NFS/RPC/GSS - oops in gss_pipe_release()]

Resending since this problem still appears to be in both the -mm and mainline kernels..

steved.

-------- Original Message --------
Subject: [PATCH] NFS/RPC/GSS - oops in gss_pipe_release()
Date: Fri, 16 Sep 2005 13:40:15 -0400
From: Steve Dickson <[email protected]>
To: [email protected]

During some recent debugging I found that an oops can occur in gss_pipe_release() because the client handle that is being passed in has already been freed.

The scenario is as follows:

1) root# mount -o sec=krb5 server:/export /mnt/export
2) user$ ls /mnt/export (which hangs because user does not have the correct credentials)
3) root# reboot

The oops occurs when the /mnt/export filesystem is unmounted. The reason being is gss_pipe_release() was already called when the ls process was killed. The stack dump of the ls process was:

  [<e09dda84>] gss_pipe_release+0x74/0xd8 [auth_rpcgss]
  [<e0a045c0>] rpc_pipe_release+0xa5/0xb9 [sunrpc]
  [<c015a9d6>] __fput+0x55/0x100
  [<c0159626>] filp_close+0x59/0x5f
  [<c012363f>] put_files_struct+0x57/0xc0
  [<c0124255>] do_exit+0x227/0x3de
  [<c01244fa>] sys_exit_group+0x0/0xd
  [<c02d120b>] syscall_call+0x7/0xb

So when the rpc_shutdown_client code is called via the umount:

  [<e09dda84>] gss_pipe_release+0x74/0xd8 [auth_rpcgss]
  [<e0a0447c>] rpc_close_pipes+0x80/0x9a [sunrpc]
  [<e0a04bb0>] rpc_depopulate+0xfb/0x142 [sunrpc]
  [<c01651bc>] cached_lookup+0xf/0x56
  [<c0166410>] __lookup_hash+0x46/0x89
  [<e0a05005>] rpc_rmdir+0x5a/0x89 [sunrpc]
  [<e09fcede>] rpcauth_free_credcache+0x87/0xd0 [sunrpc]
  [<e09f8431>] rpc_destroy_client+0x70/0xa4 [sunrpc]
  [<e09f8421>] rpc_destroy_client+0x60/0xa4 [sunrpc]
  [<e09f83ba>] rpc_shutdown_client+0xd1/0xd8 [sunrpc]
  [<c011e586>] default_wake_function+0x0/0xc
  [<e0aefe75>] nfs_kill_super+0x38/0x63 [nfs]

the client handle (which is in the rpc_inode) passed to gss_pipe_release() has already been freeded.

It appears from other places in the code (namely rpc_close_pipes()) that the only way to invalidate an rpc_inode is to set the ops pointer to NULL which is what the attached patch does.

Is there a better way to invalid an rpc_inode?

steved.

This patch stops the release_pipe() funtion from being called twice by invalidating the ops pointer in the rpc_inode when rpc_pipe_release() is called.

------------------------------------------------------

@@ -177,6 +177,8 @@ rpc_pipe_release(struct inode *inode, st
 		__rpc_purge_upcall(inode, -EPIPE);
 	if (rpci->ops->release_pipe)
 		rpci->ops->release_pipe(inode);
+	if (!rpci->nreaders && !rpci->nwriters)
+		rpci->ops = NULL;
 out:
 	up(&inode->i_sem);
 	return 0;