Re: [PATCH] iommu/arm-smmu-v3-sva: Enable Hardware Access and Hardware Dirty bits

[Pranjal Shrivastava <praan@google.com>]

On Fri, May 08, 2026 at 03:24:32PM +0100, Robin Murphy wrote:
> On 2026-05-08 2:57 pm, Pranjal Shrivastava wrote:
> > On Fri, May 08, 2026 at 02:31:11PM +0100, Robin Murphy wrote:
> > > On 2026-05-08 2:12 pm, Pranjal Shrivastava wrote:
> > > > On Fri, May 08, 2026 at 09:35:50AM -0300, Jason Gunthorpe wrote:
> > > > > On Thu, May 07, 2026 at 10:30:14PM +0000, Pranjal Shrivastava wrote:
> > > > > > > @@ -92,6 +92,16 @@ void arm_smmu_make_sva_cd(struct arm_smmu_cd *target,
> > > > > > >    		target->data[1] = cpu_to_le64(virt_to_phys(mm->pgd) &
> > > > > > >    					      CTXDESC_CD_1_TTB0_MASK);
> > > > > > > +
> > > > > > > +		/*
> > > > > > > +		 * Enable Hardware Access and Dirty updates (DBM) if supported.
> > > > > > > +		 * This is safe to enable by default, as PTE_WRITE and PTE_DBM
> > > > > > > +		 * share the same bit.
> > > > > > > +		 */
> > > > > > > +		if (master->smmu->features & ARM_SMMU_FEAT_HA)
> > > > > > > +			target->data[0] |= cpu_to_le64(CTXDESC_CD_0_TCR_HA);
> > > > > > > +		if (master->smmu->features & ARM_SMMU_FEAT_HD)
> > > > > > > +			target->data[0] |= cpu_to_le64(CTXDESC_CD_0_TCR_HD);
> > > > > > 
> > > > > > IIUC, we should be setting these if IO_PGTABLE_QUIRK_ARM_HD is present?
> > > > > 
> > > > > SVA does not use IO_PGTABLE at all, and it directly constructs its own
> > > > > CD.
> > > > > 
> > > > > No relation between those two flows.
> > > > 
> > > > I understand that but I mean we need to know if the system supports
> > > > HTTU ? Like for SMMU we use the IO_PGTABLE_QUIRK, shouldn't we be
> > > > checking if the CPU's tables support HTTU?
> > > > 
> > > > Are we assuming that if the SMMU IDR presents HTTU capability the MMU
> > > > would also have it? I think an unconditional enablement is risky as we
> > > > may not have system-wide HTTU support.
> > > > 
> > > > If we look at arm_smmu_master_sva_supported, the driver already
> > > > maintains a strict agreement between the CPU and SMMU for SVA.
> > > > It checks sanitized CPU ID registers for things like PARANGE & ASIDBITS,
> > > > and it uses system_supports_bbml2_noabort() to decide whether to enable
> > > > FEAT_BBML2.
> > > > 
> > > > Shouldn't we follow this exact same pattern for HTTU ?
> > > > We should probably be checking cpu_has_hw_af() (from asm/cpufeature.h)
> > > > in the SVA support check or here if we wanna enable HTTU.
> > > 
> > > It might make sense to depend on CONFIG_ARM64_HW_AFDBM - when that is
> > > enabled, then IIRC we already expect to cope with some CPUs not supporting
> > > hardware updates, so it should still be fine for an SMMU to make them even
> > > if no CPU does. However, if it's disabled then I'm not sure if missing
> > > access flag faults (if SMMU HA silently sets them) might be an issue - for
> > > dirty, we'd just never put down the Writeable-Clean permission so enabling
> > > SMMU HD wouldn't do anything anyway.
> > 
> > I see, so IIUC, you mean if IS_ENABLED(CONFIG_ARM64_HW_AFDBM) but CPU
> > doesn't enable HTTU, it is perfectly safe to let the SMMU do HTT updates,
> > Since the fault handlers are already expecting HW-triggered updates?
> > 
> > Which means our check would be something like:
> > 
> >     if (IS_ENABLED(CONFIG_ARM64_HW_AFDBM) {
> >     	if (smmu->features & FEAT_HA)
> > 	 ...
> >     }
> > 
> > instead of cpu_has_hw_af()?
> 
> Hmm, looking closer, cpu_has_hw_af() is the thing which actually influences
> mm behaviour (via arch_has_hw_pte_young and arch_wants_old_prefaulted_pte),
> and that can still be false at runtime if ARM64_HW_AFDBM is enabled but any
> CPU doesn't support HAFDBS, so perhaps you were right the first time :)
> 

Yea, I believe the cpu_has_hw_af() is the right gate.

> Although AFAICS from __cpu_setup(), ARM64_HW_AFDBM will still
> unconditionally enable TCR_EL1.HA on CPUs which do support it, so maybe it
> is OK anyway?
> 

I believe cpu_has_hw_af() is still the safer gate for SVA. While 
individual cores might turn on their local HA support, cpu_has_hw_af()
represents the sanitized system view.

In mismatched systems (where some cores support HAFDBS and others don't),
cpu_has_hw_af() will be false & mm shall default to software-managed AF/
Dirty for consistency across all threads. Enabling HTTU in the SMMU while
the kernel mm is in 'SW-Managed' mode could cause the SMMU to silently 
flip bits that the kernel is expecting to handle via faults, leading to a
mismatch.

Thanks,
Praan