Re: [Qemu-devel] [RFC 7/7] qxl: add allocator

[Alon Levy]

On Mon, Feb 20, 2012 at 02:18:19PM +0100, Gerd Hoffmann wrote:
>   Hi,
> 
> > I'll send a series that works with vnc+spice and sdl+spice (didn't test
> > sdl+vnc+spice), and screendumps at the same time.
> > 
> >>
> >> Setting the QEMU_ALLOCATED_FLAG flag sounds hackish too.
> 
> QEMU_ALLOCATED_FLAG is just a flag which is used by the
> defaultallocator_free_displaysurface function to figure whenever the
> displaysurface memory was allocated by qemu_alloc_display or not.
> 
> I think vga.c shouldn't look at it in the first place, and faking it
> because vga.c looks at it is even worse.
> 
> I'm also not sure it is the right approach to to have qxl register a
> display allocator in the first place.  The only other place doing this
> is sdl, which is a user interface.  None of the gfx card emulations in
> the tree do that ...

TBH I'm having problems without doing it with my own Allocator. In
particular calling ppm_save from spice_server thread seems to be a
problem. I can remove the hackish use of QEMU_ALLOCATED_FLAG by not
checking for it in vga.c in the first place, but I don't know what other
code it affects (well, I do - every machine using vga, but I'm not
testing all of these).

Also reallocating the displaysurface seems to be the wrong thing, and
that's what we do whenever we check in qxl_render_update
is_shared_buffer.

> 
> > Actually I think the right thing is to move/copy the 24bit->32bit convertion
> > from vga.c to pflib.c, what do you think?
> 
> Agree, although that easily gets a patch series of its own when you
> collect optimized format conversion functions to move them over to pflib ...

Yes, I think I'll defer that to later.

> 
> BTW: qxl insisting on a shared displaysurface isn't very clean too, it
> better should be able to fallback to just copying/converting for the
> non-shared case.

I'll try to get that working, although it seems to require having some
timer/bh to do the ppm_save.

> 
> cheers,
>   Gerd
> 
>