Re: [PATCH] iommu: Implement deferred domain attachment

[Robin Murphy]

On 2020-05-15 19:26, Joerg Roedel wrote:
> On Fri, May 15, 2020 at 05:28:53PM +0100, Robin Murphy wrote:
> > On 2020-05-15 17:14, Joerg Roedel wrote:
> > > diff --git a/drivers/iommu/dma-iommu.c b/drivers/iommu/dma-iommu.c
> > > index ba128d1cdaee..403fda04ea98 100644
> > > --- a/drivers/iommu/dma-iommu.c
> > > +++ b/drivers/iommu/dma-iommu.c
> > > @@ -362,8 +362,8 @@ static int iommu_dma_deferred_attach(struct device *dev,
> > >    		return 0;
> > >    	if (unlikely(ops->is_attach_deferred &&
> > > -			ops->is_attach_deferred(domain, dev)))
> > > -		return iommu_attach_device(domain, dev);
> > > +		     ops->is_attach_deferred(domain, dev)))
> > > +		return iommu_attach_device_no_defer(domain, dev);
> >
> > Wouldn't it be simpler to just invoke ops->attach_dev directly and avoid
> > having to formalise a public interface that nobody else should ever use
> > anyway?
>
> That would omit the ops->attach_dev != NULL check and the trace-point on
> device attach. Besides that, it would be a layering violation. But the
> function is of course entirely internal to the iommu subsytem and is a
> good canditate to be moved to a header file in drivers/iommu.

Sure, checking the pointer before calling was implied, but the 
tracepoint is a good argument, I'd forgotten about that :)

> > @@ -746,8 +747,11 @@ int iommu_group_add_device(struct iommu_group *group,
> > struct device *dev)
> >
> >          mutex_lock(&group->mutex);
> >          list_add_tail(&device->list, &group->devices);
> > -       if (group->domain)
> > -               ret = __iommu_attach_device(group->domain, dev);
> > +       domain = group->domain;
> > +       if (domain && (!domain->ops->is_attach_deferred ||
> > +                      !domain->ops->is_attach_deferred(domain, dev)))
> > +               ret = __iommu_attach_device(domain, dev);
> > +       }
> >          mutex_unlock(&group->mutex);
> >          if (ret)
> >                  goto err_put_group;
>
> No, doing this in iommu_group_add_device() doesn't solve the problem.
> The attach must not happen before a device driver took control of the
> device and silenced any DMA initiated by the old kernel. At probe time
> this isn't guaranteed.

But that's not what this is; this is (supposed to be) the exact same 
"don't actually perform the attach yet" logic as before, just 
restricting it to default domains in the one place that it actually 
needs to be, so as not to fundamentally bugger up iommu_attach_device() 
in a way that prevents it from working as expected at the correct point 
later.

Thinking a bit more, consider if the driver resets the device then 
attaches it straight to its own unmanaged domain rather than calling any 
DMA ops (e.g. VFIO?) - it looks like that would also be totally broken 
right now, and no amount of bodges in iommu-dma is going to help there.

Robin.