Return-Path: linux-nfs-owner@vger.kernel.org Received: from mx1.redhat.com ([209.132.183.28]:36631 "EHLO mx1.redhat.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1759814Ab3BUUgH (ORCPT ); Thu, 21 Feb 2013 15:36:07 -0500 Date: Thu, 21 Feb 2013 15:36:04 -0500 From: "J. Bruce Fields" To: "Myklebust, Trond" Cc: "linux-nfs@vger.kernel.org" , "chuck.lever@oracle.com" , "simo@redhat.com" Subject: Re: [PATCH 1/6] SUNRPC: make AF_LOCAL connect synchronous Message-ID: <20130221203603.GE3531@pad.fieldses.org> References: <1361464705-12340-1-git-send-email-bfields@redhat.com> <1361464705-12340-2-git-send-email-bfields@redhat.com> <4FA345DA4F4AE44899BD2B03EEEC2FA9235DA5F3@SACEXCMBX04-PRD.hq.netapp.com> <20130221194804.GC3531@pad.fieldses.org> <4FA345DA4F4AE44899BD2B03EEEC2FA9235DA946@SACEXCMBX04-PRD.hq.netapp.com> MIME-Version: 1.0 Content-Type: text/plain; charset=us-ascii In-Reply-To: <4FA345DA4F4AE44899BD2B03EEEC2FA9235DA946@SACEXCMBX04-PRD.hq.netapp.com> Sender: linux-nfs-owner@vger.kernel.org List-ID: On Thu, Feb 21, 2013 at 08:02:30PM +0000, Myklebust, Trond wrote: > On Thu, 2013-02-21 at 14:48 -0500, J. Bruce Fields wrote: > > On Thu, Feb 21, 2013 at 06:17:47PM +0000, Myklebust, Trond wrote: > > > On Thu, 2013-02-21 at 11:38 -0500, J. Bruce Fields wrote: > > > > From: "J. Bruce Fields" > > > > > > > > It doesn't appear that anyone actually needs to connect asynchronously. > > > > > > > > Also, using a workqueue for the connect means we lose the namespace > > > > information from the original process. This is a problem since there's > > > > no way to explicitly pass in a filesystem namespace for resolution of an > > > > AF_LOCAL address. > > > > > > > > Signed-off-by: J. Bruce Fields > > > > --- > > > > net/sunrpc/xprtsock.c | 35 +++++++++++++++++++++++++++-------- > > > > 1 file changed, 27 insertions(+), 8 deletions(-) > > > > > > > > diff --git a/net/sunrpc/xprtsock.c b/net/sunrpc/xprtsock.c > > > > index bbc0915..b1df874 100644 > > > > --- a/net/sunrpc/xprtsock.c > > > > +++ b/net/sunrpc/xprtsock.c > > > > @@ -1866,13 +1866,9 @@ static int xs_local_finish_connecting(struct rpc_xprt *xprt, > > > > * @xprt: RPC transport to connect > > > > * @transport: socket transport to connect > > > > * @create_sock: function to create a socket of the correct type > > > > - * > > > > - * Invoked by a work queue tasklet. > > > > */ > > > > -static void xs_local_setup_socket(struct work_struct *work) > > > > +static void xs_local_setup_socket(struct sock_xprt *transport) > > > > { > > > > - struct sock_xprt *transport = > > > > - container_of(work, struct sock_xprt, connect_worker.work); > > > > struct rpc_xprt *xprt = &transport->xprt; > > > > struct socket *sock; > > > > int status = -EIO; > > > > @@ -1919,6 +1915,31 @@ out: > > > > current->flags &= ~PF_FSTRANS; > > > > } > > > > > > > > +static void xs_local_connect(struct rpc_task *task) > > > > +{ > > > > + struct rpc_xprt *xprt = task->tk_xprt; > > > > + struct sock_xprt *transport = container_of(xprt, struct sock_xprt, xprt); > > > > + unsigned long timeout; > > > > + > > > > + if (RPC_IS_ASYNC(task)) > > > > + rpc_exit(task, -ENOTCONN); > > > > > > Needs a "return"... > > > > Fixed, thanks. > > > > > > > > > > > > + > > > > + if (transport->sock != NULL && !RPC_IS_SOFTCONN(task)) { > > > > + dprintk("RPC: xs_connect delayed xprt %p for %lu " > > > > + "seconds\n", > > > > + xprt, xprt->reestablish_timeout / HZ); > > > > + timeout = xprt->reestablish_timeout; > > > > + xprt->reestablish_timeout <<= 1; > > > > + if (xprt->reestablish_timeout < XS_TCP_INIT_REEST_TO) > > > > + xprt->reestablish_timeout = XS_TCP_INIT_REEST_TO; > > > > + if (xprt->reestablish_timeout > XS_TCP_MAX_REEST_TO) > > > > + xprt->reestablish_timeout = XS_TCP_MAX_REEST_TO; > > > > + rpc_delay(task, timeout); > > > > > > This too needs to exit in order to sleep. > > > > Oops, so maybe simplest would be just to connect first and then delay if > > there's a failure. > > > > (Or is there any problem just making that rpc_delay() an msleep() since > > we know we're in the synchronous case?) > > A 5 minute msleep() is probably a bit excessive. Making it interruptible > might help, but ... > That does however raise the issue of why we need exponential back off > here? An AF_LOCAL connect call is pretty efficient. And the only excuse for the connect not succeeding is, what, rpcbind just crashed or something? In which case I think our only responsibility is just not to spin furiously. How about just something like ret = xs_local_setup_socket(transport); if (ret && !RPC_IS_SOFTCONN(task)) msleep_interruptible(1000); return; ? Or we could keep the same exponential timeout logic and just adjust the min/max. I have no strong opinions here.... --b.