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(prev_last_bv, next_first_bv))
26 if (!BIOVEC_SEG_BOUNDARY(q, prev_last_bv, next_first_bv))
28 if (prev->bi_seg_back_size + next_first_bv->bv_len >
29 queue_max_segment_size(q))
34 static inline bool bio_will_gap(struct request_queue *q,
35 struct request *prev_rq, struct bio *prev, struct bio *next)
37 struct bio_vec pb, nb;
39 if (!bio_has_data(prev) || !queue_virt_boundary(q))
43 * Don't merge if the 1st bio starts with non-zero offset, otherwise it
44 * is quite difficult to respect the sg gap limit. We work hard to
45 * merge a huge number of small single bios in case of mkfs.
48 bio_get_first_bvec(prev_rq->bio, &pb);
50 bio_get_first_bvec(prev, &pb);
55 * We don't need to worry about the situation that the merged segment
56 * ends in unaligned virt boundary:
58 * - if 'pb' ends aligned, the merged segment ends aligned
59 * - if 'pb' ends unaligned, the next bio must include
60 * one single bvec of 'nb', otherwise the 'nb' can't
63 bio_get_last_bvec(prev, &pb);
64 bio_get_first_bvec(next, &nb);
65 if (bios_segs_mergeable(q, prev, &pb, &nb))
67 return __bvec_gap_to_prev(q, &pb, nb.bv_offset);
70 static inline bool req_gap_back_merge(struct request *req, struct bio *bio)
72 return bio_will_gap(req->q, req, req->biotail, bio);
75 static inline bool req_gap_front_merge(struct request *req, struct bio *bio)
77 return bio_will_gap(req->q, NULL, bio, req->bio);
80 static struct bio *blk_bio_discard_split(struct request_queue *q,
85 unsigned int max_discard_sectors, granularity;
88 unsigned split_sectors;
92 /* Zero-sector (unknown) and one-sector granularities are the same. */
93 granularity = max(q->limits.discard_granularity >> 9, 1U);
95 max_discard_sectors = min(q->limits.max_discard_sectors, UINT_MAX >> 9);
96 max_discard_sectors -= max_discard_sectors % granularity;
98 if (unlikely(!max_discard_sectors)) {
103 if (bio_sectors(bio) <= max_discard_sectors)
106 split_sectors = max_discard_sectors;
109 * If the next starting sector would be misaligned, stop the discard at
110 * the previous aligned sector.
112 alignment = (q->limits.discard_alignment >> 9) % granularity;
114 tmp = bio->bi_iter.bi_sector + split_sectors - alignment;
115 tmp = sector_div(tmp, granularity);
117 if (split_sectors > tmp)
118 split_sectors -= tmp;
120 return bio_split(bio, split_sectors, GFP_NOIO, bs);
123 static struct bio *blk_bio_write_zeroes_split(struct request_queue *q,
124 struct bio *bio, struct bio_set *bs, unsigned *nsegs)
128 if (!q->limits.max_write_zeroes_sectors)
131 if (bio_sectors(bio) <= q->limits.max_write_zeroes_sectors)
134 return bio_split(bio, q->limits.max_write_zeroes_sectors, GFP_NOIO, bs);
137 static struct bio *blk_bio_write_same_split(struct request_queue *q,
144 if (!q->limits.max_write_same_sectors)
147 if (bio_sectors(bio) <= q->limits.max_write_same_sectors)
150 return bio_split(bio, q->limits.max_write_same_sectors, GFP_NOIO, bs);
153 static inline unsigned get_max_io_size(struct request_queue *q,
156 unsigned sectors = blk_max_size_offset(q, bio->bi_iter.bi_sector);
157 unsigned mask = queue_logical_block_size(q) - 1;
159 /* aligned to logical block size */
160 sectors &= ~(mask >> 9);
165 static struct bio *blk_bio_segment_split(struct request_queue *q,
170 struct bio_vec bv, bvprv, *bvprvp = NULL;
171 struct bvec_iter iter;
172 unsigned seg_size = 0, nsegs = 0, sectors = 0;
173 unsigned front_seg_size = bio->bi_seg_front_size;
174 bool do_split = true;
175 struct bio *new = NULL;
176 const unsigned max_sectors = get_max_io_size(q, bio);
178 bio_for_each_segment(bv, bio, iter) {
180 * If the queue doesn't support SG gaps and adding this
181 * offset would create a gap, disallow it.
183 if (bvprvp && bvec_gap_to_prev(q, bvprvp, bv.bv_offset))
186 if (sectors + (bv.bv_len >> 9) > max_sectors) {
188 * Consider this a new segment if we're splitting in
189 * the middle of this vector.
191 if (nsegs < queue_max_segments(q) &&
192 sectors < max_sectors) {
194 sectors = max_sectors;
199 if (bvprvp && blk_queue_cluster(q)) {
200 if (seg_size + bv.bv_len > queue_max_segment_size(q))
202 if (!BIOVEC_PHYS_MERGEABLE(bvprvp, &bv))
204 if (!BIOVEC_SEG_BOUNDARY(q, bvprvp, &bv))
207 seg_size += bv.bv_len;
210 sectors += bv.bv_len >> 9;
215 if (nsegs == queue_max_segments(q))
218 if (nsegs == 1 && seg_size > front_seg_size)
219 front_seg_size = seg_size;
224 seg_size = bv.bv_len;
225 sectors += bv.bv_len >> 9;
234 new = bio_split(bio, sectors, GFP_NOIO, bs);
239 if (nsegs == 1 && seg_size > front_seg_size)
240 front_seg_size = seg_size;
241 bio->bi_seg_front_size = front_seg_size;
242 if (seg_size > bio->bi_seg_back_size)
243 bio->bi_seg_back_size = seg_size;
245 return do_split ? new : NULL;
248 void blk_queue_split(struct request_queue *q, struct bio **bio)
250 struct bio *split, *res;
253 switch (bio_op(*bio)) {
255 case REQ_OP_SECURE_ERASE:
256 split = blk_bio_discard_split(q, *bio, &q->bio_split, &nsegs);
258 case REQ_OP_WRITE_ZEROES:
259 split = blk_bio_write_zeroes_split(q, *bio, &q->bio_split, &nsegs);
261 case REQ_OP_WRITE_SAME:
262 split = blk_bio_write_same_split(q, *bio, &q->bio_split, &nsegs);
265 split = blk_bio_segment_split(q, *bio, &q->bio_split, &nsegs);
269 /* physical segments can be figured out during splitting */
270 res = split ? split : *bio;
271 res->bi_phys_segments = nsegs;
272 bio_set_flag(res, BIO_SEG_VALID);
275 /* there isn't chance to merge the splitted bio */
276 split->bi_opf |= REQ_NOMERGE;
279 * Since we're recursing into make_request here, ensure
280 * that we mark this bio as already having entered the queue.
281 * If not, and the queue is going away, we can get stuck
282 * forever on waiting for the queue reference to drop. But
283 * that will never happen, as we're already holding a
286 bio_set_flag(*bio, BIO_QUEUE_ENTERED);
288 bio_chain(split, *bio);
289 trace_block_split(q, split, (*bio)->bi_iter.bi_sector);
290 generic_make_request(*bio);
294 EXPORT_SYMBOL(blk_queue_split);
296 static unsigned int __blk_recalc_rq_segments(struct request_queue *q,
300 struct bio_vec bv, bvprv = { NULL };
301 int cluster, prev = 0;
302 unsigned int seg_size, nr_phys_segs;
303 struct bio *fbio, *bbio;
304 struct bvec_iter iter;
309 switch (bio_op(bio)) {
311 case REQ_OP_SECURE_ERASE:
312 case REQ_OP_WRITE_ZEROES:
314 case REQ_OP_WRITE_SAME:
319 cluster = blk_queue_cluster(q);
323 bio_for_each_segment(bv, bio, iter) {
325 * If SG merging is disabled, each bio vector is
331 if (prev && cluster) {
332 if (seg_size + bv.bv_len
333 > queue_max_segment_size(q))
335 if (!BIOVEC_PHYS_MERGEABLE(&bvprv, &bv))
337 if (!BIOVEC_SEG_BOUNDARY(q, &bvprv, &bv))
340 seg_size += bv.bv_len;
345 if (nr_phys_segs == 1 && seg_size >
346 fbio->bi_seg_front_size)
347 fbio->bi_seg_front_size = seg_size;
352 seg_size = bv.bv_len;
357 if (nr_phys_segs == 1 && seg_size > fbio->bi_seg_front_size)
358 fbio->bi_seg_front_size = seg_size;
359 if (seg_size > bbio->bi_seg_back_size)
360 bbio->bi_seg_back_size = seg_size;
365 void blk_recalc_rq_segments(struct request *rq)
367 bool no_sg_merge = !!test_bit(QUEUE_FLAG_NO_SG_MERGE,
368 &rq->q->queue_flags);
370 rq->nr_phys_segments = __blk_recalc_rq_segments(rq->q, rq->bio,
374 void blk_recount_segments(struct request_queue *q, struct bio *bio)
376 unsigned short seg_cnt;
378 /* estimate segment number by bi_vcnt for non-cloned bio */
379 if (bio_flagged(bio, BIO_CLONED))
380 seg_cnt = bio_segments(bio);
382 seg_cnt = bio->bi_vcnt;
384 if (test_bit(QUEUE_FLAG_NO_SG_MERGE, &q->queue_flags) &&
385 (seg_cnt < queue_max_segments(q)))
386 bio->bi_phys_segments = seg_cnt;
388 struct bio *nxt = bio->bi_next;
391 bio->bi_phys_segments = __blk_recalc_rq_segments(q, bio, false);
395 bio_set_flag(bio, BIO_SEG_VALID);
397 EXPORT_SYMBOL(blk_recount_segments);
399 static int blk_phys_contig_segment(struct request_queue *q, struct bio *bio,
402 struct bio_vec end_bv = { NULL }, nxt_bv;
404 if (!blk_queue_cluster(q))
407 if (bio->bi_seg_back_size + nxt->bi_seg_front_size >
408 queue_max_segment_size(q))
411 if (!bio_has_data(bio))
414 bio_get_last_bvec(bio, &end_bv);
415 bio_get_first_bvec(nxt, &nxt_bv);
417 if (!BIOVEC_PHYS_MERGEABLE(&end_bv, &nxt_bv))
421 * bio and nxt are contiguous in memory; check if the queue allows
422 * these two to be merged into one
424 if (BIOVEC_SEG_BOUNDARY(q, &end_bv, &nxt_bv))
431 __blk_segment_map_sg(struct request_queue *q, struct bio_vec *bvec,
432 struct scatterlist *sglist, struct bio_vec *bvprv,
433 struct scatterlist **sg, int *nsegs, int *cluster)
436 int nbytes = bvec->bv_len;
438 if (*sg && *cluster) {
439 if ((*sg)->length + nbytes > queue_max_segment_size(q))
442 if (!BIOVEC_PHYS_MERGEABLE(bvprv, bvec))
444 if (!BIOVEC_SEG_BOUNDARY(q, bvprv, bvec))
447 (*sg)->length += nbytes;
454 * If the driver previously mapped a shorter
455 * list, we could see a termination bit
456 * prematurely unless it fully inits the sg
457 * table on each mapping. We KNOW that there
458 * must be more entries here or the driver
459 * would be buggy, so force clear the
460 * termination bit to avoid doing a full
461 * sg_init_table() in drivers for each command.
467 sg_set_page(*sg, bvec->bv_page, nbytes, bvec->bv_offset);
473 static inline int __blk_bvec_map_sg(struct request_queue *q, struct bio_vec bv,
474 struct scatterlist *sglist, struct scatterlist **sg)
477 sg_set_page(*sg, bv.bv_page, bv.bv_len, bv.bv_offset);
481 static int __blk_bios_map_sg(struct request_queue *q, struct bio *bio,
482 struct scatterlist *sglist,
483 struct scatterlist **sg)
485 struct bio_vec bvec, bvprv = { NULL };
486 struct bvec_iter iter;
487 int cluster = blk_queue_cluster(q), nsegs = 0;
490 bio_for_each_segment(bvec, bio, iter)
491 __blk_segment_map_sg(q, &bvec, sglist, &bvprv, sg,
498 * map a request to scatterlist, return number of sg entries setup. Caller
499 * must make sure sg can hold rq->nr_phys_segments entries
501 int blk_rq_map_sg(struct request_queue *q, struct request *rq,
502 struct scatterlist *sglist)
504 struct scatterlist *sg = NULL;
507 if (rq->rq_flags & RQF_SPECIAL_PAYLOAD)
508 nsegs = __blk_bvec_map_sg(q, rq->special_vec, sglist, &sg);
509 else if (rq->bio && bio_op(rq->bio) == REQ_OP_WRITE_SAME)
510 nsegs = __blk_bvec_map_sg(q, bio_iovec(rq->bio), sglist, &sg);
512 nsegs = __blk_bios_map_sg(q, rq->bio, sglist, &sg);
514 if (unlikely(rq->rq_flags & RQF_COPY_USER) &&
515 (blk_rq_bytes(rq) & q->dma_pad_mask)) {
516 unsigned int pad_len =
517 (q->dma_pad_mask & ~blk_rq_bytes(rq)) + 1;
519 sg->length += pad_len;
520 rq->extra_len += pad_len;
523 if (q->dma_drain_size && q->dma_drain_needed(rq)) {
524 if (op_is_write(req_op(rq)))
525 memset(q->dma_drain_buffer, 0, q->dma_drain_size);
529 sg_set_page(sg, virt_to_page(q->dma_drain_buffer),
531 ((unsigned long)q->dma_drain_buffer) &
534 rq->extra_len += q->dma_drain_size;
541 * Something must have been wrong if the figured number of
542 * segment is bigger than number of req's physical segments
544 WARN_ON(nsegs > blk_rq_nr_phys_segments(rq));
548 EXPORT_SYMBOL(blk_rq_map_sg);
550 static inline int ll_new_hw_segment(struct request_queue *q,
554 int nr_phys_segs = bio_phys_segments(q, bio);
556 if (req->nr_phys_segments + nr_phys_segs > queue_max_segments(q))
559 if (blk_integrity_merge_bio(q, req, bio) == false)
563 * This will form the start of a new hw segment. Bump both
566 req->nr_phys_segments += nr_phys_segs;
570 req_set_nomerge(q, req);
574 int ll_back_merge_fn(struct request_queue *q, struct request *req,
577 if (req_gap_back_merge(req, bio))
579 if (blk_integrity_rq(req) &&
580 integrity_req_gap_back_merge(req, bio))
582 if (blk_rq_sectors(req) + bio_sectors(bio) >
583 blk_rq_get_max_sectors(req, blk_rq_pos(req))) {
584 req_set_nomerge(q, req);
587 if (!bio_flagged(req->biotail, BIO_SEG_VALID))
588 blk_recount_segments(q, req->biotail);
589 if (!bio_flagged(bio, BIO_SEG_VALID))
590 blk_recount_segments(q, bio);
592 return ll_new_hw_segment(q, req, bio);
595 int ll_front_merge_fn(struct request_queue *q, struct request *req,
599 if (req_gap_front_merge(req, bio))
601 if (blk_integrity_rq(req) &&
602 integrity_req_gap_front_merge(req, bio))
604 if (blk_rq_sectors(req) + bio_sectors(bio) >
605 blk_rq_get_max_sectors(req, bio->bi_iter.bi_sector)) {
606 req_set_nomerge(q, req);
609 if (!bio_flagged(bio, BIO_SEG_VALID))
610 blk_recount_segments(q, bio);
611 if (!bio_flagged(req->bio, BIO_SEG_VALID))
612 blk_recount_segments(q, req->bio);
614 return ll_new_hw_segment(q, req, bio);
618 * blk-mq uses req->special to carry normal driver per-request payload, it
619 * does not indicate a prepared command that we cannot merge with.
621 static bool req_no_special_merge(struct request *req)
623 struct request_queue *q = req->q;
625 return !q->mq_ops && req->special;
628 static bool req_attempt_discard_merge(struct request_queue *q, struct request *req,
629 struct request *next)
631 unsigned short segments = blk_rq_nr_discard_segments(req);
633 if (segments >= queue_max_discard_segments(q))
635 if (blk_rq_sectors(req) + bio_sectors(next->bio) >
636 blk_rq_get_max_sectors(req, blk_rq_pos(req)))
639 req->nr_phys_segments = segments + blk_rq_nr_discard_segments(next);
642 req_set_nomerge(q, req);
646 static int ll_merge_requests_fn(struct request_queue *q, struct request *req,
647 struct request *next)
649 int total_phys_segments;
650 unsigned int seg_size =
651 req->biotail->bi_seg_back_size + next->bio->bi_seg_front_size;
654 * First check if the either of the requests are re-queued
655 * requests. Can't merge them if they are.
657 if (req_no_special_merge(req) || req_no_special_merge(next))
660 if (req_gap_back_merge(req, next->bio))
664 * Will it become too large?
666 if ((blk_rq_sectors(req) + blk_rq_sectors(next)) >
667 blk_rq_get_max_sectors(req, blk_rq_pos(req)))
670 total_phys_segments = req->nr_phys_segments + next->nr_phys_segments;
671 if (blk_phys_contig_segment(q, req->biotail, next->bio)) {
672 if (req->nr_phys_segments == 1)
673 req->bio->bi_seg_front_size = seg_size;
674 if (next->nr_phys_segments == 1)
675 next->biotail->bi_seg_back_size = seg_size;
676 total_phys_segments--;
679 if (total_phys_segments > queue_max_segments(q))
682 if (blk_integrity_merge_rq(q, req, next) == false)
686 req->nr_phys_segments = total_phys_segments;
691 * blk_rq_set_mixed_merge - mark a request as mixed merge
692 * @rq: request to mark as mixed merge
695 * @rq is about to be mixed merged. Make sure the attributes
696 * which can be mixed are set in each bio and mark @rq as mixed
699 void blk_rq_set_mixed_merge(struct request *rq)
701 unsigned int ff = rq->cmd_flags & REQ_FAILFAST_MASK;
704 if (rq->rq_flags & RQF_MIXED_MERGE)
708 * @rq will no longer represent mixable attributes for all the
709 * contained bios. It will just track those of the first one.
710 * Distributes the attributs to each bio.
712 for (bio = rq->bio; bio; bio = bio->bi_next) {
713 WARN_ON_ONCE((bio->bi_opf & REQ_FAILFAST_MASK) &&
714 (bio->bi_opf & REQ_FAILFAST_MASK) != ff);
717 rq->rq_flags |= RQF_MIXED_MERGE;
720 static void blk_account_io_merge(struct request *req)
722 if (blk_do_io_stat(req)) {
723 struct hd_struct *part;
726 cpu = part_stat_lock();
729 part_round_stats(req->q, cpu, part);
730 part_dec_in_flight(req->q, part, rq_data_dir(req));
738 * For non-mq, this has to be called with the request spinlock acquired.
739 * For mq with scheduling, the appropriate queue wide lock should be held.
741 static struct request *attempt_merge(struct request_queue *q,
742 struct request *req, struct request *next)
745 lockdep_assert_held(q->queue_lock);
747 if (!rq_mergeable(req) || !rq_mergeable(next))
750 if (req_op(req) != req_op(next))
756 if (blk_rq_pos(req) + blk_rq_sectors(req) != blk_rq_pos(next))
759 if (rq_data_dir(req) != rq_data_dir(next)
760 || req->rq_disk != next->rq_disk
761 || req_no_special_merge(next))
764 if (req_op(req) == REQ_OP_WRITE_SAME &&
765 !blk_write_same_mergeable(req->bio, next->bio))
769 * Don't allow merge of different write hints, or for a hint with
772 if (req->write_hint != next->write_hint)
776 * If we are allowed to merge, then append bio list
777 * from next to rq and release next. merge_requests_fn
778 * will have updated segment counts, update sector
779 * counts here. Handle DISCARDs separately, as they
780 * have separate settings.
782 if (req_op(req) == REQ_OP_DISCARD) {
783 if (!req_attempt_discard_merge(q, req, next))
785 } else if (!ll_merge_requests_fn(q, req, next))
789 * If failfast settings disagree or any of the two is already
790 * a mixed merge, mark both as mixed before proceeding. This
791 * makes sure that all involved bios have mixable attributes
794 if (((req->rq_flags | next->rq_flags) & RQF_MIXED_MERGE) ||
795 (req->cmd_flags & REQ_FAILFAST_MASK) !=
796 (next->cmd_flags & REQ_FAILFAST_MASK)) {
797 blk_rq_set_mixed_merge(req);
798 blk_rq_set_mixed_merge(next);
802 * At this point we have either done a back merge or front merge. We
803 * need the smaller start_time_ns of the merged requests to be the
804 * current request for accounting purposes.
806 if (next->start_time_ns < req->start_time_ns)
807 req->start_time_ns = next->start_time_ns;
809 req->biotail->bi_next = next->bio;
810 req->biotail = next->biotail;
812 req->__data_len += blk_rq_bytes(next);
814 if (req_op(req) != REQ_OP_DISCARD)
815 elv_merge_requests(q, req, next);
818 * 'next' is going away, so update stats accordingly
820 blk_account_io_merge(next);
822 req->ioprio = ioprio_best(req->ioprio, next->ioprio);
823 if (blk_rq_cpu_valid(next))
824 req->cpu = next->cpu;
827 * ownership of bio passed from next to req, return 'next' for
834 struct request *attempt_back_merge(struct request_queue *q, struct request *rq)
836 struct request *next = elv_latter_request(q, rq);
839 return attempt_merge(q, rq, next);
844 struct request *attempt_front_merge(struct request_queue *q, struct request *rq)
846 struct request *prev = elv_former_request(q, rq);
849 return attempt_merge(q, prev, rq);
854 int blk_attempt_req_merge(struct request_queue *q, struct request *rq,
855 struct request *next)
857 struct elevator_queue *e = q->elevator;
858 struct request *free;
860 if (!e->uses_mq && e->type->ops.sq.elevator_allow_rq_merge_fn)
861 if (!e->type->ops.sq.elevator_allow_rq_merge_fn(q, rq, next))
864 free = attempt_merge(q, rq, next);
866 __blk_put_request(q, free);
873 bool blk_rq_merge_ok(struct request *rq, struct bio *bio)
875 if (!rq_mergeable(rq) || !bio_mergeable(bio))
878 if (req_op(rq) != bio_op(bio))
881 /* different data direction or already started, don't merge */
882 if (bio_data_dir(bio) != rq_data_dir(rq))
885 /* must be same device and not a special request */
886 if (rq->rq_disk != bio->bi_disk || req_no_special_merge(rq))
889 /* only merge integrity protected bio into ditto rq */
890 if (blk_integrity_merge_bio(rq->q, rq, bio) == false)
893 /* must be using the same buffer */
894 if (req_op(rq) == REQ_OP_WRITE_SAME &&
895 !blk_write_same_mergeable(rq->bio, bio))
899 * Don't allow merge of different write hints, or for a hint with
902 if (rq->write_hint != bio->bi_write_hint)
908 enum elv_merge blk_try_merge(struct request *rq, struct bio *bio)
910 if (req_op(rq) == REQ_OP_DISCARD &&
911 queue_max_discard_segments(rq->q) > 1)
912 return ELEVATOR_DISCARD_MERGE;
913 else if (blk_rq_pos(rq) + blk_rq_sectors(rq) == bio->bi_iter.bi_sector)
914 return ELEVATOR_BACK_MERGE;
915 else if (blk_rq_pos(rq) - bio_sectors(bio) == bio->bi_iter.bi_sector)
916 return ELEVATOR_FRONT_MERGE;
917 return ELEVATOR_NO_MERGE;