On Wed, Dec 21, 2016 at 10:06:36AM -0800, Stephen Hemminger wrote:
> On Wed, 21 Dec 2016 17:45:13 +0800
> Yuanhan Liu <yuanhan....@linux.intel.com> wrote:
>
> > From: Jan Wickbom <jan.wick...@ericsson.com>
> >
> > Currently select() is used to monitor file descriptors for vhostuser
> > ports. This limits the number of ports possible to create since the
> > fd number is used as index in the fd_set and we have seen fds > 1023.
> > This patch changes select() to poll(). This way we can keep an
> > packed (pollfd) array for the fds, e.g. as many fds as the size of
> > the array.
> >
> > Also see:
> > http://dpdk.org/ml/archives/dev/2016-April/037024.html
> >
> > Reported-by: Patrik Andersson <patrik.r.anders...@ericsson.com>
> > Signed-off-by: Jan Wickbom <jan.wick...@ericsson.com>
> > Signed-off-by: Yuanhan Liu <yuanhan....@linux.intel.com>
>
> Why not epoll()? It scales much better. The old System V poll
> is just as bad with 1023 fd's.
Indeed, there was a plan to use epoll() to fix this issue. It's been
delayed long enough that Jan came up with a fix with poll recently.
I don't have strong preference to epoll() over poll, for following
reasons:
- epoll() is Linux only, though I don't know other platforms (mainly
*BSD) actually care vhost-user or not.
- epoll indeed scales much better, but I don't think the scale actually
matters a lot here: vhost-user negotiation normally happens once on
QEMU startup, and it's a short lived process. Both facts make it be
non-performance critical.
- 1023 is also not a big problem so far: vhost lib hardcodes the max
vhost devices we could support to be 1024 after all.
For above reasons and Jan already came up with a solution (that would
work for most platforms), I think I'm fine with this poll() so far.
--yliu
