Re: [PATCH 2/2] iommu/vt-d: Replace intel SVM APIs with generic SVA APIs

From: Lu Baolu <baolu.lu@linux.intel.com>
To: "Raj, Ashok" <ashok.raj@intel.com>,
	Jean-Philippe Brucker <jean-philippe@linaro.org>
Cc: "Tian, Kevin" <kevin.tian@intel.com>,
	Dave Jiang <dave.jiang@intel.com>,
	"Lu, Baolu" <baolu.lu@intel.com>,
	Jean-Philippe Brucker <jean-philippe@linaro.com>,
	LKML <linux-kernel@vger.kernel.org>,
	iommu@lists.linux-foundation.org,
	David Woodhouse <dwmw2@infradead.org>
Subject: Re: [PATCH 2/2] iommu/vt-d: Replace intel SVM APIs with generic SVA APIs
Date: Tue, 24 Mar 2020 09:42:20 +0800	[thread overview]
Message-ID: <6cc2f1ea-5a7d-eeab-b50c-5b464098de6b@linux.intel.com> (raw)
In-Reply-To: <20200323230113.GA84386@otc-nc-03>

On 2020/3/24 7:01, Raj, Ashok wrote:
> Hi Jean
> 
> On Fri, Mar 20, 2020 at 10:29:55AM +0100, Jean-Philippe Brucker wrote:
>>> +#define to_intel_svm_dev(handle) container_of(handle, struct intel_svm_dev, sva)
>>> +struct iommu_sva *
>>> +intel_svm_bind(struct device *dev, struct mm_struct *mm, void *drvdata)
>>> +{
>>> +	struct iommu_sva *sva = ERR_PTR(-EINVAL);
>>> +	struct intel_svm_dev *sdev = NULL;
>>> +	int flags = 0;
>>> +	int ret;
>>> +
>>> +	/*
>>> +	 * TODO: Consolidate with generic iommu-sva bind after it is merged.
>>> +	 * It will require shared SVM data structures, i.e. combine io_mm
>>> +	 * and intel_svm etc.
>>> +	 */
>>> +	if (drvdata)
>>> +		flags = *(int *)drvdata;
>>
>> drvdata is more for storing device driver contexts that can be passed to
>> iommu_sva_ops, but I get that this is temporary.
>>
>> As usual I'm dreading supervisor mode making it into the common API. What
>> are your plans regarding SUPERVISOR_MODE and PRIVATE_PASID flags?  The
>> previous discussion on the subject [1] had me hoping that you could
>> replace supervisor mode with normal mappings (auxiliary domains?)
>> I'm less worried about PRIVATE_PASID, it would just add complexity into
> 
> We don't seem to have an immediate need for PRIVATE_PASID. There are some talks
> about potential usages, but nothing concrete. I think it might be good to
> get rid of it now and add when we really need.
> 
> For SUPERVISOR_MODE, the idea is to have aux domain. Baolu is working on
> something to replace. Certainly the entire kernel address is opening up
> the whole kimono.. so we are looking at dynamically creating mappings on demand.
> It might take some of the benefits of SVA in general with no need to create
> mappings, but for security somebody has to pay the price :-)

My thought is to reuse below aux-domain API.

int iommu_aux_attach_device(struct iommu_domain *domain,
			    struct device *dev)

Currently, it asks the vendor iommu driver to allocate a PASID and bind
the domain with it. We can change it to allow the caller to pass in an
existing supervisor PASID.

int iommu_aux_attach_device(struct iommu_domain *domain,
			    struct device *dev,
			    int *pasid)

In the vendor iommu driver, if (*pasid == INVALID_IOASID), it will
allocate a pasid (the same as current behavior); otherwise, attach
the domain with the pass-in pasid.

Best regards,
baolu
_______________________________________________
iommu mailing list
iommu@lists.linux-foundation.org
https://lists.linuxfoundation.org/mailman/listinfo/iommu