2 * Functions related to segment and merge handling
4 #include <linux/kernel.h>
5 #include <linux/module.h>
7 #include <linux/blkdev.h>
8 #include <linux/scatterlist.h>
12 static struct bio *blk_bio_discard_split(struct request_queue *q,
17 unsigned int max_discard_sectors, granularity;
20 unsigned split_sectors;
24 /* Zero-sector (unknown) and one-sector granularities are the same. */
25 granularity = max(q->limits.discard_granularity >> 9, 1U);
27 max_discard_sectors = min(q->limits.max_discard_sectors, UINT_MAX >> 9);
28 max_discard_sectors -= max_discard_sectors % granularity;
30 if (unlikely(!max_discard_sectors)) {
35 if (bio_sectors(bio) <= max_discard_sectors)
38 split_sectors = max_discard_sectors;
41 * If the next starting sector would be misaligned, stop the discard at
42 * the previous aligned sector.
44 alignment = (q->limits.discard_alignment >> 9) % granularity;
46 tmp = bio->bi_iter.bi_sector + split_sectors - alignment;
47 tmp = sector_div(tmp, granularity);
49 if (split_sectors > tmp)
52 return bio_split(bio, split_sectors, GFP_NOIO, bs);
55 static struct bio *blk_bio_write_same_split(struct request_queue *q,
62 if (!q->limits.max_write_same_sectors)
65 if (bio_sectors(bio) <= q->limits.max_write_same_sectors)
68 return bio_split(bio, q->limits.max_write_same_sectors, GFP_NOIO, bs);
71 static inline unsigned get_max_io_size(struct request_queue *q,
74 unsigned sectors = blk_max_size_offset(q, bio->bi_iter.bi_sector);
75 unsigned mask = queue_logical_block_size(q) - 1;
77 /* aligned to logical block size */
78 sectors &= ~(mask >> 9);
83 static struct bio *blk_bio_segment_split(struct request_queue *q,
88 struct bio_vec bv, bvprv, *bvprvp = NULL;
89 struct bvec_iter iter;
90 unsigned seg_size = 0, nsegs = 0, sectors = 0;
91 unsigned front_seg_size = bio->bi_seg_front_size;
93 struct bio *new = NULL;
94 const unsigned max_sectors = get_max_io_size(q, bio);
96 bio_for_each_segment(bv, bio, iter) {
98 * If the queue doesn't support SG gaps and adding this
99 * offset would create a gap, disallow it.
101 if (bvprvp && bvec_gap_to_prev(q, bvprvp, bv.bv_offset))
104 if (sectors + (bv.bv_len >> 9) > max_sectors) {
106 * Consider this a new segment if we're splitting in
107 * the middle of this vector.
109 if (nsegs < queue_max_segments(q) &&
110 sectors < max_sectors) {
112 sectors = max_sectors;
116 /* Make this single bvec as the 1st segment */
119 if (bvprvp && blk_queue_cluster(q)) {
120 if (seg_size + bv.bv_len > queue_max_segment_size(q))
122 if (!BIOVEC_PHYS_MERGEABLE(bvprvp, &bv))
124 if (!BIOVEC_SEG_BOUNDARY(q, bvprvp, &bv))
127 seg_size += bv.bv_len;
130 sectors += bv.bv_len >> 9;
132 if (nsegs == 1 && seg_size > front_seg_size)
133 front_seg_size = seg_size;
137 if (nsegs == queue_max_segments(q))
143 seg_size = bv.bv_len;
144 sectors += bv.bv_len >> 9;
146 if (nsegs == 1 && seg_size > front_seg_size)
147 front_seg_size = seg_size;
155 new = bio_split(bio, sectors, GFP_NOIO, bs);
160 bio->bi_seg_front_size = front_seg_size;
161 if (seg_size > bio->bi_seg_back_size)
162 bio->bi_seg_back_size = seg_size;
164 return do_split ? new : NULL;
167 void blk_queue_split(struct request_queue *q, struct bio **bio,
170 struct bio *split, *res;
173 if ((*bio)->bi_rw & REQ_DISCARD)
174 split = blk_bio_discard_split(q, *bio, bs, &nsegs);
175 else if ((*bio)->bi_rw & REQ_WRITE_SAME)
176 split = blk_bio_write_same_split(q, *bio, bs, &nsegs);
178 split = blk_bio_segment_split(q, *bio, q->bio_split, &nsegs);
180 /* physical segments can be figured out during splitting */
181 res = split ? split : *bio;
182 res->bi_phys_segments = nsegs;
183 bio_set_flag(res, BIO_SEG_VALID);
186 /* there isn't chance to merge the splitted bio */
187 split->bi_rw |= REQ_NOMERGE;
189 bio_chain(split, *bio);
190 generic_make_request(*bio);
194 EXPORT_SYMBOL(blk_queue_split);
196 static unsigned int __blk_recalc_rq_segments(struct request_queue *q,
200 struct bio_vec bv, bvprv = { NULL };
201 int cluster, prev = 0;
202 unsigned int seg_size, nr_phys_segs;
203 struct bio *fbio, *bbio;
204 struct bvec_iter iter;
210 * This should probably be returning 0, but blk_add_request_payload()
213 if (bio->bi_rw & REQ_DISCARD)
216 if (bio->bi_rw & REQ_WRITE_SAME)
220 cluster = blk_queue_cluster(q);
224 bio_for_each_segment(bv, bio, iter) {
226 * If SG merging is disabled, each bio vector is
232 if (prev && cluster) {
233 if (seg_size + bv.bv_len
234 > queue_max_segment_size(q))
236 if (!BIOVEC_PHYS_MERGEABLE(&bvprv, &bv))
238 if (!BIOVEC_SEG_BOUNDARY(q, &bvprv, &bv))
241 seg_size += bv.bv_len;
246 if (nr_phys_segs == 1 && seg_size >
247 fbio->bi_seg_front_size)
248 fbio->bi_seg_front_size = seg_size;
253 seg_size = bv.bv_len;
258 if (nr_phys_segs == 1 && seg_size > fbio->bi_seg_front_size)
259 fbio->bi_seg_front_size = seg_size;
260 if (seg_size > bbio->bi_seg_back_size)
261 bbio->bi_seg_back_size = seg_size;
266 void blk_recalc_rq_segments(struct request *rq)
268 bool no_sg_merge = !!test_bit(QUEUE_FLAG_NO_SG_MERGE,
269 &rq->q->queue_flags);
271 rq->nr_phys_segments = __blk_recalc_rq_segments(rq->q, rq->bio,
275 void blk_recount_segments(struct request_queue *q, struct bio *bio)
277 unsigned short seg_cnt;
279 /* estimate segment number by bi_vcnt for non-cloned bio */
280 if (bio_flagged(bio, BIO_CLONED))
281 seg_cnt = bio_segments(bio);
283 seg_cnt = bio->bi_vcnt;
285 if (test_bit(QUEUE_FLAG_NO_SG_MERGE, &q->queue_flags) &&
286 (seg_cnt < queue_max_segments(q)))
287 bio->bi_phys_segments = seg_cnt;
289 struct bio *nxt = bio->bi_next;
292 bio->bi_phys_segments = __blk_recalc_rq_segments(q, bio, false);
296 bio_set_flag(bio, BIO_SEG_VALID);
298 EXPORT_SYMBOL(blk_recount_segments);
300 static int blk_phys_contig_segment(struct request_queue *q, struct bio *bio,
303 struct bio_vec end_bv = { NULL }, nxt_bv;
304 struct bvec_iter iter;
306 if (!blk_queue_cluster(q))
309 if (bio->bi_seg_back_size + nxt->bi_seg_front_size >
310 queue_max_segment_size(q))
313 if (!bio_has_data(bio))
316 bio_for_each_segment(end_bv, bio, iter)
317 if (end_bv.bv_len == iter.bi_size)
320 nxt_bv = bio_iovec(nxt);
322 if (!BIOVEC_PHYS_MERGEABLE(&end_bv, &nxt_bv))
326 * bio and nxt are contiguous in memory; check if the queue allows
327 * these two to be merged into one
329 if (BIOVEC_SEG_BOUNDARY(q, &end_bv, &nxt_bv))
336 __blk_segment_map_sg(struct request_queue *q, struct bio_vec *bvec,
337 struct scatterlist *sglist, struct bio_vec *bvprv,
338 struct scatterlist **sg, int *nsegs, int *cluster)
341 int nbytes = bvec->bv_len;
343 if (*sg && *cluster) {
344 if ((*sg)->length + nbytes > queue_max_segment_size(q))
347 if (!BIOVEC_PHYS_MERGEABLE(bvprv, bvec))
349 if (!BIOVEC_SEG_BOUNDARY(q, bvprv, bvec))
352 (*sg)->length += nbytes;
359 * If the driver previously mapped a shorter
360 * list, we could see a termination bit
361 * prematurely unless it fully inits the sg
362 * table on each mapping. We KNOW that there
363 * must be more entries here or the driver
364 * would be buggy, so force clear the
365 * termination bit to avoid doing a full
366 * sg_init_table() in drivers for each command.
372 sg_set_page(*sg, bvec->bv_page, nbytes, bvec->bv_offset);
378 static int __blk_bios_map_sg(struct request_queue *q, struct bio *bio,
379 struct scatterlist *sglist,
380 struct scatterlist **sg)
382 struct bio_vec bvec, bvprv = { NULL };
383 struct bvec_iter iter;
387 cluster = blk_queue_cluster(q);
389 if (bio->bi_rw & REQ_DISCARD) {
391 * This is a hack - drivers should be neither modifying the
392 * biovec, nor relying on bi_vcnt - but because of
393 * blk_add_request_payload(), a discard bio may or may not have
394 * a payload we need to set up here (thank you Christoph) and
395 * bi_vcnt is really the only way of telling if we need to.
404 if (bio->bi_rw & REQ_WRITE_SAME) {
407 bvec = bio_iovec(bio);
408 sg_set_page(*sg, bvec.bv_page, bvec.bv_len, bvec.bv_offset);
413 bio_for_each_segment(bvec, bio, iter)
414 __blk_segment_map_sg(q, &bvec, sglist, &bvprv, sg,
421 * map a request to scatterlist, return number of sg entries setup. Caller
422 * must make sure sg can hold rq->nr_phys_segments entries
424 int blk_rq_map_sg(struct request_queue *q, struct request *rq,
425 struct scatterlist *sglist)
427 struct scatterlist *sg = NULL;
431 nsegs = __blk_bios_map_sg(q, rq->bio, sglist, &sg);
433 if (unlikely(rq->cmd_flags & REQ_COPY_USER) &&
434 (blk_rq_bytes(rq) & q->dma_pad_mask)) {
435 unsigned int pad_len =
436 (q->dma_pad_mask & ~blk_rq_bytes(rq)) + 1;
438 sg->length += pad_len;
439 rq->extra_len += pad_len;
442 if (q->dma_drain_size && q->dma_drain_needed(rq)) {
443 if (rq->cmd_flags & REQ_WRITE)
444 memset(q->dma_drain_buffer, 0, q->dma_drain_size);
448 sg_set_page(sg, virt_to_page(q->dma_drain_buffer),
450 ((unsigned long)q->dma_drain_buffer) &
453 rq->extra_len += q->dma_drain_size;
460 * Something must have been wrong if the figured number of
461 * segment is bigger than number of req's physical segments
463 WARN_ON(nsegs > rq->nr_phys_segments);
467 EXPORT_SYMBOL(blk_rq_map_sg);
469 static inline int ll_new_hw_segment(struct request_queue *q,
473 int nr_phys_segs = bio_phys_segments(q, bio);
475 if (req->nr_phys_segments + nr_phys_segs > queue_max_segments(q))
478 if (blk_integrity_merge_bio(q, req, bio) == false)
482 * This will form the start of a new hw segment. Bump both
485 req->nr_phys_segments += nr_phys_segs;
489 req->cmd_flags |= REQ_NOMERGE;
490 if (req == q->last_merge)
491 q->last_merge = NULL;
495 int ll_back_merge_fn(struct request_queue *q, struct request *req,
498 if (req_gap_back_merge(req, bio))
500 if (blk_integrity_rq(req) &&
501 integrity_req_gap_back_merge(req, bio))
503 if (blk_rq_sectors(req) + bio_sectors(bio) >
504 blk_rq_get_max_sectors(req)) {
505 req->cmd_flags |= REQ_NOMERGE;
506 if (req == q->last_merge)
507 q->last_merge = NULL;
510 if (!bio_flagged(req->biotail, BIO_SEG_VALID))
511 blk_recount_segments(q, req->biotail);
512 if (!bio_flagged(bio, BIO_SEG_VALID))
513 blk_recount_segments(q, bio);
515 return ll_new_hw_segment(q, req, bio);
518 int ll_front_merge_fn(struct request_queue *q, struct request *req,
522 if (req_gap_front_merge(req, bio))
524 if (blk_integrity_rq(req) &&
525 integrity_req_gap_front_merge(req, bio))
527 if (blk_rq_sectors(req) + bio_sectors(bio) >
528 blk_rq_get_max_sectors(req)) {
529 req->cmd_flags |= REQ_NOMERGE;
530 if (req == q->last_merge)
531 q->last_merge = NULL;
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 int ll_merge_requests_fn(struct request_queue *q, struct request *req,
554 struct request *next)
556 int total_phys_segments;
557 unsigned int seg_size =
558 req->biotail->bi_seg_back_size + next->bio->bi_seg_front_size;
561 * First check if the either of the requests are re-queued
562 * requests. Can't merge them if they are.
564 if (req_no_special_merge(req) || req_no_special_merge(next))
567 if (req_gap_back_merge(req, next->bio))
571 * Will it become too large?
573 if ((blk_rq_sectors(req) + blk_rq_sectors(next)) >
574 blk_rq_get_max_sectors(req))
577 total_phys_segments = req->nr_phys_segments + next->nr_phys_segments;
578 if (blk_phys_contig_segment(q, req->biotail, next->bio)) {
579 if (req->nr_phys_segments == 1)
580 req->bio->bi_seg_front_size = seg_size;
581 if (next->nr_phys_segments == 1)
582 next->biotail->bi_seg_back_size = seg_size;
583 total_phys_segments--;
586 if (total_phys_segments > queue_max_segments(q))
589 if (blk_integrity_merge_rq(q, req, next) == false)
593 req->nr_phys_segments = total_phys_segments;
598 * blk_rq_set_mixed_merge - mark a request as mixed merge
599 * @rq: request to mark as mixed merge
602 * @rq is about to be mixed merged. Make sure the attributes
603 * which can be mixed are set in each bio and mark @rq as mixed
606 void blk_rq_set_mixed_merge(struct request *rq)
608 unsigned int ff = rq->cmd_flags & REQ_FAILFAST_MASK;
611 if (rq->cmd_flags & REQ_MIXED_MERGE)
615 * @rq will no longer represent mixable attributes for all the
616 * contained bios. It will just track those of the first one.
617 * Distributes the attributs to each bio.
619 for (bio = rq->bio; bio; bio = bio->bi_next) {
620 WARN_ON_ONCE((bio->bi_rw & REQ_FAILFAST_MASK) &&
621 (bio->bi_rw & REQ_FAILFAST_MASK) != ff);
624 rq->cmd_flags |= REQ_MIXED_MERGE;
627 static void blk_account_io_merge(struct request *req)
629 if (blk_do_io_stat(req)) {
630 struct hd_struct *part;
633 cpu = part_stat_lock();
636 part_round_stats(cpu, part);
637 part_dec_in_flight(part, rq_data_dir(req));
645 * Has to be called with the request spinlock acquired
647 static int attempt_merge(struct request_queue *q, struct request *req,
648 struct request *next)
650 if (!rq_mergeable(req) || !rq_mergeable(next))
653 if (!blk_check_merge_flags(req->cmd_flags, next->cmd_flags))
659 if (blk_rq_pos(req) + blk_rq_sectors(req) != blk_rq_pos(next))
662 if (rq_data_dir(req) != rq_data_dir(next)
663 || req->rq_disk != next->rq_disk
664 || req_no_special_merge(next))
667 if (req->cmd_flags & REQ_WRITE_SAME &&
668 !blk_write_same_mergeable(req->bio, next->bio))
672 * If we are allowed to merge, then append bio list
673 * from next to rq and release next. merge_requests_fn
674 * will have updated segment counts, update sector
677 if (!ll_merge_requests_fn(q, req, next))
681 * If failfast settings disagree or any of the two is already
682 * a mixed merge, mark both as mixed before proceeding. This
683 * makes sure that all involved bios have mixable attributes
686 if ((req->cmd_flags | next->cmd_flags) & REQ_MIXED_MERGE ||
687 (req->cmd_flags & REQ_FAILFAST_MASK) !=
688 (next->cmd_flags & REQ_FAILFAST_MASK)) {
689 blk_rq_set_mixed_merge(req);
690 blk_rq_set_mixed_merge(next);
694 * At this point we have either done a back merge
695 * or front merge. We need the smaller start_time of
696 * the merged requests to be the current request
697 * for accounting purposes.
699 if (time_after(req->start_time, next->start_time))
700 req->start_time = next->start_time;
702 req->biotail->bi_next = next->bio;
703 req->biotail = next->biotail;
705 req->__data_len += blk_rq_bytes(next);
707 elv_merge_requests(q, req, next);
710 * 'next' is going away, so update stats accordingly
712 blk_account_io_merge(next);
714 req->ioprio = ioprio_best(req->ioprio, next->ioprio);
715 if (blk_rq_cpu_valid(next))
716 req->cpu = next->cpu;
718 /* owner-ship of bio passed from next to req */
720 __blk_put_request(q, next);
724 int attempt_back_merge(struct request_queue *q, struct request *rq)
726 struct request *next = elv_latter_request(q, rq);
729 return attempt_merge(q, rq, next);
734 int attempt_front_merge(struct request_queue *q, struct request *rq)
736 struct request *prev = elv_former_request(q, rq);
739 return attempt_merge(q, prev, rq);
744 int blk_attempt_req_merge(struct request_queue *q, struct request *rq,
745 struct request *next)
747 return attempt_merge(q, rq, next);
750 bool blk_rq_merge_ok(struct request *rq, struct bio *bio)
752 if (!rq_mergeable(rq) || !bio_mergeable(bio))
755 if (!blk_check_merge_flags(rq->cmd_flags, bio->bi_rw))
758 /* different data direction or already started, don't merge */
759 if (bio_data_dir(bio) != rq_data_dir(rq))
762 /* must be same device and not a special request */
763 if (rq->rq_disk != bio->bi_bdev->bd_disk || req_no_special_merge(rq))
766 /* only merge integrity protected bio into ditto rq */
767 if (blk_integrity_merge_bio(rq->q, rq, bio) == false)
770 /* must be using the same buffer */
771 if (rq->cmd_flags & REQ_WRITE_SAME &&
772 !blk_write_same_mergeable(rq->bio, bio))
778 int blk_try_merge(struct request *rq, struct bio *bio)
780 if (blk_rq_pos(rq) + blk_rq_sectors(rq) == bio->bi_iter.bi_sector)
781 return ELEVATOR_BACK_MERGE;
782 else if (blk_rq_pos(rq) - bio_sectors(bio) == bio->bi_iter.bi_sector)
783 return ELEVATOR_FRONT_MERGE;
784 return ELEVATOR_NO_MERGE;