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 *sectors, unsigned max_segs)
167 unsigned len = bv->bv_len;
168 unsigned total_len = 0;
169 unsigned new_nsegs = 0, seg_size = 0;
172 * Multi-page bvec may be too big to hold in one segment, so the
173 * current bvec has to be splitted as multiple segments.
175 while (len && new_nsegs + *nsegs < max_segs) {
176 seg_size = get_max_segment_size(q, bv->bv_offset + total_len);
177 seg_size = min(seg_size, len);
180 total_len += seg_size;
183 if ((bv->bv_offset + total_len) & queue_virt_boundary(q))
190 *sectors += total_len >> 9;
193 /* split in the middle of the bvec if len != 0 */
197 static struct bio *blk_bio_segment_split(struct request_queue *q,
202 struct bio_vec bv, bvprv, *bvprvp = NULL;
203 struct bvec_iter iter;
204 unsigned nsegs = 0, sectors = 0;
205 bool do_split = true;
206 struct bio *new = NULL;
207 const unsigned max_sectors = get_max_io_size(q, bio);
208 const unsigned max_segs = queue_max_segments(q);
210 bio_for_each_bvec(bv, bio, iter) {
212 * If the queue doesn't support SG gaps and adding this
213 * offset would create a gap, disallow it.
215 if (bvprvp && bvec_gap_to_prev(q, bvprvp, bv.bv_offset))
218 if (sectors + (bv.bv_len >> 9) > max_sectors) {
220 * Consider this a new segment if we're splitting in
221 * the middle of this vector.
223 if (nsegs < max_segs &&
224 sectors < max_sectors) {
225 /* split in the middle of bvec */
226 bv.bv_len = (max_sectors - sectors) << 9;
227 bvec_split_segs(q, &bv, &nsegs,
233 if (nsegs == max_segs)
239 if (bv.bv_offset + bv.bv_len <= PAGE_SIZE) {
241 sectors += bv.bv_len >> 9;
242 } else if (bvec_split_segs(q, &bv, &nsegs, §ors,
253 new = bio_split(bio, sectors, GFP_NOIO, bs);
258 return do_split ? new : NULL;
261 void blk_queue_split(struct request_queue *q, struct bio **bio)
263 struct bio *split, *res;
266 switch (bio_op(*bio)) {
268 case REQ_OP_SECURE_ERASE:
269 split = blk_bio_discard_split(q, *bio, &q->bio_split, &nsegs);
271 case REQ_OP_WRITE_ZEROES:
272 split = blk_bio_write_zeroes_split(q, *bio, &q->bio_split, &nsegs);
274 case REQ_OP_WRITE_SAME:
275 split = blk_bio_write_same_split(q, *bio, &q->bio_split, &nsegs);
278 split = blk_bio_segment_split(q, *bio, &q->bio_split, &nsegs);
282 /* physical segments can be figured out during splitting */
283 res = split ? split : *bio;
284 res->bi_phys_segments = nsegs;
285 bio_set_flag(res, BIO_SEG_VALID);
288 /* there isn't chance to merge the splitted bio */
289 split->bi_opf |= REQ_NOMERGE;
292 * Since we're recursing into make_request here, ensure
293 * that we mark this bio as already having entered the queue.
294 * If not, and the queue is going away, we can get stuck
295 * forever on waiting for the queue reference to drop. But
296 * that will never happen, as we're already holding a
299 bio_set_flag(*bio, BIO_QUEUE_ENTERED);
301 bio_chain(split, *bio);
302 trace_block_split(q, split, (*bio)->bi_iter.bi_sector);
303 generic_make_request(*bio);
307 EXPORT_SYMBOL(blk_queue_split);
309 static unsigned int __blk_recalc_rq_segments(struct request_queue *q,
312 unsigned int nr_phys_segs = 0;
313 struct bvec_iter iter;
319 switch (bio_op(bio)) {
321 case REQ_OP_SECURE_ERASE:
322 case REQ_OP_WRITE_ZEROES:
324 case REQ_OP_WRITE_SAME:
329 bio_for_each_bvec(bv, bio, iter)
330 bvec_split_segs(q, &bv, &nr_phys_segs, NULL, UINT_MAX);
336 void blk_recalc_rq_segments(struct request *rq)
338 rq->nr_phys_segments = __blk_recalc_rq_segments(rq->q, rq->bio);
341 void blk_recount_segments(struct request_queue *q, struct bio *bio)
343 struct bio *nxt = bio->bi_next;
346 bio->bi_phys_segments = __blk_recalc_rq_segments(q, bio);
349 bio_set_flag(bio, BIO_SEG_VALID);
352 static inline struct scatterlist *blk_next_sg(struct scatterlist **sg,
353 struct scatterlist *sglist)
359 * If the driver previously mapped a shorter list, we could see a
360 * termination bit prematurely unless it fully inits the sg table
361 * on each mapping. We KNOW that there must be more entries here
362 * or the driver would be buggy, so force clear the termination bit
363 * to avoid doing a full sg_init_table() in drivers for each command.
369 static unsigned blk_bvec_map_sg(struct request_queue *q,
370 struct bio_vec *bvec, struct scatterlist *sglist,
371 struct scatterlist **sg)
373 unsigned nbytes = bvec->bv_len;
374 unsigned nsegs = 0, total = 0;
377 unsigned offset = bvec->bv_offset + total;
378 unsigned len = min(get_max_segment_size(q, offset), nbytes);
379 struct page *page = bvec->bv_page;
382 * Unfortunately a fair number of drivers barf on scatterlists
383 * that have an offset larger than PAGE_SIZE, despite other
384 * subsystems dealing with that invariant just fine. For now
385 * stick to the legacy format where we never present those from
386 * the block layer, but the code below should be removed once
387 * these offenders (mostly MMC/SD drivers) are fixed.
389 page += (offset >> PAGE_SHIFT);
390 offset &= ~PAGE_MASK;
392 *sg = blk_next_sg(sg, sglist);
393 sg_set_page(*sg, page, len, offset);
403 static inline int __blk_bvec_map_sg(struct bio_vec bv,
404 struct scatterlist *sglist, struct scatterlist **sg)
406 *sg = blk_next_sg(sg, sglist);
407 sg_set_page(*sg, bv.bv_page, bv.bv_len, bv.bv_offset);
411 /* only try to merge bvecs into one sg if they are from two bios */
413 __blk_segment_map_sg_merge(struct request_queue *q, struct bio_vec *bvec,
414 struct bio_vec *bvprv, struct scatterlist **sg)
417 int nbytes = bvec->bv_len;
422 if ((*sg)->length + nbytes > queue_max_segment_size(q))
425 if (!biovec_phys_mergeable(q, bvprv, bvec))
428 (*sg)->length += nbytes;
433 static int __blk_bios_map_sg(struct request_queue *q, struct bio *bio,
434 struct scatterlist *sglist,
435 struct scatterlist **sg)
437 struct bio_vec uninitialized_var(bvec), bvprv = { NULL };
438 struct bvec_iter iter;
440 bool new_bio = false;
443 bio_for_each_bvec(bvec, bio, iter) {
445 * Only try to merge bvecs from two bios given we
446 * have done bio internal merge when adding pages
450 __blk_segment_map_sg_merge(q, &bvec, &bvprv, sg))
453 if (bvec.bv_offset + bvec.bv_len <= PAGE_SIZE)
454 nsegs += __blk_bvec_map_sg(bvec, sglist, sg);
456 nsegs += blk_bvec_map_sg(q, &bvec, sglist, sg);
460 if (likely(bio->bi_iter.bi_size)) {
470 * map a request to scatterlist, return number of sg entries setup. Caller
471 * must make sure sg can hold rq->nr_phys_segments entries
473 int blk_rq_map_sg(struct request_queue *q, struct request *rq,
474 struct scatterlist *sglist)
476 struct scatterlist *sg = NULL;
479 if (rq->rq_flags & RQF_SPECIAL_PAYLOAD)
480 nsegs = __blk_bvec_map_sg(rq->special_vec, sglist, &sg);
481 else if (rq->bio && bio_op(rq->bio) == REQ_OP_WRITE_SAME)
482 nsegs = __blk_bvec_map_sg(bio_iovec(rq->bio), sglist, &sg);
484 nsegs = __blk_bios_map_sg(q, rq->bio, sglist, &sg);
486 if (unlikely(rq->rq_flags & RQF_COPY_USER) &&
487 (blk_rq_bytes(rq) & q->dma_pad_mask)) {
488 unsigned int pad_len =
489 (q->dma_pad_mask & ~blk_rq_bytes(rq)) + 1;
491 sg->length += pad_len;
492 rq->extra_len += pad_len;
495 if (q->dma_drain_size && q->dma_drain_needed(rq)) {
496 if (op_is_write(req_op(rq)))
497 memset(q->dma_drain_buffer, 0, q->dma_drain_size);
501 sg_set_page(sg, virt_to_page(q->dma_drain_buffer),
503 ((unsigned long)q->dma_drain_buffer) &
506 rq->extra_len += q->dma_drain_size;
513 * Something must have been wrong if the figured number of
514 * segment is bigger than number of req's physical segments
516 WARN_ON(nsegs > blk_rq_nr_phys_segments(rq));
520 EXPORT_SYMBOL(blk_rq_map_sg);
522 static inline int ll_new_hw_segment(struct request_queue *q,
526 int nr_phys_segs = bio_phys_segments(q, bio);
528 if (req->nr_phys_segments + nr_phys_segs > queue_max_segments(q))
531 if (blk_integrity_merge_bio(q, req, bio) == false)
535 * This will form the start of a new hw segment. Bump both
538 req->nr_phys_segments += nr_phys_segs;
542 req_set_nomerge(q, req);
546 int ll_back_merge_fn(struct request_queue *q, struct request *req,
549 if (req_gap_back_merge(req, bio))
551 if (blk_integrity_rq(req) &&
552 integrity_req_gap_back_merge(req, bio))
554 if (blk_rq_sectors(req) + bio_sectors(bio) >
555 blk_rq_get_max_sectors(req, blk_rq_pos(req))) {
556 req_set_nomerge(q, req);
559 if (!bio_flagged(req->biotail, BIO_SEG_VALID))
560 blk_recount_segments(q, req->biotail);
561 if (!bio_flagged(bio, BIO_SEG_VALID))
562 blk_recount_segments(q, bio);
564 return ll_new_hw_segment(q, req, bio);
567 int ll_front_merge_fn(struct request_queue *q, struct request *req,
571 if (req_gap_front_merge(req, bio))
573 if (blk_integrity_rq(req) &&
574 integrity_req_gap_front_merge(req, bio))
576 if (blk_rq_sectors(req) + bio_sectors(bio) >
577 blk_rq_get_max_sectors(req, bio->bi_iter.bi_sector)) {
578 req_set_nomerge(q, req);
581 if (!bio_flagged(bio, BIO_SEG_VALID))
582 blk_recount_segments(q, bio);
583 if (!bio_flagged(req->bio, BIO_SEG_VALID))
584 blk_recount_segments(q, req->bio);
586 return ll_new_hw_segment(q, req, bio);
589 static bool req_attempt_discard_merge(struct request_queue *q, struct request *req,
590 struct request *next)
592 unsigned short segments = blk_rq_nr_discard_segments(req);
594 if (segments >= queue_max_discard_segments(q))
596 if (blk_rq_sectors(req) + bio_sectors(next->bio) >
597 blk_rq_get_max_sectors(req, blk_rq_pos(req)))
600 req->nr_phys_segments = segments + blk_rq_nr_discard_segments(next);
603 req_set_nomerge(q, req);
607 static int ll_merge_requests_fn(struct request_queue *q, struct request *req,
608 struct request *next)
610 int total_phys_segments;
612 if (req_gap_back_merge(req, next->bio))
616 * Will it become too large?
618 if ((blk_rq_sectors(req) + blk_rq_sectors(next)) >
619 blk_rq_get_max_sectors(req, blk_rq_pos(req)))
622 total_phys_segments = req->nr_phys_segments + next->nr_phys_segments;
623 if (total_phys_segments > queue_max_segments(q))
626 if (blk_integrity_merge_rq(q, req, next) == false)
630 req->nr_phys_segments = total_phys_segments;
635 * blk_rq_set_mixed_merge - mark a request as mixed merge
636 * @rq: request to mark as mixed merge
639 * @rq is about to be mixed merged. Make sure the attributes
640 * which can be mixed are set in each bio and mark @rq as mixed
643 void blk_rq_set_mixed_merge(struct request *rq)
645 unsigned int ff = rq->cmd_flags & REQ_FAILFAST_MASK;
648 if (rq->rq_flags & RQF_MIXED_MERGE)
652 * @rq will no longer represent mixable attributes for all the
653 * contained bios. It will just track those of the first one.
654 * Distributes the attributs to each bio.
656 for (bio = rq->bio; bio; bio = bio->bi_next) {
657 WARN_ON_ONCE((bio->bi_opf & REQ_FAILFAST_MASK) &&
658 (bio->bi_opf & REQ_FAILFAST_MASK) != ff);
661 rq->rq_flags |= RQF_MIXED_MERGE;
664 static void blk_account_io_merge(struct request *req)
666 if (blk_do_io_stat(req)) {
667 struct hd_struct *part;
672 part_dec_in_flight(req->q, part, rq_data_dir(req));
679 * Two cases of handling DISCARD merge:
680 * If max_discard_segments > 1, the driver takes every bio
681 * as a range and send them to controller together. The ranges
682 * needn't to be contiguous.
683 * Otherwise, the bios/requests will be handled as same as
684 * others which should be contiguous.
686 static inline bool blk_discard_mergable(struct request *req)
688 if (req_op(req) == REQ_OP_DISCARD &&
689 queue_max_discard_segments(req->q) > 1)
694 static enum elv_merge blk_try_req_merge(struct request *req,
695 struct request *next)
697 if (blk_discard_mergable(req))
698 return ELEVATOR_DISCARD_MERGE;
699 else if (blk_rq_pos(req) + blk_rq_sectors(req) == blk_rq_pos(next))
700 return ELEVATOR_BACK_MERGE;
702 return ELEVATOR_NO_MERGE;
706 * For non-mq, this has to be called with the request spinlock acquired.
707 * For mq with scheduling, the appropriate queue wide lock should be held.
709 static struct request *attempt_merge(struct request_queue *q,
710 struct request *req, struct request *next)
712 if (!rq_mergeable(req) || !rq_mergeable(next))
715 if (req_op(req) != req_op(next))
718 if (rq_data_dir(req) != rq_data_dir(next)
719 || req->rq_disk != next->rq_disk)
722 if (req_op(req) == REQ_OP_WRITE_SAME &&
723 !blk_write_same_mergeable(req->bio, next->bio))
727 * Don't allow merge of different write hints, or for a hint with
730 if (req->write_hint != next->write_hint)
733 if (req->ioprio != next->ioprio)
737 * If we are allowed to merge, then append bio list
738 * from next to rq and release next. merge_requests_fn
739 * will have updated segment counts, update sector
740 * counts here. Handle DISCARDs separately, as they
741 * have separate settings.
744 switch (blk_try_req_merge(req, next)) {
745 case ELEVATOR_DISCARD_MERGE:
746 if (!req_attempt_discard_merge(q, req, next))
749 case ELEVATOR_BACK_MERGE:
750 if (!ll_merge_requests_fn(q, req, next))
758 * If failfast settings disagree or any of the two is already
759 * a mixed merge, mark both as mixed before proceeding. This
760 * makes sure that all involved bios have mixable attributes
763 if (((req->rq_flags | next->rq_flags) & RQF_MIXED_MERGE) ||
764 (req->cmd_flags & REQ_FAILFAST_MASK) !=
765 (next->cmd_flags & REQ_FAILFAST_MASK)) {
766 blk_rq_set_mixed_merge(req);
767 blk_rq_set_mixed_merge(next);
771 * At this point we have either done a back merge or front merge. We
772 * need the smaller start_time_ns of the merged requests to be the
773 * current request for accounting purposes.
775 if (next->start_time_ns < req->start_time_ns)
776 req->start_time_ns = next->start_time_ns;
778 req->biotail->bi_next = next->bio;
779 req->biotail = next->biotail;
781 req->__data_len += blk_rq_bytes(next);
783 if (!blk_discard_mergable(req))
784 elv_merge_requests(q, req, next);
787 * 'next' is going away, so update stats accordingly
789 blk_account_io_merge(next);
792 * ownership of bio passed from next to req, return 'next' for
799 struct request *attempt_back_merge(struct request_queue *q, struct request *rq)
801 struct request *next = elv_latter_request(q, rq);
804 return attempt_merge(q, rq, next);
809 struct request *attempt_front_merge(struct request_queue *q, struct request *rq)
811 struct request *prev = elv_former_request(q, rq);
814 return attempt_merge(q, prev, rq);
819 int blk_attempt_req_merge(struct request_queue *q, struct request *rq,
820 struct request *next)
822 struct request *free;
824 free = attempt_merge(q, rq, next);
826 blk_put_request(free);
833 bool blk_rq_merge_ok(struct request *rq, struct bio *bio)
835 if (!rq_mergeable(rq) || !bio_mergeable(bio))
838 if (req_op(rq) != bio_op(bio))
841 /* different data direction or already started, don't merge */
842 if (bio_data_dir(bio) != rq_data_dir(rq))
845 /* must be same device */
846 if (rq->rq_disk != bio->bi_disk)
849 /* only merge integrity protected bio into ditto rq */
850 if (blk_integrity_merge_bio(rq->q, rq, bio) == false)
853 /* must be using the same buffer */
854 if (req_op(rq) == REQ_OP_WRITE_SAME &&
855 !blk_write_same_mergeable(rq->bio, bio))
859 * Don't allow merge of different write hints, or for a hint with
862 if (rq->write_hint != bio->bi_write_hint)
865 if (rq->ioprio != bio_prio(bio))
871 enum elv_merge blk_try_merge(struct request *rq, struct bio *bio)
873 if (blk_discard_mergable(rq))
874 return ELEVATOR_DISCARD_MERGE;
875 else if (blk_rq_pos(rq) + blk_rq_sectors(rq) == bio->bi_iter.bi_sector)
876 return ELEVATOR_BACK_MERGE;
877 else if (blk_rq_pos(rq) - bio_sectors(bio) == bio->bi_iter.bi_sector)
878 return ELEVATOR_FRONT_MERGE;
879 return ELEVATOR_NO_MERGE;