Re: [PATCH V4 05/18] iommu/ioasid: Redefine IOASID set and allocation APIs

Hi Jason,

On Wed, 31 Mar 2021 15:33:24 -0300, Jason Gunthorpe [off-list ref] wrote:

On Wed, Mar 31, 2021 at 11:20:30AM -0700, Jacob Pan wrote:

quoted

Hi Jason,

On Wed, 31 Mar 2021 14:31:48 -0300, Jason Gunthorpe [off-list ref]
wrote: 

quoted

quoted

quoted

We should try to avoid hidden behind the scenes kernel
interconnections between subsystems.
    

 [...]  
 [...]  

quoted

yes, this is done in this patchset.
  

 [...]  

quoted

Just to clarify, you are saying (when FREE happens before proper
teardown) there is no need to proactively notify all users of the
IOASID to drop their reference. Instead, just wait for the other
parties to naturally close and drop their references. Am I
understanding you correctly?  

Yes. What are receivers going to do when you notify them anyhow? What
will a mdev do? This is how you get into they crazy locking problems.

The receivers perform cleanup work similar to normal unbind. Drain/Abort
PASID. Locking is an issue in that the atomic notifier is under IOASID
spinlock, so I provided a common ordered workqueue to let mdev drivers
queue cleanup work that cannot be done in atomic context. Not ideal. Also
need to prevent nested notifications for certain cases.

It is an error for userspace to shutdown like this, recover sensibly
and don't crash the kernel. PCIe error TLPs are expected, supress
them. That is what we decided on the mmu notifier discussion.

quoted

I feel having the notifications can add two values:
1. Shorten the duration of errors (as you mentioned below), FD close can
take a long and unpredictable time. e.g. FD shared.  

Only if userspace exits in some uncontrolled way. In a controlled exit
it can close all the FDs in the right order.

It is OK if userspace does something weird and ends up with disabled
IOASIDs. It shouldn't do that if it cares.

Agreed.

quoted

2. Provide teardown ordering among PASID users. i.e. vCPU, IOMMU, mdev.
 

This is a hard ask too, there is no natural ordering here I can see,
obviously we want vcpu, mdev, iommu for qemu but that doesn't seem to
fall out unless we explicitly hard wire it into the kernel.

The ordering problem as I understood is that it is difficult for KVM to
rendezvous all vCPUs before updating PASID translation table. So there
could be in-flight enqcmd with the stale PASID after the PASID table update
and refcount drop.

If KVM is the last one to drop the PASID refcount, the PASID could be
immediately reused and starts a new life. The in-flight enqcmd with the
stale PASID could cause problems. The likelihood and window is very small.

If we ensure KVM does PASID table update before IOMMU and mdev driver, the
stale PASID in the in-flight enqcmd would be be drained before starting
a new life.

Perhaps Yi and Kevin can explain this better.

Doesn't kvm always kill the vCPU first based on the mmu notifier
shooting down all the memory? IIRC this happens before FD close?

I don't know the answer, Kevin & Yi?

quoted

quoted

The duration between unmapping the ioasid and releasing all HW access
will have HW see PCIE TLP errors due to the blocked access. If
userspace messes up the order it is fine to cause this. We already had
this dicussion when talking about how to deal with process exit in the
simple SVA case.  

Yes, we have disabled fault reporting during this period. The slight
differences vs. the simple SVA case is that KVM is also involved and
there might be an ordering requirement to stop vCPU first.  

KVM can continue to use the PASIDs, they are parked and DMA is
permanently blocked. When KVM reaches a natural point in its teardown
it can release them.

If you have to stop the vcpu from a iommu notifier you are in the
crazy locking world I mentioned. IMHO don't create exciting locking
problems just to avoid PCI errors in uncontrolled shutdown.

Suppress the errors instead.

I agree, this simplify things a lot. Just need to clarify the in-flight
enqcmd case.

Jason

Thanks,

Jacob