On Mon, Nov 04, 2019 at 10:05:43AM -0800, Christoph Hellwig wrote:
> diff --git a/block/blk-merge.c b/block/blk-merge.c
> index 48e6725b32ee..06eb38357b41 100644
> --- a/block/blk-merge.c
> +++ b/block/blk-merge.c
> @@ -293,7 +293,7 @@ static struct bio *blk_bio_segment_split(struct request_queue *q,
> void __blk_queue_split(struct request_queue *q, struct bio **bio,
> unsigned int *nr_segs)
> {
> - struct bio *split;
> + struct bio *split = NULL;
>
> switch (bio_op(*bio)) {
> case REQ_OP_DISCARD:
> @@ -309,6 +309,19 @@ void __blk_queue_split(struct request_queue *q, struct bio **bio,
> nr_segs);
> break;
> default:
> + /*
> + * All drivers must accept single-segments bios that are <=
> + * PAGE_SIZE. This is a quick and dirty check that relies on
> + * the fact that bi_io_vec[0] is always valid if a bio has data.
> + * The check might lead to occasional false negatives when bios
> + * are cloned, but compared to the performance impact of cloned
> + * bios themselves the loop below doesn't matter anyway.
> + */
> + if ((*bio)->bi_vcnt == 1 &&
> + (*bio)->bi_io_vec[0].bv_len <= PAGE_SIZE) {
> + *nr_segs = 1;
> + break;
> + }
> split = blk_bio_segment_split(q, *bio, &q->bio_split, nr_segs);
> break;
> }
If the device advertises a chunk boundary and this small IO happens to
cross it, skipping the split is going to harm performance.
