Re: PowerPC radeon KMS - is it possible?

["Gerhard Pircher" <gerhard_pircher@gmx.net>]

-------- Original-Nachricht --------
> Datum: Mon, 23 Apr 2012 11:56:06 +0200
> Von: "Michel Dänzer" <michel@daenzer.net>
> An: Gerhard Pircher <gerhard_pircher@gmx.net>
> CC: linuxppc-dev@lists.ozlabs.org
> Betreff: Re: PowerPC radeon KMS - is it possible?

> On Fre, 2012-04-20 at 18:14 +0200, Gerhard Pircher wrote: 
> > > Von: "Michel Dänzer" <michel@daenzer.net>
> > > On Fre, 2012-04-20 at 13:15 +0200, Gerhard Pircher wrote: 
> > > > > Von: "Michel Dänzer" <michel@daenzer.net>
> > > > > On Don, 2012-04-19 at 13:48 +0200, Gerhard Pircher wrote: 
> > > > > > 
> > > > > > The "former case" is an explanation, why I see data corruption
> > > > > > with my< AGPGART driver (more or less a copy of the uninorth
> > > > > > driver) on my non-coherent platform. There are no cache
> > > > > > flushes done for writes to already mapped pages.
> > > > > 
> > > > > As I said, the radeon driver always maps AGP memory uncacheable
> > > > > for the CPU, so no such CPU cache flushes should be necessary.
> > > > I know. We also discussed this topic over two years ago. :-)
> > > > 
> > > > What I didn't understand yet is how this uncacheable memory is
> > > > allocated (well, I never took the time to look at this again). The
> > > > functions in ttm_page_alloc.c seem to allocate normal cacheable
> > > > memory and try to set the page flags with set_pages_array_uc(),
> > > > which is more or less a no-op on powerpc. ttm_page_alloc_dma.c on
> > > > the other side is only used with SWIOTLB!?
> > > [...] 
> > > > Could it be that the memory is finally mapped uncacheable by
> > > radeon_bo_kmap()/
> > > > ttm_bo_kmap()/..some other TTM functions../vmap()?
> > > 
> > > Yeah, AFAICT, basically ttm_io_prot() defines the mapping
> > > attributes, and vmap() is used to enforce them for kernel mappings.
> > Okay, that sounds like the approach used by arch/powerpc/mm/dma-
> > noncoherent.c in my ("green") ears. What about the PCIGART mode?
> > Is the driver free to use cached memory in this mode?
> 
> Yes, it assumes PCI(e) GART to be CPU cache coherent.
Okay. I guess it should be possible to modify it so that it makes use
of uncacheable memory - just for testing!?
PCIGART was working "somehow" on my platform up to the ~2.6.39 kernel,
i.e. I could login to GNOME and open a program until the machine
locked-up. :-)

> > > > [    5.878809] [drm:radeon_test_moves] *ERROR* Incorrect VRAM->GTT copy 0: Got 0xf1416ec0, expected 0xf1570ec0 (VRAM map 0xf1480000-0xf1580000)
> > > [...]
> > > > The VRAM->GTT copy totally puzzles me, as it returns a wrong page
> > > > address, but the offset is fine!?
> > > 
> > > Maybe it's still the values from the GTT->VRAM test, i.e. either
> > > the GPU writes didn't make it to the memory mapped into the AGP
That's indeed the case. I changed the code so that gtt_map is
reinitialized with some magic value before the VRAM->GTT copy and the
debug output shows that the GPU writes don't make it to the memory -
or they are going to the wrong memory location, but that's harder to
track.

> > > GART (some AGP bridges are known to have issues with that) or the
> > > CPU doesn't see it.
> > What is the workaround for such an AGP bridge? If there is one at
> > all...
> 
> The only workaround short of not using AGP would be not doing GPU->AGP
> transfers, but that's not implemented or even planned at all.
Okay.

> > > BTW, does your driver set cant_use_aperture, or is the linear
> > > aperture accessible by the CPU?
> > The driver sets cant_use_aperture. I couldn't get it working at all
> > without it.
> 
> Does the hardware really not allow the CPU to access the linear aperture
> though? Because if it does, I wonder if that might be more reliable.
I'm afraid no, but I could try it out again.

BTW: I see that the uninorth driver defines needs_scratch_page. What
is this actually good for?

Thanks!
Gerhard
-- 
Empfehlen Sie GMX DSL Ihren Freunden und Bekannten und wir
belohnen Sie mit bis zu 50,- Euro! https://freundschaftswerbung.gmx.de