[PATCH v14 0/7] drm/i915/gvt: Dma-buf support for GVT-g

Alex Williamson alex.williamson at redhat.com
Thu Sep 21 16:08:38 UTC 2017


On Wed, 20 Sep 2017 12:54:36 +0200
Gerd Hoffmann <kraxel at redhat.com> wrote:

>   Hi,
> 
> > Which I am curious is that in the case (3) above, although the same
> > fd number is returned,
> > it should be referenced with different dmabuf (in another word #2,
> > not the #1), as it has been installed
> > to a new dmabuf_obj.  
> 
> Try applying this on top of v14:
> https://www.kraxel.org/cgit/linux/commit/?h=gvt-dmabuf-v14&id=1ecdf2d41
> b1ebac0c778e2cb87f82dbdf7080aa8
> 
> > > Try https://www.kraxel.org/cgit/qemu/log/?h=work/intel-vgpu with
> > > gtk ui (-
> > > display gtk,gl=on).  
> > 
> > I met an issue when trying this branch, which doesn't look like the
> > issue we are discussing:
> > Qemu-system-x86_64: Couldn't find current GLX or EGL context.  
> 
> Yep, something with the opengl setup.
> It easily reproduces with your sdl-based branch too.
> Just apply the patch above and watch the kernel log.
> 
> kraxel at broadwell ~# dmesg | grep "fd 60"
> [  213.090899] intel_vgpu_query_plane: new dmabuf, fd 60, obj
> ffff93e6d0fcd280, 256x256, cursor
> [  215.250901] intel_vgpu_query_plane: new dmabuf, fd 60, obj
> ffff93e6d09a3540, 256x256, cursor
> [  220.321928] intel_vgpu_query_plane: new dmabuf, fd 60, obj 
> ffff93e6d09a3d40, 256x256, cursor
>     [ note: closing qemu window here ]
> [  252.620172] intel_vgpu_gem_release: free dmabuf, fd 60, obj
> ffff93e6d09a3d40
> [  252.620191] intel_vgpu_gem_release: free dmabuf, fd 60, obj
> ffff93e6d0fcd280
> [  252.620200] intel_vgpu_gem_release: free dmabuf, fd 60, obj
> ffff93e6d09a3540
> 
> > Can we discuss in a IRC channel to quick sync our issues?  
> 
> #gvtg on freenode.

This seems like a really important discussion on the API design, but
we've dropped the rest of the world from the Cc list.  Adding back
intel-gvt-dev and Kirti, though this seems entirely focused on the fd
version of the API, trying to match the dma-buf lifecycle to the
reference we use for it, where the fd isn't actually a good fit.  Would
anyone object to me attaching the entire discussion thread from this
week to an email here so it's not private?  Thanks,

Alex


More information about the intel-gvt-dev mailing list