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
= bio_segments(bio
);
304 if (test_bit(QUEUE_FLAG_NO_SG_MERGE
, &q
->queue_flags
) &&
305 (seg_cnt
< queue_max_segments(q
)))
306 bio
->bi_phys_segments
= seg_cnt
;
308 struct bio
*nxt
= bio
->bi_next
;
311 bio
->bi_phys_segments
= __blk_recalc_rq_segments(q
, bio
, false);
315 bio_set_flag(bio
, BIO_SEG_VALID
);
317 EXPORT_SYMBOL(blk_recount_segments
);
319 static int blk_phys_contig_segment(struct request_queue
*q
, struct bio
*bio
,
322 struct bio_vec end_bv
= { NULL
}, nxt_bv
;
324 if (!blk_queue_cluster(q
))
327 if (bio
->bi_seg_back_size
+ nxt
->bi_seg_front_size
>
328 queue_max_segment_size(q
))
331 if (!bio_has_data(bio
))
334 bio_get_last_bvec(bio
, &end_bv
);
335 bio_get_first_bvec(nxt
, &nxt_bv
);
337 if (!BIOVEC_PHYS_MERGEABLE(&end_bv
, &nxt_bv
))
341 * bio and nxt are contiguous in memory; check if the queue allows
342 * these two to be merged into one
344 if (BIOVEC_SEG_BOUNDARY(q
, &end_bv
, &nxt_bv
))
351 __blk_segment_map_sg(struct request_queue
*q
, struct bio_vec
*bvec
,
352 struct scatterlist
*sglist
, struct bio_vec
*bvprv
,
353 struct scatterlist
**sg
, int *nsegs
, int *cluster
)
356 int nbytes
= bvec
->bv_len
;
358 if (*sg
&& *cluster
) {
359 if ((*sg
)->length
+ nbytes
> queue_max_segment_size(q
))
362 if (!BIOVEC_PHYS_MERGEABLE(bvprv
, bvec
))
364 if (!BIOVEC_SEG_BOUNDARY(q
, bvprv
, bvec
))
367 (*sg
)->length
+= nbytes
;
374 * If the driver previously mapped a shorter
375 * list, we could see a termination bit
376 * prematurely unless it fully inits the sg
377 * table on each mapping. We KNOW that there
378 * must be more entries here or the driver
379 * would be buggy, so force clear the
380 * termination bit to avoid doing a full
381 * sg_init_table() in drivers for each command.
387 sg_set_page(*sg
, bvec
->bv_page
, nbytes
, bvec
->bv_offset
);
393 static inline int __blk_bvec_map_sg(struct request_queue
*q
, struct bio_vec bv
,
394 struct scatterlist
*sglist
, struct scatterlist
**sg
)
397 sg_set_page(*sg
, bv
.bv_page
, bv
.bv_len
, bv
.bv_offset
);
401 static int __blk_bios_map_sg(struct request_queue
*q
, struct bio
*bio
,
402 struct scatterlist
*sglist
,
403 struct scatterlist
**sg
)
405 struct bio_vec bvec
, bvprv
= { NULL
};
406 struct bvec_iter iter
;
407 int cluster
= blk_queue_cluster(q
), nsegs
= 0;
410 bio_for_each_segment(bvec
, bio
, iter
)
411 __blk_segment_map_sg(q
, &bvec
, sglist
, &bvprv
, sg
,
418 * map a request to scatterlist, return number of sg entries setup. Caller
419 * must make sure sg can hold rq->nr_phys_segments entries
421 int blk_rq_map_sg(struct request_queue
*q
, struct request
*rq
,
422 struct scatterlist
*sglist
)
424 struct scatterlist
*sg
= NULL
;
427 if (rq
->rq_flags
& RQF_SPECIAL_PAYLOAD
)
428 nsegs
= __blk_bvec_map_sg(q
, rq
->special_vec
, sglist
, &sg
);
429 else if (rq
->bio
&& bio_op(rq
->bio
) == REQ_OP_WRITE_SAME
)
430 nsegs
= __blk_bvec_map_sg(q
, bio_iovec(rq
->bio
), sglist
, &sg
);
432 nsegs
= __blk_bios_map_sg(q
, rq
->bio
, sglist
, &sg
);
434 if (unlikely(rq
->rq_flags
& RQF_COPY_USER
) &&
435 (blk_rq_bytes(rq
) & q
->dma_pad_mask
)) {
436 unsigned int pad_len
=
437 (q
->dma_pad_mask
& ~blk_rq_bytes(rq
)) + 1;
439 sg
->length
+= pad_len
;
440 rq
->extra_len
+= pad_len
;
443 if (q
->dma_drain_size
&& q
->dma_drain_needed(rq
)) {
444 if (op_is_write(req_op(rq
)))
445 memset(q
->dma_drain_buffer
, 0, q
->dma_drain_size
);
449 sg_set_page(sg
, virt_to_page(q
->dma_drain_buffer
),
451 ((unsigned long)q
->dma_drain_buffer
) &
454 rq
->extra_len
+= q
->dma_drain_size
;
461 * Something must have been wrong if the figured number of
462 * segment is bigger than number of req's physical segments
464 WARN_ON(nsegs
> blk_rq_nr_phys_segments(rq
));
468 EXPORT_SYMBOL(blk_rq_map_sg
);
470 static inline int ll_new_hw_segment(struct request_queue
*q
,
474 int nr_phys_segs
= bio_phys_segments(q
, bio
);
476 if (req
->nr_phys_segments
+ nr_phys_segs
> queue_max_segments(q
))
479 if (blk_integrity_merge_bio(q
, req
, bio
) == false)
483 * This will form the start of a new hw segment. Bump both
486 req
->nr_phys_segments
+= nr_phys_segs
;
490 req_set_nomerge(q
, req
);
494 int ll_back_merge_fn(struct request_queue
*q
, struct request
*req
,
497 if (req_gap_back_merge(req
, bio
))
499 if (blk_integrity_rq(req
) &&
500 integrity_req_gap_back_merge(req
, bio
))
502 if (blk_rq_sectors(req
) + bio_sectors(bio
) >
503 blk_rq_get_max_sectors(req
, blk_rq_pos(req
))) {
504 req_set_nomerge(q
, req
);
507 if (!bio_flagged(req
->biotail
, BIO_SEG_VALID
))
508 blk_recount_segments(q
, req
->biotail
);
509 if (!bio_flagged(bio
, BIO_SEG_VALID
))
510 blk_recount_segments(q
, bio
);
512 return ll_new_hw_segment(q
, req
, bio
);
515 int ll_front_merge_fn(struct request_queue
*q
, struct request
*req
,
519 if (req_gap_front_merge(req
, bio
))
521 if (blk_integrity_rq(req
) &&
522 integrity_req_gap_front_merge(req
, bio
))
524 if (blk_rq_sectors(req
) + bio_sectors(bio
) >
525 blk_rq_get_max_sectors(req
, bio
->bi_iter
.bi_sector
)) {
526 req_set_nomerge(q
, req
);
529 if (!bio_flagged(bio
, BIO_SEG_VALID
))
530 blk_recount_segments(q
, bio
);
531 if (!bio_flagged(req
->bio
, BIO_SEG_VALID
))
532 blk_recount_segments(q
, req
->bio
);
534 return ll_new_hw_segment(q
, req
, bio
);
538 * blk-mq uses req->special to carry normal driver per-request payload, it
539 * does not indicate a prepared command that we cannot merge with.
541 static bool req_no_special_merge(struct request
*req
)
543 struct request_queue
*q
= req
->q
;
545 return !q
->mq_ops
&& req
->special
;
548 static bool req_attempt_discard_merge(struct request_queue
*q
, struct request
*req
,
549 struct request
*next
)
551 unsigned short segments
= blk_rq_nr_discard_segments(req
);
553 if (segments
>= queue_max_discard_segments(q
))
555 if (blk_rq_sectors(req
) + bio_sectors(next
->bio
) >
556 blk_rq_get_max_sectors(req
, blk_rq_pos(req
)))
559 req
->nr_phys_segments
= segments
+ blk_rq_nr_discard_segments(next
);
562 req_set_nomerge(q
, req
);
566 static int ll_merge_requests_fn(struct request_queue
*q
, struct request
*req
,
567 struct request
*next
)
569 int total_phys_segments
;
570 unsigned int seg_size
=
571 req
->biotail
->bi_seg_back_size
+ next
->bio
->bi_seg_front_size
;
574 * First check if the either of the requests are re-queued
575 * requests. Can't merge them if they are.
577 if (req_no_special_merge(req
) || req_no_special_merge(next
))
580 if (req_gap_back_merge(req
, next
->bio
))
584 * Will it become too large?
586 if ((blk_rq_sectors(req
) + blk_rq_sectors(next
)) >
587 blk_rq_get_max_sectors(req
, blk_rq_pos(req
)))
590 total_phys_segments
= req
->nr_phys_segments
+ next
->nr_phys_segments
;
591 if (blk_phys_contig_segment(q
, req
->biotail
, next
->bio
)) {
592 if (req
->nr_phys_segments
== 1)
593 req
->bio
->bi_seg_front_size
= seg_size
;
594 if (next
->nr_phys_segments
== 1)
595 next
->biotail
->bi_seg_back_size
= seg_size
;
596 total_phys_segments
--;
599 if (total_phys_segments
> queue_max_segments(q
))
602 if (blk_integrity_merge_rq(q
, req
, next
) == false)
606 req
->nr_phys_segments
= total_phys_segments
;
611 * blk_rq_set_mixed_merge - mark a request as mixed merge
612 * @rq: request to mark as mixed merge
615 * @rq is about to be mixed merged. Make sure the attributes
616 * which can be mixed are set in each bio and mark @rq as mixed
619 void blk_rq_set_mixed_merge(struct request
*rq
)
621 unsigned int ff
= rq
->cmd_flags
& REQ_FAILFAST_MASK
;
624 if (rq
->rq_flags
& RQF_MIXED_MERGE
)
628 * @rq will no longer represent mixable attributes for all the
629 * contained bios. It will just track those of the first one.
630 * Distributes the attributs to each bio.
632 for (bio
= rq
->bio
; bio
; bio
= bio
->bi_next
) {
633 WARN_ON_ONCE((bio
->bi_opf
& REQ_FAILFAST_MASK
) &&
634 (bio
->bi_opf
& REQ_FAILFAST_MASK
) != ff
);
637 rq
->rq_flags
|= RQF_MIXED_MERGE
;
640 static void blk_account_io_merge(struct request
*req
)
642 if (blk_do_io_stat(req
)) {
643 struct hd_struct
*part
;
646 cpu
= part_stat_lock();
649 part_round_stats(req
->q
, cpu
, part
);
650 part_dec_in_flight(req
->q
, part
, rq_data_dir(req
));
657 * Two cases of handling DISCARD merge:
658 * If max_discard_segments > 1, the driver takes every bio
659 * as a range and send them to controller together. The ranges
660 * needn't to be contiguous.
661 * Otherwise, the bios/requests will be handled as same as
662 * others which should be contiguous.
664 static inline bool blk_discard_mergable(struct request
*req
)
666 if (req_op(req
) == REQ_OP_DISCARD
&&
667 queue_max_discard_segments(req
->q
) > 1)
672 enum elv_merge
blk_try_req_merge(struct request
*req
, struct request
*next
)
674 if (blk_discard_mergable(req
))
675 return ELEVATOR_DISCARD_MERGE
;
676 else if (blk_rq_pos(req
) + blk_rq_sectors(req
) == blk_rq_pos(next
))
677 return ELEVATOR_BACK_MERGE
;
679 return ELEVATOR_NO_MERGE
;
683 * For non-mq, this has to be called with the request spinlock acquired.
684 * For mq with scheduling, the appropriate queue wide lock should be held.
686 static struct request
*attempt_merge(struct request_queue
*q
,
687 struct request
*req
, struct request
*next
)
690 lockdep_assert_held(q
->queue_lock
);
692 if (!rq_mergeable(req
) || !rq_mergeable(next
))
695 if (req_op(req
) != req_op(next
))
698 if (rq_data_dir(req
) != rq_data_dir(next
)
699 || req
->rq_disk
!= next
->rq_disk
700 || req_no_special_merge(next
))
703 if (req_op(req
) == REQ_OP_WRITE_SAME
&&
704 !blk_write_same_mergeable(req
->bio
, next
->bio
))
708 * Don't allow merge of different write hints, or for a hint with
711 if (req
->write_hint
!= next
->write_hint
)
715 * If we are allowed to merge, then append bio list
716 * from next to rq and release next. merge_requests_fn
717 * will have updated segment counts, update sector
718 * counts here. Handle DISCARDs separately, as they
719 * have separate settings.
722 switch (blk_try_req_merge(req
, next
)) {
723 case ELEVATOR_DISCARD_MERGE
:
724 if (!req_attempt_discard_merge(q
, req
, next
))
727 case ELEVATOR_BACK_MERGE
:
728 if (!ll_merge_requests_fn(q
, req
, next
))
736 * If failfast settings disagree or any of the two is already
737 * a mixed merge, mark both as mixed before proceeding. This
738 * makes sure that all involved bios have mixable attributes
741 if (((req
->rq_flags
| next
->rq_flags
) & RQF_MIXED_MERGE
) ||
742 (req
->cmd_flags
& REQ_FAILFAST_MASK
) !=
743 (next
->cmd_flags
& REQ_FAILFAST_MASK
)) {
744 blk_rq_set_mixed_merge(req
);
745 blk_rq_set_mixed_merge(next
);
749 * At this point we have either done a back merge
750 * or front merge. We need the smaller start_time of
751 * the merged requests to be the current request
752 * for accounting purposes.
754 if (time_after(req
->start_time
, next
->start_time
))
755 req
->start_time
= next
->start_time
;
757 req
->biotail
->bi_next
= next
->bio
;
758 req
->biotail
= next
->biotail
;
760 req
->__data_len
+= blk_rq_bytes(next
);
762 if (!blk_discard_mergable(req
))
763 elv_merge_requests(q
, req
, next
);
766 * 'next' is going away, so update stats accordingly
768 blk_account_io_merge(next
);
770 req
->ioprio
= ioprio_best(req
->ioprio
, next
->ioprio
);
771 if (blk_rq_cpu_valid(next
))
772 req
->cpu
= next
->cpu
;
775 * ownership of bio passed from next to req, return 'next' for
782 struct request
*attempt_back_merge(struct request_queue
*q
, struct request
*rq
)
784 struct request
*next
= elv_latter_request(q
, rq
);
787 return attempt_merge(q
, rq
, next
);
792 struct request
*attempt_front_merge(struct request_queue
*q
, struct request
*rq
)
794 struct request
*prev
= elv_former_request(q
, rq
);
797 return attempt_merge(q
, prev
, rq
);
802 int blk_attempt_req_merge(struct request_queue
*q
, struct request
*rq
,
803 struct request
*next
)
805 struct elevator_queue
*e
= q
->elevator
;
806 struct request
*free
;
808 if (!e
->uses_mq
&& e
->type
->ops
.sq
.elevator_allow_rq_merge_fn
)
809 if (!e
->type
->ops
.sq
.elevator_allow_rq_merge_fn(q
, rq
, next
))
812 free
= attempt_merge(q
, rq
, next
);
814 __blk_put_request(q
, free
);
821 bool blk_rq_merge_ok(struct request
*rq
, struct bio
*bio
)
823 if (!rq_mergeable(rq
) || !bio_mergeable(bio
))
826 if (req_op(rq
) != bio_op(bio
))
829 /* different data direction or already started, don't merge */
830 if (bio_data_dir(bio
) != rq_data_dir(rq
))
833 /* must be same device and not a special request */
834 if (rq
->rq_disk
!= bio
->bi_disk
|| req_no_special_merge(rq
))
837 /* only merge integrity protected bio into ditto rq */
838 if (blk_integrity_merge_bio(rq
->q
, rq
, bio
) == false)
841 /* must be using the same buffer */
842 if (req_op(rq
) == REQ_OP_WRITE_SAME
&&
843 !blk_write_same_mergeable(rq
->bio
, bio
))
847 * Don't allow merge of different write hints, or for a hint with
850 if (rq
->write_hint
!= bio
->bi_write_hint
)
856 enum elv_merge
blk_try_merge(struct request
*rq
, struct bio
*bio
)
858 if (blk_discard_mergable(rq
))
859 return ELEVATOR_DISCARD_MERGE
;
860 else if (blk_rq_pos(rq
) + blk_rq_sectors(rq
) == bio
->bi_iter
.bi_sector
)
861 return ELEVATOR_BACK_MERGE
;
862 else if (blk_rq_pos(rq
) - bio_sectors(bio
) == bio
->bi_iter
.bi_sector
)
863 return ELEVATOR_FRONT_MERGE
;
864 return ELEVATOR_NO_MERGE
;