Re: [PATCH 1/2] iommu: Add iommu_group_get/set_domain()

From: Robin Murphy
Date: Tue Jun 30 2020 - 12:51:37 EST


On 2020-06-30 02:03, Lu Baolu wrote:
Hi Robin,

On 6/29/20 7:56 PM, Robin Murphy wrote:
On 2020-06-27 04:15, Lu Baolu wrote:
The hardware assistant vfio mediated device is a use case of iommu
aux-domain. The interactions between vfio/mdev and iommu during mdev
creation and passthr are:

- Create a group for mdev with iommu_group_alloc();
- Add the device to the group with
ÂÂÂÂÂÂÂÂ group = iommu_group_alloc();
ÂÂÂÂÂÂÂÂ if (IS_ERR(group))
ÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂ return PTR_ERR(group);

ÂÂÂÂÂÂÂÂ ret = iommu_group_add_device(group, &mdev->dev);
ÂÂÂÂÂÂÂÂ if (!ret)
ÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂ dev_info(&mdev->dev, "MDEV: group_id = %d\n",
ÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂ iommu_group_id(group));
- Allocate an aux-domain
ÂÂÂÂiommu_domain_alloc()
- Attach the aux-domain to the physical device from which the mdev is
ÂÂ created.
ÂÂÂÂiommu_aux_attach_device()

In the whole process, an iommu group was allocated for the mdev and an
iommu domain was attached to the group, but the group->domain leaves
NULL. As the result, iommu_get_domain_for_dev() doesn't work anymore.

This adds iommu_group_get/set_domain() so that group->domain could be
managed whenever a domain is attached or detached through the aux-domain
api's.

Letting external callers poke around directly in the internals of iommu_group doesn't look right to me.

Unfortunately, it seems that the vifo iommu abstraction is deeply bound
to the IOMMU subsystem. We can easily find other examples:

iommu_group_get/set_iommudata()
iommu_group_get/set_name()
...

Sure, but those are ways for users of a group to attach useful information of their own to it, that doesn't matter to the IOMMU subsystem itself. The interface you've proposed gives callers rich new opportunities to fundamentally break correct operation of the API:

dom = iommu_domain_alloc();
iommu_attach_group(dom, grp);
...
iommu_group_set_domain(grp, NULL);
// oops, leaked and can't ever detach properly now

or perhaps:

grp = iommu_group_alloc();
iommu_group_add_device(grp, dev);
iommu_group_set_domain(grp, dom);
...
iommu_detach_group(dom, grp);
// oops, IOMMU driver might not handle this

If a regular device is attached to one or more aux domains for PASID use, iommu_get_domain_for_dev() is still going to return the primary domain, so why should it be expected to behave differently for mediated

Unlike the normal device attach, we will encounter two devices when it
comes to aux-domain.

- Parent physical device - this might be, for example, a PCIe device
with PASID feature support, hence it is able to tag an unique PASID
for DMA transfers originated from its subset. The device driver hence
is able to wrapper this subset into an isolated:

- Mediated device - a fake device created by the device driver mentioned
above.

Yes. All you mentioned are right for the parent device. But for mediated
device, iommu_get_domain_for_dev() doesn't work even it has an valid
iommu_group and iommu_domain.

iommu_get_domain_for_dev() is a necessary interface for device drivers
which want to support aux-domain. For example,

Only if they want to follow this very specific notion of using made-up devices and groups to represent aux attachments. Even if a driver managing its own aux domains entirely privately does create child devices for them, it's not like it can't keep its domain pointers in drvdata if it wants to ;)

Let's not conflate the current implementation of vfio_mdev with the general concepts involved here.

ÂÂÂÂÂÂÂÂÂ struct iommu_domain *domain;
ÂÂÂÂÂÂÂÂÂ struct device *dev = mdev_dev(mdev);
ÂÂÂÂÂ unsigned long pasid;

ÂÂÂÂÂÂÂÂÂ domain = iommu_get_domain_for_dev(dev);
ÂÂÂÂÂÂÂÂÂ if (!domain)
ÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂÂ return -ENODEV;

ÂÂÂÂÂÂÂÂÂ pasid = iommu_aux_get_pasid(domain, dev->parent);
ÂÂÂÂÂ if (pasid == IOASID_INVALID)
ÂÂÂÂÂÂÂÂÂ return -EINVAL;

ÂÂÂÂÂ /* Program the device context with the PASID value */
ÂÂÂÂÂ ....

Without this fix, iommu_get_domain_for_dev() always returns NULL and the
device driver has no means to support aux-domain.

So either the IOMMU API itself is missing the ability to do the right thing internally, or the mdev layer isn't using it appropriately. Either way, simply punching holes in the API for mdev to hack around its own mess doesn't seem like the best thing to do.

The initial impression I got was that it's implicitly assumed here that the mdev itself is attached to exactly one aux domain and nothing else, at which point I would wonder why it's using aux at all, but are you saying that in fact no attach happens with the mdev group either way, only to the parent device?

I'll admit I'm not hugely familiar with any of this, but it seems to me that the logical flow should be:

- allocate domain
- attach as aux to parent
- retrieve aux domain PASID
- create mdev child based on PASID
- attach mdev to domain (normally)

Of course that might require giving the IOMMU API a proper first-class notion of mediated devices, such that it knows the mdev represents the PASID, and can recognise the mdev attach is equivalent to the earlier parent aux attach so not just blindly hand it down to an IOMMU driver that's never heard of this new device before. Or perhaps the IOMMU drivers do their own bookkeeping for the mdev bus, such that they do handle the attach call, and just validate it internally based on the associated parent device and PASID. Either way, the inside maintains self-consistency and from the outside it looks like standard API usage without nasty hacks.

I'm pretty sure I've heard suggestions of using mediated devices beyond VFIO (e.g. within the kernel itself), so chances are this is a direction that we'll have to take at some point anyway.

And, that said, even if people do want an immediate quick fix regardless of technical debt, I'd still be a lot happier to see iommu_group_set_domain() lightly respun as iommu_attach_mdev() ;)

Robin.