pNFS: Ensure LAYOUTGET and LAYOUTRETURN are properly serialised

commit bf0291dd2267a2b9a4cd74d65249553d11bb45d6 upstream.

According to RFC5661, the client is responsible for serialising
LAYOUTGET and LAYOUTRETURN to avoid ambiguity. Consider the case
where we send both in parallel.

Client					Server
======					======
LAYOUTGET(seqid=X)
LAYOUTRETURN(seqid=X)
					LAYOUTGET return seqid=X+1
					LAYOUTRETURN return seqid=X+2
Process LAYOUTRETURN
          Forget layout stateid
Process LAYOUTGET
          Set seqid=X+1

The client processes the layoutget/layoutreturn in the wrong order,
and since the result of the layoutreturn was to clear the only
existing layout segment, the client forgets the layout stateid.

When the LAYOUTGET comes in, it is treated as having a completely
new stateid, and so the client sets the wrong sequence id...

Fix is to check if there are outstanding LAYOUTGET requests
before we send the LAYOUTRETURN (note that LAYOUGET will already
wait if it sees an outstanding LAYOUTRETURN).

Signed-off-by: Trond Myklebust <trond.myklebust@primarydata.com>
Signed-off-by: Trond Myklebust <trond.myklebust@primarydata.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>

diff --git a/fs/nfs/pnfs.c b/fs/nfs/pnfs.c
index ae340aca84c1a9128b75ca040b4c3e7bff5ea67d..5cd3568eea063a42812c03324acddfb5833975fd 100644 (file)
--- a/fs/nfs/pnfs.c
+++ b/fs/nfs/pnfs.c
@@ -365,6 +365,9 @@ pnfs_layout_need_return(struct pnfs_layout_hdr *lo,
 static bool
 pnfs_prepare_layoutreturn(struct pnfs_layout_hdr *lo)
 {
+       /* Serialise LAYOUTGET/LAYOUTRETURN */
+       if (atomic_read(&lo->plh_outstanding) != 0)
+               return false;
        if (test_and_set_bit(NFS_LAYOUT_RETURN, &lo->plh_flags))
                return false;
        lo->plh_return_iomode = 0;