Re: [Qemu-devel] QEMU and vIOMMU support for emulated VF passthrough to nested (L2) VM

[Elijah Shakkour]

> -----Original Message-----
> From: Peter Xu <address@hidden>
> Sent: Wednesday, March 27, 2019 10:43 AM
> To: Knut Omang <address@hidden>
> Cc: Elijah Shakkour <address@hidden>; Michael S. Tsirkin
> <address@hidden>; Alex Williamson <address@hidden>;
> Marcel Apfelbaum <address@hidden>; Stefan Hajnoczi
> <address@hidden>; address@hidden
> Subject: Re: QEMU and vIOMMU support for emulated VF passthrough to
> nested (L2) VM
> 
> On Wed, Mar 27, 2019 at 08:57:56AM +0100, Knut Omang wrote:
> > On Wed, 2019-03-27 at 14:41 +0800, Peter Xu wrote:
> > > On Tue, Mar 26, 2019 at 01:23:12PM +0000, Elijah Shakkour wrote:
> > > > Adding QEMU-devel
> > >
> > > Hi, Elijah,
> > >
> > > >
> > > > -----Original Message-----
> > > > From: Michael S. Tsirkin <address@hidden>
> > > > Sent: Tuesday, March 26, 2019 2:53 PM
> > > > To: Elijah Shakkour <address@hidden>
> > > > Cc: Knut Omang <address@hidden>; Alex Williamson
> > > > <address@hidden>;
> > > Marcel Apfelbaum <address@hidden>; Stefan Hajnoczi
> > > <address@hidden>; address@hidden
> > > > Subject: Re: QEMU and vIOMMU support for emulated VF passthrough
> > > > to nested (L2) VM
> > > >
> > > > I think you forgot to copy the qemu mailing list.
> > > >
> > > > On Tue, Mar 26, 2019 at 10:08:17AM +0000, Elijah Shakkour wrote:
> > > > > My questions are:
> > > > >
> > > > > - Suppose that there is an emulated NIC that supports SRIOV (I
> > > > > implemented such a
> > > NIC), now does QEMU support a scenario of an emulated NIC that
> > > supports SRIOV in Hyper-V
> > > L1 guest, that invokes VF and pass it to nested linux L2 guest?
> > >
> > > I am not an expert of SR-IOV but I can't see a limitation to not
> > > allow that to happen.
> > >
> > > > > - I'm using vIOMMU in L1, so what is needed to be done in QEMU
> > > > > or maybe in emulated
> > > NIC PF/VF to allow DMA remapping and INT remapping work as
> expected?
> > >
> > > Your below command line should work, and even it seems to be an
> > > overkill.

I didn't have DMA nor MMIO read/write working with my old command line.
But, when I removed all CPU flags and only provided "-cpu host", I see that 
MMIO works.
Still, DMA read/write from emulated device doesn't work for VF. For example:
Driver provides me a buffer pointer through MMIO write, this address (pointer) 
is GPA of L2, and when I try to call pci_dma_read() with this address I get:
"
Unassigned mem read 0000000000000000
"
I expected this address to be translated correctly in the vIOMMU.
Is there something that I'm missing here?

> > >
> > > If your device is completely emulated, IIUC you only simply need
> > > this on the latest QEMU:
> > >
> > >   -M q35 -device intel-iommu
> > >
> > > Split-irqchip and IR is on by default now, so you'll naturally gain
> > > x2apic if it's supported.  You can use x-aw-bits but only if you
> > > really need address space beyond 39 bits (which I suspect).  The
> > > rest parameters are optional too.

What do you mean by "completely emulated"?
I think that, for the sake of our talk, all I need is configuration space and 
MMIO call backs, and MSIX, PCIe capability in both PF and VF and SRIOV 
capability in PF.

> > >
> > > > > - Does the command line below -that I use to run QEMU- seem ok
> > > > > for the scenario I
> > > described to work?
> > >
> > > Before I look into details of the cmdline - I'd say MMIO in L2
> > > should have nothing to do with IOMMU...
> >
> > The addresses used in L2 is the GPAs of the L2, which would typically
> > be different from the L2 HPAs == L1 GPAs, so I think the IOMMU mappings
> must work.
> >
> > You would need something like 'intel_iommu=on iommu=pt' as boot
> parameters for L1.
> 
> Yes, the IOMMU must work to do the assignment.  What I meant was that
> IOMMU should not be in the code path of MMIO accesses even for L2.
> IIUC that's the processor who reads/writes to the memory region and if it's a
> MMIO issue then it probably has little to do with IOMMU.
> 
> Thanks,

As I said, MMIO works, but when I try to do DMA from VF device to GPA in L2 it 
fails.
I expected this DMA read to work fine without any changes to be done.

> 
> >
> > > Are you sure the MMIO traps are
> > > setup correctly?  Can the VF do IO properly even without L2?

The answer is Yes for both questions.

> >
> > I agree with Peter that just running the VF as another function in L1
> > would be good to test before trying to get L2 passthrough to work.
> >
> > I recommend you also verify that passing the PF through works as
> > expected, unless you already have done so.
> >
> > And do you see correct BAR address values in the lspci -vvv output in
> > the L2 instance?
> >
> > The SR/IOV logic is from QEMU's perspective just another device
> > instance apart from the differences in the BAR setup code, so if
> > passing through a non-virtual device works, and VF BAR addresses
> > appear right, I believe VFs should work as well.
> >
> > > Also I don't know whether there can be some tricks when you boot L2
> > > with vfio-pci when the device to assign is a VF.
> >
> > A lot has happened since I was actively using the SR/IOV patch set
> > myself so that might entirely be possible from my perspective.
> >
> > Thanks,
> > Knut
> >

As I said, my problem now is in translation of L2 GPA provided by driver, when 
I call DMA read/write for this address from VF.
Any insights?

> > > > >
> > > > > -----Original Message-----
> > > > > From: Michael S. Tsirkin <address@hidden>
> > > > > Sent: Monday, March 25, 2019 4:14 AM
> > > > > To: Elijah Shakkour <address@hidden>
> > > > > Cc: Knut Omang <address@hidden>; Alex Williamson
> > > > > <address@hidden>; Marcel Apfelbaum
> > > > > <address@hidden>; Stefan Hajnoczi
> > > > > <address@hidden>
> > > > > Subject: Re: QEMU and vIOMMU support for emulated VF
> passthrough
> > > > > to nested (L2) VM
> > > > >
> > > > > Pls post all questions on list.
> > > > > I have a policy against answering off-list mail.
> > > > > Cc Pter Xu might be a good idea, too.
> > > > >
> > > > > On Sun, Mar 24, 2019 at 09:56:26PM +0000, Elijah Shakkour wrote:
> > > > > > Hey,
> > > > > >
> > > > > > I'm emulating Mellanox ConnectX-4 in QEMU and right now, I'm
> > > > > > adding SRIOV
> > > capability.
> > > > > > I'm using Knut Omang SRIOV patches rebased to QEMU v2.12.
> > > > > > My server (L0) is Linux. L1 guest is Windows2016 Hyper-V and
> > > > > > L2 guest is Linux
> > > RH7.2.
> > > > > > I can see my device in L1 VM and I see the invocation of the
> > > > > > VF via SRIOV
> > > capability.
> > > > > > Inside L2 guest I see the virtual function in "lspci' command.
> > > > > > But when driver of L2 guest issues MMIO read/write, my MMIO ops
> don't get called.
> > > > > > I implemented my VF basically like Omang SRIOV example patch.
> > > > > >
> > > > > > Could you please shed some light on what you think I might be
> missing?
> > > > > >
> > > > > > Here is the command line I run:
> > > > > >
> > > > > > ./x86_64-softmmu/qemu-system-x86_64 \  -machine
> > > > > > q35,accel=kvm,usb=off,dump-guest-core=off,kernel-irqchip=split
> > > > > > \  -m 32G \  -smp 2 \  -enable-kvm \  -cpu
> > > > > >
> host,vmx=on,ss=on,cx16=on,x2apic=on,hypervisor=on,lahf_lm=on,h
> > > > > > v_time ,h v_relaxed,hv_vapic,hv_spinlocks=0x1fff,kvm=on \
> > > > > > -vnc 127.0.0.1:0,to=99,id=default \  -drive
> > > > > > file=$IMAGE,format=qcow2,if=none,id=drive-sata0-0-0 \
> > > > > > -chardev
> > > > > > pty,id=charserial0 \  -device
> > > > > > intel-iommu,intremap=on,caching-mode=on,device-
> iotlb=on,eim=on
> > > > > > ,x-aw-
> > > > > > bi
> > > > > > ts=48 \  -device
> > > > > > ide-hd,bus=ide.0,drive=drive-sata0-0-0,id=sata0-0-0,bootindex=
> > > > > > 0 \ -device
> > > > > > pcie-root-port,pref64-reserve=500M,slot=0,id=pcie_port.1,bus=p
> > > > > > cie.0,
> > > > > > mu
> > > > > > ltifunction=on \  -netdev
> > > > > > tap,id=tap5,ifname=tap5,script=no,downscript=no \  -device
> > > > > > connectx4,netdev=tap5,bus=pcie_port.1,multifunction=on
> > >
> > > Regards,
> > >
> >
> 
> --
> Peter Xu