private_data;
974         struct cached_dev *dc = container_of(d, struct cached_dev, disk);
975         int rw = bio_data_dir(bio);
976
977         generic_start_io_acct(q, rw, bio_sectors(bio), &d->disk->part0);
978
979         bio_set_dev(bio, dc->bdev);
980         bio->bi_iter.bi_sector += dc->sb.data_offset;
981
982         if (cached_dev_get(dc)) {
983                 s = search_alloc(bio, d);
984                 trace_bcache_request_start(s->d, bio);
985
986                 if (!bio->bi_iter.bi_size) {
987                         /*
988                          * can't call bch_journal_meta from under
989                          * generic_make_request
990                          */
991                         continue_at_nobarrier(&s->cl,
992                                               cached_dev_nodata,
993                                               bcache_wq);
994                 } else {
995                         s->iop.bypass = check_should_bypass(dc, bio);
996
997                         if (rw)
998                                 cached_dev_write(dc, s);
999                         else
1000                                 cached_dev_read(dc, s);
1001                 }
1002         } else {
1003                 if ((bio_op(bio) == REQ_OP_DISCARD) &&
1004                     !blk_queue_discard(bdev_get_queue(dc->bdev)))
1005                         bio_endio(bio);
1006                 else
1007                         generic_make_request(bio);
1008         }
1009
1010         return BLK_QC_T_NONE;
1011 }
1012
1013 static int cached_dev_ioctl(struct bcache_device *d, fmode_t mode,
1014                             unsigned int cmd, unsigned long arg)
1015 {
1016         struct cached_dev *dc = container_of(d, struct cached_dev, disk);
1017         return __blkdev_driver_ioctl(dc->bdev, mode, cmd, arg);
1018 }
1019
1020 static int cached_dev_congested(void *data, int bits)
1021 {
1022         struct bcache_device *d = data;
1023         struct cached_dev *dc = container_of(d, struct cached_dev, disk);
1024         struct request_queue *q = bdev_get_queue(dc->bdev);
1025         int ret = 0;
1026
1027         if (bdi_congested(q->backing_dev_info, bits))
1028                 return 1;
1029
1030         if (cached_dev_get(dc)) {
1031                 unsigned i;
1032                 struct cache *ca;
1033
1034                 for_each_cache(ca, d->c, i) {
1035                         q = bdev_get_queue(ca->bdev);
1036                         ret |= bdi_congested(q->backing_dev_info, bits);
1037                 }
1038
1039                 cached_dev_put(dc);
1040         }
1041
1042         return ret;
1043 }
1044
1045 void bch_cached_dev_request_init(struct cached_dev *dc)
1046 {
1047         struct gendisk *g = dc->disk.disk;
1048
1049         g->queue->make_request_fn               = cached_dev_make_request;
1050         g->queue->backing_dev_info->congested_fn = cached_dev_congested;
1051         dc->disk.cache_miss                     = cached_dev_cache_miss;
1052         dc->disk.ioctl                          = cached_dev_ioctl;
1053 }
1054
1055 /* Flash backed devices */
1056
1057 static int flash_dev_cache_miss(struct btree *b, struct search *s,
1058                                 struct bio *bio, unsigned sectors)
1059 {
1060         unsigned bytes = min(sectors, bio_sectors(bio)) << 9;
1061
1062         swap(bio->bi_iter.bi_size, bytes);
1063         zero_fill_bio(bio);
1064         swap(bio->bi_iter.bi_size, bytes);
1065
1066         bio_advance(bio, bytes);
1067
1068         if (!bio->bi_iter.bi_size)
1069                 return MAP_DONE;
1070
1071         return MAP_CONTINUE;
1072 }
1073
1074 static void flash_dev_nodata(struct closure *cl)
1075 {
1076         struct search *s = container_of(cl, struct search, cl);
1077
1078         if (s->iop.flush_journal)
1079                 bch_journal_meta(s->iop.c, cl);
1080
1081         continue_at(cl, search_free, NULL);
1082 }
1083
1084 static blk_qc_t flash_dev_make_request(struct request_queue *q,
1085                                              struct bio *bio)
1086 {
1087         struct search *s;
1088         struct closure *cl;
1089         struct bcache_device *d = bio->bi_disk->private_data;
1090         int rw = bio_data_dir(bio);
1091
1092         generic_start_io_acct(q, rw, bio_sectors(bio), &d->disk->part0);
1093
1094         s = search_alloc(bio, d);
1095         cl = &s->cl;
1096         bio = &s->bio.bio;
1097
1098         trace_bcache_request_start(s->d, bio);
1099
1100         if (!bio->bi_iter.bi_size) {
1101                 /*
1102                  * can't call bch_journal_meta from under
1103                  * generic_make_request
1104                  */
1105                 continue_at_nobarrier(&s->cl,
1106                                       flash_dev_nodata,
1107                                       bcache_wq);
1108                 return BLK_QC_T_NONE;
1109         } else if (rw) {
1110                 bch_keybuf_check_overlapping(&s->iop.c->moving_gc_keys,
1111                                         &KEY(d->id, bio->bi_iter.bi_sector, 0),
1112                                         &KEY(d->id, bio_end_sector(bio), 0));
1113
1114                 s->iop.bypass           = (bio_op(bio) == REQ_OP_DISCARD) != 0;
1115                 s->iop.writeback        = true;
1116                 s->iop.bio              = bio;
1117
1118                 closure_call(&s->iop.cl, bch_data_insert, NULL, cl);
1119         } else {
1120                 closure_call(&s->iop.cl, cache_lookup, NULL, cl);
1121         }
1122
1123         continue_at(cl, search_free, NULL);
1124         return BLK_QC_T_NONE;
1125 }
1126
1127 static int flash_dev_ioctl(struct bcache_device *d, fmode_t mode,
1128                            unsigned int cmd, unsigned long arg)
1129 {
1130         return -ENOTTY;
1131 }
1132
1133 static int flash_dev_congested(void *data, int bits)
1134 {
1135         struct bcache_device *d = data;
1136         struct request_queue *q;
1137         struct cache *ca;
1138         unsigned i;
1139         int ret = 0;
1140
1141         for_each_cache(ca, d->c, i) {
1142                 q = bdev_get_queue(ca->bdev);
1143                 ret |= bdi_congested(q->backing_dev_info, bits);
1144         }
1145
1146         return ret;
1147 }
1148
1149 void bch_flash_dev_request_init(struct bcache_device *d)
1150 {
1151         struct gendisk *g = d->disk;
1152
1153         g->queue->make_request_fn               = flash_dev_make_request;
1154         g->queue->backing_dev_info->congested_fn = flash_dev_congested;
1155         d->cache_miss                           = flash_dev_cache_miss;
1156         d->ioctl                                = flash_dev_ioctl;
1157 }
1158
1159 void bch_request_exit(void)
1160 {
1161         if (bch_search_cache)
1162                 kmem_cache_destroy(bch_search_cache);
1163 }
1164
1165 int __init bch_request_init(void)
1166 {
1167         bch_search_cache = KMEM_CACHE(search, 0);
1168         if (!bch_search_cache)
1169                 return -ENOMEM;
1170
1171         return 0;
1172 }
This page took 0.112718 seconds and 4 git commands to generate.