From mboxrd@z Thu Jan 1 00:00:00 1970 From: Jeff Layton Subject: Re: [PATCH][RFC] use after free in NLM subsystem -- how best to fix it? Date: Thu, 27 Sep 2007 21:13:07 -0400 Message-ID: <20070927211307.cfe63de9.jlayton@redhat.com> References: <20070924161236.48779845.jlayton@poochiereds.net> <1190672003.6700.34.camel@heimdal.trondhjem.org> <20070925102501.c770c202.jlayton@redhat.com> <1190739948.7330.18.camel@heimdal.trondhjem.org> <20070927135938.d5e733c6.jlayton@redhat.com> <20070927183803.GF10113@fieldses.org> <20070927150927.12d07946.jlayton@redhat.com> <20070927205533.GC21523@fieldses.org> Mime-Version: 1.0 Content-Type: text/plain; charset="us-ascii" Cc: nfs@lists.sourceforge.net To: "J. Bruce Fields" Return-path: Received: from sc8-sf-mx2-b.sourceforge.net ([10.3.1.92] helo=mail.sourceforge.net) by sc8-sf-list2-new.sourceforge.net with esmtp (Exim 4.43) id 1Ib4Pk-00059j-Dv for nfs@lists.sourceforge.net; Thu, 27 Sep 2007 18:13:08 -0700 Received: from mx1.redhat.com ([66.187.233.31]) by mail.sourceforge.net with esmtp (Exim 4.44) id 1Ib4Pn-0004tY-3I for nfs@lists.sourceforge.net; Thu, 27 Sep 2007 18:13:13 -0700 In-Reply-To: <20070927205533.GC21523@fieldses.org> List-Id: "Discussion of NFS under Linux development, interoperability, and testing." List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , Sender: nfs-bounces@lists.sourceforge.net Errors-To: nfs-bounces@lists.sourceforge.net On Thu, 27 Sep 2007 16:55:33 -0400 "J. Bruce Fields" wrote: > On Thu, Sep 27, 2007 at 03:09:27PM -0400, Jeff Layton wrote: > > On Thu, 27 Sep 2007 14:38:03 -0400 > > "J. Bruce Fields" wrote: > > > > > On Thu, Sep 27, 2007 at 01:59:38PM -0400, Jeff Layton wrote: > > > > Now that I've started really digging into this, I'm thinking that I may > > > > be wrong about the race that exists in current mainline. There was a > > > > change done ~June 2007: > > > > > > > > commit 34f52e3591f241b825353ba27def956d8487c400 > > > > Author: Trond Myklebust > > > > Date: Thu Jun 14 16:40:31 2007 -0400 > > > > > > > > SUNRPC: Convert rpc_clnt->cl_users to a kref > > > > > > > > Signed-off-by: Trond Myklebust > > > > > > > > ...this changed nlm_destroy_host from just setting cl_dead to instead > > > > use rpc_shutdown_client. So this code now actually kills active RPC > > > > tasks for the RPC client and waits for them to come down instead of > > > > just marking the client dead. This should mitigate the race that > > > > definitely exists in earlier kernels. > > > > > > Is there still a window where lockd could be killed just as someone is > > > starting a new rpc (but the task isn't yet visible to > > > rpc_shutdown_client)? > > > > > > > Perhaps, but I don't think that's the case here. Here is the oops > > message: > > > > https://bugzilla.redhat.com/show_bug.cgi?id=253754#c7 > > > > It crashed in rpciod while doing svc_wake_up from an async call. Unless > > I'm missing something, the only way that could happen is from > > nlmsvc_grant_callback. That's the rpc callback from > > nlmsvc_grant_blocked, and that function is only ever called from lockd > > itself. > > > > So that question becomes: > > > > Is there still a window where lockd could be killed just as lockd is > > starting a new rpc (but the task isn't yet visible to > > rpc_shutdown_client)? > > > > I'm thinking the answer here is no, since the call would happen near the > > top of the event loop, and nlm_shutdown_hosts occurs well after that. > > Without actually looking at the code (but going on the memory of a > similar-looking bug in the delegation callback code): is there a reason > the crash would have to occur right after the rpc_shutdown_client() > call? If the problem occurs because, say, task->tk_client points to > freed memory, it may take a while for that memory to actually be > overwritten, so it may look just OK enough for the rpc code to still > limp on a little while longer before crashing. > The kernel where I saw this had CONFIG_SLUB_DEBUG_ON, so the memory should have been poisoned after free (you can see some of the poisoning patterns in the registers in the oops). One possibility is that lockd took a long time to come down and a new lockd was started before it did. That might prevent nlmsvc_invalidate_all from being called at all on the old lockd. The new lockd then goes to process the block, async call goes out and the callback is called, but by that time b_daemon points to freed memory. Alas, I didn't get a coredump from that crash, so I can't go back and look to see if the KERN_DEBUG message got printed. IIRC though, I was thrashing lockd up and down pretty quickly, so the above race might not be too unlikely. I think for now, I'm going to focus on defining what the behavior should be as far as allowing a new lockd to start while the old one is still running. And then once we do that, to try to clean up the locking around some of the vars that define that behavior. That seems to be an obvious problem anyway and I have a suspicion it's at least related to this crash... -- Jeff Layton ------------------------------------------------------------------------- This SF.net email is sponsored by: Microsoft Defy all challenges. Microsoft(R) Visual Studio 2005. http://clk.atdmt.com/MRT/go/vse0120000070mrt/direct/01/ _______________________________________________ NFS maillist - NFS@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/nfs