Virtualization - Jun 2023 - [RFC PATCHES 00/17] IOMMUFD: Deliver IO page faults to user space

On Mon, Jun 19, 2023 at 11:35:50AM +0800, Baolu Lu
wrote:
> > Another outstanding issue was what to do for PASID stop. When the
guest
> > device driver stops using a PASID it issues a PASID stop request to
the
> > device (a device-specific mechanism). If the device is not using PRI
stop
> > markers it waits for pending PRs to complete and we're fine.
Otherwise it
> > sends a stop marker which is flushed to the PRI queue, but does not
wait
> > for pending PRs.
> > 
> > Handling stop markers is annoying. If the device issues one, then the
PRI
> > queue contains stale faults, a stop marker, followed by valid faults
for
> > the next address space bound to this PASID. The next address space
will
> > get all the spurious faults because the fault handler doesn't know
that
> > there is a stop marker coming. Linux is probably alright with spurious
> > faults, though maybe not in all cases, and other guests may not
support
> > them at all.
> > 
> > We might need to revisit supporting stop markers: request that each
device
> > driver declares whether their device uses stop markers on unbind()
("This
> > mechanism must indicate that a Stop Marker Message will be
generated."
> > says the spec, but doesn't say if the function always uses one or
the
> > other mechanism so it's per-unbind). Then we still have to
synchronize
> > unbind() with the fault handler to deal with the pending stop marker,
> > which might have already gone through or be generated later.
> 
> I don't quite follow here. Once a PASID is unbound from the device, the
> device driver should be free to release the PASID. The PASID could then
> be used for any other purpose. The device driver has no idea when the
> pending page requests are flushed after unbind(), so it cannot decide
> how long should the PASID be delayed for reuse. Therefore, I understand
> that a successful return from the unbind() function denotes that all
> pending page requests have been flushed and the PASID is viable for
> other use.
Yes that's the contract for unbind() at the moment
> 
> > 
> > Currently we ignore all that and just flush the PRI queue, followed by
the
> > IOPF queue, to get rid of any stale fault before reassigning the
PASID. A
> > guest however would also need to first flush the HW PRI queue, but
doesn't
> > have a direct way to do that. If we want to support guests that
don't deal
> > with stop markers, the host needs to flush the PRI queue when a PASID
is
> > detached. I guess on Intel detaching the PASID goes through the host
which
> > can flush the host queue. On Arm we'll probably need to flush the
queue
> > when receiving a PASID cache invalidation, which the guest issues
after
> > clearing a PASID table entry.
> 
> The Intel VT-d driver follows below steps to drain pending page requests
> when a PASID is unbound from a device.
> 
> - Tear down the device's pasid table entry for the stopped pasid.
>   This ensures that ATS/PRI will stop putting more page requests for the
>   pasid in VT-d PRQ.
Oh that's interesting, I didn't know about the implicit TLB
invalidations
on page requests for VT-d.

For Arm SMMU, clearing the PASID table entry does cause ATS Translation
Requests to return with Completer Abort, but does not affect PRI. The SMMU
pushes page requests directly into the PRI queue without reading any table
(unless the queue overflows).

We're counting on the device driver to perform the PASID stop request
before calling unbind(), described in PCIe 6.20.1 (Managing PASID Usage)
and 10.4.1.2 (Managing PASID Usage on PRG Requests). This ensures that
when unbind() is called, no more page request for the PASID is pushed into
the PRI queue. But some may still be in the queue if the device uses stop
markers.
> - Sync with the PRQ handling thread until all related page requests in
>   PRQ have been delivered.
This is what I'm concerned about. For VT-d this happens in the host which
is in charge of modifying the PASID table. For SMMU, the guest writes the
PASID table. It flushes its virtual PRI queue, but not the physical queue
that is managed by the host.

One synchronization point where the host could flush the physical PRI
queue is the PASID config invalidation (CMD_CFGI_CD). As Jason pointed out
the host may not be able to observe those if a command queue is assigned
directly to the guest (a theoretical SMMU extension), though in that case
the guest may also have direct access to a PRI queue (like the AMD vIOMMU
extension) and be able to flush the queue directly.

But we can just wait for PRI implementations and see what the drivers
need. Maybe no device will implement stop markers.

Thanks,
Jean
> - Flush the iopf queue with iopf_queue_flush_dev().
> - Follow the steps defined in VT-d spec section 7.10 to drain all page
>   requests and responses between VT-d and the endpoint device.