Return-Path: Received: from mail-gw0-f46.google.com ([74.125.83.46]:36216 "EHLO mail-gw0-f46.google.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1751201Ab1BOO7F (ORCPT ); Tue, 15 Feb 2011 09:59:05 -0500 Received: by gwj20 with SMTP id 20so103086gwj.19 for ; Tue, 15 Feb 2011 06:59:04 -0800 (PST) From: Jeff Layton To: Trond Myklebust Cc: linux-nfs@vger.kernel.org Subject: [PATCH] nfs: don't queue synchronous NFSv4 close rpc_release to nfsiod Date: Tue, 15 Feb 2011 09:58:59 -0500 Message-Id: <1297781939-1400-1-git-send-email-jlayton@redhat.com> Sender: linux-nfs-owner@vger.kernel.org List-ID: Content-Type: text/plain MIME-Version: 1.0 I recently had some of our QA people report some connectathon test failures in RHEL5 (2.6.18-based kernel). For some odd reason (maybe scheduling differences that make the race more likely?) the problem occurs more frequently on s390. The problem generally manifests itself on NFSv4 as a race where an rmdir fails because a silly-renamed file in the directory wasn't deleted in time. Looking at traces, what you usually see is the failing rmdir attempt that fails with the sillydelete of the file that prevented it very soon afterward. Silly deletes are handled via dentry_iput and in the case of a close on NFSv4, the last dentry reference is often held by the CLOSE RPC task. nfs4_do_close does the close as an async RPC task that it conditionally waits on depending on whether the close is synchronous or not. It also sets the workqueue for the task to nfsiod_workqueue. When tk_workqueue is set, the rpc_release operation is queued to that workqueue. rpc_release is where the dentry reference held by the task is put. The caller has no way to wait for that to complete, so the close(2) syscall can easily return before the rpc_release call is ever done. In some cases, that rpc_release is delayed for a long enough to prevent a subsequent rmdir of the containing directory. I believe this is a bug, or at least not ideal behavior. We should try not to have the close(2) call return in this situation until the sillydelete is done. I've been able to reproduce this more reliably by adding a 100ms sleep at the top of nfs4_free_closedata. I've not seen it "in the wild" on mainline kernels, but it seems quite possible when a machine is heavily loaded. This patch fixes this by not setting tk_workqueue in nfs4_do_close when the wait flag is set. This makes the final rpc_put_task a synchronous operation and should prevent close(2) from returning before the dentry_iput is done. Signed-off-by: Jeff Layton --- fs/nfs/nfs4proc.c | 5 ++++- 1 files changed, 4 insertions(+), 1 deletions(-) diff --git a/fs/nfs/nfs4proc.c b/fs/nfs/nfs4proc.c index 78936a8..4cabfea 100644 --- a/fs/nfs/nfs4proc.c +++ b/fs/nfs/nfs4proc.c @@ -1988,11 +1988,14 @@ int nfs4_do_close(struct path *path, struct nfs4_state *state, gfp_t gfp_mask, i .rpc_client = server->client, .rpc_message = &msg, .callback_ops = &nfs4_close_ops, - .workqueue = nfsiod_workqueue, .flags = RPC_TASK_ASYNC, }; int status = -ENOMEM; + /* rpc_release must be synchronous too if "wait" is set */ + if (!wait) + task_setup_data.workqueue = nfsiod_workqueue; + calldata = kzalloc(sizeof(*calldata), gfp_mask); if (calldata == NULL) goto out; -- 1.7.4