Currently when falling back to doing I/O through the MDS (via
pnfs_{read|write}_through_mds), the client frees the nfs_pgio_header
without releasing the reference taken on the dreq
via pnfs_generic_pg_{read|write}pages -> nfs_pgheader_init ->
nfs_direct_pgio_init.  It then takes another reference on the dreq via
nfs_generic_pg_pgios -> nfs_pgheader_init -> nfs_direct_pgio_init and
as a result the requester will become stuck in inode_dio_wait.  Once
that happens, other processes accessing the inode will become stuck as
well.

Ensure that pnfs_read_through_mds() and pnfs_write_through_mds() clean
up correctly by calling hdr->completion_ops->completion() instead of
calling hdr->release() directly.

This can be reproduced (sometimes) by performing "storage failover
takeover" commands on NetApp filer while doing direct I/O from a client.

This can also be reproduced using SystemTap to simulate a failure while
doing direct I/O from a client (from Dave Wysochanski
<dwysocha@redhat.com>):

stap -v -g -e 'probe module("nfs_layout_nfsv41_files").function("nfs4_fl_prepare_ds").return { $return=NULL; exit(); }'

Suggested-by: Trond Myklebust <trond.myklebust@primarydata.com>
Signed-off-by: Scott Mayhew <smayhew@redhat.com>
Fixes: 1ca018d2

 ("pNFS: Fix a memory leak when attempted pnfs fails")
Cc: stable@vger.kernel.org
Signed-off-by: Trond Myklebust <trond.myklebust@primarydata.com>