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);
136 * Return the maximum number of sectors from the start of a bio that may be
137 * submitted as a single request to a block device. If enough sectors remain,
138 * align the end to the physical block size. Otherwise align the end to the
139 * logical block size. This approach minimizes the number of non-aligned
140 * requests that are submitted to a block device if the start of a bio is not
141 * aligned to a physical block boundary.
143 static inline unsigned get_max_io_size(struct request_queue *q,
146 unsigned sectors = blk_max_size_offset(q, bio->bi_iter.bi_sector);
147 unsigned max_sectors = sectors;
148 unsigned pbs = queue_physical_block_size(q) >> SECTOR_SHIFT;
149 unsigned lbs = queue_logical_block_size(q) >> SECTOR_SHIFT;
150 unsigned start_offset = bio->bi_iter.bi_sector & (pbs - 1);
152 max_sectors += start_offset;
153 max_sectors &= ~(pbs - 1);
154 if (max_sectors > start_offset)
155 return max_sectors - start_offset;
157 return sectors & (lbs - 1);
160 static unsigned get_max_segment_size(const struct request_queue *q,
163 unsigned long mask = queue_segment_boundary(q);
165 /* default segment boundary mask means no boundary limit */
166 if (mask == BLK_SEG_BOUNDARY_MASK)
167 return queue_max_segment_size(q);
169 return min_t(unsigned long, mask - (mask & offset) + 1,
170 queue_max_segment_size(q));
174 * bvec_split_segs - verify whether or not a bvec should be split in the middle
175 * @q: [in] request queue associated with the bio associated with @bv
176 * @bv: [in] bvec to examine
177 * @nsegs: [in,out] Number of segments in the bio being built. Incremented
178 * by the number of segments from @bv that may be appended to that
179 * bio without exceeding @max_segs
180 * @sectors: [in,out] Number of sectors in the bio being built. Incremented
181 * by the number of sectors from @bv that may be appended to that
182 * bio without exceeding @max_sectors
183 * @max_segs: [in] upper bound for *@nsegs
184 * @max_sectors: [in] upper bound for *@sectors
186 * When splitting a bio, it can happen that a bvec is encountered that is too
187 * big to fit in a single segment and hence that it has to be split in the
188 * middle. This function verifies whether or not that should happen. The value
189 * %true is returned if and only if appending the entire @bv to a bio with
190 * *@nsegs segments and *@sectors sectors would make that bio unacceptable for
193 static bool bvec_split_segs(const struct request_queue *q,
194 const struct bio_vec *bv, unsigned *nsegs,
195 unsigned *sectors, unsigned max_segs,
196 unsigned max_sectors)
198 unsigned max_len = (min(max_sectors, UINT_MAX >> 9) - *sectors) << 9;
199 unsigned len = min(bv->bv_len, max_len);
200 unsigned total_len = 0;
201 unsigned seg_size = 0;
203 while (len && *nsegs < max_segs) {
204 seg_size = get_max_segment_size(q, bv->bv_offset + total_len);
205 seg_size = min(seg_size, len);
208 total_len += seg_size;
211 if ((bv->bv_offset + total_len) & queue_virt_boundary(q))
215 *sectors += total_len >> 9;
217 /* tell the caller to split the bvec if it is too big to fit */
218 return len > 0 || bv->bv_len > max_len;
222 * blk_bio_segment_split - split a bio in two bios
223 * @q: [in] request queue pointer
224 * @bio: [in] bio to be split
225 * @bs: [in] bio set to allocate the clone from
226 * @segs: [out] number of segments in the bio with the first half of the sectors
228 * Clone @bio, update the bi_iter of the clone to represent the first sectors
229 * of @bio and update @bio->bi_iter to represent the remaining sectors. The
230 * following is guaranteed for the cloned bio:
231 * - That it has at most get_max_io_size(@q, @bio) sectors.
232 * - That it has at most queue_max_segments(@q) segments.
234 * Except for discard requests the cloned bio will point at the bi_io_vec of
235 * the original bio. It is the responsibility of the caller to ensure that the
236 * original bio is not freed before the cloned bio. The caller is also
237 * responsible for ensuring that @bs is only destroyed after processing of the
238 * split bio has finished.
240 static struct bio *blk_bio_segment_split(struct request_queue *q,
245 struct bio_vec bv, bvprv, *bvprvp = NULL;
246 struct bvec_iter iter;
247 unsigned nsegs = 0, sectors = 0;
248 const unsigned max_sectors = get_max_io_size(q, bio);
249 const unsigned max_segs = queue_max_segments(q);
251 bio_for_each_bvec(bv, bio, iter) {
253 * If the queue doesn't support SG gaps and adding this
254 * offset would create a gap, disallow it.
256 if (bvprvp && bvec_gap_to_prev(q, bvprvp, bv.bv_offset))
259 if (nsegs < max_segs &&
260 sectors + (bv.bv_len >> 9) <= max_sectors &&
261 bv.bv_offset + bv.bv_len <= PAGE_SIZE) {
263 sectors += bv.bv_len >> 9;
264 } else if (bvec_split_segs(q, &bv, &nsegs, §ors, max_segs,
277 return bio_split(bio, sectors, GFP_NOIO, bs);
281 * __blk_queue_split - split a bio and submit the second half
282 * @q: [in] request queue pointer
283 * @bio: [in, out] bio to be split
284 * @nr_segs: [out] number of segments in the first bio
286 * Split a bio into two bios, chain the two bios, submit the second half and
287 * store a pointer to the first half in *@bio. If the second bio is still too
288 * big it will be split by a recursive call to this function. Since this
289 * function may allocate a new bio from @q->bio_split, it is the responsibility
290 * of the caller to ensure that @q is only released after processing of the
291 * split bio has finished.
293 void __blk_queue_split(struct request_queue *q, struct bio **bio,
294 unsigned int *nr_segs)
298 switch (bio_op(*bio)) {
300 case REQ_OP_SECURE_ERASE:
301 split = blk_bio_discard_split(q, *bio, &q->bio_split, nr_segs);
303 case REQ_OP_WRITE_ZEROES:
304 split = blk_bio_write_zeroes_split(q, *bio, &q->bio_split,
307 case REQ_OP_WRITE_SAME:
308 split = blk_bio_write_same_split(q, *bio, &q->bio_split,
312 split = blk_bio_segment_split(q, *bio, &q->bio_split, nr_segs);
317 /* there isn't chance to merge the splitted bio */
318 split->bi_opf |= REQ_NOMERGE;
321 * Since we're recursing into make_request here, ensure
322 * that we mark this bio as already having entered the queue.
323 * If not, and the queue is going away, we can get stuck
324 * forever on waiting for the queue reference to drop. But
325 * that will never happen, as we're already holding a
328 bio_set_flag(*bio, BIO_QUEUE_ENTERED);
330 bio_chain(split, *bio);
331 trace_block_split(q, split, (*bio)->bi_iter.bi_sector);
332 generic_make_request(*bio);
338 * blk_queue_split - split a bio and submit the second half
339 * @q: [in] request queue pointer
340 * @bio: [in, out] bio to be split
342 * Split a bio into two bios, chains the two bios, submit the second half and
343 * store a pointer to the first half in *@bio. Since this function may allocate
344 * a new bio from @q->bio_split, it is the responsibility of the caller to
345 * ensure that @q is only released after processing of the split bio has
348 void blk_queue_split(struct request_queue *q, struct bio **bio)
350 unsigned int nr_segs;
352 __blk_queue_split(q, bio, &nr_segs);
354 EXPORT_SYMBOL(blk_queue_split);
356 unsigned int blk_recalc_rq_segments(struct request *rq)
358 unsigned int nr_phys_segs = 0;
359 unsigned int nr_sectors = 0;
360 struct req_iterator iter;
366 switch (bio_op(rq->bio)) {
368 case REQ_OP_SECURE_ERASE:
369 case REQ_OP_WRITE_ZEROES:
371 case REQ_OP_WRITE_SAME:
375 rq_for_each_bvec(bv, rq, iter)
376 bvec_split_segs(rq->q, &bv, &nr_phys_segs, &nr_sectors,
381 static inline struct scatterlist *blk_next_sg(struct scatterlist **sg,
382 struct scatterlist *sglist)
388 * If the driver previously mapped a shorter list, we could see a
389 * termination bit prematurely unless it fully inits the sg table
390 * on each mapping. We KNOW that there must be more entries here
391 * or the driver would be buggy, so force clear the termination bit
392 * to avoid doing a full sg_init_table() in drivers for each command.
398 static unsigned blk_bvec_map_sg(struct request_queue *q,
399 struct bio_vec *bvec, struct scatterlist *sglist,
400 struct scatterlist **sg)
402 unsigned nbytes = bvec->bv_len;
403 unsigned nsegs = 0, total = 0;
406 unsigned offset = bvec->bv_offset + total;
407 unsigned len = min(get_max_segment_size(q, offset), nbytes);
408 struct page *page = bvec->bv_page;
411 * Unfortunately a fair number of drivers barf on scatterlists
412 * that have an offset larger than PAGE_SIZE, despite other
413 * subsystems dealing with that invariant just fine. For now
414 * stick to the legacy format where we never present those from
415 * the block layer, but the code below should be removed once
416 * these offenders (mostly MMC/SD drivers) are fixed.
418 page += (offset >> PAGE_SHIFT);
419 offset &= ~PAGE_MASK;
421 *sg = blk_next_sg(sg, sglist);
422 sg_set_page(*sg, page, len, offset);
432 static inline int __blk_bvec_map_sg(struct bio_vec bv,
433 struct scatterlist *sglist, struct scatterlist **sg)
435 *sg = blk_next_sg(sg, sglist);
436 sg_set_page(*sg, bv.bv_page, bv.bv_len, bv.bv_offset);
440 /* only try to merge bvecs into one sg if they are from two bios */
442 __blk_segment_map_sg_merge(struct request_queue *q, struct bio_vec *bvec,
443 struct bio_vec *bvprv, struct scatterlist **sg)
446 int nbytes = bvec->bv_len;
451 if ((*sg)->length + nbytes > queue_max_segment_size(q))
454 if (!biovec_phys_mergeable(q, bvprv, bvec))
457 (*sg)->length += nbytes;
462 static int __blk_bios_map_sg(struct request_queue *q, struct bio *bio,
463 struct scatterlist *sglist,
464 struct scatterlist **sg)
466 struct bio_vec uninitialized_var(bvec), bvprv = { NULL };
467 struct bvec_iter iter;
469 bool new_bio = false;
472 bio_for_each_bvec(bvec, bio, iter) {
474 * Only try to merge bvecs from two bios given we
475 * have done bio internal merge when adding pages
479 __blk_segment_map_sg_merge(q, &bvec, &bvprv, sg))
482 if (bvec.bv_offset + bvec.bv_len <= PAGE_SIZE)
483 nsegs += __blk_bvec_map_sg(bvec, sglist, sg);
485 nsegs += blk_bvec_map_sg(q, &bvec, sglist, sg);
489 if (likely(bio->bi_iter.bi_size)) {
499 * map a request to scatterlist, return number of sg entries setup. Caller
500 * must make sure sg can hold rq->nr_phys_segments entries
502 int blk_rq_map_sg(struct request_queue *q, struct request *rq,
503 struct scatterlist *sglist)
505 struct scatterlist *sg = NULL;
508 if (rq->rq_flags & RQF_SPECIAL_PAYLOAD)
509 nsegs = __blk_bvec_map_sg(rq->special_vec, sglist, &sg);
510 else if (rq->bio && bio_op(rq->bio) == REQ_OP_WRITE_SAME)
511 nsegs = __blk_bvec_map_sg(bio_iovec(rq->bio), sglist, &sg);
513 nsegs = __blk_bios_map_sg(q, rq->bio, sglist, &sg);
515 if (unlikely(rq->rq_flags & RQF_COPY_USER) &&
516 (blk_rq_bytes(rq) & q->dma_pad_mask)) {
517 unsigned int pad_len =
518 (q->dma_pad_mask & ~blk_rq_bytes(rq)) + 1;
520 sg->length += pad_len;
521 rq->extra_len += pad_len;
524 if (q->dma_drain_size && q->dma_drain_needed(rq)) {
525 if (op_is_write(req_op(rq)))
526 memset(q->dma_drain_buffer, 0, q->dma_drain_size);
530 sg_set_page(sg, virt_to_page(q->dma_drain_buffer),
532 ((unsigned long)q->dma_drain_buffer) &
535 rq->extra_len += q->dma_drain_size;
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 int ll_new_hw_segment(struct request *req, struct bio *bio,
552 unsigned int nr_phys_segs)
554 if (req->nr_phys_segments + nr_phys_segs > queue_max_segments(req->q))
557 if (blk_integrity_merge_bio(req->q, req, bio) == false)
561 * This will form the start of a new hw segment. Bump both
564 req->nr_phys_segments += nr_phys_segs;
568 req_set_nomerge(req->q, req);
572 int ll_back_merge_fn(struct request *req, struct bio *bio, unsigned int nr_segs)
574 if (req_gap_back_merge(req, bio))
576 if (blk_integrity_rq(req) &&
577 integrity_req_gap_back_merge(req, bio))
579 if (blk_rq_sectors(req) + bio_sectors(bio) >
580 blk_rq_get_max_sectors(req, blk_rq_pos(req))) {
581 req_set_nomerge(req->q, req);
585 return ll_new_hw_segment(req, bio, nr_segs);
588 int ll_front_merge_fn(struct request *req, struct bio *bio, unsigned int nr_segs)
590 if (req_gap_front_merge(req, bio))
592 if (blk_integrity_rq(req) &&
593 integrity_req_gap_front_merge(req, bio))
595 if (blk_rq_sectors(req) + bio_sectors(bio) >
596 blk_rq_get_max_sectors(req, bio->bi_iter.bi_sector)) {
597 req_set_nomerge(req->q, req);
601 return ll_new_hw_segment(req, bio, nr_segs);
604 static bool req_attempt_discard_merge(struct request_queue *q, struct request *req,
605 struct request *next)
607 unsigned short segments = blk_rq_nr_discard_segments(req);
609 if (segments >= queue_max_discard_segments(q))
611 if (blk_rq_sectors(req) + bio_sectors(next->bio) >
612 blk_rq_get_max_sectors(req, blk_rq_pos(req)))
615 req->nr_phys_segments = segments + blk_rq_nr_discard_segments(next);
618 req_set_nomerge(q, req);
622 static int ll_merge_requests_fn(struct request_queue *q, struct request *req,
623 struct request *next)
625 int total_phys_segments;
627 if (req_gap_back_merge(req, next->bio))
631 * Will it become too large?
633 if ((blk_rq_sectors(req) + blk_rq_sectors(next)) >
634 blk_rq_get_max_sectors(req, blk_rq_pos(req)))
637 total_phys_segments = req->nr_phys_segments + next->nr_phys_segments;
638 if (total_phys_segments > queue_max_segments(q))
641 if (blk_integrity_merge_rq(q, req, next) == false)
645 req->nr_phys_segments = total_phys_segments;
650 * blk_rq_set_mixed_merge - mark a request as mixed merge
651 * @rq: request to mark as mixed merge
654 * @rq is about to be mixed merged. Make sure the attributes
655 * which can be mixed are set in each bio and mark @rq as mixed
658 void blk_rq_set_mixed_merge(struct request *rq)
660 unsigned int ff = rq->cmd_flags & REQ_FAILFAST_MASK;
663 if (rq->rq_flags & RQF_MIXED_MERGE)
667 * @rq will no longer represent mixable attributes for all the
668 * contained bios. It will just track those of the first one.
669 * Distributes the attributs to each bio.
671 for (bio = rq->bio; bio; bio = bio->bi_next) {
672 WARN_ON_ONCE((bio->bi_opf & REQ_FAILFAST_MASK) &&
673 (bio->bi_opf & REQ_FAILFAST_MASK) != ff);
676 rq->rq_flags |= RQF_MIXED_MERGE;
679 static void blk_account_io_merge(struct request *req)
681 if (blk_do_io_stat(req)) {
682 struct hd_struct *part;
687 part_dec_in_flight(req->q, part, rq_data_dir(req));
694 * Two cases of handling DISCARD merge:
695 * If max_discard_segments > 1, the driver takes every bio
696 * as a range and send them to controller together. The ranges
697 * needn't to be contiguous.
698 * Otherwise, the bios/requests will be handled as same as
699 * others which should be contiguous.
701 static inline bool blk_discard_mergable(struct request *req)
703 if (req_op(req) == REQ_OP_DISCARD &&
704 queue_max_discard_segments(req->q) > 1)
709 static enum elv_merge blk_try_req_merge(struct request *req,
710 struct request *next)
712 if (blk_discard_mergable(req))
713 return ELEVATOR_DISCARD_MERGE;
714 else if (blk_rq_pos(req) + blk_rq_sectors(req) == blk_rq_pos(next))
715 return ELEVATOR_BACK_MERGE;
717 return ELEVATOR_NO_MERGE;
721 * For non-mq, this has to be called with the request spinlock acquired.
722 * For mq with scheduling, the appropriate queue wide lock should be held.
724 static struct request *attempt_merge(struct request_queue *q,
725 struct request *req, struct request *next)
727 if (!rq_mergeable(req) || !rq_mergeable(next))
730 if (req_op(req) != req_op(next))
733 if (rq_data_dir(req) != rq_data_dir(next)
734 || req->rq_disk != next->rq_disk)
737 if (req_op(req) == REQ_OP_WRITE_SAME &&
738 !blk_write_same_mergeable(req->bio, next->bio))
742 * Don't allow merge of different write hints, or for a hint with
745 if (req->write_hint != next->write_hint)
748 if (req->ioprio != next->ioprio)
752 * If we are allowed to merge, then append bio list
753 * from next to rq and release next. merge_requests_fn
754 * will have updated segment counts, update sector
755 * counts here. Handle DISCARDs separately, as they
756 * have separate settings.
759 switch (blk_try_req_merge(req, next)) {
760 case ELEVATOR_DISCARD_MERGE:
761 if (!req_attempt_discard_merge(q, req, next))
764 case ELEVATOR_BACK_MERGE:
765 if (!ll_merge_requests_fn(q, req, next))
773 * If failfast settings disagree or any of the two is already
774 * a mixed merge, mark both as mixed before proceeding. This
775 * makes sure that all involved bios have mixable attributes
778 if (((req->rq_flags | next->rq_flags) & RQF_MIXED_MERGE) ||
779 (req->cmd_flags & REQ_FAILFAST_MASK) !=
780 (next->cmd_flags & REQ_FAILFAST_MASK)) {
781 blk_rq_set_mixed_merge(req);
782 blk_rq_set_mixed_merge(next);
786 * At this point we have either done a back merge or front merge. We
787 * need the smaller start_time_ns of the merged requests to be the
788 * current request for accounting purposes.
790 if (next->start_time_ns < req->start_time_ns)
791 req->start_time_ns = next->start_time_ns;
793 req->biotail->bi_next = next->bio;
794 req->biotail = next->biotail;
796 req->__data_len += blk_rq_bytes(next);
798 if (!blk_discard_mergable(req))
799 elv_merge_requests(q, req, next);
802 * 'next' is going away, so update stats accordingly
804 blk_account_io_merge(next);
807 * ownership of bio passed from next to req, return 'next' for
814 struct request *attempt_back_merge(struct request_queue *q, struct request *rq)
816 struct request *next = elv_latter_request(q, rq);
819 return attempt_merge(q, rq, next);
824 struct request *attempt_front_merge(struct request_queue *q, struct request *rq)
826 struct request *prev = elv_former_request(q, rq);
829 return attempt_merge(q, prev, rq);
834 int blk_attempt_req_merge(struct request_queue *q, struct request *rq,
835 struct request *next)
837 struct request *free;
839 free = attempt_merge(q, rq, next);
841 blk_put_request(free);
848 bool blk_rq_merge_ok(struct request *rq, struct bio *bio)
850 if (!rq_mergeable(rq) || !bio_mergeable(bio))
853 if (req_op(rq) != bio_op(bio))
856 /* different data direction or already started, don't merge */
857 if (bio_data_dir(bio) != rq_data_dir(rq))
860 /* must be same device */
861 if (rq->rq_disk != bio->bi_disk)
864 /* only merge integrity protected bio into ditto rq */
865 if (blk_integrity_merge_bio(rq->q, rq, bio) == false)
868 /* must be using the same buffer */
869 if (req_op(rq) == REQ_OP_WRITE_SAME &&
870 !blk_write_same_mergeable(rq->bio, bio))
874 * Don't allow merge of different write hints, or for a hint with
877 if (rq->write_hint != bio->bi_write_hint)
880 if (rq->ioprio != bio_prio(bio))
886 enum elv_merge blk_try_merge(struct request *rq, struct bio *bio)
888 if (blk_discard_mergable(rq))
889 return ELEVATOR_DISCARD_MERGE;
890 else if (blk_rq_pos(rq) + blk_rq_sectors(rq) == bio->bi_iter.bi_sector)
891 return ELEVATOR_BACK_MERGE;
892 else if (blk_rq_pos(rq) - bio_sectors(bio) == bio->bi_iter.bi_sector)
893 return ELEVATOR_FRONT_MERGE;
894 return ELEVATOR_NO_MERGE;