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
, UINT_MAX
>> 9);
31 max_discard_sectors
-= max_discard_sectors
% granularity
;
33 if (unlikely(!max_discard_sectors
)) {
38 if (bio_sectors(bio
) <= max_discard_sectors
)
41 split_sectors
= max_discard_sectors
;
44 * If the next starting sector would be misaligned, stop the discard at
45 * the previous aligned sector.
47 alignment
= (q
->limits
.discard_alignment
>> 9) % granularity
;
49 tmp
= bio
->bi_iter
.bi_sector
+ split_sectors
- alignment
;
50 tmp
= sector_div(tmp
, granularity
);
52 if (split_sectors
> tmp
)
55 return bio_split(bio
, split_sectors
, GFP_NOIO
, bs
);
58 static struct bio
*blk_bio_write_zeroes_split(struct request_queue
*q
,
59 struct bio
*bio
, struct bio_set
*bs
, unsigned *nsegs
)
63 if (!q
->limits
.max_write_zeroes_sectors
)
66 if (bio_sectors(bio
) <= q
->limits
.max_write_zeroes_sectors
)
69 return bio_split(bio
, q
->limits
.max_write_zeroes_sectors
, GFP_NOIO
, bs
);
72 static struct bio
*blk_bio_write_same_split(struct request_queue
*q
,
79 if (!q
->limits
.max_write_same_sectors
)
82 if (bio_sectors(bio
) <= q
->limits
.max_write_same_sectors
)
85 return bio_split(bio
, q
->limits
.max_write_same_sectors
, GFP_NOIO
, bs
);
88 static inline unsigned get_max_io_size(struct request_queue
*q
,
91 unsigned sectors
= blk_max_size_offset(q
, bio
->bi_iter
.bi_sector
);
92 unsigned mask
= queue_logical_block_size(q
) - 1;
94 /* aligned to logical block size */
95 sectors
&= ~(mask
>> 9);
100 static struct bio
*blk_bio_segment_split(struct request_queue
*q
,
105 struct bio_vec bv
, bvprv
, *bvprvp
= NULL
;
106 struct bvec_iter iter
;
107 unsigned seg_size
= 0, nsegs
= 0, sectors
= 0;
108 unsigned front_seg_size
= bio
->bi_seg_front_size
;
109 bool do_split
= true;
110 struct bio
*new = NULL
;
111 const unsigned max_sectors
= get_max_io_size(q
, bio
);
113 bio_for_each_segment(bv
, bio
, iter
) {
115 * If the queue doesn't support SG gaps and adding this
116 * offset would create a gap, disallow it.
118 if (bvprvp
&& bvec_gap_to_prev(q
, bvprvp
, bv
.bv_offset
))
121 if (sectors
+ (bv
.bv_len
>> 9) > max_sectors
) {
123 * Consider this a new segment if we're splitting in
124 * the middle of this vector.
126 if (nsegs
< queue_max_segments(q
) &&
127 sectors
< max_sectors
) {
129 sectors
= max_sectors
;
133 /* Make this single bvec as the 1st segment */
136 if (bvprvp
&& blk_queue_cluster(q
)) {
137 if (seg_size
+ bv
.bv_len
> queue_max_segment_size(q
))
139 if (!BIOVEC_PHYS_MERGEABLE(bvprvp
, &bv
))
141 if (!BIOVEC_SEG_BOUNDARY(q
, bvprvp
, &bv
))
144 seg_size
+= bv
.bv_len
;
147 sectors
+= bv
.bv_len
>> 9;
149 if (nsegs
== 1 && seg_size
> front_seg_size
)
150 front_seg_size
= seg_size
;
154 if (nsegs
== queue_max_segments(q
))
160 seg_size
= bv
.bv_len
;
161 sectors
+= bv
.bv_len
>> 9;
163 if (nsegs
== 1 && seg_size
> front_seg_size
)
164 front_seg_size
= seg_size
;
172 new = bio_split(bio
, sectors
, GFP_NOIO
, bs
);
177 bio
->bi_seg_front_size
= front_seg_size
;
178 if (seg_size
> bio
->bi_seg_back_size
)
179 bio
->bi_seg_back_size
= seg_size
;
181 return do_split
? new : NULL
;
184 void blk_queue_split(struct request_queue
*q
, struct bio
**bio
)
186 struct bio
*split
, *res
;
189 switch (bio_op(*bio
)) {
191 case REQ_OP_SECURE_ERASE
:
192 split
= blk_bio_discard_split(q
, *bio
, q
->bio_split
, &nsegs
);
194 case REQ_OP_WRITE_ZEROES
:
195 split
= blk_bio_write_zeroes_split(q
, *bio
, q
->bio_split
, &nsegs
);
197 case REQ_OP_WRITE_SAME
:
198 split
= blk_bio_write_same_split(q
, *bio
, q
->bio_split
, &nsegs
);
201 split
= blk_bio_segment_split(q
, *bio
, q
->bio_split
, &nsegs
);
205 /* physical segments can be figured out during splitting */
206 res
= split
? split
: *bio
;
207 res
->bi_phys_segments
= nsegs
;
208 bio_set_flag(res
, BIO_SEG_VALID
);
211 /* there isn't chance to merge the splitted bio */
212 split
->bi_opf
|= REQ_NOMERGE
;
214 bio_chain(split
, *bio
);
215 trace_block_split(q
, split
, (*bio
)->bi_iter
.bi_sector
);
216 generic_make_request(*bio
);
220 EXPORT_SYMBOL(blk_queue_split
);
222 static unsigned int __blk_recalc_rq_segments(struct request_queue
*q
,
226 struct bio_vec bv
, bvprv
= { NULL
};
227 int cluster
, prev
= 0;
228 unsigned int seg_size
, nr_phys_segs
;
229 struct bio
*fbio
, *bbio
;
230 struct bvec_iter iter
;
235 switch (bio_op(bio
)) {
237 case REQ_OP_SECURE_ERASE
:
238 case REQ_OP_WRITE_ZEROES
:
240 case REQ_OP_WRITE_SAME
:
245 cluster
= blk_queue_cluster(q
);
249 bio_for_each_segment(bv
, bio
, iter
) {
251 * If SG merging is disabled, each bio vector is
257 if (prev
&& cluster
) {
258 if (seg_size
+ bv
.bv_len
259 > queue_max_segment_size(q
))
261 if (!BIOVEC_PHYS_MERGEABLE(&bvprv
, &bv
))
263 if (!BIOVEC_SEG_BOUNDARY(q
, &bvprv
, &bv
))
266 seg_size
+= bv
.bv_len
;
271 if (nr_phys_segs
== 1 && seg_size
>
272 fbio
->bi_seg_front_size
)
273 fbio
->bi_seg_front_size
= seg_size
;
278 seg_size
= bv
.bv_len
;
283 if (nr_phys_segs
== 1 && seg_size
> fbio
->bi_seg_front_size
)
284 fbio
->bi_seg_front_size
= seg_size
;
285 if (seg_size
> bbio
->bi_seg_back_size
)
286 bbio
->bi_seg_back_size
= seg_size
;
291 void blk_recalc_rq_segments(struct request
*rq
)
293 bool no_sg_merge
= !!test_bit(QUEUE_FLAG_NO_SG_MERGE
,
294 &rq
->q
->queue_flags
);
296 rq
->nr_phys_segments
= __blk_recalc_rq_segments(rq
->q
, rq
->bio
,
300 void blk_recount_segments(struct request_queue
*q
, struct bio
*bio
)
302 unsigned short seg_cnt
;
304 /* estimate segment number by bi_vcnt for non-cloned bio */
305 if (bio_flagged(bio
, BIO_CLONED
))
306 seg_cnt
= bio_segments(bio
);
308 seg_cnt
= bio
->bi_vcnt
;
310 if (test_bit(QUEUE_FLAG_NO_SG_MERGE
, &q
->queue_flags
) &&
311 (seg_cnt
< queue_max_segments(q
)))
312 bio
->bi_phys_segments
= seg_cnt
;
314 struct bio
*nxt
= bio
->bi_next
;
317 bio
->bi_phys_segments
= __blk_recalc_rq_segments(q
, bio
, false);
321 bio_set_flag(bio
, BIO_SEG_VALID
);
323 EXPORT_SYMBOL(blk_recount_segments
);
325 static int blk_phys_contig_segment(struct request_queue
*q
, struct bio
*bio
,
328 struct bio_vec end_bv
= { NULL
}, nxt_bv
;
330 if (!blk_queue_cluster(q
))
333 if (bio
->bi_seg_back_size
+ nxt
->bi_seg_front_size
>
334 queue_max_segment_size(q
))
337 if (!bio_has_data(bio
))
340 bio_get_last_bvec(bio
, &end_bv
);
341 bio_get_first_bvec(nxt
, &nxt_bv
);
343 if (!BIOVEC_PHYS_MERGEABLE(&end_bv
, &nxt_bv
))
347 * bio and nxt are contiguous in memory; check if the queue allows
348 * these two to be merged into one
350 if (BIOVEC_SEG_BOUNDARY(q
, &end_bv
, &nxt_bv
))
357 __blk_segment_map_sg(struct request_queue
*q
, struct bio_vec
*bvec
,
358 struct scatterlist
*sglist
, struct bio_vec
*bvprv
,
359 struct scatterlist
**sg
, int *nsegs
, int *cluster
)
362 int nbytes
= bvec
->bv_len
;
364 if (*sg
&& *cluster
) {
365 if ((*sg
)->length
+ nbytes
> queue_max_segment_size(q
))
368 if (!BIOVEC_PHYS_MERGEABLE(bvprv
, bvec
))
370 if (!BIOVEC_SEG_BOUNDARY(q
, bvprv
, bvec
))
373 (*sg
)->length
+= nbytes
;
380 * If the driver previously mapped a shorter
381 * list, we could see a termination bit
382 * prematurely unless it fully inits the sg
383 * table on each mapping. We KNOW that there
384 * must be more entries here or the driver
385 * would be buggy, so force clear the
386 * termination bit to avoid doing a full
387 * sg_init_table() in drivers for each command.
393 sg_set_page(*sg
, bvec
->bv_page
, nbytes
, bvec
->bv_offset
);
399 static inline int __blk_bvec_map_sg(struct request_queue
*q
, struct bio_vec bv
,
400 struct scatterlist
*sglist
, struct scatterlist
**sg
)
403 sg_set_page(*sg
, bv
.bv_page
, bv
.bv_len
, bv
.bv_offset
);
407 static int __blk_bios_map_sg(struct request_queue
*q
, struct bio
*bio
,
408 struct scatterlist
*sglist
,
409 struct scatterlist
**sg
)
411 struct bio_vec bvec
, bvprv
= { NULL
};
412 struct bvec_iter iter
;
413 int cluster
= blk_queue_cluster(q
), nsegs
= 0;
416 bio_for_each_segment(bvec
, bio
, iter
)
417 __blk_segment_map_sg(q
, &bvec
, sglist
, &bvprv
, sg
,
424 * map a request to scatterlist, return number of sg entries setup. Caller
425 * must make sure sg can hold rq->nr_phys_segments entries
427 int blk_rq_map_sg(struct request_queue
*q
, struct request
*rq
,
428 struct scatterlist
*sglist
)
430 struct scatterlist
*sg
= NULL
;
433 if (rq
->rq_flags
& RQF_SPECIAL_PAYLOAD
)
434 nsegs
= __blk_bvec_map_sg(q
, rq
->special_vec
, sglist
, &sg
);
435 else if (rq
->bio
&& bio_op(rq
->bio
) == REQ_OP_WRITE_SAME
)
436 nsegs
= __blk_bvec_map_sg(q
, bio_iovec(rq
->bio
), sglist
, &sg
);
438 nsegs
= __blk_bios_map_sg(q
, rq
->bio
, sglist
, &sg
);
440 if (unlikely(rq
->rq_flags
& RQF_COPY_USER
) &&
441 (blk_rq_bytes(rq
) & q
->dma_pad_mask
)) {
442 unsigned int pad_len
=
443 (q
->dma_pad_mask
& ~blk_rq_bytes(rq
)) + 1;
445 sg
->length
+= pad_len
;
446 rq
->extra_len
+= pad_len
;
449 if (q
->dma_drain_size
&& q
->dma_drain_needed(rq
)) {
450 if (op_is_write(req_op(rq
)))
451 memset(q
->dma_drain_buffer
, 0, q
->dma_drain_size
);
455 sg_set_page(sg
, virt_to_page(q
->dma_drain_buffer
),
457 ((unsigned long)q
->dma_drain_buffer
) &
460 rq
->extra_len
+= q
->dma_drain_size
;
467 * Something must have been wrong if the figured number of
468 * segment is bigger than number of req's physical segments
470 WARN_ON(nsegs
> blk_rq_nr_phys_segments(rq
));
474 EXPORT_SYMBOL(blk_rq_map_sg
);
476 static inline int ll_new_hw_segment(struct request_queue
*q
,
480 int nr_phys_segs
= bio_phys_segments(q
, bio
);
482 if (req
->nr_phys_segments
+ nr_phys_segs
> queue_max_segments(q
))
485 if (blk_integrity_merge_bio(q
, req
, bio
) == false)
489 * This will form the start of a new hw segment. Bump both
492 req
->nr_phys_segments
+= nr_phys_segs
;
496 req_set_nomerge(q
, req
);
500 int ll_back_merge_fn(struct request_queue
*q
, struct request
*req
,
503 if (req_gap_back_merge(req
, bio
))
505 if (blk_integrity_rq(req
) &&
506 integrity_req_gap_back_merge(req
, bio
))
508 if (blk_rq_sectors(req
) + bio_sectors(bio
) >
509 blk_rq_get_max_sectors(req
, blk_rq_pos(req
))) {
510 req_set_nomerge(q
, req
);
513 if (!bio_flagged(req
->biotail
, BIO_SEG_VALID
))
514 blk_recount_segments(q
, req
->biotail
);
515 if (!bio_flagged(bio
, BIO_SEG_VALID
))
516 blk_recount_segments(q
, bio
);
518 return ll_new_hw_segment(q
, req
, bio
);
521 int ll_front_merge_fn(struct request_queue
*q
, struct request
*req
,
525 if (req_gap_front_merge(req
, bio
))
527 if (blk_integrity_rq(req
) &&
528 integrity_req_gap_front_merge(req
, bio
))
530 if (blk_rq_sectors(req
) + bio_sectors(bio
) >
531 blk_rq_get_max_sectors(req
, bio
->bi_iter
.bi_sector
)) {
532 req_set_nomerge(q
, req
);
535 if (!bio_flagged(bio
, BIO_SEG_VALID
))
536 blk_recount_segments(q
, bio
);
537 if (!bio_flagged(req
->bio
, BIO_SEG_VALID
))
538 blk_recount_segments(q
, req
->bio
);
540 return ll_new_hw_segment(q
, req
, bio
);
544 * blk-mq uses req->special to carry normal driver per-request payload, it
545 * does not indicate a prepared command that we cannot merge with.
547 static bool req_no_special_merge(struct request
*req
)
549 struct request_queue
*q
= req
->q
;
551 return !q
->mq_ops
&& req
->special
;
554 static int ll_merge_requests_fn(struct request_queue
*q
, struct request
*req
,
555 struct request
*next
)
557 int total_phys_segments
;
558 unsigned int seg_size
=
559 req
->biotail
->bi_seg_back_size
+ next
->bio
->bi_seg_front_size
;
562 * First check if the either of the requests are re-queued
563 * requests. Can't merge them if they are.
565 if (req_no_special_merge(req
) || req_no_special_merge(next
))
568 if (req_gap_back_merge(req
, next
->bio
))
572 * Will it become too large?
574 if ((blk_rq_sectors(req
) + blk_rq_sectors(next
)) >
575 blk_rq_get_max_sectors(req
, blk_rq_pos(req
)))
578 total_phys_segments
= req
->nr_phys_segments
+ next
->nr_phys_segments
;
579 if (blk_phys_contig_segment(q
, req
->biotail
, next
->bio
)) {
580 if (req
->nr_phys_segments
== 1)
581 req
->bio
->bi_seg_front_size
= seg_size
;
582 if (next
->nr_phys_segments
== 1)
583 next
->biotail
->bi_seg_back_size
= seg_size
;
584 total_phys_segments
--;
587 if (total_phys_segments
> queue_max_segments(q
))
590 if (blk_integrity_merge_rq(q
, req
, next
) == false)
594 req
->nr_phys_segments
= total_phys_segments
;
599 * blk_rq_set_mixed_merge - mark a request as mixed merge
600 * @rq: request to mark as mixed merge
603 * @rq is about to be mixed merged. Make sure the attributes
604 * which can be mixed are set in each bio and mark @rq as mixed
607 void blk_rq_set_mixed_merge(struct request
*rq
)
609 unsigned int ff
= rq
->cmd_flags
& REQ_FAILFAST_MASK
;
612 if (rq
->rq_flags
& RQF_MIXED_MERGE
)
616 * @rq will no longer represent mixable attributes for all the
617 * contained bios. It will just track those of the first one.
618 * Distributes the attributs to each bio.
620 for (bio
= rq
->bio
; bio
; bio
= bio
->bi_next
) {
621 WARN_ON_ONCE((bio
->bi_opf
& REQ_FAILFAST_MASK
) &&
622 (bio
->bi_opf
& REQ_FAILFAST_MASK
) != ff
);
625 rq
->rq_flags
|= RQF_MIXED_MERGE
;
628 static void blk_account_io_merge(struct request
*req
)
630 if (blk_do_io_stat(req
)) {
631 struct hd_struct
*part
;
634 cpu
= part_stat_lock();
637 part_round_stats(req
->q
, cpu
, part
);
638 part_dec_in_flight(req
->q
, part
, rq_data_dir(req
));
646 * For non-mq, this has to be called with the request spinlock acquired.
647 * For mq with scheduling, the appropriate queue wide lock should be held.
649 static struct request
*attempt_merge(struct request_queue
*q
,
650 struct request
*req
, struct request
*next
)
653 lockdep_assert_held(q
->queue_lock
);
655 if (!rq_mergeable(req
) || !rq_mergeable(next
))
658 if (req_op(req
) != req_op(next
))
664 if (blk_rq_pos(req
) + blk_rq_sectors(req
) != blk_rq_pos(next
))
667 if (rq_data_dir(req
) != rq_data_dir(next
)
668 || req
->rq_disk
!= next
->rq_disk
669 || req_no_special_merge(next
))
672 if (req_op(req
) == REQ_OP_WRITE_SAME
&&
673 !blk_write_same_mergeable(req
->bio
, next
->bio
))
677 * Don't allow merge of different write hints, or for a hint with
680 if (req
->write_hint
!= next
->write_hint
)
684 * If we are allowed to merge, then append bio list
685 * from next to rq and release next. merge_requests_fn
686 * will have updated segment counts, update sector
689 if (!ll_merge_requests_fn(q
, req
, next
))
693 * If failfast settings disagree or any of the two is already
694 * a mixed merge, mark both as mixed before proceeding. This
695 * makes sure that all involved bios have mixable attributes
698 if (((req
->rq_flags
| next
->rq_flags
) & RQF_MIXED_MERGE
) ||
699 (req
->cmd_flags
& REQ_FAILFAST_MASK
) !=
700 (next
->cmd_flags
& REQ_FAILFAST_MASK
)) {
701 blk_rq_set_mixed_merge(req
);
702 blk_rq_set_mixed_merge(next
);
706 * At this point we have either done a back merge
707 * or front merge. We need the smaller start_time of
708 * the merged requests to be the current request
709 * for accounting purposes.
711 if (time_after(req
->start_time
, next
->start_time
))
712 req
->start_time
= next
->start_time
;
714 req
->biotail
->bi_next
= next
->bio
;
715 req
->biotail
= next
->biotail
;
717 req
->__data_len
+= blk_rq_bytes(next
);
719 elv_merge_requests(q
, req
, next
);
722 * 'next' is going away, so update stats accordingly
724 blk_account_io_merge(next
);
726 req
->ioprio
= ioprio_best(req
->ioprio
, next
->ioprio
);
727 if (blk_rq_cpu_valid(next
))
728 req
->cpu
= next
->cpu
;
731 * ownership of bio passed from next to req, return 'next' for
738 struct request
*attempt_back_merge(struct request_queue
*q
, struct request
*rq
)
740 struct request
*next
= elv_latter_request(q
, rq
);
743 return attempt_merge(q
, rq
, next
);
748 struct request
*attempt_front_merge(struct request_queue
*q
, struct request
*rq
)
750 struct request
*prev
= elv_former_request(q
, rq
);
753 return attempt_merge(q
, prev
, rq
);
758 int blk_attempt_req_merge(struct request_queue
*q
, struct request
*rq
,
759 struct request
*next
)
761 struct elevator_queue
*e
= q
->elevator
;
762 struct request
*free
;
764 if (!e
->uses_mq
&& e
->type
->ops
.sq
.elevator_allow_rq_merge_fn
)
765 if (!e
->type
->ops
.sq
.elevator_allow_rq_merge_fn(q
, rq
, next
))
768 free
= attempt_merge(q
, rq
, next
);
770 __blk_put_request(q
, free
);
777 bool blk_rq_merge_ok(struct request
*rq
, struct bio
*bio
)
779 if (!rq_mergeable(rq
) || !bio_mergeable(bio
))
782 if (req_op(rq
) != bio_op(bio
))
785 /* different data direction or already started, don't merge */
786 if (bio_data_dir(bio
) != rq_data_dir(rq
))
789 /* must be same device and not a special request */
790 if (rq
->rq_disk
!= bio
->bi_disk
|| req_no_special_merge(rq
))
793 /* only merge integrity protected bio into ditto rq */
794 if (blk_integrity_merge_bio(rq
->q
, rq
, bio
) == false)
797 /* must be using the same buffer */
798 if (req_op(rq
) == REQ_OP_WRITE_SAME
&&
799 !blk_write_same_mergeable(rq
->bio
, bio
))
803 * Don't allow merge of different write hints, or for a hint with
806 if (rq
->write_hint
!= bio
->bi_write_hint
)
812 enum elv_merge
blk_try_merge(struct request
*rq
, struct bio
*bio
)
814 if (req_op(rq
) == REQ_OP_DISCARD
&&
815 queue_max_discard_segments(rq
->q
) > 1)
816 return ELEVATOR_DISCARD_MERGE
;
817 else if (blk_rq_pos(rq
) + blk_rq_sectors(rq
) == bio
->bi_iter
.bi_sector
)
818 return ELEVATOR_BACK_MERGE
;
819 else if (blk_rq_pos(rq
) - bio_sectors(bio
) == bio
->bi_iter
.bi_sector
)
820 return ELEVATOR_FRONT_MERGE
;
821 return ELEVATOR_NO_MERGE
;