[PATCH] [RFC] etnaviv: Don't apply pgprot_noncached for ETNA_BO_UNCACHED

[Lucas Stach]

Hi Guido,

Am Freitag, den 14.09.2018, 14:10 +0200 schrieb Guido Günther:
> Hi,
> On Fri, Sep 14, 2018 at 11:27:08AM +0200, Lucas Stach wrote:
> > Hi Guido,
> > 
> > Am Donnerstag, den 13.09.2018, 20:00 +0200 schrieb Guido Günther:
> > > Hi,
> > > On arm64 userspace does not seem to be able to mmap memory marked as
> > > pgprot_noncached(nGnRnE) or pgprot_device(nGnRE).
> > 
> > What's the issue with a normal uncached mapping on ARM64? Is the
> > mapping not set up at all, or do you get aborts or something?
> 
> Userspace gets SIGBUS during memset like below:
> 
>   bo = etna_bo_new(dev, 0x100, flags);
>   buf = etna_bo_map(bo);
>   etna_bo_cpu_prep(bo, DRM_ETNA_PREP_WRITE);
>   memset(buf, 0, 0x100);

That is kind of expected with uncached being translated to a nG
mapping. memset as implemented by glibc on ARM64 uses some tricks with
unaligned loads/stores to avoid control flow, which depend on being
able to merge overlapping requests in a gather buffer.

Having the uncached mapping as nGnR allows for usages such as signal
pages, so I would like to avoid applying this patch if possible.
Instead I would rather make the etnaviv UAPI more strict with regard to
the memory attributes by demanding that userspace only uses UNCACHED
for usages where the access alignment can be tightly controlled and
must use WC or CACHED for any other usages. I'm not aware of any
existing userspace (aside from the WIP GC7000 Mesa support) that would
break due to such a change in the UAPI contract.

Regards,
Lucas