Message-ID: <ZSeOujvhyr8Rr8LY@Asurada-Nvidia>
Date:   Wed, 11 Oct 2023 23:14:18 -0700
From:   Nicolin Chen <nicolinc@...dia.com>
To:     Tina Zhang <tina.zhang@...el.com>
CC:     Jason Gunthorpe <jgg@...pe.ca>, Kevin Tian <kevin.tian@...el.com>,
        "Lu Baolu" <baolu.lu@...ux.intel.com>,
        Michael Shavit <mshavit@...gle.com>,
        "Vasant Hegde" <vasant.hegde@....com>,
        <linux-kernel@...r.kernel.org>
Subject: Re: [PATCH v7 0/5] Share sva domains with all devices bound to a mm

On Thu, Oct 12, 2023 at 11:01:07AM +0800, Tina Zhang wrote:

> This series is to share sva(shared virtual addressing) domains with all
> devices bound to one mm.
> 
> Problem
> -------
> In the current iommu core code, sva domain is allocated per IOMMU group,
> when device driver is binding a process address space to a device (which is
> handled in iommu_sva_bind_device()). If one than more device is bound to
> the same process address space, there must be more than one sva domain
> instance, with each device having one. In other words, the sva domain
> doesn't share between those devices bound to the same process address
> space, and that leads to two problems:
> 1) device driver has to duplicate sva domains with enqcmd, as those sva
> domains have the same PASID and are relevant to one virtual address space.
> This makes the sva domain handling complex in device drivers.
> 2) IOMMU driver cannot get sufficient info of the IOMMUs that have
> devices behind them bound to the same virtual address space, when handling
> mmu_notifier_ops callbacks. As a result, IOMMU IOTLB invalidation is
> performed per device instead of per IOMMU, and that may lead to
> superfluous IOTLB invalidation issue, especially in a virtualization
> environment where all devices may be behind one virtual IOMMU.
> 
> Solution
> --------
> This patch-set tries to fix those two problems by allowing sharing sva
> domains with all devices bound to a mm. To achieve this, a new structure
> pointer is introduced to mm to replace the old PASID field, which can keep
> the info of PASID as well as the corresponding shared sva domains.
> Besides, function iommu_sva_bind_device() is updated to ensure a new sva
> domain can only be allocated when the old ones cannot work for the IOMMU.
> With these changes, a device driver can expect one sva domain could work
> for per PASID instance(e.g., enqcmd PASID instance), and therefore may get
> rid of handling sva domain duplication. Besides, IOMMU driver (e.g., intel
> vt-d driver) can get sufficient info (e.g., the info of the IOMMUs having
> their devices bound to one virtual address space) when handling
> mmu_notifier_ops callbacks, to remove the redundant IOTLB invalidations.
> 
> Arguably there shouldn't be more than one sva_domain with the same PASID,
> and in any sane configuration there should be only 1 type of IOMMU driver
> that needs only 1 SVA domain. However, in reality, IOMMUs on one platform
> may not be identical to each other. Thus, attaching a sva domain that has
> been successfully bound to device A behind a IOMMU A, to device B behind
> IOMMU B may get failed due to the difference between IOMMU A and IOMMU
> B. In this case, a new sva domain with the same PASID needs to be
> allocated to work with IOMMU B. That's why we need a list to keep sva
> domains of one PASID. For the platform where IOMMUs are compatible to each
> other, there should be one sva domain in the list.
> 
> v7:
>   - Add mm_pasid_init() back and do zeroing mm->iommu_mm pointer in
>     mm_pasid_init() to avoid the use-after-free/double-free problem.
>   - Update the commit message of patch "iommu: Add mm_get_enqcmd_pasid()
>     helper function".

Tested this v7 running some SVA cases with SMMUv3:

Tested-by: Nicolin Chen <nicolinc@...dia.com>

Powered by blists - more mailing lists