Re: [Qemu-devel] intel-iommu and vhost: Do we need 'device-iotlb' and 'ats'?

[Peter Xu]

On Fri, Feb 23, 2018 at 06:34:04AM +0000, Jintack Lim wrote:
> On Fri, Feb 23, 2018 at 1:10 AM Peter Xu <address@hidden> wrote:
> 
> > On Fri, Feb 23, 2018 at 12:32:13AM -0500, Jintack Lim wrote:
> > > Hi Peter,
> > >
> > > Hope you had great holidays!
> > >
> > > On Thu, Feb 22, 2018 at 10:55 PM, Peter Xu <address@hidden> wrote:
> > > > On Tue, Feb 20, 2018 at 11:03:46PM -0500, Jintack Lim wrote:
> > > >> Hi,
> > > >>
> > > >> I'm using vhost with the virtual intel-iommu, and this page[1] shows
> > > >> the QEMU command line example.
> > > >>
> > > >> qemu-system-x86_64 -M q35,accel=kvm,kernel-irqchip=split -m 2G \
> > > >>                    -device intel-iommu,intremap=on,device-iotlb=on \
> > > >>                    -device ioh3420,id=pcie.1,chassis=1 \
> > > >>                    -device
> > > >>
> > virtio-net-pci,bus=pcie.1,netdev=net0,disable-legacy=on,disable-modern=off,iommu_platform=on,ats=on
> > > >> \
> > > >>                    -netdev tap,id=net0,vhostforce \
> > > >>                    $IMAGE_PATH
> > > >>
> > > >> I wonder what's the impact of using device-iotlb and ats options as
> > > >> they are described necessary.
> > > >>
> > > >> In my understanding, vhost in the kernel only looks at
> > > >> VIRTIO_F_IOMMU_PLATFORM, and when it is set, vhost uses a
> > > >> device-iotlb. In addition, vhost and QEMU communicate using vhost_msg
> > > >> basically to cache mappings correctly in the vhost, so I wonder what's
> > > >> the role of ats in this case.
> > > >
> > > > The "ats" as virtio device parameter will add ATS capability to the
> > > > PCI device.
> > > >
> > > > The "device-iotlb" as intel-iommu parameter will enable ATS in the
> > > > IOMMU device (and also report that in ACPI field).
> > > >
> > > > If both parameters are provided IIUC it means guest will know virtio
> > > > device has device-iotlb and it'll treat the device specially (e.g.,
> > > > guest will need to send device-iotlb invalidations).
> > >
> > > Oh, I see. I was focusing on how QEMU and vhost work in the host, but
> > > I think I missed the guest part! Thanks. I see that the Intel IOMMU
> > > driver has has_iotlb_device flag for that purpose.
> > >
> > > >
> > > > We'd better keep these parameters when running virtio devices with
> > > > vIOMMU.  For the rest of vhost/arm specific questions, I'll leave to
> > > > others.
> > >
> > > It seems like SMMU is not checking ATS capability - at least
> > > ats_enabled flag - but I may miss something here as well :)
> > >
> > > >
> > > > PS: Though IIUC the whole ATS thing may not really be necessary for
> > > > current VT-d emulation, since even with ATS vhost is registering UNMAP
> > > > IOMMU notifiers (see vhost_iommu_region_add()), and IIUC that means
> > > > vhost will receive IOTLB invalidations even without ATS support, and
> > > > it _might_ still work.
> > >
> > > Right. That's what I thought.
> > >
> > > Come to think of it, I'm not sure why we need to flush mappings in
> > > IOMMU and devices separately in the first place... Any thoughts?
> >
> > I don't know ATS much, neither.
> >
> > You can have a look at chap 4 of vt-d spec:
> >
> >         One approach to scaling IOTLBs is to enable I/O devices to
> >         participate in the DMA remapping with IOTLBs implemented at
> >         the devices. The Device-IOTLBs alleviate pressure for IOTLB
> >         resources in the core logic, and provide opportunities for
> >         devices to improve performance by pre-fetching address
> >         translations before issuing DMA requests. This may be useful
> >         for devices with strict DMA latency requirements (such as
> >         isochronous devices), and for devices that have large DMA
> >         working set or multiple active DMA streams.
> >
> > So I think it's for performance's sake. For example, the DMA operation
> > won't need to be translated at all if it's pre-translated, so it can
> > have less latency.  And also, that'll offload some of the translation
> > process so that workload can be more distributed.
> >
> > When with that (caches located both on IOMMU's and device's side), we
> > need to invalidate all the cache when needed.
> >
> 
> Right. I think my question was not clear. My question was that why don’t
> IOMMU invalidate device-iotlb along with its mappings in one go. Then IOMMU
> device driver doesn’t need to flush device-iotlb explicitly. Maybe the
> reason is that ATS and IOMMU are not always coupled.. but I guess it’s time
> for me to get some more background :)

Ah, I see your point.

I don't know the answer.  My wild guess is that IOMMU is just trying
to be simple and only provide most basic functionalities, leaving
complex stuff to CPU.  For example, if IOMMU takes over the ownership
to deliever device-iotlb invalidations when receiving iotlb
invalidations, it possibly needs to traverse the device tree sometimes
(e.g., for domain invalidations) to know what device is under what
domain, which is really compliated.  While it'll be simpler for CPU to
do this since it's very possible that the OS keeps a list of devices
for a domain already.

IMHO that follows the *nix philosophy too - Do One Thing And Do It
Well.  Though again, it's wild guess and I may be wrong. :)

CCing Alex, in case he has quick answers.

> 
> 
> > >
> > > Your reply was really helpful to me. I appreciate it.
> >
> > My pleasure.  Thanks,
> >
> > >
> > > Thanks,
> > > Jintack
> > >
> > > > But there can be other differences, like
> > > > performance, etc.
> > > >
> > > >>
> > > >> A related question is that if we use SMMU emulation[2] on ARM without
> > > >> those options, does vhost cache mappings as if it has a device-iotlb?
> > > >> (I guess this is the case.)
> > > >>
> > > >> I'm pretty new to QEMU code, so I might be missing something. Can
> > > >> somebody shed some light on it?
> > > >>
> > > >> [1] https://wiki.qemu.org/Features/VT-d
> > > >> [2]
> > http://lists.nongnu.org/archive/html/qemu-devel/2018-02/msg04736.html
> > > >>
> > > >> Thanks,
> > > >> Jintack
> > > >>
> > > >
> > > > --
> > > > Peter Xu
> > > >
> > >
> >
> > --
> > Peter Xu
> >
> >

-- 
Peter Xu