Re: [PATCH] vhost/vsock: add IOTLB API support

[Stefano Garzarella]

On Tue, Nov 03, 2020 at 02:46:13PM -0500, Peter Xu wrote:
> On Tue, Nov 03, 2020 at 05:04:23PM +0800, Jason Wang wrote:
> > On 2020/11/3 上午1:11, Stefano Garzarella wrote:
> > > On Fri, Oct 30, 2020 at 07:44:43PM +0800, Jason Wang wrote:
> > > >
> > > > On 2020/10/30 下午6:54, Stefano Garzarella wrote:
> > > > > On Fri, Oct 30, 2020 at 06:02:18PM +0800, Jason Wang wrote:
> > > > > >
> > > > > > On 2020/10/30 上午1:43, Stefano Garzarella wrote:
> > > > > > > This patch enables the IOTLB API support for vhost-vsock devices,
> > > > > > > allowing the userspace to emulate an IOMMU for the guest.
> > > > > > >
> > > > > > > These changes were made following vhost-net, in details this patch:
> > > > > > > - exposes VIRTIO_F_ACCESS_PLATFORM feature and inits the iotlb
> > > > > > >   device if the feature is acked
> > > > > > > - implements VHOST_GET_BACKEND_FEATURES and
> > > > > > >   VHOST_SET_BACKEND_FEATURES ioctls
> > > > > > > - calls vq_meta_prefetch() before vq processing to prefetch vq
> > > > > > >   metadata address in IOTLB
> > > > > > > - provides .read_iter, .write_iter, and .poll callbacks for the
> > > > > > >   chardev; they are used by the userspace to exchange IOTLB messages
> > > > > > >
> > > > > > > This patch was tested with QEMU and a patch applied [1] to fix a
> > > > > > > simple issue:
> > > > > > >     $ qemu -M q35,accel=kvm,kernel-irqchip=split \
> > > > > > >            -drive file=fedora.qcow2,format=qcow2,if=virtio \
> > > > > > >            -device intel-iommu,intremap=on \
> > > > > > >            -device vhost-vsock-pci,guest-cid=3,iommu_platform=on
> > > > > >
> > > > > >
> > > > > > Patch looks good, but a question:
> > > > > >
> > > > > > It looks to me you don't enable ATS which means vhost won't
> > > > > > get any invalidation request or did I miss anything?
> > > > > >
> > > > >
> > > > > You're right, I didn't see invalidation requests, only miss and
> > > > > updates.
> > > > > Now I have tried to enable 'ats' and 'device-iotlb' but I still
> > > > > don't see any invalidation.
> > > > >
> > > > > How can I test it? (Sorry but I don't have much experience yet
> > > > > with vIOMMU)
> > > >
> > > >
> > > > I guess it's because the batched unmap. Maybe you can try to use
> > > > "intel_iommu=strict" in guest kernel command line to see if it
> > > > works.
> > > >
> > > > Btw, make sure the qemu contains the patch [1]. Otherwise ATS won't
> > > > be enabled for recent Linux Kernel in the guest.
> > >
> > > The problem was my kernel, it was built with a tiny configuration.
> > > Using fedora stock kernel I can see the 'invalidate' requests, but I
> > > also had the following issues.
> > >
> > > Do they make you ring any bells?
> > >
> > > $ ./qemu -m 4G -smp 4 -M q35,accel=kvm,kernel-irqchip=split \
> > >     -drive file=fedora.qcow2,format=qcow2,if=virtio \
> > >     -device intel-iommu,intremap=on,device-iotlb=on \
> > >     -device vhost-vsock-pci,guest-cid=6,iommu_platform=on,ats=on,id=v1
> > >
> > >     qemu-system-x86_64: vtd_iova_to_slpte: detected IOVA overflow    
> > > (iova=0x1d40000030c0)
> >
> >
> > It's a hint that IOVA exceeds the AW. It might be worth to check whether the
> > missed IOVA reported from IOTLB is legal.
>
> Yeah.  By default the QEMU vIOMMU should only support 39bits width for guest
> iova address space.  To extend it, we can use:
>
>  -device intel-iommu,aw-bits=48
>
> So we'll enable 4-level iommu pgtable.
>
> Here the iova is obvious longer than this, so it'll be interesting to know why
> that iova is allocated in the guest driver since the driver should know somehow
> that this iova is beyond what's supported (guest iommu driver should be able to
> probe viommu capability on this width information too).

Peter, Jason, thanks for the hints!

I'll try to understand what is going on in the guest driver.

Stefano