Re: [PATCH 3/4] iommu: Add support to change default domain of an iommu_group

[Lu Baolu <baolu.lu@linux.intel.com>]

Hi,

On 9/4/19 11:09 AM, Prakhya, Sai Praneeth wrote:
> Hi Joerg,
> 
> Thanks a lot! for the review. I highly appreciate for sparing your time to review the patch :)
> 
>> On Tue, Aug 20, 2019 at 07:42:25PM -0700, Sai Praneeth Prakhya wrote:
>>> +	/*
>>> +	 * iommu_domain_alloc() takes "struct bus_type" as an argument which
>> is
>>> +	 * a member in "struct device". Changing a group's default domain type
>>> +	 * deals at iommu_group level rather than device level and hence there
>>> +	 * is no straight forward way to get "bus_type" of an iommu_group that
>>> +	 * could be passed to iommu_domain_alloc(). So, instead of directly
>>> +	 * calling iommu_domain_alloc(), use iommu_ops from previous default
>>> +	 * domain.
>>> +	 */
>>> +	if (!prev_domain || !prev_domain->ops ||
>>> +	    !prev_domain->ops->domain_alloc || !type)
>>> +		return -EINVAL;
>>
>> Hmm, this isn't really nice and clean, but I understand why you need it.
> 
> I agree.. It didn't look good for me either :(
> But, I didn't find any other better solution.
> 
>> I will think about a better way to get iommu_ops here.
> 
> Sure! That will be great!
> 
>>> +free_prev_domain:
>>> +	/*
>>> +	 * Free the existing default domain and replace it with the newly
>>> +	 * created default domain. No need to set group->domain because
>>> +	 * __iommu_attach_group() already does it on success.
>>> +	 */
>>> +	iommu_domain_free(prev_domain);
>>> +	group->default_domain = new_domain;
>>> +	return 0;
>>
>> It isn't obvious to me from this patch, how to are the dma_ops updated when
>> the default domain changes between identity and dma?
> 
> Thanks for raising this.
> For intel_iommu, dma_map_ops is defined at drivers/iommu/intel-iommu.c and
> all the callbacks like alloc(), map_sg() and map_page(), check if the device needs DMA mapping or not
> by calling iommu_need_mapping(). The callbacks inherently do the right thing based on the outcome.
> So, essentially the dma_ops are same for dma/identity domain.

This isn't always true as we are considering per-device dma ops.

Best regards,
Baolu

> 
> I just realized (sorry!) that other iommu drivers (Eg: AMD) doesn't do it the same way i.e. looks like the callbacks
> aren't checking if the device needs a dma mapping or identity mapping.
> I will take a look at other iommu drivers and will handle this in V2.
> 
> Please let me know if I missed something.
> 
>>> +	/* Check if any device in the group still has a driver binded to it */
>>> +	if (iommu_group_for_each_dev(group, NULL, is_driver_binded)) {
>>> +		pr_err("Active drivers exist for devices in the group\n");
>>> +		return -EBUSY;
>>> +	}
>>
>> This is racy with device driver probing code. Unfortunatly there is no clean way
>> out of that either, locking all devices in the group and then do the re-attach will
>> introduce a lock-inversion with group->mutex. But please put a comment here
>> saying that this might race with device driver probing.
> 
> Sure! Makes sense. Will add it in V2.
> 
> Regards,
> Sai
> 
_______________________________________________
iommu mailing list
iommu@lists.linux-foundation.org
https://lists.linuxfoundation.org/mailman/listinfo/iommu