splice() on unix sockets

[Mark Hills <mark@pogo.org.uk>]

I found that splice() can operate from a pipe to a unix socket, but from a 
unix socket to a pipe fails with EINVAL.

On investigation, splicing from a pipe to a unix socket (af_unix.c) gets 
routed through a regular unix_stream_sendmsg (via sock_no_sendpage). If my 
understanding is correct this means it will copy memory and not 
references, so it's not actually a splice.

If so, splice() behaves inconsistently; sometimes it fails on unsupported 
file descriptors, sometimes it falls back to a copy. The man page favours 
the former, but the latter could also be useful -- to call splice() with 
any pair of fds and have the kernel do the best it can.

Splicing between unix sockets and pipes would give a useful performance 
benefit. Is there any work in progress on this? Are there any caveats to 
implementing it?

Thanks,

-- 
Mark