[Spice-devel] synchronous io considered harmful

[Alon Levy]

On Thu, Jun 16, 2011 at 09:21:13PM +0200, Alon Levy wrote:
> Hi,
> 
>  I just had a short conversation with Marc-Andre about a bug in spice/qxl and it turns out there are a couple of possible solutions and I require some more input.
> 

Forgot: this is the bug: https://bugzilla.redhat.com/show_bug.cgi?id=700134

>  The problem: spice was stalling in flush_display_commands caused by QXL_UPDATE_AREA. Since the vcpu thread is holding the global qemu mutex then qemu is unresponsive during this time to any other clients, specifically a libvirt call via virt-manager (virDomainGetXMLDesc) was blocked.
>  
>  There are a number of wrong things here, I'll start from what I think we should fix but any of these would solve this specific problem:
>  1. QXL_IO_UPDATE_AREA is synchronous. If we made it async, by using an interrupt to notify of it's completion, then the io would complete practicaly immediately, and the mutex would be relinquished.
>  2. flush_display_commands is waiting for the client if the PIPE is too large. This is wrong - why should a slow client prevent the guest from doing an update area? Presumably we do this to keep the pipe length limited, which presumably limits the amount of memory. I don't have any numbers, but this looks wrong to me.
>  3. we don't drop qemu's global mutex when dispatching a message. Since an IO can happen from any vcpu we would need to add our own dispatcher mutex in spice.
> 
> Any comments are appreciated,
> 
> Alon
> _______________________________________________
> Spice-devel mailing list
> Spice-devel at lists.freedesktop.org
> http://lists.freedesktop.org/mailman/listinfo/spice-devel