On 17.03.2023 09:43, Roger Pau Monné wrote:
> On Tue, Mar 14, 2023 at 08:56:30PM +0000, Volodymyr Babchuk wrote:
>> vPCI MMIO handlers are accessing pdevs without protecting this
>> access with pcidevs_{lock|unlock}. This is not a problem as of now
>> as these are only used by Dom0. But, towards vPCI is used also for
>> guests, we need to properly protect pdev and pdev->vpci from being
>> removed while still in use.
>>
>> For that use pdev reference counting.
>
> I wonder whether vPCI does need to take another reference to the
> device. This all stems from me not having it fully clear how the
> reference counting is supposed to be used for pdevs.
>
> As mentioned in a previous patch, I would expect device assignation to
> take a reference, and hence vPCI won't need to take an extra refcount
> since vPCI can only be used once the device has been assigned to a
> domain, and hence already has at least an extra reference taken from
> the fact it's assigned to a domain.
>
> If anything I would add an ASSERT(pdev->refcount > 1) or equivalent.
FWIW: +1
Jan
