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>
15 static inline bool bio_will_gap(struct request_queue *q,
16 struct request *prev_rq, struct bio *prev, struct bio *next)
18 struct bio_vec pb, nb;
20 if (!bio_has_data(prev) || !queue_virt_boundary(q))
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.
29 bio_get_first_bvec(prev_rq->bio, &pb);
31 bio_get_first_bvec(prev, &pb);
32 if (pb.bv_offset & queue_virt_boundary(q))
36 * We don't need to worry about the situation that the merged segment
37 * ends in unaligned virt boundary:
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
44 bio_get_last_bvec(prev, &pb);
45 bio_get_first_bvec(next, &nb);
46 if (biovec_phys_mergeable(q, &pb, &nb))
48 return __bvec_gap_to_prev(q, &pb, nb.bv_offset);
51 static inline bool req_gap_back_merge(struct request *req, struct bio *bio)
53 return bio_will_gap(req->q, req, req->biotail, bio);
56 static inline bool req_gap_front_merge(struct request *req, struct bio *bio)
58 return bio_will_gap(req->q, NULL, bio, req->bio);
61 static struct bio *blk_bio_discard_split(struct request_queue *q,
66 unsigned int max_discard_sectors, granularity;
69 unsigned split_sectors;
73 /* Zero-sector (unknown) and one-sector granularities are the same. */
74 granularity = max(q->limits.discard_granularity >> 9, 1U);
76 max_discard_sectors = min(q->limits.max_discard_sectors,
77 bio_allowed_max_sectors(q));
78 max_discard_sectors -= max_discard_sectors % granularity;
80 if (unlikely(!max_discard_sectors)) {
85 if (bio_sectors(bio) <= max_discard_sectors)
88 split_sectors = max_discard_sectors;
91 * If the next starting sector would be misaligned, stop the discard at
92 * the previous aligned sector.
94 alignment = (q->limits.discard_alignment >> 9) % granularity;
96 tmp = bio->bi_iter.bi_sector + split_sectors - alignment;
97 tmp = sector_div(tmp, granularity);
99 if (split_sectors > tmp)
100 split_sectors -= tmp;
102 return bio_split(bio, split_sectors, GFP_NOIO, bs);
105 static struct bio *blk_bio_write_zeroes_split(struct request_queue *q,
106 struct bio *bio, struct bio_set *bs, unsigned *nsegs)
110 if (!q->limits.max_write_zeroes_sectors)
113 if (bio_sectors(bio) <= q->limits.max_write_zeroes_sectors)
116 return bio_split(bio, q->limits.max_write_zeroes_sectors, GFP_NOIO, bs);
119 static struct bio *blk_bio_write_same_split(struct request_queue *q,
126 if (!q->limits.max_write_same_sectors)
129 if (bio_sectors(bio) <= q->limits.max_write_same_sectors)
132 return bio_split(bio, q->limits.max_write_same_sectors, GFP_NOIO, bs);
135 static inline unsigned get_max_io_size(struct request_queue *q,
138 unsigned sectors = blk_max_size_offset(q, bio->bi_iter.bi_sector);
139 unsigned mask = queue_logical_block_size(q) - 1;
141 /* aligned to logical block size */
142 sectors &= ~(mask >> 9);
147 static unsigned get_max_segment_size(struct request_queue *q,
150 unsigned long mask = queue_segment_boundary(q);
152 /* default segment boundary mask means no boundary limit */
153 if (mask == BLK_SEG_BOUNDARY_MASK)
154 return queue_max_segment_size(q);
156 return min_t(unsigned long, mask - (mask & offset) + 1,
157 queue_max_segment_size(q));
161 * Split the bvec @bv into segments, and update all kinds of
164 static bool bvec_split_segs(struct request_queue *q, struct bio_vec *bv,
165 unsigned *nsegs, unsigned *last_seg_size,
166 unsigned *front_seg_size, unsigned *sectors, unsigned max_segs)
168 unsigned len = bv->bv_len;
169 unsigned total_len = 0;
170 unsigned new_nsegs = 0, seg_size = 0;
173 * Multi-page bvec may be too big to hold in one segment, so the
174 * current bvec has to be splitted as multiple segments.
176 while (len && new_nsegs + *nsegs < max_segs) {
177 seg_size = get_max_segment_size(q, bv->bv_offset + total_len);
178 seg_size = min(seg_size, len);
181 total_len += seg_size;
184 if ((bv->bv_offset + total_len) & queue_virt_boundary(q))
191 /* update front segment size */
193 unsigned first_seg_size;
196 first_seg_size = get_max_segment_size(q, bv->bv_offset);
198 first_seg_size = queue_max_segment_size(q);
200 if (*front_seg_size < first_seg_size)
201 *front_seg_size = first_seg_size;
204 /* update other varibles */
205 *last_seg_size = seg_size;
208 *sectors += total_len >> 9;
210 /* split in the middle of the bvec if len != 0 */
214 static struct bio *blk_bio_segment_split(struct request_queue *q,
219 struct bio_vec bv, bvprv, *bvprvp = NULL;
220 struct bvec_iter iter;
221 unsigned seg_size = 0, nsegs = 0, sectors = 0;
222 unsigned front_seg_size = bio->bi_seg_front_size;
223 bool do_split = true;
224 struct bio *new = NULL;
225 const unsigned max_sectors = get_max_io_size(q, bio);
226 const unsigned max_segs = queue_max_segments(q);
228 bio_for_each_bvec(bv, bio, iter) {
230 * If the queue doesn't support SG gaps and adding this
231 * offset would create a gap, disallow it.
233 if (bvprvp && bvec_gap_to_prev(q, bvprvp, bv.bv_offset))
236 if (sectors + (bv.bv_len >> 9) > max_sectors) {
238 * Consider this a new segment if we're splitting in
239 * the middle of this vector.
241 if (nsegs < max_segs &&
242 sectors < max_sectors) {
243 /* split in the middle of bvec */
244 bv.bv_len = (max_sectors - sectors) << 9;
245 bvec_split_segs(q, &bv, &nsegs,
253 if (nsegs == max_segs)
259 if (bv.bv_offset + bv.bv_len <= PAGE_SIZE) {
261 seg_size = bv.bv_len;
262 sectors += bv.bv_len >> 9;
263 if (nsegs == 1 && seg_size > front_seg_size)
264 front_seg_size = seg_size;
265 } else if (bvec_split_segs(q, &bv, &nsegs, &seg_size,
266 &front_seg_size, §ors, max_segs)) {
276 new = bio_split(bio, sectors, GFP_NOIO, bs);
281 bio->bi_seg_front_size = front_seg_size;
282 if (seg_size > bio->bi_seg_back_size)
283 bio->bi_seg_back_size = seg_size;
285 return do_split ? new : NULL;
288 void blk_queue_split(struct request_queue *q, struct bio **bio)
290 struct bio *split, *res;
293 switch (bio_op(*bio)) {
295 case REQ_OP_SECURE_ERASE:
296 split = blk_bio_discard_split(q, *bio, &q->bio_split, &nsegs);
298 case REQ_OP_WRITE_ZEROES:
299 split = blk_bio_write_zeroes_split(q, *bio, &q->bio_split, &nsegs);
301 case REQ_OP_WRITE_SAME:
302 split = blk_bio_write_same_split(q, *bio, &q->bio_split, &nsegs);
305 split = blk_bio_segment_split(q, *bio, &q->bio_split, &nsegs);
309 /* physical segments can be figured out during splitting */
310 res = split ? split : *bio;
311 res->bi_phys_segments = nsegs;
312 bio_set_flag(res, BIO_SEG_VALID);
315 /* there isn't chance to merge the splitted bio */
316 split->bi_opf |= REQ_NOMERGE;
319 * Since we're recursing into make_request here, ensure
320 * that we mark this bio as already having entered the queue.
321 * If not, and the queue is going away, we can get stuck
322 * forever on waiting for the queue reference to drop. But
323 * that will never happen, as we're already holding a
326 bio_set_flag(*bio, BIO_QUEUE_ENTERED);
328 bio_chain(split, *bio);
329 trace_block_split(q, split, (*bio)->bi_iter.bi_sector);
330 generic_make_request(*bio);
334 EXPORT_SYMBOL(blk_queue_split);
336 static unsigned int __blk_recalc_rq_segments(struct request_queue *q,
339 struct bio_vec uninitialized_var(bv), bvprv = { NULL };
340 unsigned int seg_size, nr_phys_segs;
341 unsigned front_seg_size;
342 struct bio *fbio, *bbio;
343 struct bvec_iter iter;
348 front_seg_size = bio->bi_seg_front_size;
350 switch (bio_op(bio)) {
352 case REQ_OP_SECURE_ERASE:
353 case REQ_OP_WRITE_ZEROES:
355 case REQ_OP_WRITE_SAME:
363 bio_for_each_bvec(bv, bio, iter) {
364 bvec_split_segs(q, &bv, &nr_phys_segs, &seg_size,
365 &front_seg_size, NULL, UINT_MAX);
368 if (likely(bio->bi_iter.bi_size))
372 fbio->bi_seg_front_size = front_seg_size;
373 if (seg_size > bbio->bi_seg_back_size)
374 bbio->bi_seg_back_size = seg_size;
379 void blk_recalc_rq_segments(struct request *rq)
381 rq->nr_phys_segments = __blk_recalc_rq_segments(rq->q, rq->bio);
384 void blk_recount_segments(struct request_queue *q, struct bio *bio)
386 struct bio *nxt = bio->bi_next;
389 bio->bi_phys_segments = __blk_recalc_rq_segments(q, bio);
392 bio_set_flag(bio, BIO_SEG_VALID);
395 static int blk_phys_contig_segment(struct request_queue *q, struct bio *bio,
398 struct bio_vec end_bv = { NULL }, nxt_bv;
400 if (bio->bi_seg_back_size + nxt->bi_seg_front_size >
401 queue_max_segment_size(q))
404 if (!bio_has_data(bio))
407 bio_get_last_bvec(bio, &end_bv);
408 bio_get_first_bvec(nxt, &nxt_bv);
410 return biovec_phys_mergeable(q, &end_bv, &nxt_bv);
413 static inline struct scatterlist *blk_next_sg(struct scatterlist **sg,
414 struct scatterlist *sglist)
420 * If the driver previously mapped a shorter list, we could see a
421 * termination bit prematurely unless it fully inits the sg table
422 * on each mapping. We KNOW that there must be more entries here
423 * or the driver would be buggy, so force clear the termination bit
424 * to avoid doing a full sg_init_table() in drivers for each command.
430 static unsigned blk_bvec_map_sg(struct request_queue *q,
431 struct bio_vec *bvec, struct scatterlist *sglist,
432 struct scatterlist **sg)
434 unsigned nbytes = bvec->bv_len;
435 unsigned nsegs = 0, total = 0;
438 unsigned offset = bvec->bv_offset + total;
439 unsigned len = min(get_max_segment_size(q, offset), nbytes);
440 struct page *page = bvec->bv_page;
443 * Unfortunately a fair number of drivers barf on scatterlists
444 * that have an offset larger than PAGE_SIZE, despite other
445 * subsystems dealing with that invariant just fine. For now
446 * stick to the legacy format where we never present those from
447 * the block layer, but the code below should be removed once
448 * these offenders (mostly MMC/SD drivers) are fixed.
450 page += (offset >> PAGE_SHIFT);
451 offset &= ~PAGE_MASK;
453 *sg = blk_next_sg(sg, sglist);
454 sg_set_page(*sg, page, len, offset);
464 static inline int __blk_bvec_map_sg(struct bio_vec bv,
465 struct scatterlist *sglist, struct scatterlist **sg)
467 *sg = blk_next_sg(sg, sglist);
468 sg_set_page(*sg, bv.bv_page, bv.bv_len, bv.bv_offset);
472 /* only try to merge bvecs into one sg if they are from two bios */
474 __blk_segment_map_sg_merge(struct request_queue *q, struct bio_vec *bvec,
475 struct bio_vec *bvprv, struct scatterlist **sg)
478 int nbytes = bvec->bv_len;
483 if ((*sg)->length + nbytes > queue_max_segment_size(q))
486 if (!biovec_phys_mergeable(q, bvprv, bvec))
489 (*sg)->length += nbytes;
494 static int __blk_bios_map_sg(struct request_queue *q, struct bio *bio,
495 struct scatterlist *sglist,
496 struct scatterlist **sg)
498 struct bio_vec uninitialized_var(bvec), bvprv = { NULL };
499 struct bvec_iter iter;
501 bool new_bio = false;
504 bio_for_each_bvec(bvec, bio, iter) {
506 * Only try to merge bvecs from two bios given we
507 * have done bio internal merge when adding pages
511 __blk_segment_map_sg_merge(q, &bvec, &bvprv, sg))
514 if (bvec.bv_offset + bvec.bv_len <= PAGE_SIZE)
515 nsegs += __blk_bvec_map_sg(bvec, sglist, sg);
517 nsegs += blk_bvec_map_sg(q, &bvec, sglist, sg);
521 if (likely(bio->bi_iter.bi_size)) {
531 * map a request to scatterlist, return number of sg entries setup. Caller
532 * must make sure sg can hold rq->nr_phys_segments entries
534 int blk_rq_map_sg(struct request_queue *q, struct request *rq,
535 struct scatterlist *sglist)
537 struct scatterlist *sg = NULL;
540 if (rq->rq_flags & RQF_SPECIAL_PAYLOAD)
541 nsegs = __blk_bvec_map_sg(rq->special_vec, sglist, &sg);
542 else if (rq->bio && bio_op(rq->bio) == REQ_OP_WRITE_SAME)
543 nsegs = __blk_bvec_map_sg(bio_iovec(rq->bio), sglist, &sg);
545 nsegs = __blk_bios_map_sg(q, rq->bio, sglist, &sg);
547 if (unlikely(rq->rq_flags & RQF_COPY_USER) &&
548 (blk_rq_bytes(rq) & q->dma_pad_mask)) {
549 unsigned int pad_len =
550 (q->dma_pad_mask & ~blk_rq_bytes(rq)) + 1;
552 sg->length += pad_len;
553 rq->extra_len += pad_len;
556 if (q->dma_drain_size && q->dma_drain_needed(rq)) {
557 if (op_is_write(req_op(rq)))
558 memset(q->dma_drain_buffer, 0, q->dma_drain_size);
562 sg_set_page(sg, virt_to_page(q->dma_drain_buffer),
564 ((unsigned long)q->dma_drain_buffer) &
567 rq->extra_len += q->dma_drain_size;
574 * Something must have been wrong if the figured number of
575 * segment is bigger than number of req's physical segments
577 WARN_ON(nsegs > blk_rq_nr_phys_segments(rq));
581 EXPORT_SYMBOL(blk_rq_map_sg);
583 static inline int ll_new_hw_segment(struct request_queue *q,
587 int nr_phys_segs = bio_phys_segments(q, bio);
589 if (req->nr_phys_segments + nr_phys_segs > queue_max_segments(q))
592 if (blk_integrity_merge_bio(q, req, bio) == false)
596 * This will form the start of a new hw segment. Bump both
599 req->nr_phys_segments += nr_phys_segs;
603 req_set_nomerge(q, req);
607 int ll_back_merge_fn(struct request_queue *q, struct request *req,
610 if (req_gap_back_merge(req, bio))
612 if (blk_integrity_rq(req) &&
613 integrity_req_gap_back_merge(req, bio))
615 if (blk_rq_sectors(req) + bio_sectors(bio) >
616 blk_rq_get_max_sectors(req, blk_rq_pos(req))) {
617 req_set_nomerge(q, req);
620 if (!bio_flagged(req->biotail, BIO_SEG_VALID))
621 blk_recount_segments(q, req->biotail);
622 if (!bio_flagged(bio, BIO_SEG_VALID))
623 blk_recount_segments(q, bio);
625 return ll_new_hw_segment(q, req, bio);
628 int ll_front_merge_fn(struct request_queue *q, struct request *req,
632 if (req_gap_front_merge(req, bio))
634 if (blk_integrity_rq(req) &&
635 integrity_req_gap_front_merge(req, bio))
637 if (blk_rq_sectors(req) + bio_sectors(bio) >
638 blk_rq_get_max_sectors(req, bio->bi_iter.bi_sector)) {
639 req_set_nomerge(q, req);
642 if (!bio_flagged(bio, BIO_SEG_VALID))
643 blk_recount_segments(q, bio);
644 if (!bio_flagged(req->bio, BIO_SEG_VALID))
645 blk_recount_segments(q, req->bio);
647 return ll_new_hw_segment(q, req, bio);
650 static bool req_attempt_discard_merge(struct request_queue *q, struct request *req,
651 struct request *next)
653 unsigned short segments = blk_rq_nr_discard_segments(req);
655 if (segments >= queue_max_discard_segments(q))
657 if (blk_rq_sectors(req) + bio_sectors(next->bio) >
658 blk_rq_get_max_sectors(req, blk_rq_pos(req)))
661 req->nr_phys_segments = segments + blk_rq_nr_discard_segments(next);
664 req_set_nomerge(q, req);
668 static int ll_merge_requests_fn(struct request_queue *q, struct request *req,
669 struct request *next)
671 int total_phys_segments;
672 unsigned int seg_size =
673 req->biotail->bi_seg_back_size + next->bio->bi_seg_front_size;
675 if (req_gap_back_merge(req, next->bio))
679 * Will it become too large?
681 if ((blk_rq_sectors(req) + blk_rq_sectors(next)) >
682 blk_rq_get_max_sectors(req, blk_rq_pos(req)))
685 total_phys_segments = req->nr_phys_segments + next->nr_phys_segments;
686 if (blk_phys_contig_segment(q, req->biotail, next->bio)) {
687 if (req->nr_phys_segments == 1)
688 req->bio->bi_seg_front_size = seg_size;
689 if (next->nr_phys_segments == 1)
690 next->biotail->bi_seg_back_size = seg_size;
693 if (total_phys_segments > queue_max_segments(q))
696 if (blk_integrity_merge_rq(q, req, next) == false)
700 req->nr_phys_segments = total_phys_segments;
705 * blk_rq_set_mixed_merge - mark a request as mixed merge
706 * @rq: request to mark as mixed merge
709 * @rq is about to be mixed merged. Make sure the attributes
710 * which can be mixed are set in each bio and mark @rq as mixed
713 void blk_rq_set_mixed_merge(struct request *rq)
715 unsigned int ff = rq->cmd_flags & REQ_FAILFAST_MASK;
718 if (rq->rq_flags & RQF_MIXED_MERGE)
722 * @rq will no longer represent mixable attributes for all the
723 * contained bios. It will just track those of the first one.
724 * Distributes the attributs to each bio.
726 for (bio = rq->bio; bio; bio = bio->bi_next) {
727 WARN_ON_ONCE((bio->bi_opf & REQ_FAILFAST_MASK) &&
728 (bio->bi_opf & REQ_FAILFAST_MASK) != ff);
731 rq->rq_flags |= RQF_MIXED_MERGE;
734 static void blk_account_io_merge(struct request *req)
736 if (blk_do_io_stat(req)) {
737 struct hd_struct *part;
742 part_dec_in_flight(req->q, part, rq_data_dir(req));
749 * Two cases of handling DISCARD merge:
750 * If max_discard_segments > 1, the driver takes every bio
751 * as a range and send them to controller together. The ranges
752 * needn't to be contiguous.
753 * Otherwise, the bios/requests will be handled as same as
754 * others which should be contiguous.
756 static inline bool blk_discard_mergable(struct request *req)
758 if (req_op(req) == REQ_OP_DISCARD &&
759 queue_max_discard_segments(req->q) > 1)
764 static enum elv_merge blk_try_req_merge(struct request *req,
765 struct request *next)
767 if (blk_discard_mergable(req))
768 return ELEVATOR_DISCARD_MERGE;
769 else if (blk_rq_pos(req) + blk_rq_sectors(req) == blk_rq_pos(next))
770 return ELEVATOR_BACK_MERGE;
772 return ELEVATOR_NO_MERGE;
776 * For non-mq, this has to be called with the request spinlock acquired.
777 * For mq with scheduling, the appropriate queue wide lock should be held.
779 static struct request *attempt_merge(struct request_queue *q,
780 struct request *req, struct request *next)
782 if (!rq_mergeable(req) || !rq_mergeable(next))
785 if (req_op(req) != req_op(next))
788 if (rq_data_dir(req) != rq_data_dir(next)
789 || req->rq_disk != next->rq_disk)
792 if (req_op(req) == REQ_OP_WRITE_SAME &&
793 !blk_write_same_mergeable(req->bio, next->bio))
797 * Don't allow merge of different write hints, or for a hint with
800 if (req->write_hint != next->write_hint)
803 if (req->ioprio != next->ioprio)
807 * If we are allowed to merge, then append bio list
808 * from next to rq and release next. merge_requests_fn
809 * will have updated segment counts, update sector
810 * counts here. Handle DISCARDs separately, as they
811 * have separate settings.
814 switch (blk_try_req_merge(req, next)) {
815 case ELEVATOR_DISCARD_MERGE:
816 if (!req_attempt_discard_merge(q, req, next))
819 case ELEVATOR_BACK_MERGE:
820 if (!ll_merge_requests_fn(q, req, next))
828 * If failfast settings disagree or any of the two is already
829 * a mixed merge, mark both as mixed before proceeding. This
830 * makes sure that all involved bios have mixable attributes
833 if (((req->rq_flags | next->rq_flags) & RQF_MIXED_MERGE) ||
834 (req->cmd_flags & REQ_FAILFAST_MASK) !=
835 (next->cmd_flags & REQ_FAILFAST_MASK)) {
836 blk_rq_set_mixed_merge(req);
837 blk_rq_set_mixed_merge(next);
841 * At this point we have either done a back merge or front merge. We
842 * need the smaller start_time_ns of the merged requests to be the
843 * current request for accounting purposes.
845 if (next->start_time_ns < req->start_time_ns)
846 req->start_time_ns = next->start_time_ns;
848 req->biotail->bi_next = next->bio;
849 req->biotail = next->biotail;
851 req->__data_len += blk_rq_bytes(next);
853 if (!blk_discard_mergable(req))
854 elv_merge_requests(q, req, next);
857 * 'next' is going away, so update stats accordingly
859 blk_account_io_merge(next);
862 * ownership of bio passed from next to req, return 'next' for
869 struct request *attempt_back_merge(struct request_queue *q, struct request *rq)
871 struct request *next = elv_latter_request(q, rq);
874 return attempt_merge(q, rq, next);
879 struct request *attempt_front_merge(struct request_queue *q, struct request *rq)
881 struct request *prev = elv_former_request(q, rq);
884 return attempt_merge(q, prev, rq);
889 int blk_attempt_req_merge(struct request_queue *q, struct request *rq,
890 struct request *next)
892 struct request *free;
894 free = attempt_merge(q, rq, next);
896 blk_put_request(free);
903 bool blk_rq_merge_ok(struct request *rq, struct bio *bio)
905 if (!rq_mergeable(rq) || !bio_mergeable(bio))
908 if (req_op(rq) != bio_op(bio))
911 /* different data direction or already started, don't merge */
912 if (bio_data_dir(bio) != rq_data_dir(rq))
915 /* must be same device */
916 if (rq->rq_disk != bio->bi_disk)
919 /* only merge integrity protected bio into ditto rq */
920 if (blk_integrity_merge_bio(rq->q, rq, bio) == false)
923 /* must be using the same buffer */
924 if (req_op(rq) == REQ_OP_WRITE_SAME &&
925 !blk_write_same_mergeable(rq->bio, bio))
929 * Don't allow merge of different write hints, or for a hint with
932 if (rq->write_hint != bio->bi_write_hint)
935 if (rq->ioprio != bio_prio(bio))
941 enum elv_merge blk_try_merge(struct request *rq, struct bio *bio)
943 if (blk_discard_mergable(rq))
944 return ELEVATOR_DISCARD_MERGE;
945 else if (blk_rq_pos(rq) + blk_rq_sectors(rq) == bio->bi_iter.bi_sector)
946 return ELEVATOR_BACK_MERGE;
947 else if (blk_rq_pos(rq) - bio_sectors(bio) == bio->bi_iter.bi_sector)
948 return ELEVATOR_FRONT_MERGE;
949 return ELEVATOR_NO_MERGE;