Re: [Qemu-devel] [4341] qemu: generate signals on tap I/O

[Anthony Liguori]

Anders wrote:
> Aurelien Jarno wrote:
>
> > That's why I reverted that commit, it was for my local tree only, I
> > committed it accidentally (along with a few other patches).
>
> Ah, okay. I only read the subject of your reverts, which did not 
> mention 4341.
>
> > > I posted a more complete implementation a few weeks ago, anything wrong
> > > with it?
> > >
> > > http://thread.gmane.org/gmane.comp.emulators.qemu/24567
> >
> > People saying it is not the way to go, and I have still haven't decided
> > myself if it is a correct solution or not.
>
> The current code is also not a correct solution.
>
> Well, I will just keep it in my local tree.

BTW, the biggest problem with tap networking in QEMU seems to be packet 
loss.  The fd_can_read handler is completely ignored for tap which means 
that if the device isn't ready to receive a packet, the packet is simply 
dropped.

I have a patch that fixes this but the only driver that has a proper 
fd_can_read handler right now is virtio-net.  I plan on including this 
tap fix in my next round of virtio patches.

Addressing the packet loss issue improves TCP streaming performance by 
4x.  In contrast, eliminating the copy in the virtio-net device only 
improves performance by about 10% (although it will probably be a lot 
more when tun/tap supports zero-copy).

Regards,

Anthony Liguori

> Cheers,
> Anders.