Return-Path: linux-nfs-owner@vger.kernel.org Received: from mail-la0-f42.google.com ([209.85.215.42]:38642 "EHLO mail-la0-f42.google.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1752086AbbAUW3l convert rfc822-to-8bit (ORCPT ); Wed, 21 Jan 2015 17:29:41 -0500 Received: by mail-la0-f42.google.com with SMTP id ms9so20693547lab.1 for ; Wed, 21 Jan 2015 14:29:40 -0800 (PST) MIME-Version: 1.0 Date: Wed, 21 Jan 2015 17:29:39 -0500 Message-ID: Subject: race between inode eviction with delegation and opening same file From: Olga Kornievskaia To: Trond Myklebust Cc: linux-nfs Content-Type: text/plain; charset=UTF-8 Sender: linux-nfs-owner@vger.kernel.org List-ID: There is a race between return of a delegation (due to resource constraints and inode getting evicted) and an open for the same file because…. In nfs_inode_return_delegation_noreclaim() we detach that delegation from the inode first (nfs_inode_detach_delegation()) and then we send a delegreturn. In between of nfs_inode_detach_delegation() and nfs_do_return_delegation() an open request comes and it doesn’t see any delegations for this inode and thus it sends an open request. At the same time eviction thread is working on doing a delegreturn for this inode. The server gets request for the open first and returns a delegation for it (note: server will issue the same delegation stateid but different seqnum as the one currently being returned by the client). Then the server processes a delegreturn for this file (from its perspective delegation stateid is no longer valid). Yet, on the client side as it processes a reply from the server, it adds a “new” delegation to the inode and proceeds to use it for an IO request later which errors in BAD_STATEID. I don't see how we can make detaching the delegation and the return of it atomic. Instead I propose the following solution: I propose to examine the delegation we get on open. If the sequence number is not 0 and we don't have a delegation already then we must have experienced this race. Therefore, we should return the delegation (and ignore if this errors, as the server thinks there is no more delegation). Something like this. Disclaimer: it doesn't seem to be the actual fix (as it hangs the machine) so I'm still missing something .... diff --git a/fs/nfs/delegation.c b/fs/nfs/delegation.c index b76166b..01bba63 100644 --- a/fs/nfs/delegation.c +++ b/fs/nfs/delegation.c @@ -368,5 +368,12 @@ int nfs_inode_set_delegation(struct inode *inode, struct rpc_cred *cred, struct old_delegation, clp); if (freeme == NULL) goto out; + } else { + if (be32_to_cpu(delegation->stateid.seqid) > 0) { + nfs_do_return_delegation(inode, delegation, 0); + goto out; + } } list_add_rcu(&delegation->super_list, &server->delegations); nfsi->delegation_state = delegation->type;