Re: SG does not ignore dxferp (direct io + mmap)

["Ewan D. Milne" <emilne@redhat.com>]

On Fri, 2016-11-25 at 12:56 -0500, Ewan Milne wrote:
> I think what we need to understand is what caused the regression in the
> first place, I probably should have been bisecting the original failure
> rather than trying to find where it started working.
> 

Bisecting leads to this commit:

commit 37f19e57a0de3c4a3417aa13ff4d04f1e0dee4b3
Author: Christoph Hellwig <hch@lst.de>
Date:   Sun Jan 18 16:16:33 2015 +0100

    block: merge __bio_map_user_iov into bio_map_user_iov
    
    And also remove the unused bdev argument.
    
    Signed-off-by: Christoph Hellwig <hch@lst.de>
    Reviewed-by: Ming Lei <tom.leiming@gmail.com>
    Signed-off-by: Jens Axboe <axboe@fb.com>

Specifically, the problem appears to be caused by the removal of
the setting of bio->bi_bdev, which would previously be set to NULL.
If I add:

diff --git a/block/bio.c b/block/bio.c
index 0723d4c..ecac37b 100644
--- a/block/bio.c
+++ b/block/bio.c
@@ -1351,6 +1351,7 @@ struct bio *bio_map_user_iov(struct request_queue
*q,
        if (iter->type & WRITE)
                bio->bi_rw |= REQ_WRITE;
 
+       bio->bi_bdev = NULL;
        bio->bi_flags |= (1 << BIO_USER_MAPPED);
 
        /*

The test passes (no zero byte corruption).

Setting dxferp would cause map_data.null_mapped to be set before it
is passed to blk_rq_map_user(_iov) which would cause a difference in
behavior.

-Ewan