[erlang-bugs] Possible bug in file:sendfile/5

[Vincent Siliakus]

I was wondering if anybody had the time to look in to this, as this
seems to be a bug, or at least unexpected behaviour to me and I don't
have enough knowledge of BEAM's internals to dive deeper in to this
myself?

This was my first post to the erlang-bugs mailing list, so if I missed
some guidelines how to properly report an issue to this list, please
let me know.

-vincent

On Thu, Dec 12, 2013 at 3:00 PM, Vincent Siliakus <zambal@REDACTED> wrote:
> Hi,
>
> It seems that file:sendfile/5 in some situations corrupts or unexpectedly
> closes the socket it uses. I stumbled up on this while using Elli, a small
> but flexible http server (https://github.com/knutin/elli).
>
> The issue manifests itself when some clients are making single requests (ie.
> they close the socket after receiving a response) in a short timespan, where
> the response is a file send back to the clients, using file:sendfile/5. Once
> in a while, after an acceptor has responded and tries to reuse the
> connection and calls gen_tcp:recv/3 again, instead of receiving {error,
> closed}, it receives {error, ebadf}
>
> I created a small example, that imitates Elli's tcp acceptor loop, without
> all the http parsing stuff. When running this example I can reliably
> reproduce this issue on two systems.
>
> System one is a dual core Mac Book Pro, running Ubuntu 13.10 and Erlang
> 16B03. System two is single core VMWare VM, running Ubuntu 12.04 LTS and
> Erlang 15B03. I tried different async-thread settings on both systems (also
> disabling async-threads altogether, by using +A 0), which doesn't seem to
> make a difference.
>
> I published the example as a gist at github, which can be found here:
> https://gist.github.com/zambal/7927631
>
> In my example, sendfile:server/0 creates a listener, with {active, false}
> and spawns 10 processes as acceptors on that listening socket.
> sendfile:clients/0 spawns 50 clients that are all performing a single
> request, after which they close the connection.
>
> If you run the example and no {error, ebadf} manifests itself, you can
> uncomment the timer:sleep(100) call on line 34, as that seems to make it
> more likely that the error will popup. Note also that you can replace the
> sendfile call with an alternative implementation that uses file:read_file
> and gen_tcp:send, by changing the sendfile(Socket, Path) call with
> sendfile2(Socket, Path) on line 32. Using the alternative implementation, I
> can not reproduce the issue and everything seems to work as expected.
>
>
> -vincent