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
);
98 bio_for_each_segment(bv
, bio
, iter
) {
100 * If the queue doesn't support SG gaps and adding this
101 * offset would create a gap, disallow it.
103 if (bvprvp
&& bvec_gap_to_prev(q
, bvprvp
, bv
.bv_offset
))
106 if (sectors
+ (bv
.bv_len
>> 9) > max_sectors
) {
108 * Consider this a new segment if we're splitting in
109 * the middle of this vector.
111 if (nsegs
< queue_max_segments(q
) &&
112 sectors
< max_sectors
) {
114 sectors
= max_sectors
;
118 /* Make this single bvec as the 1st segment */
121 if (bvprvp
&& blk_queue_cluster(q
)) {
122 if (seg_size
+ bv
.bv_len
> queue_max_segment_size(q
))
124 if (!BIOVEC_PHYS_MERGEABLE(bvprvp
, &bv
))
126 if (!BIOVEC_SEG_BOUNDARY(q
, bvprvp
, &bv
))
129 seg_size
+= bv
.bv_len
;
132 sectors
+= bv
.bv_len
>> 9;
134 if (nsegs
== 1 && seg_size
> front_seg_size
)
135 front_seg_size
= seg_size
;
139 if (nsegs
== queue_max_segments(q
))
145 seg_size
= bv
.bv_len
;
146 sectors
+= bv
.bv_len
>> 9;
148 if (nsegs
== 1 && seg_size
> front_seg_size
)
149 front_seg_size
= seg_size
;
157 new = bio_split(bio
, sectors
, GFP_NOIO
, bs
);
162 bio
->bi_seg_front_size
= front_seg_size
;
163 if (seg_size
> bio
->bi_seg_back_size
)
164 bio
->bi_seg_back_size
= seg_size
;
166 return do_split
? new : NULL
;
169 void blk_queue_split(struct request_queue
*q
, struct bio
**bio
,
172 struct bio
*split
, *res
;
175 if (bio_op(*bio
) == REQ_OP_DISCARD
)
176 split
= blk_bio_discard_split(q
, *bio
, bs
, &nsegs
);
177 else if (bio_op(*bio
) == REQ_OP_WRITE_SAME
)
178 split
= blk_bio_write_same_split(q
, *bio
, bs
, &nsegs
);
180 split
= blk_bio_segment_split(q
, *bio
, q
->bio_split
, &nsegs
);
182 /* physical segments can be figured out during splitting */
183 res
= split
? split
: *bio
;
184 res
->bi_phys_segments
= nsegs
;
185 bio_set_flag(res
, BIO_SEG_VALID
);
188 /* there isn't chance to merge the splitted bio */
189 split
->bi_opf
|= REQ_NOMERGE
;
191 bio_chain(split
, *bio
);
192 trace_block_split(q
, split
, (*bio
)->bi_iter
.bi_sector
);
193 generic_make_request(*bio
);
197 EXPORT_SYMBOL(blk_queue_split
);
199 static unsigned int __blk_recalc_rq_segments(struct request_queue
*q
,
203 struct bio_vec bv
, bvprv
= { NULL
};
204 int cluster
, prev
= 0;
205 unsigned int seg_size
, nr_phys_segs
;
206 struct bio
*fbio
, *bbio
;
207 struct bvec_iter iter
;
213 * This should probably be returning 0, but blk_add_request_payload()
216 if (bio_op(bio
) == REQ_OP_DISCARD
)
219 if (bio_op(bio
) == REQ_OP_WRITE_SAME
)
223 cluster
= blk_queue_cluster(q
);
227 bio_for_each_segment(bv
, bio
, iter
) {
229 * If SG merging is disabled, each bio vector is
235 if (prev
&& cluster
) {
236 if (seg_size
+ bv
.bv_len
237 > queue_max_segment_size(q
))
239 if (!BIOVEC_PHYS_MERGEABLE(&bvprv
, &bv
))
241 if (!BIOVEC_SEG_BOUNDARY(q
, &bvprv
, &bv
))
244 seg_size
+= bv
.bv_len
;
249 if (nr_phys_segs
== 1 && seg_size
>
250 fbio
->bi_seg_front_size
)
251 fbio
->bi_seg_front_size
= seg_size
;
256 seg_size
= bv
.bv_len
;
261 if (nr_phys_segs
== 1 && seg_size
> fbio
->bi_seg_front_size
)
262 fbio
->bi_seg_front_size
= seg_size
;
263 if (seg_size
> bbio
->bi_seg_back_size
)
264 bbio
->bi_seg_back_size
= seg_size
;
269 void blk_recalc_rq_segments(struct request
*rq
)
271 bool no_sg_merge
= !!test_bit(QUEUE_FLAG_NO_SG_MERGE
,
272 &rq
->q
->queue_flags
);
274 rq
->nr_phys_segments
= __blk_recalc_rq_segments(rq
->q
, rq
->bio
,
278 void blk_recount_segments(struct request_queue
*q
, struct bio
*bio
)
280 unsigned short seg_cnt
;
282 /* estimate segment number by bi_vcnt for non-cloned bio */
283 if (bio_flagged(bio
, BIO_CLONED
))
284 seg_cnt
= bio_segments(bio
);
286 seg_cnt
= bio
->bi_vcnt
;
288 if (test_bit(QUEUE_FLAG_NO_SG_MERGE
, &q
->queue_flags
) &&
289 (seg_cnt
< queue_max_segments(q
)))
290 bio
->bi_phys_segments
= seg_cnt
;
292 struct bio
*nxt
= bio
->bi_next
;
295 bio
->bi_phys_segments
= __blk_recalc_rq_segments(q
, bio
, false);
299 bio_set_flag(bio
, BIO_SEG_VALID
);
301 EXPORT_SYMBOL(blk_recount_segments
);
303 static int blk_phys_contig_segment(struct request_queue
*q
, struct bio
*bio
,
306 struct bio_vec end_bv
= { NULL
}, nxt_bv
;
308 if (!blk_queue_cluster(q
))
311 if (bio
->bi_seg_back_size
+ nxt
->bi_seg_front_size
>
312 queue_max_segment_size(q
))
315 if (!bio_has_data(bio
))
318 bio_get_last_bvec(bio
, &end_bv
);
319 bio_get_first_bvec(nxt
, &nxt_bv
);
321 if (!BIOVEC_PHYS_MERGEABLE(&end_bv
, &nxt_bv
))
325 * bio and nxt are contiguous in memory; check if the queue allows
326 * these two to be merged into one
328 if (BIOVEC_SEG_BOUNDARY(q
, &end_bv
, &nxt_bv
))
335 __blk_segment_map_sg(struct request_queue
*q
, struct bio_vec
*bvec
,
336 struct scatterlist
*sglist
, struct bio_vec
*bvprv
,
337 struct scatterlist
**sg
, int *nsegs
, int *cluster
)
340 int nbytes
= bvec
->bv_len
;
342 if (*sg
&& *cluster
) {
343 if ((*sg
)->length
+ nbytes
> queue_max_segment_size(q
))
346 if (!BIOVEC_PHYS_MERGEABLE(bvprv
, bvec
))
348 if (!BIOVEC_SEG_BOUNDARY(q
, bvprv
, bvec
))
351 (*sg
)->length
+= nbytes
;
358 * If the driver previously mapped a shorter
359 * list, we could see a termination bit
360 * prematurely unless it fully inits the sg
361 * table on each mapping. We KNOW that there
362 * must be more entries here or the driver
363 * would be buggy, so force clear the
364 * termination bit to avoid doing a full
365 * sg_init_table() in drivers for each command.
371 sg_set_page(*sg
, bvec
->bv_page
, nbytes
, bvec
->bv_offset
);
377 static int __blk_bios_map_sg(struct request_queue
*q
, struct bio
*bio
,
378 struct scatterlist
*sglist
,
379 struct scatterlist
**sg
)
381 struct bio_vec bvec
, bvprv
= { NULL
};
382 struct bvec_iter iter
;
386 cluster
= blk_queue_cluster(q
);
388 if (bio_op(bio
) == REQ_OP_DISCARD
) {
390 * This is a hack - drivers should be neither modifying the
391 * biovec, nor relying on bi_vcnt - but because of
392 * blk_add_request_payload(), a discard bio may or may not have
393 * a payload we need to set up here (thank you Christoph) and
394 * bi_vcnt is really the only way of telling if we need to.
403 if (bio_op(bio
) == REQ_OP_WRITE_SAME
) {
406 bvec
= bio_iovec(bio
);
407 sg_set_page(*sg
, bvec
.bv_page
, bvec
.bv_len
, bvec
.bv_offset
);
412 bio_for_each_segment(bvec
, bio
, iter
)
413 __blk_segment_map_sg(q
, &bvec
, sglist
, &bvprv
, sg
,
420 * map a request to scatterlist, return number of sg entries setup. Caller
421 * must make sure sg can hold rq->nr_phys_segments entries
423 int blk_rq_map_sg(struct request_queue
*q
, struct request
*rq
,
424 struct scatterlist
*sglist
)
426 struct scatterlist
*sg
= NULL
;
430 nsegs
= __blk_bios_map_sg(q
, rq
->bio
, sglist
, &sg
);
432 if (unlikely(rq
->cmd_flags
& REQ_COPY_USER
) &&
433 (blk_rq_bytes(rq
) & q
->dma_pad_mask
)) {
434 unsigned int pad_len
=
435 (q
->dma_pad_mask
& ~blk_rq_bytes(rq
)) + 1;
437 sg
->length
+= pad_len
;
438 rq
->extra_len
+= pad_len
;
441 if (q
->dma_drain_size
&& q
->dma_drain_needed(rq
)) {
442 if (op_is_write(req_op(rq
)))
443 memset(q
->dma_drain_buffer
, 0, q
->dma_drain_size
);
447 sg_set_page(sg
, virt_to_page(q
->dma_drain_buffer
),
449 ((unsigned long)q
->dma_drain_buffer
) &
452 rq
->extra_len
+= q
->dma_drain_size
;
459 * Something must have been wrong if the figured number of
460 * segment is bigger than number of req's physical segments
462 WARN_ON(nsegs
> rq
->nr_phys_segments
);
466 EXPORT_SYMBOL(blk_rq_map_sg
);
468 static inline int ll_new_hw_segment(struct request_queue
*q
,
472 int nr_phys_segs
= bio_phys_segments(q
, bio
);
474 if (req
->nr_phys_segments
+ nr_phys_segs
> queue_max_segments(q
))
477 if (blk_integrity_merge_bio(q
, req
, bio
) == false)
481 * This will form the start of a new hw segment. Bump both
484 req
->nr_phys_segments
+= nr_phys_segs
;
488 req
->cmd_flags
|= REQ_NOMERGE
;
489 if (req
== q
->last_merge
)
490 q
->last_merge
= NULL
;
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
->cmd_flags
|= REQ_NOMERGE
;
505 if (req
== q
->last_merge
)
506 q
->last_merge
= NULL
;
509 if (!bio_flagged(req
->biotail
, BIO_SEG_VALID
))
510 blk_recount_segments(q
, req
->biotail
);
511 if (!bio_flagged(bio
, BIO_SEG_VALID
))
512 blk_recount_segments(q
, bio
);
514 return ll_new_hw_segment(q
, req
, bio
);
517 int ll_front_merge_fn(struct request_queue
*q
, struct request
*req
,
521 if (req_gap_front_merge(req
, bio
))
523 if (blk_integrity_rq(req
) &&
524 integrity_req_gap_front_merge(req
, bio
))
526 if (blk_rq_sectors(req
) + bio_sectors(bio
) >
527 blk_rq_get_max_sectors(req
, bio
->bi_iter
.bi_sector
)) {
528 req
->cmd_flags
|= REQ_NOMERGE
;
529 if (req
== q
->last_merge
)
530 q
->last_merge
= NULL
;
533 if (!bio_flagged(bio
, BIO_SEG_VALID
))
534 blk_recount_segments(q
, bio
);
535 if (!bio_flagged(req
->bio
, BIO_SEG_VALID
))
536 blk_recount_segments(q
, req
->bio
);
538 return ll_new_hw_segment(q
, req
, bio
);
542 * blk-mq uses req->special to carry normal driver per-request payload, it
543 * does not indicate a prepared command that we cannot merge with.
545 static bool req_no_special_merge(struct request
*req
)
547 struct request_queue
*q
= req
->q
;
549 return !q
->mq_ops
&& req
->special
;
552 static int ll_merge_requests_fn(struct request_queue
*q
, struct request
*req
,
553 struct request
*next
)
555 int total_phys_segments
;
556 unsigned int seg_size
=
557 req
->biotail
->bi_seg_back_size
+ next
->bio
->bi_seg_front_size
;
560 * First check if the either of the requests are re-queued
561 * requests. Can't merge them if they are.
563 if (req_no_special_merge(req
) || req_no_special_merge(next
))
566 if (req_gap_back_merge(req
, next
->bio
))
570 * Will it become too large?
572 if ((blk_rq_sectors(req
) + blk_rq_sectors(next
)) >
573 blk_rq_get_max_sectors(req
, blk_rq_pos(req
)))
576 total_phys_segments
= req
->nr_phys_segments
+ next
->nr_phys_segments
;
577 if (blk_phys_contig_segment(q
, req
->biotail
, next
->bio
)) {
578 if (req
->nr_phys_segments
== 1)
579 req
->bio
->bi_seg_front_size
= seg_size
;
580 if (next
->nr_phys_segments
== 1)
581 next
->biotail
->bi_seg_back_size
= seg_size
;
582 total_phys_segments
--;
585 if (total_phys_segments
> queue_max_segments(q
))
588 if (blk_integrity_merge_rq(q
, req
, next
) == false)
592 req
->nr_phys_segments
= total_phys_segments
;
597 * blk_rq_set_mixed_merge - mark a request as mixed merge
598 * @rq: request to mark as mixed merge
601 * @rq is about to be mixed merged. Make sure the attributes
602 * which can be mixed are set in each bio and mark @rq as mixed
605 void blk_rq_set_mixed_merge(struct request
*rq
)
607 unsigned int ff
= rq
->cmd_flags
& REQ_FAILFAST_MASK
;
610 if (rq
->cmd_flags
& REQ_MIXED_MERGE
)
614 * @rq will no longer represent mixable attributes for all the
615 * contained bios. It will just track those of the first one.
616 * Distributes the attributs to each bio.
618 for (bio
= rq
->bio
; bio
; bio
= bio
->bi_next
) {
619 WARN_ON_ONCE((bio
->bi_opf
& REQ_FAILFAST_MASK
) &&
620 (bio
->bi_opf
& REQ_FAILFAST_MASK
) != ff
);
623 rq
->cmd_flags
|= REQ_MIXED_MERGE
;
626 static void blk_account_io_merge(struct request
*req
)
628 if (blk_do_io_stat(req
)) {
629 struct hd_struct
*part
;
632 cpu
= part_stat_lock();
635 part_round_stats(cpu
, part
);
636 part_dec_in_flight(part
, rq_data_dir(req
));
644 * Has to be called with the request spinlock acquired
646 static int attempt_merge(struct request_queue
*q
, struct request
*req
,
647 struct request
*next
)
649 if (!rq_mergeable(req
) || !rq_mergeable(next
))
652 if (req_op(req
) != req_op(next
))
658 if (blk_rq_pos(req
) + blk_rq_sectors(req
) != blk_rq_pos(next
))
661 if (rq_data_dir(req
) != rq_data_dir(next
)
662 || req
->rq_disk
!= next
->rq_disk
663 || req_no_special_merge(next
))
666 if (req_op(req
) == REQ_OP_WRITE_SAME
&&
667 !blk_write_same_mergeable(req
->bio
, next
->bio
))
671 * If we are allowed to merge, then append bio list
672 * from next to rq and release next. merge_requests_fn
673 * will have updated segment counts, update sector
676 if (!ll_merge_requests_fn(q
, req
, next
))
680 * If failfast settings disagree or any of the two is already
681 * a mixed merge, mark both as mixed before proceeding. This
682 * makes sure that all involved bios have mixable attributes
685 if ((req
->cmd_flags
| next
->cmd_flags
) & REQ_MIXED_MERGE
||
686 (req
->cmd_flags
& REQ_FAILFAST_MASK
) !=
687 (next
->cmd_flags
& REQ_FAILFAST_MASK
)) {
688 blk_rq_set_mixed_merge(req
);
689 blk_rq_set_mixed_merge(next
);
693 * At this point we have either done a back merge
694 * or front merge. We need the smaller start_time of
695 * the merged requests to be the current request
696 * for accounting purposes.
698 if (time_after(req
->start_time
, next
->start_time
))
699 req
->start_time
= next
->start_time
;
701 req
->biotail
->bi_next
= next
->bio
;
702 req
->biotail
= next
->biotail
;
704 req
->__data_len
+= blk_rq_bytes(next
);
706 elv_merge_requests(q
, req
, next
);
709 * 'next' is going away, so update stats accordingly
711 blk_account_io_merge(next
);
713 req
->ioprio
= ioprio_best(req
->ioprio
, next
->ioprio
);
714 if (blk_rq_cpu_valid(next
))
715 req
->cpu
= next
->cpu
;
717 /* owner-ship of bio passed from next to req */
719 __blk_put_request(q
, next
);
723 int attempt_back_merge(struct request_queue
*q
, struct request
*rq
)
725 struct request
*next
= elv_latter_request(q
, rq
);
728 return attempt_merge(q
, rq
, next
);
733 int attempt_front_merge(struct request_queue
*q
, struct request
*rq
)
735 struct request
*prev
= elv_former_request(q
, rq
);
738 return attempt_merge(q
, prev
, rq
);
743 int blk_attempt_req_merge(struct request_queue
*q
, struct request
*rq
,
744 struct request
*next
)
746 struct elevator_queue
*e
= q
->elevator
;
748 if (e
->type
->ops
.elevator_allow_rq_merge_fn
)
749 if (!e
->type
->ops
.elevator_allow_rq_merge_fn(q
, rq
, next
))
752 return attempt_merge(q
, rq
, next
);
755 bool blk_rq_merge_ok(struct request
*rq
, struct bio
*bio
)
757 if (!rq_mergeable(rq
) || !bio_mergeable(bio
))
760 if (req_op(rq
) != bio_op(bio
))
763 /* different data direction or already started, don't merge */
764 if (bio_data_dir(bio
) != rq_data_dir(rq
))
767 /* must be same device and not a special request */
768 if (rq
->rq_disk
!= bio
->bi_bdev
->bd_disk
|| req_no_special_merge(rq
))
771 /* only merge integrity protected bio into ditto rq */
772 if (blk_integrity_merge_bio(rq
->q
, rq
, bio
) == false)
775 /* must be using the same buffer */
776 if (req_op(rq
) == REQ_OP_WRITE_SAME
&&
777 !blk_write_same_mergeable(rq
->bio
, bio
))
783 int blk_try_merge(struct request
*rq
, struct bio
*bio
)
785 if (blk_rq_pos(rq
) + blk_rq_sectors(rq
) == bio
->bi_iter
.bi_sector
)
786 return ELEVATOR_BACK_MERGE
;
787 else if (blk_rq_pos(rq
) - bio_sectors(bio
) == bio
->bi_iter
.bi_sector
)
788 return ELEVATOR_FRONT_MERGE
;
789 return ELEVATOR_NO_MERGE
;