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 inline unsigned get_max_io_size(struct request_queue
*q
,
76 unsigned sectors
= blk_max_size_offset(q
, bio
->bi_iter
.bi_sector
);
77 unsigned mask
= queue_logical_block_size(q
) - 1;
79 /* aligned to logical block size */
80 sectors
&= ~(mask
>> 9);
85 static struct bio
*blk_bio_segment_split(struct request_queue
*q
,
90 struct bio_vec bv
, bvprv
, *bvprvp
= NULL
;
91 struct bvec_iter iter
;
92 unsigned seg_size
= 0, nsegs
= 0, sectors
= 0;
93 unsigned front_seg_size
= bio
->bi_seg_front_size
;
95 struct bio
*new = NULL
;
96 const unsigned max_sectors
= get_max_io_size(q
, bio
);
99 bio_for_each_segment(bv
, bio
, iter
) {
101 * With arbitrary bio size, the incoming bio may be very
102 * big. We have to split the bio into small bios so that
103 * each holds at most BIO_MAX_PAGES bvecs because
104 * bio_clone() can fail to allocate big bvecs.
106 * It should have been better to apply the limit per
107 * request queue in which bio_clone() is involved,
108 * instead of globally. The biggest blocker is the
109 * bio_clone() in bio bounce.
111 * If bio is splitted by this reason, we should have
112 * allowed to continue bios merging, but don't do
113 * that now for making the change simple.
115 * TODO: deal with bio bounce's bio_clone() gracefully
116 * and convert the global limit into per-queue limit.
118 if (bvecs
++ >= BIO_MAX_PAGES
)
122 * If the queue doesn't support SG gaps and adding this
123 * offset would create a gap, disallow it.
125 if (bvprvp
&& bvec_gap_to_prev(q
, bvprvp
, bv
.bv_offset
))
128 if (sectors
+ (bv
.bv_len
>> 9) > max_sectors
) {
130 * Consider this a new segment if we're splitting in
131 * the middle of this vector.
133 if (nsegs
< queue_max_segments(q
) &&
134 sectors
< max_sectors
) {
136 sectors
= max_sectors
;
140 /* Make this single bvec as the 1st segment */
143 if (bvprvp
&& blk_queue_cluster(q
)) {
144 if (seg_size
+ bv
.bv_len
> queue_max_segment_size(q
))
146 if (!BIOVEC_PHYS_MERGEABLE(bvprvp
, &bv
))
148 if (!BIOVEC_SEG_BOUNDARY(q
, bvprvp
, &bv
))
151 seg_size
+= bv
.bv_len
;
154 sectors
+= bv
.bv_len
>> 9;
156 if (nsegs
== 1 && seg_size
> front_seg_size
)
157 front_seg_size
= seg_size
;
161 if (nsegs
== queue_max_segments(q
))
167 seg_size
= bv
.bv_len
;
168 sectors
+= bv
.bv_len
>> 9;
170 if (nsegs
== 1 && seg_size
> front_seg_size
)
171 front_seg_size
= seg_size
;
179 new = bio_split(bio
, sectors
, GFP_NOIO
, bs
);
184 bio
->bi_seg_front_size
= front_seg_size
;
185 if (seg_size
> bio
->bi_seg_back_size
)
186 bio
->bi_seg_back_size
= seg_size
;
188 return do_split
? new : NULL
;
191 void blk_queue_split(struct request_queue
*q
, struct bio
**bio
,
194 struct bio
*split
, *res
;
197 switch (bio_op(*bio
)) {
199 case REQ_OP_SECURE_ERASE
:
200 split
= blk_bio_discard_split(q
, *bio
, bs
, &nsegs
);
202 case REQ_OP_WRITE_SAME
:
203 split
= blk_bio_write_same_split(q
, *bio
, bs
, &nsegs
);
206 split
= blk_bio_segment_split(q
, *bio
, q
->bio_split
, &nsegs
);
210 /* physical segments can be figured out during splitting */
211 res
= split
? split
: *bio
;
212 res
->bi_phys_segments
= nsegs
;
213 bio_set_flag(res
, BIO_SEG_VALID
);
216 /* there isn't chance to merge the splitted bio */
217 split
->bi_opf
|= REQ_NOMERGE
;
219 bio_chain(split
, *bio
);
220 trace_block_split(q
, split
, (*bio
)->bi_iter
.bi_sector
);
221 generic_make_request(*bio
);
225 EXPORT_SYMBOL(blk_queue_split
);
227 static unsigned int __blk_recalc_rq_segments(struct request_queue
*q
,
231 struct bio_vec bv
, bvprv
= { NULL
};
232 int cluster
, prev
= 0;
233 unsigned int seg_size
, nr_phys_segs
;
234 struct bio
*fbio
, *bbio
;
235 struct bvec_iter iter
;
241 * This should probably be returning 0, but blk_add_request_payload()
244 if (bio_op(bio
) == REQ_OP_DISCARD
|| bio_op(bio
) == REQ_OP_SECURE_ERASE
)
247 if (bio_op(bio
) == REQ_OP_WRITE_SAME
)
251 cluster
= blk_queue_cluster(q
);
255 bio_for_each_segment(bv
, bio
, iter
) {
257 * If SG merging is disabled, each bio vector is
263 if (prev
&& cluster
) {
264 if (seg_size
+ bv
.bv_len
265 > queue_max_segment_size(q
))
267 if (!BIOVEC_PHYS_MERGEABLE(&bvprv
, &bv
))
269 if (!BIOVEC_SEG_BOUNDARY(q
, &bvprv
, &bv
))
272 seg_size
+= bv
.bv_len
;
277 if (nr_phys_segs
== 1 && seg_size
>
278 fbio
->bi_seg_front_size
)
279 fbio
->bi_seg_front_size
= seg_size
;
284 seg_size
= bv
.bv_len
;
289 if (nr_phys_segs
== 1 && seg_size
> fbio
->bi_seg_front_size
)
290 fbio
->bi_seg_front_size
= seg_size
;
291 if (seg_size
> bbio
->bi_seg_back_size
)
292 bbio
->bi_seg_back_size
= seg_size
;
297 void blk_recalc_rq_segments(struct request
*rq
)
299 bool no_sg_merge
= !!test_bit(QUEUE_FLAG_NO_SG_MERGE
,
300 &rq
->q
->queue_flags
);
302 rq
->nr_phys_segments
= __blk_recalc_rq_segments(rq
->q
, rq
->bio
,
306 void blk_recount_segments(struct request_queue
*q
, struct bio
*bio
)
308 unsigned short seg_cnt
= bio_segments(bio
);
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 int __blk_bios_map_sg(struct request_queue
*q
, struct bio
*bio
,
400 struct scatterlist
*sglist
,
401 struct scatterlist
**sg
)
403 struct bio_vec bvec
, bvprv
= { NULL
};
404 struct bvec_iter iter
;
408 cluster
= blk_queue_cluster(q
);
410 switch (bio_op(bio
)) {
412 case REQ_OP_SECURE_ERASE
:
414 * This is a hack - drivers should be neither modifying the
415 * biovec, nor relying on bi_vcnt - but because of
416 * blk_add_request_payload(), a discard bio may or may not have
417 * a payload we need to set up here (thank you Christoph) and
418 * bi_vcnt is really the only way of telling if we need to.
423 case REQ_OP_WRITE_SAME
:
425 bvec
= bio_iovec(bio
);
426 sg_set_page(*sg
, bvec
.bv_page
, bvec
.bv_len
, bvec
.bv_offset
);
433 bio_for_each_segment(bvec
, bio
, iter
)
434 __blk_segment_map_sg(q
, &bvec
, sglist
, &bvprv
, sg
,
441 * map a request to scatterlist, return number of sg entries setup. Caller
442 * must make sure sg can hold rq->nr_phys_segments entries
444 int blk_rq_map_sg(struct request_queue
*q
, struct request
*rq
,
445 struct scatterlist
*sglist
)
447 struct scatterlist
*sg
= NULL
;
451 nsegs
= __blk_bios_map_sg(q
, rq
->bio
, sglist
, &sg
);
453 if (unlikely(rq
->cmd_flags
& REQ_COPY_USER
) &&
454 (blk_rq_bytes(rq
) & q
->dma_pad_mask
)) {
455 unsigned int pad_len
=
456 (q
->dma_pad_mask
& ~blk_rq_bytes(rq
)) + 1;
458 sg
->length
+= pad_len
;
459 rq
->extra_len
+= pad_len
;
462 if (q
->dma_drain_size
&& q
->dma_drain_needed(rq
)) {
463 if (op_is_write(req_op(rq
)))
464 memset(q
->dma_drain_buffer
, 0, q
->dma_drain_size
);
468 sg_set_page(sg
, virt_to_page(q
->dma_drain_buffer
),
470 ((unsigned long)q
->dma_drain_buffer
) &
473 rq
->extra_len
+= q
->dma_drain_size
;
480 * Something must have been wrong if the figured number of
481 * segment is bigger than number of req's physical segments
483 WARN_ON(nsegs
> rq
->nr_phys_segments
);
487 EXPORT_SYMBOL(blk_rq_map_sg
);
489 static inline int ll_new_hw_segment(struct request_queue
*q
,
493 int nr_phys_segs
= bio_phys_segments(q
, bio
);
495 if (req
->nr_phys_segments
+ nr_phys_segs
> queue_max_segments(q
))
498 if (blk_integrity_merge_bio(q
, req
, bio
) == false)
502 * This will form the start of a new hw segment. Bump both
505 req
->nr_phys_segments
+= nr_phys_segs
;
509 req
->cmd_flags
|= REQ_NOMERGE
;
510 if (req
== q
->last_merge
)
511 q
->last_merge
= NULL
;
515 int ll_back_merge_fn(struct request_queue
*q
, struct request
*req
,
518 if (req_gap_back_merge(req
, bio
))
520 if (blk_integrity_rq(req
) &&
521 integrity_req_gap_back_merge(req
, bio
))
523 if (blk_rq_sectors(req
) + bio_sectors(bio
) >
524 blk_rq_get_max_sectors(req
, blk_rq_pos(req
))) {
525 req
->cmd_flags
|= REQ_NOMERGE
;
526 if (req
== q
->last_merge
)
527 q
->last_merge
= NULL
;
530 if (!bio_flagged(req
->biotail
, BIO_SEG_VALID
))
531 blk_recount_segments(q
, req
->biotail
);
532 if (!bio_flagged(bio
, BIO_SEG_VALID
))
533 blk_recount_segments(q
, bio
);
535 return ll_new_hw_segment(q
, req
, bio
);
538 int ll_front_merge_fn(struct request_queue
*q
, struct request
*req
,
542 if (req_gap_front_merge(req
, bio
))
544 if (blk_integrity_rq(req
) &&
545 integrity_req_gap_front_merge(req
, bio
))
547 if (blk_rq_sectors(req
) + bio_sectors(bio
) >
548 blk_rq_get_max_sectors(req
, bio
->bi_iter
.bi_sector
)) {
549 req
->cmd_flags
|= REQ_NOMERGE
;
550 if (req
== q
->last_merge
)
551 q
->last_merge
= NULL
;
554 if (!bio_flagged(bio
, BIO_SEG_VALID
))
555 blk_recount_segments(q
, bio
);
556 if (!bio_flagged(req
->bio
, BIO_SEG_VALID
))
557 blk_recount_segments(q
, req
->bio
);
559 return ll_new_hw_segment(q
, req
, bio
);
563 * blk-mq uses req->special to carry normal driver per-request payload, it
564 * does not indicate a prepared command that we cannot merge with.
566 static bool req_no_special_merge(struct request
*req
)
568 struct request_queue
*q
= req
->q
;
570 return !q
->mq_ops
&& req
->special
;
573 static int ll_merge_requests_fn(struct request_queue
*q
, struct request
*req
,
574 struct request
*next
)
576 int total_phys_segments
;
577 unsigned int seg_size
=
578 req
->biotail
->bi_seg_back_size
+ next
->bio
->bi_seg_front_size
;
581 * First check if the either of the requests are re-queued
582 * requests. Can't merge them if they are.
584 if (req_no_special_merge(req
) || req_no_special_merge(next
))
587 if (req_gap_back_merge(req
, next
->bio
))
591 * Will it become too large?
593 if ((blk_rq_sectors(req
) + blk_rq_sectors(next
)) >
594 blk_rq_get_max_sectors(req
, blk_rq_pos(req
)))
597 total_phys_segments
= req
->nr_phys_segments
+ next
->nr_phys_segments
;
598 if (blk_phys_contig_segment(q
, req
->biotail
, next
->bio
)) {
599 if (req
->nr_phys_segments
== 1)
600 req
->bio
->bi_seg_front_size
= seg_size
;
601 if (next
->nr_phys_segments
== 1)
602 next
->biotail
->bi_seg_back_size
= seg_size
;
603 total_phys_segments
--;
606 if (total_phys_segments
> queue_max_segments(q
))
609 if (blk_integrity_merge_rq(q
, req
, next
) == false)
613 req
->nr_phys_segments
= total_phys_segments
;
618 * blk_rq_set_mixed_merge - mark a request as mixed merge
619 * @rq: request to mark as mixed merge
622 * @rq is about to be mixed merged. Make sure the attributes
623 * which can be mixed are set in each bio and mark @rq as mixed
626 void blk_rq_set_mixed_merge(struct request
*rq
)
628 unsigned int ff
= rq
->cmd_flags
& REQ_FAILFAST_MASK
;
631 if (rq
->cmd_flags
& REQ_MIXED_MERGE
)
635 * @rq will no longer represent mixable attributes for all the
636 * contained bios. It will just track those of the first one.
637 * Distributes the attributs to each bio.
639 for (bio
= rq
->bio
; bio
; bio
= bio
->bi_next
) {
640 WARN_ON_ONCE((bio
->bi_opf
& REQ_FAILFAST_MASK
) &&
641 (bio
->bi_opf
& REQ_FAILFAST_MASK
) != ff
);
644 rq
->cmd_flags
|= REQ_MIXED_MERGE
;
647 static void blk_account_io_merge(struct request
*req
)
649 if (blk_do_io_stat(req
)) {
650 struct hd_struct
*part
;
653 cpu
= part_stat_lock();
656 part_round_stats(cpu
, part
);
657 part_dec_in_flight(part
, rq_data_dir(req
));
665 * Has to be called with the request spinlock acquired
667 static int attempt_merge(struct request_queue
*q
, struct request
*req
,
668 struct request
*next
)
670 if (!rq_mergeable(req
) || !rq_mergeable(next
))
673 if (req_op(req
) != req_op(next
))
679 if (blk_rq_pos(req
) + blk_rq_sectors(req
) != blk_rq_pos(next
))
682 if (rq_data_dir(req
) != rq_data_dir(next
)
683 || req
->rq_disk
!= next
->rq_disk
684 || req_no_special_merge(next
))
687 if (req_op(req
) == REQ_OP_WRITE_SAME
&&
688 !blk_write_same_mergeable(req
->bio
, next
->bio
))
692 * If we are allowed to merge, then append bio list
693 * from next to rq and release next. merge_requests_fn
694 * will have updated segment counts, update sector
697 if (!ll_merge_requests_fn(q
, req
, next
))
701 * If failfast settings disagree or any of the two is already
702 * a mixed merge, mark both as mixed before proceeding. This
703 * makes sure that all involved bios have mixable attributes
706 if ((req
->cmd_flags
| next
->cmd_flags
) & REQ_MIXED_MERGE
||
707 (req
->cmd_flags
& REQ_FAILFAST_MASK
) !=
708 (next
->cmd_flags
& REQ_FAILFAST_MASK
)) {
709 blk_rq_set_mixed_merge(req
);
710 blk_rq_set_mixed_merge(next
);
714 * At this point we have either done a back merge
715 * or front merge. We need the smaller start_time of
716 * the merged requests to be the current request
717 * for accounting purposes.
719 if (time_after(req
->start_time
, next
->start_time
))
720 req
->start_time
= next
->start_time
;
722 req
->biotail
->bi_next
= next
->bio
;
723 req
->biotail
= next
->biotail
;
725 req
->__data_len
+= blk_rq_bytes(next
);
727 elv_merge_requests(q
, req
, next
);
730 * 'next' is going away, so update stats accordingly
732 blk_account_io_merge(next
);
734 req
->ioprio
= ioprio_best(req
->ioprio
, next
->ioprio
);
735 if (blk_rq_cpu_valid(next
))
736 req
->cpu
= next
->cpu
;
738 /* owner-ship of bio passed from next to req */
740 __blk_put_request(q
, next
);
744 int attempt_back_merge(struct request_queue
*q
, struct request
*rq
)
746 struct request
*next
= elv_latter_request(q
, rq
);
749 return attempt_merge(q
, rq
, next
);
754 int attempt_front_merge(struct request_queue
*q
, struct request
*rq
)
756 struct request
*prev
= elv_former_request(q
, rq
);
759 return attempt_merge(q
, prev
, rq
);
764 int blk_attempt_req_merge(struct request_queue
*q
, struct request
*rq
,
765 struct request
*next
)
767 struct elevator_queue
*e
= q
->elevator
;
769 if (e
->type
->ops
.elevator_allow_rq_merge_fn
)
770 if (!e
->type
->ops
.elevator_allow_rq_merge_fn(q
, rq
, next
))
773 return attempt_merge(q
, rq
, next
);
776 bool blk_rq_merge_ok(struct request
*rq
, struct bio
*bio
)
778 if (!rq_mergeable(rq
) || !bio_mergeable(bio
))
781 if (req_op(rq
) != bio_op(bio
))
784 /* different data direction or already started, don't merge */
785 if (bio_data_dir(bio
) != rq_data_dir(rq
))
788 /* must be same device and not a special request */
789 if (rq
->rq_disk
!= bio
->bi_bdev
->bd_disk
|| req_no_special_merge(rq
))
792 /* only merge integrity protected bio into ditto rq */
793 if (blk_integrity_merge_bio(rq
->q
, rq
, bio
) == false)
796 /* must be using the same buffer */
797 if (req_op(rq
) == REQ_OP_WRITE_SAME
&&
798 !blk_write_same_mergeable(rq
->bio
, bio
))
804 int blk_try_merge(struct request
*rq
, struct bio
*bio
)
806 if (blk_rq_pos(rq
) + blk_rq_sectors(rq
) == bio
->bi_iter
.bi_sector
)
807 return ELEVATOR_BACK_MERGE
;
808 else if (blk_rq_pos(rq
) - bio_sectors(bio
) == bio
->bi_iter
.bi_sector
)
809 return ELEVATOR_FRONT_MERGE
;
810 return ELEVATOR_NO_MERGE
;