[RFC PATCH] drm: disable WC optimization for cache coherent devices on non-x86
Ard Biesheuvel
ard.biesheuvel at linaro.org
Mon Jan 21 18:28:30 UTC 2019
On Mon, 21 Jan 2019 at 19:24, Michel Dänzer <michel at daenzer.net> wrote:
>
> On 2019-01-21 7:20 p.m., Ard Biesheuvel wrote:
> > On Mon, 21 Jan 2019 at 19:04, Michel Dänzer <michel at daenzer.net> wrote:
> >>
> >> On 2019-01-21 6:59 p.m., Ard Biesheuvel wrote:
> >>> On Mon, 21 Jan 2019 at 18:55, Michel Dänzer <michel at daenzer.net> wrote:
> >>>>
> >>>> On 2019-01-21 5:30 p.m., Ard Biesheuvel wrote:
> >>>>> On Mon, 21 Jan 2019 at 17:22, Christoph Hellwig <hch at infradead.org> wrote:
> >>>>>
> >>>>>> Until that happens we should just change the driver ifdefs to default
> >>>>>> the hacks to off and only enable them on setups where we 100%
> >>>>>> positively know that they actually work. And document that fact
> >>>>>> in big fat comments.
> >>>>>
> >>>>> Well, as I mentioned in my commit log as well, if we default to off
> >>>>> unless CONFIG_X86, we may break working setups on MIPS and Power where
> >>>>> the device is in fact non-cache coherent, and relies on this
> >>>>> 'optimization' to get things working.
> >>>>
> >>>> FWIW, the amdgpu driver doesn't rely on non-snooped transfers for
> >>>> correct basic operation (the scenario Christian brought up is a very
> >>>> specialized use-case), so that shouldn't be an issue.
> >>>>
> >>>
> >>> The point is that this is only true for x86.
> >>>
> >>> On other architectures, the use of non-cached mappings on the CPU side
> >>> means that you /do/ rely on non-snooped transfers, since if those
> >>> transfers turn out not to snoop inadvertently, the accesses are
> >>> incoherent with the CPU's view of memory.
> >>
> >> The driver generally only uses non-cached mappings if
> >> drm_arch/device_can_wc_memory returns true.
> >>
> >
> > Indeed. And so we should take care to only return 'true' from that
> > function if it is guaranteed that non-cached CPU mappings are coherent
> > with the mappings used by the GPU, either because that is always the
> > case (like on x86), or because we know that the platform in question
> > implements NoSnoop correctly throughout the interconnect.
> >
> > What seems to be complicating matters is that in some cases, the
> > device is non-cache coherent to begin with, so regardless of whether
> > the NoSnoop attribute is used or not, those accesses will not snoop in
> > the caches and be coherent with the non-cached mappings used by the
> > CPU. So if we restrict this optimization [on non-X86] to platforms
> > that are known to implement NoSnoop correctly, we may break platforms
> > that are implicitly NoSnoop all the time.
>
> Since the driver generally doesn't rely on non-snooped accesses for
> correctness, that couldn't "break" anything that hasn't always been broken.
>
Again, that is only true on x86.
On other architectures, DMA writes from the device may allocate in the
caches, and be invisible to the CPU when it uses non-cached mappings.
More information about the amd-gfx
mailing list