Re: [PATCH] remote-curl: don't hang when a server dies before any output

From: Jeff King <peff@peff.net>
Date: Nov 14, 2016, 18:24 UTC
Message-ID: <20161114182431.e7jjnq422c4xobdb@sigill.intra.peff.net>
In-Reply-To: <1478729910-26232-1-git-send-email-dturner@twosigma.com>

On Wed, Nov 09, 2016 at 05:18:30PM -0500, David Turner wrote:

> In the event that a HTTP server closes the connection after giving a
> 200 but before giving any packets, we don't want to hang forever
> waiting for a response that will never come.  Instead, we should die
> immediately.

I agree we don't want to hang forever, but this leaves open the question: what is hanging?

My guess is that fetch-pack is waiting for more data from the server, and remote-curl is waiting for fetch-pack to tell us what to send for the next request. Neither will make forward progress because they are effectively waiting on each other.

Which means this is likely a special case of malformed input from the server. A server which likewise sends a partial response could end up in the same deadlock, I would think (e.g., a half-finished pktline, or a pktline but no trailing flush).

That doesn't make it wrong to fix this specific case (especially if it's a common one), but I wonder if we could do better.

The root of the issue is that only fetch-pack understands the protocol, and remote-curl is blindly proxying the data. But only remote-curl knows that the HTTP request has ended, and it doesn't relay that information to fetch-pack. So I can think of two solutions:

  1. Some way of remote-curl communicating the EOF to fetch-pack. It
     can't just close the descriptor, since we need to pass more data
     over it for the followup requests. You'd need something
     out-of-band, or to frame the HTTP data inside another layer of
     pktlines, both of which are kind of gross.

  2. Have remote-curl understand enough of the protocol that it can
     abort rather than hang.

     I think that's effectively the approach of your patch, but for one
     specific case. But could we, for example, make sure that everything
     we proxy is a complete set of pktlines and ends with a flush? And
     if not, then we hang up on fetch-pack.

     I _think_ that would work, because even the pack is always encased
     in pktlines for smart-http.

Show 6 quoted lines

> @@ -659,6 +662,8 @@ static int post_rpc(struct rpc_state *rpc)
>  	curl_easy_setopt(slot->curl, CURLOPT_WRITEFUNCTION, rpc_in);
>  	curl_easy_setopt(slot->curl, CURLOPT_FILE, rpc);
>  
> +
> +	rpc->any_written = 0;

Extra blank line here?

Show 7 quoted lines

> @@ -667,6 +672,9 @@ static int post_rpc(struct rpc_state *rpc)
>  	if (err != HTTP_OK)
>  		err = -1;
>  
> +	if (!rpc->any_written)
> +		err = -1;
> +

I wondered if there were any cases where it was normal for the server to return zero bytes. Possibly the ref advertisement is one, but this is _just_ handling post_rpc(), so that's OK. And I think by definition every response has to at least return a flush packet, or we would make no forward progress (i.e., the exact case you are dealing with here).

-Peff