I'm bisecting now, there are a couple of patches from Ming in
the area of the bio splitting code...

CC'ing Ming, Linux-block and Linux-nvme as iser is identical to nvme
wrt the virtual boundary so I think nvme will break as well.

Bisection reveals that this one is the culprit:

commit 52cc6eead9095e2faf2ec7afc013aa3af1f01ac5
Author: Ming Lei [off-list ref]
Date:   Thu Sep 17 09:58:38 2015 -0600

     block: blk-merge: fast-clone bio when splitting rw bios

     biovecs has become immutable since v3.13, so it isn't necessary
     to allocate biovecs for the new cloned bios, then we can save
     one extra biovecs allocation/copy, and the allocation is often
     not fixed-length and a bit more expensive.

     For example, if the 'max_sectors_kb' of null blk's queue is set
     as 16(32 sectors) via sysfs just for making more splits, this patch
     can increase throught about ~70% in the sequential read test over
     null_blk(direct io, bs: 1M).

     Cc: Christoph Hellwig [off-list ref]
     Cc: Kent Overstreet [off-list ref]
     Cc: Ming Lin [off-list ref]
     Cc: Dongsu Park [off-list ref]
     Signed-off-by: Ming Lei [off-list ref]

     This fixes a performance regression introduced by commit 54efd50bfd,
     and allows us to take full advantage of the fact that we have immutable
     bio_vecs. Hand applied, as it rejected violently with commit
     5014c311baa2.

     Signed-off-by: Jens Axboe [off-list ref]
--
--
To unsubscribe from this list: send the line "unsubscribe linux-rdma" in
the body of a message to majordomo-u79uwXL29TY76Z2rM5mHXA@public.gmane.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html