[PATCH 2/2] iommu: Get DT/ACPI parsing into the proper probe path

[Jason Gunthorpe]

On Mon, Feb 17, 2025 at 03:00:46PM +0000, Robin Murphy wrote:
> On 14/02/2025 8:14 pm, Jason Gunthorpe wrote:
> > On Thu, Feb 13, 2025 at 11:49:00PM +0000, Robin Murphy wrote:
> > 
> > > much just calling the same path twice. At client driver probe time,
> > > dev->driver is obviously set; conversely at device_add(), or a
> > > subsequent bus_iommu_probe(), any device waiting for an IOMMU really
> > 
> > Could you put the dev->driver test into iommu_device_use_default_domain()?
> > 
> > It looks like many of the cases are just guarding that call.
> > 
> > > should *not* have a driver already, so we can use that as a condition to
> > > disambiguate the two cases, and avoid recursing back into the IOMMU core
> > > at the wrong times.
> > 
> > Which sounds like this:
> > 
> > > +		mutex_unlock(&iommu_probe_device_lock);
> > > +		dev->bus->dma_configure(dev);
> > > +		mutex_lock(&iommu_probe_device_lock);
> > > +	}
> > 
> > Shouldn't call iommu_device_use_default_domain() ?
> 
> Semantically it shouldn't really be called at this stage, but it won't be
> anyway since "to_<x>_driver(NULL)->driver_managed_dma" is not false -
> trouble is it's also not true ;)

That case in PCI I understood, but the other cases seemed like they
would be OK, especially if group is NULL

> > This is the test I mean, if iommu_group is set then
> > dev->iommu->iommu_dev->ops is supposed to be valid too. It seems like
> > it should be done earlier..
> 
> Yeah, looking at it now I'm really not sure why this ended up in this order
> - I guess I was effectively adding the dma_configure() call to the front of
> the existing iommu_fwspec_ops() check, and then I moved the lockdep_assert()
> up to make more sense. But then the ops check probably should have been
> after the group check to begin with, for much the same reasoning as above.
> I'll sort that out for v2.

I guess check it at the top and then check it again after re-locking.

> > > +	 * And if we do now see any replay calls, they would indicate someone
> > > +	 * misusing the dma_configure path outside bus code.
> > > +	 */
> > > +	if (dev_iommu_fwspec_get(dev) && dev->driver)
> > > +		dev_WARN(dev, "late IOMMU probe at driver bind, something fishy here!\n");
> > 
> > WARN_ON_ONCE or dump_stack() to get the stack trace out?
> 
> Indeed, hence dev_WARN() (!= dev_warn())

Oh, I've never seen that variation before!

Jason