1 // SPDX-License-Identifier: GPL-2.0
3 * Functions related to segment and merge handling
5 #include <linux/kernel.h>
6 #include <linux/module.h>
8 #include <linux/blkdev.h>
9 #include <linux/scatterlist.h>
11 #include <trace/events/block.h>
16 * Check if the two bvecs from two bios can be merged to one segment. If yes,
17 * no need to check gap between the two bios since the 1st bio and the 1st bvec
18 * in the 2nd bio can be handled in one segment.
20 static inline bool bios_segs_mergeable(struct request_queue *q,
21 struct bio *prev, struct bio_vec *prev_last_bv,
22 struct bio_vec *next_first_bv)
24 if (!biovec_phys_mergeable(q, prev_last_bv, next_first_bv))
26 if (prev->bi_seg_back_size + next_first_bv->bv_len >
27 queue_max_segment_size(q))
32 static inline bool bio_will_gap(struct request_queue *q,
33 struct request *prev_rq, struct bio *prev, struct bio *next)
35 struct bio_vec pb, nb;
37 if (!bio_has_data(prev) || !queue_virt_boundary(q))
41 * Don't merge if the 1st bio starts with non-zero offset, otherwise it
42 * is quite difficult to respect the sg gap limit. We work hard to
43 * merge a huge number of small single bios in case of mkfs.
46 bio_get_first_bvec(prev_rq->bio, &pb);
48 bio_get_first_bvec(prev, &pb);
49 if (pb.bv_offset & queue_virt_boundary(q))
53 * We don't need to worry about the situation that the merged segment
54 * ends in unaligned virt boundary:
56 * - if 'pb' ends aligned, the merged segment ends aligned
57 * - if 'pb' ends unaligned, the next bio must include
58 * one single bvec of 'nb', otherwise the 'nb' can't
61 bio_get_last_bvec(prev, &pb);
62 bio_get_first_bvec(next, &nb);
63 if (bios_segs_mergeable(q, prev, &pb, &nb))
65 return __bvec_gap_to_prev(q, &pb, nb.bv_offset);
68 static inline bool req_gap_back_merge(struct request *req, struct bio *bio)
70 return bio_will_gap(req->q, req, req->biotail, bio);
73 static inline bool req_gap_front_merge(struct request *req, struct bio *bio)
75 return bio_will_gap(req->q, NULL, bio, req->bio);
78 static struct bio *blk_bio_discard_split(struct request_queue *q,
83 unsigned int max_discard_sectors, granularity;
86 unsigned split_sectors;
90 /* Zero-sector (unknown) and one-sector granularities are the same. */
91 granularity = max(q->limits.discard_granularity >> 9, 1U);
93 max_discard_sectors = min(q->limits.max_discard_sectors,
94 bio_allowed_max_sectors(q));
95 max_discard_sectors -= max_discard_sectors % granularity;
97 if (unlikely(!max_discard_sectors)) {
102 if (bio_sectors(bio) <= max_discard_sectors)
105 split_sectors = max_discard_sectors;
108 * If the next starting sector would be misaligned, stop the discard at
109 * the previous aligned sector.
111 alignment = (q->limits.discard_alignment >> 9) % granularity;
113 tmp = bio->bi_iter.bi_sector + split_sectors - alignment;
114 tmp = sector_div(tmp, granularity);
116 if (split_sectors > tmp)
117 split_sectors -= tmp;
119 return bio_split(bio, split_sectors, GFP_NOIO, bs);
122 static struct bio *blk_bio_write_zeroes_split(struct request_queue *q,
123 struct bio *bio, struct bio_set *bs, unsigned *nsegs)
127 if (!q->limits.max_write_zeroes_sectors)
130 if (bio_sectors(bio) <= q->limits.max_write_zeroes_sectors)
133 return bio_split(bio, q->limits.max_write_zeroes_sectors, GFP_NOIO, bs);
136 static struct bio *blk_bio_write_same_split(struct request_queue *q,
143 if (!q->limits.max_write_same_sectors)
146 if (bio_sectors(bio) <= q->limits.max_write_same_sectors)
149 return bio_split(bio, q->limits.max_write_same_sectors, GFP_NOIO, bs);
152 static inline unsigned get_max_io_size(struct request_queue *q,
155 unsigned sectors = blk_max_size_offset(q, bio->bi_iter.bi_sector);
156 unsigned mask = queue_logical_block_size(q) - 1;
158 /* aligned to logical block size */
159 sectors &= ~(mask >> 9);
164 static struct bio *blk_bio_segment_split(struct request_queue *q,
169 struct bio_vec bv, bvprv, *bvprvp = NULL;
170 struct bvec_iter iter;
171 unsigned seg_size = 0, nsegs = 0, sectors = 0;
172 unsigned front_seg_size = bio->bi_seg_front_size;
173 bool do_split = true;
174 struct bio *new = NULL;
175 const unsigned max_sectors = get_max_io_size(q, bio);
177 bio_for_each_segment(bv, bio, iter) {
179 * If the queue doesn't support SG gaps and adding this
180 * offset would create a gap, disallow it.
182 if (bvprvp && bvec_gap_to_prev(q, bvprvp, bv.bv_offset))
185 if (sectors + (bv.bv_len >> 9) > max_sectors) {
187 * Consider this a new segment if we're splitting in
188 * the middle of this vector.
190 if (nsegs < queue_max_segments(q) &&
191 sectors < max_sectors) {
193 sectors = max_sectors;
198 if (bvprvp && blk_queue_cluster(q)) {
199 if (seg_size + bv.bv_len > queue_max_segment_size(q))
201 if (!biovec_phys_mergeable(q, bvprvp, &bv))
204 seg_size += bv.bv_len;
207 sectors += bv.bv_len >> 9;
212 if (nsegs == queue_max_segments(q))
215 if (nsegs == 1 && seg_size > front_seg_size)
216 front_seg_size = seg_size;
221 seg_size = bv.bv_len;
222 sectors += bv.bv_len >> 9;
231 new = bio_split(bio, sectors, GFP_NOIO, bs);
236 if (nsegs == 1 && seg_size > front_seg_size)
237 front_seg_size = seg_size;
238 bio->bi_seg_front_size = front_seg_size;
239 if (seg_size > bio->bi_seg_back_size)
240 bio->bi_seg_back_size = seg_size;
242 return do_split ? new : NULL;
245 void blk_queue_split(struct request_queue *q, struct bio **bio)
247 struct bio *split, *res;
250 switch (bio_op(*bio)) {
252 case REQ_OP_SECURE_ERASE:
253 split = blk_bio_discard_split(q, *bio, &q->bio_split, &nsegs);
255 case REQ_OP_WRITE_ZEROES:
256 split = blk_bio_write_zeroes_split(q, *bio, &q->bio_split, &nsegs);
258 case REQ_OP_WRITE_SAME:
259 split = blk_bio_write_same_split(q, *bio, &q->bio_split, &nsegs);
262 split = blk_bio_segment_split(q, *bio, &q->bio_split, &nsegs);
266 /* physical segments can be figured out during splitting */
267 res = split ? split : *bio;
268 res->bi_phys_segments = nsegs;
269 bio_set_flag(res, BIO_SEG_VALID);
272 /* there isn't chance to merge the splitted bio */
273 split->bi_opf |= REQ_NOMERGE;
276 * Since we're recursing into make_request here, ensure
277 * that we mark this bio as already having entered the queue.
278 * If not, and the queue is going away, we can get stuck
279 * forever on waiting for the queue reference to drop. But
280 * that will never happen, as we're already holding a
283 bio_set_flag(*bio, BIO_QUEUE_ENTERED);
285 bio_chain(split, *bio);
286 trace_block_split(q, split, (*bio)->bi_iter.bi_sector);
287 generic_make_request(*bio);
291 EXPORT_SYMBOL(blk_queue_split);
293 static unsigned int __blk_recalc_rq_segments(struct request_queue *q,
297 struct bio_vec bv, bvprv = { NULL };
298 int cluster, prev = 0;
299 unsigned int seg_size, nr_phys_segs;
300 struct bio *fbio, *bbio;
301 struct bvec_iter iter;
306 switch (bio_op(bio)) {
308 case REQ_OP_SECURE_ERASE:
309 case REQ_OP_WRITE_ZEROES:
311 case REQ_OP_WRITE_SAME:
316 cluster = blk_queue_cluster(q);
320 bio_for_each_segment(bv, bio, iter) {
322 * If SG merging is disabled, each bio vector is
328 if (prev && cluster) {
329 if (seg_size + bv.bv_len
330 > queue_max_segment_size(q))
332 if (!biovec_phys_mergeable(q, &bvprv, &bv))
335 seg_size += bv.bv_len;
340 if (nr_phys_segs == 1 && seg_size >
341 fbio->bi_seg_front_size)
342 fbio->bi_seg_front_size = seg_size;
347 seg_size = bv.bv_len;
352 if (nr_phys_segs == 1 && seg_size > fbio->bi_seg_front_size)
353 fbio->bi_seg_front_size = seg_size;
354 if (seg_size > bbio->bi_seg_back_size)
355 bbio->bi_seg_back_size = seg_size;
360 void blk_recalc_rq_segments(struct request *rq)
362 bool no_sg_merge = !!test_bit(QUEUE_FLAG_NO_SG_MERGE,
363 &rq->q->queue_flags);
365 rq->nr_phys_segments = __blk_recalc_rq_segments(rq->q, rq->bio,
369 void blk_recount_segments(struct request_queue *q, struct bio *bio)
371 unsigned short seg_cnt;
373 /* estimate segment number by bi_vcnt for non-cloned bio */
374 if (bio_flagged(bio, BIO_CLONED))
375 seg_cnt = bio_segments(bio);
377 seg_cnt = bio->bi_vcnt;
379 if (test_bit(QUEUE_FLAG_NO_SG_MERGE, &q->queue_flags) &&
380 (seg_cnt < queue_max_segments(q)))
381 bio->bi_phys_segments = seg_cnt;
383 struct bio *nxt = bio->bi_next;
386 bio->bi_phys_segments = __blk_recalc_rq_segments(q, bio, false);
390 bio_set_flag(bio, BIO_SEG_VALID);
392 EXPORT_SYMBOL(blk_recount_segments);
394 static int blk_phys_contig_segment(struct request_queue *q, struct bio *bio,
397 struct bio_vec end_bv = { NULL }, nxt_bv;
399 if (!blk_queue_cluster(q))
402 if (bio->bi_seg_back_size + nxt->bi_seg_front_size >
403 queue_max_segment_size(q))
406 if (!bio_has_data(bio))
409 bio_get_last_bvec(bio, &end_bv);
410 bio_get_first_bvec(nxt, &nxt_bv);
412 return biovec_phys_mergeable(q, &end_bv, &nxt_bv);
416 __blk_segment_map_sg(struct request_queue *q, struct bio_vec *bvec,
417 struct scatterlist *sglist, struct bio_vec *bvprv,
418 struct scatterlist **sg, int *nsegs, int *cluster)
421 int nbytes = bvec->bv_len;
423 if (*sg && *cluster) {
424 if ((*sg)->length + nbytes > queue_max_segment_size(q))
426 if (!biovec_phys_mergeable(q, bvprv, bvec))
429 (*sg)->length += nbytes;
436 * If the driver previously mapped a shorter
437 * list, we could see a termination bit
438 * prematurely unless it fully inits the sg
439 * table on each mapping. We KNOW that there
440 * must be more entries here or the driver
441 * would be buggy, so force clear the
442 * termination bit to avoid doing a full
443 * sg_init_table() in drivers for each command.
449 sg_set_page(*sg, bvec->bv_page, nbytes, bvec->bv_offset);
455 static inline int __blk_bvec_map_sg(struct request_queue *q, struct bio_vec bv,
456 struct scatterlist *sglist, struct scatterlist **sg)
459 sg_set_page(*sg, bv.bv_page, bv.bv_len, bv.bv_offset);
463 static int __blk_bios_map_sg(struct request_queue *q, struct bio *bio,
464 struct scatterlist *sglist,
465 struct scatterlist **sg)
467 struct bio_vec bvec, bvprv = { NULL };
468 struct bvec_iter iter;
469 int cluster = blk_queue_cluster(q), nsegs = 0;
472 bio_for_each_segment(bvec, bio, iter)
473 __blk_segment_map_sg(q, &bvec, sglist, &bvprv, sg,
480 * map a request to scatterlist, return number of sg entries setup. Caller
481 * must make sure sg can hold rq->nr_phys_segments entries
483 int blk_rq_map_sg(struct request_queue *q, struct request *rq,
484 struct scatterlist *sglist)
486 struct scatterlist *sg = NULL;
489 if (rq->rq_flags & RQF_SPECIAL_PAYLOAD)
490 nsegs = __blk_bvec_map_sg(q, rq->special_vec, sglist, &sg);
491 else if (rq->bio && bio_op(rq->bio) == REQ_OP_WRITE_SAME)
492 nsegs = __blk_bvec_map_sg(q, bio_iovec(rq->bio), sglist, &sg);
494 nsegs = __blk_bios_map_sg(q, rq->bio, sglist, &sg);
496 if (unlikely(rq->rq_flags & RQF_COPY_USER) &&
497 (blk_rq_bytes(rq) & q->dma_pad_mask)) {
498 unsigned int pad_len =
499 (q->dma_pad_mask & ~blk_rq_bytes(rq)) + 1;
501 sg->length += pad_len;
502 rq->extra_len += pad_len;
505 if (q->dma_drain_size && q->dma_drain_needed(rq)) {
506 if (op_is_write(req_op(rq)))
507 memset(q->dma_drain_buffer, 0, q->dma_drain_size);
511 sg_set_page(sg, virt_to_page(q->dma_drain_buffer),
513 ((unsigned long)q->dma_drain_buffer) &
516 rq->extra_len += q->dma_drain_size;
523 * Something must have been wrong if the figured number of
524 * segment is bigger than number of req's physical segments
526 WARN_ON(nsegs > blk_rq_nr_phys_segments(rq));
530 EXPORT_SYMBOL(blk_rq_map_sg);
532 static inline int ll_new_hw_segment(struct request_queue *q,
536 int nr_phys_segs = bio_phys_segments(q, bio);
538 if (req->nr_phys_segments + nr_phys_segs > queue_max_segments(q))
541 if (blk_integrity_merge_bio(q, req, bio) == false)
545 * This will form the start of a new hw segment. Bump both
548 req->nr_phys_segments += nr_phys_segs;
552 req_set_nomerge(q, req);
556 int ll_back_merge_fn(struct request_queue *q, struct request *req,
559 if (req_gap_back_merge(req, bio))
561 if (blk_integrity_rq(req) &&
562 integrity_req_gap_back_merge(req, bio))
564 if (blk_rq_sectors(req) + bio_sectors(bio) >
565 blk_rq_get_max_sectors(req, blk_rq_pos(req))) {
566 req_set_nomerge(q, req);
569 if (!bio_flagged(req->biotail, BIO_SEG_VALID))
570 blk_recount_segments(q, req->biotail);
571 if (!bio_flagged(bio, BIO_SEG_VALID))
572 blk_recount_segments(q, bio);
574 return ll_new_hw_segment(q, req, bio);
577 int ll_front_merge_fn(struct request_queue *q, struct request *req,
581 if (req_gap_front_merge(req, bio))
583 if (blk_integrity_rq(req) &&
584 integrity_req_gap_front_merge(req, bio))
586 if (blk_rq_sectors(req) + bio_sectors(bio) >
587 blk_rq_get_max_sectors(req, bio->bi_iter.bi_sector)) {
588 req_set_nomerge(q, req);
591 if (!bio_flagged(bio, BIO_SEG_VALID))
592 blk_recount_segments(q, bio);
593 if (!bio_flagged(req->bio, BIO_SEG_VALID))
594 blk_recount_segments(q, req->bio);
596 return ll_new_hw_segment(q, req, bio);
600 * blk-mq uses req->special to carry normal driver per-request payload, it
601 * does not indicate a prepared command that we cannot merge with.
603 static bool req_no_special_merge(struct request *req)
605 struct request_queue *q = req->q;
607 return !q->mq_ops && req->special;
610 static bool req_attempt_discard_merge(struct request_queue *q, struct request *req,
611 struct request *next)
613 unsigned short segments = blk_rq_nr_discard_segments(req);
615 if (segments >= queue_max_discard_segments(q))
617 if (blk_rq_sectors(req) + bio_sectors(next->bio) >
618 blk_rq_get_max_sectors(req, blk_rq_pos(req)))
621 req->nr_phys_segments = segments + blk_rq_nr_discard_segments(next);
624 req_set_nomerge(q, req);
628 static int ll_merge_requests_fn(struct request_queue *q, struct request *req,
629 struct request *next)
631 int total_phys_segments;
632 unsigned int seg_size =
633 req->biotail->bi_seg_back_size + next->bio->bi_seg_front_size;
636 * First check if the either of the requests are re-queued
637 * requests. Can't merge them if they are.
639 if (req_no_special_merge(req) || req_no_special_merge(next))
642 if (req_gap_back_merge(req, next->bio))
646 * Will it become too large?
648 if ((blk_rq_sectors(req) + blk_rq_sectors(next)) >
649 blk_rq_get_max_sectors(req, blk_rq_pos(req)))
652 total_phys_segments = req->nr_phys_segments + next->nr_phys_segments;
653 if (blk_phys_contig_segment(q, req->biotail, next->bio)) {
654 if (req->nr_phys_segments == 1)
655 req->bio->bi_seg_front_size = seg_size;
656 if (next->nr_phys_segments == 1)
657 next->biotail->bi_seg_back_size = seg_size;
658 total_phys_segments--;
661 if (total_phys_segments > queue_max_segments(q))
664 if (blk_integrity_merge_rq(q, req, next) == false)
668 req->nr_phys_segments = total_phys_segments;
673 * blk_rq_set_mixed_merge - mark a request as mixed merge
674 * @rq: request to mark as mixed merge
677 * @rq is about to be mixed merged. Make sure the attributes
678 * which can be mixed are set in each bio and mark @rq as mixed
681 void blk_rq_set_mixed_merge(struct request *rq)
683 unsigned int ff = rq->cmd_flags & REQ_FAILFAST_MASK;
686 if (rq->rq_flags & RQF_MIXED_MERGE)
690 * @rq will no longer represent mixable attributes for all the
691 * contained bios. It will just track those of the first one.
692 * Distributes the attributs to each bio.
694 for (bio = rq->bio; bio; bio = bio->bi_next) {
695 WARN_ON_ONCE((bio->bi_opf & REQ_FAILFAST_MASK) &&
696 (bio->bi_opf & REQ_FAILFAST_MASK) != ff);
699 rq->rq_flags |= RQF_MIXED_MERGE;
702 static void blk_account_io_merge(struct request *req)
704 if (blk_do_io_stat(req)) {
705 struct hd_struct *part;
708 cpu = part_stat_lock();
711 part_round_stats(req->q, cpu, part);
712 part_dec_in_flight(req->q, part, rq_data_dir(req));
719 * Two cases of handling DISCARD merge:
720 * If max_discard_segments > 1, the driver takes every bio
721 * as a range and send them to controller together. The ranges
722 * needn't to be contiguous.
723 * Otherwise, the bios/requests will be handled as same as
724 * others which should be contiguous.
726 static inline bool blk_discard_mergable(struct request *req)
728 if (req_op(req) == REQ_OP_DISCARD &&
729 queue_max_discard_segments(req->q) > 1)
734 enum elv_merge blk_try_req_merge(struct request *req, struct request *next)
736 if (blk_discard_mergable(req))
737 return ELEVATOR_DISCARD_MERGE;
738 else if (blk_rq_pos(req) + blk_rq_sectors(req) == blk_rq_pos(next))
739 return ELEVATOR_BACK_MERGE;
741 return ELEVATOR_NO_MERGE;
745 * For non-mq, this has to be called with the request spinlock acquired.
746 * For mq with scheduling, the appropriate queue wide lock should be held.
748 static struct request *attempt_merge(struct request_queue *q,
749 struct request *req, struct request *next)
752 lockdep_assert_held(q->queue_lock);
754 if (!rq_mergeable(req) || !rq_mergeable(next))
757 if (req_op(req) != req_op(next))
760 if (rq_data_dir(req) != rq_data_dir(next)
761 || req->rq_disk != next->rq_disk
762 || req_no_special_merge(next))
765 if (req_op(req) == REQ_OP_WRITE_SAME &&
766 !blk_write_same_mergeable(req->bio, next->bio))
770 * Don't allow merge of different write hints, or for a hint with
773 if (req->write_hint != next->write_hint)
777 * If we are allowed to merge, then append bio list
778 * from next to rq and release next. merge_requests_fn
779 * will have updated segment counts, update sector
780 * counts here. Handle DISCARDs separately, as they
781 * have separate settings.
784 switch (blk_try_req_merge(req, next)) {
785 case ELEVATOR_DISCARD_MERGE:
786 if (!req_attempt_discard_merge(q, req, next))
789 case ELEVATOR_BACK_MERGE:
790 if (!ll_merge_requests_fn(q, req, next))
798 * If failfast settings disagree or any of the two is already
799 * a mixed merge, mark both as mixed before proceeding. This
800 * makes sure that all involved bios have mixable attributes
803 if (((req->rq_flags | next->rq_flags) & RQF_MIXED_MERGE) ||
804 (req->cmd_flags & REQ_FAILFAST_MASK) !=
805 (next->cmd_flags & REQ_FAILFAST_MASK)) {
806 blk_rq_set_mixed_merge(req);
807 blk_rq_set_mixed_merge(next);
811 * At this point we have either done a back merge or front merge. We
812 * need the smaller start_time_ns of the merged requests to be the
813 * current request for accounting purposes.
815 if (next->start_time_ns < req->start_time_ns)
816 req->start_time_ns = next->start_time_ns;
818 req->biotail->bi_next = next->bio;
819 req->biotail = next->biotail;
821 req->__data_len += blk_rq_bytes(next);
823 if (req_op(req) != REQ_OP_DISCARD)
824 elv_merge_requests(q, req, next);
827 * 'next' is going away, so update stats accordingly
829 blk_account_io_merge(next);
831 req->ioprio = ioprio_best(req->ioprio, next->ioprio);
832 if (blk_rq_cpu_valid(next))
833 req->cpu = next->cpu;
836 * ownership of bio passed from next to req, return 'next' for
843 struct request *attempt_back_merge(struct request_queue *q, struct request *rq)
845 struct request *next = elv_latter_request(q, rq);
848 return attempt_merge(q, rq, next);
853 struct request *attempt_front_merge(struct request_queue *q, struct request *rq)
855 struct request *prev = elv_former_request(q, rq);
858 return attempt_merge(q, prev, rq);
863 int blk_attempt_req_merge(struct request_queue *q, struct request *rq,
864 struct request *next)
866 struct elevator_queue *e = q->elevator;
867 struct request *free;
869 if (!e->uses_mq && e->type->ops.sq.elevator_allow_rq_merge_fn)
870 if (!e->type->ops.sq.elevator_allow_rq_merge_fn(q, rq, next))
873 free = attempt_merge(q, rq, next);
875 __blk_put_request(q, free);
882 bool blk_rq_merge_ok(struct request *rq, struct bio *bio)
884 if (!rq_mergeable(rq) || !bio_mergeable(bio))
887 if (req_op(rq) != bio_op(bio))
890 /* different data direction or already started, don't merge */
891 if (bio_data_dir(bio) != rq_data_dir(rq))
894 /* must be same device and not a special request */
895 if (rq->rq_disk != bio->bi_disk || req_no_special_merge(rq))
898 /* only merge integrity protected bio into ditto rq */
899 if (blk_integrity_merge_bio(rq->q, rq, bio) == false)
902 /* must be using the same buffer */
903 if (req_op(rq) == REQ_OP_WRITE_SAME &&
904 !blk_write_same_mergeable(rq->bio, bio))
908 * Don't allow merge of different write hints, or for a hint with
911 if (rq->write_hint != bio->bi_write_hint)
917 enum elv_merge blk_try_merge(struct request *rq, struct bio *bio)
919 if (blk_discard_mergable(rq))
920 return ELEVATOR_DISCARD_MERGE;
921 else if (blk_rq_pos(rq) + blk_rq_sectors(rq) == bio->bi_iter.bi_sector)
922 return ELEVATOR_BACK_MERGE;
923 else if (blk_rq_pos(rq) - bio_sectors(bio) == bio->bi_iter.bi_sector)
924 return ELEVATOR_FRONT_MERGE;
925 return ELEVATOR_NO_MERGE;