1 // SPDX-License-Identifier: GPL-2.0
3 * Functions related to segment and merge handling
5 #include <linux/kernel.h>
6 #include <linux/module.h>
8 #include <linux/blkdev.h>
9 #include <linux/scatterlist.h>
11 #include <trace/events/block.h>
15 static struct bio
*blk_bio_discard_split(struct request_queue
*q
,
20 unsigned int max_discard_sectors
, granularity
;
23 unsigned split_sectors
;
27 /* Zero-sector (unknown) and one-sector granularities are the same. */
28 granularity
= max(q
->limits
.discard_granularity
>> 9, 1U);
30 max_discard_sectors
= min(q
->limits
.max_discard_sectors
,
31 bio_allowed_max_sectors(q
));
32 max_discard_sectors
-= max_discard_sectors
% granularity
;
34 if (unlikely(!max_discard_sectors
)) {
39 if (bio_sectors(bio
) <= max_discard_sectors
)
42 split_sectors
= max_discard_sectors
;
45 * If the next starting sector would be misaligned, stop the discard at
46 * the previous aligned sector.
48 alignment
= (q
->limits
.discard_alignment
>> 9) % granularity
;
50 tmp
= bio
->bi_iter
.bi_sector
+ split_sectors
- alignment
;
51 tmp
= sector_div(tmp
, granularity
);
53 if (split_sectors
> tmp
)
56 return bio_split(bio
, split_sectors
, GFP_NOIO
, bs
);
59 static struct bio
*blk_bio_write_zeroes_split(struct request_queue
*q
,
60 struct bio
*bio
, struct bio_set
*bs
, unsigned *nsegs
)
64 if (!q
->limits
.max_write_zeroes_sectors
)
67 if (bio_sectors(bio
) <= q
->limits
.max_write_zeroes_sectors
)
70 return bio_split(bio
, q
->limits
.max_write_zeroes_sectors
, GFP_NOIO
, bs
);
73 static struct bio
*blk_bio_write_same_split(struct request_queue
*q
,
80 if (!q
->limits
.max_write_same_sectors
)
83 if (bio_sectors(bio
) <= q
->limits
.max_write_same_sectors
)
86 return bio_split(bio
, q
->limits
.max_write_same_sectors
, GFP_NOIO
, bs
);
89 static inline unsigned get_max_io_size(struct request_queue
*q
,
92 unsigned sectors
= blk_max_size_offset(q
, bio
->bi_iter
.bi_sector
);
93 unsigned mask
= queue_logical_block_size(q
) - 1;
95 /* aligned to logical block size */
96 sectors
&= ~(mask
>> 9);
101 static struct bio
*blk_bio_segment_split(struct request_queue
*q
,
106 struct bio_vec bv
, bvprv
, *bvprvp
= NULL
;
107 struct bvec_iter iter
;
108 unsigned seg_size
= 0, nsegs
= 0, sectors
= 0;
109 unsigned front_seg_size
= bio
->bi_seg_front_size
;
110 bool do_split
= true;
111 struct bio
*new = NULL
;
112 const unsigned max_sectors
= get_max_io_size(q
, bio
);
114 bio_for_each_segment(bv
, bio
, iter
) {
116 * If the queue doesn't support SG gaps and adding this
117 * offset would create a gap, disallow it.
119 if (bvprvp
&& bvec_gap_to_prev(q
, bvprvp
, bv
.bv_offset
))
122 if (sectors
+ (bv
.bv_len
>> 9) > max_sectors
) {
124 * Consider this a new segment if we're splitting in
125 * the middle of this vector.
127 if (nsegs
< queue_max_segments(q
) &&
128 sectors
< max_sectors
) {
130 sectors
= max_sectors
;
134 /* Make this single bvec as the 1st segment */
137 if (bvprvp
&& blk_queue_cluster(q
)) {
138 if (seg_size
+ bv
.bv_len
> queue_max_segment_size(q
))
140 if (!BIOVEC_PHYS_MERGEABLE(bvprvp
, &bv
))
142 if (!BIOVEC_SEG_BOUNDARY(q
, bvprvp
, &bv
))
145 seg_size
+= bv
.bv_len
;
148 sectors
+= bv
.bv_len
>> 9;
150 if (nsegs
== 1 && seg_size
> front_seg_size
)
151 front_seg_size
= seg_size
;
155 if (nsegs
== queue_max_segments(q
))
161 seg_size
= bv
.bv_len
;
162 sectors
+= bv
.bv_len
>> 9;
164 if (nsegs
== 1 && seg_size
> front_seg_size
)
165 front_seg_size
= seg_size
;
173 new = bio_split(bio
, sectors
, GFP_NOIO
, bs
);
178 bio
->bi_seg_front_size
= front_seg_size
;
179 if (seg_size
> bio
->bi_seg_back_size
)
180 bio
->bi_seg_back_size
= seg_size
;
182 return do_split
? new : NULL
;
185 void blk_queue_split(struct request_queue
*q
, struct bio
**bio
)
187 struct bio
*split
, *res
;
190 switch (bio_op(*bio
)) {
192 case REQ_OP_SECURE_ERASE
:
193 split
= blk_bio_discard_split(q
, *bio
, q
->bio_split
, &nsegs
);
195 case REQ_OP_WRITE_ZEROES
:
196 split
= blk_bio_write_zeroes_split(q
, *bio
, q
->bio_split
, &nsegs
);
198 case REQ_OP_WRITE_SAME
:
199 split
= blk_bio_write_same_split(q
, *bio
, q
->bio_split
, &nsegs
);
202 split
= blk_bio_segment_split(q
, *bio
, q
->bio_split
, &nsegs
);
206 /* physical segments can be figured out during splitting */
207 res
= split
? split
: *bio
;
208 res
->bi_phys_segments
= nsegs
;
209 bio_set_flag(res
, BIO_SEG_VALID
);
212 /* there isn't chance to merge the splitted bio */
213 split
->bi_opf
|= REQ_NOMERGE
;
215 bio_chain(split
, *bio
);
216 trace_block_split(q
, split
, (*bio
)->bi_iter
.bi_sector
);
217 generic_make_request(*bio
);
221 EXPORT_SYMBOL(blk_queue_split
);
223 static unsigned int __blk_recalc_rq_segments(struct request_queue
*q
,
227 struct bio_vec bv
, bvprv
= { NULL
};
228 int cluster
, prev
= 0;
229 unsigned int seg_size
, nr_phys_segs
;
230 struct bio
*fbio
, *bbio
;
231 struct bvec_iter iter
;
236 switch (bio_op(bio
)) {
238 case REQ_OP_SECURE_ERASE
:
239 case REQ_OP_WRITE_ZEROES
:
241 case REQ_OP_WRITE_SAME
:
246 cluster
= blk_queue_cluster(q
);
250 bio_for_each_segment(bv
, bio
, iter
) {
252 * If SG merging is disabled, each bio vector is
258 if (prev
&& cluster
) {
259 if (seg_size
+ bv
.bv_len
260 > queue_max_segment_size(q
))
262 if (!BIOVEC_PHYS_MERGEABLE(&bvprv
, &bv
))
264 if (!BIOVEC_SEG_BOUNDARY(q
, &bvprv
, &bv
))
267 seg_size
+= bv
.bv_len
;
272 if (nr_phys_segs
== 1 && seg_size
>
273 fbio
->bi_seg_front_size
)
274 fbio
->bi_seg_front_size
= seg_size
;
279 seg_size
= bv
.bv_len
;
284 if (nr_phys_segs
== 1 && seg_size
> fbio
->bi_seg_front_size
)
285 fbio
->bi_seg_front_size
= seg_size
;
286 if (seg_size
> bbio
->bi_seg_back_size
)
287 bbio
->bi_seg_back_size
= seg_size
;
292 void blk_recalc_rq_segments(struct request
*rq
)
294 bool no_sg_merge
= !!test_bit(QUEUE_FLAG_NO_SG_MERGE
,
295 &rq
->q
->queue_flags
);
297 rq
->nr_phys_segments
= __blk_recalc_rq_segments(rq
->q
, rq
->bio
,
301 void blk_recount_segments(struct request_queue
*q
, struct bio
*bio
)
303 unsigned short seg_cnt
= bio_segments(bio
);
305 if (test_bit(QUEUE_FLAG_NO_SG_MERGE
, &q
->queue_flags
) &&
306 (seg_cnt
< queue_max_segments(q
)))
307 bio
->bi_phys_segments
= seg_cnt
;
309 struct bio
*nxt
= bio
->bi_next
;
312 bio
->bi_phys_segments
= __blk_recalc_rq_segments(q
, bio
, false);
316 bio_set_flag(bio
, BIO_SEG_VALID
);
318 EXPORT_SYMBOL(blk_recount_segments
);
320 static int blk_phys_contig_segment(struct request_queue
*q
, struct bio
*bio
,
323 struct bio_vec end_bv
= { NULL
}, nxt_bv
;
325 if (!blk_queue_cluster(q
))
328 if (bio
->bi_seg_back_size
+ nxt
->bi_seg_front_size
>
329 queue_max_segment_size(q
))
332 if (!bio_has_data(bio
))
335 bio_get_last_bvec(bio
, &end_bv
);
336 bio_get_first_bvec(nxt
, &nxt_bv
);
338 if (!BIOVEC_PHYS_MERGEABLE(&end_bv
, &nxt_bv
))
342 * bio and nxt are contiguous in memory; check if the queue allows
343 * these two to be merged into one
345 if (BIOVEC_SEG_BOUNDARY(q
, &end_bv
, &nxt_bv
))
352 __blk_segment_map_sg(struct request_queue
*q
, struct bio_vec
*bvec
,
353 struct scatterlist
*sglist
, struct bio_vec
*bvprv
,
354 struct scatterlist
**sg
, int *nsegs
, int *cluster
)
357 int nbytes
= bvec
->bv_len
;
359 if (*sg
&& *cluster
) {
360 if ((*sg
)->length
+ nbytes
> queue_max_segment_size(q
))
363 if (!BIOVEC_PHYS_MERGEABLE(bvprv
, bvec
))
365 if (!BIOVEC_SEG_BOUNDARY(q
, bvprv
, bvec
))
368 (*sg
)->length
+= nbytes
;
375 * If the driver previously mapped a shorter
376 * list, we could see a termination bit
377 * prematurely unless it fully inits the sg
378 * table on each mapping. We KNOW that there
379 * must be more entries here or the driver
380 * would be buggy, so force clear the
381 * termination bit to avoid doing a full
382 * sg_init_table() in drivers for each command.
388 sg_set_page(*sg
, bvec
->bv_page
, nbytes
, bvec
->bv_offset
);
394 static inline int __blk_bvec_map_sg(struct request_queue
*q
, struct bio_vec bv
,
395 struct scatterlist
*sglist
, struct scatterlist
**sg
)
398 sg_set_page(*sg
, bv
.bv_page
, bv
.bv_len
, bv
.bv_offset
);
402 static int __blk_bios_map_sg(struct request_queue
*q
, struct bio
*bio
,
403 struct scatterlist
*sglist
,
404 struct scatterlist
**sg
)
406 struct bio_vec bvec
, bvprv
= { NULL
};
407 struct bvec_iter iter
;
408 int cluster
= blk_queue_cluster(q
), nsegs
= 0;
411 bio_for_each_segment(bvec
, bio
, iter
)
412 __blk_segment_map_sg(q
, &bvec
, sglist
, &bvprv
, sg
,
419 * map a request to scatterlist, return number of sg entries setup. Caller
420 * must make sure sg can hold rq->nr_phys_segments entries
422 int blk_rq_map_sg(struct request_queue
*q
, struct request
*rq
,
423 struct scatterlist
*sglist
)
425 struct scatterlist
*sg
= NULL
;
428 if (rq
->rq_flags
& RQF_SPECIAL_PAYLOAD
)
429 nsegs
= __blk_bvec_map_sg(q
, rq
->special_vec
, sglist
, &sg
);
430 else if (rq
->bio
&& bio_op(rq
->bio
) == REQ_OP_WRITE_SAME
)
431 nsegs
= __blk_bvec_map_sg(q
, bio_iovec(rq
->bio
), sglist
, &sg
);
433 nsegs
= __blk_bios_map_sg(q
, rq
->bio
, sglist
, &sg
);
435 if (unlikely(rq
->rq_flags
& RQF_COPY_USER
) &&
436 (blk_rq_bytes(rq
) & q
->dma_pad_mask
)) {
437 unsigned int pad_len
=
438 (q
->dma_pad_mask
& ~blk_rq_bytes(rq
)) + 1;
440 sg
->length
+= pad_len
;
441 rq
->extra_len
+= pad_len
;
444 if (q
->dma_drain_size
&& q
->dma_drain_needed(rq
)) {
445 if (op_is_write(req_op(rq
)))
446 memset(q
->dma_drain_buffer
, 0, q
->dma_drain_size
);
450 sg_set_page(sg
, virt_to_page(q
->dma_drain_buffer
),
452 ((unsigned long)q
->dma_drain_buffer
) &
455 rq
->extra_len
+= q
->dma_drain_size
;
462 * Something must have been wrong if the figured number of
463 * segment is bigger than number of req's physical segments
465 WARN_ON(nsegs
> blk_rq_nr_phys_segments(rq
));
469 EXPORT_SYMBOL(blk_rq_map_sg
);
471 static inline int ll_new_hw_segment(struct request_queue
*q
,
475 int nr_phys_segs
= bio_phys_segments(q
, bio
);
477 if (req
->nr_phys_segments
+ nr_phys_segs
> queue_max_segments(q
))
480 if (blk_integrity_merge_bio(q
, req
, bio
) == false)
484 * This will form the start of a new hw segment. Bump both
487 req
->nr_phys_segments
+= nr_phys_segs
;
491 req_set_nomerge(q
, req
);
495 int ll_back_merge_fn(struct request_queue
*q
, struct request
*req
,
498 if (req_gap_back_merge(req
, bio
))
500 if (blk_integrity_rq(req
) &&
501 integrity_req_gap_back_merge(req
, bio
))
503 if (blk_rq_sectors(req
) + bio_sectors(bio
) >
504 blk_rq_get_max_sectors(req
, blk_rq_pos(req
))) {
505 req_set_nomerge(q
, req
);
508 if (!bio_flagged(req
->biotail
, BIO_SEG_VALID
))
509 blk_recount_segments(q
, req
->biotail
);
510 if (!bio_flagged(bio
, BIO_SEG_VALID
))
511 blk_recount_segments(q
, bio
);
513 return ll_new_hw_segment(q
, req
, bio
);
516 int ll_front_merge_fn(struct request_queue
*q
, struct request
*req
,
520 if (req_gap_front_merge(req
, bio
))
522 if (blk_integrity_rq(req
) &&
523 integrity_req_gap_front_merge(req
, bio
))
525 if (blk_rq_sectors(req
) + bio_sectors(bio
) >
526 blk_rq_get_max_sectors(req
, bio
->bi_iter
.bi_sector
)) {
527 req_set_nomerge(q
, req
);
530 if (!bio_flagged(bio
, BIO_SEG_VALID
))
531 blk_recount_segments(q
, bio
);
532 if (!bio_flagged(req
->bio
, BIO_SEG_VALID
))
533 blk_recount_segments(q
, req
->bio
);
535 return ll_new_hw_segment(q
, req
, bio
);
539 * blk-mq uses req->special to carry normal driver per-request payload, it
540 * does not indicate a prepared command that we cannot merge with.
542 static bool req_no_special_merge(struct request
*req
)
544 struct request_queue
*q
= req
->q
;
546 return !q
->mq_ops
&& req
->special
;
549 static bool req_attempt_discard_merge(struct request_queue
*q
, struct request
*req
,
550 struct request
*next
)
552 unsigned short segments
= blk_rq_nr_discard_segments(req
);
554 if (segments
>= queue_max_discard_segments(q
))
556 if (blk_rq_sectors(req
) + bio_sectors(next
->bio
) >
557 blk_rq_get_max_sectors(req
, blk_rq_pos(req
)))
560 req
->nr_phys_segments
= segments
+ blk_rq_nr_discard_segments(next
);
563 req_set_nomerge(q
, req
);
567 static int ll_merge_requests_fn(struct request_queue
*q
, struct request
*req
,
568 struct request
*next
)
570 int total_phys_segments
;
571 unsigned int seg_size
=
572 req
->biotail
->bi_seg_back_size
+ next
->bio
->bi_seg_front_size
;
575 * First check if the either of the requests are re-queued
576 * requests. Can't merge them if they are.
578 if (req_no_special_merge(req
) || req_no_special_merge(next
))
581 if (req_gap_back_merge(req
, next
->bio
))
585 * Will it become too large?
587 if ((blk_rq_sectors(req
) + blk_rq_sectors(next
)) >
588 blk_rq_get_max_sectors(req
, blk_rq_pos(req
)))
591 total_phys_segments
= req
->nr_phys_segments
+ next
->nr_phys_segments
;
592 if (blk_phys_contig_segment(q
, req
->biotail
, next
->bio
)) {
593 if (req
->nr_phys_segments
== 1)
594 req
->bio
->bi_seg_front_size
= seg_size
;
595 if (next
->nr_phys_segments
== 1)
596 next
->biotail
->bi_seg_back_size
= seg_size
;
597 total_phys_segments
--;
600 if (total_phys_segments
> queue_max_segments(q
))
603 if (blk_integrity_merge_rq(q
, req
, next
) == false)
607 req
->nr_phys_segments
= total_phys_segments
;
612 * blk_rq_set_mixed_merge - mark a request as mixed merge
613 * @rq: request to mark as mixed merge
616 * @rq is about to be mixed merged. Make sure the attributes
617 * which can be mixed are set in each bio and mark @rq as mixed
620 void blk_rq_set_mixed_merge(struct request
*rq
)
622 unsigned int ff
= rq
->cmd_flags
& REQ_FAILFAST_MASK
;
625 if (rq
->rq_flags
& RQF_MIXED_MERGE
)
629 * @rq will no longer represent mixable attributes for all the
630 * contained bios. It will just track those of the first one.
631 * Distributes the attributs to each bio.
633 for (bio
= rq
->bio
; bio
; bio
= bio
->bi_next
) {
634 WARN_ON_ONCE((bio
->bi_opf
& REQ_FAILFAST_MASK
) &&
635 (bio
->bi_opf
& REQ_FAILFAST_MASK
) != ff
);
638 rq
->rq_flags
|= RQF_MIXED_MERGE
;
641 static void blk_account_io_merge(struct request
*req
)
643 if (blk_do_io_stat(req
)) {
644 struct hd_struct
*part
;
647 cpu
= part_stat_lock();
650 part_round_stats(req
->q
, cpu
, part
);
651 part_dec_in_flight(req
->q
, part
, rq_data_dir(req
));
658 * Two cases of handling DISCARD merge:
659 * If max_discard_segments > 1, the driver takes every bio
660 * as a range and send them to controller together. The ranges
661 * needn't to be contiguous.
662 * Otherwise, the bios/requests will be handled as same as
663 * others which should be contiguous.
665 static inline bool blk_discard_mergable(struct request
*req
)
667 if (req_op(req
) == REQ_OP_DISCARD
&&
668 queue_max_discard_segments(req
->q
) > 1)
673 enum elv_merge
blk_try_req_merge(struct request
*req
, struct request
*next
)
675 if (blk_discard_mergable(req
))
676 return ELEVATOR_DISCARD_MERGE
;
677 else if (blk_rq_pos(req
) + blk_rq_sectors(req
) == blk_rq_pos(next
))
678 return ELEVATOR_BACK_MERGE
;
680 return ELEVATOR_NO_MERGE
;
684 * For non-mq, this has to be called with the request spinlock acquired.
685 * For mq with scheduling, the appropriate queue wide lock should be held.
687 static struct request
*attempt_merge(struct request_queue
*q
,
688 struct request
*req
, struct request
*next
)
691 lockdep_assert_held(q
->queue_lock
);
693 if (!rq_mergeable(req
) || !rq_mergeable(next
))
696 if (req_op(req
) != req_op(next
))
699 if (rq_data_dir(req
) != rq_data_dir(next
)
700 || req
->rq_disk
!= next
->rq_disk
701 || req_no_special_merge(next
))
704 if (req_op(req
) == REQ_OP_WRITE_SAME
&&
705 !blk_write_same_mergeable(req
->bio
, next
->bio
))
709 * Don't allow merge of different write hints, or for a hint with
712 if (req
->write_hint
!= next
->write_hint
)
716 * If we are allowed to merge, then append bio list
717 * from next to rq and release next. merge_requests_fn
718 * will have updated segment counts, update sector
719 * counts here. Handle DISCARDs separately, as they
720 * have separate settings.
723 switch (blk_try_req_merge(req
, next
)) {
724 case ELEVATOR_DISCARD_MERGE
:
725 if (!req_attempt_discard_merge(q
, req
, next
))
728 case ELEVATOR_BACK_MERGE
:
729 if (!ll_merge_requests_fn(q
, req
, next
))
737 * If failfast settings disagree or any of the two is already
738 * a mixed merge, mark both as mixed before proceeding. This
739 * makes sure that all involved bios have mixable attributes
742 if (((req
->rq_flags
| next
->rq_flags
) & RQF_MIXED_MERGE
) ||
743 (req
->cmd_flags
& REQ_FAILFAST_MASK
) !=
744 (next
->cmd_flags
& REQ_FAILFAST_MASK
)) {
745 blk_rq_set_mixed_merge(req
);
746 blk_rq_set_mixed_merge(next
);
750 * At this point we have either done a back merge
751 * or front merge. We need the smaller start_time of
752 * the merged requests to be the current request
753 * for accounting purposes.
755 if (time_after(req
->start_time
, next
->start_time
))
756 req
->start_time
= next
->start_time
;
758 req
->biotail
->bi_next
= next
->bio
;
759 req
->biotail
= next
->biotail
;
761 req
->__data_len
+= blk_rq_bytes(next
);
763 if (!blk_discard_mergable(req
))
764 elv_merge_requests(q
, req
, next
);
767 * 'next' is going away, so update stats accordingly
769 blk_account_io_merge(next
);
771 req
->ioprio
= ioprio_best(req
->ioprio
, next
->ioprio
);
772 if (blk_rq_cpu_valid(next
))
773 req
->cpu
= next
->cpu
;
776 * ownership of bio passed from next to req, return 'next' for
783 struct request
*attempt_back_merge(struct request_queue
*q
, struct request
*rq
)
785 struct request
*next
= elv_latter_request(q
, rq
);
788 return attempt_merge(q
, rq
, next
);
793 struct request
*attempt_front_merge(struct request_queue
*q
, struct request
*rq
)
795 struct request
*prev
= elv_former_request(q
, rq
);
798 return attempt_merge(q
, prev
, rq
);
803 int blk_attempt_req_merge(struct request_queue
*q
, struct request
*rq
,
804 struct request
*next
)
806 struct elevator_queue
*e
= q
->elevator
;
807 struct request
*free
;
809 if (!e
->uses_mq
&& e
->type
->ops
.sq
.elevator_allow_rq_merge_fn
)
810 if (!e
->type
->ops
.sq
.elevator_allow_rq_merge_fn(q
, rq
, next
))
813 free
= attempt_merge(q
, rq
, next
);
815 __blk_put_request(q
, free
);
822 bool blk_rq_merge_ok(struct request
*rq
, struct bio
*bio
)
824 if (!rq_mergeable(rq
) || !bio_mergeable(bio
))
827 if (req_op(rq
) != bio_op(bio
))
830 /* different data direction or already started, don't merge */
831 if (bio_data_dir(bio
) != rq_data_dir(rq
))
834 /* must be same device and not a special request */
835 if (rq
->rq_disk
!= bio
->bi_disk
|| req_no_special_merge(rq
))
838 /* only merge integrity protected bio into ditto rq */
839 if (blk_integrity_merge_bio(rq
->q
, rq
, bio
) == false)
842 /* must be using the same buffer */
843 if (req_op(rq
) == REQ_OP_WRITE_SAME
&&
844 !blk_write_same_mergeable(rq
->bio
, bio
))
848 * Don't allow merge of different write hints, or for a hint with
851 if (rq
->write_hint
!= bio
->bi_write_hint
)
857 enum elv_merge
blk_try_merge(struct request
*rq
, struct bio
*bio
)
859 if (blk_discard_mergable(rq
))
860 return ELEVATOR_DISCARD_MERGE
;
861 else if (blk_rq_pos(rq
) + blk_rq_sectors(rq
) == bio
->bi_iter
.bi_sector
)
862 return ELEVATOR_BACK_MERGE
;
863 else if (blk_rq_pos(rq
) - bio_sectors(bio
) == bio
->bi_iter
.bi_sector
)
864 return ELEVATOR_FRONT_MERGE
;
865 return ELEVATOR_NO_MERGE
;