dm mpath: switch paths in dm_blk_ioctl() code path

[ Upstream commit 2361ae595352dec015d14292f1b539242d8446d6 ]

SCSI LUN passthrough code such as qemu's "scsi-block" device model
pass every IO to the host via SG_IO ioctls. Currently, dm-multipath
calls choose_pgpath() only in the block IO code path, not in the ioctl
code path (unless current_pgpath is NULL). This has the effect that no
path switching and thus no load balancing is done for SCSI-passthrough
IO, unless the active path fails.

Fix this by using the same logic in multipath_prepare_ioctl() as in
multipath_clone_and_map().

Note: The allegedly best path selection algorithm, service-time,
still wouldn't work perfectly, because the io size of the current
request is always set to 0. Changing that for the IO passthrough
case would require the ioctl cmd and arg to be passed to dm's
prepare_ioctl() method.

Signed-off-by: Martin Wilck <mwilck@suse.com>
Reviewed-by: Hannes Reinecke <hare@suse.de>
Signed-off-by: Mike Snitzer <snitzer@redhat.com>
Signed-off-by: Sasha Levin <sashal@kernel.org>

diff --git a/drivers/md/dm-mpath.c b/drivers/md/dm-mpath.c
index 456d790c918c00a32a1e3b8451144b5ae7d311ef..f2de4c73cc8fa534e09615f6df9510897016e9b0 100644 (file)
--- a/drivers/md/dm-mpath.c
+++ b/drivers/md/dm-mpath.c
@@ -1856,7 +1856,7 @@ static int multipath_prepare_ioctl(struct dm_target *ti,
        int r;
 
        current_pgpath = READ_ONCE(m->current_pgpath);
-       if (!current_pgpath)
+       if (!current_pgpath || !test_bit(MPATHF_QUEUE_IO, &m->flags))
                current_pgpath = choose_pgpath(m, 0);
 
        if (current_pgpath) {