Re: [PATCH v3 4/7] iommu/vt-d: Fold dmar_remove_one_dev_info() into its caller

[Baolu Lu]

On 2022/11/16 17:15, Tian, Kevin wrote:
> > From: Baolu Lu <baolu.lu@xxxxxxxxxxxxxxx>
> > Sent: Wednesday, November 16, 2022 4:03 PM
> >
> > On 2022/11/16 13:35, Tian, Kevin wrote:
> > > > From: Baolu Lu<baolu.lu@xxxxxxxxxxxxxxx>
> > > > Sent: Wednesday, November 16, 2022 12:36 PM
> > > >
> > > > On 11/16/22 11:53 AM, Tian, Kevin wrote:
> > > > > > From: Lu Baolu<baolu.lu@xxxxxxxxxxxxxxx>
> > > > > > Sent: Monday, November 14, 2022 9:41 AM
> > > > > > @@ -4562,7 +4538,10 @@ static void
> > intel_iommu_release_device(struct
> > > > > > device *dev)
> > > > > >     {
> > > > > >     	struct device_domain_info *info = dev_iommu_priv_get(dev);
> > > > > >
> > > > > > -	dmar_remove_one_dev_info(dev);
> > > > > > +	iommu_disable_pci_caps(info);
> > > > > > +	domain_context_clear(info);
> > > > > > +	device_block_translation(dev);
> > > > > clear context after blocking translation.
> > > > Unfortunately domain_context_clear() needs reference to info->domain
> > > > (for domain id when flushing cache), which is cleared in
> > > > device_block_translation().
> > > this sounds an ordering problem. clearing context should be after
> > > blocking translation in concept.
> >
> > At present, when the default domain is attached to the device, we first
> > populate the pasid table entry, and then populate the device context
> > entry. Above code is just the reverse operation.
> >
> > Can you see any practical problems caused by this sequence? If so, it
> > seems that we should carefully consider whether such problems already
> > exist.
>
> there is no problem with existing code. Just after this patch the order
> looks weird based on the literal name of those functions.
>
> domain_context_clear() is a big hammer to disable the context entry,
> implying translation must be blocked. Then calling another block
> translation afterwards becomes unnecessary.
>
> Probably it should be split into two functions with one requiring
> info->domain called before block translation and the rest which
> actually clears the context entry being the last step?

This is what the existing code does. Perhaps I should drop this patch,
or only rename iommu_disable_dev_iotlb() to iommu_disable_pci_caps().

Best regards,
baolu