Re: sunrpc: what prevents an xprt from being freed before task_cleanup runs?

[Trond Myklebust <trond.myklebust@fys.uio.no>]

On Tue, 2010-08-03 at 14:48 -0400, J. Bruce Fields wrote:
> On Tue, Aug 03, 2010 at 02:40:15PM -0400, Trond Myklebust wrote:
> > On Tue, 2010-08-03 at 13:24 -0400, Jeff Layton wrote:
> > > We got a report recently about a panic in RHEL5 (2.6.18 based kernel).
> > > The problem appears to be that a task_cleanup workqueue job ran and got
> > > passed a pointer to an xprt that had been freed. The bug is here in
> > > case anyone is interested in the details:
> > > 
> > >     https://bugzilla.redhat.com/show_bug.cgi?id=611938
> > > 
> > > The situation seems to be pretty difficult to reproduce, but I don't
> > > see anything that's intended to ensure that this doesn't occur in RHEL5
> > > or mainline. The task_cleanup workqueue job doesn't hold a reference to
> > > the xprt, and the job isn't canceled when the xprt is torn down.
> > > 
> > > Bruce had a look and suggested that we may need something like the
> > > patch below (pasted in, so it probably won't apply correctly). I've
> > > tested a backported version of it on RHEL5 and it seems to work fine.
> > > 
> > > Is it reasonable to cancel task_cleanup when destroying the xprt? Or,
> > > am I missing something that should prevent this situation in mainline
> > > (and perhaps isn't in RHEL5's kernel).
> > > 
> > > Any help is appreciated...
> > > 
> > > -----------------------------[snip]---------------------------------
> > > diff --git a/net/sunrpc/xprt.c b/net/sunrpc/xprt.c
> > > index dcd0132..2a1f664 100644
> > > --- a/net/sunrpc/xprt.c
> > > +++ b/net/sunrpc/xprt.c
> > > @@ -1129,6 +1129,7 @@ static void xprt_destroy(struct kref *kref)
> > >   rpc_destroy_wait_queue(&xprt->sending);
> > >   rpc_destroy_wait_queue(&xprt->resend);
> > >   rpc_destroy_wait_queue(&xprt->backlog);
> > > + cancel_work_sync(&xprt->task_cleanup);
> > >   /*
> > >    * Tear down transport state and free the rpc_xprt
> > >    */
> > > -----------------------------[snip]---------------------------------
> > > 
> > > Thanks,
> > 
> > How about doing a wait_on_bit_lock(&xprt->state, XPRT_LOCKED,...)
> > instead?
> 
> I'll believe you, but I don't get it: what guarantees that XPRT_LOCKED
> isn't cleared while xprt_autoclose autoclose is still running?

Look more closely: the callers that enqueue task_cleanup on the
workqueue all ensure that XPRT_LOCKED is held, and it isn't released
until the call to xprt_release_write() at the very end of
xprt_autoclose().

The rpciod_workqueue deadlocking is an issue, though, so unfortunately,
we won't be able to make use of the above.