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) {
85 * If the queue doesn't support SG gaps and adding this
86 * offset would create a gap, disallow it.
88 if (bvprvp && bvec_gap_to_prev(q, bvprvp, bv.bv_offset))
91 if (sectors + (bv.bv_len >> 9) >
92 blk_max_size_offset(q, bio->bi_iter.bi_sector)) {
94 * Consider this a new segment if we're splitting in
95 * the middle of this vector.
97 if (nsegs < queue_max_segments(q) &&
98 sectors < blk_max_size_offset(q,
99 bio->bi_iter.bi_sector)) {
101 sectors = blk_max_size_offset(q,
102 bio->bi_iter.bi_sector);
107 if (bvprvp && blk_queue_cluster(q)) {
108 if (seg_size + bv.bv_len > queue_max_segment_size(q))
110 if (!BIOVEC_PHYS_MERGEABLE(bvprvp, &bv))
112 if (!BIOVEC_SEG_BOUNDARY(q, bvprvp, &bv))
115 seg_size += bv.bv_len;
118 sectors += bv.bv_len >> 9;
120 if (nsegs == 1 && seg_size > front_seg_size)
121 front_seg_size = seg_size;
125 if (nsegs == queue_max_segments(q))
131 seg_size = bv.bv_len;
132 sectors += bv.bv_len >> 9;
134 if (nsegs == 1 && seg_size > front_seg_size)
135 front_seg_size = seg_size;
143 new = bio_split(bio, sectors, GFP_NOIO, bs);
148 bio->bi_seg_front_size = front_seg_size;
149 if (seg_size > bio->bi_seg_back_size)
150 bio->bi_seg_back_size = seg_size;
152 return do_split ? new : NULL;
155 void blk_queue_split(struct request_queue *q, struct bio **bio,
158 struct bio *split, *res;
161 if ((*bio)->bi_rw & REQ_DISCARD)
162 split = blk_bio_discard_split(q, *bio, bs, &nsegs);
163 else if ((*bio)->bi_rw & REQ_WRITE_SAME)
164 split = blk_bio_write_same_split(q, *bio, bs, &nsegs);
166 split = blk_bio_segment_split(q, *bio, q->bio_split, &nsegs);
168 /* physical segments can be figured out during splitting */
169 res = split ? split : *bio;
170 res->bi_phys_segments = nsegs;
171 bio_set_flag(res, BIO_SEG_VALID);
174 /* there isn't chance to merge the splitted bio */
175 split->bi_rw |= REQ_NOMERGE;
177 bio_chain(split, *bio);
178 generic_make_request(*bio);
182 EXPORT_SYMBOL(blk_queue_split);
184 static unsigned int __blk_recalc_rq_segments(struct request_queue *q,
188 struct bio_vec bv, bvprv = { NULL };
189 int cluster, prev = 0;
190 unsigned int seg_size, nr_phys_segs;
191 struct bio *fbio, *bbio;
192 struct bvec_iter iter;
198 * This should probably be returning 0, but blk_add_request_payload()
201 if (bio->bi_rw & REQ_DISCARD)
204 if (bio->bi_rw & REQ_WRITE_SAME)
208 cluster = blk_queue_cluster(q);
212 bio_for_each_segment(bv, bio, iter) {
214 * If SG merging is disabled, each bio vector is
220 if (prev && cluster) {
221 if (seg_size + bv.bv_len
222 > queue_max_segment_size(q))
224 if (!BIOVEC_PHYS_MERGEABLE(&bvprv, &bv))
226 if (!BIOVEC_SEG_BOUNDARY(q, &bvprv, &bv))
229 seg_size += bv.bv_len;
234 if (nr_phys_segs == 1 && seg_size >
235 fbio->bi_seg_front_size)
236 fbio->bi_seg_front_size = seg_size;
241 seg_size = bv.bv_len;
246 if (nr_phys_segs == 1 && seg_size > fbio->bi_seg_front_size)
247 fbio->bi_seg_front_size = seg_size;
248 if (seg_size > bbio->bi_seg_back_size)
249 bbio->bi_seg_back_size = seg_size;
254 void blk_recalc_rq_segments(struct request *rq)
256 bool no_sg_merge = !!test_bit(QUEUE_FLAG_NO_SG_MERGE,
257 &rq->q->queue_flags);
259 rq->nr_phys_segments = __blk_recalc_rq_segments(rq->q, rq->bio,
263 void blk_recount_segments(struct request_queue *q, struct bio *bio)
265 unsigned short seg_cnt;
267 /* estimate segment number by bi_vcnt for non-cloned bio */
268 if (bio_flagged(bio, BIO_CLONED))
269 seg_cnt = bio_segments(bio);
271 seg_cnt = bio->bi_vcnt;
273 if (test_bit(QUEUE_FLAG_NO_SG_MERGE, &q->queue_flags) &&
274 (seg_cnt < queue_max_segments(q)))
275 bio->bi_phys_segments = seg_cnt;
277 struct bio *nxt = bio->bi_next;
280 bio->bi_phys_segments = __blk_recalc_rq_segments(q, bio, false);
284 bio_set_flag(bio, BIO_SEG_VALID);
286 EXPORT_SYMBOL(blk_recount_segments);
288 static int blk_phys_contig_segment(struct request_queue *q, struct bio *bio,
291 struct bio_vec end_bv = { NULL }, nxt_bv;
292 struct bvec_iter iter;
294 if (!blk_queue_cluster(q))
297 if (bio->bi_seg_back_size + nxt->bi_seg_front_size >
298 queue_max_segment_size(q))
301 if (!bio_has_data(bio))
304 bio_for_each_segment(end_bv, bio, iter)
305 if (end_bv.bv_len == iter.bi_size)
308 nxt_bv = bio_iovec(nxt);
310 if (!BIOVEC_PHYS_MERGEABLE(&end_bv, &nxt_bv))
314 * bio and nxt are contiguous in memory; check if the queue allows
315 * these two to be merged into one
317 if (BIOVEC_SEG_BOUNDARY(q, &end_bv, &nxt_bv))
324 __blk_segment_map_sg(struct request_queue *q, struct bio_vec *bvec,
325 struct scatterlist *sglist, struct bio_vec *bvprv,
326 struct scatterlist **sg, int *nsegs, int *cluster)
329 int nbytes = bvec->bv_len;
331 if (*sg && *cluster) {
332 if ((*sg)->length + nbytes > queue_max_segment_size(q))
335 if (!BIOVEC_PHYS_MERGEABLE(bvprv, bvec))
337 if (!BIOVEC_SEG_BOUNDARY(q, bvprv, bvec))
340 (*sg)->length += nbytes;
347 * If the driver previously mapped a shorter
348 * list, we could see a termination bit
349 * prematurely unless it fully inits the sg
350 * table on each mapping. We KNOW that there
351 * must be more entries here or the driver
352 * would be buggy, so force clear the
353 * termination bit to avoid doing a full
354 * sg_init_table() in drivers for each command.
360 sg_set_page(*sg, bvec->bv_page, nbytes, bvec->bv_offset);
366 static int __blk_bios_map_sg(struct request_queue *q, struct bio *bio,
367 struct scatterlist *sglist,
368 struct scatterlist **sg)
370 struct bio_vec bvec, bvprv = { NULL };
371 struct bvec_iter iter;
375 cluster = blk_queue_cluster(q);
377 if (bio->bi_rw & REQ_DISCARD) {
379 * This is a hack - drivers should be neither modifying the
380 * biovec, nor relying on bi_vcnt - but because of
381 * blk_add_request_payload(), a discard bio may or may not have
382 * a payload we need to set up here (thank you Christoph) and
383 * bi_vcnt is really the only way of telling if we need to.
392 if (bio->bi_rw & REQ_WRITE_SAME) {
395 bvec = bio_iovec(bio);
396 sg_set_page(*sg, bvec.bv_page, bvec.bv_len, bvec.bv_offset);
401 bio_for_each_segment(bvec, bio, iter)
402 __blk_segment_map_sg(q, &bvec, sglist, &bvprv, sg,
409 * map a request to scatterlist, return number of sg entries setup. Caller
410 * must make sure sg can hold rq->nr_phys_segments entries
412 int blk_rq_map_sg(struct request_queue *q, struct request *rq,
413 struct scatterlist *sglist)
415 struct scatterlist *sg = NULL;
419 nsegs = __blk_bios_map_sg(q, rq->bio, sglist, &sg);
421 if (unlikely(rq->cmd_flags & REQ_COPY_USER) &&
422 (blk_rq_bytes(rq) & q->dma_pad_mask)) {
423 unsigned int pad_len =
424 (q->dma_pad_mask & ~blk_rq_bytes(rq)) + 1;
426 sg->length += pad_len;
427 rq->extra_len += pad_len;
430 if (q->dma_drain_size && q->dma_drain_needed(rq)) {
431 if (rq->cmd_flags & REQ_WRITE)
432 memset(q->dma_drain_buffer, 0, q->dma_drain_size);
436 sg_set_page(sg, virt_to_page(q->dma_drain_buffer),
438 ((unsigned long)q->dma_drain_buffer) &
441 rq->extra_len += q->dma_drain_size;
448 * Something must have been wrong if the figured number of
449 * segment is bigger than number of req's physical segments
451 WARN_ON(nsegs > rq->nr_phys_segments);
455 EXPORT_SYMBOL(blk_rq_map_sg);
457 static inline int ll_new_hw_segment(struct request_queue *q,
461 int nr_phys_segs = bio_phys_segments(q, bio);
463 if (req->nr_phys_segments + nr_phys_segs > queue_max_segments(q))
466 if (blk_integrity_merge_bio(q, req, bio) == false)
470 * This will form the start of a new hw segment. Bump both
473 req->nr_phys_segments += nr_phys_segs;
477 req->cmd_flags |= REQ_NOMERGE;
478 if (req == q->last_merge)
479 q->last_merge = NULL;
483 int ll_back_merge_fn(struct request_queue *q, struct request *req,
486 if (req_gap_back_merge(req, bio))
488 if (blk_integrity_rq(req) &&
489 integrity_req_gap_back_merge(req, bio))
491 if (blk_rq_sectors(req) + bio_sectors(bio) >
492 blk_rq_get_max_sectors(req)) {
493 req->cmd_flags |= REQ_NOMERGE;
494 if (req == q->last_merge)
495 q->last_merge = NULL;
498 if (!bio_flagged(req->biotail, BIO_SEG_VALID))
499 blk_recount_segments(q, req->biotail);
500 if (!bio_flagged(bio, BIO_SEG_VALID))
501 blk_recount_segments(q, bio);
503 return ll_new_hw_segment(q, req, bio);
506 int ll_front_merge_fn(struct request_queue *q, struct request *req,
510 if (req_gap_front_merge(req, bio))
512 if (blk_integrity_rq(req) &&
513 integrity_req_gap_front_merge(req, bio))
515 if (blk_rq_sectors(req) + bio_sectors(bio) >
516 blk_rq_get_max_sectors(req)) {
517 req->cmd_flags |= REQ_NOMERGE;
518 if (req == q->last_merge)
519 q->last_merge = NULL;
522 if (!bio_flagged(bio, BIO_SEG_VALID))
523 blk_recount_segments(q, bio);
524 if (!bio_flagged(req->bio, BIO_SEG_VALID))
525 blk_recount_segments(q, req->bio);
527 return ll_new_hw_segment(q, req, bio);
531 * blk-mq uses req->special to carry normal driver per-request payload, it
532 * does not indicate a prepared command that we cannot merge with.
534 static bool req_no_special_merge(struct request *req)
536 struct request_queue *q = req->q;
538 return !q->mq_ops && req->special;
541 static int ll_merge_requests_fn(struct request_queue *q, struct request *req,
542 struct request *next)
544 int total_phys_segments;
545 unsigned int seg_size =
546 req->biotail->bi_seg_back_size + next->bio->bi_seg_front_size;
549 * First check if the either of the requests are re-queued
550 * requests. Can't merge them if they are.
552 if (req_no_special_merge(req) || req_no_special_merge(next))
555 if (req_gap_back_merge(req, next->bio))
559 * Will it become too large?
561 if ((blk_rq_sectors(req) + blk_rq_sectors(next)) >
562 blk_rq_get_max_sectors(req))
565 total_phys_segments = req->nr_phys_segments + next->nr_phys_segments;
566 if (blk_phys_contig_segment(q, req->biotail, next->bio)) {
567 if (req->nr_phys_segments == 1)
568 req->bio->bi_seg_front_size = seg_size;
569 if (next->nr_phys_segments == 1)
570 next->biotail->bi_seg_back_size = seg_size;
571 total_phys_segments--;
574 if (total_phys_segments > queue_max_segments(q))
577 if (blk_integrity_merge_rq(q, req, next) == false)
581 req->nr_phys_segments = total_phys_segments;
586 * blk_rq_set_mixed_merge - mark a request as mixed merge
587 * @rq: request to mark as mixed merge
590 * @rq is about to be mixed merged. Make sure the attributes
591 * which can be mixed are set in each bio and mark @rq as mixed
594 void blk_rq_set_mixed_merge(struct request *rq)
596 unsigned int ff = rq->cmd_flags & REQ_FAILFAST_MASK;
599 if (rq->cmd_flags & REQ_MIXED_MERGE)
603 * @rq will no longer represent mixable attributes for all the
604 * contained bios. It will just track those of the first one.
605 * Distributes the attributs to each bio.
607 for (bio = rq->bio; bio; bio = bio->bi_next) {
608 WARN_ON_ONCE((bio->bi_rw & REQ_FAILFAST_MASK) &&
609 (bio->bi_rw & REQ_FAILFAST_MASK) != ff);
612 rq->cmd_flags |= REQ_MIXED_MERGE;
615 static void blk_account_io_merge(struct request *req)
617 if (blk_do_io_stat(req)) {
618 struct hd_struct *part;
621 cpu = part_stat_lock();
624 part_round_stats(cpu, part);
625 part_dec_in_flight(part, rq_data_dir(req));
633 * Has to be called with the request spinlock acquired
635 static int attempt_merge(struct request_queue *q, struct request *req,
636 struct request *next)
638 if (!rq_mergeable(req) || !rq_mergeable(next))
641 if (!blk_check_merge_flags(req->cmd_flags, next->cmd_flags))
647 if (blk_rq_pos(req) + blk_rq_sectors(req) != blk_rq_pos(next))
650 if (rq_data_dir(req) != rq_data_dir(next)
651 || req->rq_disk != next->rq_disk
652 || req_no_special_merge(next))
655 if (req->cmd_flags & REQ_WRITE_SAME &&
656 !blk_write_same_mergeable(req->bio, next->bio))
660 * If we are allowed to merge, then append bio list
661 * from next to rq and release next. merge_requests_fn
662 * will have updated segment counts, update sector
665 if (!ll_merge_requests_fn(q, req, next))
669 * If failfast settings disagree or any of the two is already
670 * a mixed merge, mark both as mixed before proceeding. This
671 * makes sure that all involved bios have mixable attributes
674 if ((req->cmd_flags | next->cmd_flags) & REQ_MIXED_MERGE ||
675 (req->cmd_flags & REQ_FAILFAST_MASK) !=
676 (next->cmd_flags & REQ_FAILFAST_MASK)) {
677 blk_rq_set_mixed_merge(req);
678 blk_rq_set_mixed_merge(next);
682 * At this point we have either done a back merge
683 * or front merge. We need the smaller start_time of
684 * the merged requests to be the current request
685 * for accounting purposes.
687 if (time_after(req->start_time, next->start_time))
688 req->start_time = next->start_time;
690 req->biotail->bi_next = next->bio;
691 req->biotail = next->biotail;
693 req->__data_len += blk_rq_bytes(next);
695 elv_merge_requests(q, req, next);
698 * 'next' is going away, so update stats accordingly
700 blk_account_io_merge(next);
702 req->ioprio = ioprio_best(req->ioprio, next->ioprio);
703 if (blk_rq_cpu_valid(next))
704 req->cpu = next->cpu;
706 /* owner-ship of bio passed from next to req */
708 __blk_put_request(q, next);
712 int attempt_back_merge(struct request_queue *q, struct request *rq)
714 struct request *next = elv_latter_request(q, rq);
717 return attempt_merge(q, rq, next);
722 int attempt_front_merge(struct request_queue *q, struct request *rq)
724 struct request *prev = elv_former_request(q, rq);
727 return attempt_merge(q, prev, rq);
732 int blk_attempt_req_merge(struct request_queue *q, struct request *rq,
733 struct request *next)
735 return attempt_merge(q, rq, next);
738 bool blk_rq_merge_ok(struct request *rq, struct bio *bio)
740 if (!rq_mergeable(rq) || !bio_mergeable(bio))
743 if (!blk_check_merge_flags(rq->cmd_flags, bio->bi_rw))
746 /* different data direction or already started, don't merge */
747 if (bio_data_dir(bio) != rq_data_dir(rq))
750 /* must be same device and not a special request */
751 if (rq->rq_disk != bio->bi_bdev->bd_disk || req_no_special_merge(rq))
754 /* only merge integrity protected bio into ditto rq */
755 if (blk_integrity_merge_bio(rq->q, rq, bio) == false)
758 /* must be using the same buffer */
759 if (rq->cmd_flags & REQ_WRITE_SAME &&
760 !blk_write_same_mergeable(rq->bio, bio))
766 int blk_try_merge(struct request *rq, struct bio *bio)
768 if (blk_rq_pos(rq) + blk_rq_sectors(rq) == bio->bi_iter.bi_sector)
769 return ELEVATOR_BACK_MERGE;
770 else if (blk_rq_pos(rq) - bio_sectors(bio) == bio->bi_iter.bi_sector)
771 return ELEVATOR_FRONT_MERGE;
772 return ELEVATOR_NO_MERGE;