Re: [PATCH rc v4] iommu: Fix nested pci_dev_reset_iommu_prepare/done()

[Nicolin Chen <nicolinc@nvidia.com>]

On Thu, Apr 02, 2026 at 07:34:24AM +0000, Tian, Kevin wrote:
> But with it there might be corner cases to pay attention in case iommu
> driver calls it in other places. e.g. Sashiko [1] pointed out one potential issue:
> 
> "
> Incrementing gdev->reset_depth at the beginning of the function prematurely
> exposes the blocking_domain to the driver. When __iommu_attach_device() is
> called above, if the driver's attach_dev callback uses
> iommu_driver_get_domain_for_dev() to query its current domain, it will observe
> reset_depth == 1 and receive the blocking_domain even though it is not yet
> attached to it.
> Could reset_depth be set to 1 only after a successful attachment, while nested
> calls still check for gdev->reset_depth > 0 at the beginning?
> "

Ah, that seems a good catch.

I noticed the same problem in done() but overlooked prepare().

Yea, it should do something likewise:

	if (gdev->reset_depth) {
		gdev->reset_depth++);
		return 0;
	}
	...
	attach(blocking_domain);
	set_pasid(blocking_domain);
	gdev->reset_depth = 1;

Nicolin