Re: [RFC 10/20] iommu/iommufd: Add IOMMU_DEVICE_GET_INFO

[Jean-Philippe Brucker]

On Thu, Sep 23, 2021 at 03:10:47AM +0000, Tian, Kevin wrote:
> > From: Jason Gunthorpe <jgg@xxxxxxxxxx>
> > Sent: Thursday, September 23, 2021 7:50 AM
> > 
> > On Wed, Sep 22, 2021 at 03:24:07PM -0600, Alex Williamson wrote:
> > > On Sun, 19 Sep 2021 14:38:38 +0800
> > > Liu Yi L <yi.l.liu@xxxxxxxxx> wrote:
> > >
> > > > +struct iommu_device_info {
> > > > +	__u32	argsz;
> > > > +	__u32	flags;
> > > > +#define IOMMU_DEVICE_INFO_ENFORCE_SNOOP	(1 << 0) /* IOMMU
> > enforced snoop */
> > >
> > > Is this too PCI specific, or perhaps too much of the mechanism rather
> 
> Isn't snoop vs. !snoop a general concept not pci specific?
> 
> > > than the result?  ie. should we just indicate if the IOMMU guarantees
> > > coherent DMA?  Thanks,
> > 
> > I think the name of "coherent DMA" for this feature inside the kernel
> > is very, very confusing. We already have something called coherent dma
> > and this usage on Intel has nothing at all to do with that.
> > 
> > In fact it looks like this confusing name has already caused
> > implementation problems as I see dma-iommu, is connecting
> > dev->dma_coherent to IOMMU_CACHE! eg in dma_info_to_prot(). This is
> > completely wrong if IOMMU_CACHE is linked to no_snoop.
> > 
> > And ARM seems to have fallen out of step with x86 as the ARM IOMMU
> > drivers are mapping IOMMU_CACHE to ARM_LPAE_PTE_MEMATTR_OIWB,
> > ARM_LPAE_MAIR_ATTR_IDX_CACHE
> > 
> > The SMMU spec for ARMv8 is pretty clear:
> > 
> >  13.6.1.1 No_snoop
> > 
> >  Support for No_snoop is system-dependent and, if implemented, No_snoop
> >  transforms a final access attribute of a Normal cacheable type to
> >  Normal-iNC-oNC-OSH downstream of (or appearing to be performed
> >  downstream of) the SMMU. No_snoop does not transform a final access
> >  attribute of any-Device.
> > 
> > Meaning setting ARM_LPAE_MAIR_ATTR_IDX_CACHE from IOMMU_CACHE
> > does NOT
> > block non-snoop, in fact it *enables* it - the reverse of what Intel
> > is doing!
> 
> Checking the code:
> 
>         if (data->iop.fmt == ARM_64_LPAE_S2 ||
>             data->iop.fmt == ARM_32_LPAE_S2) {
>                 if (prot & IOMMU_MMIO)
>                         pte |= ARM_LPAE_PTE_MEMATTR_DEV;
>                 else if (prot & IOMMU_CACHE)
>                         pte |= ARM_LPAE_PTE_MEMATTR_OIWB;
>                 else
>                         pte |= ARM_LPAE_PTE_MEMATTR_NC;
> 
> It does set attribute to WB for IOMMU_CACHE and then NC (Non-cacheable)
> for !IOMMU_CACHE. The main difference between Intel and ARM is that Intel
> by default allows both snoop and non-snoop traffic with one additional bit
> to enforce snoop, while ARM requires explicit SMMU configuration for snoop
> and non-snoop respectively.
> 
>         } else {
>                 if (prot & IOMMU_MMIO)
>                         pte |= (ARM_LPAE_MAIR_ATTR_IDX_DEV
>                                 << ARM_LPAE_PTE_ATTRINDX_SHIFT);
>                 else if (prot & IOMMU_CACHE)
>                         pte |= (ARM_LPAE_MAIR_ATTR_IDX_CACHE
>                                 << ARM_LPAE_PTE_ATTRINDX_SHIFT);
>         }
> 
> same for this one. MAIR_ELx register is programmed to ARM_LPAE_MAIR_
> ATTR_WBRWA for IDX_CACHE bit. I'm not sure why it doesn't use 
> IDX_NC though, when !IOMMU_CACHE.

It is in effect since IDX_NC == 0

> 
> > 
> > So this is all a mess.
> > 
> > Better to start clear and unambiguous names in the uAPI and someone
> > can try to clean up the kernel eventually.
> > 
> > The required behavior for iommufd is to have the IOMMU ignore the
> > no-snoop bit so that Intel HW can disable wbinvd. This bit should be
> > clearly documented for its exact purpose and if other arches also have
> > instructions that need to be disabled if snoop TLPs are allowed then
> > they can re-use this bit. It appears ARM does not have this issue and
> > does not need the bit.
> 
> Disabling wbinvd is one purpose. imo the more important intention
> is that iommu vendor uses different PTE formats between snoop and
> !snoop. As long as we want allow userspace to opt in case of isoch 
> performance requirement (unlike current vfio which always choose
> snoop format if available), such mechanism is required for all vendors.
> 
> When creating an ioas there could be three snoop modes:
> 
> 1) snoop for all attached devices;
> 2) non-snoop for all attached devices;
> 3) device-selected snoop;
> 
> Intel supports 1) <enforce-snoop on> and 3) <enforce-snoop off>. snoop
> and nonsnoop devices can be attached to a same ioas in 3).
> 
> ARM supports 1) <snoop format> and 2) <nonsnoop format>. snoop devices
> and nonsnoop devices must be attached to different ioas's in 1) and 2)
> respectively.

I think Arm mainly supports 3), ie. No_snoop PCI transactions on pages
mapped cacheable become non-cacheable memory accesses.

But the Arm Base System Architecture 1.0
(https://developer.arm.com/documentation/den0094/a) states that it's
implementation dependent whether the system supports No_snoop.

    In the case where the system has a System MMU translating and
    attributing the transactions from the root complex, the PCI Express
    transactions must keep the memory attributes assigned by the System
    MMU. If the System MMU-assigned attribute is cacheable then it is
    IMPLEMENTATION DEFINED if No_snoop transactions replace the attribute
    with non-cached.

So we can only tell userspace "No_snoop is not supported" (provided we
even want to allow them to enable No_snoop). Users in control of stage-1
tables can create non-cacheable mappings through MAIR attributes.

Thanks,
Jean

> 
> Then the device info should reports:
> 
> /* iommu enforced snoop */
> +#define IOMMU_DEVICE_INFO_ENFORCE_SNOOP	(1 << 0)
> /* iommu enforced nonsnoop */
> +#define IOMMU_DEVICE_INFO_ENFORCE_NONSNOOP	(1 << 1)
> /* device selected snoop */
> +#define IOMMU_DEVICE_INFO_DEVICE_SNOOP	(1 << 2)
> 
> > 
> > What ARM is doing with IOMMU_CACHE is unclear to me, and I'm unclear
> > if/how iommufd should expose it as a controllable PTE flag. The ARM
> > 
> 
> Based on above analysis I think the ARM usage with IOMMU_CACHE
> doesn't change. 
> 
> Thanks
> Kevin