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>
12 static struct bio
*blk_bio_discard_split(struct request_queue
*q
,
17 unsigned int max_discard_sectors
, granularity
;
20 unsigned split_sectors
;
24 /* Zero-sector (unknown) and one-sector granularities are the same. */
25 granularity
= max(q
->limits
.discard_granularity
>> 9, 1U);
27 max_discard_sectors
= min(q
->limits
.max_discard_sectors
, UINT_MAX
>> 9);
28 max_discard_sectors
-= max_discard_sectors
% granularity
;
30 if (unlikely(!max_discard_sectors
)) {
35 if (bio_sectors(bio
) <= max_discard_sectors
)
38 split_sectors
= max_discard_sectors
;
41 * If the next starting sector would be misaligned, stop the discard at
42 * the previous aligned sector.
44 alignment
= (q
->limits
.discard_alignment
>> 9) % granularity
;
46 tmp
= bio
->bi_iter
.bi_sector
+ split_sectors
- alignment
;
47 tmp
= sector_div(tmp
, granularity
);
49 if (split_sectors
> tmp
)
52 return bio_split(bio
, split_sectors
, GFP_NOIO
, bs
);
55 static struct bio
*blk_bio_write_same_split(struct request_queue
*q
,
62 if (!q
->limits
.max_write_same_sectors
)
65 if (bio_sectors(bio
) <= q
->limits
.max_write_same_sectors
)
68 return bio_split(bio
, q
->limits
.max_write_same_sectors
, GFP_NOIO
, bs
);
71 static inline unsigned get_max_io_size(struct request_queue
*q
,
74 unsigned sectors
= blk_max_size_offset(q
, bio
->bi_iter
.bi_sector
);
75 unsigned mask
= queue_logical_block_size(q
) - 1;
77 /* aligned to logical block size */
78 sectors
&= ~(mask
>> 9);
83 static struct bio
*blk_bio_segment_split(struct request_queue
*q
,
88 struct bio_vec bv
, bvprv
, *bvprvp
= NULL
;
89 struct bvec_iter iter
;
90 unsigned seg_size
= 0, nsegs
= 0, sectors
= 0;
91 unsigned front_seg_size
= bio
->bi_seg_front_size
;
93 struct bio
*new = NULL
;
94 const unsigned max_sectors
= get_max_io_size(q
, bio
);
97 bio_for_each_segment(bv
, bio
, iter
) {
99 * With arbitrary bio size, the incoming bio may be very
100 * big. We have to split the bio into small bios so that
101 * each holds at most BIO_MAX_PAGES bvecs because
102 * bio_clone() can fail to allocate big bvecs.
104 * It should have been better to apply the limit per
105 * request queue in which bio_clone() is involved,
106 * instead of globally. The biggest blocker is the
107 * bio_clone() in bio bounce.
109 * If bio is splitted by this reason, we should have
110 * allowed to continue bios merging, but don't do
111 * that now for making the change simple.
113 * TODO: deal with bio bounce's bio_clone() gracefully
114 * and convert the global limit into per-queue limit.
116 if (bvecs
++ >= BIO_MAX_PAGES
)
120 * If the queue doesn't support SG gaps and adding this
121 * offset would create a gap, disallow it.
123 if (bvprvp
&& bvec_gap_to_prev(q
, bvprvp
, bv
.bv_offset
))
126 if (sectors
+ (bv
.bv_len
>> 9) > max_sectors
) {
128 * Consider this a new segment if we're splitting in
129 * the middle of this vector.
131 if (nsegs
< queue_max_segments(q
) &&
132 sectors
< max_sectors
) {
134 sectors
= max_sectors
;
138 /* Make this single bvec as the 1st segment */
141 if (bvprvp
&& blk_queue_cluster(q
)) {
142 if (seg_size
+ bv
.bv_len
> queue_max_segment_size(q
))
144 if (!BIOVEC_PHYS_MERGEABLE(bvprvp
, &bv
))
146 if (!BIOVEC_SEG_BOUNDARY(q
, bvprvp
, &bv
))
149 seg_size
+= bv
.bv_len
;
152 sectors
+= bv
.bv_len
>> 9;
154 if (nsegs
== 1 && seg_size
> front_seg_size
)
155 front_seg_size
= seg_size
;
159 if (nsegs
== queue_max_segments(q
))
165 seg_size
= bv
.bv_len
;
166 sectors
+= bv
.bv_len
>> 9;
168 if (nsegs
== 1 && seg_size
> front_seg_size
)
169 front_seg_size
= seg_size
;
177 new = bio_split(bio
, sectors
, GFP_NOIO
, bs
);
182 bio
->bi_seg_front_size
= front_seg_size
;
183 if (seg_size
> bio
->bi_seg_back_size
)
184 bio
->bi_seg_back_size
= seg_size
;
186 return do_split
? new : NULL
;
189 void blk_queue_split(struct request_queue
*q
, struct bio
**bio
,
192 struct bio
*split
, *res
;
195 if ((*bio
)->bi_rw
& REQ_DISCARD
)
196 split
= blk_bio_discard_split(q
, *bio
, bs
, &nsegs
);
197 else if ((*bio
)->bi_rw
& REQ_WRITE_SAME
)
198 split
= blk_bio_write_same_split(q
, *bio
, bs
, &nsegs
);
200 split
= blk_bio_segment_split(q
, *bio
, q
->bio_split
, &nsegs
);
202 /* physical segments can be figured out during splitting */
203 res
= split
? split
: *bio
;
204 res
->bi_phys_segments
= nsegs
;
205 bio_set_flag(res
, BIO_SEG_VALID
);
208 /* there isn't chance to merge the splitted bio */
209 split
->bi_rw
|= REQ_NOMERGE
;
211 bio_chain(split
, *bio
);
212 generic_make_request(*bio
);
216 EXPORT_SYMBOL(blk_queue_split
);
218 static unsigned int __blk_recalc_rq_segments(struct request_queue
*q
,
222 struct bio_vec bv
, bvprv
= { NULL
};
223 int cluster
, prev
= 0;
224 unsigned int seg_size
, nr_phys_segs
;
225 struct bio
*fbio
, *bbio
;
226 struct bvec_iter iter
;
232 * This should probably be returning 0, but blk_add_request_payload()
235 if (bio
->bi_rw
& REQ_DISCARD
)
238 if (bio
->bi_rw
& REQ_WRITE_SAME
)
242 cluster
= blk_queue_cluster(q
);
246 bio_for_each_segment(bv
, bio
, iter
) {
248 * If SG merging is disabled, each bio vector is
254 if (prev
&& cluster
) {
255 if (seg_size
+ bv
.bv_len
256 > queue_max_segment_size(q
))
258 if (!BIOVEC_PHYS_MERGEABLE(&bvprv
, &bv
))
260 if (!BIOVEC_SEG_BOUNDARY(q
, &bvprv
, &bv
))
263 seg_size
+= bv
.bv_len
;
268 if (nr_phys_segs
== 1 && seg_size
>
269 fbio
->bi_seg_front_size
)
270 fbio
->bi_seg_front_size
= seg_size
;
275 seg_size
= bv
.bv_len
;
280 if (nr_phys_segs
== 1 && seg_size
> fbio
->bi_seg_front_size
)
281 fbio
->bi_seg_front_size
= seg_size
;
282 if (seg_size
> bbio
->bi_seg_back_size
)
283 bbio
->bi_seg_back_size
= seg_size
;
288 void blk_recalc_rq_segments(struct request
*rq
)
290 bool no_sg_merge
= !!test_bit(QUEUE_FLAG_NO_SG_MERGE
,
291 &rq
->q
->queue_flags
);
293 rq
->nr_phys_segments
= __blk_recalc_rq_segments(rq
->q
, rq
->bio
,
297 void blk_recount_segments(struct request_queue
*q
, struct bio
*bio
)
299 unsigned short seg_cnt
;
301 /* estimate segment number by bi_vcnt for non-cloned bio */
302 if (bio_flagged(bio
, BIO_CLONED
))
303 seg_cnt
= bio_segments(bio
);
305 seg_cnt
= bio
->bi_vcnt
;
307 if (test_bit(QUEUE_FLAG_NO_SG_MERGE
, &q
->queue_flags
) &&
308 (seg_cnt
< queue_max_segments(q
)))
309 bio
->bi_phys_segments
= seg_cnt
;
311 struct bio
*nxt
= bio
->bi_next
;
314 bio
->bi_phys_segments
= __blk_recalc_rq_segments(q
, bio
, false);
318 bio_set_flag(bio
, BIO_SEG_VALID
);
320 EXPORT_SYMBOL(blk_recount_segments
);
322 static int blk_phys_contig_segment(struct request_queue
*q
, struct bio
*bio
,
325 struct bio_vec end_bv
= { NULL
}, nxt_bv
;
326 struct bvec_iter iter
;
328 if (!blk_queue_cluster(q
))
331 if (bio
->bi_seg_back_size
+ nxt
->bi_seg_front_size
>
332 queue_max_segment_size(q
))
335 if (!bio_has_data(bio
))
338 bio_for_each_segment(end_bv
, bio
, iter
)
339 if (end_bv
.bv_len
== iter
.bi_size
)
342 nxt_bv
= bio_iovec(nxt
);
344 if (!BIOVEC_PHYS_MERGEABLE(&end_bv
, &nxt_bv
))
348 * bio and nxt are contiguous in memory; check if the queue allows
349 * these two to be merged into one
351 if (BIOVEC_SEG_BOUNDARY(q
, &end_bv
, &nxt_bv
))
358 __blk_segment_map_sg(struct request_queue
*q
, struct bio_vec
*bvec
,
359 struct scatterlist
*sglist
, struct bio_vec
*bvprv
,
360 struct scatterlist
**sg
, int *nsegs
, int *cluster
)
363 int nbytes
= bvec
->bv_len
;
365 if (*sg
&& *cluster
) {
366 if ((*sg
)->length
+ nbytes
> queue_max_segment_size(q
))
369 if (!BIOVEC_PHYS_MERGEABLE(bvprv
, bvec
))
371 if (!BIOVEC_SEG_BOUNDARY(q
, bvprv
, bvec
))
374 (*sg
)->length
+= nbytes
;
381 * If the driver previously mapped a shorter
382 * list, we could see a termination bit
383 * prematurely unless it fully inits the sg
384 * table on each mapping. We KNOW that there
385 * must be more entries here or the driver
386 * would be buggy, so force clear the
387 * termination bit to avoid doing a full
388 * sg_init_table() in drivers for each command.
394 sg_set_page(*sg
, bvec
->bv_page
, nbytes
, bvec
->bv_offset
);
400 static int __blk_bios_map_sg(struct request_queue
*q
, struct bio
*bio
,
401 struct scatterlist
*sglist
,
402 struct scatterlist
**sg
)
404 struct bio_vec bvec
, bvprv
= { NULL
};
405 struct bvec_iter iter
;
409 cluster
= blk_queue_cluster(q
);
411 if (bio
->bi_rw
& REQ_DISCARD
) {
413 * This is a hack - drivers should be neither modifying the
414 * biovec, nor relying on bi_vcnt - but because of
415 * blk_add_request_payload(), a discard bio may or may not have
416 * a payload we need to set up here (thank you Christoph) and
417 * bi_vcnt is really the only way of telling if we need to.
426 if (bio
->bi_rw
& REQ_WRITE_SAME
) {
429 bvec
= bio_iovec(bio
);
430 sg_set_page(*sg
, bvec
.bv_page
, bvec
.bv_len
, bvec
.bv_offset
);
435 bio_for_each_segment(bvec
, bio
, iter
)
436 __blk_segment_map_sg(q
, &bvec
, sglist
, &bvprv
, sg
,
443 * map a request to scatterlist, return number of sg entries setup. Caller
444 * must make sure sg can hold rq->nr_phys_segments entries
446 int blk_rq_map_sg(struct request_queue
*q
, struct request
*rq
,
447 struct scatterlist
*sglist
)
449 struct scatterlist
*sg
= NULL
;
453 nsegs
= __blk_bios_map_sg(q
, rq
->bio
, sglist
, &sg
);
455 if (unlikely(rq
->cmd_flags
& REQ_COPY_USER
) &&
456 (blk_rq_bytes(rq
) & q
->dma_pad_mask
)) {
457 unsigned int pad_len
=
458 (q
->dma_pad_mask
& ~blk_rq_bytes(rq
)) + 1;
460 sg
->length
+= pad_len
;
461 rq
->extra_len
+= pad_len
;
464 if (q
->dma_drain_size
&& q
->dma_drain_needed(rq
)) {
465 if (rq
->cmd_flags
& REQ_WRITE
)
466 memset(q
->dma_drain_buffer
, 0, q
->dma_drain_size
);
470 sg_set_page(sg
, virt_to_page(q
->dma_drain_buffer
),
472 ((unsigned long)q
->dma_drain_buffer
) &
475 rq
->extra_len
+= q
->dma_drain_size
;
482 * Something must have been wrong if the figured number of
483 * segment is bigger than number of req's physical segments
485 WARN_ON(nsegs
> rq
->nr_phys_segments
);
489 EXPORT_SYMBOL(blk_rq_map_sg
);
491 static inline int ll_new_hw_segment(struct request_queue
*q
,
495 int nr_phys_segs
= bio_phys_segments(q
, bio
);
497 if (req
->nr_phys_segments
+ nr_phys_segs
> queue_max_segments(q
))
500 if (blk_integrity_merge_bio(q
, req
, bio
) == false)
504 * This will form the start of a new hw segment. Bump both
507 req
->nr_phys_segments
+= nr_phys_segs
;
511 req
->cmd_flags
|= REQ_NOMERGE
;
512 if (req
== q
->last_merge
)
513 q
->last_merge
= NULL
;
517 int ll_back_merge_fn(struct request_queue
*q
, struct request
*req
,
520 if (req_gap_back_merge(req
, bio
))
522 if (blk_integrity_rq(req
) &&
523 integrity_req_gap_back_merge(req
, bio
))
525 if (blk_rq_sectors(req
) + bio_sectors(bio
) >
526 blk_rq_get_max_sectors(req
)) {
527 req
->cmd_flags
|= REQ_NOMERGE
;
528 if (req
== q
->last_merge
)
529 q
->last_merge
= NULL
;
532 if (!bio_flagged(req
->biotail
, BIO_SEG_VALID
))
533 blk_recount_segments(q
, req
->biotail
);
534 if (!bio_flagged(bio
, BIO_SEG_VALID
))
535 blk_recount_segments(q
, bio
);
537 return ll_new_hw_segment(q
, req
, bio
);
540 int ll_front_merge_fn(struct request_queue
*q
, struct request
*req
,
544 if (req_gap_front_merge(req
, bio
))
546 if (blk_integrity_rq(req
) &&
547 integrity_req_gap_front_merge(req
, bio
))
549 if (blk_rq_sectors(req
) + bio_sectors(bio
) >
550 blk_rq_get_max_sectors(req
)) {
551 req
->cmd_flags
|= REQ_NOMERGE
;
552 if (req
== q
->last_merge
)
553 q
->last_merge
= NULL
;
556 if (!bio_flagged(bio
, BIO_SEG_VALID
))
557 blk_recount_segments(q
, bio
);
558 if (!bio_flagged(req
->bio
, BIO_SEG_VALID
))
559 blk_recount_segments(q
, req
->bio
);
561 return ll_new_hw_segment(q
, req
, bio
);
565 * blk-mq uses req->special to carry normal driver per-request payload, it
566 * does not indicate a prepared command that we cannot merge with.
568 static bool req_no_special_merge(struct request
*req
)
570 struct request_queue
*q
= req
->q
;
572 return !q
->mq_ops
&& req
->special
;
575 static int ll_merge_requests_fn(struct request_queue
*q
, struct request
*req
,
576 struct request
*next
)
578 int total_phys_segments
;
579 unsigned int seg_size
=
580 req
->biotail
->bi_seg_back_size
+ next
->bio
->bi_seg_front_size
;
583 * First check if the either of the requests are re-queued
584 * requests. Can't merge them if they are.
586 if (req_no_special_merge(req
) || req_no_special_merge(next
))
589 if (req_gap_back_merge(req
, next
->bio
))
593 * Will it become too large?
595 if ((blk_rq_sectors(req
) + blk_rq_sectors(next
)) >
596 blk_rq_get_max_sectors(req
))
599 total_phys_segments
= req
->nr_phys_segments
+ next
->nr_phys_segments
;
600 if (blk_phys_contig_segment(q
, req
->biotail
, next
->bio
)) {
601 if (req
->nr_phys_segments
== 1)
602 req
->bio
->bi_seg_front_size
= seg_size
;
603 if (next
->nr_phys_segments
== 1)
604 next
->biotail
->bi_seg_back_size
= seg_size
;
605 total_phys_segments
--;
608 if (total_phys_segments
> queue_max_segments(q
))
611 if (blk_integrity_merge_rq(q
, req
, next
) == false)
615 req
->nr_phys_segments
= total_phys_segments
;
620 * blk_rq_set_mixed_merge - mark a request as mixed merge
621 * @rq: request to mark as mixed merge
624 * @rq is about to be mixed merged. Make sure the attributes
625 * which can be mixed are set in each bio and mark @rq as mixed
628 void blk_rq_set_mixed_merge(struct request
*rq
)
630 unsigned int ff
= rq
->cmd_flags
& REQ_FAILFAST_MASK
;
633 if (rq
->cmd_flags
& REQ_MIXED_MERGE
)
637 * @rq will no longer represent mixable attributes for all the
638 * contained bios. It will just track those of the first one.
639 * Distributes the attributs to each bio.
641 for (bio
= rq
->bio
; bio
; bio
= bio
->bi_next
) {
642 WARN_ON_ONCE((bio
->bi_rw
& REQ_FAILFAST_MASK
) &&
643 (bio
->bi_rw
& REQ_FAILFAST_MASK
) != ff
);
646 rq
->cmd_flags
|= REQ_MIXED_MERGE
;
649 static void blk_account_io_merge(struct request
*req
)
651 if (blk_do_io_stat(req
)) {
652 struct hd_struct
*part
;
655 cpu
= part_stat_lock();
658 part_round_stats(cpu
, part
);
659 part_dec_in_flight(part
, rq_data_dir(req
));
667 * Has to be called with the request spinlock acquired
669 static int attempt_merge(struct request_queue
*q
, struct request
*req
,
670 struct request
*next
)
672 if (!rq_mergeable(req
) || !rq_mergeable(next
))
675 if (!blk_check_merge_flags(req
->cmd_flags
, next
->cmd_flags
))
681 if (blk_rq_pos(req
) + blk_rq_sectors(req
) != blk_rq_pos(next
))
684 if (rq_data_dir(req
) != rq_data_dir(next
)
685 || req
->rq_disk
!= next
->rq_disk
686 || req_no_special_merge(next
))
689 if (req
->cmd_flags
& REQ_WRITE_SAME
&&
690 !blk_write_same_mergeable(req
->bio
, next
->bio
))
694 * If we are allowed to merge, then append bio list
695 * from next to rq and release next. merge_requests_fn
696 * will have updated segment counts, update sector
699 if (!ll_merge_requests_fn(q
, req
, next
))
703 * If failfast settings disagree or any of the two is already
704 * a mixed merge, mark both as mixed before proceeding. This
705 * makes sure that all involved bios have mixable attributes
708 if ((req
->cmd_flags
| next
->cmd_flags
) & REQ_MIXED_MERGE
||
709 (req
->cmd_flags
& REQ_FAILFAST_MASK
) !=
710 (next
->cmd_flags
& REQ_FAILFAST_MASK
)) {
711 blk_rq_set_mixed_merge(req
);
712 blk_rq_set_mixed_merge(next
);
716 * At this point we have either done a back merge
717 * or front merge. We need the smaller start_time of
718 * the merged requests to be the current request
719 * for accounting purposes.
721 if (time_after(req
->start_time
, next
->start_time
))
722 req
->start_time
= next
->start_time
;
724 req
->biotail
->bi_next
= next
->bio
;
725 req
->biotail
= next
->biotail
;
727 req
->__data_len
+= blk_rq_bytes(next
);
729 elv_merge_requests(q
, req
, next
);
732 * 'next' is going away, so update stats accordingly
734 blk_account_io_merge(next
);
736 req
->ioprio
= ioprio_best(req
->ioprio
, next
->ioprio
);
737 if (blk_rq_cpu_valid(next
))
738 req
->cpu
= next
->cpu
;
740 /* owner-ship of bio passed from next to req */
742 __blk_put_request(q
, next
);
746 int attempt_back_merge(struct request_queue
*q
, struct request
*rq
)
748 struct request
*next
= elv_latter_request(q
, rq
);
751 return attempt_merge(q
, rq
, next
);
756 int attempt_front_merge(struct request_queue
*q
, struct request
*rq
)
758 struct request
*prev
= elv_former_request(q
, rq
);
761 return attempt_merge(q
, prev
, rq
);
766 int blk_attempt_req_merge(struct request_queue
*q
, struct request
*rq
,
767 struct request
*next
)
769 return attempt_merge(q
, rq
, next
);
772 bool blk_rq_merge_ok(struct request
*rq
, struct bio
*bio
)
774 if (!rq_mergeable(rq
) || !bio_mergeable(bio
))
777 if (!blk_check_merge_flags(rq
->cmd_flags
, bio
->bi_rw
))
780 /* different data direction or already started, don't merge */
781 if (bio_data_dir(bio
) != rq_data_dir(rq
))
784 /* must be same device and not a special request */
785 if (rq
->rq_disk
!= bio
->bi_bdev
->bd_disk
|| req_no_special_merge(rq
))
788 /* only merge integrity protected bio into ditto rq */
789 if (blk_integrity_merge_bio(rq
->q
, rq
, bio
) == false)
792 /* must be using the same buffer */
793 if (rq
->cmd_flags
& REQ_WRITE_SAME
&&
794 !blk_write_same_mergeable(rq
->bio
, bio
))
800 int blk_try_merge(struct request
*rq
, struct bio
*bio
)
802 if (blk_rq_pos(rq
) + blk_rq_sectors(rq
) == bio
->bi_iter
.bi_sector
)
803 return ELEVATOR_BACK_MERGE
;
804 else if (blk_rq_pos(rq
) - bio_sectors(bio
) == bio
->bi_iter
.bi_sector
)
805 return ELEVATOR_FRONT_MERGE
;
806 return ELEVATOR_NO_MERGE
;