Re: virtio,iommu_platform=on

[Michael Roth]

Quoting Alexey Kardashevskiy (2019-11-11 21:53:49)
> Hi!
> 
> I am enabling IOMMU for virtio in the pseries firmware (SLOF) and seeing
> problems, one of them is SLOF does SCSI bus scan, then it stops the
> virtio-scsi by clearing MMIO|IO|BUSMASTER from PCI_COMMAND (as SLOF
> stopped using the devices) and when this happens, I see unassigned
> memory access (see below) which happens because disabling busmaster
> disables IOMMU and QEMU cannot access the rings to do some shutdown. And
> when this happens, the device does not come back even if SLOF re-enables it.
> 
> Hacking SLOF to not clear BUSMASTER makes virtio-scsi work but it is
> hardly a right fix.

I hit the same issue enabling IOMMU for virtio-blk using this branch:

  https://github.com/mdroth/SLOF/commits/virtio-iommu

I just sent a tentative fix for QEMU as:

  "virtio-pci: disable vring processing when bus-mastering is disabled"

It's an RFC since piggy-backing off dev->broken seems a bit hacky, but
it seems to fix the issue at least.

BTW, the SLOF branch above needs cleanup, but it's booting guests okay
and I was planning to post this week. Where are you at on yours? Maybe
we should sync up...