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>
10 #include <trace/events/block.h>
14 static struct bio *blk_bio_discard_split(struct request_queue *q,
19 unsigned int max_discard_sectors, granularity;
22 unsigned split_sectors;
26 /* Zero-sector (unknown) and one-sector granularities are the same. */
27 granularity = max(q->limits.discard_granularity >> 9, 1U);
29 max_discard_sectors = min(q->limits.max_discard_sectors, UINT_MAX >> 9);
30 max_discard_sectors -= max_discard_sectors % granularity;
32 if (unlikely(!max_discard_sectors)) {
37 if (bio_sectors(bio) <= max_discard_sectors)
40 split_sectors = max_discard_sectors;
43 * If the next starting sector would be misaligned, stop the discard at
44 * the previous aligned sector.
46 alignment = (q->limits.discard_alignment >> 9) % granularity;
48 tmp = bio->bi_iter.bi_sector + split_sectors - alignment;
49 tmp = sector_div(tmp, granularity);
51 if (split_sectors > tmp)
54 return bio_split(bio, split_sectors, GFP_NOIO, bs);
57 static struct bio *blk_bio_write_same_split(struct request_queue *q,
64 if (!q->limits.max_write_same_sectors)
67 if (bio_sectors(bio) <= q->limits.max_write_same_sectors)
70 return bio_split(bio, q->limits.max_write_same_sectors, GFP_NOIO, bs);
73 static inline unsigned get_max_io_size(struct request_queue *q,
76 unsigned sectors = blk_max_size_offset(q, bio->bi_iter.bi_sector);
77 unsigned mask = queue_logical_block_size(q) - 1;
79 /* aligned to logical block size */
80 sectors &= ~(mask >> 9);
85 static struct bio *blk_bio_segment_split(struct request_queue *q,
90 struct bio_vec bv, bvprv, *bvprvp = NULL;
91 struct bvec_iter iter;
92 unsigned seg_size = 0, nsegs = 0, sectors = 0;
93 unsigned front_seg_size = bio->bi_seg_front_size;
95 struct bio *new = NULL;
96 const unsigned max_sectors = get_max_io_size(q, bio);
99 bio_for_each_segment(bv, bio, iter) {
101 * With arbitrary bio size, the incoming bio may be very
102 * big. We have to split the bio into small bios so that
103 * each holds at most BIO_MAX_PAGES bvecs because
104 * bio_clone() can fail to allocate big bvecs.
106 * It should have been better to apply the limit per
107 * request queue in which bio_clone() is involved,
108 * instead of globally. The biggest blocker is the
109 * bio_clone() in bio bounce.
111 * If bio is splitted by this reason, we should have
112 * allowed to continue bios merging, but don't do
113 * that now for making the change simple.
115 * TODO: deal with bio bounce's bio_clone() gracefully
116 * and convert the global limit into per-queue limit.
118 if (bvecs++ >= BIO_MAX_PAGES)
122 * If the queue doesn't support SG gaps and adding this
123 * offset would create a gap, disallow it.
125 if (bvprvp && bvec_gap_to_prev(q, bvprvp, bv.bv_offset))
128 if (sectors + (bv.bv_len >> 9) > max_sectors) {
130 * Consider this a new segment if we're splitting in
131 * the middle of this vector.
133 if (nsegs < queue_max_segments(q) &&
134 sectors < max_sectors) {
136 sectors = max_sectors;
140 /* Make this single bvec as the 1st segment */
143 if (bvprvp && blk_queue_cluster(q)) {
144 if (seg_size + bv.bv_len > queue_max_segment_size(q))
146 if (!BIOVEC_PHYS_MERGEABLE(bvprvp, &bv))
148 if (!BIOVEC_SEG_BOUNDARY(q, bvprvp, &bv))
151 seg_size += bv.bv_len;
154 sectors += bv.bv_len >> 9;
156 if (nsegs == 1 && seg_size > front_seg_size)
157 front_seg_size = seg_size;
161 if (nsegs == queue_max_segments(q))
167 seg_size = bv.bv_len;
168 sectors += bv.bv_len >> 9;
170 if (nsegs == 1 && seg_size > front_seg_size)
171 front_seg_size = seg_size;
179 new = bio_split(bio, sectors, GFP_NOIO, bs);
184 bio->bi_seg_front_size = front_seg_size;
185 if (seg_size > bio->bi_seg_back_size)
186 bio->bi_seg_back_size = seg_size;
188 return do_split ? new : NULL;
191 void blk_queue_split(struct request_queue *q, struct bio **bio,
194 struct bio *split, *res;
197 switch (bio_op(*bio)) {
199 case REQ_OP_SECURE_ERASE:
200 split = blk_bio_discard_split(q, *bio, bs, &nsegs);
202 case REQ_OP_WRITE_SAME:
203 split = blk_bio_write_same_split(q, *bio, bs, &nsegs);
206 split = blk_bio_segment_split(q, *bio, q->bio_split, &nsegs);
210 /* physical segments can be figured out during splitting */
211 res = split ? split : *bio;
212 res->bi_phys_segments = nsegs;
213 bio_set_flag(res, BIO_SEG_VALID);
216 /* there isn't chance to merge the splitted bio */
217 split->bi_opf |= REQ_NOMERGE;
219 bio_chain(split, *bio);
220 trace_block_split(q, split, (*bio)->bi_iter.bi_sector);
221 generic_make_request(*bio);
225 EXPORT_SYMBOL(blk_queue_split);
227 static unsigned int __blk_recalc_rq_segments(struct request_queue *q,
231 struct bio_vec bv, bvprv = { NULL };
232 int cluster, prev = 0;
233 unsigned int seg_size, nr_phys_segs;
234 struct bio *fbio, *bbio;
235 struct bvec_iter iter;
241 * This should probably be returning 0, but blk_add_request_payload()
244 if (bio_op(bio) == REQ_OP_DISCARD || bio_op(bio) == REQ_OP_SECURE_ERASE)
247 if (bio_op(bio) == REQ_OP_WRITE_SAME)
251 cluster = blk_queue_cluster(q);
255 bio_for_each_segment(bv, bio, iter) {
257 * If SG merging is disabled, each bio vector is
263 if (prev && cluster) {
264 if (seg_size + bv.bv_len
265 > queue_max_segment_size(q))
267 if (!BIOVEC_PHYS_MERGEABLE(&bvprv, &bv))
269 if (!BIOVEC_SEG_BOUNDARY(q, &bvprv, &bv))
272 seg_size += bv.bv_len;
277 if (nr_phys_segs == 1 && seg_size >
278 fbio->bi_seg_front_size)
279 fbio->bi_seg_front_size = seg_size;
284 seg_size = bv.bv_len;
289 if (nr_phys_segs == 1 && seg_size > fbio->bi_seg_front_size)
290 fbio->bi_seg_front_size = seg_size;
291 if (seg_size > bbio->bi_seg_back_size)
292 bbio->bi_seg_back_size = seg_size;
297 void blk_recalc_rq_segments(struct request *rq)
299 bool no_sg_merge = !!test_bit(QUEUE_FLAG_NO_SG_MERGE,
300 &rq->q->queue_flags);
302 rq->nr_phys_segments = __blk_recalc_rq_segments(rq->q, rq->bio,
306 void blk_recount_segments(struct request_queue *q, struct bio *bio)
308 unsigned short seg_cnt;
310 /* estimate segment number by bi_vcnt for non-cloned bio */
311 if (bio_flagged(bio, BIO_CLONED))
312 seg_cnt = bio_segments(bio);
314 seg_cnt = bio->bi_vcnt;
316 if (test_bit(QUEUE_FLAG_NO_SG_MERGE, &q->queue_flags) &&
317 (seg_cnt < queue_max_segments(q)))
318 bio->bi_phys_segments = seg_cnt;
320 struct bio *nxt = bio->bi_next;
323 bio->bi_phys_segments = __blk_recalc_rq_segments(q, bio, false);
327 bio_set_flag(bio, BIO_SEG_VALID);
329 EXPORT_SYMBOL(blk_recount_segments);
331 static int blk_phys_contig_segment(struct request_queue *q, struct bio *bio,
334 struct bio_vec end_bv = { NULL }, nxt_bv;
336 if (!blk_queue_cluster(q))
339 if (bio->bi_seg_back_size + nxt->bi_seg_front_size >
340 queue_max_segment_size(q))
343 if (!bio_has_data(bio))
346 bio_get_last_bvec(bio, &end_bv);
347 bio_get_first_bvec(nxt, &nxt_bv);
349 if (!BIOVEC_PHYS_MERGEABLE(&end_bv, &nxt_bv))
353 * bio and nxt are contiguous in memory; check if the queue allows
354 * these two to be merged into one
356 if (BIOVEC_SEG_BOUNDARY(q, &end_bv, &nxt_bv))
363 __blk_segment_map_sg(struct request_queue *q, struct bio_vec *bvec,
364 struct scatterlist *sglist, struct bio_vec *bvprv,
365 struct scatterlist **sg, int *nsegs, int *cluster)
368 int nbytes = bvec->bv_len;
370 if (*sg && *cluster) {
371 if ((*sg)->length + nbytes > queue_max_segment_size(q))
374 if (!BIOVEC_PHYS_MERGEABLE(bvprv, bvec))
376 if (!BIOVEC_SEG_BOUNDARY(q, bvprv, bvec))
379 (*sg)->length += nbytes;
386 * If the driver previously mapped a shorter
387 * list, we could see a termination bit
388 * prematurely unless it fully inits the sg
389 * table on each mapping. We KNOW that there
390 * must be more entries here or the driver
391 * would be buggy, so force clear the
392 * termination bit to avoid doing a full
393 * sg_init_table() in drivers for each command.
399 sg_set_page(*sg, bvec->bv_page, nbytes, bvec->bv_offset);
405 static int __blk_bios_map_sg(struct request_queue *q, struct bio *bio,
406 struct scatterlist *sglist,
407 struct scatterlist **sg)
409 struct bio_vec bvec, bvprv = { NULL };
410 struct bvec_iter iter;
414 cluster = blk_queue_cluster(q);
416 switch (bio_op(bio)) {
418 case REQ_OP_SECURE_ERASE:
420 * This is a hack - drivers should be neither modifying the
421 * biovec, nor relying on bi_vcnt - but because of
422 * blk_add_request_payload(), a discard bio may or may not have
423 * a payload we need to set up here (thank you Christoph) and
424 * bi_vcnt is really the only way of telling if we need to.
429 case REQ_OP_WRITE_SAME:
431 bvec = bio_iovec(bio);
432 sg_set_page(*sg, bvec.bv_page, bvec.bv_len, bvec.bv_offset);
439 bio_for_each_segment(bvec, bio, iter)
440 __blk_segment_map_sg(q, &bvec, sglist, &bvprv, sg,
447 * map a request to scatterlist, return number of sg entries setup. Caller
448 * must make sure sg can hold rq->nr_phys_segments entries
450 int blk_rq_map_sg(struct request_queue *q, struct request *rq,
451 struct scatterlist *sglist)
453 struct scatterlist *sg = NULL;
457 nsegs = __blk_bios_map_sg(q, rq->bio, sglist, &sg);
459 if (unlikely(rq->cmd_flags & REQ_COPY_USER) &&
460 (blk_rq_bytes(rq) & q->dma_pad_mask)) {
461 unsigned int pad_len =
462 (q->dma_pad_mask & ~blk_rq_bytes(rq)) + 1;
464 sg->length += pad_len;
465 rq->extra_len += pad_len;
468 if (q->dma_drain_size && q->dma_drain_needed(rq)) {
469 if (op_is_write(req_op(rq)))
470 memset(q->dma_drain_buffer, 0, q->dma_drain_size);
474 sg_set_page(sg, virt_to_page(q->dma_drain_buffer),
476 ((unsigned long)q->dma_drain_buffer) &
479 rq->extra_len += q->dma_drain_size;
486 * Something must have been wrong if the figured number of
487 * segment is bigger than number of req's physical segments
489 WARN_ON(nsegs > rq->nr_phys_segments);
493 EXPORT_SYMBOL(blk_rq_map_sg);
495 static inline int ll_new_hw_segment(struct request_queue *q,
499 int nr_phys_segs = bio_phys_segments(q, bio);
501 if (req->nr_phys_segments + nr_phys_segs > queue_max_segments(q))
504 if (blk_integrity_merge_bio(q, req, bio) == false)
508 * This will form the start of a new hw segment. Bump both
511 req->nr_phys_segments += nr_phys_segs;
515 req->cmd_flags |= REQ_NOMERGE;
516 if (req == q->last_merge)
517 q->last_merge = NULL;
521 int ll_back_merge_fn(struct request_queue *q, struct request *req,
524 if (req_gap_back_merge(req, bio))
526 if (blk_integrity_rq(req) &&
527 integrity_req_gap_back_merge(req, bio))
529 if (blk_rq_sectors(req) + bio_sectors(bio) >
530 blk_rq_get_max_sectors(req, blk_rq_pos(req))) {
531 req->cmd_flags |= REQ_NOMERGE;
532 if (req == q->last_merge)
533 q->last_merge = NULL;
536 if (!bio_flagged(req->biotail, BIO_SEG_VALID))
537 blk_recount_segments(q, req->biotail);
538 if (!bio_flagged(bio, BIO_SEG_VALID))
539 blk_recount_segments(q, bio);
541 return ll_new_hw_segment(q, req, bio);
544 int ll_front_merge_fn(struct request_queue *q, struct request *req,
548 if (req_gap_front_merge(req, bio))
550 if (blk_integrity_rq(req) &&
551 integrity_req_gap_front_merge(req, bio))
553 if (blk_rq_sectors(req) + bio_sectors(bio) >
554 blk_rq_get_max_sectors(req, bio->bi_iter.bi_sector)) {
555 req->cmd_flags |= REQ_NOMERGE;
556 if (req == q->last_merge)
557 q->last_merge = NULL;
560 if (!bio_flagged(bio, BIO_SEG_VALID))
561 blk_recount_segments(q, bio);
562 if (!bio_flagged(req->bio, BIO_SEG_VALID))
563 blk_recount_segments(q, req->bio);
565 return ll_new_hw_segment(q, req, bio);
569 * blk-mq uses req->special to carry normal driver per-request payload, it
570 * does not indicate a prepared command that we cannot merge with.
572 static bool req_no_special_merge(struct request *req)
574 struct request_queue *q = req->q;
576 return !q->mq_ops && req->special;
579 static int ll_merge_requests_fn(struct request_queue *q, struct request *req,
580 struct request *next)
582 int total_phys_segments;
583 unsigned int seg_size =
584 req->biotail->bi_seg_back_size + next->bio->bi_seg_front_size;
587 * First check if the either of the requests are re-queued
588 * requests. Can't merge them if they are.
590 if (req_no_special_merge(req) || req_no_special_merge(next))
593 if (req_gap_back_merge(req, next->bio))
597 * Will it become too large?
599 if ((blk_rq_sectors(req) + blk_rq_sectors(next)) >
600 blk_rq_get_max_sectors(req, blk_rq_pos(req)))
603 total_phys_segments = req->nr_phys_segments + next->nr_phys_segments;
604 if (blk_phys_contig_segment(q, req->biotail, next->bio)) {
605 if (req->nr_phys_segments == 1)
606 req->bio->bi_seg_front_size = seg_size;
607 if (next->nr_phys_segments == 1)
608 next->biotail->bi_seg_back_size = seg_size;
609 total_phys_segments--;
612 if (total_phys_segments > queue_max_segments(q))
615 if (blk_integrity_merge_rq(q, req, next) == false)
619 req->nr_phys_segments = total_phys_segments;
624 * blk_rq_set_mixed_merge - mark a request as mixed merge
625 * @rq: request to mark as mixed merge
628 * @rq is about to be mixed merged. Make sure the attributes
629 * which can be mixed are set in each bio and mark @rq as mixed
632 void blk_rq_set_mixed_merge(struct request *rq)
634 unsigned int ff = rq->cmd_flags & REQ_FAILFAST_MASK;
637 if (rq->cmd_flags & REQ_MIXED_MERGE)
641 * @rq will no longer represent mixable attributes for all the
642 * contained bios. It will just track those of the first one.
643 * Distributes the attributs to each bio.
645 for (bio = rq->bio; bio; bio = bio->bi_next) {
646 WARN_ON_ONCE((bio->bi_opf & REQ_FAILFAST_MASK) &&
647 (bio->bi_opf & REQ_FAILFAST_MASK) != ff);
650 rq->cmd_flags |= REQ_MIXED_MERGE;
653 static void blk_account_io_merge(struct request *req)
655 if (blk_do_io_stat(req)) {
656 struct hd_struct *part;
659 cpu = part_stat_lock();
662 part_round_stats(cpu, part);
663 part_dec_in_flight(part, rq_data_dir(req));
671 * Has to be called with the request spinlock acquired
673 static int attempt_merge(struct request_queue *q, struct request *req,
674 struct request *next)
676 if (!rq_mergeable(req) || !rq_mergeable(next))
679 if (req_op(req) != req_op(next))
685 if (blk_rq_pos(req) + blk_rq_sectors(req) != blk_rq_pos(next))
688 if (rq_data_dir(req) != rq_data_dir(next)
689 || req->rq_disk != next->rq_disk
690 || req_no_special_merge(next))
693 if (req_op(req) == REQ_OP_WRITE_SAME &&
694 !blk_write_same_mergeable(req->bio, next->bio))
698 * If we are allowed to merge, then append bio list
699 * from next to rq and release next. merge_requests_fn
700 * will have updated segment counts, update sector
703 if (!ll_merge_requests_fn(q, req, next))
707 * If failfast settings disagree or any of the two is already
708 * a mixed merge, mark both as mixed before proceeding. This
709 * makes sure that all involved bios have mixable attributes
712 if ((req->cmd_flags | next->cmd_flags) & REQ_MIXED_MERGE ||
713 (req->cmd_flags & REQ_FAILFAST_MASK) !=
714 (next->cmd_flags & REQ_FAILFAST_MASK)) {
715 blk_rq_set_mixed_merge(req);
716 blk_rq_set_mixed_merge(next);
720 * At this point we have either done a back merge
721 * or front merge. We need the smaller start_time of
722 * the merged requests to be the current request
723 * for accounting purposes.
725 if (time_after(req->start_time, next->start_time))
726 req->start_time = next->start_time;
728 req->biotail->bi_next = next->bio;
729 req->biotail = next->biotail;
731 req->__data_len += blk_rq_bytes(next);
733 elv_merge_requests(q, req, next);
736 * 'next' is going away, so update stats accordingly
738 blk_account_io_merge(next);
740 req->ioprio = ioprio_best(req->ioprio, next->ioprio);
741 if (blk_rq_cpu_valid(next))
742 req->cpu = next->cpu;
744 /* owner-ship of bio passed from next to req */
746 __blk_put_request(q, next);
750 int attempt_back_merge(struct request_queue *q, struct request *rq)
752 struct request *next = elv_latter_request(q, rq);
755 return attempt_merge(q, rq, next);
760 int attempt_front_merge(struct request_queue *q, struct request *rq)
762 struct request *prev = elv_former_request(q, rq);
765 return attempt_merge(q, prev, rq);
770 int blk_attempt_req_merge(struct request_queue *q, struct request *rq,
771 struct request *next)
773 struct elevator_queue *e = q->elevator;
775 if (e->type->ops.elevator_allow_rq_merge_fn)
776 if (!e->type->ops.elevator_allow_rq_merge_fn(q, rq, next))
779 return attempt_merge(q, rq, next);
782 bool blk_rq_merge_ok(struct request *rq, struct bio *bio)
784 if (!rq_mergeable(rq) || !bio_mergeable(bio))
787 if (req_op(rq) != bio_op(bio))
790 /* different data direction or already started, don't merge */
791 if (bio_data_dir(bio) != rq_data_dir(rq))
794 /* must be same device and not a special request */
795 if (rq->rq_disk != bio->bi_bdev->bd_disk || req_no_special_merge(rq))
798 /* only merge integrity protected bio into ditto rq */
799 if (blk_integrity_merge_bio(rq->q, rq, bio) == false)
802 /* must be using the same buffer */
803 if (req_op(rq) == REQ_OP_WRITE_SAME &&
804 !blk_write_same_mergeable(rq->bio, bio))
810 int blk_try_merge(struct request *rq, struct bio *bio)
812 if (blk_rq_pos(rq) + blk_rq_sectors(rq) == bio->bi_iter.bi_sector)
813 return ELEVATOR_BACK_MERGE;
814 else if (blk_rq_pos(rq) - bio_sectors(bio) == bio->bi_iter.bi_sector)
815 return ELEVATOR_FRONT_MERGE;
816 return ELEVATOR_NO_MERGE;