Re: [PATCH v2 7/8] iommu: Export iommu_get_dma_domain

[Jacob Pan]

Hi Kevin,

On Tue, 28 Mar 2023 07:52:23 +0000, "Tian, Kevin" <kevin.tian@xxxxxxxxx>
wrote:

> > From: Baolu Lu <baolu.lu@xxxxxxxxxxxxxxx>
> > Sent: Tuesday, March 28, 2023 2:04 PM
> > 
> > On 3/28/23 7:21 AM, Jacob Pan wrote:  
> > > Devices that use ENQCMDS to submit work needs to retrieve its DMA
> > > domain. It can then attach PASID to the DMA domain for shared mapping
> > > (with RID) established by DMA API.
> > >
> > > Signed-off-by: Jacob Pan<jacob.jun.pan@xxxxxxxxxxxxxxx>
> > > ---
> > >   drivers/iommu/iommu.c | 1 +
> > >   include/linux/iommu.h | 5 +++++
> > >   2 files changed, 6 insertions(+)
> > >
> > > diff --git a/drivers/iommu/iommu.c b/drivers/iommu/iommu.c
> > > index 10db680acaed..c51d343a75d2 100644
> > > --- a/drivers/iommu/iommu.c
> > > +++ b/drivers/iommu/iommu.c
> > > @@ -2118,6 +2118,7 @@ struct iommu_domain  
> > *iommu_get_dma_domain(struct device *dev)  
> > >   {
> > >   	return dev->iommu_group->default_domain;
> > >   }
> > > +EXPORT_SYMBOL_GPL(iommu_get_dma_domain);  
> > 
> > Directly exporting this function for external use seems unsafe. If the
> > caller is the kernel driver for this device, it's fine because default
> > domain remains unchanged during the life cycle of the driver. Otherwise,
> > using this function may cause UAF. Keep in mind that group's default
> > domain could be changed through sysfs.
> > 
> > However, iommu_get_domain_for_dev() has already done so and has been
> > exported. Maybe I'm worried too much. :-)
> >   
> 
> Agree. The kernel driver managing the device wants to get the current 
> domain of the device then iommu_get_domain_for_dev() is the right
> interface. It knows the domain is the dma domain.
This goes back to v1 then :)

I thought the comments from v1 is that we don't want to check the domain
type is DMA domain returned by iommu_get_domain_for_dev()

If the driver "knows" the domain is dma domain, why can't it use
iommu_get_dma_domain()? seems paradoxical.

Thanks,

Jacob