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;
109 if (nsegs == queue_max_segments(q))
115 seg_size = bv.bv_len;
116 sectors += bv.bv_len >> 9;
118 if (nsegs == 1 && seg_size > front_seg_size)
119 front_seg_size = seg_size;
127 new = bio_split(bio, sectors, GFP_NOIO, bs);
132 bio->bi_seg_front_size = front_seg_size;
133 if (seg_size > bio->bi_seg_back_size)
134 bio->bi_seg_back_size = seg_size;
136 return do_split ? new : NULL;
139 void blk_queue_split(struct request_queue *q, struct bio **bio,
142 struct bio *split, *res;
145 if ((*bio)->bi_rw & REQ_DISCARD)
146 split = blk_bio_discard_split(q, *bio, bs, &nsegs);
147 else if ((*bio)->bi_rw & REQ_WRITE_SAME)
148 split = blk_bio_write_same_split(q, *bio, bs, &nsegs);
150 split = blk_bio_segment_split(q, *bio, q->bio_split, &nsegs);
152 /* physical segments can be figured out during splitting */
153 res = split ? split : *bio;
154 res->bi_phys_segments = nsegs;
155 bio_set_flag(res, BIO_SEG_VALID);
158 /* there isn't chance to merge the splitted bio */
159 split->bi_rw |= REQ_NOMERGE;
161 bio_chain(split, *bio);
162 generic_make_request(*bio);
166 EXPORT_SYMBOL(blk_queue_split);
168 static unsigned int __blk_recalc_rq_segments(struct request_queue *q,
172 struct bio_vec bv, bvprv = { NULL };
173 int cluster, prev = 0;
174 unsigned int seg_size, nr_phys_segs;
175 struct bio *fbio, *bbio;
176 struct bvec_iter iter;
182 * This should probably be returning 0, but blk_add_request_payload()
185 if (bio->bi_rw & REQ_DISCARD)
188 if (bio->bi_rw & REQ_WRITE_SAME)
192 cluster = blk_queue_cluster(q);
196 bio_for_each_segment(bv, bio, iter) {
198 * If SG merging is disabled, each bio vector is
204 if (prev && cluster) {
205 if (seg_size + bv.bv_len
206 > queue_max_segment_size(q))
208 if (!BIOVEC_PHYS_MERGEABLE(&bvprv, &bv))
210 if (!BIOVEC_SEG_BOUNDARY(q, &bvprv, &bv))
213 seg_size += bv.bv_len;
218 if (nr_phys_segs == 1 && seg_size >
219 fbio->bi_seg_front_size)
220 fbio->bi_seg_front_size = seg_size;
225 seg_size = bv.bv_len;
230 if (nr_phys_segs == 1 && seg_size > fbio->bi_seg_front_size)
231 fbio->bi_seg_front_size = seg_size;
232 if (seg_size > bbio->bi_seg_back_size)
233 bbio->bi_seg_back_size = seg_size;
238 void blk_recalc_rq_segments(struct request *rq)
240 bool no_sg_merge = !!test_bit(QUEUE_FLAG_NO_SG_MERGE,
241 &rq->q->queue_flags);
243 rq->nr_phys_segments = __blk_recalc_rq_segments(rq->q, rq->bio,
247 void blk_recount_segments(struct request_queue *q, struct bio *bio)
249 unsigned short seg_cnt;
251 /* estimate segment number by bi_vcnt for non-cloned bio */
252 if (bio_flagged(bio, BIO_CLONED))
253 seg_cnt = bio_segments(bio);
255 seg_cnt = bio->bi_vcnt;
257 if (test_bit(QUEUE_FLAG_NO_SG_MERGE, &q->queue_flags) &&
258 (seg_cnt < queue_max_segments(q)))
259 bio->bi_phys_segments = seg_cnt;
261 struct bio *nxt = bio->bi_next;
264 bio->bi_phys_segments = __blk_recalc_rq_segments(q, bio, false);
268 bio_set_flag(bio, BIO_SEG_VALID);
270 EXPORT_SYMBOL(blk_recount_segments);
272 static int blk_phys_contig_segment(struct request_queue *q, struct bio *bio,
275 struct bio_vec end_bv = { NULL }, nxt_bv;
276 struct bvec_iter iter;
278 if (!blk_queue_cluster(q))
281 if (bio->bi_seg_back_size + nxt->bi_seg_front_size >
282 queue_max_segment_size(q))
285 if (!bio_has_data(bio))
288 bio_for_each_segment(end_bv, bio, iter)
289 if (end_bv.bv_len == iter.bi_size)
292 nxt_bv = bio_iovec(nxt);
294 if (!BIOVEC_PHYS_MERGEABLE(&end_bv, &nxt_bv))
298 * bio and nxt are contiguous in memory; check if the queue allows
299 * these two to be merged into one
301 if (BIOVEC_SEG_BOUNDARY(q, &end_bv, &nxt_bv))
308 __blk_segment_map_sg(struct request_queue *q, struct bio_vec *bvec,
309 struct scatterlist *sglist, struct bio_vec *bvprv,
310 struct scatterlist **sg, int *nsegs, int *cluster)
313 int nbytes = bvec->bv_len;
315 if (*sg && *cluster) {
316 if ((*sg)->length + nbytes > queue_max_segment_size(q))
319 if (!BIOVEC_PHYS_MERGEABLE(bvprv, bvec))
321 if (!BIOVEC_SEG_BOUNDARY(q, bvprv, bvec))
324 (*sg)->length += nbytes;
331 * If the driver previously mapped a shorter
332 * list, we could see a termination bit
333 * prematurely unless it fully inits the sg
334 * table on each mapping. We KNOW that there
335 * must be more entries here or the driver
336 * would be buggy, so force clear the
337 * termination bit to avoid doing a full
338 * sg_init_table() in drivers for each command.
344 sg_set_page(*sg, bvec->bv_page, nbytes, bvec->bv_offset);
350 static int __blk_bios_map_sg(struct request_queue *q, struct bio *bio,
351 struct scatterlist *sglist,
352 struct scatterlist **sg)
354 struct bio_vec bvec, bvprv = { NULL };
355 struct bvec_iter iter;
359 cluster = blk_queue_cluster(q);
361 if (bio->bi_rw & REQ_DISCARD) {
363 * This is a hack - drivers should be neither modifying the
364 * biovec, nor relying on bi_vcnt - but because of
365 * blk_add_request_payload(), a discard bio may or may not have
366 * a payload we need to set up here (thank you Christoph) and
367 * bi_vcnt is really the only way of telling if we need to.
376 if (bio->bi_rw & REQ_WRITE_SAME) {
379 bvec = bio_iovec(bio);
380 sg_set_page(*sg, bvec.bv_page, bvec.bv_len, bvec.bv_offset);
385 bio_for_each_segment(bvec, bio, iter)
386 __blk_segment_map_sg(q, &bvec, sglist, &bvprv, sg,
393 * map a request to scatterlist, return number of sg entries setup. Caller
394 * must make sure sg can hold rq->nr_phys_segments entries
396 int blk_rq_map_sg(struct request_queue *q, struct request *rq,
397 struct scatterlist *sglist)
399 struct scatterlist *sg = NULL;
403 nsegs = __blk_bios_map_sg(q, rq->bio, sglist, &sg);
405 if (unlikely(rq->cmd_flags & REQ_COPY_USER) &&
406 (blk_rq_bytes(rq) & q->dma_pad_mask)) {
407 unsigned int pad_len =
408 (q->dma_pad_mask & ~blk_rq_bytes(rq)) + 1;
410 sg->length += pad_len;
411 rq->extra_len += pad_len;
414 if (q->dma_drain_size && q->dma_drain_needed(rq)) {
415 if (rq->cmd_flags & REQ_WRITE)
416 memset(q->dma_drain_buffer, 0, q->dma_drain_size);
420 sg_set_page(sg, virt_to_page(q->dma_drain_buffer),
422 ((unsigned long)q->dma_drain_buffer) &
425 rq->extra_len += q->dma_drain_size;
433 EXPORT_SYMBOL(blk_rq_map_sg);
435 static inline int ll_new_hw_segment(struct request_queue *q,
439 int nr_phys_segs = bio_phys_segments(q, bio);
441 if (req->nr_phys_segments + nr_phys_segs > queue_max_segments(q))
444 if (blk_integrity_merge_bio(q, req, bio) == false)
448 * This will form the start of a new hw segment. Bump both
451 req->nr_phys_segments += nr_phys_segs;
455 req->cmd_flags |= REQ_NOMERGE;
456 if (req == q->last_merge)
457 q->last_merge = NULL;
461 int ll_back_merge_fn(struct request_queue *q, struct request *req,
464 if (req_gap_back_merge(req, bio))
466 if (blk_integrity_rq(req) &&
467 integrity_req_gap_back_merge(req, bio))
469 if (blk_rq_sectors(req) + bio_sectors(bio) >
470 blk_rq_get_max_sectors(req)) {
471 req->cmd_flags |= REQ_NOMERGE;
472 if (req == q->last_merge)
473 q->last_merge = NULL;
476 if (!bio_flagged(req->biotail, BIO_SEG_VALID))
477 blk_recount_segments(q, req->biotail);
478 if (!bio_flagged(bio, BIO_SEG_VALID))
479 blk_recount_segments(q, bio);
481 return ll_new_hw_segment(q, req, bio);
484 int ll_front_merge_fn(struct request_queue *q, struct request *req,
488 if (req_gap_front_merge(req, bio))
490 if (blk_integrity_rq(req) &&
491 integrity_req_gap_front_merge(req, bio))
493 if (blk_rq_sectors(req) + bio_sectors(bio) >
494 blk_rq_get_max_sectors(req)) {
495 req->cmd_flags |= REQ_NOMERGE;
496 if (req == q->last_merge)
497 q->last_merge = NULL;
500 if (!bio_flagged(bio, BIO_SEG_VALID))
501 blk_recount_segments(q, bio);
502 if (!bio_flagged(req->bio, BIO_SEG_VALID))
503 blk_recount_segments(q, req->bio);
505 return ll_new_hw_segment(q, req, bio);
509 * blk-mq uses req->special to carry normal driver per-request payload, it
510 * does not indicate a prepared command that we cannot merge with.
512 static bool req_no_special_merge(struct request *req)
514 struct request_queue *q = req->q;
516 return !q->mq_ops && req->special;
519 static int ll_merge_requests_fn(struct request_queue *q, struct request *req,
520 struct request *next)
522 int total_phys_segments;
523 unsigned int seg_size =
524 req->biotail->bi_seg_back_size + next->bio->bi_seg_front_size;
527 * First check if the either of the requests are re-queued
528 * requests. Can't merge them if they are.
530 if (req_no_special_merge(req) || req_no_special_merge(next))
533 if (req_gap_back_merge(req, next->bio))
537 * Will it become too large?
539 if ((blk_rq_sectors(req) + blk_rq_sectors(next)) >
540 blk_rq_get_max_sectors(req))
543 total_phys_segments = req->nr_phys_segments + next->nr_phys_segments;
544 if (blk_phys_contig_segment(q, req->biotail, next->bio)) {
545 if (req->nr_phys_segments == 1)
546 req->bio->bi_seg_front_size = seg_size;
547 if (next->nr_phys_segments == 1)
548 next->biotail->bi_seg_back_size = seg_size;
549 total_phys_segments--;
552 if (total_phys_segments > queue_max_segments(q))
555 if (blk_integrity_merge_rq(q, req, next) == false)
559 req->nr_phys_segments = total_phys_segments;
564 * blk_rq_set_mixed_merge - mark a request as mixed merge
565 * @rq: request to mark as mixed merge
568 * @rq is about to be mixed merged. Make sure the attributes
569 * which can be mixed are set in each bio and mark @rq as mixed
572 void blk_rq_set_mixed_merge(struct request *rq)
574 unsigned int ff = rq->cmd_flags & REQ_FAILFAST_MASK;
577 if (rq->cmd_flags & REQ_MIXED_MERGE)
581 * @rq will no longer represent mixable attributes for all the
582 * contained bios. It will just track those of the first one.
583 * Distributes the attributs to each bio.
585 for (bio = rq->bio; bio; bio = bio->bi_next) {
586 WARN_ON_ONCE((bio->bi_rw & REQ_FAILFAST_MASK) &&
587 (bio->bi_rw & REQ_FAILFAST_MASK) != ff);
590 rq->cmd_flags |= REQ_MIXED_MERGE;
593 static void blk_account_io_merge(struct request *req)
595 if (blk_do_io_stat(req)) {
596 struct hd_struct *part;
599 cpu = part_stat_lock();
602 part_round_stats(cpu, part);
603 part_dec_in_flight(part, rq_data_dir(req));
611 * Has to be called with the request spinlock acquired
613 static int attempt_merge(struct request_queue *q, struct request *req,
614 struct request *next)
616 if (!rq_mergeable(req) || !rq_mergeable(next))
619 if (!blk_check_merge_flags(req->cmd_flags, next->cmd_flags))
625 if (blk_rq_pos(req) + blk_rq_sectors(req) != blk_rq_pos(next))
628 if (rq_data_dir(req) != rq_data_dir(next)
629 || req->rq_disk != next->rq_disk
630 || req_no_special_merge(next))
633 if (req->cmd_flags & REQ_WRITE_SAME &&
634 !blk_write_same_mergeable(req->bio, next->bio))
638 * If we are allowed to merge, then append bio list
639 * from next to rq and release next. merge_requests_fn
640 * will have updated segment counts, update sector
643 if (!ll_merge_requests_fn(q, req, next))
647 * If failfast settings disagree or any of the two is already
648 * a mixed merge, mark both as mixed before proceeding. This
649 * makes sure that all involved bios have mixable attributes
652 if ((req->cmd_flags | next->cmd_flags) & REQ_MIXED_MERGE ||
653 (req->cmd_flags & REQ_FAILFAST_MASK) !=
654 (next->cmd_flags & REQ_FAILFAST_MASK)) {
655 blk_rq_set_mixed_merge(req);
656 blk_rq_set_mixed_merge(next);
660 * At this point we have either done a back merge
661 * or front merge. We need the smaller start_time of
662 * the merged requests to be the current request
663 * for accounting purposes.
665 if (time_after(req->start_time, next->start_time))
666 req->start_time = next->start_time;
668 req->biotail->bi_next = next->bio;
669 req->biotail = next->biotail;
671 req->__data_len += blk_rq_bytes(next);
673 elv_merge_requests(q, req, next);
676 * 'next' is going away, so update stats accordingly
678 blk_account_io_merge(next);
680 req->ioprio = ioprio_best(req->ioprio, next->ioprio);
681 if (blk_rq_cpu_valid(next))
682 req->cpu = next->cpu;
684 /* owner-ship of bio passed from next to req */
686 __blk_put_request(q, next);
690 int attempt_back_merge(struct request_queue *q, struct request *rq)
692 struct request *next = elv_latter_request(q, rq);
695 return attempt_merge(q, rq, next);
700 int attempt_front_merge(struct request_queue *q, struct request *rq)
702 struct request *prev = elv_former_request(q, rq);
705 return attempt_merge(q, prev, rq);
710 int blk_attempt_req_merge(struct request_queue *q, struct request *rq,
711 struct request *next)
713 return attempt_merge(q, rq, next);
716 bool blk_rq_merge_ok(struct request *rq, struct bio *bio)
718 if (!rq_mergeable(rq) || !bio_mergeable(bio))
721 if (!blk_check_merge_flags(rq->cmd_flags, bio->bi_rw))
724 /* different data direction or already started, don't merge */
725 if (bio_data_dir(bio) != rq_data_dir(rq))
728 /* must be same device and not a special request */
729 if (rq->rq_disk != bio->bi_bdev->bd_disk || req_no_special_merge(rq))
732 /* only merge integrity protected bio into ditto rq */
733 if (blk_integrity_merge_bio(rq->q, rq, bio) == false)
736 /* must be using the same buffer */
737 if (rq->cmd_flags & REQ_WRITE_SAME &&
738 !blk_write_same_mergeable(rq->bio, bio))
744 int blk_try_merge(struct request *rq, struct bio *bio)
746 if (blk_rq_pos(rq) + blk_rq_sectors(rq) == bio->bi_iter.bi_sector)
747 return ELEVATOR_BACK_MERGE;
748 else if (blk_rq_pos(rq) - bio_sectors(bio) == bio->bi_iter.bi_sector)
749 return ELEVATOR_FRONT_MERGE;
750 return ELEVATOR_NO_MERGE;