Re: [PATCH v6 07/10] iommufd: Fault-capable hwpt attach/detach/replace

[Baolu Lu]

On 6/7/24 5:30 PM, Tian, Kevin wrote:
> > From: Lu Baolu <baolu.lu@xxxxxxxxxxxxxxx>
> > Sent: Monday, May 27, 2024 12:05 PM
> >
> > Add iopf-capable hw page table attach/detach/replace helpers. The pointer
> > to iommufd_device is stored in the domain attachment handle, so that it
> > can be echo'ed back in the iopf_group.
>
> this message needs an update. now the device pointer is not in the
> attach handle.

The iommufd_device pointer is in the attach handle provided by iommufd 
in attach or replace path.

> and there worths a explanation about VF in the commit msg.
>
> > @@ -376,7 +377,10 @@ int iommufd_hw_pagetable_attach(struct
> > iommufd_hw_pagetable *hwpt,
> >   	 * attachment.
> >   	 */
> >   	if (list_empty(&idev->igroup->device_list)) {
> > -		rc = iommu_attach_group(hwpt->domain, idev->igroup->group);
> > +		if (hwpt->fault)
> > +			rc = iommufd_fault_domain_attach_dev(hwpt, idev);
> > +		else
> > +			rc = iommu_attach_group(hwpt->domain, idev-
> > > igroup->group);
>
> Does it read better to have a iommufd_attach_device() wrapper with
> above branches handled internally?

Yes. Will do this in the next version.

> > +static int iommufd_fault_iopf_enable(struct iommufd_device *idev)
> > +{
> > +	struct device *dev = idev->dev;
> > +	int ret;
> > +
> > +	/*
> > +	 * Once we turn on PCI/PRI support for VF, the response failure code
> > +	 * could not be forwarded to the hardware due to PRI being a shared
>
> you could but just doing so is incorrect. 😊
>
> s/could/should/

Done.

> > +	 * resource between PF and VFs. There is no coordination for this
> > +	 * shared capability. This waits for a vPRI reset to recover.
> > +	 */
>
> this may go a bit further to talk about supporting it requires an emulation
> in iommufd (i.e. pause any further fault delivery until vPRI reset). It is a
> future work so disable it for VF at this point.

Yes.

> > +void iommufd_fault_domain_detach_dev(struct iommufd_hw_pagetable
> > *hwpt,
> > +				     struct iommufd_device *idev)
> > +{
> > +	struct iommufd_attach_handle *handle;
> > +
> > +	handle = iommufd_device_get_attach_handle(idev);
> > +	iommu_detach_group_handle(hwpt->domain, idev->igroup->group);
> > +	iommufd_auto_response_faults(hwpt, handle);
> > +	iommufd_fault_iopf_disable(idev);
> > +	kfree(handle);
>
> this assumes that the detach callback of the iommu driver needs to drain
> in-fly page requests so no further reference to handle or queue new req
> to the deliver queue when it returns, otherwise there is a use-after-free
> race or stale requests in the fault queue which auto response doesn't
> cleanly handle.
>
> iirc previous discussion reveals that only intel-iommu driver guarantees
> that behavior. In any case this should be documented to avoid this being
> used in a non-conforming iommu driver.
>
> If I misunderstood, some comment why no race in this window is also
> appreciated. 😊

Yes. The iommu driver needs to guarantee that there will be no iopf
request for a RID or PASID after the domain has been detached. This
implies that:

- IOMMU hardware should not put further iopf in its hardware queue if
  the domain has been detached.
- Before the domain detachment is complete, the iommu driver should
  flush all iopf targeting the detached domain in the hardware page
  request queue.

> > +}
> > +
> > +static int __fault_domain_replace_dev(struct iommufd_device *idev,
> > +				      struct iommufd_hw_pagetable *hwpt,
> > +				      struct iommufd_hw_pagetable *old)
> > +{
> > +	struct iommufd_attach_handle *handle, *curr = NULL;
> > +	int ret;
> > +
> > +	if (old->fault)
> > +		curr = iommufd_device_get_attach_handle(idev);
> > +
> > +	if (hwpt->fault) {
> > +		handle = kzalloc(sizeof(*handle), GFP_KERNEL);
> > +		if (!handle)
> > +			return -ENOMEM;
> > +
> > +		handle->handle.domain = hwpt->domain;
> > +		handle->idev = idev;
> > +		ret = iommu_replace_group_handle(idev->igroup->group,
> > +						 hwpt->domain, &handle-
> > > handle);
> > +	} else {
> > +		ret = iommu_replace_group_handle(idev->igroup->group,
> > +						 hwpt->domain, NULL);
> > +	}
> > +
> > +	if (!ret && curr) {
> > +		iommufd_auto_response_faults(old, curr);
> > +		kfree(curr);
> > +	}
>
> In last version you said auto response is required only when switching
> from fault-capable old to fault-incapable new. But above code doesn't
> reflect that description?

What the current code does is auto-respond to all page faults targeting
the old fault-capable hwpt. I'm also okay if we decide to limit this to
flushing page faults only if the new hwpt is *not* fault-capable.

Best regards,
baolu