[Network] allow setting socket options

[azjezz]

see: https://github.com/facebook/hhvm/issues/8451

[fredemmott]

Unless I'm missing something, the HHVM issue is not relevant here - the sockets are created with socket_create, which means that TCP_NODELAY and SO_KEEPALIVE can be set with socket_set_option

[azjezz]

HHVM doesn't support no-dely / keep-alive :disappointed:

[fredemmott]

That doesn't appear to be the case

hphpd> $sock = socket_create(AF_INET, SOCK_STREAM, SOL_TCP)
hphpd> socket_set_option($sock, SOL_SOCKET, /* random number */ 12345, 1)
Warning: unable to set socket option [42]: Protocol not available
hphpd> =socket_get_option($sock, SOL_SOCKET, SO_KEEPALIVE)
0
hphpd> =socket_get_option($sock, SOL_TCP, TCP_NODELAY)
=socket_get_option($sock, SOL_TCP, TCP_NODELAY)
0
hphpd> socket_set_option($sock, SOL_SOCKET, SO_KEEPALIVE, 1)
hphpd> socket_set_option($sock, SOL_TCP, TCP_NODELAY, 1)
hphpd> =socket_get_option($sock, SOL_SOCKET, SO_KEEPALIVE)
8
hphpd> =socket_get_option($sock, SOL_TCP, TCP_NODELAY)
4

[fredemmott]

that said, retitled as I won't be doing these in small batches; there's many SO and TCP options, and I'm hoping to avoid adding something as rough as mixed, but that might end up being necessary. At the least, there should be an enum for socket options, a different enum for tcp options, etc.

[azjezz]

That doesn't appear to be the case

:scream: i clearly remember they didn't exist back when i was working on #36

At the least, there should be an enum for socket options, a different enum for tcp options, etc.

wouldn't using a shape be "stricter" ? e.g, keep-alive value should be bool, while timeout should be float

[fredemmott]

perhaps - probably separate shapes, then TCP\Conenct would take shape('socket_options' => SocketOptionsShape', 'tcp_options' => TcpOptionsShape) or something like that or options -> socket, options -> tcp

[azjezz]

or maybe tcp options shape can have all socket options therefore it can be passed as both ? :)

[fredemmott]

while timeout should be float

When we move to native implementaitons, I'm probably going to change all timeouts to be ints, or something like the timeval C struct - the native implementaiton is an int seconds and int microseconds. It's not as convenient, but avoids issues like https://github.com/hhvm/hsl-experimental/pull/104

or maybe tcp options shape can have all socket options therefore it can be passed as both ? :)

• I wouldn't want to do this unless we get shape splat
• doesn't really fit with the type system - shape('TCP_NODELAY' => bool, 'SO_KEEPALIVE' => bool) is not a subtype of shape('SO_KEEPALIVE' => bool), and ... should definitely not be part of these definitions. Perhaps records could make that reasonable.

[fredemmott]

😱 i clearly remember they didn't exist back when i was working on #36

Looks like they were added in 2015; for both #36 and the HHVM issue, you were specific about the stream_socket_*() functions - they still appear to be unsupported for stream_context_set_option() - could be this is what you were referring to, instead of socket_set_option() ?

[azjezz]

probably :thinking:

[fredemmott]

Splat would be really nice here; there's a few (SO_NREAD, SO_ERROR, SO_NWRITE in particular) that it would be nice to type as invalid for setsockopt, but valid for getsockopt, without duplciating the entire shape