Re: [PATCH] iommu/io-pgtable-arm: Allow non-coherent masters to use system cache

[isaacm@codeaurora.org]

On 2021-01-08 10:18, Will Deacon wrote:
> On Fri, Jan 08, 2021 at 11:17:25AM +0530, Sai Prakash Ranjan wrote:
>> On 2021-01-07 22:27, isaacm@codeaurora.org wrote:
>> > On 2021-01-06 03:56, Will Deacon wrote:
>> > > On Thu, Dec 24, 2020 at 12:10:07PM +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 mark the memory as normal sys-cached for GPU to use
>> > > > system cache.
>> > > > Without this, the system cache lines are not allocated for GPU.
>> > > > We use
>> > > > the IO_PGTABLE_QUIRK_ARM_OUTER_WBWA quirk instead of a page
>> > > > protection
>> > > > flag as the flag cannot be exposed via DMA api because of no in-tree
>> > > > users.
>> > > >
>> > > > Signed-off-by: Sai Prakash Ranjan <saiprakash.ranjan@codeaurora.org>
>> > > > ---
>> > > >  drivers/iommu/io-pgtable-arm.c | 3 +++
>> > > >  1 file changed, 3 insertions(+)
>> > > >
>> > > > diff --git a/drivers/iommu/io-pgtable-arm.c
>> > > > b/drivers/iommu/io-pgtable-arm.c
>> > > > index 7c9ea9d7874a..3fb7de8304a2 100644
>> > > > --- a/drivers/iommu/io-pgtable-arm.c
>> > > > +++ b/drivers/iommu/io-pgtable-arm.c
>> > > > @@ -415,6 +415,9 @@ static arm_lpae_iopte
>> > > > arm_lpae_prot_to_pte(struct arm_lpae_io_pgtable *data,
>> > > >  		else if (prot & IOMMU_CACHE)
>> > > >  			pte |= (ARM_LPAE_MAIR_ATTR_IDX_CACHE
>> > > >  				<< ARM_LPAE_PTE_ATTRINDX_SHIFT);
>> > > > +		else if (data->iop.cfg.quirks & IO_PGTABLE_QUIRK_ARM_OUTER_WBWA)
>> > > > +			pte |= (ARM_LPAE_MAIR_ATTR_IDX_INC_OCACHE
>> > > > +				<< ARM_LPAE_PTE_ATTRINDX_SHIFT);
>> > > >  	}
>> > >
>> > While this approach of enabling system cache globally for both page
>> > tables and other buffers
>> > works for the GPU usecase, this isn't ideal for other clients that use
>> > system cache. For example,
>> > video clients only want to cache a subset of their buffers in the
>> > system cache, due to the sizing constraint
>> > imposed by how much of the system cache they can use. So, it would be
>> > ideal to have
>> > a way of expressing the desire to use the system cache on a per-buffer
>> > basis. Additionally,
>> > our video clients use the DMA layer, and since the requirement is for
>> > caching in the system cache
>> > to be a per buffer attribute, it seems like we would have to have a
>> > DMA attribute to express
>> > this on a per-buffer basis.
>> >
>> 
>> I did bring this up initially [1], also where is this video client
>> in upstream? AFAIK, only system cache user in upstream is GPU.
>> We cannot add any DMA attribute unless there is any user upstream
>> as per [2], so when the support for such a client is added, wouldn't
>> ((data->iop.cfg.quirks & IO_PGTABLE_QUIRK_ARM_OUTER_WBWA) || 
>> PROT_FLAG)
>> work?
> 
> Hmm, I think this is another case where we need to separate out the
> page-table walker attributes from the access attributes. Currently,
> IO_PGTABLE_QUIRK_ARM_OUTER_WBWA applies _only_ to the page-table walker
> and I don't think it makes any sense for that to be per-buffer (how 
> would
> you even manage that?). However, if we want to extend this to data 
> accesses
> and we know that there are valid use-cases where this should be 
> per-buffer,
> then shoe-horning it in with the walker quirk does not feel like the 
> best
> thing to do.
Right, I agree that this seems something that merits the same level of 
separation
that exists for the page table walker attributes with respect to 
coherency, and
data buffer attributes with respect to coherency (i.e page table walker 
coherency
does not imply data buffer coherency--that is driven through 
IOMMU_CACHE).
> 
> As a starting point, we could:
> 
>   1. Rename IO_PGTABLE_QUIRK_ARM_OUTER_WBWA to IO_PGTABLE_QUIRK_PTW_LLC
>   2. Add a new prot flag IOMMU_LLC
>   3. Have the GPU pass the new prot for its buffer mappings
> 
> Does that work? One thing I'm not sure about is whether IOMMU_CACHE 
> should
Yes, that should work, as that'll leave the door open for there to be a 
DMA attribute
that can be wired up to IOMMU_LLC.
> imply IOMMU_LLC, or whether there is a use-case for inner-cacheable, 
> outer
> non-cacheable mappings for a coherent device. Have you ever seen that 
> sort
I'm not aware of such a usecase, but I believe that a coherent device 
will
have their buffers cached in the system cache anyway, as well as the CPU 
caches.

--Isaac
> of thing before?
> 
> Will
> 
> _______________________________________________
> linux-arm-kernel mailing list
> linux-arm-kernel@lists.infradead.org
> http://lists.infradead.org/mailman/listinfo/linux-arm-kernel