[Intel-gfx] [PATCH] drmmode: unreserve userspace fences

Daniel Vetter daniel at ffwll.ch
Sun Jan 31 11:59:56 CET 2010


Hi Eric,

I've justed noticed that you pushed a patch to fix the kernel. But just
patching the kernel is not enough because new userspaces is still broken
for older kernels. And currently "older kernels" means _any_ released
mainline kernel.

Without this patch I'm experiencing unnecessary uxa fallbacks (due to
failed execbuf ioctls due to fence starvation).

Please reconsider applying.

Thanks, Daniel

On Fri, Jan 15, 2010 at 10:33:52PM +0100, Daniel Vetter wrote:
> On Fri, Jan 15, 2010 at 01:13:25PM -0800, Eric Anholt wrote:
> > On Thu, 14 Jan 2010 10:13:30 +0100, Daniel Vetter <daniel.vetter at ffwll.ch> wrote:
> > > With drmmode, the kernel should control all fence regs. Fence
> > > starved i8xx with tiled glyph caches actually managed to max out the
> > > 5 fences available by default: 1 target, 4 glyph cache bos + one
> > > used by kms modeset for scanout.
> > 
> > Hmm.  KMS support and fence management both came in 2.6.29.  At that
> > point, did the userland code still write fence regs for KMS
> > framebuffers?  It would be nice for KMS to assume the correct number of
> > userland fence regs (0) instead of the old UMS compat number (3),
> > instead of new userland having to explicitly say that it's not broken.
> 
> IIRC, these reserved fences where only used with ums and kernel gem. So
> their not needed with kms. But even when we change the kernel to set to to
> 0 when using kms we need this patch for older kernels. And before
> userspace enabled the tiled glyph cache there was no real need for more
> fences (as evidenced by no one reporting it as a bug). Furthermore
> changing this would strickly speaking break the kernel api, i.e. at least
> some audition of all released drivers is called for. Which is not
> something I'd like to waste time on.
> 
> So I'd say no one's gonna suffer by not changing the kernel, and it's the
> safer path.
> 
> -Daniel
> -- 
> Daniel Vetter
> Mail: daniel at ffwll.ch
> Mobile: +41 (0)79 365 57 48
> 
> 
> -- 
> Daniel Vetter
> Mail: daniel at ffwll.ch
> Mobile: +41 (0)79 365 57 48

-- 
Daniel Vetter
Mail: daniel at ffwll.ch
Mobile: +41 (0)79 365 57 48



More information about the Intel-gfx mailing list