Re: [PATCH v3 01/18] iommu: Add device dma ownership set/release interfaces

[Lu Baolu]

On 12/6/21 11:01 PM, Jason Gunthorpe wrote:
> On Mon, Dec 06, 2021 at 02:35:55PM +0100, Joerg Roedel wrote:
> > On Mon, Dec 06, 2021 at 09:58:46AM +0800, Lu Baolu wrote:
> > > >From the perspective of who is initiating the device to do DMA, device
> > > DMA could be divided into the following types:
> > >
> > >          DMA_OWNER_DMA_API: Device DMAs are initiated by a kernel driver
> > > 			through the kernel DMA API.
> > >          DMA_OWNER_PRIVATE_DOMAIN: Device DMAs are initiated by a kernel
> > > 			driver with its own PRIVATE domain.
> > > 	DMA_OWNER_PRIVATE_DOMAIN_USER: Device DMAs are initiated by
> > > 			userspace.
> >
> > I have looked at the other iommu patches in this series, but I still
> > don't quite get what the difference in the code flow is between
> > DMA_OWNER_PRIVATE_DOMAIN and DMA_OWNER_PRIVATE_DOMAIN_USER. What are the
> > differences in the iommu core behavior based on this setting?
>
> USER causes the IOMMU code to spend extra work to never assign the
> default domain. Lu, it would be good to update the comment with this
> detail
>
> Once in USER mode the domain is always a /dev/null domain or a domain
> controlled by userspace. Never a domain pointing at kernel memory.

Yes. The __iommu_detach_group() re-attaches the default domain
automatically. This is not allowed once in USER mode.

I will update the comments whit this detail.

> > >   struct group_device {
> > > @@ -621,6 +624,7 @@ struct iommu_group *iommu_group_alloc(void)
> > >   	INIT_LIST_HEAD(&group->devices);
> > >   	INIT_LIST_HEAD(&group->entry);
> > >   	BLOCKING_INIT_NOTIFIER_HEAD(&group->notifier);
> > > +	group->dma_owner = DMA_OWNER_NONE;
> >
> >
> > DMA_OWNER_NONE is also questionable. All devices are always in one
> > domain, and the default domain is always the one used for DMA-API, so
> > why isn't the initial value DMA_OWNER_DMA_API?
>
> 'NONE' means the group is in the default domain but no driver is bound
> and thus DMA isn't being used. Seeing NONE is the only condition when
> it is OK to change the domain.
>
> This could be reworked to instead rely on the refcount == 0 as the
> signal to know it is OK to change the domain and then we never have
> NONE at all. Lu?

NONE is just a parking state. It's okay to rely on the "refcount == 0"
for state transition as far as I see. I will work towards this.

Best regards,
baolu