RE: [RFC v3 2/8] vfio/type1: Make per-application (VM) PASID quota tunable

["Liu, Yi L" <yi.l.liu@intel.com>]

Hi Jacob,

> From: Jacob Pan [mailto:jacob.jun.pan@linux.intel.com]
> Sent: Saturday, February 8, 2020 3:44 AM
> To: Liu, Yi L <yi.l.liu@intel.com>
> Subject: Re: [RFC v3 2/8] vfio/type1: Make per-application (VM) PASID quota tunable
> 
> On Wed, 29 Jan 2020 04:11:46 -0800
> "Liu, Yi L" <yi.l.liu@intel.com> wrote:
> 
> > From: Liu Yi L <yi.l.liu@intel.com>
> >
> > The PASID quota is per-application (VM) according to vfio's PASID
> > management rule. For better flexibility, quota shall be user tunable
> > . This patch provides a VFIO based user interface for which quota can
> > be adjusted. However, quota cannot be adjusted downward below the
> > number of outstanding PASIDs.
> >
> > This patch only makes the per-VM PASID quota tunable. While for the
> > way to tune the default PASID quota, it may require a new vfio module
> > option or other way. This may be another patchset in future.
> >
> One issue we need to solve is how to share PASIDs at the system
> level, e.g. Both VMs and baremetal drivers could use PASIDs.
> 
> This patch is granting quota to a guest w/o knowing the remaining
> system capacity. So guest PASID allocation could fail even within its
> quota.

that's true.

> The solution I am thinking is to enforce quota at IOASID common
> code, since IOASID APIs already used to manage system-wide allocation.
> How about the following changes to IOASID?
> 1. introduce quota in ioasid_set (could have a soft limit for better
> sharing)
>
> 2. introduce an API to create a set with quota before allocation, e.g.
> ioasid_set_id = ioasid_alloc_set(size, token)
> set_id will be used for ioasid_alloc() instead of token.

Is the token the mm pointer? I guess you may want to add one more
API like ioasid_get_set_id(token), thus that other ioasid user could get
set_id with their token. If token is the same give them the same set_id.

> 
> 3. introduce API to adjust set quota ioasid_adjust_set_size(set_id,
> size)
> 
> 4. API to check remaining PASIDs ioasid_get_capacity(set_id); //return
> system capacity if set_id == 0;
> 
> 5. API to set system capacity, ioasid_set_capacity(nr_pasids), e.g. if
> system has 20 bit PASIDs, IOMMU driver needs to call
> ioasid_set_capacity(1<<20) during boot.

yes, this is definitely necessary.

> 6. Optional set level APIs. e.g. ioasid_free_set(set_id), frees all
> IOASIDs in the set.

If this is provided. I think VFIO may be not necessary to track allocated
PASIDs. When VM is down or crashed, VFIO just use this API to reclaim
allocated PASIDs.

> With these APIs, this patch could query PASID capacity at both system
> and set level and adjust quota within range. i.e.
> 1. IOMMU vendor driver(or other driver to use PASID w/o IOMMU) sets
> system wide capacity during boot.
> 2. VFIO Call ioasid_alloc_set() when allocating vfio_mm(), set default
> quota
> 3. Adjust quota per set with ioasid_adjust_set_size() as the tunable in
> this patch.

I think this is abstraction of the allocated PASID track logic in a common
layer. It would simplify user logic.

Regards,
Yi Liu
_______________________________________________
iommu mailing list
iommu@lists.linux-foundation.org
https://lists.linuxfoundation.org/mailman/listinfo/iommu