Re: [Qemu-devel] [PATCH 2/2] qxl: add QXL_IO_UPDATE_MEM for guest S3&S4 support

[Alon Levy]

On Wed, Jun 29, 2011 at 12:25:00PM +0200, Gerd Hoffmann wrote:
> On 06/29/11 11:21, Alon Levy wrote:
> >On Wed, Jun 29, 2011 at 11:01:11AM +0200, Gerd Hoffmann wrote:
> >>   Hi,
> >>
> >>>>I think it will receive them after migration, since the command ring
> >>>>was stored.
> >>>Our confusion here is because you think there is still seemless migration. 
> >>>Unfortunately
> >>>it doesn't work right now, unless you plan to fix it the only form of 
> >>>migration right
> >>>now is switch-host, and for that those commands will be lost, the new 
> >>>connection will receive
> >>>images for each surface. If you treat the client as seemless you are 
> >>>completely right.
> >>
> >>The spice server needs this too so it can render the surfaces
> >>correctly before sending the surface images to the client (or send
> >>the old surfaces and the commands on top of that).
> >>
> >>That is one difference between qemu migration and S3 state: For qemu
> >>migration it is no problem to have unprocessed commands in the
> >>rings, they will simply be processed once the spice server state is
> >>restored. When the guest driver restores the state when it comes
> >>back from S3 it needs the command rings to do so, thats why they
> >>must be flushed before entering S3 ...
> >
> >You mean it needs the command rings to be empty before, since they are lost
> >during the reset, right?
> 
> One more reason.  Wasn't aware there is a reset anyway, was thinking
hah. The reset is the whole mess.. otherwise S3 would have been trivial,
and actually disabling the reset was the first thing we did, but of course
it doesn't solve S4 in that case.

> more about the command ordering.  Without reset spice-server would
> first process the old commands (which may reference non-existing
> surfaces), then the new commands which re-recreate all state, which
> is simply the wrong order.  With reset the old commands just get
> lost which causes rendering bugs.
> 
> Is it an option to have the driver just remove the commands from the
> ring (and resubmit after resume)?  I suspect it isn't as there is no
> race-free way to do that, right?
Right - the whole ring assumes that the same side removes. of course we
can add an IO for that (two - FREEZE and UNFREEZE). But I think this is
the wrong approach. Instead, rendering all the commands, and dropping the
wait for the client. Right now if we flush we do actually wait for the client,
but I plan to remove this later. (we do this right now for update_area as
well and that's much higher frequency).

> 
> cheers,
>   Gerd
> 
>