xref: /openbmc/linux/drivers/md/bcache/request.c (revision 99fee508)
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"
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, struct bio *bio)
30 {
31 	return BDEV_CACHE_MODE(&dc->sb);
32 }
33 
34 static bool verify(struct cached_dev *dc, struct bio *bio)
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, bio);
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 	if (bio->bi_iter.bi_sector & (c->sb.block_size - 1) ||
388 	    bio_sectors(bio) & (c->sb.block_size - 1)) {
389 		pr_debug("skipping unaligned io");
390 		goto skip;
391 	}
392 
393 	if (bypass_torture_test(dc)) {
394 		if ((get_random_int() & 3) == 3)
395 			goto skip;
396 		else
397 			goto rescale;
398 	}
399 
400 	if (!congested && !dc->sequential_cutoff)
401 		goto rescale;
402 
403 	spin_lock(&dc->io_lock);
404 
405 	hlist_for_each_entry(i, iohash(dc, bio->bi_iter.bi_sector), hash)
406 		if (i->last == bio->bi_iter.bi_sector &&
407 		    time_before(jiffies, i->jiffies))
408 			goto found;
409 
410 	i = list_first_entry(&dc->io_lru, struct io, lru);
411 
412 	add_sequential(task);
413 	i->sequential = 0;
414 found:
415 	if (i->sequential + bio->bi_iter.bi_size > i->sequential)
416 		i->sequential	+= bio->bi_iter.bi_size;
417 
418 	i->last			 = bio_end_sector(bio);
419 	i->jiffies		 = jiffies + msecs_to_jiffies(5000);
420 	task->sequential_io	 = i->sequential;
421 
422 	hlist_del(&i->hash);
423 	hlist_add_head(&i->hash, iohash(dc, i->last));
424 	list_move_tail(&i->lru, &dc->io_lru);
425 
426 	spin_unlock(&dc->io_lock);
427 
428 	sectors = max(task->sequential_io,
429 		      task->sequential_io_avg) >> 9;
430 
431 	if (dc->sequential_cutoff &&
432 	    sectors >= dc->sequential_cutoff >> 9) {
433 		trace_bcache_bypass_sequential(bio);
434 		goto skip;
435 	}
436 
437 	if (congested && sectors >= congested) {
438 		trace_bcache_bypass_congested(bio);
439 		goto skip;
440 	}
441 
442 rescale:
443 	bch_rescale_priorities(c, bio_sectors(bio));
444 	return false;
445 skip:
446 	bch_mark_sectors_bypassed(c, dc, bio_sectors(bio));
447 	return true;
448 }
449 
450 /* Cache lookup */
451 
452 struct search {
453 	/* Stack frame for bio_complete */
454 	struct closure		cl;
455 
456 	struct bbio		bio;
457 	struct bio		*orig_bio;
458 	struct bio		*cache_miss;
459 	struct bcache_device	*d;
460 
461 	unsigned		insert_bio_sectors;
462 	unsigned		recoverable:1;
463 	unsigned		write:1;
464 	unsigned		read_dirty_data:1;
465 
466 	unsigned long		start_time;
467 
468 	struct btree_op		op;
469 	struct data_insert_op	iop;
470 };
471 
472 static void bch_cache_read_endio(struct bio *bio)
473 {
474 	struct bbio *b = container_of(bio, struct bbio, bio);
475 	struct closure *cl = bio->bi_private;
476 	struct search *s = container_of(cl, struct search, cl);
477 
478 	/*
479 	 * If the bucket was reused while our bio was in flight, we might have
480 	 * read the wrong data. Set s->error but not error so it doesn't get
481 	 * counted against the cache device, but we'll still reread the data
482 	 * from the backing device.
483 	 */
484 
485 	if (bio->bi_status)
486 		s->iop.status = bio->bi_status;
487 	else if (!KEY_DIRTY(&b->key) &&
488 		 ptr_stale(s->iop.c, &b->key, 0)) {
489 		atomic_long_inc(&s->iop.c->cache_read_races);
490 		s->iop.status = BLK_STS_IOERR;
491 	}
492 
493 	bch_bbio_endio(s->iop.c, bio, bio->bi_status, "reading from cache");
494 }
495 
496 /*
497  * Read from a single key, handling the initial cache miss if the key starts in
498  * the middle of the bio
499  */
500 static int cache_lookup_fn(struct btree_op *op, struct btree *b, struct bkey *k)
501 {
502 	struct search *s = container_of(op, struct search, op);
503 	struct bio *n, *bio = &s->bio.bio;
504 	struct bkey *bio_key;
505 	unsigned ptr;
506 
507 	if (bkey_cmp(k, &KEY(s->iop.inode, bio->bi_iter.bi_sector, 0)) <= 0)
508 		return MAP_CONTINUE;
509 
510 	if (KEY_INODE(k) != s->iop.inode ||
511 	    KEY_START(k) > bio->bi_iter.bi_sector) {
512 		unsigned bio_sectors = bio_sectors(bio);
513 		unsigned sectors = KEY_INODE(k) == s->iop.inode
514 			? min_t(uint64_t, INT_MAX,
515 				KEY_START(k) - bio->bi_iter.bi_sector)
516 			: INT_MAX;
517 
518 		int ret = s->d->cache_miss(b, s, bio, sectors);
519 		if (ret != MAP_CONTINUE)
520 			return ret;
521 
522 		/* if this was a complete miss we shouldn't get here */
523 		BUG_ON(bio_sectors <= sectors);
524 	}
525 
526 	if (!KEY_SIZE(k))
527 		return MAP_CONTINUE;
528 
529 	/* XXX: figure out best pointer - for multiple cache devices */
530 	ptr = 0;
531 
532 	PTR_BUCKET(b->c, k, ptr)->prio = INITIAL_PRIO;
533 
534 	if (KEY_DIRTY(k))
535 		s->read_dirty_data = true;
536 
537 	n = bio_next_split(bio, min_t(uint64_t, INT_MAX,
538 				      KEY_OFFSET(k) - bio->bi_iter.bi_sector),
539 			   GFP_NOIO, s->d->bio_split);
540 
541 	bio_key = &container_of(n, struct bbio, bio)->key;
542 	bch_bkey_copy_single_ptr(bio_key, k, ptr);
543 
544 	bch_cut_front(&KEY(s->iop.inode, n->bi_iter.bi_sector, 0), bio_key);
545 	bch_cut_back(&KEY(s->iop.inode, bio_end_sector(n), 0), bio_key);
546 
547 	n->bi_end_io	= bch_cache_read_endio;
548 	n->bi_private	= &s->cl;
549 
550 	/*
551 	 * The bucket we're reading from might be reused while our bio
552 	 * is in flight, and we could then end up reading the wrong
553 	 * data.
554 	 *
555 	 * We guard against this by checking (in cache_read_endio()) if
556 	 * the pointer is stale again; if so, we treat it as an error
557 	 * and reread from the backing device (but we don't pass that
558 	 * error up anywhere).
559 	 */
560 
561 	__bch_submit_bbio(n, b->c);
562 	return n == bio ? MAP_DONE : MAP_CONTINUE;
563 }
564 
565 static void cache_lookup(struct closure *cl)
566 {
567 	struct search *s = container_of(cl, struct search, iop.cl);
568 	struct bio *bio = &s->bio.bio;
569 	int ret;
570 
571 	bch_btree_op_init(&s->op, -1);
572 
573 	ret = bch_btree_map_keys(&s->op, s->iop.c,
574 				 &KEY(s->iop.inode, bio->bi_iter.bi_sector, 0),
575 				 cache_lookup_fn, MAP_END_KEY);
576 	if (ret == -EAGAIN) {
577 		continue_at(cl, cache_lookup, bcache_wq);
578 		return;
579 	}
580 
581 	closure_return(cl);
582 }
583 
584 /* Common code for the make_request functions */
585 
586 static void request_endio(struct bio *bio)
587 {
588 	struct closure *cl = bio->bi_private;
589 
590 	if (bio->bi_status) {
591 		struct search *s = container_of(cl, struct search, cl);
592 		s->iop.status = bio->bi_status;
593 		/* Only cache read errors are recoverable */
594 		s->recoverable = false;
595 	}
596 
597 	bio_put(bio);
598 	closure_put(cl);
599 }
600 
601 static void bio_complete(struct search *s)
602 {
603 	if (s->orig_bio) {
604 		struct request_queue *q = s->orig_bio->bi_disk->queue;
605 		generic_end_io_acct(q, bio_data_dir(s->orig_bio),
606 				    &s->d->disk->part0, s->start_time);
607 
608 		trace_bcache_request_end(s->d, s->orig_bio);
609 		s->orig_bio->bi_status = s->iop.status;
610 		bio_endio(s->orig_bio);
611 		s->orig_bio = NULL;
612 	}
613 }
614 
615 static void do_bio_hook(struct search *s, struct bio *orig_bio)
616 {
617 	struct bio *bio = &s->bio.bio;
618 
619 	bio_init(bio, NULL, 0);
620 	__bio_clone_fast(bio, orig_bio);
621 	bio->bi_end_io		= request_endio;
622 	bio->bi_private		= &s->cl;
623 
624 	bio_cnt_set(bio, 3);
625 }
626 
627 static void search_free(struct closure *cl)
628 {
629 	struct search *s = container_of(cl, struct search, cl);
630 	bio_complete(s);
631 
632 	if (s->iop.bio)
633 		bio_put(s->iop.bio);
634 
635 	closure_debug_destroy(cl);
636 	mempool_free(s, s->d->c->search);
637 }
638 
639 static inline struct search *search_alloc(struct bio *bio,
640 					  struct bcache_device *d)
641 {
642 	struct search *s;
643 
644 	s = mempool_alloc(d->c->search, GFP_NOIO);
645 
646 	closure_init(&s->cl, NULL);
647 	do_bio_hook(s, bio);
648 
649 	s->orig_bio		= bio;
650 	s->cache_miss		= NULL;
651 	s->d			= d;
652 	s->recoverable		= 1;
653 	s->write		= op_is_write(bio_op(bio));
654 	s->read_dirty_data	= 0;
655 	s->start_time		= jiffies;
656 
657 	s->iop.c		= d->c;
658 	s->iop.bio		= NULL;
659 	s->iop.inode		= d->id;
660 	s->iop.write_point	= hash_long((unsigned long) current, 16);
661 	s->iop.write_prio	= 0;
662 	s->iop.status		= 0;
663 	s->iop.flags		= 0;
664 	s->iop.flush_journal	= op_is_flush(bio->bi_opf);
665 	s->iop.wq		= bcache_wq;
666 
667 	return s;
668 }
669 
670 /* Cached devices */
671 
672 static void cached_dev_bio_complete(struct closure *cl)
673 {
674 	struct search *s = container_of(cl, struct search, cl);
675 	struct cached_dev *dc = container_of(s->d, struct cached_dev, disk);
676 
677 	search_free(cl);
678 	cached_dev_put(dc);
679 }
680 
681 /* Process reads */
682 
683 static void cached_dev_cache_miss_done(struct closure *cl)
684 {
685 	struct search *s = container_of(cl, struct search, cl);
686 
687 	if (s->iop.replace_collision)
688 		bch_mark_cache_miss_collision(s->iop.c, s->d);
689 
690 	if (s->iop.bio)
691 		bio_free_pages(s->iop.bio);
692 
693 	cached_dev_bio_complete(cl);
694 }
695 
696 static void cached_dev_read_error(struct closure *cl)
697 {
698 	struct search *s = container_of(cl, struct search, cl);
699 	struct bio *bio = &s->bio.bio;
700 
701 	if (s->recoverable) {
702 		/* Retry from the backing device: */
703 		trace_bcache_read_retry(s->orig_bio);
704 
705 		s->iop.status = 0;
706 		do_bio_hook(s, s->orig_bio);
707 
708 		/* XXX: invalidate cache */
709 
710 		closure_bio_submit(bio, cl);
711 	}
712 
713 	continue_at(cl, cached_dev_cache_miss_done, NULL);
714 }
715 
716 static void cached_dev_read_done(struct closure *cl)
717 {
718 	struct search *s = container_of(cl, struct search, cl);
719 	struct cached_dev *dc = container_of(s->d, struct cached_dev, disk);
720 
721 	/*
722 	 * We had a cache miss; cache_bio now contains data ready to be inserted
723 	 * into the cache.
724 	 *
725 	 * First, we copy the data we just read from cache_bio's bounce buffers
726 	 * to the buffers the original bio pointed to:
727 	 */
728 
729 	if (s->iop.bio) {
730 		bio_reset(s->iop.bio);
731 		s->iop.bio->bi_iter.bi_sector = s->cache_miss->bi_iter.bi_sector;
732 		bio_copy_dev(s->iop.bio, s->cache_miss);
733 		s->iop.bio->bi_iter.bi_size = s->insert_bio_sectors << 9;
734 		bch_bio_map(s->iop.bio, NULL);
735 
736 		bio_copy_data(s->cache_miss, s->iop.bio);
737 
738 		bio_put(s->cache_miss);
739 		s->cache_miss = NULL;
740 	}
741 
742 	if (verify(dc, &s->bio.bio) && s->recoverable && !s->read_dirty_data)
743 		bch_data_verify(dc, s->orig_bio);
744 
745 	bio_complete(s);
746 
747 	if (s->iop.bio &&
748 	    !test_bit(CACHE_SET_STOPPING, &s->iop.c->flags)) {
749 		BUG_ON(!s->iop.replace);
750 		closure_call(&s->iop.cl, bch_data_insert, NULL, cl);
751 	}
752 
753 	continue_at(cl, cached_dev_cache_miss_done, NULL);
754 }
755 
756 static void cached_dev_read_done_bh(struct closure *cl)
757 {
758 	struct search *s = container_of(cl, struct search, cl);
759 	struct cached_dev *dc = container_of(s->d, struct cached_dev, disk);
760 
761 	bch_mark_cache_accounting(s->iop.c, s->d,
762 				  !s->cache_miss, s->iop.bypass);
763 	trace_bcache_read(s->orig_bio, !s->cache_miss, s->iop.bypass);
764 
765 	if (s->iop.status)
766 		continue_at_nobarrier(cl, cached_dev_read_error, bcache_wq);
767 	else if (s->iop.bio || verify(dc, &s->bio.bio))
768 		continue_at_nobarrier(cl, cached_dev_read_done, bcache_wq);
769 	else
770 		continue_at_nobarrier(cl, cached_dev_bio_complete, NULL);
771 }
772 
773 static int cached_dev_cache_miss(struct btree *b, struct search *s,
774 				 struct bio *bio, unsigned sectors)
775 {
776 	int ret = MAP_CONTINUE;
777 	unsigned reada = 0;
778 	struct cached_dev *dc = container_of(s->d, struct cached_dev, disk);
779 	struct bio *miss, *cache_bio;
780 
781 	if (s->cache_miss || s->iop.bypass) {
782 		miss = bio_next_split(bio, sectors, GFP_NOIO, s->d->bio_split);
783 		ret = miss == bio ? MAP_DONE : MAP_CONTINUE;
784 		goto out_submit;
785 	}
786 
787 	if (!(bio->bi_opf & REQ_RAHEAD) &&
788 	    !(bio->bi_opf & REQ_META) &&
789 	    s->iop.c->gc_stats.in_use < CUTOFF_CACHE_READA)
790 		reada = min_t(sector_t, dc->readahead >> 9,
791 			      get_capacity(bio->bi_disk) - bio_end_sector(bio));
792 
793 	s->insert_bio_sectors = min(sectors, bio_sectors(bio) + reada);
794 
795 	s->iop.replace_key = KEY(s->iop.inode,
796 				 bio->bi_iter.bi_sector + s->insert_bio_sectors,
797 				 s->insert_bio_sectors);
798 
799 	ret = bch_btree_insert_check_key(b, &s->op, &s->iop.replace_key);
800 	if (ret)
801 		return ret;
802 
803 	s->iop.replace = true;
804 
805 	miss = bio_next_split(bio, sectors, GFP_NOIO, s->d->bio_split);
806 
807 	/* btree_search_recurse()'s btree iterator is no good anymore */
808 	ret = miss == bio ? MAP_DONE : -EINTR;
809 
810 	cache_bio = bio_alloc_bioset(GFP_NOWAIT,
811 			DIV_ROUND_UP(s->insert_bio_sectors, PAGE_SECTORS),
812 			dc->disk.bio_split);
813 	if (!cache_bio)
814 		goto out_submit;
815 
816 	cache_bio->bi_iter.bi_sector	= miss->bi_iter.bi_sector;
817 	bio_copy_dev(cache_bio, miss);
818 	cache_bio->bi_iter.bi_size	= s->insert_bio_sectors << 9;
819 
820 	cache_bio->bi_end_io	= request_endio;
821 	cache_bio->bi_private	= &s->cl;
822 
823 	bch_bio_map(cache_bio, NULL);
824 	if (bio_alloc_pages(cache_bio, __GFP_NOWARN|GFP_NOIO))
825 		goto out_put;
826 
827 	if (reada)
828 		bch_mark_cache_readahead(s->iop.c, s->d);
829 
830 	s->cache_miss	= miss;
831 	s->iop.bio	= cache_bio;
832 	bio_get(cache_bio);
833 	closure_bio_submit(cache_bio, &s->cl);
834 
835 	return ret;
836 out_put:
837 	bio_put(cache_bio);
838 out_submit:
839 	miss->bi_end_io		= request_endio;
840 	miss->bi_private	= &s->cl;
841 	closure_bio_submit(miss, &s->cl);
842 	return ret;
843 }
844 
845 static void cached_dev_read(struct cached_dev *dc, struct search *s)
846 {
847 	struct closure *cl = &s->cl;
848 
849 	closure_call(&s->iop.cl, cache_lookup, NULL, cl);
850 	continue_at(cl, cached_dev_read_done_bh, NULL);
851 }
852 
853 /* Process writes */
854 
855 static void cached_dev_write_complete(struct closure *cl)
856 {
857 	struct search *s = container_of(cl, struct search, cl);
858 	struct cached_dev *dc = container_of(s->d, struct cached_dev, disk);
859 
860 	up_read_non_owner(&dc->writeback_lock);
861 	cached_dev_bio_complete(cl);
862 }
863 
864 static void cached_dev_write(struct cached_dev *dc, struct search *s)
865 {
866 	struct closure *cl = &s->cl;
867 	struct bio *bio = &s->bio.bio;
868 	struct bkey start = KEY(dc->disk.id, bio->bi_iter.bi_sector, 0);
869 	struct bkey end = KEY(dc->disk.id, bio_end_sector(bio), 0);
870 
871 	bch_keybuf_check_overlapping(&s->iop.c->moving_gc_keys, &start, &end);
872 
873 	down_read_non_owner(&dc->writeback_lock);
874 	if (bch_keybuf_check_overlapping(&dc->writeback_keys, &start, &end)) {
875 		/*
876 		 * We overlap with some dirty data undergoing background
877 		 * writeback, force this write to writeback
878 		 */
879 		s->iop.bypass = false;
880 		s->iop.writeback = true;
881 	}
882 
883 	/*
884 	 * Discards aren't _required_ to do anything, so skipping if
885 	 * check_overlapping returned true is ok
886 	 *
887 	 * But check_overlapping drops dirty keys for which io hasn't started,
888 	 * so we still want to call it.
889 	 */
890 	if (bio_op(bio) == REQ_OP_DISCARD)
891 		s->iop.bypass = true;
892 
893 	if (should_writeback(dc, s->orig_bio,
894 			     cache_mode(dc, bio),
895 			     s->iop.bypass)) {
896 		s->iop.bypass = false;
897 		s->iop.writeback = true;
898 	}
899 
900 	if (s->iop.bypass) {
901 		s->iop.bio = s->orig_bio;
902 		bio_get(s->iop.bio);
903 
904 		if ((bio_op(bio) != REQ_OP_DISCARD) ||
905 		    blk_queue_discard(bdev_get_queue(dc->bdev)))
906 			closure_bio_submit(bio, cl);
907 	} else if (s->iop.writeback) {
908 		bch_writeback_add(dc);
909 		s->iop.bio = bio;
910 
911 		if (bio->bi_opf & REQ_PREFLUSH) {
912 			/* Also need to send a flush to the backing device */
913 			struct bio *flush = bio_alloc_bioset(GFP_NOIO, 0,
914 							     dc->disk.bio_split);
915 
916 			bio_copy_dev(flush, bio);
917 			flush->bi_end_io = request_endio;
918 			flush->bi_private = cl;
919 			flush->bi_opf = REQ_OP_WRITE | REQ_PREFLUSH;
920 
921 			closure_bio_submit(flush, cl);
922 		}
923 	} else {
924 		s->iop.bio = bio_clone_fast(bio, GFP_NOIO, dc->disk.bio_split);
925 
926 		closure_bio_submit(bio, cl);
927 	}
928 
929 	closure_call(&s->iop.cl, bch_data_insert, NULL, cl);
930 	continue_at(cl, cached_dev_write_complete, NULL);
931 }
932 
933 static void cached_dev_nodata(struct closure *cl)
934 {
935 	struct search *s = container_of(cl, struct search, cl);
936 	struct bio *bio = &s->bio.bio;
937 
938 	if (s->iop.flush_journal)
939 		bch_journal_meta(s->iop.c, cl);
940 
941 	/* If it's a flush, we send the flush to the backing device too */
942 	closure_bio_submit(bio, cl);
943 
944 	continue_at(cl, cached_dev_bio_complete, NULL);
945 }
946 
947 /* Cached devices - read & write stuff */
948 
949 static blk_qc_t cached_dev_make_request(struct request_queue *q,
950 					struct bio *bio)
951 {
952 	struct search *s;
953 	struct bcache_device *d = bio->bi_disk->private_data;
954 	struct cached_dev *dc = container_of(d, struct cached_dev, disk);
955 	int rw = bio_data_dir(bio);
956 
957 	generic_start_io_acct(q, rw, bio_sectors(bio), &d->disk->part0);
958 
959 	bio_set_dev(bio, dc->bdev);
960 	bio->bi_iter.bi_sector += dc->sb.data_offset;
961 
962 	if (cached_dev_get(dc)) {
963 		s = search_alloc(bio, d);
964 		trace_bcache_request_start(s->d, bio);
965 
966 		if (!bio->bi_iter.bi_size) {
967 			/*
968 			 * can't call bch_journal_meta from under
969 			 * generic_make_request
970 			 */
971 			continue_at_nobarrier(&s->cl,
972 					      cached_dev_nodata,
973 					      bcache_wq);
974 		} else {
975 			s->iop.bypass = check_should_bypass(dc, bio);
976 
977 			if (rw)
978 				cached_dev_write(dc, s);
979 			else
980 				cached_dev_read(dc, s);
981 		}
982 	} else {
983 		if ((bio_op(bio) == REQ_OP_DISCARD) &&
984 		    !blk_queue_discard(bdev_get_queue(dc->bdev)))
985 			bio_endio(bio);
986 		else
987 			generic_make_request(bio);
988 	}
989 
990 	return BLK_QC_T_NONE;
991 }
992 
993 static int cached_dev_ioctl(struct bcache_device *d, fmode_t mode,
994 			    unsigned int cmd, unsigned long arg)
995 {
996 	struct cached_dev *dc = container_of(d, struct cached_dev, disk);
997 	return __blkdev_driver_ioctl(dc->bdev, mode, cmd, arg);
998 }
999 
1000 static int cached_dev_congested(void *data, int bits)
1001 {
1002 	struct bcache_device *d = data;
1003 	struct cached_dev *dc = container_of(d, struct cached_dev, disk);
1004 	struct request_queue *q = bdev_get_queue(dc->bdev);
1005 	int ret = 0;
1006 
1007 	if (bdi_congested(q->backing_dev_info, bits))
1008 		return 1;
1009 
1010 	if (cached_dev_get(dc)) {
1011 		unsigned i;
1012 		struct cache *ca;
1013 
1014 		for_each_cache(ca, d->c, i) {
1015 			q = bdev_get_queue(ca->bdev);
1016 			ret |= bdi_congested(q->backing_dev_info, bits);
1017 		}
1018 
1019 		cached_dev_put(dc);
1020 	}
1021 
1022 	return ret;
1023 }
1024 
1025 void bch_cached_dev_request_init(struct cached_dev *dc)
1026 {
1027 	struct gendisk *g = dc->disk.disk;
1028 
1029 	g->queue->make_request_fn		= cached_dev_make_request;
1030 	g->queue->backing_dev_info->congested_fn = cached_dev_congested;
1031 	dc->disk.cache_miss			= cached_dev_cache_miss;
1032 	dc->disk.ioctl				= cached_dev_ioctl;
1033 }
1034 
1035 /* Flash backed devices */
1036 
1037 static int flash_dev_cache_miss(struct btree *b, struct search *s,
1038 				struct bio *bio, unsigned sectors)
1039 {
1040 	unsigned bytes = min(sectors, bio_sectors(bio)) << 9;
1041 
1042 	swap(bio->bi_iter.bi_size, bytes);
1043 	zero_fill_bio(bio);
1044 	swap(bio->bi_iter.bi_size, bytes);
1045 
1046 	bio_advance(bio, bytes);
1047 
1048 	if (!bio->bi_iter.bi_size)
1049 		return MAP_DONE;
1050 
1051 	return MAP_CONTINUE;
1052 }
1053 
1054 static void flash_dev_nodata(struct closure *cl)
1055 {
1056 	struct search *s = container_of(cl, struct search, cl);
1057 
1058 	if (s->iop.flush_journal)
1059 		bch_journal_meta(s->iop.c, cl);
1060 
1061 	continue_at(cl, search_free, NULL);
1062 }
1063 
1064 static blk_qc_t flash_dev_make_request(struct request_queue *q,
1065 					     struct bio *bio)
1066 {
1067 	struct search *s;
1068 	struct closure *cl;
1069 	struct bcache_device *d = bio->bi_disk->private_data;
1070 	int rw = bio_data_dir(bio);
1071 
1072 	generic_start_io_acct(q, rw, bio_sectors(bio), &d->disk->part0);
1073 
1074 	s = search_alloc(bio, d);
1075 	cl = &s->cl;
1076 	bio = &s->bio.bio;
1077 
1078 	trace_bcache_request_start(s->d, bio);
1079 
1080 	if (!bio->bi_iter.bi_size) {
1081 		/*
1082 		 * can't call bch_journal_meta from under
1083 		 * generic_make_request
1084 		 */
1085 		continue_at_nobarrier(&s->cl,
1086 				      flash_dev_nodata,
1087 				      bcache_wq);
1088 		return BLK_QC_T_NONE;
1089 	} else if (rw) {
1090 		bch_keybuf_check_overlapping(&s->iop.c->moving_gc_keys,
1091 					&KEY(d->id, bio->bi_iter.bi_sector, 0),
1092 					&KEY(d->id, bio_end_sector(bio), 0));
1093 
1094 		s->iop.bypass		= (bio_op(bio) == REQ_OP_DISCARD) != 0;
1095 		s->iop.writeback	= true;
1096 		s->iop.bio		= bio;
1097 
1098 		closure_call(&s->iop.cl, bch_data_insert, NULL, cl);
1099 	} else {
1100 		closure_call(&s->iop.cl, cache_lookup, NULL, cl);
1101 	}
1102 
1103 	continue_at(cl, search_free, NULL);
1104 	return BLK_QC_T_NONE;
1105 }
1106 
1107 static int flash_dev_ioctl(struct bcache_device *d, fmode_t mode,
1108 			   unsigned int cmd, unsigned long arg)
1109 {
1110 	return -ENOTTY;
1111 }
1112 
1113 static int flash_dev_congested(void *data, int bits)
1114 {
1115 	struct bcache_device *d = data;
1116 	struct request_queue *q;
1117 	struct cache *ca;
1118 	unsigned i;
1119 	int ret = 0;
1120 
1121 	for_each_cache(ca, d->c, i) {
1122 		q = bdev_get_queue(ca->bdev);
1123 		ret |= bdi_congested(q->backing_dev_info, bits);
1124 	}
1125 
1126 	return ret;
1127 }
1128 
1129 void bch_flash_dev_request_init(struct bcache_device *d)
1130 {
1131 	struct gendisk *g = d->disk;
1132 
1133 	g->queue->make_request_fn		= flash_dev_make_request;
1134 	g->queue->backing_dev_info->congested_fn = flash_dev_congested;
1135 	d->cache_miss				= flash_dev_cache_miss;
1136 	d->ioctl				= flash_dev_ioctl;
1137 }
1138 
1139 void bch_request_exit(void)
1140 {
1141 	if (bch_search_cache)
1142 		kmem_cache_destroy(bch_search_cache);
1143 }
1144 
1145 int __init bch_request_init(void)
1146 {
1147 	bch_search_cache = KMEM_CACHE(search, 0);
1148 	if (!bch_search_cache)
1149 		return -ENOMEM;
1150 
1151 	return 0;
1152 }
1153