Re: [BUG] iommufd/selftest: KASAN slab-use-after-free in iommu_report_device_fault

[Jason Gunthorpe <jgg@ziepe.ca>]

On Tue, Jul 28, 2026 at 01:32:09PM +0800, Peiyang He wrote:
> On 2026/7/27 17:41, Tian, Kevin wrote:
> >> From: Peiyang He <peiyang_he@smail.nju.edu.cn>
> >> Sent: Wednesday, July 22, 2026 5:52 PM
> >>
> > [...]
> >  > Since iommu_attach_handle_get() explicitly says that the caller should be
> >> responsible for "synchronize the call of iommu_attach_handle_get()
> >> with domain attachment and detachment", maybe we shouldn't add extra
> >> lifetime
> >> management mechanism around iommu_attach_handle, instead we should
> >> fix this at the caller side.
> >>
> >> Fault report can run concurrently with PASID detach and replacement, but
> >> before the old attach handle is removed and freed, mockdev must block new
> >> reports
> >> for that PASID and wait for in-flight iommu_report_device_fault() calls to
> >> finish.
> >>
> > 
> > yes the mock driver should drain in-flight fault requests before the
> > actual detachment, just like other iommu drivers do.
> 
> Hi Kevin,
> 
> Thanks for your reply. I looked into the code in more detail. Would this
> approach make sense?
> 
> In iommufd_test_trigger_iopf(), call iommu_report_device_fault() under the protection
> of idev->igroup->lock. Since detach and replace also require the same lock, they can
> be serialized with fault reporting. This would ensure that a mock fault report either
> completes before detach/replace starts, or starts only after the old handle has been removed.
> Thus it should prevent the UAF.
> 
> Would this simple serialization meet the design intention of mock dev? Or would you prefer
> an explicit quiesce/drain mechanism instead?

The mock driver should follow the expected design of a real driver, so
I'd put an internal rwsem or something like that to protect this

Jason