xref: /openbmc/linux/drivers/md/bcache/request.c (revision 82003e04)
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->error		= -ENOMEM;
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_error) {
182 		/* TODO: We could try to recover from this. */
183 		if (op->writeback)
184 			op->error = bio->bi_error;
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_error, "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 (atomic_sub_return(bio_sectors(bio), &op->c->sectors_to_gc) < 0) {
200 		set_gc_sectors(op->c);
201 		wake_up_gc(op->c);
202 	}
203 
204 	if (op->bypass)
205 		return bch_data_invalidate(cl);
206 
207 	/*
208 	 * Journal writes are marked REQ_PREFLUSH; if the original write was a
209 	 * flush, it'll wait on the journal write.
210 	 */
211 	bio->bi_opf &= ~(REQ_PREFLUSH|REQ_FUA);
212 
213 	do {
214 		unsigned i;
215 		struct bkey *k;
216 		struct bio_set *split = op->c->bio_split;
217 
218 		/* 1 for the device pointer and 1 for the chksum */
219 		if (bch_keylist_realloc(&op->insert_keys,
220 					3 + (op->csum ? 1 : 0),
221 					op->c)) {
222 			continue_at(cl, bch_data_insert_keys, op->wq);
223 			return;
224 		}
225 
226 		k = op->insert_keys.top;
227 		bkey_init(k);
228 		SET_KEY_INODE(k, op->inode);
229 		SET_KEY_OFFSET(k, bio->bi_iter.bi_sector);
230 
231 		if (!bch_alloc_sectors(op->c, k, bio_sectors(bio),
232 				       op->write_point, op->write_prio,
233 				       op->writeback))
234 			goto err;
235 
236 		n = bio_next_split(bio, KEY_SIZE(k), GFP_NOIO, split);
237 
238 		n->bi_end_io	= bch_data_insert_endio;
239 		n->bi_private	= cl;
240 
241 		if (op->writeback) {
242 			SET_KEY_DIRTY(k, true);
243 
244 			for (i = 0; i < KEY_PTRS(k); i++)
245 				SET_GC_MARK(PTR_BUCKET(op->c, k, i),
246 					    GC_MARK_DIRTY);
247 		}
248 
249 		SET_KEY_CSUM(k, op->csum);
250 		if (KEY_CSUM(k))
251 			bio_csum(n, k);
252 
253 		trace_bcache_cache_insert(k);
254 		bch_keylist_push(&op->insert_keys);
255 
256 		bio_set_op_attrs(n, REQ_OP_WRITE, 0);
257 		bch_submit_bbio(n, op->c, k, 0);
258 	} while (n != bio);
259 
260 	op->insert_data_done = true;
261 	continue_at(cl, bch_data_insert_keys, op->wq);
262 	return;
263 err:
264 	/* bch_alloc_sectors() blocks if s->writeback = true */
265 	BUG_ON(op->writeback);
266 
267 	/*
268 	 * But if it's not a writeback write we'd rather just bail out if
269 	 * there aren't any buckets ready to write to - it might take awhile and
270 	 * we might be starving btree writes for gc or something.
271 	 */
272 
273 	if (!op->replace) {
274 		/*
275 		 * Writethrough write: We can't complete the write until we've
276 		 * updated the index. But we don't want to delay the write while
277 		 * we wait for buckets to be freed up, so just invalidate the
278 		 * rest of the write.
279 		 */
280 		op->bypass = true;
281 		return bch_data_invalidate(cl);
282 	} else {
283 		/*
284 		 * From a cache miss, we can just insert the keys for the data
285 		 * we have written or bail out if we didn't do anything.
286 		 */
287 		op->insert_data_done = true;
288 		bio_put(bio);
289 
290 		if (!bch_keylist_empty(&op->insert_keys))
291 			continue_at(cl, bch_data_insert_keys, op->wq);
292 		else
293 			closure_return(cl);
294 	}
295 }
296 
297 /**
298  * bch_data_insert - stick some data in the cache
299  *
300  * This is the starting point for any data to end up in a cache device; it could
301  * be from a normal write, or a writeback write, or a write to a flash only
302  * volume - it's also used by the moving garbage collector to compact data in
303  * mostly empty buckets.
304  *
305  * It first writes the data to the cache, creating a list of keys to be inserted
306  * (if the data had to be fragmented there will be multiple keys); after the
307  * data is written it calls bch_journal, and after the keys have been added to
308  * the next journal write they're inserted into the btree.
309  *
310  * It inserts the data in s->cache_bio; bi_sector is used for the key offset,
311  * and op->inode is used for the key inode.
312  *
313  * If s->bypass is true, instead of inserting the data it invalidates the
314  * region of the cache represented by s->cache_bio and op->inode.
315  */
316 void bch_data_insert(struct closure *cl)
317 {
318 	struct data_insert_op *op = container_of(cl, struct data_insert_op, cl);
319 
320 	trace_bcache_write(op->c, op->inode, op->bio,
321 			   op->writeback, op->bypass);
322 
323 	bch_keylist_init(&op->insert_keys);
324 	bio_get(op->bio);
325 	bch_data_insert_start(cl);
326 }
327 
328 /* Congested? */
329 
330 unsigned bch_get_congested(struct cache_set *c)
331 {
332 	int i;
333 	long rand;
334 
335 	if (!c->congested_read_threshold_us &&
336 	    !c->congested_write_threshold_us)
337 		return 0;
338 
339 	i = (local_clock_us() - c->congested_last_us) / 1024;
340 	if (i < 0)
341 		return 0;
342 
343 	i += atomic_read(&c->congested);
344 	if (i >= 0)
345 		return 0;
346 
347 	i += CONGESTED_MAX;
348 
349 	if (i > 0)
350 		i = fract_exp_two(i, 6);
351 
352 	rand = get_random_int();
353 	i -= bitmap_weight(&rand, BITS_PER_LONG);
354 
355 	return i > 0 ? i : 1;
356 }
357 
358 static void add_sequential(struct task_struct *t)
359 {
360 	ewma_add(t->sequential_io_avg,
361 		 t->sequential_io, 8, 0);
362 
363 	t->sequential_io = 0;
364 }
365 
366 static struct hlist_head *iohash(struct cached_dev *dc, uint64_t k)
367 {
368 	return &dc->io_hash[hash_64(k, RECENT_IO_BITS)];
369 }
370 
371 static bool check_should_bypass(struct cached_dev *dc, struct bio *bio)
372 {
373 	struct cache_set *c = dc->disk.c;
374 	unsigned mode = cache_mode(dc, bio);
375 	unsigned sectors, congested = bch_get_congested(c);
376 	struct task_struct *task = current;
377 	struct io *i;
378 
379 	if (test_bit(BCACHE_DEV_DETACHING, &dc->disk.flags) ||
380 	    c->gc_stats.in_use > CUTOFF_CACHE_ADD ||
381 	    (bio_op(bio) == REQ_OP_DISCARD))
382 		goto skip;
383 
384 	if (mode == CACHE_MODE_NONE ||
385 	    (mode == CACHE_MODE_WRITEAROUND &&
386 	     op_is_write(bio_op(bio))))
387 		goto skip;
388 
389 	if (bio->bi_iter.bi_sector & (c->sb.block_size - 1) ||
390 	    bio_sectors(bio) & (c->sb.block_size - 1)) {
391 		pr_debug("skipping unaligned io");
392 		goto skip;
393 	}
394 
395 	if (bypass_torture_test(dc)) {
396 		if ((get_random_int() & 3) == 3)
397 			goto skip;
398 		else
399 			goto rescale;
400 	}
401 
402 	if (!congested && !dc->sequential_cutoff)
403 		goto rescale;
404 
405 	if (!congested &&
406 	    mode == CACHE_MODE_WRITEBACK &&
407 	    op_is_write(bio_op(bio)) &&
408 	    (bio->bi_opf & REQ_SYNC))
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 
474 	unsigned long		start_time;
475 
476 	struct btree_op		op;
477 	struct data_insert_op	iop;
478 };
479 
480 static void bch_cache_read_endio(struct bio *bio)
481 {
482 	struct bbio *b = container_of(bio, struct bbio, bio);
483 	struct closure *cl = bio->bi_private;
484 	struct search *s = container_of(cl, struct search, cl);
485 
486 	/*
487 	 * If the bucket was reused while our bio was in flight, we might have
488 	 * read the wrong data. Set s->error but not error so it doesn't get
489 	 * counted against the cache device, but we'll still reread the data
490 	 * from the backing device.
491 	 */
492 
493 	if (bio->bi_error)
494 		s->iop.error = bio->bi_error;
495 	else if (!KEY_DIRTY(&b->key) &&
496 		 ptr_stale(s->iop.c, &b->key, 0)) {
497 		atomic_long_inc(&s->iop.c->cache_read_races);
498 		s->iop.error = -EINTR;
499 	}
500 
501 	bch_bbio_endio(s->iop.c, bio, bio->bi_error, "reading from cache");
502 }
503 
504 /*
505  * Read from a single key, handling the initial cache miss if the key starts in
506  * the middle of the bio
507  */
508 static int cache_lookup_fn(struct btree_op *op, struct btree *b, struct bkey *k)
509 {
510 	struct search *s = container_of(op, struct search, op);
511 	struct bio *n, *bio = &s->bio.bio;
512 	struct bkey *bio_key;
513 	unsigned ptr;
514 
515 	if (bkey_cmp(k, &KEY(s->iop.inode, bio->bi_iter.bi_sector, 0)) <= 0)
516 		return MAP_CONTINUE;
517 
518 	if (KEY_INODE(k) != s->iop.inode ||
519 	    KEY_START(k) > bio->bi_iter.bi_sector) {
520 		unsigned bio_sectors = bio_sectors(bio);
521 		unsigned sectors = KEY_INODE(k) == s->iop.inode
522 			? min_t(uint64_t, INT_MAX,
523 				KEY_START(k) - bio->bi_iter.bi_sector)
524 			: INT_MAX;
525 
526 		int ret = s->d->cache_miss(b, s, bio, sectors);
527 		if (ret != MAP_CONTINUE)
528 			return ret;
529 
530 		/* if this was a complete miss we shouldn't get here */
531 		BUG_ON(bio_sectors <= sectors);
532 	}
533 
534 	if (!KEY_SIZE(k))
535 		return MAP_CONTINUE;
536 
537 	/* XXX: figure out best pointer - for multiple cache devices */
538 	ptr = 0;
539 
540 	PTR_BUCKET(b->c, k, ptr)->prio = INITIAL_PRIO;
541 
542 	if (KEY_DIRTY(k))
543 		s->read_dirty_data = true;
544 
545 	n = bio_next_split(bio, min_t(uint64_t, INT_MAX,
546 				      KEY_OFFSET(k) - bio->bi_iter.bi_sector),
547 			   GFP_NOIO, s->d->bio_split);
548 
549 	bio_key = &container_of(n, struct bbio, bio)->key;
550 	bch_bkey_copy_single_ptr(bio_key, k, ptr);
551 
552 	bch_cut_front(&KEY(s->iop.inode, n->bi_iter.bi_sector, 0), bio_key);
553 	bch_cut_back(&KEY(s->iop.inode, bio_end_sector(n), 0), bio_key);
554 
555 	n->bi_end_io	= bch_cache_read_endio;
556 	n->bi_private	= &s->cl;
557 
558 	/*
559 	 * The bucket we're reading from might be reused while our bio
560 	 * is in flight, and we could then end up reading the wrong
561 	 * data.
562 	 *
563 	 * We guard against this by checking (in cache_read_endio()) if
564 	 * the pointer is stale again; if so, we treat it as an error
565 	 * and reread from the backing device (but we don't pass that
566 	 * error up anywhere).
567 	 */
568 
569 	__bch_submit_bbio(n, b->c);
570 	return n == bio ? MAP_DONE : MAP_CONTINUE;
571 }
572 
573 static void cache_lookup(struct closure *cl)
574 {
575 	struct search *s = container_of(cl, struct search, iop.cl);
576 	struct bio *bio = &s->bio.bio;
577 	int ret;
578 
579 	bch_btree_op_init(&s->op, -1);
580 
581 	ret = bch_btree_map_keys(&s->op, s->iop.c,
582 				 &KEY(s->iop.inode, bio->bi_iter.bi_sector, 0),
583 				 cache_lookup_fn, MAP_END_KEY);
584 	if (ret == -EAGAIN) {
585 		continue_at(cl, cache_lookup, bcache_wq);
586 		return;
587 	}
588 
589 	closure_return(cl);
590 }
591 
592 /* Common code for the make_request functions */
593 
594 static void request_endio(struct bio *bio)
595 {
596 	struct closure *cl = bio->bi_private;
597 
598 	if (bio->bi_error) {
599 		struct search *s = container_of(cl, struct search, cl);
600 		s->iop.error = bio->bi_error;
601 		/* Only cache read errors are recoverable */
602 		s->recoverable = false;
603 	}
604 
605 	bio_put(bio);
606 	closure_put(cl);
607 }
608 
609 static void bio_complete(struct search *s)
610 {
611 	if (s->orig_bio) {
612 		generic_end_io_acct(bio_data_dir(s->orig_bio),
613 				    &s->d->disk->part0, s->start_time);
614 
615 		trace_bcache_request_end(s->d, s->orig_bio);
616 		s->orig_bio->bi_error = s->iop.error;
617 		bio_endio(s->orig_bio);
618 		s->orig_bio = NULL;
619 	}
620 }
621 
622 static void do_bio_hook(struct search *s, struct bio *orig_bio)
623 {
624 	struct bio *bio = &s->bio.bio;
625 
626 	bio_init(bio);
627 	__bio_clone_fast(bio, orig_bio);
628 	bio->bi_end_io		= request_endio;
629 	bio->bi_private		= &s->cl;
630 
631 	bio_cnt_set(bio, 3);
632 }
633 
634 static void search_free(struct closure *cl)
635 {
636 	struct search *s = container_of(cl, struct search, cl);
637 	bio_complete(s);
638 
639 	if (s->iop.bio)
640 		bio_put(s->iop.bio);
641 
642 	closure_debug_destroy(cl);
643 	mempool_free(s, s->d->c->search);
644 }
645 
646 static inline struct search *search_alloc(struct bio *bio,
647 					  struct bcache_device *d)
648 {
649 	struct search *s;
650 
651 	s = mempool_alloc(d->c->search, GFP_NOIO);
652 
653 	closure_init(&s->cl, NULL);
654 	do_bio_hook(s, bio);
655 
656 	s->orig_bio		= bio;
657 	s->cache_miss		= NULL;
658 	s->d			= d;
659 	s->recoverable		= 1;
660 	s->write		= op_is_write(bio_op(bio));
661 	s->read_dirty_data	= 0;
662 	s->start_time		= jiffies;
663 
664 	s->iop.c		= d->c;
665 	s->iop.bio		= NULL;
666 	s->iop.inode		= d->id;
667 	s->iop.write_point	= hash_long((unsigned long) current, 16);
668 	s->iop.write_prio	= 0;
669 	s->iop.error		= 0;
670 	s->iop.flags		= 0;
671 	s->iop.flush_journal	= (bio->bi_opf & (REQ_PREFLUSH|REQ_FUA)) != 0;
672 	s->iop.wq		= bcache_wq;
673 
674 	return s;
675 }
676 
677 /* Cached devices */
678 
679 static void cached_dev_bio_complete(struct closure *cl)
680 {
681 	struct search *s = container_of(cl, struct search, cl);
682 	struct cached_dev *dc = container_of(s->d, struct cached_dev, disk);
683 
684 	search_free(cl);
685 	cached_dev_put(dc);
686 }
687 
688 /* Process reads */
689 
690 static void cached_dev_cache_miss_done(struct closure *cl)
691 {
692 	struct search *s = container_of(cl, struct search, cl);
693 
694 	if (s->iop.replace_collision)
695 		bch_mark_cache_miss_collision(s->iop.c, s->d);
696 
697 	if (s->iop.bio)
698 		bio_free_pages(s->iop.bio);
699 
700 	cached_dev_bio_complete(cl);
701 }
702 
703 static void cached_dev_read_error(struct closure *cl)
704 {
705 	struct search *s = container_of(cl, struct search, cl);
706 	struct bio *bio = &s->bio.bio;
707 
708 	if (s->recoverable) {
709 		/* Retry from the backing device: */
710 		trace_bcache_read_retry(s->orig_bio);
711 
712 		s->iop.error = 0;
713 		do_bio_hook(s, s->orig_bio);
714 
715 		/* XXX: invalidate cache */
716 
717 		closure_bio_submit(bio, cl);
718 	}
719 
720 	continue_at(cl, cached_dev_cache_miss_done, NULL);
721 }
722 
723 static void cached_dev_read_done(struct closure *cl)
724 {
725 	struct search *s = container_of(cl, struct search, cl);
726 	struct cached_dev *dc = container_of(s->d, struct cached_dev, disk);
727 
728 	/*
729 	 * We had a cache miss; cache_bio now contains data ready to be inserted
730 	 * into the cache.
731 	 *
732 	 * First, we copy the data we just read from cache_bio's bounce buffers
733 	 * to the buffers the original bio pointed to:
734 	 */
735 
736 	if (s->iop.bio) {
737 		bio_reset(s->iop.bio);
738 		s->iop.bio->bi_iter.bi_sector = s->cache_miss->bi_iter.bi_sector;
739 		s->iop.bio->bi_bdev = s->cache_miss->bi_bdev;
740 		s->iop.bio->bi_iter.bi_size = s->insert_bio_sectors << 9;
741 		bch_bio_map(s->iop.bio, NULL);
742 
743 		bio_copy_data(s->cache_miss, s->iop.bio);
744 
745 		bio_put(s->cache_miss);
746 		s->cache_miss = NULL;
747 	}
748 
749 	if (verify(dc, &s->bio.bio) && s->recoverable && !s->read_dirty_data)
750 		bch_data_verify(dc, s->orig_bio);
751 
752 	bio_complete(s);
753 
754 	if (s->iop.bio &&
755 	    !test_bit(CACHE_SET_STOPPING, &s->iop.c->flags)) {
756 		BUG_ON(!s->iop.replace);
757 		closure_call(&s->iop.cl, bch_data_insert, NULL, cl);
758 	}
759 
760 	continue_at(cl, cached_dev_cache_miss_done, NULL);
761 }
762 
763 static void cached_dev_read_done_bh(struct closure *cl)
764 {
765 	struct search *s = container_of(cl, struct search, cl);
766 	struct cached_dev *dc = container_of(s->d, struct cached_dev, disk);
767 
768 	bch_mark_cache_accounting(s->iop.c, s->d,
769 				  !s->cache_miss, s->iop.bypass);
770 	trace_bcache_read(s->orig_bio, !s->cache_miss, s->iop.bypass);
771 
772 	if (s->iop.error)
773 		continue_at_nobarrier(cl, cached_dev_read_error, bcache_wq);
774 	else if (s->iop.bio || verify(dc, &s->bio.bio))
775 		continue_at_nobarrier(cl, cached_dev_read_done, bcache_wq);
776 	else
777 		continue_at_nobarrier(cl, cached_dev_bio_complete, NULL);
778 }
779 
780 static int cached_dev_cache_miss(struct btree *b, struct search *s,
781 				 struct bio *bio, unsigned sectors)
782 {
783 	int ret = MAP_CONTINUE;
784 	unsigned reada = 0;
785 	struct cached_dev *dc = container_of(s->d, struct cached_dev, disk);
786 	struct bio *miss, *cache_bio;
787 
788 	if (s->cache_miss || s->iop.bypass) {
789 		miss = bio_next_split(bio, sectors, GFP_NOIO, s->d->bio_split);
790 		ret = miss == bio ? MAP_DONE : MAP_CONTINUE;
791 		goto out_submit;
792 	}
793 
794 	if (!(bio->bi_opf & REQ_RAHEAD) &&
795 	    !(bio->bi_opf & REQ_META) &&
796 	    s->iop.c->gc_stats.in_use < CUTOFF_CACHE_READA)
797 		reada = min_t(sector_t, dc->readahead >> 9,
798 			      bdev_sectors(bio->bi_bdev) - bio_end_sector(bio));
799 
800 	s->insert_bio_sectors = min(sectors, bio_sectors(bio) + reada);
801 
802 	s->iop.replace_key = KEY(s->iop.inode,
803 				 bio->bi_iter.bi_sector + s->insert_bio_sectors,
804 				 s->insert_bio_sectors);
805 
806 	ret = bch_btree_insert_check_key(b, &s->op, &s->iop.replace_key);
807 	if (ret)
808 		return ret;
809 
810 	s->iop.replace = true;
811 
812 	miss = bio_next_split(bio, sectors, GFP_NOIO, s->d->bio_split);
813 
814 	/* btree_search_recurse()'s btree iterator is no good anymore */
815 	ret = miss == bio ? MAP_DONE : -EINTR;
816 
817 	cache_bio = bio_alloc_bioset(GFP_NOWAIT,
818 			DIV_ROUND_UP(s->insert_bio_sectors, PAGE_SECTORS),
819 			dc->disk.bio_split);
820 	if (!cache_bio)
821 		goto out_submit;
822 
823 	cache_bio->bi_iter.bi_sector	= miss->bi_iter.bi_sector;
824 	cache_bio->bi_bdev		= miss->bi_bdev;
825 	cache_bio->bi_iter.bi_size	= s->insert_bio_sectors << 9;
826 
827 	cache_bio->bi_end_io	= request_endio;
828 	cache_bio->bi_private	= &s->cl;
829 
830 	bch_bio_map(cache_bio, NULL);
831 	if (bio_alloc_pages(cache_bio, __GFP_NOWARN|GFP_NOIO))
832 		goto out_put;
833 
834 	if (reada)
835 		bch_mark_cache_readahead(s->iop.c, s->d);
836 
837 	s->cache_miss	= miss;
838 	s->iop.bio	= cache_bio;
839 	bio_get(cache_bio);
840 	closure_bio_submit(cache_bio, &s->cl);
841 
842 	return ret;
843 out_put:
844 	bio_put(cache_bio);
845 out_submit:
846 	miss->bi_end_io		= request_endio;
847 	miss->bi_private	= &s->cl;
848 	closure_bio_submit(miss, &s->cl);
849 	return ret;
850 }
851 
852 static void cached_dev_read(struct cached_dev *dc, struct search *s)
853 {
854 	struct closure *cl = &s->cl;
855 
856 	closure_call(&s->iop.cl, cache_lookup, NULL, cl);
857 	continue_at(cl, cached_dev_read_done_bh, NULL);
858 }
859 
860 /* Process writes */
861 
862 static void cached_dev_write_complete(struct closure *cl)
863 {
864 	struct search *s = container_of(cl, struct search, cl);
865 	struct cached_dev *dc = container_of(s->d, struct cached_dev, disk);
866 
867 	up_read_non_owner(&dc->writeback_lock);
868 	cached_dev_bio_complete(cl);
869 }
870 
871 static void cached_dev_write(struct cached_dev *dc, struct search *s)
872 {
873 	struct closure *cl = &s->cl;
874 	struct bio *bio = &s->bio.bio;
875 	struct bkey start = KEY(dc->disk.id, bio->bi_iter.bi_sector, 0);
876 	struct bkey end = KEY(dc->disk.id, bio_end_sector(bio), 0);
877 
878 	bch_keybuf_check_overlapping(&s->iop.c->moving_gc_keys, &start, &end);
879 
880 	down_read_non_owner(&dc->writeback_lock);
881 	if (bch_keybuf_check_overlapping(&dc->writeback_keys, &start, &end)) {
882 		/*
883 		 * We overlap with some dirty data undergoing background
884 		 * writeback, force this write to writeback
885 		 */
886 		s->iop.bypass = false;
887 		s->iop.writeback = true;
888 	}
889 
890 	/*
891 	 * Discards aren't _required_ to do anything, so skipping if
892 	 * check_overlapping returned true is ok
893 	 *
894 	 * But check_overlapping drops dirty keys for which io hasn't started,
895 	 * so we still want to call it.
896 	 */
897 	if (bio_op(bio) == REQ_OP_DISCARD)
898 		s->iop.bypass = true;
899 
900 	if (should_writeback(dc, s->orig_bio,
901 			     cache_mode(dc, bio),
902 			     s->iop.bypass)) {
903 		s->iop.bypass = false;
904 		s->iop.writeback = true;
905 	}
906 
907 	if (s->iop.bypass) {
908 		s->iop.bio = s->orig_bio;
909 		bio_get(s->iop.bio);
910 
911 		if ((bio_op(bio) != REQ_OP_DISCARD) ||
912 		    blk_queue_discard(bdev_get_queue(dc->bdev)))
913 			closure_bio_submit(bio, cl);
914 	} else if (s->iop.writeback) {
915 		bch_writeback_add(dc);
916 		s->iop.bio = bio;
917 
918 		if (bio->bi_opf & REQ_PREFLUSH) {
919 			/* Also need to send a flush to the backing device */
920 			struct bio *flush = bio_alloc_bioset(GFP_NOIO, 0,
921 							     dc->disk.bio_split);
922 
923 			flush->bi_bdev	= bio->bi_bdev;
924 			flush->bi_end_io = request_endio;
925 			flush->bi_private = cl;
926 			bio_set_op_attrs(flush, REQ_OP_WRITE, WRITE_FLUSH);
927 
928 			closure_bio_submit(flush, cl);
929 		}
930 	} else {
931 		s->iop.bio = bio_clone_fast(bio, GFP_NOIO, dc->disk.bio_split);
932 
933 		closure_bio_submit(bio, cl);
934 	}
935 
936 	closure_call(&s->iop.cl, bch_data_insert, NULL, cl);
937 	continue_at(cl, cached_dev_write_complete, NULL);
938 }
939 
940 static void cached_dev_nodata(struct closure *cl)
941 {
942 	struct search *s = container_of(cl, struct search, cl);
943 	struct bio *bio = &s->bio.bio;
944 
945 	if (s->iop.flush_journal)
946 		bch_journal_meta(s->iop.c, cl);
947 
948 	/* If it's a flush, we send the flush to the backing device too */
949 	closure_bio_submit(bio, cl);
950 
951 	continue_at(cl, cached_dev_bio_complete, NULL);
952 }
953 
954 /* Cached devices - read & write stuff */
955 
956 static blk_qc_t cached_dev_make_request(struct request_queue *q,
957 					struct bio *bio)
958 {
959 	struct search *s;
960 	struct bcache_device *d = bio->bi_bdev->bd_disk->private_data;
961 	struct cached_dev *dc = container_of(d, struct cached_dev, disk);
962 	int rw = bio_data_dir(bio);
963 
964 	generic_start_io_acct(rw, bio_sectors(bio), &d->disk->part0);
965 
966 	bio->bi_bdev = dc->bdev;
967 	bio->bi_iter.bi_sector += dc->sb.data_offset;
968 
969 	if (cached_dev_get(dc)) {
970 		s = search_alloc(bio, d);
971 		trace_bcache_request_start(s->d, bio);
972 
973 		if (!bio->bi_iter.bi_size) {
974 			/*
975 			 * can't call bch_journal_meta from under
976 			 * generic_make_request
977 			 */
978 			continue_at_nobarrier(&s->cl,
979 					      cached_dev_nodata,
980 					      bcache_wq);
981 		} else {
982 			s->iop.bypass = check_should_bypass(dc, bio);
983 
984 			if (rw)
985 				cached_dev_write(dc, s);
986 			else
987 				cached_dev_read(dc, s);
988 		}
989 	} else {
990 		if ((bio_op(bio) == REQ_OP_DISCARD) &&
991 		    !blk_queue_discard(bdev_get_queue(dc->bdev)))
992 			bio_endio(bio);
993 		else
994 			generic_make_request(bio);
995 	}
996 
997 	return BLK_QC_T_NONE;
998 }
999 
1000 static int cached_dev_ioctl(struct bcache_device *d, fmode_t mode,
1001 			    unsigned int cmd, unsigned long arg)
1002 {
1003 	struct cached_dev *dc = container_of(d, struct cached_dev, disk);
1004 	return __blkdev_driver_ioctl(dc->bdev, mode, cmd, arg);
1005 }
1006 
1007 static int cached_dev_congested(void *data, int bits)
1008 {
1009 	struct bcache_device *d = data;
1010 	struct cached_dev *dc = container_of(d, struct cached_dev, disk);
1011 	struct request_queue *q = bdev_get_queue(dc->bdev);
1012 	int ret = 0;
1013 
1014 	if (bdi_congested(&q->backing_dev_info, bits))
1015 		return 1;
1016 
1017 	if (cached_dev_get(dc)) {
1018 		unsigned i;
1019 		struct cache *ca;
1020 
1021 		for_each_cache(ca, d->c, i) {
1022 			q = bdev_get_queue(ca->bdev);
1023 			ret |= bdi_congested(&q->backing_dev_info, bits);
1024 		}
1025 
1026 		cached_dev_put(dc);
1027 	}
1028 
1029 	return ret;
1030 }
1031 
1032 void bch_cached_dev_request_init(struct cached_dev *dc)
1033 {
1034 	struct gendisk *g = dc->disk.disk;
1035 
1036 	g->queue->make_request_fn		= cached_dev_make_request;
1037 	g->queue->backing_dev_info.congested_fn = cached_dev_congested;
1038 	dc->disk.cache_miss			= cached_dev_cache_miss;
1039 	dc->disk.ioctl				= cached_dev_ioctl;
1040 }
1041 
1042 /* Flash backed devices */
1043 
1044 static int flash_dev_cache_miss(struct btree *b, struct search *s,
1045 				struct bio *bio, unsigned sectors)
1046 {
1047 	unsigned bytes = min(sectors, bio_sectors(bio)) << 9;
1048 
1049 	swap(bio->bi_iter.bi_size, bytes);
1050 	zero_fill_bio(bio);
1051 	swap(bio->bi_iter.bi_size, bytes);
1052 
1053 	bio_advance(bio, bytes);
1054 
1055 	if (!bio->bi_iter.bi_size)
1056 		return MAP_DONE;
1057 
1058 	return MAP_CONTINUE;
1059 }
1060 
1061 static void flash_dev_nodata(struct closure *cl)
1062 {
1063 	struct search *s = container_of(cl, struct search, cl);
1064 
1065 	if (s->iop.flush_journal)
1066 		bch_journal_meta(s->iop.c, cl);
1067 
1068 	continue_at(cl, search_free, NULL);
1069 }
1070 
1071 static blk_qc_t flash_dev_make_request(struct request_queue *q,
1072 					     struct bio *bio)
1073 {
1074 	struct search *s;
1075 	struct closure *cl;
1076 	struct bcache_device *d = bio->bi_bdev->bd_disk->private_data;
1077 	int rw = bio_data_dir(bio);
1078 
1079 	generic_start_io_acct(rw, bio_sectors(bio), &d->disk->part0);
1080 
1081 	s = search_alloc(bio, d);
1082 	cl = &s->cl;
1083 	bio = &s->bio.bio;
1084 
1085 	trace_bcache_request_start(s->d, bio);
1086 
1087 	if (!bio->bi_iter.bi_size) {
1088 		/*
1089 		 * can't call bch_journal_meta from under
1090 		 * generic_make_request
1091 		 */
1092 		continue_at_nobarrier(&s->cl,
1093 				      flash_dev_nodata,
1094 				      bcache_wq);
1095 		return BLK_QC_T_NONE;
1096 	} else if (rw) {
1097 		bch_keybuf_check_overlapping(&s->iop.c->moving_gc_keys,
1098 					&KEY(d->id, bio->bi_iter.bi_sector, 0),
1099 					&KEY(d->id, bio_end_sector(bio), 0));
1100 
1101 		s->iop.bypass		= (bio_op(bio) == REQ_OP_DISCARD) != 0;
1102 		s->iop.writeback	= true;
1103 		s->iop.bio		= bio;
1104 
1105 		closure_call(&s->iop.cl, bch_data_insert, NULL, cl);
1106 	} else {
1107 		closure_call(&s->iop.cl, cache_lookup, NULL, cl);
1108 	}
1109 
1110 	continue_at(cl, search_free, NULL);
1111 	return BLK_QC_T_NONE;
1112 }
1113 
1114 static int flash_dev_ioctl(struct bcache_device *d, fmode_t mode,
1115 			   unsigned int cmd, unsigned long arg)
1116 {
1117 	return -ENOTTY;
1118 }
1119 
1120 static int flash_dev_congested(void *data, int bits)
1121 {
1122 	struct bcache_device *d = data;
1123 	struct request_queue *q;
1124 	struct cache *ca;
1125 	unsigned i;
1126 	int ret = 0;
1127 
1128 	for_each_cache(ca, d->c, i) {
1129 		q = bdev_get_queue(ca->bdev);
1130 		ret |= bdi_congested(&q->backing_dev_info, bits);
1131 	}
1132 
1133 	return ret;
1134 }
1135 
1136 void bch_flash_dev_request_init(struct bcache_device *d)
1137 {
1138 	struct gendisk *g = d->disk;
1139 
1140 	g->queue->make_request_fn		= flash_dev_make_request;
1141 	g->queue->backing_dev_info.congested_fn = flash_dev_congested;
1142 	d->cache_miss				= flash_dev_cache_miss;
1143 	d->ioctl				= flash_dev_ioctl;
1144 }
1145 
1146 void bch_request_exit(void)
1147 {
1148 	if (bch_search_cache)
1149 		kmem_cache_destroy(bch_search_cache);
1150 }
1151 
1152 int __init bch_request_init(void)
1153 {
1154 	bch_search_cache = KMEM_CACHE(search, 0);
1155 	if (!bch_search_cache)
1156 		return -ENOMEM;
1157 
1158 	return 0;
1159 }
1160