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;
80 bio_for_each_segment(bv, bio, iter) {
81 if (sectors + (bv.bv_len >> 9) > queue_max_sectors(q))
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 (bvprvp && blk_queue_cluster(q)) {
92 if (seg_size + bv.bv_len > queue_max_segment_size(q))
94 if (!BIOVEC_PHYS_MERGEABLE(bvprvp, &bv))
96 if (!BIOVEC_SEG_BOUNDARY(q, bvprvp, &bv))
99 seg_size += bv.bv_len;
102 sectors += bv.bv_len >> 9;
106 if (nsegs == queue_max_segments(q))
112 seg_size = bv.bv_len;
113 sectors += bv.bv_len >> 9;
120 return bio_split(bio, sectors, GFP_NOIO, bs);
123 void blk_queue_split(struct request_queue *q, struct bio **bio,
126 struct bio *split, *res;
129 if ((*bio)->bi_rw & REQ_DISCARD)
130 split = blk_bio_discard_split(q, *bio, bs, &nsegs);
131 else if ((*bio)->bi_rw & REQ_WRITE_SAME)
132 split = blk_bio_write_same_split(q, *bio, bs, &nsegs);
134 split = blk_bio_segment_split(q, *bio, q->bio_split, &nsegs);
136 /* physical segments can be figured out during splitting */
137 res = split ? split : *bio;
138 res->bi_phys_segments = nsegs;
139 bio_set_flag(res, BIO_SEG_VALID);
142 /* there isn't chance to merge the splitted bio */
143 split->bi_rw |= REQ_NOMERGE;
145 bio_chain(split, *bio);
146 generic_make_request(*bio);
150 EXPORT_SYMBOL(blk_queue_split);
152 static unsigned int __blk_recalc_rq_segments(struct request_queue *q,
156 struct bio_vec bv, bvprv = { NULL };
157 int cluster, prev = 0;
158 unsigned int seg_size, nr_phys_segs;
159 struct bio *fbio, *bbio;
160 struct bvec_iter iter;
166 * This should probably be returning 0, but blk_add_request_payload()
169 if (bio->bi_rw & REQ_DISCARD)
172 if (bio->bi_rw & REQ_WRITE_SAME)
176 cluster = blk_queue_cluster(q);
180 bio_for_each_segment(bv, bio, iter) {
182 * If SG merging is disabled, each bio vector is
188 if (prev && cluster) {
189 if (seg_size + bv.bv_len
190 > queue_max_segment_size(q))
192 if (!BIOVEC_PHYS_MERGEABLE(&bvprv, &bv))
194 if (!BIOVEC_SEG_BOUNDARY(q, &bvprv, &bv))
197 seg_size += bv.bv_len;
202 if (nr_phys_segs == 1 && seg_size >
203 fbio->bi_seg_front_size)
204 fbio->bi_seg_front_size = seg_size;
209 seg_size = bv.bv_len;
214 if (nr_phys_segs == 1 && seg_size > fbio->bi_seg_front_size)
215 fbio->bi_seg_front_size = seg_size;
216 if (seg_size > bbio->bi_seg_back_size)
217 bbio->bi_seg_back_size = seg_size;
222 void blk_recalc_rq_segments(struct request *rq)
224 bool no_sg_merge = !!test_bit(QUEUE_FLAG_NO_SG_MERGE,
225 &rq->q->queue_flags);
227 rq->nr_phys_segments = __blk_recalc_rq_segments(rq->q, rq->bio,
231 void blk_recount_segments(struct request_queue *q, struct bio *bio)
233 unsigned short seg_cnt;
235 /* estimate segment number by bi_vcnt for non-cloned bio */
236 if (bio_flagged(bio, BIO_CLONED))
237 seg_cnt = bio_segments(bio);
239 seg_cnt = bio->bi_vcnt;
241 if (test_bit(QUEUE_FLAG_NO_SG_MERGE, &q->queue_flags) &&
242 (seg_cnt < queue_max_segments(q)))
243 bio->bi_phys_segments = seg_cnt;
245 struct bio *nxt = bio->bi_next;
248 bio->bi_phys_segments = __blk_recalc_rq_segments(q, bio, false);
252 bio_set_flag(bio, BIO_SEG_VALID);
254 EXPORT_SYMBOL(blk_recount_segments);
256 static int blk_phys_contig_segment(struct request_queue *q, struct bio *bio,
259 struct bio_vec end_bv = { NULL }, nxt_bv;
260 struct bvec_iter iter;
262 if (!blk_queue_cluster(q))
265 if (bio->bi_seg_back_size + nxt->bi_seg_front_size >
266 queue_max_segment_size(q))
269 if (!bio_has_data(bio))
272 bio_for_each_segment(end_bv, bio, iter)
273 if (end_bv.bv_len == iter.bi_size)
276 nxt_bv = bio_iovec(nxt);
278 if (!BIOVEC_PHYS_MERGEABLE(&end_bv, &nxt_bv))
282 * bio and nxt are contiguous in memory; check if the queue allows
283 * these two to be merged into one
285 if (BIOVEC_SEG_BOUNDARY(q, &end_bv, &nxt_bv))
292 __blk_segment_map_sg(struct request_queue *q, struct bio_vec *bvec,
293 struct scatterlist *sglist, struct bio_vec *bvprv,
294 struct scatterlist **sg, int *nsegs, int *cluster)
297 int nbytes = bvec->bv_len;
299 if (*sg && *cluster) {
300 if ((*sg)->length + nbytes > queue_max_segment_size(q))
303 if (!BIOVEC_PHYS_MERGEABLE(bvprv, bvec))
305 if (!BIOVEC_SEG_BOUNDARY(q, bvprv, bvec))
308 (*sg)->length += nbytes;
315 * If the driver previously mapped a shorter
316 * list, we could see a termination bit
317 * prematurely unless it fully inits the sg
318 * table on each mapping. We KNOW that there
319 * must be more entries here or the driver
320 * would be buggy, so force clear the
321 * termination bit to avoid doing a full
322 * sg_init_table() in drivers for each command.
328 sg_set_page(*sg, bvec->bv_page, nbytes, bvec->bv_offset);
334 static int __blk_bios_map_sg(struct request_queue *q, struct bio *bio,
335 struct scatterlist *sglist,
336 struct scatterlist **sg)
338 struct bio_vec bvec, bvprv = { NULL };
339 struct bvec_iter iter;
343 cluster = blk_queue_cluster(q);
345 if (bio->bi_rw & REQ_DISCARD) {
347 * This is a hack - drivers should be neither modifying the
348 * biovec, nor relying on bi_vcnt - but because of
349 * blk_add_request_payload(), a discard bio may or may not have
350 * a payload we need to set up here (thank you Christoph) and
351 * bi_vcnt is really the only way of telling if we need to.
360 if (bio->bi_rw & REQ_WRITE_SAME) {
363 bvec = bio_iovec(bio);
364 sg_set_page(*sg, bvec.bv_page, bvec.bv_len, bvec.bv_offset);
369 bio_for_each_segment(bvec, bio, iter)
370 __blk_segment_map_sg(q, &bvec, sglist, &bvprv, sg,
377 * map a request to scatterlist, return number of sg entries setup. Caller
378 * must make sure sg can hold rq->nr_phys_segments entries
380 int blk_rq_map_sg(struct request_queue *q, struct request *rq,
381 struct scatterlist *sglist)
383 struct scatterlist *sg = NULL;
387 nsegs = __blk_bios_map_sg(q, rq->bio, sglist, &sg);
389 if (unlikely(rq->cmd_flags & REQ_COPY_USER) &&
390 (blk_rq_bytes(rq) & q->dma_pad_mask)) {
391 unsigned int pad_len =
392 (q->dma_pad_mask & ~blk_rq_bytes(rq)) + 1;
394 sg->length += pad_len;
395 rq->extra_len += pad_len;
398 if (q->dma_drain_size && q->dma_drain_needed(rq)) {
399 if (rq->cmd_flags & REQ_WRITE)
400 memset(q->dma_drain_buffer, 0, q->dma_drain_size);
404 sg_set_page(sg, virt_to_page(q->dma_drain_buffer),
406 ((unsigned long)q->dma_drain_buffer) &
409 rq->extra_len += q->dma_drain_size;
417 EXPORT_SYMBOL(blk_rq_map_sg);
419 static inline int ll_new_hw_segment(struct request_queue *q,
423 int nr_phys_segs = bio_phys_segments(q, bio);
425 if (req->nr_phys_segments + nr_phys_segs > queue_max_segments(q))
428 if (blk_integrity_merge_bio(q, req, bio) == false)
432 * This will form the start of a new hw segment. Bump both
435 req->nr_phys_segments += nr_phys_segs;
439 req->cmd_flags |= REQ_NOMERGE;
440 if (req == q->last_merge)
441 q->last_merge = NULL;
445 int ll_back_merge_fn(struct request_queue *q, struct request *req,
448 if (req_gap_back_merge(req, bio))
450 if (blk_integrity_rq(req) &&
451 integrity_req_gap_back_merge(req, bio))
453 if (blk_rq_sectors(req) + bio_sectors(bio) >
454 blk_rq_get_max_sectors(req)) {
455 req->cmd_flags |= REQ_NOMERGE;
456 if (req == q->last_merge)
457 q->last_merge = NULL;
460 if (!bio_flagged(req->biotail, BIO_SEG_VALID))
461 blk_recount_segments(q, req->biotail);
462 if (!bio_flagged(bio, BIO_SEG_VALID))
463 blk_recount_segments(q, bio);
465 return ll_new_hw_segment(q, req, bio);
468 int ll_front_merge_fn(struct request_queue *q, struct request *req,
472 if (req_gap_front_merge(req, bio))
474 if (blk_integrity_rq(req) &&
475 integrity_req_gap_front_merge(req, bio))
477 if (blk_rq_sectors(req) + bio_sectors(bio) >
478 blk_rq_get_max_sectors(req)) {
479 req->cmd_flags |= REQ_NOMERGE;
480 if (req == q->last_merge)
481 q->last_merge = NULL;
484 if (!bio_flagged(bio, BIO_SEG_VALID))
485 blk_recount_segments(q, bio);
486 if (!bio_flagged(req->bio, BIO_SEG_VALID))
487 blk_recount_segments(q, req->bio);
489 return ll_new_hw_segment(q, req, bio);
493 * blk-mq uses req->special to carry normal driver per-request payload, it
494 * does not indicate a prepared command that we cannot merge with.
496 static bool req_no_special_merge(struct request *req)
498 struct request_queue *q = req->q;
500 return !q->mq_ops && req->special;
503 static int ll_merge_requests_fn(struct request_queue *q, struct request *req,
504 struct request *next)
506 int total_phys_segments;
507 unsigned int seg_size =
508 req->biotail->bi_seg_back_size + next->bio->bi_seg_front_size;
511 * First check if the either of the requests are re-queued
512 * requests. Can't merge them if they are.
514 if (req_no_special_merge(req) || req_no_special_merge(next))
517 if (req_gap_back_merge(req, next->bio))
521 * Will it become too large?
523 if ((blk_rq_sectors(req) + blk_rq_sectors(next)) >
524 blk_rq_get_max_sectors(req))
527 total_phys_segments = req->nr_phys_segments + next->nr_phys_segments;
528 if (blk_phys_contig_segment(q, req->biotail, next->bio)) {
529 if (req->nr_phys_segments == 1)
530 req->bio->bi_seg_front_size = seg_size;
531 if (next->nr_phys_segments == 1)
532 next->biotail->bi_seg_back_size = seg_size;
533 total_phys_segments--;
536 if (total_phys_segments > queue_max_segments(q))
539 if (blk_integrity_merge_rq(q, req, next) == false)
543 req->nr_phys_segments = total_phys_segments;
548 * blk_rq_set_mixed_merge - mark a request as mixed merge
549 * @rq: request to mark as mixed merge
552 * @rq is about to be mixed merged. Make sure the attributes
553 * which can be mixed are set in each bio and mark @rq as mixed
556 void blk_rq_set_mixed_merge(struct request *rq)
558 unsigned int ff = rq->cmd_flags & REQ_FAILFAST_MASK;
561 if (rq->cmd_flags & REQ_MIXED_MERGE)
565 * @rq will no longer represent mixable attributes for all the
566 * contained bios. It will just track those of the first one.
567 * Distributes the attributs to each bio.
569 for (bio = rq->bio; bio; bio = bio->bi_next) {
570 WARN_ON_ONCE((bio->bi_rw & REQ_FAILFAST_MASK) &&
571 (bio->bi_rw & REQ_FAILFAST_MASK) != ff);
574 rq->cmd_flags |= REQ_MIXED_MERGE;
577 static void blk_account_io_merge(struct request *req)
579 if (blk_do_io_stat(req)) {
580 struct hd_struct *part;
583 cpu = part_stat_lock();
586 part_round_stats(cpu, part);
587 part_dec_in_flight(part, rq_data_dir(req));
595 * Has to be called with the request spinlock acquired
597 static int attempt_merge(struct request_queue *q, struct request *req,
598 struct request *next)
600 if (!rq_mergeable(req) || !rq_mergeable(next))
603 if (!blk_check_merge_flags(req->cmd_flags, next->cmd_flags))
609 if (blk_rq_pos(req) + blk_rq_sectors(req) != blk_rq_pos(next))
612 if (rq_data_dir(req) != rq_data_dir(next)
613 || req->rq_disk != next->rq_disk
614 || req_no_special_merge(next))
617 if (req->cmd_flags & REQ_WRITE_SAME &&
618 !blk_write_same_mergeable(req->bio, next->bio))
622 * If we are allowed to merge, then append bio list
623 * from next to rq and release next. merge_requests_fn
624 * will have updated segment counts, update sector
627 if (!ll_merge_requests_fn(q, req, next))
631 * If failfast settings disagree or any of the two is already
632 * a mixed merge, mark both as mixed before proceeding. This
633 * makes sure that all involved bios have mixable attributes
636 if ((req->cmd_flags | next->cmd_flags) & REQ_MIXED_MERGE ||
637 (req->cmd_flags & REQ_FAILFAST_MASK) !=
638 (next->cmd_flags & REQ_FAILFAST_MASK)) {
639 blk_rq_set_mixed_merge(req);
640 blk_rq_set_mixed_merge(next);
644 * At this point we have either done a back merge
645 * or front merge. We need the smaller start_time of
646 * the merged requests to be the current request
647 * for accounting purposes.
649 if (time_after(req->start_time, next->start_time))
650 req->start_time = next->start_time;
652 req->biotail->bi_next = next->bio;
653 req->biotail = next->biotail;
655 req->__data_len += blk_rq_bytes(next);
657 elv_merge_requests(q, req, next);
660 * 'next' is going away, so update stats accordingly
662 blk_account_io_merge(next);
664 req->ioprio = ioprio_best(req->ioprio, next->ioprio);
665 if (blk_rq_cpu_valid(next))
666 req->cpu = next->cpu;
668 /* owner-ship of bio passed from next to req */
670 __blk_put_request(q, next);
674 int attempt_back_merge(struct request_queue *q, struct request *rq)
676 struct request *next = elv_latter_request(q, rq);
679 return attempt_merge(q, rq, next);
684 int attempt_front_merge(struct request_queue *q, struct request *rq)
686 struct request *prev = elv_former_request(q, rq);
689 return attempt_merge(q, prev, rq);
694 int blk_attempt_req_merge(struct request_queue *q, struct request *rq,
695 struct request *next)
697 return attempt_merge(q, rq, next);
700 bool blk_rq_merge_ok(struct request *rq, struct bio *bio)
702 if (!rq_mergeable(rq) || !bio_mergeable(bio))
705 if (!blk_check_merge_flags(rq->cmd_flags, bio->bi_rw))
708 /* different data direction or already started, don't merge */
709 if (bio_data_dir(bio) != rq_data_dir(rq))
712 /* must be same device and not a special request */
713 if (rq->rq_disk != bio->bi_bdev->bd_disk || req_no_special_merge(rq))
716 /* only merge integrity protected bio into ditto rq */
717 if (blk_integrity_merge_bio(rq->q, rq, bio) == false)
720 /* must be using the same buffer */
721 if (rq->cmd_flags & REQ_WRITE_SAME &&
722 !blk_write_same_mergeable(rq->bio, bio))
728 int blk_try_merge(struct request *rq, struct bio *bio)
730 if (blk_rq_pos(rq) + blk_rq_sectors(rq) == bio->bi_iter.bi_sector)
731 return ELEVATOR_BACK_MERGE;
732 else if (blk_rq_pos(rq) - bio_sectors(bio) == bio->bi_iter.bi_sector)
733 return ELEVATOR_FRONT_MERGE;
734 return ELEVATOR_NO_MERGE;