Re: [PATCH] iommu: Remove functions that support private domain

[Joerg Roedel <joro@8bytes.org>]

On Sun, May 17, 2020 at 08:29:17AM +0000, Prakhya, Sai Praneeth wrote:
> iommu_bus_notifier()
> -> iommu_release_device()
>  -> ops->release_device() (Eg: intel_iommu_release_device())
>   -> iommu_group_remove_device()
>    -> mutex_lock()
> 
> But, I added print statements to iommu_bus_notifier() and
> iommu_group_remove_device() and noticed that
> iommu_group_remove_device() wasn't being called upon modprobe -r
> <driver_name> and iommu_probe_device() isn't called upon modprobe
> <driver_name> because

Calling modprobe is the device driver binding path, the release_device
event is called when a device is going away, e.g. on hotunplug or when a
VF of a PCI device is released.

This could also happen at runtime, and the code needs to protect against
that. My suggestion is still to limit runtime-changing of default
domains to groups with only one device. The flow would be as follows:


	1. device_lock(dev);
	   // Device can't be bound to a driver or hot-removed from now
	   // on.

	2. if (device_is_bound_to_some_driver(dev))
	           goto 6;

	3. mutex_lock(&group->mutex);
	   // group of dev

	4. Change the default domain

	5. mutex_unlock(&group->mutex);

	6. device_unlock(dev);

This avoids lock inversion and should be safe with regard to hotplug and
device driver probing.

Regards,

	Joerg
_______________________________________________
iommu mailing list
iommu@lists.linux-foundation.org
https://lists.linuxfoundation.org/mailman/listinfo/iommu