NFSv4.1: Prevent a 3-way deadlock between layoutreturn, open and state recovery

Andy Adamson reports:

The state manager is recovering expired state and recovery OPENs are being
processed. If kswapd is pruning inodes at the same time, a deadlock can occur
when kswapd calls evict_inode on an NFSv4.1 inode with a layout, and the
resultant layoutreturn gets an error that the state mangager is to handle,
causing the layoutreturn to wait on the (NFS client) cl_rpcwaitq.

At the same time an open is waiting for the inode deletion to complete in
__wait_on_freeing_inode.

If the open is either the open called by the state manager, or an open from
the same open owner that is holding the NFSv4 sequence id which causes the
OPEN from the state manager to wait for the sequence id on the Seqid_waitqueue,
then the state is deadlocked with kswapd.

The fix is simply to have layoutreturn ignore all errors except NFS4ERR_DELAY.
We already know that layouts are dropped on all server reboots, and that
it has to be coded to deal with the "forgetful client model" that doesn't
send layoutreturns.

Reported-by: Andy Adamson <andros@netapp.com>
Link: http://lkml.kernel.org/r/1385402270-14284-1-git-send-email-andros@netapp.com
Signed-off-by: Trond Myklebust <Trond.Myklebust@primarydata.com>

diff --git a/fs/nfs/nfs4proc.c b/fs/nfs/nfs4proc.c
index f01e2aa53210ca2aa1ad493f1f45097bed2e22c9..e040359983ce5a65d9def1f9bf40d3c128272c00 100644 (file)
--- a/fs/nfs/nfs4proc.c
+++ b/fs/nfs/nfs4proc.c
@@ -7599,7 +7599,14 @@ static void nfs4_layoutreturn_done(struct rpc_task *task, void *calldata)
                return;
 
        server = NFS_SERVER(lrp->args.inode);
-       if (nfs4_async_handle_error(task, server, NULL) == -EAGAIN) {
+       switch (task->tk_status) {
+       default:
+               task->tk_status = 0;
+       case 0:
+               break;
+       case -NFS4ERR_DELAY:
+               if (nfs4_async_handle_error(task, server, NULL) != -EAGAIN)
+                       break;
                rpc_restart_call_prepare(task);
                return;
        }