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,
16 unsigned int max_discard_sectors, granularity;
19 unsigned split_sectors;
21 /* Zero-sector (unknown) and one-sector granularities are the same. */
22 granularity = max(q->limits.discard_granularity >> 9, 1U);
24 max_discard_sectors = min(q->limits.max_discard_sectors, UINT_MAX >> 9);
25 max_discard_sectors -= max_discard_sectors % granularity;
27 if (unlikely(!max_discard_sectors)) {
32 if (bio_sectors(bio) <= max_discard_sectors)
35 split_sectors = max_discard_sectors;
38 * If the next starting sector would be misaligned, stop the discard at
39 * the previous aligned sector.
41 alignment = (q->limits.discard_alignment >> 9) % granularity;
43 tmp = bio->bi_iter.bi_sector + split_sectors - alignment;
44 tmp = sector_div(tmp, granularity);
46 if (split_sectors > tmp)
49 return bio_split(bio, split_sectors, GFP_NOIO, bs);
52 static struct bio *blk_bio_write_same_split(struct request_queue *q,
56 if (!q->limits.max_write_same_sectors)
59 if (bio_sectors(bio) <= q->limits.max_write_same_sectors)
62 return bio_split(bio, q->limits.max_write_same_sectors, GFP_NOIO, bs);
65 static struct bio *blk_bio_segment_split(struct request_queue *q,
70 struct bio_vec bv, bvprv, *bvprvp = NULL;
71 struct bvec_iter iter;
72 unsigned seg_size = 0, nsegs = 0, sectors = 0;
74 bio_for_each_segment(bv, bio, iter) {
75 sectors += bv.bv_len >> 9;
77 if (sectors > queue_max_sectors(q))
81 * If the queue doesn't support SG gaps and adding this
82 * offset would create a gap, disallow it.
84 if (bvprvp && bvec_gap_to_prev(q, bvprvp, bv.bv_offset))
87 if (bvprvp && blk_queue_cluster(q)) {
88 if (seg_size + bv.bv_len > queue_max_segment_size(q))
90 if (!BIOVEC_PHYS_MERGEABLE(bvprvp, &bv))
92 if (!BIOVEC_SEG_BOUNDARY(q, bvprvp, &bv))
95 seg_size += bv.bv_len;
101 if (nsegs == queue_max_segments(q))
107 seg_size = bv.bv_len;
112 split = bio_clone_bioset(bio, GFP_NOIO, bs);
114 split->bi_iter.bi_size -= iter.bi_size;
117 if (bio_integrity(bio)) {
118 bio_integrity_advance(bio, split->bi_iter.bi_size);
119 bio_integrity_trim(split, 0, bio_sectors(split));
125 void blk_queue_split(struct request_queue *q, struct bio **bio,
130 if ((*bio)->bi_rw & REQ_DISCARD)
131 split = blk_bio_discard_split(q, *bio, bs);
132 else if ((*bio)->bi_rw & REQ_WRITE_SAME)
133 split = blk_bio_write_same_split(q, *bio, bs);
135 split = blk_bio_segment_split(q, *bio, q->bio_split);
138 bio_chain(split, *bio);
139 generic_make_request(*bio);
143 EXPORT_SYMBOL(blk_queue_split);
145 static unsigned int __blk_recalc_rq_segments(struct request_queue *q,
149 struct bio_vec bv, bvprv = { NULL };
150 int cluster, prev = 0;
151 unsigned int seg_size, nr_phys_segs;
152 struct bio *fbio, *bbio;
153 struct bvec_iter iter;
159 * This should probably be returning 0, but blk_add_request_payload()
162 if (bio->bi_rw & REQ_DISCARD)
165 if (bio->bi_rw & REQ_WRITE_SAME)
169 cluster = blk_queue_cluster(q);
173 bio_for_each_segment(bv, bio, iter) {
175 * If SG merging is disabled, each bio vector is
181 if (prev && cluster) {
182 if (seg_size + bv.bv_len
183 > queue_max_segment_size(q))
185 if (!BIOVEC_PHYS_MERGEABLE(&bvprv, &bv))
187 if (!BIOVEC_SEG_BOUNDARY(q, &bvprv, &bv))
190 seg_size += bv.bv_len;
195 if (nr_phys_segs == 1 && seg_size >
196 fbio->bi_seg_front_size)
197 fbio->bi_seg_front_size = seg_size;
202 seg_size = bv.bv_len;
207 if (nr_phys_segs == 1 && seg_size > fbio->bi_seg_front_size)
208 fbio->bi_seg_front_size = seg_size;
209 if (seg_size > bbio->bi_seg_back_size)
210 bbio->bi_seg_back_size = seg_size;
215 void blk_recalc_rq_segments(struct request *rq)
217 bool no_sg_merge = !!test_bit(QUEUE_FLAG_NO_SG_MERGE,
218 &rq->q->queue_flags);
220 rq->nr_phys_segments = __blk_recalc_rq_segments(rq->q, rq->bio,
224 void blk_recount_segments(struct request_queue *q, struct bio *bio)
226 unsigned short seg_cnt;
228 /* estimate segment number by bi_vcnt for non-cloned bio */
229 if (bio_flagged(bio, BIO_CLONED))
230 seg_cnt = bio_segments(bio);
232 seg_cnt = bio->bi_vcnt;
234 if (test_bit(QUEUE_FLAG_NO_SG_MERGE, &q->queue_flags) &&
235 (seg_cnt < queue_max_segments(q)))
236 bio->bi_phys_segments = seg_cnt;
238 struct bio *nxt = bio->bi_next;
241 bio->bi_phys_segments = __blk_recalc_rq_segments(q, bio, false);
245 bio_set_flag(bio, BIO_SEG_VALID);
247 EXPORT_SYMBOL(blk_recount_segments);
249 static int blk_phys_contig_segment(struct request_queue *q, struct bio *bio,
252 struct bio_vec end_bv = { NULL }, nxt_bv;
253 struct bvec_iter iter;
255 if (!blk_queue_cluster(q))
258 if (bio->bi_seg_back_size + nxt->bi_seg_front_size >
259 queue_max_segment_size(q))
262 if (!bio_has_data(bio))
265 bio_for_each_segment(end_bv, bio, iter)
266 if (end_bv.bv_len == iter.bi_size)
269 nxt_bv = bio_iovec(nxt);
271 if (!BIOVEC_PHYS_MERGEABLE(&end_bv, &nxt_bv))
275 * bio and nxt are contiguous in memory; check if the queue allows
276 * these two to be merged into one
278 if (BIOVEC_SEG_BOUNDARY(q, &end_bv, &nxt_bv))
285 __blk_segment_map_sg(struct request_queue *q, struct bio_vec *bvec,
286 struct scatterlist *sglist, struct bio_vec *bvprv,
287 struct scatterlist **sg, int *nsegs, int *cluster)
290 int nbytes = bvec->bv_len;
292 if (*sg && *cluster) {
293 if ((*sg)->length + nbytes > queue_max_segment_size(q))
296 if (!BIOVEC_PHYS_MERGEABLE(bvprv, bvec))
298 if (!BIOVEC_SEG_BOUNDARY(q, bvprv, bvec))
301 (*sg)->length += nbytes;
308 * If the driver previously mapped a shorter
309 * list, we could see a termination bit
310 * prematurely unless it fully inits the sg
311 * table on each mapping. We KNOW that there
312 * must be more entries here or the driver
313 * would be buggy, so force clear the
314 * termination bit to avoid doing a full
315 * sg_init_table() in drivers for each command.
321 sg_set_page(*sg, bvec->bv_page, nbytes, bvec->bv_offset);
327 static int __blk_bios_map_sg(struct request_queue *q, struct bio *bio,
328 struct scatterlist *sglist,
329 struct scatterlist **sg)
331 struct bio_vec bvec, bvprv = { NULL };
332 struct bvec_iter iter;
336 cluster = blk_queue_cluster(q);
338 if (bio->bi_rw & REQ_DISCARD) {
340 * This is a hack - drivers should be neither modifying the
341 * biovec, nor relying on bi_vcnt - but because of
342 * blk_add_request_payload(), a discard bio may or may not have
343 * a payload we need to set up here (thank you Christoph) and
344 * bi_vcnt is really the only way of telling if we need to.
353 if (bio->bi_rw & REQ_WRITE_SAME) {
356 bvec = bio_iovec(bio);
357 sg_set_page(*sg, bvec.bv_page, bvec.bv_len, bvec.bv_offset);
362 bio_for_each_segment(bvec, bio, iter)
363 __blk_segment_map_sg(q, &bvec, sglist, &bvprv, sg,
370 * map a request to scatterlist, return number of sg entries setup. Caller
371 * must make sure sg can hold rq->nr_phys_segments entries
373 int blk_rq_map_sg(struct request_queue *q, struct request *rq,
374 struct scatterlist *sglist)
376 struct scatterlist *sg = NULL;
380 nsegs = __blk_bios_map_sg(q, rq->bio, sglist, &sg);
382 if (unlikely(rq->cmd_flags & REQ_COPY_USER) &&
383 (blk_rq_bytes(rq) & q->dma_pad_mask)) {
384 unsigned int pad_len =
385 (q->dma_pad_mask & ~blk_rq_bytes(rq)) + 1;
387 sg->length += pad_len;
388 rq->extra_len += pad_len;
391 if (q->dma_drain_size && q->dma_drain_needed(rq)) {
392 if (rq->cmd_flags & REQ_WRITE)
393 memset(q->dma_drain_buffer, 0, q->dma_drain_size);
397 sg_set_page(sg, virt_to_page(q->dma_drain_buffer),
399 ((unsigned long)q->dma_drain_buffer) &
402 rq->extra_len += q->dma_drain_size;
410 EXPORT_SYMBOL(blk_rq_map_sg);
412 static inline int ll_new_hw_segment(struct request_queue *q,
416 int nr_phys_segs = bio_phys_segments(q, bio);
418 if (req->nr_phys_segments + nr_phys_segs > queue_max_segments(q))
421 if (blk_integrity_merge_bio(q, req, bio) == false)
425 * This will form the start of a new hw segment. Bump both
428 req->nr_phys_segments += nr_phys_segs;
432 req->cmd_flags |= REQ_NOMERGE;
433 if (req == q->last_merge)
434 q->last_merge = NULL;
438 int ll_back_merge_fn(struct request_queue *q, struct request *req,
441 if (blk_rq_sectors(req) + bio_sectors(bio) >
442 blk_rq_get_max_sectors(req)) {
443 req->cmd_flags |= REQ_NOMERGE;
444 if (req == q->last_merge)
445 q->last_merge = NULL;
448 if (!bio_flagged(req->biotail, BIO_SEG_VALID))
449 blk_recount_segments(q, req->biotail);
450 if (!bio_flagged(bio, BIO_SEG_VALID))
451 blk_recount_segments(q, bio);
453 return ll_new_hw_segment(q, req, bio);
456 int ll_front_merge_fn(struct request_queue *q, struct request *req,
459 if (blk_rq_sectors(req) + bio_sectors(bio) >
460 blk_rq_get_max_sectors(req)) {
461 req->cmd_flags |= REQ_NOMERGE;
462 if (req == q->last_merge)
463 q->last_merge = NULL;
466 if (!bio_flagged(bio, BIO_SEG_VALID))
467 blk_recount_segments(q, bio);
468 if (!bio_flagged(req->bio, BIO_SEG_VALID))
469 blk_recount_segments(q, req->bio);
471 return ll_new_hw_segment(q, req, bio);
475 * blk-mq uses req->special to carry normal driver per-request payload, it
476 * does not indicate a prepared command that we cannot merge with.
478 static bool req_no_special_merge(struct request *req)
480 struct request_queue *q = req->q;
482 return !q->mq_ops && req->special;
485 static int req_gap_to_prev(struct request *req, struct bio *next)
487 struct bio *prev = req->biotail;
489 return bvec_gap_to_prev(req->q, &prev->bi_io_vec[prev->bi_vcnt - 1],
490 next->bi_io_vec[0].bv_offset);
493 static int ll_merge_requests_fn(struct request_queue *q, struct request *req,
494 struct request *next)
496 int total_phys_segments;
497 unsigned int seg_size =
498 req->biotail->bi_seg_back_size + next->bio->bi_seg_front_size;
501 * First check if the either of the requests are re-queued
502 * requests. Can't merge them if they are.
504 if (req_no_special_merge(req) || req_no_special_merge(next))
507 if (req_gap_to_prev(req, next->bio))
511 * Will it become too large?
513 if ((blk_rq_sectors(req) + blk_rq_sectors(next)) >
514 blk_rq_get_max_sectors(req))
517 total_phys_segments = req->nr_phys_segments + next->nr_phys_segments;
518 if (blk_phys_contig_segment(q, req->biotail, next->bio)) {
519 if (req->nr_phys_segments == 1)
520 req->bio->bi_seg_front_size = seg_size;
521 if (next->nr_phys_segments == 1)
522 next->biotail->bi_seg_back_size = seg_size;
523 total_phys_segments--;
526 if (total_phys_segments > queue_max_segments(q))
529 if (blk_integrity_merge_rq(q, req, next) == false)
533 req->nr_phys_segments = total_phys_segments;
538 * blk_rq_set_mixed_merge - mark a request as mixed merge
539 * @rq: request to mark as mixed merge
542 * @rq is about to be mixed merged. Make sure the attributes
543 * which can be mixed are set in each bio and mark @rq as mixed
546 void blk_rq_set_mixed_merge(struct request *rq)
548 unsigned int ff = rq->cmd_flags & REQ_FAILFAST_MASK;
551 if (rq->cmd_flags & REQ_MIXED_MERGE)
555 * @rq will no longer represent mixable attributes for all the
556 * contained bios. It will just track those of the first one.
557 * Distributes the attributs to each bio.
559 for (bio = rq->bio; bio; bio = bio->bi_next) {
560 WARN_ON_ONCE((bio->bi_rw & REQ_FAILFAST_MASK) &&
561 (bio->bi_rw & REQ_FAILFAST_MASK) != ff);
564 rq->cmd_flags |= REQ_MIXED_MERGE;
567 static void blk_account_io_merge(struct request *req)
569 if (blk_do_io_stat(req)) {
570 struct hd_struct *part;
573 cpu = part_stat_lock();
576 part_round_stats(cpu, part);
577 part_dec_in_flight(part, rq_data_dir(req));
585 * Has to be called with the request spinlock acquired
587 static int attempt_merge(struct request_queue *q, struct request *req,
588 struct request *next)
590 if (!rq_mergeable(req) || !rq_mergeable(next))
593 if (!blk_check_merge_flags(req->cmd_flags, next->cmd_flags))
599 if (blk_rq_pos(req) + blk_rq_sectors(req) != blk_rq_pos(next))
602 if (rq_data_dir(req) != rq_data_dir(next)
603 || req->rq_disk != next->rq_disk
604 || req_no_special_merge(next))
607 if (req->cmd_flags & REQ_WRITE_SAME &&
608 !blk_write_same_mergeable(req->bio, next->bio))
612 * If we are allowed to merge, then append bio list
613 * from next to rq and release next. merge_requests_fn
614 * will have updated segment counts, update sector
617 if (!ll_merge_requests_fn(q, req, next))
621 * If failfast settings disagree or any of the two is already
622 * a mixed merge, mark both as mixed before proceeding. This
623 * makes sure that all involved bios have mixable attributes
626 if ((req->cmd_flags | next->cmd_flags) & REQ_MIXED_MERGE ||
627 (req->cmd_flags & REQ_FAILFAST_MASK) !=
628 (next->cmd_flags & REQ_FAILFAST_MASK)) {
629 blk_rq_set_mixed_merge(req);
630 blk_rq_set_mixed_merge(next);
634 * At this point we have either done a back merge
635 * or front merge. We need the smaller start_time of
636 * the merged requests to be the current request
637 * for accounting purposes.
639 if (time_after(req->start_time, next->start_time))
640 req->start_time = next->start_time;
642 req->biotail->bi_next = next->bio;
643 req->biotail = next->biotail;
645 req->__data_len += blk_rq_bytes(next);
647 elv_merge_requests(q, req, next);
650 * 'next' is going away, so update stats accordingly
652 blk_account_io_merge(next);
654 req->ioprio = ioprio_best(req->ioprio, next->ioprio);
655 if (blk_rq_cpu_valid(next))
656 req->cpu = next->cpu;
658 /* owner-ship of bio passed from next to req */
660 __blk_put_request(q, next);
664 int attempt_back_merge(struct request_queue *q, struct request *rq)
666 struct request *next = elv_latter_request(q, rq);
669 return attempt_merge(q, rq, next);
674 int attempt_front_merge(struct request_queue *q, struct request *rq)
676 struct request *prev = elv_former_request(q, rq);
679 return attempt_merge(q, prev, rq);
684 int blk_attempt_req_merge(struct request_queue *q, struct request *rq,
685 struct request *next)
687 return attempt_merge(q, rq, next);
690 bool blk_rq_merge_ok(struct request *rq, struct bio *bio)
692 if (!rq_mergeable(rq) || !bio_mergeable(bio))
695 if (!blk_check_merge_flags(rq->cmd_flags, bio->bi_rw))
698 /* different data direction or already started, don't merge */
699 if (bio_data_dir(bio) != rq_data_dir(rq))
702 /* must be same device and not a special request */
703 if (rq->rq_disk != bio->bi_bdev->bd_disk || req_no_special_merge(rq))
706 /* only merge integrity protected bio into ditto rq */
707 if (blk_integrity_merge_bio(rq->q, rq, bio) == false)
710 /* must be using the same buffer */
711 if (rq->cmd_flags & REQ_WRITE_SAME &&
712 !blk_write_same_mergeable(rq->bio, bio))
715 /* Only check gaps if the bio carries data */
716 if (bio_has_data(bio) && req_gap_to_prev(rq, bio))
722 int blk_try_merge(struct request *rq, struct bio *bio)
724 if (blk_rq_pos(rq) + blk_rq_sectors(rq) == bio->bi_iter.bi_sector)
725 return ELEVATOR_BACK_MERGE;
726 else if (blk_rq_pos(rq) - bio_sectors(bio) == bio->bi_iter.bi_sector)
727 return ELEVATOR_FRONT_MERGE;
728 return ELEVATOR_NO_MERGE;