Re: [PATCH v6 3/4] iommu/vt-d: Add set_dev_pasid callback for dma domain

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

On 5/31/23 12:55 AM, Jason Gunthorpe wrote:
> On Tue, May 30, 2023 at 10:19:05AM +0800, Baolu Lu wrote:
>> On 5/30/23 3:48 AM, Jason Gunthorpe wrote:
>>> On Fri, May 19, 2023 at 01:32:22PM -0700, Jacob Pan wrote:
>>>
>>>> @@ -4720,25 +4762,99 @@ static void intel_iommu_iotlb_sync_map(struct iommu_domain *domain,
>>>>    static void intel_iommu_remove_dev_pasid(struct device *dev, ioasid_t pasid)
>>>>    {
>>>>    	struct intel_iommu *iommu = device_to_iommu(dev, NULL, NULL);
>>>> +	struct dev_pasid_info *curr, *dev_pasid = NULL;
>>>> +	struct dmar_domain *dmar_domain;
>>>>    	struct iommu_domain *domain;
>>>> +	unsigned long flags;
>>>> -	/* Domain type specific cleanup: */
>>>>    	domain = iommu_get_domain_for_dev_pasid(dev, pasid, 0);
>>>> -	if (domain) {
>>>> -		switch (domain->type) {
>>>> -		case IOMMU_DOMAIN_SVA:
>>>> -			intel_svm_remove_dev_pasid(dev, pasid);
>>>> -			break;
>>>> -		default:
>>>> -			/* should never reach here */
>>>> -			WARN_ON(1);
>>>> +	if (!domain)
>>>> +		goto out_tear_down;
>>>> +
>>>> +	/*
>>>> +	 * The SVA implementation needs to stop mm notification, drain the
>>>> +	 * pending page fault requests before tearing down the pasid entry.
>>>> +	 * The VT-d spec (section 6.2.3.1) also recommends that software
>>>> +	 * could use a reserved domain id for all first-only and pass-through
>>>> +	 * translations. Hence there's no need to call domain_detach_iommu()
>>>> +	 * in the sva domain case.
>>>> +	 */
>>>> +	if (domain->type == IOMMU_DOMAIN_SVA) {
>>>> +		intel_svm_remove_dev_pasid(dev, pasid);
>>>> +		goto out_tear_down;
>>>> +	}
>>>
>>> But why don't you need to do all the other
>>> intel_pasid_tear_down_entry(), intel_svm_drain_prq() (which is
>>> misnamed) and other stuff from intel_svm_remove_dev_pasid() ?
>>
>> Perhaps,
>>
>> 	if (domain->type == IOMMU_DOMAIN_SVA) {
>> 		intel_svm_remove_dev_pasid(dev, pasid);
>> 		return;
>> 	}
>>
>> ?
> 
> I would expect only stuff directly connected to SVM be in the SVM
> function.
> 
> De-initalizing PRI and any other pasid destruction should be in this
> function.
> 
>>> There still seems to be waaay too much "SVM" in the PASID code.
>>
>> This segment of code is destined to be temporary. From a long-term
>> perspective, I hope to move SVA specific staffs such as mm notification,
>> prq draining, etc. to the iommu core. They are generic rather than Intel
>> iommu specific.
> 
> Yes, sort of, but.. That is just the mmu notifier bits
> 
> All the PRI/PASID teardown needs to be unlinked from SVM

Get your point now. Yes. PRI and PASID teardown are not SVA-specific.
Sorry that we should rename SVM to SVA to unify the Linux terminology.

> 
>>> It would be nice if the different domain types had their own ops..
>>
>> Good suggestion!
>>
>> We can add a domain ops in the Intel domain structure which is
>> responsible for how to install an Intel iommu domain onto the VT-d
>> hardware.
> 
> We should have seperate iommu_domain_ops at least, I think that would
> cover alot of it?

Are you suggesting adding this ops in common iommu_domain or intel's
dmar_domain? My understanding is the latter. To do so, probably we need
to define various callbacks for different type of domains: identity,
blocking, dma remapping, sva and possibly nested. Also need to care
about legacy vs. scalable mode.

That's the reason why I hoped to do all these in separated series with
carefully reviewing and testing.

Best regards,
baolu