Re: [PATCH v11 11/16] iommu/arm-smmu-v3: Add CMDQ_PROD_STOP_FLAG to gate CMDQ submissions

[Pranjal Shrivastava <praan@google.com>]

On Wed, Sep 30, 2026 at 01:33:58PM -0700, Nicolin Chen wrote:
> On Tue, Sep 29, 2026 at 03:45:05AM +0000, Pranjal Shrivastava wrote:
> > @@ -726,13 +727,42 @@ int __arm_smmu_cmdq_issue_cmdlist(struct arm_smmu_device *smmu,
> >  	do {
> >  		u64 old;
> >  
> > +		/*
> > +		 * If the SMMU is suspended/suspending, any new CMDs are elided.
> > +		 * This loop is the Point of Commitment. If we haven't cmpxchg'd
> > +		 * our new indices yet, we can safely bail. Once the indices are
> > +		 * committed, we MUST write valid commands to those slots to
> > +		 * avoid indefinite polling in the drain function.
> > +		 */
> > +		if (Q_STOP(llq.prod)) {
> > +			local_irq_restore(flags);
> > +			return 0;
> > +		}
> 
> Sashiko pointed out this:
> "
> Can this early return cause memory corruption by silently dropping ATC
> invalidations?
> 
> When the SMMU suspends, Q_STOP(llq.prod) becomes true. If a driver or
> background thread then calls dma_unmap() to free a buffer while a PCIe
> endpoint (with ATS enabled) is suspended to a state like D0, the SMMU
> driver will observe the stop flag here in __arm_smmu_cmdq_issue_cmdlist().
> 
> By bailing out and returning 0 (success) without actually submitting the
> CMDQ_OP_ATC_INV command to the hardware, the IOMMU core is misled into
> freeing the memory while the PCIe endpoint's Address Translation Cache
> (ATC) retains the stale translation. When the PCI device resumes, or if
> it issues a TLP while in D0, it could use the stale ATC entry to access
> the now-freed memory, bypassing IOMMU protections.
> 
> Are we assuming endpoint drivers clear their own ATC or that a hardware
> reset handles it? Client endpoint drivers typically lack an API to
> manually clear the ATC, and SMMU hardware resets do not broadcast ATC
> invalidations to endpoints.
> "
> 
> We may get away from the TLB maintenance. But ATC can be the case
> broken by the stop flag?
> 

Not really, I guess Sashiko missed the comment in a later patch. This
was discussed with Jason in v9 [1].

The key thing is that ATC_INVs can't be issued by the time we set this
flag anyway.. if we're in the suspend callback, the PCIe EP is already
suspended, so none of the ATC invalidation TLPs would be responded to.

I'd addressed this in the RPM patch with a comment right above this check:

         * Note that eliding ATC invalidations (CMDQ_OP_ATC_INV) is safe
         * because client PCIe endpoints are guaranteed to be suspended
         * (via device links) before the SMMU is suspended. With the PCIe
         * links in a low-power state, no new TLPs can be transmitted.
         * It is strictly the responsibility of the client/endpoint driver
         * to quiesce DMA and ensure that the ATC state is cleared across
         * power state transitions.

i.e. we only start eliding once *all* consumers are RPM suspended, so
the "endpoint in D0" scenario can't happen.. and a suspended PCIe
function shouldn't be issuing translated requests anyway.

So we can't issue ATC_INVs during suspend (the EP is already down), nor
during resume (the SMMU resumes *before* the EP is made active). The EP
can't use its ATC while suspended, and if it loses power/resets on the
way back to D0 (from D3cold, or D3hot with No_Soft_Reset=0), it comes 
back with an empty ATC.. same assumption the PCI reset path makes today
(pci_dev_reset_iommu_prepare()).

I can add a pointer in this patch's comment to the RPM patch to make
that clearer.

[1] https://lore.kernel.org/all/20260826134757.GK3325090@nvidia.com/

Thanks,
Praan