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 unsigned get_max_segment_size(struct request_queue *q,
167 unsigned long mask = queue_segment_boundary(q);
169 /* default segment boundary mask means no boundary limit */
170 if (mask == BLK_SEG_BOUNDARY_MASK)
171 return queue_max_segment_size(q);
173 return min_t(unsigned long, mask - (mask & offset) + 1,
174 queue_max_segment_size(q));
178 * Split the bvec @bv into segments, and update all kinds of
181 static bool bvec_split_segs(struct request_queue *q, struct bio_vec *bv,
182 unsigned *nsegs, unsigned *last_seg_size,
183 unsigned *front_seg_size, unsigned *sectors)
185 unsigned len = bv->bv_len;
186 unsigned total_len = 0;
187 unsigned new_nsegs = 0, seg_size = 0;
190 * Multi-page bvec may be too big to hold in one segment, so the
191 * current bvec has to be splitted as multiple segments.
193 while (len && new_nsegs + *nsegs < queue_max_segments(q)) {
194 seg_size = get_max_segment_size(q, bv->bv_offset + total_len);
195 seg_size = min(seg_size, len);
198 total_len += seg_size;
201 if ((bv->bv_offset + total_len) & queue_virt_boundary(q))
208 /* update front segment size */
210 unsigned first_seg_size;
213 first_seg_size = get_max_segment_size(q, bv->bv_offset);
215 first_seg_size = queue_max_segment_size(q);
217 if (*front_seg_size < first_seg_size)
218 *front_seg_size = first_seg_size;
221 /* update other varibles */
222 *last_seg_size = seg_size;
225 *sectors += total_len >> 9;
227 /* split in the middle of the bvec if len != 0 */
231 static struct bio *blk_bio_segment_split(struct request_queue *q,
236 struct bio_vec bv, bvprv, *bvprvp = NULL;
237 struct bvec_iter iter;
238 unsigned seg_size = 0, nsegs = 0, sectors = 0;
239 unsigned front_seg_size = bio->bi_seg_front_size;
240 bool do_split = true;
241 struct bio *new = NULL;
242 const unsigned max_sectors = get_max_io_size(q, bio);
244 bio_for_each_bvec(bv, bio, iter) {
246 * If the queue doesn't support SG gaps and adding this
247 * offset would create a gap, disallow it.
249 if (bvprvp && bvec_gap_to_prev(q, bvprvp, bv.bv_offset))
252 if (sectors + (bv.bv_len >> 9) > max_sectors) {
254 * Consider this a new segment if we're splitting in
255 * the middle of this vector.
257 if (nsegs < queue_max_segments(q) &&
258 sectors < max_sectors) {
259 /* split in the middle of bvec */
260 bv.bv_len = (max_sectors - sectors) << 9;
261 bvec_split_segs(q, &bv, &nsegs,
270 if (seg_size + bv.bv_len > queue_max_segment_size(q))
272 if (!biovec_phys_mergeable(q, bvprvp, &bv))
275 seg_size += bv.bv_len;
278 sectors += bv.bv_len >> 9;
283 if (nsegs == queue_max_segments(q))
289 if (bvec_split_segs(q, &bv, &nsegs, &seg_size,
290 &front_seg_size, §ors))
300 new = bio_split(bio, sectors, GFP_NOIO, bs);
305 bio->bi_seg_front_size = front_seg_size;
306 if (seg_size > bio->bi_seg_back_size)
307 bio->bi_seg_back_size = seg_size;
309 return do_split ? new : NULL;
312 void blk_queue_split(struct request_queue *q, struct bio **bio)
314 struct bio *split, *res;
317 switch (bio_op(*bio)) {
319 case REQ_OP_SECURE_ERASE:
320 split = blk_bio_discard_split(q, *bio, &q->bio_split, &nsegs);
322 case REQ_OP_WRITE_ZEROES:
323 split = blk_bio_write_zeroes_split(q, *bio, &q->bio_split, &nsegs);
325 case REQ_OP_WRITE_SAME:
326 split = blk_bio_write_same_split(q, *bio, &q->bio_split, &nsegs);
329 split = blk_bio_segment_split(q, *bio, &q->bio_split, &nsegs);
333 /* physical segments can be figured out during splitting */
334 res = split ? split : *bio;
335 res->bi_phys_segments = nsegs;
336 bio_set_flag(res, BIO_SEG_VALID);
339 /* there isn't chance to merge the splitted bio */
340 split->bi_opf |= REQ_NOMERGE;
343 * Since we're recursing into make_request here, ensure
344 * that we mark this bio as already having entered the queue.
345 * If not, and the queue is going away, we can get stuck
346 * forever on waiting for the queue reference to drop. But
347 * that will never happen, as we're already holding a
350 bio_set_flag(*bio, BIO_QUEUE_ENTERED);
352 bio_chain(split, *bio);
353 trace_block_split(q, split, (*bio)->bi_iter.bi_sector);
354 generic_make_request(*bio);
358 EXPORT_SYMBOL(blk_queue_split);
360 static unsigned int __blk_recalc_rq_segments(struct request_queue *q,
364 struct bio_vec bv, bvprv = { NULL };
366 unsigned int seg_size, nr_phys_segs;
367 unsigned front_seg_size = bio->bi_seg_front_size;
368 struct bio *fbio, *bbio;
369 struct bvec_iter iter;
374 switch (bio_op(bio)) {
376 case REQ_OP_SECURE_ERASE:
377 case REQ_OP_WRITE_ZEROES:
379 case REQ_OP_WRITE_SAME:
387 bio_for_each_bvec(bv, bio, iter) {
389 * If SG merging is disabled, each bio vector is
396 if (seg_size + bv.bv_len
397 > queue_max_segment_size(q))
399 if (!biovec_phys_mergeable(q, &bvprv, &bv))
402 seg_size += bv.bv_len;
409 bvec_split_segs(q, &bv, &nr_phys_segs, &seg_size,
410 &front_seg_size, NULL);
415 fbio->bi_seg_front_size = front_seg_size;
416 if (seg_size > bbio->bi_seg_back_size)
417 bbio->bi_seg_back_size = seg_size;
422 void blk_recalc_rq_segments(struct request *rq)
424 bool no_sg_merge = !!test_bit(QUEUE_FLAG_NO_SG_MERGE,
425 &rq->q->queue_flags);
427 rq->nr_phys_segments = __blk_recalc_rq_segments(rq->q, rq->bio,
431 void blk_recount_segments(struct request_queue *q, struct bio *bio)
433 unsigned short seg_cnt = bio_segments(bio);
435 if (test_bit(QUEUE_FLAG_NO_SG_MERGE, &q->queue_flags) &&
436 (seg_cnt < queue_max_segments(q)))
437 bio->bi_phys_segments = seg_cnt;
439 struct bio *nxt = bio->bi_next;
442 bio->bi_phys_segments = __blk_recalc_rq_segments(q, bio, false);
446 bio_set_flag(bio, BIO_SEG_VALID);
449 static int blk_phys_contig_segment(struct request_queue *q, struct bio *bio,
452 struct bio_vec end_bv = { NULL }, nxt_bv;
454 if (bio->bi_seg_back_size + nxt->bi_seg_front_size >
455 queue_max_segment_size(q))
458 if (!bio_has_data(bio))
461 bio_get_last_bvec(bio, &end_bv);
462 bio_get_first_bvec(nxt, &nxt_bv);
464 return biovec_phys_mergeable(q, &end_bv, &nxt_bv);
468 __blk_segment_map_sg(struct request_queue *q, struct bio_vec *bvec,
469 struct scatterlist *sglist, struct bio_vec *bvprv,
470 struct scatterlist **sg, int *nsegs)
473 int nbytes = bvec->bv_len;
476 if ((*sg)->length + nbytes > queue_max_segment_size(q))
478 if (!biovec_phys_mergeable(q, bvprv, bvec))
481 (*sg)->length += nbytes;
488 * If the driver previously mapped a shorter
489 * list, we could see a termination bit
490 * prematurely unless it fully inits the sg
491 * table on each mapping. We KNOW that there
492 * must be more entries here or the driver
493 * would be buggy, so force clear the
494 * termination bit to avoid doing a full
495 * sg_init_table() in drivers for each command.
501 sg_set_page(*sg, bvec->bv_page, nbytes, bvec->bv_offset);
507 static inline int __blk_bvec_map_sg(struct request_queue *q, struct bio_vec bv,
508 struct scatterlist *sglist, struct scatterlist **sg)
511 sg_set_page(*sg, bv.bv_page, bv.bv_len, bv.bv_offset);
515 static int __blk_bios_map_sg(struct request_queue *q, struct bio *bio,
516 struct scatterlist *sglist,
517 struct scatterlist **sg)
519 struct bio_vec bvec, bvprv = { NULL };
520 struct bvec_iter iter;
524 bio_for_each_segment(bvec, bio, iter)
525 __blk_segment_map_sg(q, &bvec, sglist, &bvprv, sg,
532 * map a request to scatterlist, return number of sg entries setup. Caller
533 * must make sure sg can hold rq->nr_phys_segments entries
535 int blk_rq_map_sg(struct request_queue *q, struct request *rq,
536 struct scatterlist *sglist)
538 struct scatterlist *sg = NULL;
541 if (rq->rq_flags & RQF_SPECIAL_PAYLOAD)
542 nsegs = __blk_bvec_map_sg(q, rq->special_vec, sglist, &sg);
543 else if (rq->bio && bio_op(rq->bio) == REQ_OP_WRITE_SAME)
544 nsegs = __blk_bvec_map_sg(q, bio_iovec(rq->bio), sglist, &sg);
546 nsegs = __blk_bios_map_sg(q, rq->bio, sglist, &sg);
548 if (unlikely(rq->rq_flags & RQF_COPY_USER) &&
549 (blk_rq_bytes(rq) & q->dma_pad_mask)) {
550 unsigned int pad_len =
551 (q->dma_pad_mask & ~blk_rq_bytes(rq)) + 1;
553 sg->length += pad_len;
554 rq->extra_len += pad_len;
557 if (q->dma_drain_size && q->dma_drain_needed(rq)) {
558 if (op_is_write(req_op(rq)))
559 memset(q->dma_drain_buffer, 0, q->dma_drain_size);
563 sg_set_page(sg, virt_to_page(q->dma_drain_buffer),
565 ((unsigned long)q->dma_drain_buffer) &
568 rq->extra_len += q->dma_drain_size;
575 * Something must have been wrong if the figured number of
576 * segment is bigger than number of req's physical segments
578 WARN_ON(nsegs > blk_rq_nr_phys_segments(rq));
582 EXPORT_SYMBOL(blk_rq_map_sg);
584 static inline int ll_new_hw_segment(struct request_queue *q,
588 int nr_phys_segs = bio_phys_segments(q, bio);
590 if (req->nr_phys_segments + nr_phys_segs > queue_max_segments(q))
593 if (blk_integrity_merge_bio(q, req, bio) == false)
597 * This will form the start of a new hw segment. Bump both
600 req->nr_phys_segments += nr_phys_segs;
604 req_set_nomerge(q, req);
608 int ll_back_merge_fn(struct request_queue *q, struct request *req,
611 if (req_gap_back_merge(req, bio))
613 if (blk_integrity_rq(req) &&
614 integrity_req_gap_back_merge(req, bio))
616 if (blk_rq_sectors(req) + bio_sectors(bio) >
617 blk_rq_get_max_sectors(req, blk_rq_pos(req))) {
618 req_set_nomerge(q, req);
621 if (!bio_flagged(req->biotail, BIO_SEG_VALID))
622 blk_recount_segments(q, req->biotail);
623 if (!bio_flagged(bio, BIO_SEG_VALID))
624 blk_recount_segments(q, bio);
626 return ll_new_hw_segment(q, req, bio);
629 int ll_front_merge_fn(struct request_queue *q, struct request *req,
633 if (req_gap_front_merge(req, bio))
635 if (blk_integrity_rq(req) &&
636 integrity_req_gap_front_merge(req, bio))
638 if (blk_rq_sectors(req) + bio_sectors(bio) >
639 blk_rq_get_max_sectors(req, bio->bi_iter.bi_sector)) {
640 req_set_nomerge(q, req);
643 if (!bio_flagged(bio, BIO_SEG_VALID))
644 blk_recount_segments(q, bio);
645 if (!bio_flagged(req->bio, BIO_SEG_VALID))
646 blk_recount_segments(q, req->bio);
648 return ll_new_hw_segment(q, req, bio);
651 static bool req_attempt_discard_merge(struct request_queue *q, struct request *req,
652 struct request *next)
654 unsigned short segments = blk_rq_nr_discard_segments(req);
656 if (segments >= queue_max_discard_segments(q))
658 if (blk_rq_sectors(req) + bio_sectors(next->bio) >
659 blk_rq_get_max_sectors(req, blk_rq_pos(req)))
662 req->nr_phys_segments = segments + blk_rq_nr_discard_segments(next);
665 req_set_nomerge(q, req);
669 static int ll_merge_requests_fn(struct request_queue *q, struct request *req,
670 struct request *next)
672 int total_phys_segments;
673 unsigned int seg_size =
674 req->biotail->bi_seg_back_size + next->bio->bi_seg_front_size;
676 if (req_gap_back_merge(req, next->bio))
680 * Will it become too large?
682 if ((blk_rq_sectors(req) + blk_rq_sectors(next)) >
683 blk_rq_get_max_sectors(req, blk_rq_pos(req)))
686 total_phys_segments = req->nr_phys_segments + next->nr_phys_segments;
687 if (blk_phys_contig_segment(q, req->biotail, next->bio)) {
688 if (req->nr_phys_segments == 1)
689 req->bio->bi_seg_front_size = seg_size;
690 if (next->nr_phys_segments == 1)
691 next->biotail->bi_seg_back_size = seg_size;
692 total_phys_segments--;
695 if (total_phys_segments > queue_max_segments(q))
698 if (blk_integrity_merge_rq(q, req, next) == false)
702 req->nr_phys_segments = total_phys_segments;
707 * blk_rq_set_mixed_merge - mark a request as mixed merge
708 * @rq: request to mark as mixed merge
711 * @rq is about to be mixed merged. Make sure the attributes
712 * which can be mixed are set in each bio and mark @rq as mixed
715 void blk_rq_set_mixed_merge(struct request *rq)
717 unsigned int ff = rq->cmd_flags & REQ_FAILFAST_MASK;
720 if (rq->rq_flags & RQF_MIXED_MERGE)
724 * @rq will no longer represent mixable attributes for all the
725 * contained bios. It will just track those of the first one.
726 * Distributes the attributs to each bio.
728 for (bio = rq->bio; bio; bio = bio->bi_next) {
729 WARN_ON_ONCE((bio->bi_opf & REQ_FAILFAST_MASK) &&
730 (bio->bi_opf & REQ_FAILFAST_MASK) != ff);
733 rq->rq_flags |= RQF_MIXED_MERGE;
736 static void blk_account_io_merge(struct request *req)
738 if (blk_do_io_stat(req)) {
739 struct hd_struct *part;
744 part_dec_in_flight(req->q, part, rq_data_dir(req));
751 * Two cases of handling DISCARD merge:
752 * If max_discard_segments > 1, the driver takes every bio
753 * as a range and send them to controller together. The ranges
754 * needn't to be contiguous.
755 * Otherwise, the bios/requests will be handled as same as
756 * others which should be contiguous.
758 static inline bool blk_discard_mergable(struct request *req)
760 if (req_op(req) == REQ_OP_DISCARD &&
761 queue_max_discard_segments(req->q) > 1)
766 static enum elv_merge blk_try_req_merge(struct request *req,
767 struct request *next)
769 if (blk_discard_mergable(req))
770 return ELEVATOR_DISCARD_MERGE;
771 else if (blk_rq_pos(req) + blk_rq_sectors(req) == blk_rq_pos(next))
772 return ELEVATOR_BACK_MERGE;
774 return ELEVATOR_NO_MERGE;
778 * For non-mq, this has to be called with the request spinlock acquired.
779 * For mq with scheduling, the appropriate queue wide lock should be held.
781 static struct request *attempt_merge(struct request_queue *q,
782 struct request *req, struct request *next)
784 if (!rq_mergeable(req) || !rq_mergeable(next))
787 if (req_op(req) != req_op(next))
790 if (rq_data_dir(req) != rq_data_dir(next)
791 || req->rq_disk != next->rq_disk)
794 if (req_op(req) == REQ_OP_WRITE_SAME &&
795 !blk_write_same_mergeable(req->bio, next->bio))
799 * Don't allow merge of different write hints, or for a hint with
802 if (req->write_hint != next->write_hint)
805 if (req->ioprio != next->ioprio)
809 * If we are allowed to merge, then append bio list
810 * from next to rq and release next. merge_requests_fn
811 * will have updated segment counts, update sector
812 * counts here. Handle DISCARDs separately, as they
813 * have separate settings.
816 switch (blk_try_req_merge(req, next)) {
817 case ELEVATOR_DISCARD_MERGE:
818 if (!req_attempt_discard_merge(q, req, next))
821 case ELEVATOR_BACK_MERGE:
822 if (!ll_merge_requests_fn(q, req, next))
830 * If failfast settings disagree or any of the two is already
831 * a mixed merge, mark both as mixed before proceeding. This
832 * makes sure that all involved bios have mixable attributes
835 if (((req->rq_flags | next->rq_flags) & RQF_MIXED_MERGE) ||
836 (req->cmd_flags & REQ_FAILFAST_MASK) !=
837 (next->cmd_flags & REQ_FAILFAST_MASK)) {
838 blk_rq_set_mixed_merge(req);
839 blk_rq_set_mixed_merge(next);
843 * At this point we have either done a back merge or front merge. We
844 * need the smaller start_time_ns of the merged requests to be the
845 * current request for accounting purposes.
847 if (next->start_time_ns < req->start_time_ns)
848 req->start_time_ns = next->start_time_ns;
850 req->biotail->bi_next = next->bio;
851 req->biotail = next->biotail;
853 req->__data_len += blk_rq_bytes(next);
855 if (!blk_discard_mergable(req))
856 elv_merge_requests(q, req, next);
859 * 'next' is going away, so update stats accordingly
861 blk_account_io_merge(next);
864 * ownership of bio passed from next to req, return 'next' for
871 struct request *attempt_back_merge(struct request_queue *q, struct request *rq)
873 struct request *next = elv_latter_request(q, rq);
876 return attempt_merge(q, rq, next);
881 struct request *attempt_front_merge(struct request_queue *q, struct request *rq)
883 struct request *prev = elv_former_request(q, rq);
886 return attempt_merge(q, prev, rq);
891 int blk_attempt_req_merge(struct request_queue *q, struct request *rq,
892 struct request *next)
894 struct request *free;
896 free = attempt_merge(q, rq, next);
898 blk_put_request(free);
905 bool blk_rq_merge_ok(struct request *rq, struct bio *bio)
907 if (!rq_mergeable(rq) || !bio_mergeable(bio))
910 if (req_op(rq) != bio_op(bio))
913 /* different data direction or already started, don't merge */
914 if (bio_data_dir(bio) != rq_data_dir(rq))
917 /* must be same device */
918 if (rq->rq_disk != bio->bi_disk)
921 /* only merge integrity protected bio into ditto rq */
922 if (blk_integrity_merge_bio(rq->q, rq, bio) == false)
925 /* must be using the same buffer */
926 if (req_op(rq) == REQ_OP_WRITE_SAME &&
927 !blk_write_same_mergeable(rq->bio, bio))
931 * Don't allow merge of different write hints, or for a hint with
934 if (rq->write_hint != bio->bi_write_hint)
937 if (rq->ioprio != bio_prio(bio))
943 enum elv_merge blk_try_merge(struct request *rq, struct bio *bio)
945 if (blk_discard_mergable(rq))
946 return ELEVATOR_DISCARD_MERGE;
947 else if (blk_rq_pos(rq) + blk_rq_sectors(rq) == bio->bi_iter.bi_sector)
948 return ELEVATOR_BACK_MERGE;
949 else if (blk_rq_pos(rq) - bio_sectors(bio) == bio->bi_iter.bi_sector)
950 return ELEVATOR_FRONT_MERGE;
951 return ELEVATOR_NO_MERGE;