Re: [Freedreno] [PATCH 0/3] iommu/drm/msm: Allow non-coherent masters to use system cache

[Rob Clark <robdclark@gmail.com>]

On Mon, Aug 9, 2021 at 11:11 AM Sai Prakash Ranjan
<saiprakash.ranjan@codeaurora.org> wrote:
>
> On 2021-08-09 23:37, Rob Clark wrote:
> > On Mon, Aug 9, 2021 at 10:47 AM Sai Prakash Ranjan
> > <saiprakash.ranjan@codeaurora.org> wrote:
> >>
> >> On 2021-08-09 23:10, Will Deacon wrote:
> >> > On Mon, Aug 09, 2021 at 10:18:21AM -0700, Rob Clark wrote:
> >> >> On Mon, Aug 9, 2021 at 10:05 AM Will Deacon <will@kernel.org> wrote:
> >> >> >
> >> >> > On Mon, Aug 09, 2021 at 09:57:08AM -0700, Rob Clark wrote:
> >> >> > > On Mon, Aug 9, 2021 at 7:56 AM Will Deacon <will@kernel.org> wrote:
> >> >> > > > On Mon, Aug 02, 2021 at 06:36:04PM -0700, Rob Clark wrote:
> >> >> > > > > On Mon, Aug 2, 2021 at 8:14 AM Will Deacon <will@kernel.org> wrote:
> >> >> > > > > > On Mon, Aug 02, 2021 at 08:08:07AM -0700, Rob Clark wrote:
> >> >> > > > > > > On Mon, Aug 2, 2021 at 3:55 AM Will Deacon <will@kernel.org> wrote:
> >> >> > > > > > > > On Thu, Jul 29, 2021 at 10:08:22AM +0530, Sai Prakash Ranjan wrote:
> >> >> > > > > > > > > On 2021-07-28 19:30, Georgi Djakov wrote:
> >> >> > > > > > > > > > On Mon, Jan 11, 2021 at 07:45:02PM +0530, Sai Prakash Ranjan wrote:
> >> >> > > > > > > > > > > commit ecd7274fb4cd ("iommu: Remove unused IOMMU_SYS_CACHE_ONLY flag")
> >> >> > > > > > > > > > > removed unused IOMMU_SYS_CACHE_ONLY prot flag and along with it went
> >> >> > > > > > > > > > > the memory type setting required for the non-coherent masters to use
> >> >> > > > > > > > > > > system cache. Now that system cache support for GPU is added, we will
> >> >> > > > > > > > > > > need to set the right PTE attribute for GPU buffers to be sys cached.
> >> >> > > > > > > > > > > Without this, the system cache lines are not allocated for GPU.
> >> >> > > > > > > > > > >
> >> >> > > > > > > > > > > So the patches in this series introduces a new prot flag IOMMU_LLC,
> >> >> > > > > > > > > > > renames IO_PGTABLE_QUIRK_ARM_OUTER_WBWA to IO_PGTABLE_QUIRK_PTW_LLC
> >> >> > > > > > > > > > > and makes GPU the user of this protection flag.
> >> >> > > > > > > > > >
> >> >> > > > > > > > > > Thank you for the patchset! Are you planning to refresh it, as it does
> >> >> > > > > > > > > > not apply anymore?
> >> >> > > > > > > > > >
> >> >> > > > > > > > >
> >> >> > > > > > > > > I was waiting on Will's reply [1]. If there are no changes needed, then
> >> >> > > > > > > > > I can repost the patch.
> >> >> > > > > > > >
> >> >> > > > > > > > I still think you need to handle the mismatched alias, no? You're adding
> >> >> > > > > > > > a new memory type to the SMMU which doesn't exist on the CPU side. That
> >> >> > > > > > > > can't be right.
> >> >> > > > > > > >
> >> >> > > > > > >
> >> >> > > > > > > Just curious, and maybe this is a dumb question, but what is your
> >> >> > > > > > > concern about mismatched aliases?  I mean the cache hierarchy on the
> >> >> > > > > > > GPU device side (anything beyond the LLC) is pretty different and
> >> >> > > > > > > doesn't really care about the smmu pgtable attributes..
> >> >> > > > > >
> >> >> > > > > > If the CPU accesses a shared buffer with different attributes to those which
> >> >> > > > > > the device is using then you fall into the "mismatched memory attributes"
> >> >> > > > > > part of the Arm architecture. It's reasonably unforgiving (you should go and
> >> >> > > > > > read it) and in some cases can apply to speculative accesses as well, but
> >> >> > > > > > the end result is typically loss of coherency.
> >> >> > > > >
> >> >> > > > > Ok, I might have a few other sections to read first to decipher the
> >> >> > > > > terminology..
> >> >> > > > >
> >> >> > > > > But my understanding of LLC is that it looks just like system memory
> >> >> > > > > to the CPU and GPU (I think that would make it "the point of
> >> >> > > > > coherence" between the GPU and CPU?)  If that is true, shouldn't it be
> >> >> > > > > invisible from the point of view of different CPU mapping options?
> >> >> > > >
> >> >> > > > You could certainly build a system where mismatched attributes don't cause
> >> >> > > > loss of coherence, but as it's not guaranteed by the architecture and the
> >> >> > > > changes proposed here affect APIs which are exposed across SoCs, then I
> >> >> > > > don't think it helps much.
> >> >> > > >
> >> >> > >
> >> >> > > Hmm, the description of the new mapping flag is that it applies only
> >> >> > > to transparent outer level cache:
> >> >> > >
> >> >> > > +/*
> >> >> > > + * Non-coherent masters can use this page protection flag to set cacheable
> >> >> > > + * memory attributes for only a transparent outer level of cache, also known as
> >> >> > > + * the last-level or system cache.
> >> >> > > + */
> >> >> > > +#define IOMMU_LLC      (1 << 6)
> >> >> > >
> >> >> > > But I suppose we could call it instead IOMMU_QCOM_LLC or something
> >> >> > > like that to make it more clear that it is not necessarily something
> >> >> > > that would work with a different outer level cache implementation?
> >> >> >
> >> >> > ... or we could just deal with the problem so that other people can reuse
> >> >> > the code. I haven't really understood the reluctance to solve this properly.
> >> >> >
> >> >> > Am I missing some reason this isn't solvable?
> >> >>
> >> >> Oh, was there another way to solve it (other than foregoing setting
> >> >> INC_OCACHE in the pgtables)?  Maybe I misunderstood, is there a
> >> >> corresponding setting on the MMU pgtables side of things?
> >> >
> >> > Right -- we just need to program the CPU's MMU with the matching memory
> >> > attributes! It's a bit more fiddly if you're just using ioremap_wc()
> >> > though, as it's usually the DMA API which handles the attributes under
> >> > the
> >> > hood.
> >> >
> >> > Anyway, sorry, I should've said that explicitly earlier on. We've done
> >> > this
> >> > sort of thing in the Android tree so I assumed Sai knew what needed to
> >> > be
> >> > done and then I didn't think to explain to you :(
> >> >
> >>
> >> Right I was aware of that but even in the android tree there is no
> >> user
> >> :)
> >> I think we can't have a new memory type without any user right in
> >> upstream
> >> like android tree?
> >>
> >> @Rob, I think you  already tried adding a new MT and used
> >> pgprot_syscached()
> >> in GPU driver but it was crashing?
> >
> > Correct, but IIRC there were some differences in the code for memory
> > types compared to the android tree.. I couldn't figure out the
> > necessary patches to cherry-pick to get the android patch to apply
> > cleanly, so I tried re-implementing it without having much of a clue
> > about how that code works (which was probably the issue) ;-)
> >
>
> Hehe no, even I get the same crash after porting/modifying the required
> patches from android ;) and I think crashes would be seen in android as
> well, its just that they don't have any user exercising that code.
>
> Thing is I can't make head and tail of the GPU crash logs, maybe you
> know
> how to decode those errors, if not I can start a thread with QC GPU team
> and ask them to decode?
>

If you have a gpu devcore dump, I can take a look at it with
crashdec.. otherwise I can try to find the branch where I had that
patch backported.

I'm more familiar with using crashdec to figure out mesa bugs, but
maybe I could spot something where what the GPU is seeing disagrees
with what the CPU expects it to be seeing.

BR,
-R