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(const 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(const struct request_queue *q,
165 const struct bio_vec *bv, unsigned *nsegs,
166 unsigned *sectors, unsigned max_segs)
168 unsigned len = bv->bv_len;
169 unsigned total_len = 0;
170 unsigned 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 && *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))
188 *sectors += total_len >> 9;
190 /* split in the middle of the bvec if len != 0 */
195 * blk_bio_segment_split - split a bio in two bios
196 * @q: [in] request queue pointer
197 * @bio: [in] bio to be split
198 * @bs: [in] bio set to allocate the clone from
199 * @segs: [out] number of segments in the bio with the first half of the sectors
201 * Clone @bio, update the bi_iter of the clone to represent the first sectors
202 * of @bio and update @bio->bi_iter to represent the remaining sectors. The
203 * following is guaranteed for the cloned bio:
204 * - That it has at most get_max_io_size(@q, @bio) sectors.
205 * - That it has at most queue_max_segments(@q) segments.
207 * Except for discard requests the cloned bio will point at the bi_io_vec of
208 * the original bio. It is the responsibility of the caller to ensure that the
209 * original bio is not freed before the cloned bio. The caller is also
210 * responsible for ensuring that @bs is only destroyed after processing of the
211 * split bio has finished.
213 static struct bio *blk_bio_segment_split(struct request_queue *q,
218 struct bio_vec bv, bvprv, *bvprvp = NULL;
219 struct bvec_iter iter;
220 unsigned nsegs = 0, sectors = 0;
221 const unsigned max_sectors = get_max_io_size(q, bio);
222 const unsigned max_segs = queue_max_segments(q);
224 bio_for_each_bvec(bv, bio, iter) {
226 * If the queue doesn't support SG gaps and adding this
227 * offset would create a gap, disallow it.
229 if (bvprvp && bvec_gap_to_prev(q, bvprvp, bv.bv_offset))
232 if (sectors + (bv.bv_len >> 9) > max_sectors) {
234 * Consider this a new segment if we're splitting in
235 * the middle of this vector.
237 if (nsegs < max_segs &&
238 sectors < max_sectors) {
239 /* split in the middle of bvec */
240 bv.bv_len = (max_sectors - sectors) << 9;
241 bvec_split_segs(q, &bv, &nsegs,
247 if (nsegs == max_segs)
253 if (bv.bv_offset + bv.bv_len <= PAGE_SIZE) {
255 sectors += bv.bv_len >> 9;
256 } else if (bvec_split_segs(q, &bv, &nsegs, §ors,
266 return bio_split(bio, sectors, GFP_NOIO, bs);
270 * __blk_queue_split - split a bio and submit the second half
271 * @q: [in] request queue pointer
272 * @bio: [in, out] bio to be split
273 * @nr_segs: [out] number of segments in the first bio
275 * Split a bio into two bios, chain the two bios, submit the second half and
276 * store a pointer to the first half in *@bio. If the second bio is still too
277 * big it will be split by a recursive call to this function. Since this
278 * function may allocate a new bio from @q->bio_split, it is the responsibility
279 * of the caller to ensure that @q is only released after processing of the
280 * split bio has finished.
282 void __blk_queue_split(struct request_queue *q, struct bio **bio,
283 unsigned int *nr_segs)
287 switch (bio_op(*bio)) {
289 case REQ_OP_SECURE_ERASE:
290 split = blk_bio_discard_split(q, *bio, &q->bio_split, nr_segs);
292 case REQ_OP_WRITE_ZEROES:
293 split = blk_bio_write_zeroes_split(q, *bio, &q->bio_split,
296 case REQ_OP_WRITE_SAME:
297 split = blk_bio_write_same_split(q, *bio, &q->bio_split,
301 split = blk_bio_segment_split(q, *bio, &q->bio_split, nr_segs);
306 /* there isn't chance to merge the splitted bio */
307 split->bi_opf |= REQ_NOMERGE;
310 * Since we're recursing into make_request here, ensure
311 * that we mark this bio as already having entered the queue.
312 * If not, and the queue is going away, we can get stuck
313 * forever on waiting for the queue reference to drop. But
314 * that will never happen, as we're already holding a
317 bio_set_flag(*bio, BIO_QUEUE_ENTERED);
319 bio_chain(split, *bio);
320 trace_block_split(q, split, (*bio)->bi_iter.bi_sector);
321 generic_make_request(*bio);
327 * blk_queue_split - split a bio and submit the second half
328 * @q: [in] request queue pointer
329 * @bio: [in, out] bio to be split
331 * Split a bio into two bios, chains the two bios, submit the second half and
332 * store a pointer to the first half in *@bio. Since this function may allocate
333 * a new bio from @q->bio_split, it is the responsibility of the caller to
334 * ensure that @q is only released after processing of the split bio has
337 void blk_queue_split(struct request_queue *q, struct bio **bio)
339 unsigned int nr_segs;
341 __blk_queue_split(q, bio, &nr_segs);
343 EXPORT_SYMBOL(blk_queue_split);
345 unsigned int blk_recalc_rq_segments(struct request *rq)
347 unsigned int nr_phys_segs = 0;
348 unsigned int nr_sectors = 0;
349 struct req_iterator iter;
355 switch (bio_op(rq->bio)) {
357 case REQ_OP_SECURE_ERASE:
358 case REQ_OP_WRITE_ZEROES:
360 case REQ_OP_WRITE_SAME:
364 rq_for_each_bvec(bv, rq, iter)
365 bvec_split_segs(rq->q, &bv, &nr_phys_segs, &nr_sectors,
370 static inline struct scatterlist *blk_next_sg(struct scatterlist **sg,
371 struct scatterlist *sglist)
377 * If the driver previously mapped a shorter list, we could see a
378 * termination bit prematurely unless it fully inits the sg table
379 * on each mapping. We KNOW that there must be more entries here
380 * or the driver would be buggy, so force clear the termination bit
381 * to avoid doing a full sg_init_table() in drivers for each command.
387 static unsigned blk_bvec_map_sg(struct request_queue *q,
388 struct bio_vec *bvec, struct scatterlist *sglist,
389 struct scatterlist **sg)
391 unsigned nbytes = bvec->bv_len;
392 unsigned nsegs = 0, total = 0;
395 unsigned offset = bvec->bv_offset + total;
396 unsigned len = min(get_max_segment_size(q, offset), nbytes);
397 struct page *page = bvec->bv_page;
400 * Unfortunately a fair number of drivers barf on scatterlists
401 * that have an offset larger than PAGE_SIZE, despite other
402 * subsystems dealing with that invariant just fine. For now
403 * stick to the legacy format where we never present those from
404 * the block layer, but the code below should be removed once
405 * these offenders (mostly MMC/SD drivers) are fixed.
407 page += (offset >> PAGE_SHIFT);
408 offset &= ~PAGE_MASK;
410 *sg = blk_next_sg(sg, sglist);
411 sg_set_page(*sg, page, len, offset);
421 static inline int __blk_bvec_map_sg(struct bio_vec bv,
422 struct scatterlist *sglist, struct scatterlist **sg)
424 *sg = blk_next_sg(sg, sglist);
425 sg_set_page(*sg, bv.bv_page, bv.bv_len, bv.bv_offset);
429 /* only try to merge bvecs into one sg if they are from two bios */
431 __blk_segment_map_sg_merge(struct request_queue *q, struct bio_vec *bvec,
432 struct bio_vec *bvprv, struct scatterlist **sg)
435 int nbytes = bvec->bv_len;
440 if ((*sg)->length + nbytes > queue_max_segment_size(q))
443 if (!biovec_phys_mergeable(q, bvprv, bvec))
446 (*sg)->length += nbytes;
451 static int __blk_bios_map_sg(struct request_queue *q, struct bio *bio,
452 struct scatterlist *sglist,
453 struct scatterlist **sg)
455 struct bio_vec uninitialized_var(bvec), bvprv = { NULL };
456 struct bvec_iter iter;
458 bool new_bio = false;
461 bio_for_each_bvec(bvec, bio, iter) {
463 * Only try to merge bvecs from two bios given we
464 * have done bio internal merge when adding pages
468 __blk_segment_map_sg_merge(q, &bvec, &bvprv, sg))
471 if (bvec.bv_offset + bvec.bv_len <= PAGE_SIZE)
472 nsegs += __blk_bvec_map_sg(bvec, sglist, sg);
474 nsegs += blk_bvec_map_sg(q, &bvec, sglist, sg);
478 if (likely(bio->bi_iter.bi_size)) {
488 * map a request to scatterlist, return number of sg entries setup. Caller
489 * must make sure sg can hold rq->nr_phys_segments entries
491 int blk_rq_map_sg(struct request_queue *q, struct request *rq,
492 struct scatterlist *sglist)
494 struct scatterlist *sg = NULL;
497 if (rq->rq_flags & RQF_SPECIAL_PAYLOAD)
498 nsegs = __blk_bvec_map_sg(rq->special_vec, sglist, &sg);
499 else if (rq->bio && bio_op(rq->bio) == REQ_OP_WRITE_SAME)
500 nsegs = __blk_bvec_map_sg(bio_iovec(rq->bio), sglist, &sg);
502 nsegs = __blk_bios_map_sg(q, rq->bio, sglist, &sg);
504 if (unlikely(rq->rq_flags & RQF_COPY_USER) &&
505 (blk_rq_bytes(rq) & q->dma_pad_mask)) {
506 unsigned int pad_len =
507 (q->dma_pad_mask & ~blk_rq_bytes(rq)) + 1;
509 sg->length += pad_len;
510 rq->extra_len += pad_len;
513 if (q->dma_drain_size && q->dma_drain_needed(rq)) {
514 if (op_is_write(req_op(rq)))
515 memset(q->dma_drain_buffer, 0, q->dma_drain_size);
519 sg_set_page(sg, virt_to_page(q->dma_drain_buffer),
521 ((unsigned long)q->dma_drain_buffer) &
524 rq->extra_len += q->dma_drain_size;
531 * Something must have been wrong if the figured number of
532 * segment is bigger than number of req's physical segments
534 WARN_ON(nsegs > blk_rq_nr_phys_segments(rq));
538 EXPORT_SYMBOL(blk_rq_map_sg);
540 static inline int ll_new_hw_segment(struct request *req, struct bio *bio,
541 unsigned int nr_phys_segs)
543 if (req->nr_phys_segments + nr_phys_segs > queue_max_segments(req->q))
546 if (blk_integrity_merge_bio(req->q, req, bio) == false)
550 * This will form the start of a new hw segment. Bump both
553 req->nr_phys_segments += nr_phys_segs;
557 req_set_nomerge(req->q, req);
561 int ll_back_merge_fn(struct request *req, struct bio *bio, unsigned int nr_segs)
563 if (req_gap_back_merge(req, bio))
565 if (blk_integrity_rq(req) &&
566 integrity_req_gap_back_merge(req, bio))
568 if (blk_rq_sectors(req) + bio_sectors(bio) >
569 blk_rq_get_max_sectors(req, blk_rq_pos(req))) {
570 req_set_nomerge(req->q, req);
574 return ll_new_hw_segment(req, bio, nr_segs);
577 int ll_front_merge_fn(struct request *req, struct bio *bio, unsigned int nr_segs)
579 if (req_gap_front_merge(req, bio))
581 if (blk_integrity_rq(req) &&
582 integrity_req_gap_front_merge(req, bio))
584 if (blk_rq_sectors(req) + bio_sectors(bio) >
585 blk_rq_get_max_sectors(req, bio->bi_iter.bi_sector)) {
586 req_set_nomerge(req->q, req);
590 return ll_new_hw_segment(req, bio, nr_segs);
593 static bool req_attempt_discard_merge(struct request_queue *q, struct request *req,
594 struct request *next)
596 unsigned short segments = blk_rq_nr_discard_segments(req);
598 if (segments >= queue_max_discard_segments(q))
600 if (blk_rq_sectors(req) + bio_sectors(next->bio) >
601 blk_rq_get_max_sectors(req, blk_rq_pos(req)))
604 req->nr_phys_segments = segments + blk_rq_nr_discard_segments(next);
607 req_set_nomerge(q, req);
611 static int ll_merge_requests_fn(struct request_queue *q, struct request *req,
612 struct request *next)
614 int total_phys_segments;
616 if (req_gap_back_merge(req, next->bio))
620 * Will it become too large?
622 if ((blk_rq_sectors(req) + blk_rq_sectors(next)) >
623 blk_rq_get_max_sectors(req, blk_rq_pos(req)))
626 total_phys_segments = req->nr_phys_segments + next->nr_phys_segments;
627 if (total_phys_segments > queue_max_segments(q))
630 if (blk_integrity_merge_rq(q, req, next) == false)
634 req->nr_phys_segments = total_phys_segments;
639 * blk_rq_set_mixed_merge - mark a request as mixed merge
640 * @rq: request to mark as mixed merge
643 * @rq is about to be mixed merged. Make sure the attributes
644 * which can be mixed are set in each bio and mark @rq as mixed
647 void blk_rq_set_mixed_merge(struct request *rq)
649 unsigned int ff = rq->cmd_flags & REQ_FAILFAST_MASK;
652 if (rq->rq_flags & RQF_MIXED_MERGE)
656 * @rq will no longer represent mixable attributes for all the
657 * contained bios. It will just track those of the first one.
658 * Distributes the attributs to each bio.
660 for (bio = rq->bio; bio; bio = bio->bi_next) {
661 WARN_ON_ONCE((bio->bi_opf & REQ_FAILFAST_MASK) &&
662 (bio->bi_opf & REQ_FAILFAST_MASK) != ff);
665 rq->rq_flags |= RQF_MIXED_MERGE;
668 static void blk_account_io_merge(struct request *req)
670 if (blk_do_io_stat(req)) {
671 struct hd_struct *part;
676 part_dec_in_flight(req->q, part, rq_data_dir(req));
683 * Two cases of handling DISCARD merge:
684 * If max_discard_segments > 1, the driver takes every bio
685 * as a range and send them to controller together. The ranges
686 * needn't to be contiguous.
687 * Otherwise, the bios/requests will be handled as same as
688 * others which should be contiguous.
690 static inline bool blk_discard_mergable(struct request *req)
692 if (req_op(req) == REQ_OP_DISCARD &&
693 queue_max_discard_segments(req->q) > 1)
698 static enum elv_merge blk_try_req_merge(struct request *req,
699 struct request *next)
701 if (blk_discard_mergable(req))
702 return ELEVATOR_DISCARD_MERGE;
703 else if (blk_rq_pos(req) + blk_rq_sectors(req) == blk_rq_pos(next))
704 return ELEVATOR_BACK_MERGE;
706 return ELEVATOR_NO_MERGE;
710 * For non-mq, this has to be called with the request spinlock acquired.
711 * For mq with scheduling, the appropriate queue wide lock should be held.
713 static struct request *attempt_merge(struct request_queue *q,
714 struct request *req, struct request *next)
716 if (!rq_mergeable(req) || !rq_mergeable(next))
719 if (req_op(req) != req_op(next))
722 if (rq_data_dir(req) != rq_data_dir(next)
723 || req->rq_disk != next->rq_disk)
726 if (req_op(req) == REQ_OP_WRITE_SAME &&
727 !blk_write_same_mergeable(req->bio, next->bio))
731 * Don't allow merge of different write hints, or for a hint with
734 if (req->write_hint != next->write_hint)
737 if (req->ioprio != next->ioprio)
741 * If we are allowed to merge, then append bio list
742 * from next to rq and release next. merge_requests_fn
743 * will have updated segment counts, update sector
744 * counts here. Handle DISCARDs separately, as they
745 * have separate settings.
748 switch (blk_try_req_merge(req, next)) {
749 case ELEVATOR_DISCARD_MERGE:
750 if (!req_attempt_discard_merge(q, req, next))
753 case ELEVATOR_BACK_MERGE:
754 if (!ll_merge_requests_fn(q, req, next))
762 * If failfast settings disagree or any of the two is already
763 * a mixed merge, mark both as mixed before proceeding. This
764 * makes sure that all involved bios have mixable attributes
767 if (((req->rq_flags | next->rq_flags) & RQF_MIXED_MERGE) ||
768 (req->cmd_flags & REQ_FAILFAST_MASK) !=
769 (next->cmd_flags & REQ_FAILFAST_MASK)) {
770 blk_rq_set_mixed_merge(req);
771 blk_rq_set_mixed_merge(next);
775 * At this point we have either done a back merge or front merge. We
776 * need the smaller start_time_ns of the merged requests to be the
777 * current request for accounting purposes.
779 if (next->start_time_ns < req->start_time_ns)
780 req->start_time_ns = next->start_time_ns;
782 req->biotail->bi_next = next->bio;
783 req->biotail = next->biotail;
785 req->__data_len += blk_rq_bytes(next);
787 if (!blk_discard_mergable(req))
788 elv_merge_requests(q, req, next);
791 * 'next' is going away, so update stats accordingly
793 blk_account_io_merge(next);
796 * ownership of bio passed from next to req, return 'next' for
803 struct request *attempt_back_merge(struct request_queue *q, struct request *rq)
805 struct request *next = elv_latter_request(q, rq);
808 return attempt_merge(q, rq, next);
813 struct request *attempt_front_merge(struct request_queue *q, struct request *rq)
815 struct request *prev = elv_former_request(q, rq);
818 return attempt_merge(q, prev, rq);
823 int blk_attempt_req_merge(struct request_queue *q, struct request *rq,
824 struct request *next)
826 struct request *free;
828 free = attempt_merge(q, rq, next);
830 blk_put_request(free);
837 bool blk_rq_merge_ok(struct request *rq, struct bio *bio)
839 if (!rq_mergeable(rq) || !bio_mergeable(bio))
842 if (req_op(rq) != bio_op(bio))
845 /* different data direction or already started, don't merge */
846 if (bio_data_dir(bio) != rq_data_dir(rq))
849 /* must be same device */
850 if (rq->rq_disk != bio->bi_disk)
853 /* only merge integrity protected bio into ditto rq */
854 if (blk_integrity_merge_bio(rq->q, rq, bio) == false)
857 /* must be using the same buffer */
858 if (req_op(rq) == REQ_OP_WRITE_SAME &&
859 !blk_write_same_mergeable(rq->bio, bio))
863 * Don't allow merge of different write hints, or for a hint with
866 if (rq->write_hint != bio->bi_write_hint)
869 if (rq->ioprio != bio_prio(bio))
875 enum elv_merge blk_try_merge(struct request *rq, struct bio *bio)
877 if (blk_discard_mergable(rq))
878 return ELEVATOR_DISCARD_MERGE;
879 else if (blk_rq_pos(rq) + blk_rq_sectors(rq) == bio->bi_iter.bi_sector)
880 return ELEVATOR_BACK_MERGE;
881 else if (blk_rq_pos(rq) - bio_sectors(bio) == bio->bi_iter.bi_sector)
882 return ELEVATOR_FRONT_MERGE;
883 return ELEVATOR_NO_MERGE;