Re: [PATCH 1/1] iommu/arm-smmu: Add support to use Last level cache

[Will Deacon]

Hi Vivek,

On Thu, Sep 20, 2018 at 05:11:53PM +0530, Vivek Gautam wrote:
> On Wed, Jun 27, 2018 at 10:07 PM Will Deacon <will.deacon@xxxxxxx> wrote:
> > On Tue, Jun 19, 2018 at 02:04:44PM +0530, Vivek Gautam wrote:
> > > On Fri, Jun 15, 2018 at 10:22 PM, Will Deacon <will.deacon@xxxxxxx> wrote:
> > > > On Fri, Jun 15, 2018 at 04:23:29PM +0530, Vivek Gautam wrote:
> > > >> Qualcomm SoCs have an additional level of cache called as
> > > >> System cache or Last level cache[1]. This cache sits right
> > > >> before the DDR, and is tightly coupled with the memory
> > > >> controller.
> > > >> The cache is available to all the clients present in the
> > > >> SoC system. The clients request their slices from this system
> > > >> cache, make it active, and can then start using it. For these
> > > >> clients with smmu, to start using the system cache for
> > > >> dma buffers and related page tables [2], few of the memory
> > > >> attributes need to be set accordingly.
> > > >> This change makes the related memory Outer-Shareable, and
> > > >> updates the MAIR with necessary protection.
> > > >>
> > > >> The MAIR attribute requirements are:
> > > >>     Inner Cacheablity = 0
> > > >>     Outer Cacheablity = 1, Write-Back Write Allocate
> > > >>     Outer Shareablity = 1
> > > >
> > > > Hmm, so is this cache coherent with the CPU or not?
> > >
> > > Thanks for reviewing.
> > > Yes, this LLC is cache coherent with CPU, so we mark for Outer-cacheable.
> > > The different masters such as GPU as able to allocated and activate a slice
> > > in this Last Level Cache.
> >
> > What I mean is, for example, if the CPU writes some data using Normal, Inner
> > Shareable, Inner/Outer Cacheable, Inner/Outer Write-back, Non-transient
> > Read/Write-allocate and a device reads that data using your MAIR encoding
> > above, is the device guaranteed to see the CPU writes after the CPU has
> > executed a DSB instruction?
> 
> No, these MAIR configurations don't guarantee that devices will have
> coherent view
> of what CPU writes. Not all devices can snoop into CPU caches (only IO-Coherent
> devices can).
> So a normal cached memory configuration in CPU MMU tables, and SMMU page tables
> is valid only for few devices that are IO-coherent.
> 
> Moreover, CPU can lookup in system cache, and so do all devices;
> allocation will depend on h/w configurations and memory attributes.
> So anything that CPU caches in system cache will be coherently visible
> to devices.
> 
> >
> > I don't think so, because the ARM ARM would say that there's a mismatch on
> > the Inner Cacheability attribute.
> >
> > > > Why don't normal
> > > > non-cacheable mappings allocated in the LLC by default?
> > >
> > > Sorry, I couldn't fully understand your question here.
> > > Few of the masters on qcom socs are not io-coherent, so for them
> > > the IC has to be marked as 0.
> >
> > By IC you mean Inner Cacheability? In your MAIR encoding above, it is zero
> > so I don't understand the problem. What goes wrong if non-coherent devices
> > use your MAIR encoding for their DMA buffers?
> >
> > > But they are able to use the LLC with OC marked as 1.
> >
> > The issue here is that whatever attributes we put in the SMMU need to align
> > with the attributes used by the CPU in order to avoid introducing mismatched
> > aliases.
> 
> Not really, right?
> Devices can use Inner non-Cacheable, Outer-cacheable (IC=0, OC=1) to allocate
> into the system cache (as these devices don't want to allocate in
> their inner caches),
> and the CPU will have a coherent view of these buffers/page-tables.
> This should be
> a normal cached non-IO-Coherent memory.
> 
> But anything that CPU writes using Normal, Inner Shareable,
> Inner/Outer Cacheable,
> Inner/Outer Write-back, Non-transient Read/Write-allocate, may not be visible
> to the device.
> 
> Also added Jordan, and Pratik to this thread.

Sorry, but I'm still completely confused.

If you only end up with mismatched memory attributes in the non-coherent
case, then why can't you just follow my suggestion to override the
attributes for non-coherent mappings on your SoC?

Will