OSDN Git Service

block: avoid blk_bio_segment_split for small I/O operations
authorChristoph Hellwig <hch@lst.de>
Mon, 4 Nov 2019 18:05:43 +0000 (10:05 -0800)
committerJens Axboe <axboe@kernel.dk>
Tue, 5 Nov 2019 00:13:54 +0000 (17:13 -0700)
__blk_queue_split() adds significant overhead for small I/O operations.
Add a shortcut to avoid it for cases where we know we never need to
split.

Based on a patch from Ming Lei.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
block/blk-merge.c

index 48e6725..f22cb62 100644 (file)
@@ -293,7 +293,7 @@ split:
 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 +309,20 @@ 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 (!q->limits.chunk_sectors &&
+                   (*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;
        }