That's as designed (and epoll behaves that way, too, doesn't it?);
when applications start using sigio or epoll on a socket, they
have to assume it's readable/writable initially. There was a huge
argu^H^H^H^H thread about that recently with subject
"epoll (was Re: [PATCH] async poll for 2.5)"
> It doesn't change when I read the HTTP header.
> So I added a kludge that calls poll() when the application wants to
> switch from reading to writing or vice versa. That is quite ugly but it
> works.
Sounds fishy... but I'd need to see your code to know more.
> The second problem is that once I start hammering the server with
> request (as opposed to running wget manually from the command line), the
> server just stops serving requests. strace shows this pattern:
>
> sigtimedwait signals an event on fd #3 (the listening socket)
> accept is called, returns #4
> fd 4 is set non-blocking
> F_SETOWN, F_SETSIG, SETFL O_ASYNC
> sigtimedwait times out.
> sigtimedwait is called again, times out again.
>
> Why is that? Googling seemed to indicate that there could be a race
> condition here after the accept. Should I be running poll on the socket
> right away? Or just blindly call the read handler?
Yes. Blindly call the handler, and do I/O until you get an
EWOULDBLOCK. That's precisely what you're supposed to do when
you start using sigio on an fd.
> Is anyone actually successfully using sigio for anything? So far it
> does not look very reliable to me.
It works ok, once you get used to its oddities.
sys_epoll works better, but sigio isn't terrible as a fallback.
- Dan
-- Dan Kegel Linux User #78045 http://www.kegel.com- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/