On 03/14/2011 03:44 PM, Eric Blake wrote:
> No, that's exactly the_wrong_ reason for TCP_NODELAY. You simply
> cannot expect message boundaries to be respected when using SOCK_STREAM.
>
> So, either sendfd/recvfd must be documented to work only on SOCK_DGRAM
> sockets, or they have to be rethought (if possible at all).
Note that just last week libvirt found an issue with SOCK_STREAM hanging
forever on recvfd when the sendfd side was skipped, but SOCK_DGRAM was
able to reliably detect when the sending side of the socket is closed.
I'm perfectly fine with documenting that sendfd/recvfd must be used on
SOCK_DGRAM only.
But that was a different problem. That was not related to sendfd/recvfd.
However, there are cases in which you want to send a file descriptor as
out-of-band messages on a stream socket, and libvirt also has one of those.
From a quick experiment, BTW, on a SOCK_DGRAM Unix socket you don't
need to send fake data at the same time, but I may be wrong and/or that
may not be portable.
Paolo
