Loading...
1/*
2 * Functions related to segment and merge handling
3 */
4#include <linux/kernel.h>
5#include <linux/module.h>
6#include <linux/bio.h>
7#include <linux/blkdev.h>
8#include <linux/scatterlist.h>
9
10#include <trace/events/block.h>
11
12#include "blk.h"
13
14static struct bio *blk_bio_discard_split(struct request_queue *q,
15 struct bio *bio,
16 struct bio_set *bs,
17 unsigned *nsegs)
18{
19 unsigned int max_discard_sectors, granularity;
20 int alignment;
21 sector_t tmp;
22 unsigned split_sectors;
23
24 *nsegs = 1;
25
26 /* Zero-sector (unknown) and one-sector granularities are the same. */
27 granularity = max(q->limits.discard_granularity >> 9, 1U);
28
29 max_discard_sectors = min(q->limits.max_discard_sectors, UINT_MAX >> 9);
30 max_discard_sectors -= max_discard_sectors % granularity;
31
32 if (unlikely(!max_discard_sectors)) {
33 /* XXX: warn */
34 return NULL;
35 }
36
37 if (bio_sectors(bio) <= max_discard_sectors)
38 return NULL;
39
40 split_sectors = max_discard_sectors;
41
42 /*
43 * If the next starting sector would be misaligned, stop the discard at
44 * the previous aligned sector.
45 */
46 alignment = (q->limits.discard_alignment >> 9) % granularity;
47
48 tmp = bio->bi_iter.bi_sector + split_sectors - alignment;
49 tmp = sector_div(tmp, granularity);
50
51 if (split_sectors > tmp)
52 split_sectors -= tmp;
53
54 return bio_split(bio, split_sectors, GFP_NOIO, bs);
55}
56
57static struct bio *blk_bio_write_same_split(struct request_queue *q,
58 struct bio *bio,
59 struct bio_set *bs,
60 unsigned *nsegs)
61{
62 *nsegs = 1;
63
64 if (!q->limits.max_write_same_sectors)
65 return NULL;
66
67 if (bio_sectors(bio) <= q->limits.max_write_same_sectors)
68 return NULL;
69
70 return bio_split(bio, q->limits.max_write_same_sectors, GFP_NOIO, bs);
71}
72
73static inline unsigned get_max_io_size(struct request_queue *q,
74 struct bio *bio)
75{
76 unsigned sectors = blk_max_size_offset(q, bio->bi_iter.bi_sector);
77 unsigned mask = queue_logical_block_size(q) - 1;
78
79 /* aligned to logical block size */
80 sectors &= ~(mask >> 9);
81
82 return sectors;
83}
84
85static struct bio *blk_bio_segment_split(struct request_queue *q,
86 struct bio *bio,
87 struct bio_set *bs,
88 unsigned *segs)
89{
90 struct bio_vec bv, bvprv, *bvprvp = NULL;
91 struct bvec_iter iter;
92 unsigned seg_size = 0, nsegs = 0, sectors = 0;
93 unsigned front_seg_size = bio->bi_seg_front_size;
94 bool do_split = true;
95 struct bio *new = NULL;
96 const unsigned max_sectors = get_max_io_size(q, bio);
97
98 bio_for_each_segment(bv, bio, iter) {
99 /*
100 * If the queue doesn't support SG gaps and adding this
101 * offset would create a gap, disallow it.
102 */
103 if (bvprvp && bvec_gap_to_prev(q, bvprvp, bv.bv_offset))
104 goto split;
105
106 if (sectors + (bv.bv_len >> 9) > max_sectors) {
107 /*
108 * Consider this a new segment if we're splitting in
109 * the middle of this vector.
110 */
111 if (nsegs < queue_max_segments(q) &&
112 sectors < max_sectors) {
113 nsegs++;
114 sectors = max_sectors;
115 }
116 if (sectors)
117 goto split;
118 /* Make this single bvec as the 1st segment */
119 }
120
121 if (bvprvp && blk_queue_cluster(q)) {
122 if (seg_size + bv.bv_len > queue_max_segment_size(q))
123 goto new_segment;
124 if (!BIOVEC_PHYS_MERGEABLE(bvprvp, &bv))
125 goto new_segment;
126 if (!BIOVEC_SEG_BOUNDARY(q, bvprvp, &bv))
127 goto new_segment;
128
129 seg_size += bv.bv_len;
130 bvprv = bv;
131 bvprvp = &bvprv;
132 sectors += bv.bv_len >> 9;
133
134 if (nsegs == 1 && seg_size > front_seg_size)
135 front_seg_size = seg_size;
136 continue;
137 }
138new_segment:
139 if (nsegs == queue_max_segments(q))
140 goto split;
141
142 nsegs++;
143 bvprv = bv;
144 bvprvp = &bvprv;
145 seg_size = bv.bv_len;
146 sectors += bv.bv_len >> 9;
147
148 if (nsegs == 1 && seg_size > front_seg_size)
149 front_seg_size = seg_size;
150 }
151
152 do_split = false;
153split:
154 *segs = nsegs;
155
156 if (do_split) {
157 new = bio_split(bio, sectors, GFP_NOIO, bs);
158 if (new)
159 bio = new;
160 }
161
162 bio->bi_seg_front_size = front_seg_size;
163 if (seg_size > bio->bi_seg_back_size)
164 bio->bi_seg_back_size = seg_size;
165
166 return do_split ? new : NULL;
167}
168
169void blk_queue_split(struct request_queue *q, struct bio **bio,
170 struct bio_set *bs)
171{
172 struct bio *split, *res;
173 unsigned nsegs;
174
175 if ((*bio)->bi_rw & REQ_DISCARD)
176 split = blk_bio_discard_split(q, *bio, bs, &nsegs);
177 else if ((*bio)->bi_rw & REQ_WRITE_SAME)
178 split = blk_bio_write_same_split(q, *bio, bs, &nsegs);
179 else
180 split = blk_bio_segment_split(q, *bio, q->bio_split, &nsegs);
181
182 /* physical segments can be figured out during splitting */
183 res = split ? split : *bio;
184 res->bi_phys_segments = nsegs;
185 bio_set_flag(res, BIO_SEG_VALID);
186
187 if (split) {
188 /* there isn't chance to merge the splitted bio */
189 split->bi_rw |= REQ_NOMERGE;
190
191 bio_chain(split, *bio);
192 trace_block_split(q, split, (*bio)->bi_iter.bi_sector);
193 generic_make_request(*bio);
194 *bio = split;
195 }
196}
197EXPORT_SYMBOL(blk_queue_split);
198
199static unsigned int __blk_recalc_rq_segments(struct request_queue *q,
200 struct bio *bio,
201 bool no_sg_merge)
202{
203 struct bio_vec bv, bvprv = { NULL };
204 int cluster, prev = 0;
205 unsigned int seg_size, nr_phys_segs;
206 struct bio *fbio, *bbio;
207 struct bvec_iter iter;
208
209 if (!bio)
210 return 0;
211
212 /*
213 * This should probably be returning 0, but blk_add_request_payload()
214 * (Christoph!!!!)
215 */
216 if (bio->bi_rw & REQ_DISCARD)
217 return 1;
218
219 if (bio->bi_rw & REQ_WRITE_SAME)
220 return 1;
221
222 fbio = bio;
223 cluster = blk_queue_cluster(q);
224 seg_size = 0;
225 nr_phys_segs = 0;
226 for_each_bio(bio) {
227 bio_for_each_segment(bv, bio, iter) {
228 /*
229 * If SG merging is disabled, each bio vector is
230 * a segment
231 */
232 if (no_sg_merge)
233 goto new_segment;
234
235 if (prev && cluster) {
236 if (seg_size + bv.bv_len
237 > queue_max_segment_size(q))
238 goto new_segment;
239 if (!BIOVEC_PHYS_MERGEABLE(&bvprv, &bv))
240 goto new_segment;
241 if (!BIOVEC_SEG_BOUNDARY(q, &bvprv, &bv))
242 goto new_segment;
243
244 seg_size += bv.bv_len;
245 bvprv = bv;
246 continue;
247 }
248new_segment:
249 if (nr_phys_segs == 1 && seg_size >
250 fbio->bi_seg_front_size)
251 fbio->bi_seg_front_size = seg_size;
252
253 nr_phys_segs++;
254 bvprv = bv;
255 prev = 1;
256 seg_size = bv.bv_len;
257 }
258 bbio = bio;
259 }
260
261 if (nr_phys_segs == 1 && seg_size > fbio->bi_seg_front_size)
262 fbio->bi_seg_front_size = seg_size;
263 if (seg_size > bbio->bi_seg_back_size)
264 bbio->bi_seg_back_size = seg_size;
265
266 return nr_phys_segs;
267}
268
269void blk_recalc_rq_segments(struct request *rq)
270{
271 bool no_sg_merge = !!test_bit(QUEUE_FLAG_NO_SG_MERGE,
272 &rq->q->queue_flags);
273
274 rq->nr_phys_segments = __blk_recalc_rq_segments(rq->q, rq->bio,
275 no_sg_merge);
276}
277
278void blk_recount_segments(struct request_queue *q, struct bio *bio)
279{
280 unsigned short seg_cnt;
281
282 /* estimate segment number by bi_vcnt for non-cloned bio */
283 if (bio_flagged(bio, BIO_CLONED))
284 seg_cnt = bio_segments(bio);
285 else
286 seg_cnt = bio->bi_vcnt;
287
288 if (test_bit(QUEUE_FLAG_NO_SG_MERGE, &q->queue_flags) &&
289 (seg_cnt < queue_max_segments(q)))
290 bio->bi_phys_segments = seg_cnt;
291 else {
292 struct bio *nxt = bio->bi_next;
293
294 bio->bi_next = NULL;
295 bio->bi_phys_segments = __blk_recalc_rq_segments(q, bio, false);
296 bio->bi_next = nxt;
297 }
298
299 bio_set_flag(bio, BIO_SEG_VALID);
300}
301EXPORT_SYMBOL(blk_recount_segments);
302
303static int blk_phys_contig_segment(struct request_queue *q, struct bio *bio,
304 struct bio *nxt)
305{
306 struct bio_vec end_bv = { NULL }, nxt_bv;
307
308 if (!blk_queue_cluster(q))
309 return 0;
310
311 if (bio->bi_seg_back_size + nxt->bi_seg_front_size >
312 queue_max_segment_size(q))
313 return 0;
314
315 if (!bio_has_data(bio))
316 return 1;
317
318 bio_get_last_bvec(bio, &end_bv);
319 bio_get_first_bvec(nxt, &nxt_bv);
320
321 if (!BIOVEC_PHYS_MERGEABLE(&end_bv, &nxt_bv))
322 return 0;
323
324 /*
325 * bio and nxt are contiguous in memory; check if the queue allows
326 * these two to be merged into one
327 */
328 if (BIOVEC_SEG_BOUNDARY(q, &end_bv, &nxt_bv))
329 return 1;
330
331 return 0;
332}
333
334static inline void
335__blk_segment_map_sg(struct request_queue *q, struct bio_vec *bvec,
336 struct scatterlist *sglist, struct bio_vec *bvprv,
337 struct scatterlist **sg, int *nsegs, int *cluster)
338{
339
340 int nbytes = bvec->bv_len;
341
342 if (*sg && *cluster) {
343 if ((*sg)->length + nbytes > queue_max_segment_size(q))
344 goto new_segment;
345
346 if (!BIOVEC_PHYS_MERGEABLE(bvprv, bvec))
347 goto new_segment;
348 if (!BIOVEC_SEG_BOUNDARY(q, bvprv, bvec))
349 goto new_segment;
350
351 (*sg)->length += nbytes;
352 } else {
353new_segment:
354 if (!*sg)
355 *sg = sglist;
356 else {
357 /*
358 * If the driver previously mapped a shorter
359 * list, we could see a termination bit
360 * prematurely unless it fully inits the sg
361 * table on each mapping. We KNOW that there
362 * must be more entries here or the driver
363 * would be buggy, so force clear the
364 * termination bit to avoid doing a full
365 * sg_init_table() in drivers for each command.
366 */
367 sg_unmark_end(*sg);
368 *sg = sg_next(*sg);
369 }
370
371 sg_set_page(*sg, bvec->bv_page, nbytes, bvec->bv_offset);
372 (*nsegs)++;
373 }
374 *bvprv = *bvec;
375}
376
377static int __blk_bios_map_sg(struct request_queue *q, struct bio *bio,
378 struct scatterlist *sglist,
379 struct scatterlist **sg)
380{
381 struct bio_vec bvec, bvprv = { NULL };
382 struct bvec_iter iter;
383 int nsegs, cluster;
384
385 nsegs = 0;
386 cluster = blk_queue_cluster(q);
387
388 if (bio->bi_rw & REQ_DISCARD) {
389 /*
390 * This is a hack - drivers should be neither modifying the
391 * biovec, nor relying on bi_vcnt - but because of
392 * blk_add_request_payload(), a discard bio may or may not have
393 * a payload we need to set up here (thank you Christoph) and
394 * bi_vcnt is really the only way of telling if we need to.
395 */
396
397 if (bio->bi_vcnt)
398 goto single_segment;
399
400 return 0;
401 }
402
403 if (bio->bi_rw & REQ_WRITE_SAME) {
404single_segment:
405 *sg = sglist;
406 bvec = bio_iovec(bio);
407 sg_set_page(*sg, bvec.bv_page, bvec.bv_len, bvec.bv_offset);
408 return 1;
409 }
410
411 for_each_bio(bio)
412 bio_for_each_segment(bvec, bio, iter)
413 __blk_segment_map_sg(q, &bvec, sglist, &bvprv, sg,
414 &nsegs, &cluster);
415
416 return nsegs;
417}
418
419/*
420 * map a request to scatterlist, return number of sg entries setup. Caller
421 * must make sure sg can hold rq->nr_phys_segments entries
422 */
423int blk_rq_map_sg(struct request_queue *q, struct request *rq,
424 struct scatterlist *sglist)
425{
426 struct scatterlist *sg = NULL;
427 int nsegs = 0;
428
429 if (rq->bio)
430 nsegs = __blk_bios_map_sg(q, rq->bio, sglist, &sg);
431
432 if (unlikely(rq->cmd_flags & REQ_COPY_USER) &&
433 (blk_rq_bytes(rq) & q->dma_pad_mask)) {
434 unsigned int pad_len =
435 (q->dma_pad_mask & ~blk_rq_bytes(rq)) + 1;
436
437 sg->length += pad_len;
438 rq->extra_len += pad_len;
439 }
440
441 if (q->dma_drain_size && q->dma_drain_needed(rq)) {
442 if (rq->cmd_flags & REQ_WRITE)
443 memset(q->dma_drain_buffer, 0, q->dma_drain_size);
444
445 sg_unmark_end(sg);
446 sg = sg_next(sg);
447 sg_set_page(sg, virt_to_page(q->dma_drain_buffer),
448 q->dma_drain_size,
449 ((unsigned long)q->dma_drain_buffer) &
450 (PAGE_SIZE - 1));
451 nsegs++;
452 rq->extra_len += q->dma_drain_size;
453 }
454
455 if (sg)
456 sg_mark_end(sg);
457
458 /*
459 * Something must have been wrong if the figured number of
460 * segment is bigger than number of req's physical segments
461 */
462 WARN_ON(nsegs > rq->nr_phys_segments);
463
464 return nsegs;
465}
466EXPORT_SYMBOL(blk_rq_map_sg);
467
468static inline int ll_new_hw_segment(struct request_queue *q,
469 struct request *req,
470 struct bio *bio)
471{
472 int nr_phys_segs = bio_phys_segments(q, bio);
473
474 if (req->nr_phys_segments + nr_phys_segs > queue_max_segments(q))
475 goto no_merge;
476
477 if (blk_integrity_merge_bio(q, req, bio) == false)
478 goto no_merge;
479
480 /*
481 * This will form the start of a new hw segment. Bump both
482 * counters.
483 */
484 req->nr_phys_segments += nr_phys_segs;
485 return 1;
486
487no_merge:
488 req->cmd_flags |= REQ_NOMERGE;
489 if (req == q->last_merge)
490 q->last_merge = NULL;
491 return 0;
492}
493
494int ll_back_merge_fn(struct request_queue *q, struct request *req,
495 struct bio *bio)
496{
497 if (req_gap_back_merge(req, bio))
498 return 0;
499 if (blk_integrity_rq(req) &&
500 integrity_req_gap_back_merge(req, bio))
501 return 0;
502 if (blk_rq_sectors(req) + bio_sectors(bio) >
503 blk_rq_get_max_sectors(req)) {
504 req->cmd_flags |= REQ_NOMERGE;
505 if (req == q->last_merge)
506 q->last_merge = NULL;
507 return 0;
508 }
509 if (!bio_flagged(req->biotail, BIO_SEG_VALID))
510 blk_recount_segments(q, req->biotail);
511 if (!bio_flagged(bio, BIO_SEG_VALID))
512 blk_recount_segments(q, bio);
513
514 return ll_new_hw_segment(q, req, bio);
515}
516
517int ll_front_merge_fn(struct request_queue *q, struct request *req,
518 struct bio *bio)
519{
520
521 if (req_gap_front_merge(req, bio))
522 return 0;
523 if (blk_integrity_rq(req) &&
524 integrity_req_gap_front_merge(req, bio))
525 return 0;
526 if (blk_rq_sectors(req) + bio_sectors(bio) >
527 blk_rq_get_max_sectors(req)) {
528 req->cmd_flags |= REQ_NOMERGE;
529 if (req == q->last_merge)
530 q->last_merge = NULL;
531 return 0;
532 }
533 if (!bio_flagged(bio, BIO_SEG_VALID))
534 blk_recount_segments(q, bio);
535 if (!bio_flagged(req->bio, BIO_SEG_VALID))
536 blk_recount_segments(q, req->bio);
537
538 return ll_new_hw_segment(q, req, bio);
539}
540
541/*
542 * blk-mq uses req->special to carry normal driver per-request payload, it
543 * does not indicate a prepared command that we cannot merge with.
544 */
545static bool req_no_special_merge(struct request *req)
546{
547 struct request_queue *q = req->q;
548
549 return !q->mq_ops && req->special;
550}
551
552static int ll_merge_requests_fn(struct request_queue *q, struct request *req,
553 struct request *next)
554{
555 int total_phys_segments;
556 unsigned int seg_size =
557 req->biotail->bi_seg_back_size + next->bio->bi_seg_front_size;
558
559 /*
560 * First check if the either of the requests are re-queued
561 * requests. Can't merge them if they are.
562 */
563 if (req_no_special_merge(req) || req_no_special_merge(next))
564 return 0;
565
566 if (req_gap_back_merge(req, next->bio))
567 return 0;
568
569 /*
570 * Will it become too large?
571 */
572 if ((blk_rq_sectors(req) + blk_rq_sectors(next)) >
573 blk_rq_get_max_sectors(req))
574 return 0;
575
576 total_phys_segments = req->nr_phys_segments + next->nr_phys_segments;
577 if (blk_phys_contig_segment(q, req->biotail, next->bio)) {
578 if (req->nr_phys_segments == 1)
579 req->bio->bi_seg_front_size = seg_size;
580 if (next->nr_phys_segments == 1)
581 next->biotail->bi_seg_back_size = seg_size;
582 total_phys_segments--;
583 }
584
585 if (total_phys_segments > queue_max_segments(q))
586 return 0;
587
588 if (blk_integrity_merge_rq(q, req, next) == false)
589 return 0;
590
591 /* Merge is OK... */
592 req->nr_phys_segments = total_phys_segments;
593 return 1;
594}
595
596/**
597 * blk_rq_set_mixed_merge - mark a request as mixed merge
598 * @rq: request to mark as mixed merge
599 *
600 * Description:
601 * @rq is about to be mixed merged. Make sure the attributes
602 * which can be mixed are set in each bio and mark @rq as mixed
603 * merged.
604 */
605void blk_rq_set_mixed_merge(struct request *rq)
606{
607 unsigned int ff = rq->cmd_flags & REQ_FAILFAST_MASK;
608 struct bio *bio;
609
610 if (rq->cmd_flags & REQ_MIXED_MERGE)
611 return;
612
613 /*
614 * @rq will no longer represent mixable attributes for all the
615 * contained bios. It will just track those of the first one.
616 * Distributes the attributs to each bio.
617 */
618 for (bio = rq->bio; bio; bio = bio->bi_next) {
619 WARN_ON_ONCE((bio->bi_rw & REQ_FAILFAST_MASK) &&
620 (bio->bi_rw & REQ_FAILFAST_MASK) != ff);
621 bio->bi_rw |= ff;
622 }
623 rq->cmd_flags |= REQ_MIXED_MERGE;
624}
625
626static void blk_account_io_merge(struct request *req)
627{
628 if (blk_do_io_stat(req)) {
629 struct hd_struct *part;
630 int cpu;
631
632 cpu = part_stat_lock();
633 part = req->part;
634
635 part_round_stats(cpu, part);
636 part_dec_in_flight(part, rq_data_dir(req));
637
638 hd_struct_put(part);
639 part_stat_unlock();
640 }
641}
642
643/*
644 * Has to be called with the request spinlock acquired
645 */
646static int attempt_merge(struct request_queue *q, struct request *req,
647 struct request *next)
648{
649 if (!rq_mergeable(req) || !rq_mergeable(next))
650 return 0;
651
652 if (!blk_check_merge_flags(req->cmd_flags, next->cmd_flags))
653 return 0;
654
655 /*
656 * not contiguous
657 */
658 if (blk_rq_pos(req) + blk_rq_sectors(req) != blk_rq_pos(next))
659 return 0;
660
661 if (rq_data_dir(req) != rq_data_dir(next)
662 || req->rq_disk != next->rq_disk
663 || req_no_special_merge(next))
664 return 0;
665
666 if (req->cmd_flags & REQ_WRITE_SAME &&
667 !blk_write_same_mergeable(req->bio, next->bio))
668 return 0;
669
670 /*
671 * If we are allowed to merge, then append bio list
672 * from next to rq and release next. merge_requests_fn
673 * will have updated segment counts, update sector
674 * counts here.
675 */
676 if (!ll_merge_requests_fn(q, req, next))
677 return 0;
678
679 /*
680 * If failfast settings disagree or any of the two is already
681 * a mixed merge, mark both as mixed before proceeding. This
682 * makes sure that all involved bios have mixable attributes
683 * set properly.
684 */
685 if ((req->cmd_flags | next->cmd_flags) & REQ_MIXED_MERGE ||
686 (req->cmd_flags & REQ_FAILFAST_MASK) !=
687 (next->cmd_flags & REQ_FAILFAST_MASK)) {
688 blk_rq_set_mixed_merge(req);
689 blk_rq_set_mixed_merge(next);
690 }
691
692 /*
693 * At this point we have either done a back merge
694 * or front merge. We need the smaller start_time of
695 * the merged requests to be the current request
696 * for accounting purposes.
697 */
698 if (time_after(req->start_time, next->start_time))
699 req->start_time = next->start_time;
700
701 req->biotail->bi_next = next->bio;
702 req->biotail = next->biotail;
703
704 req->__data_len += blk_rq_bytes(next);
705
706 elv_merge_requests(q, req, next);
707
708 /*
709 * 'next' is going away, so update stats accordingly
710 */
711 blk_account_io_merge(next);
712
713 req->ioprio = ioprio_best(req->ioprio, next->ioprio);
714 if (blk_rq_cpu_valid(next))
715 req->cpu = next->cpu;
716
717 /* owner-ship of bio passed from next to req */
718 next->bio = NULL;
719 __blk_put_request(q, next);
720 return 1;
721}
722
723int attempt_back_merge(struct request_queue *q, struct request *rq)
724{
725 struct request *next = elv_latter_request(q, rq);
726
727 if (next)
728 return attempt_merge(q, rq, next);
729
730 return 0;
731}
732
733int attempt_front_merge(struct request_queue *q, struct request *rq)
734{
735 struct request *prev = elv_former_request(q, rq);
736
737 if (prev)
738 return attempt_merge(q, prev, rq);
739
740 return 0;
741}
742
743int blk_attempt_req_merge(struct request_queue *q, struct request *rq,
744 struct request *next)
745{
746 return attempt_merge(q, rq, next);
747}
748
749bool blk_rq_merge_ok(struct request *rq, struct bio *bio)
750{
751 if (!rq_mergeable(rq) || !bio_mergeable(bio))
752 return false;
753
754 if (!blk_check_merge_flags(rq->cmd_flags, bio->bi_rw))
755 return false;
756
757 /* different data direction or already started, don't merge */
758 if (bio_data_dir(bio) != rq_data_dir(rq))
759 return false;
760
761 /* must be same device and not a special request */
762 if (rq->rq_disk != bio->bi_bdev->bd_disk || req_no_special_merge(rq))
763 return false;
764
765 /* only merge integrity protected bio into ditto rq */
766 if (blk_integrity_merge_bio(rq->q, rq, bio) == false)
767 return false;
768
769 /* must be using the same buffer */
770 if (rq->cmd_flags & REQ_WRITE_SAME &&
771 !blk_write_same_mergeable(rq->bio, bio))
772 return false;
773
774 return true;
775}
776
777int blk_try_merge(struct request *rq, struct bio *bio)
778{
779 if (blk_rq_pos(rq) + blk_rq_sectors(rq) == bio->bi_iter.bi_sector)
780 return ELEVATOR_BACK_MERGE;
781 else if (blk_rq_pos(rq) - bio_sectors(bio) == bio->bi_iter.bi_sector)
782 return ELEVATOR_FRONT_MERGE;
783 return ELEVATOR_NO_MERGE;
784}
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Functions related to segment and merge handling
4 */
5#include <linux/kernel.h>
6#include <linux/module.h>
7#include <linux/bio.h>
8#include <linux/blkdev.h>
9#include <linux/scatterlist.h>
10
11#include <trace/events/block.h>
12
13#include "blk.h"
14
15static inline bool bio_will_gap(struct request_queue *q,
16 struct request *prev_rq, struct bio *prev, struct bio *next)
17{
18 struct bio_vec pb, nb;
19
20 if (!bio_has_data(prev) || !queue_virt_boundary(q))
21 return false;
22
23 /*
24 * Don't merge if the 1st bio starts with non-zero offset, otherwise it
25 * is quite difficult to respect the sg gap limit. We work hard to
26 * merge a huge number of small single bios in case of mkfs.
27 */
28 if (prev_rq)
29 bio_get_first_bvec(prev_rq->bio, &pb);
30 else
31 bio_get_first_bvec(prev, &pb);
32 if (pb.bv_offset & queue_virt_boundary(q))
33 return true;
34
35 /*
36 * We don't need to worry about the situation that the merged segment
37 * ends in unaligned virt boundary:
38 *
39 * - if 'pb' ends aligned, the merged segment ends aligned
40 * - if 'pb' ends unaligned, the next bio must include
41 * one single bvec of 'nb', otherwise the 'nb' can't
42 * merge with 'pb'
43 */
44 bio_get_last_bvec(prev, &pb);
45 bio_get_first_bvec(next, &nb);
46 if (biovec_phys_mergeable(q, &pb, &nb))
47 return false;
48 return __bvec_gap_to_prev(q, &pb, nb.bv_offset);
49}
50
51static inline bool req_gap_back_merge(struct request *req, struct bio *bio)
52{
53 return bio_will_gap(req->q, req, req->biotail, bio);
54}
55
56static inline bool req_gap_front_merge(struct request *req, struct bio *bio)
57{
58 return bio_will_gap(req->q, NULL, bio, req->bio);
59}
60
61static struct bio *blk_bio_discard_split(struct request_queue *q,
62 struct bio *bio,
63 struct bio_set *bs,
64 unsigned *nsegs)
65{
66 unsigned int max_discard_sectors, granularity;
67 int alignment;
68 sector_t tmp;
69 unsigned split_sectors;
70
71 *nsegs = 1;
72
73 /* Zero-sector (unknown) and one-sector granularities are the same. */
74 granularity = max(q->limits.discard_granularity >> 9, 1U);
75
76 max_discard_sectors = min(q->limits.max_discard_sectors,
77 bio_allowed_max_sectors(q));
78 max_discard_sectors -= max_discard_sectors % granularity;
79
80 if (unlikely(!max_discard_sectors)) {
81 /* XXX: warn */
82 return NULL;
83 }
84
85 if (bio_sectors(bio) <= max_discard_sectors)
86 return NULL;
87
88 split_sectors = max_discard_sectors;
89
90 /*
91 * If the next starting sector would be misaligned, stop the discard at
92 * the previous aligned sector.
93 */
94 alignment = (q->limits.discard_alignment >> 9) % granularity;
95
96 tmp = bio->bi_iter.bi_sector + split_sectors - alignment;
97 tmp = sector_div(tmp, granularity);
98
99 if (split_sectors > tmp)
100 split_sectors -= tmp;
101
102 return bio_split(bio, split_sectors, GFP_NOIO, bs);
103}
104
105static struct bio *blk_bio_write_zeroes_split(struct request_queue *q,
106 struct bio *bio, struct bio_set *bs, unsigned *nsegs)
107{
108 *nsegs = 0;
109
110 if (!q->limits.max_write_zeroes_sectors)
111 return NULL;
112
113 if (bio_sectors(bio) <= q->limits.max_write_zeroes_sectors)
114 return NULL;
115
116 return bio_split(bio, q->limits.max_write_zeroes_sectors, GFP_NOIO, bs);
117}
118
119static struct bio *blk_bio_write_same_split(struct request_queue *q,
120 struct bio *bio,
121 struct bio_set *bs,
122 unsigned *nsegs)
123{
124 *nsegs = 1;
125
126 if (!q->limits.max_write_same_sectors)
127 return NULL;
128
129 if (bio_sectors(bio) <= q->limits.max_write_same_sectors)
130 return NULL;
131
132 return bio_split(bio, q->limits.max_write_same_sectors, GFP_NOIO, bs);
133}
134
135/*
136 * Return the maximum number of sectors from the start of a bio that may be
137 * submitted as a single request to a block device. If enough sectors remain,
138 * align the end to the physical block size. Otherwise align the end to the
139 * logical block size. This approach minimizes the number of non-aligned
140 * requests that are submitted to a block device if the start of a bio is not
141 * aligned to a physical block boundary.
142 */
143static inline unsigned get_max_io_size(struct request_queue *q,
144 struct bio *bio)
145{
146 unsigned sectors = blk_max_size_offset(q, bio->bi_iter.bi_sector);
147 unsigned max_sectors = sectors;
148 unsigned pbs = queue_physical_block_size(q) >> SECTOR_SHIFT;
149 unsigned lbs = queue_logical_block_size(q) >> SECTOR_SHIFT;
150 unsigned start_offset = bio->bi_iter.bi_sector & (pbs - 1);
151
152 max_sectors += start_offset;
153 max_sectors &= ~(pbs - 1);
154 if (max_sectors > start_offset)
155 return max_sectors - start_offset;
156
157 return sectors & ~(lbs - 1);
158}
159
160static inline unsigned get_max_segment_size(const struct request_queue *q,
161 struct page *start_page,
162 unsigned long offset)
163{
164 unsigned long mask = queue_segment_boundary(q);
165
166 offset = mask & (page_to_phys(start_page) + offset);
167
168 /*
169 * overflow may be triggered in case of zero page physical address
170 * on 32bit arch, use queue's max segment size when that happens.
171 */
172 return min_not_zero(mask - offset + 1,
173 (unsigned long)queue_max_segment_size(q));
174}
175
176/**
177 * bvec_split_segs - verify whether or not a bvec should be split in the middle
178 * @q: [in] request queue associated with the bio associated with @bv
179 * @bv: [in] bvec to examine
180 * @nsegs: [in,out] Number of segments in the bio being built. Incremented
181 * by the number of segments from @bv that may be appended to that
182 * bio without exceeding @max_segs
183 * @sectors: [in,out] Number of sectors in the bio being built. Incremented
184 * by the number of sectors from @bv that may be appended to that
185 * bio without exceeding @max_sectors
186 * @max_segs: [in] upper bound for *@nsegs
187 * @max_sectors: [in] upper bound for *@sectors
188 *
189 * When splitting a bio, it can happen that a bvec is encountered that is too
190 * big to fit in a single segment and hence that it has to be split in the
191 * middle. This function verifies whether or not that should happen. The value
192 * %true is returned if and only if appending the entire @bv to a bio with
193 * *@nsegs segments and *@sectors sectors would make that bio unacceptable for
194 * the block driver.
195 */
196static bool bvec_split_segs(const struct request_queue *q,
197 const struct bio_vec *bv, unsigned *nsegs,
198 unsigned *sectors, unsigned max_segs,
199 unsigned max_sectors)
200{
201 unsigned max_len = (min(max_sectors, UINT_MAX >> 9) - *sectors) << 9;
202 unsigned len = min(bv->bv_len, max_len);
203 unsigned total_len = 0;
204 unsigned seg_size = 0;
205
206 while (len && *nsegs < max_segs) {
207 seg_size = get_max_segment_size(q, bv->bv_page,
208 bv->bv_offset + total_len);
209 seg_size = min(seg_size, len);
210
211 (*nsegs)++;
212 total_len += seg_size;
213 len -= seg_size;
214
215 if ((bv->bv_offset + total_len) & queue_virt_boundary(q))
216 break;
217 }
218
219 *sectors += total_len >> 9;
220
221 /* tell the caller to split the bvec if it is too big to fit */
222 return len > 0 || bv->bv_len > max_len;
223}
224
225/**
226 * blk_bio_segment_split - split a bio in two bios
227 * @q: [in] request queue pointer
228 * @bio: [in] bio to be split
229 * @bs: [in] bio set to allocate the clone from
230 * @segs: [out] number of segments in the bio with the first half of the sectors
231 *
232 * Clone @bio, update the bi_iter of the clone to represent the first sectors
233 * of @bio and update @bio->bi_iter to represent the remaining sectors. The
234 * following is guaranteed for the cloned bio:
235 * - That it has at most get_max_io_size(@q, @bio) sectors.
236 * - That it has at most queue_max_segments(@q) segments.
237 *
238 * Except for discard requests the cloned bio will point at the bi_io_vec of
239 * the original bio. It is the responsibility of the caller to ensure that the
240 * original bio is not freed before the cloned bio. The caller is also
241 * responsible for ensuring that @bs is only destroyed after processing of the
242 * split bio has finished.
243 */
244static struct bio *blk_bio_segment_split(struct request_queue *q,
245 struct bio *bio,
246 struct bio_set *bs,
247 unsigned *segs)
248{
249 struct bio_vec bv, bvprv, *bvprvp = NULL;
250 struct bvec_iter iter;
251 unsigned nsegs = 0, sectors = 0;
252 const unsigned max_sectors = get_max_io_size(q, bio);
253 const unsigned max_segs = queue_max_segments(q);
254
255 bio_for_each_bvec(bv, bio, iter) {
256 /*
257 * If the queue doesn't support SG gaps and adding this
258 * offset would create a gap, disallow it.
259 */
260 if (bvprvp && bvec_gap_to_prev(q, bvprvp, bv.bv_offset))
261 goto split;
262
263 if (nsegs < max_segs &&
264 sectors + (bv.bv_len >> 9) <= max_sectors &&
265 bv.bv_offset + bv.bv_len <= PAGE_SIZE) {
266 nsegs++;
267 sectors += bv.bv_len >> 9;
268 } else if (bvec_split_segs(q, &bv, &nsegs, §ors, max_segs,
269 max_sectors)) {
270 goto split;
271 }
272
273 bvprv = bv;
274 bvprvp = &bvprv;
275 }
276
277 *segs = nsegs;
278 return NULL;
279split:
280 *segs = nsegs;
281 return bio_split(bio, sectors, GFP_NOIO, bs);
282}
283
284/**
285 * __blk_queue_split - split a bio and submit the second half
286 * @bio: [in, out] bio to be split
287 * @nr_segs: [out] number of segments in the first bio
288 *
289 * Split a bio into two bios, chain the two bios, submit the second half and
290 * store a pointer to the first half in *@bio. If the second bio is still too
291 * big it will be split by a recursive call to this function. Since this
292 * function may allocate a new bio from @bio->bi_disk->queue->bio_split, it is
293 * the responsibility of the caller to ensure that
294 * @bio->bi_disk->queue->bio_split is only released after processing of the
295 * split bio has finished.
296 */
297void __blk_queue_split(struct bio **bio, unsigned int *nr_segs)
298{
299 struct request_queue *q = (*bio)->bi_disk->queue;
300 struct bio *split = NULL;
301
302 switch (bio_op(*bio)) {
303 case REQ_OP_DISCARD:
304 case REQ_OP_SECURE_ERASE:
305 split = blk_bio_discard_split(q, *bio, &q->bio_split, nr_segs);
306 break;
307 case REQ_OP_WRITE_ZEROES:
308 split = blk_bio_write_zeroes_split(q, *bio, &q->bio_split,
309 nr_segs);
310 break;
311 case REQ_OP_WRITE_SAME:
312 split = blk_bio_write_same_split(q, *bio, &q->bio_split,
313 nr_segs);
314 break;
315 default:
316 /*
317 * All drivers must accept single-segments bios that are <=
318 * PAGE_SIZE. This is a quick and dirty check that relies on
319 * the fact that bi_io_vec[0] is always valid if a bio has data.
320 * The check might lead to occasional false negatives when bios
321 * are cloned, but compared to the performance impact of cloned
322 * bios themselves the loop below doesn't matter anyway.
323 */
324 if (!q->limits.chunk_sectors &&
325 (*bio)->bi_vcnt == 1 &&
326 ((*bio)->bi_io_vec[0].bv_len +
327 (*bio)->bi_io_vec[0].bv_offset) <= PAGE_SIZE) {
328 *nr_segs = 1;
329 break;
330 }
331 split = blk_bio_segment_split(q, *bio, &q->bio_split, nr_segs);
332 break;
333 }
334
335 if (split) {
336 /* there isn't chance to merge the splitted bio */
337 split->bi_opf |= REQ_NOMERGE;
338
339 bio_chain(split, *bio);
340 trace_block_split(q, split, (*bio)->bi_iter.bi_sector);
341 submit_bio_noacct(*bio);
342 *bio = split;
343 }
344}
345
346/**
347 * blk_queue_split - split a bio and submit the second half
348 * @bio: [in, out] bio to be split
349 *
350 * Split a bio into two bios, chains the two bios, submit the second half and
351 * store a pointer to the first half in *@bio. Since this function may allocate
352 * a new bio from @bio->bi_disk->queue->bio_split, it is the responsibility of
353 * the caller to ensure that @bio->bi_disk->queue->bio_split is only released
354 * after processing of the split bio has finished.
355 */
356void blk_queue_split(struct bio **bio)
357{
358 unsigned int nr_segs;
359
360 __blk_queue_split(bio, &nr_segs);
361}
362EXPORT_SYMBOL(blk_queue_split);
363
364unsigned int blk_recalc_rq_segments(struct request *rq)
365{
366 unsigned int nr_phys_segs = 0;
367 unsigned int nr_sectors = 0;
368 struct req_iterator iter;
369 struct bio_vec bv;
370
371 if (!rq->bio)
372 return 0;
373
374 switch (bio_op(rq->bio)) {
375 case REQ_OP_DISCARD:
376 case REQ_OP_SECURE_ERASE:
377 case REQ_OP_WRITE_ZEROES:
378 return 0;
379 case REQ_OP_WRITE_SAME:
380 return 1;
381 }
382
383 rq_for_each_bvec(bv, rq, iter)
384 bvec_split_segs(rq->q, &bv, &nr_phys_segs, &nr_sectors,
385 UINT_MAX, UINT_MAX);
386 return nr_phys_segs;
387}
388
389static inline struct scatterlist *blk_next_sg(struct scatterlist **sg,
390 struct scatterlist *sglist)
391{
392 if (!*sg)
393 return sglist;
394
395 /*
396 * If the driver previously mapped a shorter list, we could see a
397 * termination bit prematurely unless it fully inits the sg table
398 * on each mapping. We KNOW that there must be more entries here
399 * or the driver would be buggy, so force clear the termination bit
400 * to avoid doing a full sg_init_table() in drivers for each command.
401 */
402 sg_unmark_end(*sg);
403 return sg_next(*sg);
404}
405
406static unsigned blk_bvec_map_sg(struct request_queue *q,
407 struct bio_vec *bvec, struct scatterlist *sglist,
408 struct scatterlist **sg)
409{
410 unsigned nbytes = bvec->bv_len;
411 unsigned nsegs = 0, total = 0;
412
413 while (nbytes > 0) {
414 unsigned offset = bvec->bv_offset + total;
415 unsigned len = min(get_max_segment_size(q, bvec->bv_page,
416 offset), nbytes);
417 struct page *page = bvec->bv_page;
418
419 /*
420 * Unfortunately a fair number of drivers barf on scatterlists
421 * that have an offset larger than PAGE_SIZE, despite other
422 * subsystems dealing with that invariant just fine. For now
423 * stick to the legacy format where we never present those from
424 * the block layer, but the code below should be removed once
425 * these offenders (mostly MMC/SD drivers) are fixed.
426 */
427 page += (offset >> PAGE_SHIFT);
428 offset &= ~PAGE_MASK;
429
430 *sg = blk_next_sg(sg, sglist);
431 sg_set_page(*sg, page, len, offset);
432
433 total += len;
434 nbytes -= len;
435 nsegs++;
436 }
437
438 return nsegs;
439}
440
441static inline int __blk_bvec_map_sg(struct bio_vec bv,
442 struct scatterlist *sglist, struct scatterlist **sg)
443{
444 *sg = blk_next_sg(sg, sglist);
445 sg_set_page(*sg, bv.bv_page, bv.bv_len, bv.bv_offset);
446 return 1;
447}
448
449/* only try to merge bvecs into one sg if they are from two bios */
450static inline bool
451__blk_segment_map_sg_merge(struct request_queue *q, struct bio_vec *bvec,
452 struct bio_vec *bvprv, struct scatterlist **sg)
453{
454
455 int nbytes = bvec->bv_len;
456
457 if (!*sg)
458 return false;
459
460 if ((*sg)->length + nbytes > queue_max_segment_size(q))
461 return false;
462
463 if (!biovec_phys_mergeable(q, bvprv, bvec))
464 return false;
465
466 (*sg)->length += nbytes;
467
468 return true;
469}
470
471static int __blk_bios_map_sg(struct request_queue *q, struct bio *bio,
472 struct scatterlist *sglist,
473 struct scatterlist **sg)
474{
475 struct bio_vec bvec, bvprv = { NULL };
476 struct bvec_iter iter;
477 int nsegs = 0;
478 bool new_bio = false;
479
480 for_each_bio(bio) {
481 bio_for_each_bvec(bvec, bio, iter) {
482 /*
483 * Only try to merge bvecs from two bios given we
484 * have done bio internal merge when adding pages
485 * to bio
486 */
487 if (new_bio &&
488 __blk_segment_map_sg_merge(q, &bvec, &bvprv, sg))
489 goto next_bvec;
490
491 if (bvec.bv_offset + bvec.bv_len <= PAGE_SIZE)
492 nsegs += __blk_bvec_map_sg(bvec, sglist, sg);
493 else
494 nsegs += blk_bvec_map_sg(q, &bvec, sglist, sg);
495 next_bvec:
496 new_bio = false;
497 }
498 if (likely(bio->bi_iter.bi_size)) {
499 bvprv = bvec;
500 new_bio = true;
501 }
502 }
503
504 return nsegs;
505}
506
507/*
508 * map a request to scatterlist, return number of sg entries setup. Caller
509 * must make sure sg can hold rq->nr_phys_segments entries
510 */
511int __blk_rq_map_sg(struct request_queue *q, struct request *rq,
512 struct scatterlist *sglist, struct scatterlist **last_sg)
513{
514 int nsegs = 0;
515
516 if (rq->rq_flags & RQF_SPECIAL_PAYLOAD)
517 nsegs = __blk_bvec_map_sg(rq->special_vec, sglist, last_sg);
518 else if (rq->bio && bio_op(rq->bio) == REQ_OP_WRITE_SAME)
519 nsegs = __blk_bvec_map_sg(bio_iovec(rq->bio), sglist, last_sg);
520 else if (rq->bio)
521 nsegs = __blk_bios_map_sg(q, rq->bio, sglist, last_sg);
522
523 if (*last_sg)
524 sg_mark_end(*last_sg);
525
526 /*
527 * Something must have been wrong if the figured number of
528 * segment is bigger than number of req's physical segments
529 */
530 WARN_ON(nsegs > blk_rq_nr_phys_segments(rq));
531
532 return nsegs;
533}
534EXPORT_SYMBOL(__blk_rq_map_sg);
535
536static inline unsigned int blk_rq_get_max_segments(struct request *rq)
537{
538 if (req_op(rq) == REQ_OP_DISCARD)
539 return queue_max_discard_segments(rq->q);
540 return queue_max_segments(rq->q);
541}
542
543static inline int ll_new_hw_segment(struct request *req, struct bio *bio,
544 unsigned int nr_phys_segs)
545{
546 if (req->nr_phys_segments + nr_phys_segs > blk_rq_get_max_segments(req))
547 goto no_merge;
548
549 if (blk_integrity_merge_bio(req->q, req, bio) == false)
550 goto no_merge;
551
552 /*
553 * This will form the start of a new hw segment. Bump both
554 * counters.
555 */
556 req->nr_phys_segments += nr_phys_segs;
557 return 1;
558
559no_merge:
560 req_set_nomerge(req->q, req);
561 return 0;
562}
563
564int ll_back_merge_fn(struct request *req, struct bio *bio, unsigned int nr_segs)
565{
566 if (req_gap_back_merge(req, bio))
567 return 0;
568 if (blk_integrity_rq(req) &&
569 integrity_req_gap_back_merge(req, bio))
570 return 0;
571 if (!bio_crypt_ctx_back_mergeable(req, bio))
572 return 0;
573 if (blk_rq_sectors(req) + bio_sectors(bio) >
574 blk_rq_get_max_sectors(req, blk_rq_pos(req))) {
575 req_set_nomerge(req->q, req);
576 return 0;
577 }
578
579 return ll_new_hw_segment(req, bio, nr_segs);
580}
581
582int ll_front_merge_fn(struct request *req, struct bio *bio, unsigned int nr_segs)
583{
584 if (req_gap_front_merge(req, bio))
585 return 0;
586 if (blk_integrity_rq(req) &&
587 integrity_req_gap_front_merge(req, bio))
588 return 0;
589 if (!bio_crypt_ctx_front_mergeable(req, bio))
590 return 0;
591 if (blk_rq_sectors(req) + bio_sectors(bio) >
592 blk_rq_get_max_sectors(req, bio->bi_iter.bi_sector)) {
593 req_set_nomerge(req->q, req);
594 return 0;
595 }
596
597 return ll_new_hw_segment(req, bio, nr_segs);
598}
599
600static bool req_attempt_discard_merge(struct request_queue *q, struct request *req,
601 struct request *next)
602{
603 unsigned short segments = blk_rq_nr_discard_segments(req);
604
605 if (segments >= queue_max_discard_segments(q))
606 goto no_merge;
607 if (blk_rq_sectors(req) + bio_sectors(next->bio) >
608 blk_rq_get_max_sectors(req, blk_rq_pos(req)))
609 goto no_merge;
610
611 req->nr_phys_segments = segments + blk_rq_nr_discard_segments(next);
612 return true;
613no_merge:
614 req_set_nomerge(q, req);
615 return false;
616}
617
618static int ll_merge_requests_fn(struct request_queue *q, struct request *req,
619 struct request *next)
620{
621 int total_phys_segments;
622
623 if (req_gap_back_merge(req, next->bio))
624 return 0;
625
626 /*
627 * Will it become too large?
628 */
629 if ((blk_rq_sectors(req) + blk_rq_sectors(next)) >
630 blk_rq_get_max_sectors(req, blk_rq_pos(req)))
631 return 0;
632
633 total_phys_segments = req->nr_phys_segments + next->nr_phys_segments;
634 if (total_phys_segments > blk_rq_get_max_segments(req))
635 return 0;
636
637 if (blk_integrity_merge_rq(q, req, next) == false)
638 return 0;
639
640 if (!bio_crypt_ctx_merge_rq(req, next))
641 return 0;
642
643 /* Merge is OK... */
644 req->nr_phys_segments = total_phys_segments;
645 return 1;
646}
647
648/**
649 * blk_rq_set_mixed_merge - mark a request as mixed merge
650 * @rq: request to mark as mixed merge
651 *
652 * Description:
653 * @rq is about to be mixed merged. Make sure the attributes
654 * which can be mixed are set in each bio and mark @rq as mixed
655 * merged.
656 */
657void blk_rq_set_mixed_merge(struct request *rq)
658{
659 unsigned int ff = rq->cmd_flags & REQ_FAILFAST_MASK;
660 struct bio *bio;
661
662 if (rq->rq_flags & RQF_MIXED_MERGE)
663 return;
664
665 /*
666 * @rq will no longer represent mixable attributes for all the
667 * contained bios. It will just track those of the first one.
668 * Distributes the attributs to each bio.
669 */
670 for (bio = rq->bio; bio; bio = bio->bi_next) {
671 WARN_ON_ONCE((bio->bi_opf & REQ_FAILFAST_MASK) &&
672 (bio->bi_opf & REQ_FAILFAST_MASK) != ff);
673 bio->bi_opf |= ff;
674 }
675 rq->rq_flags |= RQF_MIXED_MERGE;
676}
677
678static void blk_account_io_merge_request(struct request *req)
679{
680 if (blk_do_io_stat(req)) {
681 part_stat_lock();
682 part_stat_inc(req->part, merges[op_stat_group(req_op(req))]);
683 part_stat_unlock();
684
685 hd_struct_put(req->part);
686 }
687}
688
689/*
690 * Two cases of handling DISCARD merge:
691 * If max_discard_segments > 1, the driver takes every bio
692 * as a range and send them to controller together. The ranges
693 * needn't to be contiguous.
694 * Otherwise, the bios/requests will be handled as same as
695 * others which should be contiguous.
696 */
697static inline bool blk_discard_mergable(struct request *req)
698{
699 if (req_op(req) == REQ_OP_DISCARD &&
700 queue_max_discard_segments(req->q) > 1)
701 return true;
702 return false;
703}
704
705static enum elv_merge blk_try_req_merge(struct request *req,
706 struct request *next)
707{
708 if (blk_discard_mergable(req))
709 return ELEVATOR_DISCARD_MERGE;
710 else if (blk_rq_pos(req) + blk_rq_sectors(req) == blk_rq_pos(next))
711 return ELEVATOR_BACK_MERGE;
712
713 return ELEVATOR_NO_MERGE;
714}
715
716/*
717 * For non-mq, this has to be called with the request spinlock acquired.
718 * For mq with scheduling, the appropriate queue wide lock should be held.
719 */
720static struct request *attempt_merge(struct request_queue *q,
721 struct request *req, struct request *next)
722{
723 if (!rq_mergeable(req) || !rq_mergeable(next))
724 return NULL;
725
726 if (req_op(req) != req_op(next))
727 return NULL;
728
729 if (rq_data_dir(req) != rq_data_dir(next)
730 || req->rq_disk != next->rq_disk)
731 return NULL;
732
733 if (req_op(req) == REQ_OP_WRITE_SAME &&
734 !blk_write_same_mergeable(req->bio, next->bio))
735 return NULL;
736
737 /*
738 * Don't allow merge of different write hints, or for a hint with
739 * non-hint IO.
740 */
741 if (req->write_hint != next->write_hint)
742 return NULL;
743
744 if (req->ioprio != next->ioprio)
745 return NULL;
746
747 /*
748 * If we are allowed to merge, then append bio list
749 * from next to rq and release next. merge_requests_fn
750 * will have updated segment counts, update sector
751 * counts here. Handle DISCARDs separately, as they
752 * have separate settings.
753 */
754
755 switch (blk_try_req_merge(req, next)) {
756 case ELEVATOR_DISCARD_MERGE:
757 if (!req_attempt_discard_merge(q, req, next))
758 return NULL;
759 break;
760 case ELEVATOR_BACK_MERGE:
761 if (!ll_merge_requests_fn(q, req, next))
762 return NULL;
763 break;
764 default:
765 return NULL;
766 }
767
768 /*
769 * If failfast settings disagree or any of the two is already
770 * a mixed merge, mark both as mixed before proceeding. This
771 * makes sure that all involved bios have mixable attributes
772 * set properly.
773 */
774 if (((req->rq_flags | next->rq_flags) & RQF_MIXED_MERGE) ||
775 (req->cmd_flags & REQ_FAILFAST_MASK) !=
776 (next->cmd_flags & REQ_FAILFAST_MASK)) {
777 blk_rq_set_mixed_merge(req);
778 blk_rq_set_mixed_merge(next);
779 }
780
781 /*
782 * At this point we have either done a back merge or front merge. We
783 * need the smaller start_time_ns of the merged requests to be the
784 * current request for accounting purposes.
785 */
786 if (next->start_time_ns < req->start_time_ns)
787 req->start_time_ns = next->start_time_ns;
788
789 req->biotail->bi_next = next->bio;
790 req->biotail = next->biotail;
791
792 req->__data_len += blk_rq_bytes(next);
793
794 if (!blk_discard_mergable(req))
795 elv_merge_requests(q, req, next);
796
797 /*
798 * 'next' is going away, so update stats accordingly
799 */
800 blk_account_io_merge_request(next);
801
802 trace_block_rq_merge(q, next);
803
804 /*
805 * ownership of bio passed from next to req, return 'next' for
806 * the caller to free
807 */
808 next->bio = NULL;
809 return next;
810}
811
812struct request *attempt_back_merge(struct request_queue *q, struct request *rq)
813{
814 struct request *next = elv_latter_request(q, rq);
815
816 if (next)
817 return attempt_merge(q, rq, next);
818
819 return NULL;
820}
821
822struct request *attempt_front_merge(struct request_queue *q, struct request *rq)
823{
824 struct request *prev = elv_former_request(q, rq);
825
826 if (prev)
827 return attempt_merge(q, prev, rq);
828
829 return NULL;
830}
831
832int blk_attempt_req_merge(struct request_queue *q, struct request *rq,
833 struct request *next)
834{
835 struct request *free;
836
837 free = attempt_merge(q, rq, next);
838 if (free) {
839 blk_put_request(free);
840 return 1;
841 }
842
843 return 0;
844}
845
846bool blk_rq_merge_ok(struct request *rq, struct bio *bio)
847{
848 if (!rq_mergeable(rq) || !bio_mergeable(bio))
849 return false;
850
851 if (req_op(rq) != bio_op(bio))
852 return false;
853
854 /* different data direction or already started, don't merge */
855 if (bio_data_dir(bio) != rq_data_dir(rq))
856 return false;
857
858 /* must be same device */
859 if (rq->rq_disk != bio->bi_disk)
860 return false;
861
862 /* only merge integrity protected bio into ditto rq */
863 if (blk_integrity_merge_bio(rq->q, rq, bio) == false)
864 return false;
865
866 /* Only merge if the crypt contexts are compatible */
867 if (!bio_crypt_rq_ctx_compatible(rq, bio))
868 return false;
869
870 /* must be using the same buffer */
871 if (req_op(rq) == REQ_OP_WRITE_SAME &&
872 !blk_write_same_mergeable(rq->bio, bio))
873 return false;
874
875 /*
876 * Don't allow merge of different write hints, or for a hint with
877 * non-hint IO.
878 */
879 if (rq->write_hint != bio->bi_write_hint)
880 return false;
881
882 if (rq->ioprio != bio_prio(bio))
883 return false;
884
885 return true;
886}
887
888enum elv_merge blk_try_merge(struct request *rq, struct bio *bio)
889{
890 if (blk_discard_mergable(rq))
891 return ELEVATOR_DISCARD_MERGE;
892 else if (blk_rq_pos(rq) + blk_rq_sectors(rq) == bio->bi_iter.bi_sector)
893 return ELEVATOR_BACK_MERGE;
894 else if (blk_rq_pos(rq) - bio_sectors(bio) == bio->bi_iter.bi_sector)
895 return ELEVATOR_FRONT_MERGE;
896 return ELEVATOR_NO_MERGE;
897}