Re: [PATCH v1] genirq/msi: fix crash when handling Multi-MSI

From: Thomas Gleixner
Date: Tue Jan 18 2022 - 19:44:56 EST


On Tue, Jan 18 2022 at 15:39, Thomas Gleixner wrote:
> On Mon, Jan 17 2022 at 11:36, Marc Zyngier wrote:
>> On Mon, 17 Jan 2022 10:10:13 +0000,
>> Tong Zhang <ztong0001@xxxxxxxxx> wrote:
>>> pci_msi_domain_check_cap (used by ops->msi_check(domain, info, dev))
>>> msi_domain_prepare_irqs
>>> __msi_domain_alloc_irqs
>>> msi_domain_alloc_irqs_descs_locked
>>>
>>> What I am suggesting is commit 0f62d941acf9 changed how this return
>>> value is being handled and created a UAF
>>
>> OK, this makes more sense.
>>
>> But msi_domain_prepare_irqs() shouldn't fail in this case, and we
>> should proceed with the allocation of at least one vector, which isn't
>> happening here.
>>
>> Also, if __msi_domain_alloc_irqs() is supposed to return the number of
>> irqs allocated, it isn't doing it consistently.
>>
>> Thomas, can you shed some light on what is the intended behaviour
>> here?
>
> Let me stare at it.

It's a subtle issue I overlooked. The UAF is due to

err:
pci_msi_unmask(entry, msi_multi_mask(entry));

in msi_capability_init() because the core has torn down and freed the
entry already.

The proposed patch "fixes" the issue for the PCI/MSI case, but could
cause a memory leak for other callers.

Not sure yet what the proper fix is, but that has to wait until tomorrow
when brain becomes awake again.

Thanks,

tglx