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/blk-integrity.h>
10 #include <linux/scatterlist.h>
11 #include <linux/part_stat.h>
12 #include <linux/blk-cgroup.h>
14 #include <trace/events/block.h>
17 #include "blk-mq-sched.h"
18 #include "blk-rq-qos.h"
19 #include "blk-throttle.h"
21 static inline void bio_get_first_bvec(struct bio *bio, struct bio_vec *bv)
23 *bv = mp_bvec_iter_bvec(bio->bi_io_vec, bio->bi_iter);
26 static inline void bio_get_last_bvec(struct bio *bio, struct bio_vec *bv)
28 struct bvec_iter iter = bio->bi_iter;
31 bio_get_first_bvec(bio, bv);
32 if (bv->bv_len == bio->bi_iter.bi_size)
33 return; /* this bio only has a single bvec */
35 bio_advance_iter(bio, &iter, iter.bi_size);
37 if (!iter.bi_bvec_done)
38 idx = iter.bi_idx - 1;
39 else /* in the middle of bvec */
42 *bv = bio->bi_io_vec[idx];
45 * iter.bi_bvec_done records actual length of the last bvec
46 * if this bio ends in the middle of one io vector
48 if (iter.bi_bvec_done)
49 bv->bv_len = iter.bi_bvec_done;
52 static inline bool bio_will_gap(struct request_queue *q,
53 struct request *prev_rq, struct bio *prev, struct bio *next)
55 struct bio_vec pb, nb;
57 if (!bio_has_data(prev) || !queue_virt_boundary(q))
61 * Don't merge if the 1st bio starts with non-zero offset, otherwise it
62 * is quite difficult to respect the sg gap limit. We work hard to
63 * merge a huge number of small single bios in case of mkfs.
66 bio_get_first_bvec(prev_rq->bio, &pb);
68 bio_get_first_bvec(prev, &pb);
69 if (pb.bv_offset & queue_virt_boundary(q))
73 * We don't need to worry about the situation that the merged segment
74 * ends in unaligned virt boundary:
76 * - if 'pb' ends aligned, the merged segment ends aligned
77 * - if 'pb' ends unaligned, the next bio must include
78 * one single bvec of 'nb', otherwise the 'nb' can't
81 bio_get_last_bvec(prev, &pb);
82 bio_get_first_bvec(next, &nb);
83 if (biovec_phys_mergeable(q, &pb, &nb))
85 return __bvec_gap_to_prev(q, &pb, nb.bv_offset);
88 static inline bool req_gap_back_merge(struct request *req, struct bio *bio)
90 return bio_will_gap(req->q, req, req->biotail, bio);
93 static inline bool req_gap_front_merge(struct request *req, struct bio *bio)
95 return bio_will_gap(req->q, NULL, bio, req->bio);
98 static struct bio *blk_bio_discard_split(struct request_queue *q,
103 unsigned int max_discard_sectors, granularity;
106 unsigned split_sectors;
110 /* Zero-sector (unknown) and one-sector granularities are the same. */
111 granularity = max(q->limits.discard_granularity >> 9, 1U);
113 max_discard_sectors = min(q->limits.max_discard_sectors,
114 bio_allowed_max_sectors(q));
115 max_discard_sectors -= max_discard_sectors % granularity;
117 if (unlikely(!max_discard_sectors)) {
122 if (bio_sectors(bio) <= max_discard_sectors)
125 split_sectors = max_discard_sectors;
128 * If the next starting sector would be misaligned, stop the discard at
129 * the previous aligned sector.
131 alignment = (q->limits.discard_alignment >> 9) % granularity;
133 tmp = bio->bi_iter.bi_sector + split_sectors - alignment;
134 tmp = sector_div(tmp, granularity);
136 if (split_sectors > tmp)
137 split_sectors -= tmp;
139 return bio_split(bio, split_sectors, GFP_NOIO, bs);
142 static struct bio *blk_bio_write_zeroes_split(struct request_queue *q,
143 struct bio *bio, struct bio_set *bs, unsigned *nsegs)
147 if (!q->limits.max_write_zeroes_sectors)
150 if (bio_sectors(bio) <= q->limits.max_write_zeroes_sectors)
153 return bio_split(bio, q->limits.max_write_zeroes_sectors, GFP_NOIO, bs);
157 * Return the maximum number of sectors from the start of a bio that may be
158 * submitted as a single request to a block device. If enough sectors remain,
159 * align the end to the physical block size. Otherwise align the end to the
160 * logical block size. This approach minimizes the number of non-aligned
161 * requests that are submitted to a block device if the start of a bio is not
162 * aligned to a physical block boundary.
164 static inline unsigned get_max_io_size(struct request_queue *q,
167 unsigned sectors = blk_max_size_offset(q, bio->bi_iter.bi_sector, 0);
168 unsigned max_sectors = sectors;
169 unsigned pbs = queue_physical_block_size(q) >> SECTOR_SHIFT;
170 unsigned lbs = queue_logical_block_size(q) >> SECTOR_SHIFT;
171 unsigned start_offset = bio->bi_iter.bi_sector & (pbs - 1);
173 max_sectors += start_offset;
174 max_sectors &= ~(pbs - 1);
175 if (max_sectors > start_offset)
176 return max_sectors - start_offset;
178 return sectors & ~(lbs - 1);
181 static inline unsigned get_max_segment_size(const struct request_queue *q,
182 struct page *start_page,
183 unsigned long offset)
185 unsigned long mask = queue_segment_boundary(q);
187 offset = mask & (page_to_phys(start_page) + offset);
190 * overflow may be triggered in case of zero page physical address
191 * on 32bit arch, use queue's max segment size when that happens.
193 return min_not_zero(mask - offset + 1,
194 (unsigned long)queue_max_segment_size(q));
198 * bvec_split_segs - verify whether or not a bvec should be split in the middle
199 * @q: [in] request queue associated with the bio associated with @bv
200 * @bv: [in] bvec to examine
201 * @nsegs: [in,out] Number of segments in the bio being built. Incremented
202 * by the number of segments from @bv that may be appended to that
203 * bio without exceeding @max_segs
204 * @sectors: [in,out] Number of sectors in the bio being built. Incremented
205 * by the number of sectors from @bv that may be appended to that
206 * bio without exceeding @max_sectors
207 * @max_segs: [in] upper bound for *@nsegs
208 * @max_sectors: [in] upper bound for *@sectors
210 * When splitting a bio, it can happen that a bvec is encountered that is too
211 * big to fit in a single segment and hence that it has to be split in the
212 * middle. This function verifies whether or not that should happen. The value
213 * %true is returned if and only if appending the entire @bv to a bio with
214 * *@nsegs segments and *@sectors sectors would make that bio unacceptable for
217 static bool bvec_split_segs(const struct request_queue *q,
218 const struct bio_vec *bv, unsigned *nsegs,
219 unsigned *sectors, unsigned max_segs,
220 unsigned max_sectors)
222 unsigned max_len = (min(max_sectors, UINT_MAX >> 9) - *sectors) << 9;
223 unsigned len = min(bv->bv_len, max_len);
224 unsigned total_len = 0;
225 unsigned seg_size = 0;
227 while (len && *nsegs < max_segs) {
228 seg_size = get_max_segment_size(q, bv->bv_page,
229 bv->bv_offset + total_len);
230 seg_size = min(seg_size, len);
233 total_len += seg_size;
236 if ((bv->bv_offset + total_len) & queue_virt_boundary(q))
240 *sectors += total_len >> 9;
242 /* tell the caller to split the bvec if it is too big to fit */
243 return len > 0 || bv->bv_len > max_len;
247 * blk_bio_segment_split - split a bio in two bios
248 * @q: [in] request queue pointer
249 * @bio: [in] bio to be split
250 * @bs: [in] bio set to allocate the clone from
251 * @segs: [out] number of segments in the bio with the first half of the sectors
253 * Clone @bio, update the bi_iter of the clone to represent the first sectors
254 * of @bio and update @bio->bi_iter to represent the remaining sectors. The
255 * following is guaranteed for the cloned bio:
256 * - That it has at most get_max_io_size(@q, @bio) sectors.
257 * - That it has at most queue_max_segments(@q) segments.
259 * Except for discard requests the cloned bio will point at the bi_io_vec of
260 * the original bio. It is the responsibility of the caller to ensure that the
261 * original bio is not freed before the cloned bio. The caller is also
262 * responsible for ensuring that @bs is only destroyed after processing of the
263 * split bio has finished.
265 static struct bio *blk_bio_segment_split(struct request_queue *q,
270 struct bio_vec bv, bvprv, *bvprvp = NULL;
271 struct bvec_iter iter;
272 unsigned nsegs = 0, sectors = 0;
273 const unsigned max_sectors = get_max_io_size(q, bio);
274 const unsigned max_segs = queue_max_segments(q);
276 bio_for_each_bvec(bv, bio, iter) {
278 * If the queue doesn't support SG gaps and adding this
279 * offset would create a gap, disallow it.
281 if (bvprvp && bvec_gap_to_prev(q, bvprvp, bv.bv_offset))
284 if (nsegs < max_segs &&
285 sectors + (bv.bv_len >> 9) <= max_sectors &&
286 bv.bv_offset + bv.bv_len <= PAGE_SIZE) {
288 sectors += bv.bv_len >> 9;
289 } else if (bvec_split_segs(q, &bv, &nsegs, §ors, max_segs,
304 * Bio splitting may cause subtle trouble such as hang when doing sync
305 * iopoll in direct IO routine. Given performance gain of iopoll for
306 * big IO can be trival, disable iopoll when split needed.
308 bio_clear_polled(bio);
309 return bio_split(bio, sectors, GFP_NOIO, bs);
313 * __blk_queue_split - split a bio and submit the second half
314 * @q: [in] request_queue new bio is being queued at
315 * @bio: [in, out] bio to be split
316 * @nr_segs: [out] number of segments in the first bio
318 * Split a bio into two bios, chain the two bios, submit the second half and
319 * store a pointer to the first half in *@bio. If the second bio is still too
320 * big it will be split by a recursive call to this function. Since this
321 * function may allocate a new bio from q->bio_split, it is the responsibility
322 * of the caller to ensure that q->bio_split is only released after processing
323 * of the split bio has finished.
325 void __blk_queue_split(struct request_queue *q, struct bio **bio,
326 unsigned int *nr_segs)
328 struct bio *split = NULL;
330 switch (bio_op(*bio)) {
332 case REQ_OP_SECURE_ERASE:
333 split = blk_bio_discard_split(q, *bio, &q->bio_split, nr_segs);
335 case REQ_OP_WRITE_ZEROES:
336 split = blk_bio_write_zeroes_split(q, *bio, &q->bio_split,
340 split = blk_bio_segment_split(q, *bio, &q->bio_split, nr_segs);
345 /* there isn't chance to merge the splitted bio */
346 split->bi_opf |= REQ_NOMERGE;
348 bio_chain(split, *bio);
349 trace_block_split(split, (*bio)->bi_iter.bi_sector);
350 submit_bio_noacct(*bio);
356 * blk_queue_split - split a bio and submit the second half
357 * @bio: [in, out] bio to be split
359 * Split a bio into two bios, chains the two bios, submit the second half and
360 * store a pointer to the first half in *@bio. Since this function may allocate
361 * a new bio from q->bio_split, it is the responsibility of the caller to ensure
362 * that q->bio_split is only released after processing of the split bio has
365 void blk_queue_split(struct bio **bio)
367 struct request_queue *q = bdev_get_queue((*bio)->bi_bdev);
368 unsigned int nr_segs;
370 if (blk_may_split(q, *bio))
371 __blk_queue_split(q, bio, &nr_segs);
373 EXPORT_SYMBOL(blk_queue_split);
375 unsigned int blk_recalc_rq_segments(struct request *rq)
377 unsigned int nr_phys_segs = 0;
378 unsigned int nr_sectors = 0;
379 struct req_iterator iter;
385 switch (bio_op(rq->bio)) {
387 case REQ_OP_SECURE_ERASE:
388 if (queue_max_discard_segments(rq->q) > 1) {
389 struct bio *bio = rq->bio;
396 case REQ_OP_WRITE_ZEROES:
400 rq_for_each_bvec(bv, rq, iter)
401 bvec_split_segs(rq->q, &bv, &nr_phys_segs, &nr_sectors,
406 static inline struct scatterlist *blk_next_sg(struct scatterlist **sg,
407 struct scatterlist *sglist)
413 * If the driver previously mapped a shorter list, we could see a
414 * termination bit prematurely unless it fully inits the sg table
415 * on each mapping. We KNOW that there must be more entries here
416 * or the driver would be buggy, so force clear the termination bit
417 * to avoid doing a full sg_init_table() in drivers for each command.
423 static unsigned blk_bvec_map_sg(struct request_queue *q,
424 struct bio_vec *bvec, struct scatterlist *sglist,
425 struct scatterlist **sg)
427 unsigned nbytes = bvec->bv_len;
428 unsigned nsegs = 0, total = 0;
431 unsigned offset = bvec->bv_offset + total;
432 unsigned len = min(get_max_segment_size(q, bvec->bv_page,
434 struct page *page = bvec->bv_page;
437 * Unfortunately a fair number of drivers barf on scatterlists
438 * that have an offset larger than PAGE_SIZE, despite other
439 * subsystems dealing with that invariant just fine. For now
440 * stick to the legacy format where we never present those from
441 * the block layer, but the code below should be removed once
442 * these offenders (mostly MMC/SD drivers) are fixed.
444 page += (offset >> PAGE_SHIFT);
445 offset &= ~PAGE_MASK;
447 *sg = blk_next_sg(sg, sglist);
448 sg_set_page(*sg, page, len, offset);
458 static inline int __blk_bvec_map_sg(struct bio_vec bv,
459 struct scatterlist *sglist, struct scatterlist **sg)
461 *sg = blk_next_sg(sg, sglist);
462 sg_set_page(*sg, bv.bv_page, bv.bv_len, bv.bv_offset);
466 /* only try to merge bvecs into one sg if they are from two bios */
468 __blk_segment_map_sg_merge(struct request_queue *q, struct bio_vec *bvec,
469 struct bio_vec *bvprv, struct scatterlist **sg)
472 int nbytes = bvec->bv_len;
477 if ((*sg)->length + nbytes > queue_max_segment_size(q))
480 if (!biovec_phys_mergeable(q, bvprv, bvec))
483 (*sg)->length += nbytes;
488 static int __blk_bios_map_sg(struct request_queue *q, struct bio *bio,
489 struct scatterlist *sglist,
490 struct scatterlist **sg)
492 struct bio_vec bvec, bvprv = { NULL };
493 struct bvec_iter iter;
495 bool new_bio = false;
498 bio_for_each_bvec(bvec, bio, iter) {
500 * Only try to merge bvecs from two bios given we
501 * have done bio internal merge when adding pages
505 __blk_segment_map_sg_merge(q, &bvec, &bvprv, sg))
508 if (bvec.bv_offset + bvec.bv_len <= PAGE_SIZE)
509 nsegs += __blk_bvec_map_sg(bvec, sglist, sg);
511 nsegs += blk_bvec_map_sg(q, &bvec, sglist, sg);
515 if (likely(bio->bi_iter.bi_size)) {
525 * map a request to scatterlist, return number of sg entries setup. Caller
526 * must make sure sg can hold rq->nr_phys_segments entries
528 int __blk_rq_map_sg(struct request_queue *q, struct request *rq,
529 struct scatterlist *sglist, struct scatterlist **last_sg)
533 if (rq->rq_flags & RQF_SPECIAL_PAYLOAD)
534 nsegs = __blk_bvec_map_sg(rq->special_vec, sglist, last_sg);
536 nsegs = __blk_bios_map_sg(q, rq->bio, sglist, last_sg);
539 sg_mark_end(*last_sg);
542 * Something must have been wrong if the figured number of
543 * segment is bigger than number of req's physical segments
545 WARN_ON(nsegs > blk_rq_nr_phys_segments(rq));
549 EXPORT_SYMBOL(__blk_rq_map_sg);
551 static inline unsigned int blk_rq_get_max_segments(struct request *rq)
553 if (req_op(rq) == REQ_OP_DISCARD)
554 return queue_max_discard_segments(rq->q);
555 return queue_max_segments(rq->q);
558 static inline unsigned int blk_rq_get_max_sectors(struct request *rq,
561 struct request_queue *q = rq->q;
563 if (blk_rq_is_passthrough(rq))
564 return q->limits.max_hw_sectors;
566 if (!q->limits.chunk_sectors ||
567 req_op(rq) == REQ_OP_DISCARD ||
568 req_op(rq) == REQ_OP_SECURE_ERASE)
569 return blk_queue_get_max_sectors(q, req_op(rq));
571 return min(blk_max_size_offset(q, offset, 0),
572 blk_queue_get_max_sectors(q, req_op(rq)));
575 static inline int ll_new_hw_segment(struct request *req, struct bio *bio,
576 unsigned int nr_phys_segs)
578 if (!blk_cgroup_mergeable(req, bio))
581 if (blk_integrity_merge_bio(req->q, req, bio) == false)
584 /* discard request merge won't add new segment */
585 if (req_op(req) == REQ_OP_DISCARD)
588 if (req->nr_phys_segments + nr_phys_segs > blk_rq_get_max_segments(req))
592 * This will form the start of a new hw segment. Bump both
595 req->nr_phys_segments += nr_phys_segs;
599 req_set_nomerge(req->q, req);
603 int ll_back_merge_fn(struct request *req, struct bio *bio, unsigned int nr_segs)
605 if (req_gap_back_merge(req, bio))
607 if (blk_integrity_rq(req) &&
608 integrity_req_gap_back_merge(req, bio))
610 if (!bio_crypt_ctx_back_mergeable(req, bio))
612 if (blk_rq_sectors(req) + bio_sectors(bio) >
613 blk_rq_get_max_sectors(req, blk_rq_pos(req))) {
614 req_set_nomerge(req->q, req);
618 return ll_new_hw_segment(req, bio, nr_segs);
621 static int ll_front_merge_fn(struct request *req, struct bio *bio,
622 unsigned int nr_segs)
624 if (req_gap_front_merge(req, bio))
626 if (blk_integrity_rq(req) &&
627 integrity_req_gap_front_merge(req, bio))
629 if (!bio_crypt_ctx_front_mergeable(req, bio))
631 if (blk_rq_sectors(req) + bio_sectors(bio) >
632 blk_rq_get_max_sectors(req, bio->bi_iter.bi_sector)) {
633 req_set_nomerge(req->q, req);
637 return ll_new_hw_segment(req, bio, nr_segs);
640 static bool req_attempt_discard_merge(struct request_queue *q, struct request *req,
641 struct request *next)
643 unsigned short segments = blk_rq_nr_discard_segments(req);
645 if (segments >= queue_max_discard_segments(q))
647 if (blk_rq_sectors(req) + bio_sectors(next->bio) >
648 blk_rq_get_max_sectors(req, blk_rq_pos(req)))
651 req->nr_phys_segments = segments + blk_rq_nr_discard_segments(next);
654 req_set_nomerge(q, req);
658 static int ll_merge_requests_fn(struct request_queue *q, struct request *req,
659 struct request *next)
661 int total_phys_segments;
663 if (req_gap_back_merge(req, next->bio))
667 * Will it become too large?
669 if ((blk_rq_sectors(req) + blk_rq_sectors(next)) >
670 blk_rq_get_max_sectors(req, blk_rq_pos(req)))
673 total_phys_segments = req->nr_phys_segments + next->nr_phys_segments;
674 if (total_phys_segments > blk_rq_get_max_segments(req))
677 if (!blk_cgroup_mergeable(req, next->bio))
680 if (blk_integrity_merge_rq(q, req, next) == false)
683 if (!bio_crypt_ctx_merge_rq(req, next))
687 req->nr_phys_segments = total_phys_segments;
692 * blk_rq_set_mixed_merge - mark a request as mixed merge
693 * @rq: request to mark as mixed merge
696 * @rq is about to be mixed merged. Make sure the attributes
697 * which can be mixed are set in each bio and mark @rq as mixed
700 void blk_rq_set_mixed_merge(struct request *rq)
702 unsigned int ff = rq->cmd_flags & REQ_FAILFAST_MASK;
705 if (rq->rq_flags & RQF_MIXED_MERGE)
709 * @rq will no longer represent mixable attributes for all the
710 * contained bios. It will just track those of the first one.
711 * Distributes the attributs to each bio.
713 for (bio = rq->bio; bio; bio = bio->bi_next) {
714 WARN_ON_ONCE((bio->bi_opf & REQ_FAILFAST_MASK) &&
715 (bio->bi_opf & REQ_FAILFAST_MASK) != ff);
718 rq->rq_flags |= RQF_MIXED_MERGE;
721 static void blk_account_io_merge_request(struct request *req)
723 if (blk_do_io_stat(req)) {
725 part_stat_inc(req->part, merges[op_stat_group(req_op(req))]);
730 static enum elv_merge blk_try_req_merge(struct request *req,
731 struct request *next)
733 if (blk_discard_mergable(req))
734 return ELEVATOR_DISCARD_MERGE;
735 else if (blk_rq_pos(req) + blk_rq_sectors(req) == blk_rq_pos(next))
736 return ELEVATOR_BACK_MERGE;
738 return ELEVATOR_NO_MERGE;
742 * For non-mq, this has to be called with the request spinlock acquired.
743 * For mq with scheduling, the appropriate queue wide lock should be held.
745 static struct request *attempt_merge(struct request_queue *q,
746 struct request *req, struct request *next)
748 if (!rq_mergeable(req) || !rq_mergeable(next))
751 if (req_op(req) != req_op(next))
754 if (rq_data_dir(req) != rq_data_dir(next))
757 if (req->ioprio != next->ioprio)
761 * If we are allowed to merge, then append bio list
762 * from next to rq and release next. merge_requests_fn
763 * will have updated segment counts, update sector
764 * counts here. Handle DISCARDs separately, as they
765 * have separate settings.
768 switch (blk_try_req_merge(req, next)) {
769 case ELEVATOR_DISCARD_MERGE:
770 if (!req_attempt_discard_merge(q, req, next))
773 case ELEVATOR_BACK_MERGE:
774 if (!ll_merge_requests_fn(q, req, next))
782 * If failfast settings disagree or any of the two is already
783 * a mixed merge, mark both as mixed before proceeding. This
784 * makes sure that all involved bios have mixable attributes
787 if (((req->rq_flags | next->rq_flags) & RQF_MIXED_MERGE) ||
788 (req->cmd_flags & REQ_FAILFAST_MASK) !=
789 (next->cmd_flags & REQ_FAILFAST_MASK)) {
790 blk_rq_set_mixed_merge(req);
791 blk_rq_set_mixed_merge(next);
795 * At this point we have either done a back merge or front merge. We
796 * need the smaller start_time_ns of the merged requests to be the
797 * current request for accounting purposes.
799 if (next->start_time_ns < req->start_time_ns)
800 req->start_time_ns = next->start_time_ns;
802 req->biotail->bi_next = next->bio;
803 req->biotail = next->biotail;
805 req->__data_len += blk_rq_bytes(next);
807 if (!blk_discard_mergable(req))
808 elv_merge_requests(q, req, next);
811 * 'next' is going away, so update stats accordingly
813 blk_account_io_merge_request(next);
815 trace_block_rq_merge(next);
818 * ownership of bio passed from next to req, return 'next' for
825 static struct request *attempt_back_merge(struct request_queue *q,
828 struct request *next = elv_latter_request(q, rq);
831 return attempt_merge(q, rq, next);
836 static struct request *attempt_front_merge(struct request_queue *q,
839 struct request *prev = elv_former_request(q, rq);
842 return attempt_merge(q, prev, rq);
848 * Try to merge 'next' into 'rq'. Return true if the merge happened, false
849 * otherwise. The caller is responsible for freeing 'next' if the merge
852 bool blk_attempt_req_merge(struct request_queue *q, struct request *rq,
853 struct request *next)
855 return attempt_merge(q, rq, next);
858 bool blk_rq_merge_ok(struct request *rq, struct bio *bio)
860 if (!rq_mergeable(rq) || !bio_mergeable(bio))
863 if (req_op(rq) != bio_op(bio))
866 /* different data direction or already started, don't merge */
867 if (bio_data_dir(bio) != rq_data_dir(rq))
870 /* don't merge across cgroup boundaries */
871 if (!blk_cgroup_mergeable(rq, bio))
874 /* only merge integrity protected bio into ditto rq */
875 if (blk_integrity_merge_bio(rq->q, rq, bio) == false)
878 /* Only merge if the crypt contexts are compatible */
879 if (!bio_crypt_rq_ctx_compatible(rq, bio))
882 if (rq->ioprio != bio_prio(bio))
888 enum elv_merge blk_try_merge(struct request *rq, struct bio *bio)
890 if (blk_discard_mergable(rq))
891 return ELEVATOR_DISCARD_MERGE;
892 else if (blk_rq_pos(rq) + blk_rq_sectors(rq) == bio->bi_iter.bi_sector)
893 return ELEVATOR_BACK_MERGE;
894 else if (blk_rq_pos(rq) - bio_sectors(bio) == bio->bi_iter.bi_sector)
895 return ELEVATOR_FRONT_MERGE;
896 return ELEVATOR_NO_MERGE;
899 static void blk_account_io_merge_bio(struct request *req)
901 if (!blk_do_io_stat(req))
905 part_stat_inc(req->part, merges[op_stat_group(req_op(req))]);
909 enum bio_merge_status {
915 static enum bio_merge_status bio_attempt_back_merge(struct request *req,
916 struct bio *bio, unsigned int nr_segs)
918 const int ff = bio->bi_opf & REQ_FAILFAST_MASK;
920 if (!ll_back_merge_fn(req, bio, nr_segs))
921 return BIO_MERGE_FAILED;
923 trace_block_bio_backmerge(bio);
924 rq_qos_merge(req->q, req, bio);
926 if ((req->cmd_flags & REQ_FAILFAST_MASK) != ff)
927 blk_rq_set_mixed_merge(req);
929 req->biotail->bi_next = bio;
931 req->__data_len += bio->bi_iter.bi_size;
933 bio_crypt_free_ctx(bio);
935 blk_account_io_merge_bio(req);
939 static enum bio_merge_status bio_attempt_front_merge(struct request *req,
940 struct bio *bio, unsigned int nr_segs)
942 const int ff = bio->bi_opf & REQ_FAILFAST_MASK;
944 if (!ll_front_merge_fn(req, bio, nr_segs))
945 return BIO_MERGE_FAILED;
947 trace_block_bio_frontmerge(bio);
948 rq_qos_merge(req->q, req, bio);
950 if ((req->cmd_flags & REQ_FAILFAST_MASK) != ff)
951 blk_rq_set_mixed_merge(req);
953 bio->bi_next = req->bio;
956 req->__sector = bio->bi_iter.bi_sector;
957 req->__data_len += bio->bi_iter.bi_size;
959 bio_crypt_do_front_merge(req, bio);
961 blk_account_io_merge_bio(req);
965 static enum bio_merge_status bio_attempt_discard_merge(struct request_queue *q,
966 struct request *req, struct bio *bio)
968 unsigned short segments = blk_rq_nr_discard_segments(req);
970 if (segments >= queue_max_discard_segments(q))
972 if (blk_rq_sectors(req) + bio_sectors(bio) >
973 blk_rq_get_max_sectors(req, blk_rq_pos(req)))
976 rq_qos_merge(q, req, bio);
978 req->biotail->bi_next = bio;
980 req->__data_len += bio->bi_iter.bi_size;
981 req->nr_phys_segments = segments + 1;
983 blk_account_io_merge_bio(req);
986 req_set_nomerge(q, req);
987 return BIO_MERGE_FAILED;
990 static enum bio_merge_status blk_attempt_bio_merge(struct request_queue *q,
993 unsigned int nr_segs,
994 bool sched_allow_merge)
996 if (!blk_rq_merge_ok(rq, bio))
997 return BIO_MERGE_NONE;
999 switch (blk_try_merge(rq, bio)) {
1000 case ELEVATOR_BACK_MERGE:
1001 if (!sched_allow_merge || blk_mq_sched_allow_merge(q, rq, bio))
1002 return bio_attempt_back_merge(rq, bio, nr_segs);
1004 case ELEVATOR_FRONT_MERGE:
1005 if (!sched_allow_merge || blk_mq_sched_allow_merge(q, rq, bio))
1006 return bio_attempt_front_merge(rq, bio, nr_segs);
1008 case ELEVATOR_DISCARD_MERGE:
1009 return bio_attempt_discard_merge(q, rq, bio);
1011 return BIO_MERGE_NONE;
1014 return BIO_MERGE_FAILED;
1018 * blk_attempt_plug_merge - try to merge with %current's plugged list
1019 * @q: request_queue new bio is being queued at
1020 * @bio: new bio being queued
1021 * @nr_segs: number of segments in @bio
1022 * from the passed in @q already in the plug list
1024 * Determine whether @bio being queued on @q can be merged with the previous
1025 * request on %current's plugged list. Returns %true if merge was successful,
1028 * Plugging coalesces IOs from the same issuer for the same purpose without
1029 * going through @q->queue_lock. As such it's more of an issuing mechanism
1030 * than scheduling, and the request, while may have elvpriv data, is not
1031 * added on the elevator at this point. In addition, we don't have
1032 * reliable access to the elevator outside queue lock. Only check basic
1033 * merging parameters without querying the elevator.
1035 * Caller must ensure !blk_queue_nomerges(q) beforehand.
1037 bool blk_attempt_plug_merge(struct request_queue *q, struct bio *bio,
1038 unsigned int nr_segs)
1040 struct blk_plug *plug;
1043 plug = blk_mq_plug(q, bio);
1044 if (!plug || rq_list_empty(plug->mq_list))
1047 rq_list_for_each(&plug->mq_list, rq) {
1049 if (blk_attempt_bio_merge(q, rq, bio, nr_segs, false) ==
1056 * Only keep iterating plug list for merges if we have multiple
1059 if (!plug->multiple_queues)
1066 * Iterate list of requests and see if we can merge this bio with any
1069 bool blk_bio_list_merge(struct request_queue *q, struct list_head *list,
1070 struct bio *bio, unsigned int nr_segs)
1075 list_for_each_entry_reverse(rq, list, queuelist) {
1079 switch (blk_attempt_bio_merge(q, rq, bio, nr_segs, true)) {
1080 case BIO_MERGE_NONE:
1084 case BIO_MERGE_FAILED:
1092 EXPORT_SYMBOL_GPL(blk_bio_list_merge);
1094 bool blk_mq_sched_try_merge(struct request_queue *q, struct bio *bio,
1095 unsigned int nr_segs, struct request **merged_request)
1099 switch (elv_merge(q, &rq, bio)) {
1100 case ELEVATOR_BACK_MERGE:
1101 if (!blk_mq_sched_allow_merge(q, rq, bio))
1103 if (bio_attempt_back_merge(rq, bio, nr_segs) != BIO_MERGE_OK)
1105 *merged_request = attempt_back_merge(q, rq);
1106 if (!*merged_request)
1107 elv_merged_request(q, rq, ELEVATOR_BACK_MERGE);
1109 case ELEVATOR_FRONT_MERGE:
1110 if (!blk_mq_sched_allow_merge(q, rq, bio))
1112 if (bio_attempt_front_merge(rq, bio, nr_segs) != BIO_MERGE_OK)
1114 *merged_request = attempt_front_merge(q, rq);
1115 if (!*merged_request)
1116 elv_merged_request(q, rq, ELEVATOR_FRONT_MERGE);
1118 case ELEVATOR_DISCARD_MERGE:
1119 return bio_attempt_discard_merge(q, rq, bio) == BIO_MERGE_OK;
1124 EXPORT_SYMBOL_GPL(blk_mq_sched_try_merge);