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 struct bio *blk_bio_discard_split(struct request_queue *q,
20 unsigned int max_discard_sectors, granularity;
23 unsigned split_sectors;
27 /* Zero-sector (unknown) and one-sector granularities are the same. */
28 granularity = max(q->limits.discard_granularity >> 9, 1U);
30 max_discard_sectors = min(q->limits.max_discard_sectors, UINT_MAX >> 9);
31 max_discard_sectors -= max_discard_sectors % granularity;
33 if (unlikely(!max_discard_sectors)) {
38 if (bio_sectors(bio) <= max_discard_sectors)
41 split_sectors = max_discard_sectors;
44 * If the next starting sector would be misaligned, stop the discard at
45 * the previous aligned sector.
47 alignment = (q->limits.discard_alignment >> 9) % granularity;
49 tmp = bio->bi_iter.bi_sector + split_sectors - alignment;
50 tmp = sector_div(tmp, granularity);
52 if (split_sectors > tmp)
55 return bio_split(bio, split_sectors, GFP_NOIO, bs);
58 static struct bio *blk_bio_write_zeroes_split(struct request_queue *q,
59 struct bio *bio, struct bio_set *bs, unsigned *nsegs)
63 if (!q->limits.max_write_zeroes_sectors)
66 if (bio_sectors(bio) <= q->limits.max_write_zeroes_sectors)
69 return bio_split(bio, q->limits.max_write_zeroes_sectors, GFP_NOIO, bs);
72 static struct bio *blk_bio_write_same_split(struct request_queue *q,
79 if (!q->limits.max_write_same_sectors)
82 if (bio_sectors(bio) <= q->limits.max_write_same_sectors)
85 return bio_split(bio, q->limits.max_write_same_sectors, GFP_NOIO, bs);
88 static inline unsigned get_max_io_size(struct request_queue *q,
91 unsigned sectors = blk_max_size_offset(q, bio->bi_iter.bi_sector);
92 unsigned mask = queue_logical_block_size(q) - 1;
94 /* aligned to logical block size */
95 sectors &= ~(mask >> 9);
100 static struct bio *blk_bio_segment_split(struct request_queue *q,
105 struct bio_vec bv, bvprv, *bvprvp = NULL;
106 struct bvec_iter iter;
107 unsigned seg_size = 0, nsegs = 0, sectors = 0;
108 unsigned front_seg_size = bio->bi_seg_front_size;
109 bool do_split = true;
110 struct bio *new = NULL;
111 const unsigned max_sectors = get_max_io_size(q, bio);
113 bio_for_each_segment(bv, bio, iter) {
115 * If the queue doesn't support SG gaps and adding this
116 * offset would create a gap, disallow it.
118 if (bvprvp && bvec_gap_to_prev(q, bvprvp, bv.bv_offset))
121 if (sectors + (bv.bv_len >> 9) > max_sectors) {
123 * Consider this a new segment if we're splitting in
124 * the middle of this vector.
126 if (nsegs < queue_max_segments(q) &&
127 sectors < max_sectors) {
129 sectors = max_sectors;
134 if (bvprvp && blk_queue_cluster(q)) {
135 if (seg_size + bv.bv_len > queue_max_segment_size(q))
137 if (!BIOVEC_PHYS_MERGEABLE(bvprvp, &bv))
139 if (!BIOVEC_SEG_BOUNDARY(q, bvprvp, &bv))
142 seg_size += bv.bv_len;
145 sectors += bv.bv_len >> 9;
150 if (nsegs == queue_max_segments(q))
153 if (nsegs == 1 && seg_size > front_seg_size)
154 front_seg_size = seg_size;
159 seg_size = bv.bv_len;
160 sectors += bv.bv_len >> 9;
169 new = bio_split(bio, sectors, GFP_NOIO, bs);
174 if (nsegs == 1 && seg_size > front_seg_size)
175 front_seg_size = seg_size;
176 bio->bi_seg_front_size = front_seg_size;
177 if (seg_size > bio->bi_seg_back_size)
178 bio->bi_seg_back_size = seg_size;
180 return do_split ? new : NULL;
183 void blk_queue_split(struct request_queue *q, struct bio **bio)
185 struct bio *split, *res;
188 switch (bio_op(*bio)) {
190 case REQ_OP_SECURE_ERASE:
191 split = blk_bio_discard_split(q, *bio, q->bio_split, &nsegs);
193 case REQ_OP_WRITE_ZEROES:
194 split = blk_bio_write_zeroes_split(q, *bio, q->bio_split, &nsegs);
196 case REQ_OP_WRITE_SAME:
197 split = blk_bio_write_same_split(q, *bio, q->bio_split, &nsegs);
200 split = blk_bio_segment_split(q, *bio, q->bio_split, &nsegs);
204 /* physical segments can be figured out during splitting */
205 res = split ? split : *bio;
206 res->bi_phys_segments = nsegs;
207 bio_set_flag(res, BIO_SEG_VALID);
210 /* there isn't chance to merge the splitted bio */
211 split->bi_opf |= REQ_NOMERGE;
213 bio_chain(split, *bio);
214 trace_block_split(q, split, (*bio)->bi_iter.bi_sector);
215 generic_make_request(*bio);
219 EXPORT_SYMBOL(blk_queue_split);
221 static unsigned int __blk_recalc_rq_segments(struct request_queue *q,
225 struct bio_vec bv, bvprv = { NULL };
226 int cluster, prev = 0;
227 unsigned int seg_size, nr_phys_segs;
228 struct bio *fbio, *bbio;
229 struct bvec_iter iter;
234 switch (bio_op(bio)) {
236 case REQ_OP_SECURE_ERASE:
237 case REQ_OP_WRITE_ZEROES:
239 case REQ_OP_WRITE_SAME:
244 cluster = blk_queue_cluster(q);
248 bio_for_each_segment(bv, bio, iter) {
250 * If SG merging is disabled, each bio vector is
256 if (prev && cluster) {
257 if (seg_size + bv.bv_len
258 > queue_max_segment_size(q))
260 if (!BIOVEC_PHYS_MERGEABLE(&bvprv, &bv))
262 if (!BIOVEC_SEG_BOUNDARY(q, &bvprv, &bv))
265 seg_size += bv.bv_len;
270 if (nr_phys_segs == 1 && seg_size >
271 fbio->bi_seg_front_size)
272 fbio->bi_seg_front_size = seg_size;
277 seg_size = bv.bv_len;
282 if (nr_phys_segs == 1 && seg_size > fbio->bi_seg_front_size)
283 fbio->bi_seg_front_size = seg_size;
284 if (seg_size > bbio->bi_seg_back_size)
285 bbio->bi_seg_back_size = seg_size;
290 void blk_recalc_rq_segments(struct request *rq)
292 bool no_sg_merge = !!test_bit(QUEUE_FLAG_NO_SG_MERGE,
293 &rq->q->queue_flags);
295 rq->nr_phys_segments = __blk_recalc_rq_segments(rq->q, rq->bio,
299 void blk_recount_segments(struct request_queue *q, struct bio *bio)
301 unsigned short seg_cnt;
303 /* estimate segment number by bi_vcnt for non-cloned bio */
304 if (bio_flagged(bio, BIO_CLONED))
305 seg_cnt = bio_segments(bio);
307 seg_cnt = bio->bi_vcnt;
309 if (test_bit(QUEUE_FLAG_NO_SG_MERGE, &q->queue_flags) &&
310 (seg_cnt < queue_max_segments(q)))
311 bio->bi_phys_segments = seg_cnt;
313 struct bio *nxt = bio->bi_next;
316 bio->bi_phys_segments = __blk_recalc_rq_segments(q, bio, false);
320 bio_set_flag(bio, BIO_SEG_VALID);
322 EXPORT_SYMBOL(blk_recount_segments);
324 static int blk_phys_contig_segment(struct request_queue *q, struct bio *bio,
327 struct bio_vec end_bv = { NULL }, nxt_bv;
329 if (!blk_queue_cluster(q))
332 if (bio->bi_seg_back_size + nxt->bi_seg_front_size >
333 queue_max_segment_size(q))
336 if (!bio_has_data(bio))
339 bio_get_last_bvec(bio, &end_bv);
340 bio_get_first_bvec(nxt, &nxt_bv);
342 if (!BIOVEC_PHYS_MERGEABLE(&end_bv, &nxt_bv))
346 * bio and nxt are contiguous in memory; check if the queue allows
347 * these two to be merged into one
349 if (BIOVEC_SEG_BOUNDARY(q, &end_bv, &nxt_bv))
356 __blk_segment_map_sg(struct request_queue *q, struct bio_vec *bvec,
357 struct scatterlist *sglist, struct bio_vec *bvprv,
358 struct scatterlist **sg, int *nsegs, int *cluster)
361 int nbytes = bvec->bv_len;
363 if (*sg && *cluster) {
364 if ((*sg)->length + nbytes > queue_max_segment_size(q))
367 if (!BIOVEC_PHYS_MERGEABLE(bvprv, bvec))
369 if (!BIOVEC_SEG_BOUNDARY(q, bvprv, bvec))
372 (*sg)->length += nbytes;
379 * If the driver previously mapped a shorter
380 * list, we could see a termination bit
381 * prematurely unless it fully inits the sg
382 * table on each mapping. We KNOW that there
383 * must be more entries here or the driver
384 * would be buggy, so force clear the
385 * termination bit to avoid doing a full
386 * sg_init_table() in drivers for each command.
392 sg_set_page(*sg, bvec->bv_page, nbytes, bvec->bv_offset);
398 static inline int __blk_bvec_map_sg(struct request_queue *q, struct bio_vec bv,
399 struct scatterlist *sglist, struct scatterlist **sg)
402 sg_set_page(*sg, bv.bv_page, bv.bv_len, bv.bv_offset);
406 static int __blk_bios_map_sg(struct request_queue *q, struct bio *bio,
407 struct scatterlist *sglist,
408 struct scatterlist **sg)
410 struct bio_vec bvec, bvprv = { NULL };
411 struct bvec_iter iter;
412 int cluster = blk_queue_cluster(q), nsegs = 0;
415 bio_for_each_segment(bvec, bio, iter)
416 __blk_segment_map_sg(q, &bvec, sglist, &bvprv, sg,
423 * map a request to scatterlist, return number of sg entries setup. Caller
424 * must make sure sg can hold rq->nr_phys_segments entries
426 int blk_rq_map_sg(struct request_queue *q, struct request *rq,
427 struct scatterlist *sglist)
429 struct scatterlist *sg = NULL;
432 if (rq->rq_flags & RQF_SPECIAL_PAYLOAD)
433 nsegs = __blk_bvec_map_sg(q, rq->special_vec, sglist, &sg);
434 else if (rq->bio && bio_op(rq->bio) == REQ_OP_WRITE_SAME)
435 nsegs = __blk_bvec_map_sg(q, bio_iovec(rq->bio), sglist, &sg);
437 nsegs = __blk_bios_map_sg(q, rq->bio, sglist, &sg);
439 if (unlikely(rq->rq_flags & RQF_COPY_USER) &&
440 (blk_rq_bytes(rq) & q->dma_pad_mask)) {
441 unsigned int pad_len =
442 (q->dma_pad_mask & ~blk_rq_bytes(rq)) + 1;
444 sg->length += pad_len;
445 rq->extra_len += pad_len;
448 if (q->dma_drain_size && q->dma_drain_needed(rq)) {
449 if (op_is_write(req_op(rq)))
450 memset(q->dma_drain_buffer, 0, q->dma_drain_size);
454 sg_set_page(sg, virt_to_page(q->dma_drain_buffer),
456 ((unsigned long)q->dma_drain_buffer) &
459 rq->extra_len += q->dma_drain_size;
466 * Something must have been wrong if the figured number of
467 * segment is bigger than number of req's physical segments
469 WARN_ON(nsegs > blk_rq_nr_phys_segments(rq));
473 EXPORT_SYMBOL(blk_rq_map_sg);
475 static inline int ll_new_hw_segment(struct request_queue *q,
479 int nr_phys_segs = bio_phys_segments(q, bio);
481 if (req->nr_phys_segments + nr_phys_segs > queue_max_segments(q))
484 if (blk_integrity_merge_bio(q, req, bio) == false)
488 * This will form the start of a new hw segment. Bump both
491 req->nr_phys_segments += nr_phys_segs;
495 req_set_nomerge(q, req);
499 int ll_back_merge_fn(struct request_queue *q, struct request *req,
502 if (req_gap_back_merge(req, bio))
504 if (blk_integrity_rq(req) &&
505 integrity_req_gap_back_merge(req, bio))
507 if (blk_rq_sectors(req) + bio_sectors(bio) >
508 blk_rq_get_max_sectors(req, blk_rq_pos(req))) {
509 req_set_nomerge(q, req);
512 if (!bio_flagged(req->biotail, BIO_SEG_VALID))
513 blk_recount_segments(q, req->biotail);
514 if (!bio_flagged(bio, BIO_SEG_VALID))
515 blk_recount_segments(q, bio);
517 return ll_new_hw_segment(q, req, bio);
520 int ll_front_merge_fn(struct request_queue *q, struct request *req,
524 if (req_gap_front_merge(req, bio))
526 if (blk_integrity_rq(req) &&
527 integrity_req_gap_front_merge(req, bio))
529 if (blk_rq_sectors(req) + bio_sectors(bio) >
530 blk_rq_get_max_sectors(req, bio->bi_iter.bi_sector)) {
531 req_set_nomerge(q, req);
534 if (!bio_flagged(bio, BIO_SEG_VALID))
535 blk_recount_segments(q, bio);
536 if (!bio_flagged(req->bio, BIO_SEG_VALID))
537 blk_recount_segments(q, req->bio);
539 return ll_new_hw_segment(q, req, bio);
543 * blk-mq uses req->special to carry normal driver per-request payload, it
544 * does not indicate a prepared command that we cannot merge with.
546 static bool req_no_special_merge(struct request *req)
548 struct request_queue *q = req->q;
550 return !q->mq_ops && req->special;
553 static bool req_attempt_discard_merge(struct request_queue *q, struct request *req,
554 struct request *next)
556 unsigned short segments = blk_rq_nr_discard_segments(req);
558 if (segments >= queue_max_discard_segments(q))
560 if (blk_rq_sectors(req) + bio_sectors(next->bio) >
561 blk_rq_get_max_sectors(req, blk_rq_pos(req)))
564 req->nr_phys_segments = segments + blk_rq_nr_discard_segments(next);
567 req_set_nomerge(q, req);
571 static int ll_merge_requests_fn(struct request_queue *q, struct request *req,
572 struct request *next)
574 int total_phys_segments;
575 unsigned int seg_size =
576 req->biotail->bi_seg_back_size + next->bio->bi_seg_front_size;
579 * First check if the either of the requests are re-queued
580 * requests. Can't merge them if they are.
582 if (req_no_special_merge(req) || req_no_special_merge(next))
585 if (req_gap_back_merge(req, next->bio))
589 * Will it become too large?
591 if ((blk_rq_sectors(req) + blk_rq_sectors(next)) >
592 blk_rq_get_max_sectors(req, blk_rq_pos(req)))
595 total_phys_segments = req->nr_phys_segments + next->nr_phys_segments;
596 if (blk_phys_contig_segment(q, req->biotail, next->bio)) {
597 if (req->nr_phys_segments == 1)
598 req->bio->bi_seg_front_size = seg_size;
599 if (next->nr_phys_segments == 1)
600 next->biotail->bi_seg_back_size = seg_size;
601 total_phys_segments--;
604 if (total_phys_segments > queue_max_segments(q))
607 if (blk_integrity_merge_rq(q, req, next) == false)
611 req->nr_phys_segments = total_phys_segments;
616 * blk_rq_set_mixed_merge - mark a request as mixed merge
617 * @rq: request to mark as mixed merge
620 * @rq is about to be mixed merged. Make sure the attributes
621 * which can be mixed are set in each bio and mark @rq as mixed
624 void blk_rq_set_mixed_merge(struct request *rq)
626 unsigned int ff = rq->cmd_flags & REQ_FAILFAST_MASK;
629 if (rq->rq_flags & RQF_MIXED_MERGE)
633 * @rq will no longer represent mixable attributes for all the
634 * contained bios. It will just track those of the first one.
635 * Distributes the attributs to each bio.
637 for (bio = rq->bio; bio; bio = bio->bi_next) {
638 WARN_ON_ONCE((bio->bi_opf & REQ_FAILFAST_MASK) &&
639 (bio->bi_opf & REQ_FAILFAST_MASK) != ff);
642 rq->rq_flags |= RQF_MIXED_MERGE;
645 static void blk_account_io_merge(struct request *req)
647 if (blk_do_io_stat(req)) {
648 struct hd_struct *part;
651 cpu = part_stat_lock();
654 part_round_stats(req->q, cpu, part);
655 part_dec_in_flight(req->q, part, rq_data_dir(req));
663 * For non-mq, this has to be called with the request spinlock acquired.
664 * For mq with scheduling, the appropriate queue wide lock should be held.
666 static struct request *attempt_merge(struct request_queue *q,
667 struct request *req, struct request *next)
670 lockdep_assert_held(q->queue_lock);
672 if (!rq_mergeable(req) || !rq_mergeable(next))
675 if (req_op(req) != req_op(next))
681 if (blk_rq_pos(req) + blk_rq_sectors(req) != blk_rq_pos(next))
684 if (rq_data_dir(req) != rq_data_dir(next)
685 || req->rq_disk != next->rq_disk
686 || req_no_special_merge(next))
689 if (req_op(req) == REQ_OP_WRITE_SAME &&
690 !blk_write_same_mergeable(req->bio, next->bio))
694 * Don't allow merge of different write hints, or for a hint with
697 if (req->write_hint != next->write_hint)
701 * If we are allowed to merge, then append bio list
702 * from next to rq and release next. merge_requests_fn
703 * will have updated segment counts, update sector
704 * counts here. Handle DISCARDs separately, as they
705 * have separate settings.
707 if (req_op(req) == REQ_OP_DISCARD) {
708 if (!req_attempt_discard_merge(q, req, next))
710 } else if (!ll_merge_requests_fn(q, req, next))
714 * If failfast settings disagree or any of the two is already
715 * a mixed merge, mark both as mixed before proceeding. This
716 * makes sure that all involved bios have mixable attributes
719 if (((req->rq_flags | next->rq_flags) & RQF_MIXED_MERGE) ||
720 (req->cmd_flags & REQ_FAILFAST_MASK) !=
721 (next->cmd_flags & REQ_FAILFAST_MASK)) {
722 blk_rq_set_mixed_merge(req);
723 blk_rq_set_mixed_merge(next);
727 * At this point we have either done a back merge
728 * or front merge. We need the smaller start_time of
729 * the merged requests to be the current request
730 * for accounting purposes.
732 if (time_after(req->start_time, next->start_time))
733 req->start_time = next->start_time;
735 req->biotail->bi_next = next->bio;
736 req->biotail = next->biotail;
738 req->__data_len += blk_rq_bytes(next);
740 if (req_op(req) != REQ_OP_DISCARD)
741 elv_merge_requests(q, req, next);
744 * 'next' is going away, so update stats accordingly
746 blk_account_io_merge(next);
748 req->ioprio = ioprio_best(req->ioprio, next->ioprio);
749 if (blk_rq_cpu_valid(next))
750 req->cpu = next->cpu;
753 * ownership of bio passed from next to req, return 'next' for
760 struct request *attempt_back_merge(struct request_queue *q, struct request *rq)
762 struct request *next = elv_latter_request(q, rq);
765 return attempt_merge(q, rq, next);
770 struct request *attempt_front_merge(struct request_queue *q, struct request *rq)
772 struct request *prev = elv_former_request(q, rq);
775 return attempt_merge(q, prev, rq);
780 int blk_attempt_req_merge(struct request_queue *q, struct request *rq,
781 struct request *next)
783 struct elevator_queue *e = q->elevator;
784 struct request *free;
786 if (!e->uses_mq && e->type->ops.sq.elevator_allow_rq_merge_fn)
787 if (!e->type->ops.sq.elevator_allow_rq_merge_fn(q, rq, next))
790 free = attempt_merge(q, rq, next);
792 __blk_put_request(q, free);
799 bool blk_rq_merge_ok(struct request *rq, struct bio *bio)
801 if (!rq_mergeable(rq) || !bio_mergeable(bio))
804 if (req_op(rq) != bio_op(bio))
807 /* different data direction or already started, don't merge */
808 if (bio_data_dir(bio) != rq_data_dir(rq))
811 /* must be same device and not a special request */
812 if (rq->rq_disk != bio->bi_disk || req_no_special_merge(rq))
815 /* only merge integrity protected bio into ditto rq */
816 if (blk_integrity_merge_bio(rq->q, rq, bio) == false)
819 /* must be using the same buffer */
820 if (req_op(rq) == REQ_OP_WRITE_SAME &&
821 !blk_write_same_mergeable(rq->bio, bio))
825 * Don't allow merge of different write hints, or for a hint with
828 if (rq->write_hint != bio->bi_write_hint)
834 enum elv_merge blk_try_merge(struct request *rq, struct bio *bio)
836 if (req_op(rq) == REQ_OP_DISCARD &&
837 queue_max_discard_segments(rq->q) > 1)
838 return ELEVATOR_DISCARD_MERGE;
839 else if (blk_rq_pos(rq) + blk_rq_sectors(rq) == bio->bi_iter.bi_sector)
840 return ELEVATOR_BACK_MERGE;
841 else if (blk_rq_pos(rq) - bio_sectors(bio) == bio->bi_iter.bi_sector)
842 return ELEVATOR_FRONT_MERGE;
843 return ELEVATOR_NO_MERGE;