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 struct bio *blk_bio_segment_split(struct request_queue *q,
76 struct bio_vec bv, bvprv, *bvprvp = NULL;
77 struct bvec_iter iter;
78 unsigned seg_size = 0, nsegs = 0, sectors = 0;
79 unsigned front_seg_size = bio->bi_seg_front_size;
81 struct bio *new = NULL;
83 bio_for_each_segment(bv, bio, iter) {
84 if (sectors + (bv.bv_len >> 9) > queue_max_sectors(q))
88 * If the queue doesn't support SG gaps and adding this
89 * offset would create a gap, disallow it.
91 if (bvprvp && bvec_gap_to_prev(q, bvprvp, bv.bv_offset))
94 if (bvprvp && blk_queue_cluster(q)) {
95 if (seg_size + bv.bv_len > queue_max_segment_size(q))
97 if (!BIOVEC_PHYS_MERGEABLE(bvprvp, &bv))
99 if (!BIOVEC_SEG_BOUNDARY(q, bvprvp, &bv))
102 seg_size += bv.bv_len;
105 sectors += bv.bv_len >> 9;
107 if (nsegs == 1 && seg_size > front_seg_size)
108 front_seg_size = seg_size;
112 if (nsegs == queue_max_segments(q))
118 seg_size = bv.bv_len;
119 sectors += bv.bv_len >> 9;
121 if (nsegs == 1 && seg_size > front_seg_size)
122 front_seg_size = seg_size;
130 new = bio_split(bio, sectors, GFP_NOIO, bs);
135 bio->bi_seg_front_size = front_seg_size;
136 if (seg_size > bio->bi_seg_back_size)
137 bio->bi_seg_back_size = seg_size;
139 return do_split ? new : NULL;
142 void blk_queue_split(struct request_queue *q, struct bio **bio,
145 struct bio *split, *res;
148 if ((*bio)->bi_rw & REQ_DISCARD)
149 split = blk_bio_discard_split(q, *bio, bs, &nsegs);
150 else if ((*bio)->bi_rw & REQ_WRITE_SAME)
151 split = blk_bio_write_same_split(q, *bio, bs, &nsegs);
153 split = blk_bio_segment_split(q, *bio, q->bio_split, &nsegs);
155 /* physical segments can be figured out during splitting */
156 res = split ? split : *bio;
157 res->bi_phys_segments = nsegs;
158 bio_set_flag(res, BIO_SEG_VALID);
161 /* there isn't chance to merge the splitted bio */
162 split->bi_rw |= REQ_NOMERGE;
164 bio_chain(split, *bio);
165 generic_make_request(*bio);
169 EXPORT_SYMBOL(blk_queue_split);
171 static unsigned int __blk_recalc_rq_segments(struct request_queue *q,
175 struct bio_vec bv, bvprv = { NULL };
176 int cluster, prev = 0;
177 unsigned int seg_size, nr_phys_segs;
178 struct bio *fbio, *bbio;
179 struct bvec_iter iter;
185 * This should probably be returning 0, but blk_add_request_payload()
188 if (bio->bi_rw & REQ_DISCARD)
191 if (bio->bi_rw & REQ_WRITE_SAME)
195 cluster = blk_queue_cluster(q);
199 bio_for_each_segment(bv, bio, iter) {
201 * If SG merging is disabled, each bio vector is
207 if (prev && cluster) {
208 if (seg_size + bv.bv_len
209 > queue_max_segment_size(q))
211 if (!BIOVEC_PHYS_MERGEABLE(&bvprv, &bv))
213 if (!BIOVEC_SEG_BOUNDARY(q, &bvprv, &bv))
216 seg_size += bv.bv_len;
221 if (nr_phys_segs == 1 && seg_size >
222 fbio->bi_seg_front_size)
223 fbio->bi_seg_front_size = seg_size;
228 seg_size = bv.bv_len;
233 if (nr_phys_segs == 1 && seg_size > fbio->bi_seg_front_size)
234 fbio->bi_seg_front_size = seg_size;
235 if (seg_size > bbio->bi_seg_back_size)
236 bbio->bi_seg_back_size = seg_size;
241 void blk_recalc_rq_segments(struct request *rq)
243 bool no_sg_merge = !!test_bit(QUEUE_FLAG_NO_SG_MERGE,
244 &rq->q->queue_flags);
246 rq->nr_phys_segments = __blk_recalc_rq_segments(rq->q, rq->bio,
250 void blk_recount_segments(struct request_queue *q, struct bio *bio)
252 unsigned short seg_cnt;
254 /* estimate segment number by bi_vcnt for non-cloned bio */
255 if (bio_flagged(bio, BIO_CLONED))
256 seg_cnt = bio_segments(bio);
258 seg_cnt = bio->bi_vcnt;
260 if (test_bit(QUEUE_FLAG_NO_SG_MERGE, &q->queue_flags) &&
261 (seg_cnt < queue_max_segments(q)))
262 bio->bi_phys_segments = seg_cnt;
264 struct bio *nxt = bio->bi_next;
267 bio->bi_phys_segments = __blk_recalc_rq_segments(q, bio, false);
271 bio_set_flag(bio, BIO_SEG_VALID);
273 EXPORT_SYMBOL(blk_recount_segments);
275 static int blk_phys_contig_segment(struct request_queue *q, struct bio *bio,
278 struct bio_vec end_bv = { NULL }, nxt_bv;
279 struct bvec_iter iter;
281 if (!blk_queue_cluster(q))
284 if (bio->bi_seg_back_size + nxt->bi_seg_front_size >
285 queue_max_segment_size(q))
288 if (!bio_has_data(bio))
291 bio_for_each_segment(end_bv, bio, iter)
292 if (end_bv.bv_len == iter.bi_size)
295 nxt_bv = bio_iovec(nxt);
297 if (!BIOVEC_PHYS_MERGEABLE(&end_bv, &nxt_bv))
301 * bio and nxt are contiguous in memory; check if the queue allows
302 * these two to be merged into one
304 if (BIOVEC_SEG_BOUNDARY(q, &end_bv, &nxt_bv))
311 __blk_segment_map_sg(struct request_queue *q, struct bio_vec *bvec,
312 struct scatterlist *sglist, struct bio_vec *bvprv,
313 struct scatterlist **sg, int *nsegs, int *cluster)
316 int nbytes = bvec->bv_len;
318 if (*sg && *cluster) {
319 if ((*sg)->length + nbytes > queue_max_segment_size(q))
322 if (!BIOVEC_PHYS_MERGEABLE(bvprv, bvec))
324 if (!BIOVEC_SEG_BOUNDARY(q, bvprv, bvec))
327 (*sg)->length += nbytes;
334 * If the driver previously mapped a shorter
335 * list, we could see a termination bit
336 * prematurely unless it fully inits the sg
337 * table on each mapping. We KNOW that there
338 * must be more entries here or the driver
339 * would be buggy, so force clear the
340 * termination bit to avoid doing a full
341 * sg_init_table() in drivers for each command.
347 sg_set_page(*sg, bvec->bv_page, nbytes, bvec->bv_offset);
353 static int __blk_bios_map_sg(struct request_queue *q, struct bio *bio,
354 struct scatterlist *sglist,
355 struct scatterlist **sg)
357 struct bio_vec bvec, bvprv = { NULL };
358 struct bvec_iter iter;
362 cluster = blk_queue_cluster(q);
364 if (bio->bi_rw & REQ_DISCARD) {
366 * This is a hack - drivers should be neither modifying the
367 * biovec, nor relying on bi_vcnt - but because of
368 * blk_add_request_payload(), a discard bio may or may not have
369 * a payload we need to set up here (thank you Christoph) and
370 * bi_vcnt is really the only way of telling if we need to.
379 if (bio->bi_rw & REQ_WRITE_SAME) {
382 bvec = bio_iovec(bio);
383 sg_set_page(*sg, bvec.bv_page, bvec.bv_len, bvec.bv_offset);
388 bio_for_each_segment(bvec, bio, iter)
389 __blk_segment_map_sg(q, &bvec, sglist, &bvprv, sg,
396 * map a request to scatterlist, return number of sg entries setup. Caller
397 * must make sure sg can hold rq->nr_phys_segments entries
399 int blk_rq_map_sg(struct request_queue *q, struct request *rq,
400 struct scatterlist *sglist)
402 struct scatterlist *sg = NULL;
406 nsegs = __blk_bios_map_sg(q, rq->bio, sglist, &sg);
408 if (unlikely(rq->cmd_flags & REQ_COPY_USER) &&
409 (blk_rq_bytes(rq) & q->dma_pad_mask)) {
410 unsigned int pad_len =
411 (q->dma_pad_mask & ~blk_rq_bytes(rq)) + 1;
413 sg->length += pad_len;
414 rq->extra_len += pad_len;
417 if (q->dma_drain_size && q->dma_drain_needed(rq)) {
418 if (rq->cmd_flags & REQ_WRITE)
419 memset(q->dma_drain_buffer, 0, q->dma_drain_size);
423 sg_set_page(sg, virt_to_page(q->dma_drain_buffer),
425 ((unsigned long)q->dma_drain_buffer) &
428 rq->extra_len += q->dma_drain_size;
435 * Something must have been wrong if the figured number of
436 * segment is bigger than number of req's physical segments
438 WARN_ON(nsegs > rq->nr_phys_segments);
442 EXPORT_SYMBOL(blk_rq_map_sg);
444 static inline int ll_new_hw_segment(struct request_queue *q,
448 int nr_phys_segs = bio_phys_segments(q, bio);
450 if (req->nr_phys_segments + nr_phys_segs > queue_max_segments(q))
453 if (blk_integrity_merge_bio(q, req, bio) == false)
457 * This will form the start of a new hw segment. Bump both
460 req->nr_phys_segments += nr_phys_segs;
464 req->cmd_flags |= REQ_NOMERGE;
465 if (req == q->last_merge)
466 q->last_merge = NULL;
470 int ll_back_merge_fn(struct request_queue *q, struct request *req,
473 if (req_gap_back_merge(req, bio))
475 if (blk_integrity_rq(req) &&
476 integrity_req_gap_back_merge(req, bio))
478 if (blk_rq_sectors(req) + bio_sectors(bio) >
479 blk_rq_get_max_sectors(req)) {
480 req->cmd_flags |= REQ_NOMERGE;
481 if (req == q->last_merge)
482 q->last_merge = NULL;
485 if (!bio_flagged(req->biotail, BIO_SEG_VALID))
486 blk_recount_segments(q, req->biotail);
487 if (!bio_flagged(bio, BIO_SEG_VALID))
488 blk_recount_segments(q, bio);
490 return ll_new_hw_segment(q, req, bio);
493 int ll_front_merge_fn(struct request_queue *q, struct request *req,
497 if (req_gap_front_merge(req, bio))
499 if (blk_integrity_rq(req) &&
500 integrity_req_gap_front_merge(req, bio))
502 if (blk_rq_sectors(req) + bio_sectors(bio) >
503 blk_rq_get_max_sectors(req)) {
504 req->cmd_flags |= REQ_NOMERGE;
505 if (req == q->last_merge)
506 q->last_merge = NULL;
509 if (!bio_flagged(bio, BIO_SEG_VALID))
510 blk_recount_segments(q, bio);
511 if (!bio_flagged(req->bio, BIO_SEG_VALID))
512 blk_recount_segments(q, req->bio);
514 return ll_new_hw_segment(q, req, bio);
518 * blk-mq uses req->special to carry normal driver per-request payload, it
519 * does not indicate a prepared command that we cannot merge with.
521 static bool req_no_special_merge(struct request *req)
523 struct request_queue *q = req->q;
525 return !q->mq_ops && req->special;
528 static int ll_merge_requests_fn(struct request_queue *q, struct request *req,
529 struct request *next)
531 int total_phys_segments;
532 unsigned int seg_size =
533 req->biotail->bi_seg_back_size + next->bio->bi_seg_front_size;
536 * First check if the either of the requests are re-queued
537 * requests. Can't merge them if they are.
539 if (req_no_special_merge(req) || req_no_special_merge(next))
542 if (req_gap_back_merge(req, next->bio))
546 * Will it become too large?
548 if ((blk_rq_sectors(req) + blk_rq_sectors(next)) >
549 blk_rq_get_max_sectors(req))
552 total_phys_segments = req->nr_phys_segments + next->nr_phys_segments;
553 if (blk_phys_contig_segment(q, req->biotail, next->bio)) {
554 if (req->nr_phys_segments == 1)
555 req->bio->bi_seg_front_size = seg_size;
556 if (next->nr_phys_segments == 1)
557 next->biotail->bi_seg_back_size = seg_size;
558 total_phys_segments--;
561 if (total_phys_segments > queue_max_segments(q))
564 if (blk_integrity_merge_rq(q, req, next) == false)
568 req->nr_phys_segments = total_phys_segments;
573 * blk_rq_set_mixed_merge - mark a request as mixed merge
574 * @rq: request to mark as mixed merge
577 * @rq is about to be mixed merged. Make sure the attributes
578 * which can be mixed are set in each bio and mark @rq as mixed
581 void blk_rq_set_mixed_merge(struct request *rq)
583 unsigned int ff = rq->cmd_flags & REQ_FAILFAST_MASK;
586 if (rq->cmd_flags & REQ_MIXED_MERGE)
590 * @rq will no longer represent mixable attributes for all the
591 * contained bios. It will just track those of the first one.
592 * Distributes the attributs to each bio.
594 for (bio = rq->bio; bio; bio = bio->bi_next) {
595 WARN_ON_ONCE((bio->bi_rw & REQ_FAILFAST_MASK) &&
596 (bio->bi_rw & REQ_FAILFAST_MASK) != ff);
599 rq->cmd_flags |= REQ_MIXED_MERGE;
602 static void blk_account_io_merge(struct request *req)
604 if (blk_do_io_stat(req)) {
605 struct hd_struct *part;
608 cpu = part_stat_lock();
611 part_round_stats(cpu, part);
612 part_dec_in_flight(part, rq_data_dir(req));
620 * Has to be called with the request spinlock acquired
622 static int attempt_merge(struct request_queue *q, struct request *req,
623 struct request *next)
625 if (!rq_mergeable(req) || !rq_mergeable(next))
628 if (!blk_check_merge_flags(req->cmd_flags, next->cmd_flags))
634 if (blk_rq_pos(req) + blk_rq_sectors(req) != blk_rq_pos(next))
637 if (rq_data_dir(req) != rq_data_dir(next)
638 || req->rq_disk != next->rq_disk
639 || req_no_special_merge(next))
642 if (req->cmd_flags & REQ_WRITE_SAME &&
643 !blk_write_same_mergeable(req->bio, next->bio))
647 * If we are allowed to merge, then append bio list
648 * from next to rq and release next. merge_requests_fn
649 * will have updated segment counts, update sector
652 if (!ll_merge_requests_fn(q, req, next))
656 * If failfast settings disagree or any of the two is already
657 * a mixed merge, mark both as mixed before proceeding. This
658 * makes sure that all involved bios have mixable attributes
661 if ((req->cmd_flags | next->cmd_flags) & REQ_MIXED_MERGE ||
662 (req->cmd_flags & REQ_FAILFAST_MASK) !=
663 (next->cmd_flags & REQ_FAILFAST_MASK)) {
664 blk_rq_set_mixed_merge(req);
665 blk_rq_set_mixed_merge(next);
669 * At this point we have either done a back merge
670 * or front merge. We need the smaller start_time of
671 * the merged requests to be the current request
672 * for accounting purposes.
674 if (time_after(req->start_time, next->start_time))
675 req->start_time = next->start_time;
677 req->biotail->bi_next = next->bio;
678 req->biotail = next->biotail;
680 req->__data_len += blk_rq_bytes(next);
682 elv_merge_requests(q, req, next);
685 * 'next' is going away, so update stats accordingly
687 blk_account_io_merge(next);
689 req->ioprio = ioprio_best(req->ioprio, next->ioprio);
690 if (blk_rq_cpu_valid(next))
691 req->cpu = next->cpu;
693 /* owner-ship of bio passed from next to req */
695 __blk_put_request(q, next);
699 int attempt_back_merge(struct request_queue *q, struct request *rq)
701 struct request *next = elv_latter_request(q, rq);
704 return attempt_merge(q, rq, next);
709 int attempt_front_merge(struct request_queue *q, struct request *rq)
711 struct request *prev = elv_former_request(q, rq);
714 return attempt_merge(q, prev, rq);
719 int blk_attempt_req_merge(struct request_queue *q, struct request *rq,
720 struct request *next)
722 return attempt_merge(q, rq, next);
725 bool blk_rq_merge_ok(struct request *rq, struct bio *bio)
727 if (!rq_mergeable(rq) || !bio_mergeable(bio))
730 if (!blk_check_merge_flags(rq->cmd_flags, bio->bi_rw))
733 /* different data direction or already started, don't merge */
734 if (bio_data_dir(bio) != rq_data_dir(rq))
737 /* must be same device and not a special request */
738 if (rq->rq_disk != bio->bi_bdev->bd_disk || req_no_special_merge(rq))
741 /* only merge integrity protected bio into ditto rq */
742 if (blk_integrity_merge_bio(rq->q, rq, bio) == false)
745 /* must be using the same buffer */
746 if (rq->cmd_flags & REQ_WRITE_SAME &&
747 !blk_write_same_mergeable(rq->bio, bio))
753 int blk_try_merge(struct request *rq, struct bio *bio)
755 if (blk_rq_pos(rq) + blk_rq_sectors(rq) == bio->bi_iter.bi_sector)
756 return ELEVATOR_BACK_MERGE;
757 else if (blk_rq_pos(rq) - bio_sectors(bio) == bio->bi_iter.bi_sector)
758 return ELEVATOR_FRONT_MERGE;
759 return ELEVATOR_NO_MERGE;