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_zeroes_split(struct request_queue
*q
,
58 struct bio
*bio
, struct bio_set
*bs
, unsigned *nsegs
)
62 if (!q
->limits
.max_write_zeroes_sectors
)
65 if (bio_sectors(bio
) <= q
->limits
.max_write_zeroes_sectors
)
68 return bio_split(bio
, q
->limits
.max_write_zeroes_sectors
, GFP_NOIO
, bs
);
71 static struct bio
*blk_bio_write_same_split(struct request_queue
*q
,
78 if (!q
->limits
.max_write_same_sectors
)
81 if (bio_sectors(bio
) <= q
->limits
.max_write_same_sectors
)
84 return bio_split(bio
, q
->limits
.max_write_same_sectors
, GFP_NOIO
, bs
);
87 static inline unsigned get_max_io_size(struct request_queue
*q
,
90 unsigned sectors
= blk_max_size_offset(q
, bio
->bi_iter
.bi_sector
);
91 unsigned mask
= queue_logical_block_size(q
) - 1;
93 /* aligned to logical block size */
94 sectors
&= ~(mask
>> 9);
99 static struct bio
*blk_bio_segment_split(struct request_queue
*q
,
104 struct bio_vec bv
, bvprv
, *bvprvp
= NULL
;
105 struct bvec_iter iter
;
106 unsigned seg_size
= 0, nsegs
= 0, sectors
= 0;
107 unsigned front_seg_size
= bio
->bi_seg_front_size
;
108 bool do_split
= true;
109 struct bio
*new = NULL
;
110 const unsigned max_sectors
= get_max_io_size(q
, bio
);
113 bio_for_each_segment(bv
, bio
, iter
) {
115 * With arbitrary bio size, the incoming bio may be very
116 * big. We have to split the bio into small bios so that
117 * each holds at most BIO_MAX_PAGES bvecs because
118 * bio_clone() can fail to allocate big bvecs.
120 * It should have been better to apply the limit per
121 * request queue in which bio_clone() is involved,
122 * instead of globally. The biggest blocker is the
123 * bio_clone() in bio bounce.
125 * If bio is splitted by this reason, we should have
126 * allowed to continue bios merging, but don't do
127 * that now for making the change simple.
129 * TODO: deal with bio bounce's bio_clone() gracefully
130 * and convert the global limit into per-queue limit.
132 if (bvecs
++ >= BIO_MAX_PAGES
)
136 * If the queue doesn't support SG gaps and adding this
137 * offset would create a gap, disallow it.
139 if (bvprvp
&& bvec_gap_to_prev(q
, bvprvp
, bv
.bv_offset
))
142 if (sectors
+ (bv
.bv_len
>> 9) > max_sectors
) {
144 * Consider this a new segment if we're splitting in
145 * the middle of this vector.
147 if (nsegs
< queue_max_segments(q
) &&
148 sectors
< max_sectors
) {
150 sectors
= max_sectors
;
154 /* Make this single bvec as the 1st segment */
157 if (bvprvp
&& blk_queue_cluster(q
)) {
158 if (seg_size
+ bv
.bv_len
> queue_max_segment_size(q
))
160 if (!BIOVEC_PHYS_MERGEABLE(bvprvp
, &bv
))
162 if (!BIOVEC_SEG_BOUNDARY(q
, bvprvp
, &bv
))
165 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
;
175 if (nsegs
== queue_max_segments(q
))
181 seg_size
= bv
.bv_len
;
182 sectors
+= bv
.bv_len
>> 9;
184 if (nsegs
== 1 && seg_size
> front_seg_size
)
185 front_seg_size
= seg_size
;
193 new = bio_split(bio
, sectors
, GFP_NOIO
, bs
);
198 bio
->bi_seg_front_size
= front_seg_size
;
199 if (seg_size
> bio
->bi_seg_back_size
)
200 bio
->bi_seg_back_size
= seg_size
;
202 return do_split
? new : NULL
;
205 void blk_queue_split(struct request_queue
*q
, struct bio
**bio
,
208 struct bio
*split
, *res
;
211 switch (bio_op(*bio
)) {
213 case REQ_OP_SECURE_ERASE
:
214 split
= blk_bio_discard_split(q
, *bio
, bs
, &nsegs
);
216 case REQ_OP_WRITE_ZEROES
:
217 split
= blk_bio_write_zeroes_split(q
, *bio
, bs
, &nsegs
);
219 case REQ_OP_WRITE_SAME
:
220 split
= blk_bio_write_same_split(q
, *bio
, bs
, &nsegs
);
223 split
= blk_bio_segment_split(q
, *bio
, q
->bio_split
, &nsegs
);
227 /* physical segments can be figured out during splitting */
228 res
= split
? split
: *bio
;
229 res
->bi_phys_segments
= nsegs
;
230 bio_set_flag(res
, BIO_SEG_VALID
);
233 /* there isn't chance to merge the splitted bio */
234 split
->bi_opf
|= REQ_NOMERGE
;
236 bio_chain(split
, *bio
);
237 trace_block_split(q
, split
, (*bio
)->bi_iter
.bi_sector
);
238 generic_make_request(*bio
);
242 EXPORT_SYMBOL(blk_queue_split
);
244 static unsigned int __blk_recalc_rq_segments(struct request_queue
*q
,
248 struct bio_vec bv
, bvprv
= { NULL
};
249 int cluster
, prev
= 0;
250 unsigned int seg_size
, nr_phys_segs
;
251 struct bio
*fbio
, *bbio
;
252 struct bvec_iter iter
;
257 switch (bio_op(bio
)) {
259 case REQ_OP_SECURE_ERASE
:
260 case REQ_OP_WRITE_ZEROES
:
262 case REQ_OP_WRITE_SAME
:
267 cluster
= blk_queue_cluster(q
);
271 bio_for_each_segment(bv
, bio
, iter
) {
273 * If SG merging is disabled, each bio vector is
279 if (prev
&& cluster
) {
280 if (seg_size
+ bv
.bv_len
281 > queue_max_segment_size(q
))
283 if (!BIOVEC_PHYS_MERGEABLE(&bvprv
, &bv
))
285 if (!BIOVEC_SEG_BOUNDARY(q
, &bvprv
, &bv
))
288 seg_size
+= bv
.bv_len
;
293 if (nr_phys_segs
== 1 && seg_size
>
294 fbio
->bi_seg_front_size
)
295 fbio
->bi_seg_front_size
= seg_size
;
300 seg_size
= bv
.bv_len
;
305 if (nr_phys_segs
== 1 && seg_size
> fbio
->bi_seg_front_size
)
306 fbio
->bi_seg_front_size
= seg_size
;
307 if (seg_size
> bbio
->bi_seg_back_size
)
308 bbio
->bi_seg_back_size
= seg_size
;
313 void blk_recalc_rq_segments(struct request
*rq
)
315 bool no_sg_merge
= !!test_bit(QUEUE_FLAG_NO_SG_MERGE
,
316 &rq
->q
->queue_flags
);
318 rq
->nr_phys_segments
= __blk_recalc_rq_segments(rq
->q
, rq
->bio
,
322 void blk_recount_segments(struct request_queue
*q
, struct bio
*bio
)
324 unsigned short seg_cnt
;
326 /* estimate segment number by bi_vcnt for non-cloned bio */
327 if (bio_flagged(bio
, BIO_CLONED
))
328 seg_cnt
= bio_segments(bio
);
330 seg_cnt
= bio
->bi_vcnt
;
332 if (test_bit(QUEUE_FLAG_NO_SG_MERGE
, &q
->queue_flags
) &&
333 (seg_cnt
< queue_max_segments(q
)))
334 bio
->bi_phys_segments
= seg_cnt
;
336 struct bio
*nxt
= bio
->bi_next
;
339 bio
->bi_phys_segments
= __blk_recalc_rq_segments(q
, bio
, false);
343 bio_set_flag(bio
, BIO_SEG_VALID
);
345 EXPORT_SYMBOL(blk_recount_segments
);
347 static int blk_phys_contig_segment(struct request_queue
*q
, struct bio
*bio
,
350 struct bio_vec end_bv
= { NULL
}, nxt_bv
;
352 if (!blk_queue_cluster(q
))
355 if (bio
->bi_seg_back_size
+ nxt
->bi_seg_front_size
>
356 queue_max_segment_size(q
))
359 if (!bio_has_data(bio
))
362 bio_get_last_bvec(bio
, &end_bv
);
363 bio_get_first_bvec(nxt
, &nxt_bv
);
365 if (!BIOVEC_PHYS_MERGEABLE(&end_bv
, &nxt_bv
))
369 * bio and nxt are contiguous in memory; check if the queue allows
370 * these two to be merged into one
372 if (BIOVEC_SEG_BOUNDARY(q
, &end_bv
, &nxt_bv
))
379 __blk_segment_map_sg(struct request_queue
*q
, struct bio_vec
*bvec
,
380 struct scatterlist
*sglist
, struct bio_vec
*bvprv
,
381 struct scatterlist
**sg
, int *nsegs
, int *cluster
)
384 int nbytes
= bvec
->bv_len
;
386 if (*sg
&& *cluster
) {
387 if ((*sg
)->length
+ nbytes
> queue_max_segment_size(q
))
390 if (!BIOVEC_PHYS_MERGEABLE(bvprv
, bvec
))
392 if (!BIOVEC_SEG_BOUNDARY(q
, bvprv
, bvec
))
395 (*sg
)->length
+= nbytes
;
402 * If the driver previously mapped a shorter
403 * list, we could see a termination bit
404 * prematurely unless it fully inits the sg
405 * table on each mapping. We KNOW that there
406 * must be more entries here or the driver
407 * would be buggy, so force clear the
408 * termination bit to avoid doing a full
409 * sg_init_table() in drivers for each command.
415 sg_set_page(*sg
, bvec
->bv_page
, nbytes
, bvec
->bv_offset
);
421 static inline int __blk_bvec_map_sg(struct request_queue
*q
, struct bio_vec bv
,
422 struct scatterlist
*sglist
, struct scatterlist
**sg
)
425 sg_set_page(*sg
, bv
.bv_page
, bv
.bv_len
, bv
.bv_offset
);
429 static int __blk_bios_map_sg(struct request_queue
*q
, struct bio
*bio
,
430 struct scatterlist
*sglist
,
431 struct scatterlist
**sg
)
433 struct bio_vec bvec
, bvprv
= { NULL
};
434 struct bvec_iter iter
;
435 int cluster
= blk_queue_cluster(q
), nsegs
= 0;
438 bio_for_each_segment(bvec
, bio
, iter
)
439 __blk_segment_map_sg(q
, &bvec
, sglist
, &bvprv
, sg
,
446 * map a request to scatterlist, return number of sg entries setup. Caller
447 * must make sure sg can hold rq->nr_phys_segments entries
449 int blk_rq_map_sg(struct request_queue
*q
, struct request
*rq
,
450 struct scatterlist
*sglist
)
452 struct scatterlist
*sg
= NULL
;
455 if (rq
->rq_flags
& RQF_SPECIAL_PAYLOAD
)
456 nsegs
= __blk_bvec_map_sg(q
, rq
->special_vec
, sglist
, &sg
);
457 else if (rq
->bio
&& bio_op(rq
->bio
) == REQ_OP_WRITE_SAME
)
458 nsegs
= __blk_bvec_map_sg(q
, bio_iovec(rq
->bio
), sglist
, &sg
);
460 nsegs
= __blk_bios_map_sg(q
, rq
->bio
, sglist
, &sg
);
462 if (unlikely(rq
->rq_flags
& RQF_COPY_USER
) &&
463 (blk_rq_bytes(rq
) & q
->dma_pad_mask
)) {
464 unsigned int pad_len
=
465 (q
->dma_pad_mask
& ~blk_rq_bytes(rq
)) + 1;
467 sg
->length
+= pad_len
;
468 rq
->extra_len
+= pad_len
;
471 if (q
->dma_drain_size
&& q
->dma_drain_needed(rq
)) {
472 if (op_is_write(req_op(rq
)))
473 memset(q
->dma_drain_buffer
, 0, q
->dma_drain_size
);
477 sg_set_page(sg
, virt_to_page(q
->dma_drain_buffer
),
479 ((unsigned long)q
->dma_drain_buffer
) &
482 rq
->extra_len
+= q
->dma_drain_size
;
489 * Something must have been wrong if the figured number of
490 * segment is bigger than number of req's physical segments
492 WARN_ON(nsegs
> blk_rq_nr_phys_segments(rq
));
496 EXPORT_SYMBOL(blk_rq_map_sg
);
498 static inline int ll_new_hw_segment(struct request_queue
*q
,
502 int nr_phys_segs
= bio_phys_segments(q
, bio
);
504 if (req
->nr_phys_segments
+ nr_phys_segs
> queue_max_segments(q
))
507 if (blk_integrity_merge_bio(q
, req
, bio
) == false)
511 * This will form the start of a new hw segment. Bump both
514 req
->nr_phys_segments
+= nr_phys_segs
;
518 req_set_nomerge(q
, req
);
522 int ll_back_merge_fn(struct request_queue
*q
, struct request
*req
,
525 if (req_gap_back_merge(req
, bio
))
527 if (blk_integrity_rq(req
) &&
528 integrity_req_gap_back_merge(req
, bio
))
530 if (blk_rq_sectors(req
) + bio_sectors(bio
) >
531 blk_rq_get_max_sectors(req
, blk_rq_pos(req
))) {
532 req_set_nomerge(q
, req
);
535 if (!bio_flagged(req
->biotail
, BIO_SEG_VALID
))
536 blk_recount_segments(q
, req
->biotail
);
537 if (!bio_flagged(bio
, BIO_SEG_VALID
))
538 blk_recount_segments(q
, bio
);
540 return ll_new_hw_segment(q
, req
, bio
);
543 int ll_front_merge_fn(struct request_queue
*q
, struct request
*req
,
547 if (req_gap_front_merge(req
, bio
))
549 if (blk_integrity_rq(req
) &&
550 integrity_req_gap_front_merge(req
, bio
))
552 if (blk_rq_sectors(req
) + bio_sectors(bio
) >
553 blk_rq_get_max_sectors(req
, bio
->bi_iter
.bi_sector
)) {
554 req_set_nomerge(q
, req
);
557 if (!bio_flagged(bio
, BIO_SEG_VALID
))
558 blk_recount_segments(q
, bio
);
559 if (!bio_flagged(req
->bio
, BIO_SEG_VALID
))
560 blk_recount_segments(q
, req
->bio
);
562 return ll_new_hw_segment(q
, req
, bio
);
566 * blk-mq uses req->special to carry normal driver per-request payload, it
567 * does not indicate a prepared command that we cannot merge with.
569 static bool req_no_special_merge(struct request
*req
)
571 struct request_queue
*q
= req
->q
;
573 return !q
->mq_ops
&& req
->special
;
576 static int ll_merge_requests_fn(struct request_queue
*q
, struct request
*req
,
577 struct request
*next
)
579 int total_phys_segments
;
580 unsigned int seg_size
=
581 req
->biotail
->bi_seg_back_size
+ next
->bio
->bi_seg_front_size
;
584 * First check if the either of the requests are re-queued
585 * requests. Can't merge them if they are.
587 if (req_no_special_merge(req
) || req_no_special_merge(next
))
590 if (req_gap_back_merge(req
, next
->bio
))
594 * Will it become too large?
596 if ((blk_rq_sectors(req
) + blk_rq_sectors(next
)) >
597 blk_rq_get_max_sectors(req
, blk_rq_pos(req
)))
600 total_phys_segments
= req
->nr_phys_segments
+ next
->nr_phys_segments
;
601 if (blk_phys_contig_segment(q
, req
->biotail
, next
->bio
)) {
602 if (req
->nr_phys_segments
== 1)
603 req
->bio
->bi_seg_front_size
= seg_size
;
604 if (next
->nr_phys_segments
== 1)
605 next
->biotail
->bi_seg_back_size
= seg_size
;
606 total_phys_segments
--;
609 if (total_phys_segments
> queue_max_segments(q
))
612 if (blk_integrity_merge_rq(q
, req
, next
) == false)
616 req
->nr_phys_segments
= total_phys_segments
;
621 * blk_rq_set_mixed_merge - mark a request as mixed merge
622 * @rq: request to mark as mixed merge
625 * @rq is about to be mixed merged. Make sure the attributes
626 * which can be mixed are set in each bio and mark @rq as mixed
629 void blk_rq_set_mixed_merge(struct request
*rq
)
631 unsigned int ff
= rq
->cmd_flags
& REQ_FAILFAST_MASK
;
634 if (rq
->rq_flags
& RQF_MIXED_MERGE
)
638 * @rq will no longer represent mixable attributes for all the
639 * contained bios. It will just track those of the first one.
640 * Distributes the attributs to each bio.
642 for (bio
= rq
->bio
; bio
; bio
= bio
->bi_next
) {
643 WARN_ON_ONCE((bio
->bi_opf
& REQ_FAILFAST_MASK
) &&
644 (bio
->bi_opf
& REQ_FAILFAST_MASK
) != ff
);
647 rq
->rq_flags
|= RQF_MIXED_MERGE
;
650 static void blk_account_io_merge(struct request
*req
)
652 if (blk_do_io_stat(req
)) {
653 struct hd_struct
*part
;
656 cpu
= part_stat_lock();
659 part_round_stats(cpu
, part
);
660 part_dec_in_flight(part
, rq_data_dir(req
));
668 * For non-mq, this has to be called with the request spinlock acquired.
669 * For mq with scheduling, the appropriate queue wide lock should be held.
671 static struct request
*attempt_merge(struct request_queue
*q
,
672 struct request
*req
, struct request
*next
)
674 if (!rq_mergeable(req
) || !rq_mergeable(next
))
677 if (req_op(req
) != req_op(next
))
683 if (blk_rq_pos(req
) + blk_rq_sectors(req
) != blk_rq_pos(next
))
686 if (rq_data_dir(req
) != rq_data_dir(next
)
687 || req
->rq_disk
!= next
->rq_disk
688 || req_no_special_merge(next
))
691 if (req_op(req
) == REQ_OP_WRITE_SAME
&&
692 !blk_write_same_mergeable(req
->bio
, next
->bio
))
696 * If we are allowed to merge, then append bio list
697 * from next to rq and release next. merge_requests_fn
698 * will have updated segment counts, update sector
701 if (!ll_merge_requests_fn(q
, req
, next
))
705 * If failfast settings disagree or any of the two is already
706 * a mixed merge, mark both as mixed before proceeding. This
707 * makes sure that all involved bios have mixable attributes
710 if (((req
->rq_flags
| next
->rq_flags
) & RQF_MIXED_MERGE
) ||
711 (req
->cmd_flags
& REQ_FAILFAST_MASK
) !=
712 (next
->cmd_flags
& REQ_FAILFAST_MASK
)) {
713 blk_rq_set_mixed_merge(req
);
714 blk_rq_set_mixed_merge(next
);
718 * At this point we have either done a back merge
719 * or front merge. We need the smaller start_time of
720 * the merged requests to be the current request
721 * for accounting purposes.
723 if (time_after(req
->start_time
, next
->start_time
))
724 req
->start_time
= next
->start_time
;
726 req
->biotail
->bi_next
= next
->bio
;
727 req
->biotail
= next
->biotail
;
729 req
->__data_len
+= blk_rq_bytes(next
);
731 elv_merge_requests(q
, req
, next
);
734 * 'next' is going away, so update stats accordingly
736 blk_account_io_merge(next
);
738 req
->ioprio
= ioprio_best(req
->ioprio
, next
->ioprio
);
739 if (blk_rq_cpu_valid(next
))
740 req
->cpu
= next
->cpu
;
743 * ownership of bio passed from next to req, return 'next' for
750 struct request
*attempt_back_merge(struct request_queue
*q
, struct request
*rq
)
752 struct request
*next
= elv_latter_request(q
, rq
);
755 return attempt_merge(q
, rq
, next
);
760 struct request
*attempt_front_merge(struct request_queue
*q
, struct request
*rq
)
762 struct request
*prev
= elv_former_request(q
, rq
);
765 return attempt_merge(q
, prev
, rq
);
770 int blk_attempt_req_merge(struct request_queue
*q
, struct request
*rq
,
771 struct request
*next
)
773 struct elevator_queue
*e
= q
->elevator
;
774 struct request
*free
;
776 if (!e
->uses_mq
&& e
->type
->ops
.sq
.elevator_allow_rq_merge_fn
)
777 if (!e
->type
->ops
.sq
.elevator_allow_rq_merge_fn(q
, rq
, next
))
780 free
= attempt_merge(q
, rq
, next
);
782 __blk_put_request(q
, free
);
789 bool blk_rq_merge_ok(struct request
*rq
, struct bio
*bio
)
791 if (!rq_mergeable(rq
) || !bio_mergeable(bio
))
794 if (req_op(rq
) != bio_op(bio
))
797 /* different data direction or already started, don't merge */
798 if (bio_data_dir(bio
) != rq_data_dir(rq
))
801 /* must be same device and not a special request */
802 if (rq
->rq_disk
!= bio
->bi_bdev
->bd_disk
|| req_no_special_merge(rq
))
805 /* only merge integrity protected bio into ditto rq */
806 if (blk_integrity_merge_bio(rq
->q
, rq
, bio
) == false)
809 /* must be using the same buffer */
810 if (req_op(rq
) == REQ_OP_WRITE_SAME
&&
811 !blk_write_same_mergeable(rq
->bio
, bio
))
817 enum elv_merge
blk_try_merge(struct request
*rq
, struct bio
*bio
)
819 if (req_op(rq
) == REQ_OP_DISCARD
&&
820 queue_max_discard_segments(rq
->q
) > 1)
821 return ELEVATOR_DISCARD_MERGE
;
822 else if (blk_rq_pos(rq
) + blk_rq_sectors(rq
) == bio
->bi_iter
.bi_sector
)
823 return ELEVATOR_BACK_MERGE
;
824 else if (blk_rq_pos(rq
) - bio_sectors(bio
) == bio
->bi_iter
.bi_sector
)
825 return ELEVATOR_FRONT_MERGE
;
826 return ELEVATOR_NO_MERGE
;