Re: [PATCH 5/5] ioeventfd: Introduce KVM_IOEVENTFD_FLAG_SOCKET

["Michael S. Tsirkin" <mst@redhat.com>]

On Sun, Jul 10, 2011 at 08:34:43AM +0300, Sasha Levin wrote:
> On Wed, 2011-07-06 at 20:58 +0300, Michael S. Tsirkin wrote:
> > On Wed, Jul 06, 2011 at 06:01:46PM +0300, Sasha Levin wrote:
> > > On Wed, 2011-07-06 at 14:42 +0300, Michael S. Tsirkin wrote:
> > > > On Wed, Jul 06, 2011 at 07:37:58AM +0300, Sasha Levin wrote:
> > > > > +	if (p->sock) {
> > > > > +		socket_write(p->sock, &data, sizeof(data));
> > > > > +		socket_read(p->sock, &data, sizeof(data));
> > > > > +		set_val(val, len, data.data);
> > > > 
> > > > Same here.
> > > 
> > > The socket_read() here I should leave blocking, and spin on it until I
> > > read something - right?
> > 
> > I think it's best to exit to userspace.
> 
> Since sock_recvmsg for AF_UNIX SEQPACKET is interruptible, if we fail
> the read here we'll take a regular MMIO exit and will allow the usermode
> to deal with the MMIO in a regular way.
> 
> I've discussed the issue of usermode might having to handle the same
> MMIO read request twice with Michael, and the solution proposed was to
> add a new type of exit to handle this special case.
> 
> After working on that solution a bit I saw it's adding a lot of code and
> complexity for this small issue, and I'm now thinking we may be better
> off with just handling reads twice in case of a signal just between
> socket_write() and socket_read() - once through the socket and once
> through a regular MMIO exit.

The problem with this approach would be reads with a side-effect though:
for example, the IRQ status read in virtio clears the status register.

I don't insist on a new type of exit, just pointing out the problem.


> -- 
> 
> Sasha.