Re: [PATCH v3 4/5] iommu/hyperv: Add para-virtualized IOMMU support for Hyper-V guest

On Thu, Aug 13, 2026 at 07:46:27PM -0300, Jason Gunthorpe wrote:

On Thu, Aug 13, 2026 at 03:29:13PM -0700, Jacob Pan wrote:

quoted

quoted

+static int hv_iommu_blocking_attach_dev(struct iommu_domain *domain,
+					struct device *dev,
+					struct iommu_domain *old)
+{
+	struct hv_iommu_endpoint *vdev = dev_iommu_priv_get(dev);
+	struct hv_iommu_domain *hv_domain =
to_hv_iommu_domain(domain);
+	struct pci_dev *pdev = to_pci_dev(dev);
+	int ret;
+
+	if (pdev->ats_enabled)
+		pci_disable_ats(pdev);
+
+	if (vdev->hv_domain == hv_domain)
+		return 0;

It may not be needed since the core code will check this. If you keep
this as extra precaution, should this be moved up before ats? since ats
must be disabled already if blocking domain is attached.

@Jacob In v2 (https://lore.kernel.org/all/20260706095510.00007ce1@linux.microsoft.com/ (local)), 
this check was in the common attach helper. It was redundant because the normal
group domain path already skips same-domain attaches.
                                                                                                                                                        
In this version, I removed it from the common helper and kept it only in the 
blocking attach callback. iommu_deinit_device() directly attaches the
device to release_domain without going through the group-level check.
IIUC, for a device which is alreading parked in the blocking domain, the
releasing might invoke another blocking-domain attach.
                                                                                                                                                        
Sorry, I should have explained this change and its rationale in the changelog.

Why check it at all? Is there an issue with attaching the same thing
twice? Will it blip the DMA or something?

Actually, I do not think there should be any issue with attaching twice.
Hyper-V *should* be able to accept such attachment. But I do not want to
make such assumption just for a redundant hypercall.

Does this reasoning make sense, or would you prefer that I remove the
check?

B.R.
Yu

Jason