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 struct bio *blk_bio_segment_split(struct request_queue *q,
78 struct bio_vec bv, bvprv, *bvprvp = NULL;
79 struct bvec_iter iter;
80 unsigned seg_size = 0, nsegs = 0, sectors = 0;
81 unsigned front_seg_size = bio->bi_seg_front_size;
83 struct bio *new = NULL;
85 bio_for_each_segment(bv, bio, iter) {
87 * If the queue doesn't support SG gaps and adding this
88 * offset would create a gap, disallow it.
90 if (bvprvp && bvec_gap_to_prev(q, bvprvp, bv.bv_offset))
93 if (sectors + (bv.bv_len >> 9) >
94 blk_max_size_offset(q, bio->bi_iter.bi_sector)) {
96 * Consider this a new segment if we're splitting in
97 * the middle of this vector.
99 if (nsegs < queue_max_segments(q) &&
100 sectors < blk_max_size_offset(q,
101 bio->bi_iter.bi_sector)) {
103 sectors = blk_max_size_offset(q,
104 bio->bi_iter.bi_sector);
109 if (bvprvp && blk_queue_cluster(q)) {
110 if (seg_size + bv.bv_len > queue_max_segment_size(q))
112 if (!BIOVEC_PHYS_MERGEABLE(bvprvp, &bv))
114 if (!BIOVEC_SEG_BOUNDARY(q, bvprvp, &bv))
117 seg_size += bv.bv_len;
120 sectors += bv.bv_len >> 9;
122 if (nsegs == 1 && seg_size > front_seg_size)
123 front_seg_size = seg_size;
127 if (nsegs == queue_max_segments(q))
133 seg_size = bv.bv_len;
134 sectors += bv.bv_len >> 9;
136 if (nsegs == 1 && seg_size > front_seg_size)
137 front_seg_size = seg_size;
145 new = bio_split(bio, sectors, GFP_NOIO, bs);
150 bio->bi_seg_front_size = front_seg_size;
151 if (seg_size > bio->bi_seg_back_size)
152 bio->bi_seg_back_size = seg_size;
154 return do_split ? new : NULL;
157 void blk_queue_split(struct request_queue *q, struct bio **bio,
160 struct bio *split, *res;
163 if ((*bio)->bi_rw & REQ_DISCARD)
164 split = blk_bio_discard_split(q, *bio, bs, &nsegs);
165 else if ((*bio)->bi_rw & REQ_WRITE_SAME)
166 split = blk_bio_write_same_split(q, *bio, bs, &nsegs);
168 split = blk_bio_segment_split(q, *bio, q->bio_split, &nsegs);
170 /* physical segments can be figured out during splitting */
171 res = split ? split : *bio;
172 res->bi_phys_segments = nsegs;
173 bio_set_flag(res, BIO_SEG_VALID);
176 /* there isn't chance to merge the splitted bio */
177 split->bi_rw |= REQ_NOMERGE;
179 bio_chain(split, *bio);
180 trace_block_split(q, split, (*bio)->bi_iter.bi_sector);
181 generic_make_request(*bio);
185 EXPORT_SYMBOL(blk_queue_split);
187 static unsigned int __blk_recalc_rq_segments(struct request_queue *q,
191 struct bio_vec bv, bvprv = { NULL };
192 int cluster, prev = 0;
193 unsigned int seg_size, nr_phys_segs;
194 struct bio *fbio, *bbio;
195 struct bvec_iter iter;
201 * This should probably be returning 0, but blk_add_request_payload()
204 if (bio->bi_rw & REQ_DISCARD)
207 if (bio->bi_rw & REQ_WRITE_SAME)
211 cluster = blk_queue_cluster(q);
215 bio_for_each_segment(bv, bio, iter) {
217 * If SG merging is disabled, each bio vector is
223 if (prev && cluster) {
224 if (seg_size + bv.bv_len
225 > queue_max_segment_size(q))
227 if (!BIOVEC_PHYS_MERGEABLE(&bvprv, &bv))
229 if (!BIOVEC_SEG_BOUNDARY(q, &bvprv, &bv))
232 seg_size += bv.bv_len;
237 if (nr_phys_segs == 1 && seg_size >
238 fbio->bi_seg_front_size)
239 fbio->bi_seg_front_size = seg_size;
244 seg_size = bv.bv_len;
249 if (nr_phys_segs == 1 && seg_size > fbio->bi_seg_front_size)
250 fbio->bi_seg_front_size = seg_size;
251 if (seg_size > bbio->bi_seg_back_size)
252 bbio->bi_seg_back_size = seg_size;
257 void blk_recalc_rq_segments(struct request *rq)
259 bool no_sg_merge = !!test_bit(QUEUE_FLAG_NO_SG_MERGE,
260 &rq->q->queue_flags);
262 rq->nr_phys_segments = __blk_recalc_rq_segments(rq->q, rq->bio,
266 void blk_recount_segments(struct request_queue *q, struct bio *bio)
268 unsigned short seg_cnt;
270 /* estimate segment number by bi_vcnt for non-cloned bio */
271 if (bio_flagged(bio, BIO_CLONED))
272 seg_cnt = bio_segments(bio);
274 seg_cnt = bio->bi_vcnt;
276 if (test_bit(QUEUE_FLAG_NO_SG_MERGE, &q->queue_flags) &&
277 (seg_cnt < queue_max_segments(q)))
278 bio->bi_phys_segments = seg_cnt;
280 struct bio *nxt = bio->bi_next;
283 bio->bi_phys_segments = __blk_recalc_rq_segments(q, bio, false);
287 bio_set_flag(bio, BIO_SEG_VALID);
289 EXPORT_SYMBOL(blk_recount_segments);
291 static int blk_phys_contig_segment(struct request_queue *q, struct bio *bio,
294 struct bio_vec end_bv = { NULL }, nxt_bv;
295 struct bvec_iter iter;
297 if (!blk_queue_cluster(q))
300 if (bio->bi_seg_back_size + nxt->bi_seg_front_size >
301 queue_max_segment_size(q))
304 if (!bio_has_data(bio))
307 bio_for_each_segment(end_bv, bio, iter)
308 if (end_bv.bv_len == iter.bi_size)
311 nxt_bv = bio_iovec(nxt);
313 if (!BIOVEC_PHYS_MERGEABLE(&end_bv, &nxt_bv))
317 * bio and nxt are contiguous in memory; check if the queue allows
318 * these two to be merged into one
320 if (BIOVEC_SEG_BOUNDARY(q, &end_bv, &nxt_bv))
327 __blk_segment_map_sg(struct request_queue *q, struct bio_vec *bvec,
328 struct scatterlist *sglist, struct bio_vec *bvprv,
329 struct scatterlist **sg, int *nsegs, int *cluster)
332 int nbytes = bvec->bv_len;
334 if (*sg && *cluster) {
335 if ((*sg)->length + nbytes > queue_max_segment_size(q))
338 if (!BIOVEC_PHYS_MERGEABLE(bvprv, bvec))
340 if (!BIOVEC_SEG_BOUNDARY(q, bvprv, bvec))
343 (*sg)->length += nbytes;
350 * If the driver previously mapped a shorter
351 * list, we could see a termination bit
352 * prematurely unless it fully inits the sg
353 * table on each mapping. We KNOW that there
354 * must be more entries here or the driver
355 * would be buggy, so force clear the
356 * termination bit to avoid doing a full
357 * sg_init_table() in drivers for each command.
363 sg_set_page(*sg, bvec->bv_page, nbytes, bvec->bv_offset);
369 static int __blk_bios_map_sg(struct request_queue *q, struct bio *bio,
370 struct scatterlist *sglist,
371 struct scatterlist **sg)
373 struct bio_vec bvec, bvprv = { NULL };
374 struct bvec_iter iter;
378 cluster = blk_queue_cluster(q);
380 if (bio->bi_rw & REQ_DISCARD) {
382 * This is a hack - drivers should be neither modifying the
383 * biovec, nor relying on bi_vcnt - but because of
384 * blk_add_request_payload(), a discard bio may or may not have
385 * a payload we need to set up here (thank you Christoph) and
386 * bi_vcnt is really the only way of telling if we need to.
395 if (bio->bi_rw & REQ_WRITE_SAME) {
398 bvec = bio_iovec(bio);
399 sg_set_page(*sg, bvec.bv_page, bvec.bv_len, bvec.bv_offset);
404 bio_for_each_segment(bvec, bio, iter)
405 __blk_segment_map_sg(q, &bvec, sglist, &bvprv, sg,
412 * map a request to scatterlist, return number of sg entries setup. Caller
413 * must make sure sg can hold rq->nr_phys_segments entries
415 int blk_rq_map_sg(struct request_queue *q, struct request *rq,
416 struct scatterlist *sglist)
418 struct scatterlist *sg = NULL;
422 nsegs = __blk_bios_map_sg(q, rq->bio, sglist, &sg);
424 if (unlikely(rq->cmd_flags & REQ_COPY_USER) &&
425 (blk_rq_bytes(rq) & q->dma_pad_mask)) {
426 unsigned int pad_len =
427 (q->dma_pad_mask & ~blk_rq_bytes(rq)) + 1;
429 sg->length += pad_len;
430 rq->extra_len += pad_len;
433 if (q->dma_drain_size && q->dma_drain_needed(rq)) {
434 if (rq->cmd_flags & REQ_WRITE)
435 memset(q->dma_drain_buffer, 0, q->dma_drain_size);
439 sg_set_page(sg, virt_to_page(q->dma_drain_buffer),
441 ((unsigned long)q->dma_drain_buffer) &
444 rq->extra_len += q->dma_drain_size;
451 * Something must have been wrong if the figured number of
452 * segment is bigger than number of req's physical segments
454 WARN_ON(nsegs > rq->nr_phys_segments);
458 EXPORT_SYMBOL(blk_rq_map_sg);
460 static inline int ll_new_hw_segment(struct request_queue *q,
464 int nr_phys_segs = bio_phys_segments(q, bio);
466 if (req->nr_phys_segments + nr_phys_segs > queue_max_segments(q))
469 if (blk_integrity_merge_bio(q, req, bio) == false)
473 * This will form the start of a new hw segment. Bump both
476 req->nr_phys_segments += nr_phys_segs;
480 req->cmd_flags |= REQ_NOMERGE;
481 if (req == q->last_merge)
482 q->last_merge = NULL;
486 int ll_back_merge_fn(struct request_queue *q, struct request *req,
489 if (req_gap_back_merge(req, bio))
491 if (blk_integrity_rq(req) &&
492 integrity_req_gap_back_merge(req, bio))
494 if (blk_rq_sectors(req) + bio_sectors(bio) >
495 blk_rq_get_max_sectors(req)) {
496 req->cmd_flags |= REQ_NOMERGE;
497 if (req == q->last_merge)
498 q->last_merge = NULL;
501 if (!bio_flagged(req->biotail, BIO_SEG_VALID))
502 blk_recount_segments(q, req->biotail);
503 if (!bio_flagged(bio, BIO_SEG_VALID))
504 blk_recount_segments(q, bio);
506 return ll_new_hw_segment(q, req, bio);
509 int ll_front_merge_fn(struct request_queue *q, struct request *req,
513 if (req_gap_front_merge(req, bio))
515 if (blk_integrity_rq(req) &&
516 integrity_req_gap_front_merge(req, bio))
518 if (blk_rq_sectors(req) + bio_sectors(bio) >
519 blk_rq_get_max_sectors(req)) {
520 req->cmd_flags |= REQ_NOMERGE;
521 if (req == q->last_merge)
522 q->last_merge = NULL;
525 if (!bio_flagged(bio, BIO_SEG_VALID))
526 blk_recount_segments(q, bio);
527 if (!bio_flagged(req->bio, BIO_SEG_VALID))
528 blk_recount_segments(q, req->bio);
530 return ll_new_hw_segment(q, req, bio);
534 * blk-mq uses req->special to carry normal driver per-request payload, it
535 * does not indicate a prepared command that we cannot merge with.
537 static bool req_no_special_merge(struct request *req)
539 struct request_queue *q = req->q;
541 return !q->mq_ops && req->special;
544 static int ll_merge_requests_fn(struct request_queue *q, struct request *req,
545 struct request *next)
547 int total_phys_segments;
548 unsigned int seg_size =
549 req->biotail->bi_seg_back_size + next->bio->bi_seg_front_size;
552 * First check if the either of the requests are re-queued
553 * requests. Can't merge them if they are.
555 if (req_no_special_merge(req) || req_no_special_merge(next))
558 if (req_gap_back_merge(req, next->bio))
562 * Will it become too large?
564 if ((blk_rq_sectors(req) + blk_rq_sectors(next)) >
565 blk_rq_get_max_sectors(req))
568 total_phys_segments = req->nr_phys_segments + next->nr_phys_segments;
569 if (blk_phys_contig_segment(q, req->biotail, next->bio)) {
570 if (req->nr_phys_segments == 1)
571 req->bio->bi_seg_front_size = seg_size;
572 if (next->nr_phys_segments == 1)
573 next->biotail->bi_seg_back_size = seg_size;
574 total_phys_segments--;
577 if (total_phys_segments > queue_max_segments(q))
580 if (blk_integrity_merge_rq(q, req, next) == false)
584 req->nr_phys_segments = total_phys_segments;
589 * blk_rq_set_mixed_merge - mark a request as mixed merge
590 * @rq: request to mark as mixed merge
593 * @rq is about to be mixed merged. Make sure the attributes
594 * which can be mixed are set in each bio and mark @rq as mixed
597 void blk_rq_set_mixed_merge(struct request *rq)
599 unsigned int ff = rq->cmd_flags & REQ_FAILFAST_MASK;
602 if (rq->cmd_flags & REQ_MIXED_MERGE)
606 * @rq will no longer represent mixable attributes for all the
607 * contained bios. It will just track those of the first one.
608 * Distributes the attributs to each bio.
610 for (bio = rq->bio; bio; bio = bio->bi_next) {
611 WARN_ON_ONCE((bio->bi_rw & REQ_FAILFAST_MASK) &&
612 (bio->bi_rw & REQ_FAILFAST_MASK) != ff);
615 rq->cmd_flags |= REQ_MIXED_MERGE;
618 static void blk_account_io_merge(struct request *req)
620 if (blk_do_io_stat(req)) {
621 struct hd_struct *part;
624 cpu = part_stat_lock();
627 part_round_stats(cpu, part);
628 part_dec_in_flight(part, rq_data_dir(req));
636 * Has to be called with the request spinlock acquired
638 static int attempt_merge(struct request_queue *q, struct request *req,
639 struct request *next)
641 if (!rq_mergeable(req) || !rq_mergeable(next))
644 if (!blk_check_merge_flags(req->cmd_flags, next->cmd_flags))
650 if (blk_rq_pos(req) + blk_rq_sectors(req) != blk_rq_pos(next))
653 if (rq_data_dir(req) != rq_data_dir(next)
654 || req->rq_disk != next->rq_disk
655 || req_no_special_merge(next))
658 if (req->cmd_flags & REQ_WRITE_SAME &&
659 !blk_write_same_mergeable(req->bio, next->bio))
663 * If we are allowed to merge, then append bio list
664 * from next to rq and release next. merge_requests_fn
665 * will have updated segment counts, update sector
668 if (!ll_merge_requests_fn(q, req, next))
672 * If failfast settings disagree or any of the two is already
673 * a mixed merge, mark both as mixed before proceeding. This
674 * makes sure that all involved bios have mixable attributes
677 if ((req->cmd_flags | next->cmd_flags) & REQ_MIXED_MERGE ||
678 (req->cmd_flags & REQ_FAILFAST_MASK) !=
679 (next->cmd_flags & REQ_FAILFAST_MASK)) {
680 blk_rq_set_mixed_merge(req);
681 blk_rq_set_mixed_merge(next);
685 * At this point we have either done a back merge
686 * or front merge. We need the smaller start_time of
687 * the merged requests to be the current request
688 * for accounting purposes.
690 if (time_after(req->start_time, next->start_time))
691 req->start_time = next->start_time;
693 req->biotail->bi_next = next->bio;
694 req->biotail = next->biotail;
696 req->__data_len += blk_rq_bytes(next);
698 elv_merge_requests(q, req, next);
701 * 'next' is going away, so update stats accordingly
703 blk_account_io_merge(next);
705 req->ioprio = ioprio_best(req->ioprio, next->ioprio);
706 if (blk_rq_cpu_valid(next))
707 req->cpu = next->cpu;
709 /* owner-ship of bio passed from next to req */
711 __blk_put_request(q, next);
715 int attempt_back_merge(struct request_queue *q, struct request *rq)
717 struct request *next = elv_latter_request(q, rq);
720 return attempt_merge(q, rq, next);
725 int attempt_front_merge(struct request_queue *q, struct request *rq)
727 struct request *prev = elv_former_request(q, rq);
730 return attempt_merge(q, prev, rq);
735 int blk_attempt_req_merge(struct request_queue *q, struct request *rq,
736 struct request *next)
738 return attempt_merge(q, rq, next);
741 bool blk_rq_merge_ok(struct request *rq, struct bio *bio)
743 if (!rq_mergeable(rq) || !bio_mergeable(bio))
746 if (!blk_check_merge_flags(rq->cmd_flags, bio->bi_rw))
749 /* different data direction or already started, don't merge */
750 if (bio_data_dir(bio) != rq_data_dir(rq))
753 /* must be same device and not a special request */
754 if (rq->rq_disk != bio->bi_bdev->bd_disk || req_no_special_merge(rq))
757 /* only merge integrity protected bio into ditto rq */
758 if (blk_integrity_merge_bio(rq->q, rq, bio) == false)
761 /* must be using the same buffer */
762 if (rq->cmd_flags & REQ_WRITE_SAME &&
763 !blk_write_same_mergeable(rq->bio, bio))
769 int blk_try_merge(struct request *rq, struct bio *bio)
771 if (blk_rq_pos(rq) + blk_rq_sectors(rq) == bio->bi_iter.bi_sector)
772 return ELEVATOR_BACK_MERGE;
773 else if (blk_rq_pos(rq) - bio_sectors(bio) == bio->bi_iter.bi_sector)
774 return ELEVATOR_FRONT_MERGE;
775 return ELEVATOR_NO_MERGE;