block: don't decrement nr_phys_segments for physically contigous segments
authorChristoph Hellwig <hch@lst.de>
Tue, 21 May 2019 07:01:40 +0000 (09:01 +0200)
committerJens Axboe <axboe@kernel.dk>
Thu, 23 May 2019 16:25:26 +0000 (10:25 -0600)
Currently ll_merge_requests_fn, unlike all other merge functions,
reduces nr_phys_segments by one if the last segment of the previous,
and the first segment of the next segement are contigous.  While this
seems like a nice solution to avoid building smaller than possible
requests it causes a mismatch between the segments actually present
in the request and those iterated over by the bvec iterators, including
__rq_for_each_bio.  This can for example mistrigger the single segment
optimization in the nvme-pci driver, and might lead to mismatching
nr_phys_segments number when recalculating the number of request
when inserting a cloned request.

We could possibly work around this by making the bvec iterators take
the front and back segment size into account, but that would require
moving them from the bio to the bio_iter and spreading this mess
over all users of bvecs.  Or we could simply remove this optimization
under the assumption that most users already build good enough bvecs,
and that the bio merge patch never cared about this optimization
either.  The latter is what this patch does.

dff824b2aadb ("nvme-pci: optimize mapping of small single segment requests").
Reviewed-by: Ming Lei <ming.lei@redhat.com>
Reviewed-by: Hannes Reinecke <hare@suse.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
block/blk-merge.c

index 21e87a7..80a5a0f 100644 (file)
@@ -358,7 +358,6 @@ static unsigned int __blk_recalc_rq_segments(struct request_queue *q,
        unsigned front_seg_size;
        struct bio *fbio, *bbio;
        struct bvec_iter iter;
-       bool new_bio = false;
 
        if (!bio)
                return 0;
@@ -379,31 +378,12 @@ static unsigned int __blk_recalc_rq_segments(struct request_queue *q,
        nr_phys_segs = 0;
        for_each_bio(bio) {
                bio_for_each_bvec(bv, bio, iter) {
-                       if (new_bio) {
-                               if (seg_size + bv.bv_len
-                                   > queue_max_segment_size(q))
-                                       goto new_segment;
-                               if (!biovec_phys_mergeable(q, &bvprv, &bv))
-                                       goto new_segment;
-
-                               seg_size += bv.bv_len;
-
-                               if (nr_phys_segs == 1 && seg_size >
-                                               front_seg_size)
-                                       front_seg_size = seg_size;
-
-                               continue;
-                       }
-new_segment:
                        bvec_split_segs(q, &bv, &nr_phys_segs, &seg_size,
                                        &front_seg_size, NULL, UINT_MAX);
-                       new_bio = false;
                }
                bbio = bio;
-               if (likely(bio->bi_iter.bi_size)) {
+               if (likely(bio->bi_iter.bi_size))
                        bvprv = bv;
-                       new_bio = true;
-               }
        }
 
        fbio->bi_seg_front_size = front_seg_size;
@@ -725,7 +705,6 @@ static int ll_merge_requests_fn(struct request_queue *q, struct request *req,
                        req->bio->bi_seg_front_size = seg_size;
                if (next->nr_phys_segments == 1)
                        next->biotail->bi_seg_back_size = seg_size;
-               total_phys_segments--;
        }
 
        if (total_phys_segments > queue_max_segments(q))