Re: [Qemu-devel] iommu emulation

[Peter Xu <peterx@redhat.com>]

On Thu, Feb 09, 2017 at 08:01:14AM -0500, Jintack Lim wrote:
> On Wed, Feb 8, 2017 at 10:52 PM, Peter Xu <peterx@redhat.com> wrote:
> > (cc qemu-devel and Alex)
> >
> > On Wed, Feb 08, 2017 at 09:14:03PM -0500, Jintack Lim wrote:
> >> On Wed, Feb 8, 2017 at 10:49 AM, Jintack Lim <jintack@cs.columbia.edu> wrote:
> >> > Hi Peter,
> >> >
> >> > On Tue, Feb 7, 2017 at 10:12 PM, Peter Xu <peterx@redhat.com> wrote:
> >> >> On Tue, Feb 07, 2017 at 02:16:29PM -0500, Jintack Lim wrote:
> >> >>> Hi Peter and Michael,
> >> >>
> >> >> Hi, Jintack,
> >> >>
> >> >>>
> >> >>> I would like to get some help to run a VM with the emulated iommu. I
> >> >>> have tried for a few days to make it work, but I couldn't.
> >> >>>
> >> >>> What I want to do eventually is to assign a network device to the
> >> >>> nested VM so that I can measure the performance of applications
> >> >>> running in the nested VM.
> >> >>
> >> >> Good to know that you are going to use [4] to do something useful. :-)
> >> >>
> >> >> However, could I ask why you want to measure the performance of
> >> >> application inside nested VM rather than host? That's something I am
> >> >> just curious about, considering that virtualization stack will
> >> >> definitely introduce overhead along the way, and I don't know whether
> >> >> that'll affect your measurement to the application.
> >> >
> >> > I have added nested virtualization support to KVM/ARM, which is under
> >> > review now. I found that application performance running inside the
> >> > nested VM is really bad both on ARM and x86, and I'm trying to figure
> >> > out what's the real overhead. I think one way to figure that out is to
> >> > see if the direct device assignment to L2 helps to reduce the overhead
> >> > or not.
> >
> > I see. IIUC you are trying to use an assigned device to replace your
> > old emulated device in L2 guest to see whether performance will drop
> > as well, right? Then at least I can know that you won't need a nested
> > VT-d here (so we should not need a vIOMMU in L2 guest).
> 
> That's right.
> 
> >
> > In that case, I think we can give it a shot, considering that L1 guest
> > will use vfio-pci for that assigned device as well, and when L2 guest
> > QEMU uses this assigned device, it'll use a static mapping (just to
> > map the whole GPA for L2 guest) there, so even if you are using a
> > kernel driver in L2 guest with your to-be-tested application, we
> > should still be having a static mapping in vIOMMU in L1 guest, which
> > is IMHO fine from performance POV.
> >
> > I cced Alex in case I missed anything here.
> >
> >> >
> >> >>
> >> >> Another thing to mention is that (in case you don't know that), device
> >> >> assignment with VT-d protection would be even slower than generic VMs
> >> >> (without Intel IOMMU protection) if you are using generic kernel
> >> >> drivers in the guest, since we may need real-time DMA translation on
> >> >> data path.
> >> >>
> >> >
> >> > So, this is the comparison between using virtio and using the device
> >> > assignment for L1? I have tested application performance running
> >> > inside L1 with and without iommu, and I found that the performance is
> >> > better with iommu.

Here iiuc you mean that "L1 guest with vIOMMU performs better than
when without vIOMMU", while ...

> >> > I thought whether the device is assigned to L1 or
> >> > L2, the DMA translation is done by iommu, which is pretty fast? Maybe
> >> > I misunderstood what you said?
> >
> > I failed to understand why an vIOMMU could help boost performance. :(
> > Could you provide your command line here so that I can try to
> > reproduce?
> 
> Sure. This is the command line to launch L1 VM
> 
> qemu-system-x86_64 -M q35,accel=kvm,kernel-irqchip=split \
> -m 12G -device intel-iommu,intremap=on,eim=off,caching-mode=on \
> -drive file=/mydata/guest0.img,format=raw --nographic -cpu host \
> -smp 4,sockets=4,cores=1,threads=1 \
> -device vfio-pci,host=08:00.0,id=net0
> 
> And this is for L2 VM.
> 
> ./qemu-system-x86_64 -M q35,accel=kvm \
> -m 8G \
> -drive file=/vm/l2guest.img,format=raw --nographic -cpu host \
> -device vfio-pci,host=00:03.0,id=net0

... here looks like these are command lines for L1/L2 guest, rather
than L1 guest with/without vIOMMU?

> 
> >
> > Besides, what I mentioned above is just in case you don't know that
> > vIOMMU will drag down the performance in most cases.
> >
> > I think here to be more explicit, the overhead of vIOMMU is different
> > for assigned devices and emulated ones.
> >
> >   (1) For emulated devices, the overhead is when we do the
> >       translation, or say when we do the DMA operation. We need
> >       real-time translation which should drag down the performance.
> >
> >   (2) For assigned devices (our case), the overhead is when we setup
> >       the pages (since we are trapping the setup procedures via CM
> >       bit). However, after it's setup, we should have no much
> >       performance drag when we really do the data transfer (during
> >       DMA) since that'll all be done in the hardware IOMMU (no matter
> >       whether the device is assigned to L1/L2 guest).
> >
> > Now, after I know your use case now (use vIOMMU in L1 guest, don't use
> > vIOMMU in L2 guest, only use assigned devices), I suspect we would
> > have no big problem according to (2).
> >
> >> >
> >> >>>
> >> >>> First, I am having trouble to boot a VM with the emulated iommu. I
> >> >>> have posted my problem to the qemu user mailing list[1],
> >> >>
> >> >> Here I would suggest that you cc qemu-devel as well next time:
> >> >>
> >> >>   qemu-devel@nongnu.org
> >> >>
> >> >> Since I guess not all people are registered to qemu-discuss, at least
> >> >> I am not in that loop. Imho cc qemu-devel could let the question
> >> >> spread to more people, and it'll get a higher chance to be answered.
> >> >
> >> > Thanks. I'll cc qemu-devel next time.
> >> >
> >> >>
> >> >>> but to put it
> >> >>> in a nutshell, I'd like to know the setting I can reuse to boot a VM
> >> >>> with the emulated iommu. (e.g. how to create a VM with q35 chipset
> >> >>> and/or libvirt xml if you use virsh).
> >> >>
> >> >> IIUC you are looking for device assignment for the nested VM case. So,
> >> >> firstly, you may need my tree to run this (see below). Then, maybe you
> >> >> can try to boot a L1 guest with assigned device (under VT-d
> >> >> protection), with command:
> >> >>
> >> >> $qemu -M q35,accel=kvm,kernel-irqchip=split -m 1G \
> >> >>       -device intel-iommu,intremap=on,eim=off,caching-mode=on \
> >> >>       -device vfio-pci,host=$HOST_PCI_ADDR \
> >> >>       $YOUR_IMAGE_PATH
> >> >>
> >> >
> >> > Thanks! I'll try this right away.
> >> >
> >> >> Here $HOST_PCI_ADDR should be something like 05:00.0, which is the
> >> >> host PCI address of the device to be assigned to guest.
> >> >>
> >> >> (If you go over the cover letter in [4], you'll see similar command
> >> >>  line there, though with some more devices assigned, and with traces)
> >> >>
> >> >> If you are playing with nested VM, you'll also need a L2 guest, which
> >> >> will be run inside the L1 guest. It'll require similar command line,
> >> >> but I would suggest you first try a L2 guest without intel-iommu
> >> >> device. Frankly speaking I haven't played with that yet, so just let
> >> >> me know if you got any problem, which is possible. :-)
> >> >>
> >>
> >> I was able to boot L2 guest without assigning a network device
> >> successfully. (host iommu was on, L1 iommu was on, and the network
> >> device was assigned to L1)
> >>
> >> Then, I unbound the network device in L1 and bound it to vfio-pci.
> >> When I try to run L2 with the following command, I got an assertion.
> >>
> >> # ./qemu-system-x86_64 -M q35,accel=kvm \
> >> -m 8G \
> >> -drive file=/vm/l2guest.img,format=raw --nographic -cpu host \
> >> -device vfio-pci,host=00:03.0,id=net0
> >>
> >> qemu-system-x86_64: hw/pci/pcie.c:686: pcie_add_capability: Assertion
> >> `prev >= 0x100' failed.
> >> Aborted (core dumped)
> >>
> >> Thoughts?
> >
> > I don't know whether it'll has anything to do with how vfio-pci works,
> > anyway I cced Alex and the list in case there is quick answer.
> >
> > I'll reproduce this nested case and update when I got anything.
> 
> Thanks!

I tried to reproduce this issue with the following 10g network card:

00:03.0 Ethernet controller: Intel Corporation Ethernet Controller 10-Gigabit X540-AT2 (rev 01)

In my case, both L1/L2 guests can boot with the assigned device. I
also did a quick netperf TCP STREAM test, the result is (in case you
are interested):

   L1 guest: 1.12Gbps
   L2 guest: 8.26Gbps

First of all, just to confirm that you were using the same qemu binary
in both host and L1 guest, right?

Then, I *think* above assertion you encountered would fail only if
prev == 0 here, but I still don't quite sure why was that happening.
Btw, could you paste me your "lspci -vvv -s 00:03.0" result in your L1
guest?

Thanks,

-- peterx