Re: [PATCH 1/2] SUNRPC: Ensure that call_connect times out correctly

[Jeff Layton <jlayton@redhat.com>]

On Mon, 17 Mar 2014 14:40:44 -0400
Trond Myklebust <trond.myklebust@primarydata.com> wrote:

> When the server is unavailable due to a networking error, etc, we want
> the RPC client to respect the timeout delays when attempting to reconnect.
> 
> Fixes: 561ec1603171 (SUNRPC: call_connect_status should recheck bind..)
> Signed-off-by: Trond Myklebust <trond.myklebust@primarydata.com>
> ---
>  net/sunrpc/clnt.c | 8 +++-----
>  1 file changed, 3 insertions(+), 5 deletions(-)
> 
> diff --git a/net/sunrpc/clnt.c b/net/sunrpc/clnt.c
> index 0edada973434..f22d3a115fda 100644
> --- a/net/sunrpc/clnt.c
> +++ b/net/sunrpc/clnt.c
> @@ -1798,10 +1798,6 @@ call_connect_status(struct rpc_task *task)
>  	trace_rpc_connect_status(task, status);
>  	task->tk_status = 0;
>  	switch (status) {
> -		/* if soft mounted, test if we've timed out */
> -	case -ETIMEDOUT:
> -		task->tk_action = call_timeout;
> -		return;
>  	case -ECONNREFUSED:
>  	case -ECONNRESET:
>  	case -ECONNABORTED:
> @@ -1812,7 +1808,9 @@ call_connect_status(struct rpc_task *task)
>  		if (RPC_IS_SOFTCONN(task))
>  			break;
>  	case -EAGAIN:
> -		task->tk_action = call_bind;
> +	case -ETIMEDOUT:
> +		/* Check if we've timed out before looping back to call_bind */
> +		task->tk_action = call_timeout;
>  		return;
>  	case 0:
>  		clnt->cl_stats->netreconn++;

I believe this patch may have broken the v4.0 callback channel
establishment code in nfsd. I think what's happening is this:

nfsd tries to create a RPC_TASK_SOFTCONN call to probe the cb channel
with a CB_NULL. It queues the connect_worker to the workqueue. That
establishes the socket and then gets a callback from the socket layer
into xs_tcp_state_change for TCP_ESTABLISHED.

That code does:

    xprt_wake_pending_tasks(xprt, -EAGAIN);

...that wakes the task up, and sets the tk_status to -EAGAIN, and it
then moves on to call_timeout due to this patch. That code then does
this:

        if (RPC_IS_SOFTCONN(task)) {
                rpc_exit(task, -ETIMEDOUT);
                return;
        }

...and the callback ping then fails with an error. Reverting this patch
seems to fix it. I see several ways that we could fix this, but I'm not
clear on the right way. Maybe we shouldn't be waking up the tasks with
-EAGAIN in the TCP_ESTABLISHED case?

-- 
Jeff Layton <jlayton@redhat.com>