[ovs-dev] [ PATCH v2] dpdk: Use DPDK 17.11.2 release.

[Stokes, Ian]

> On 05/02/2018 11:16 AM, Maxime Coquelin wrote:
> > Hi Ian,
> >
> > On 05/02/2018 11:42 AM, Stokes, Ian wrote:
> >>> On 04/25/2018 10:17 AM, Stokes, Ian wrote:
> >>>>> Modify travis linux build script to use the latest DPDK stable
> >>>>> release 17.11.2. Update docs for latest DPDK stable releases.
> >>>>>
> >>>>
> >>>> Thanks for this Kevin, I'm lining this up for the next pull request.
> >>>>
> >>>
> >>> Thanks - note, this is relevant for master and branch-2.9.
> >>>
> >>
> >> Hi Kevin,
> >>
> >> I was running validation for the this week's pull request and I
> >> spotted vhost reconnect was failing.
> >>
> >> I traced it back to the move to 17.11.2.
> >>
> >> The test case is a vm running testpmd forwarding traffic in p2p.
> >>
> >> The vswitch itself is reset so traffic stops arriving to the vm,
> >> after restarting the vswitch reconnects to the VM. At this point
> >> testpmd within the VM segfaults.
> >>
> 
> Thanks for finding it. Which version of guest DPDK and QEMU do you use?
> Also what are settings for rx mergable buffers, vIOMMU, MTU, num queues?

Sorry for the delay, wanted to do a little more testing.

DPDK 17.11.2 (on both host and guest.)
QEMU 2.9.1.
Mergable buffer off.
vIOMMU disabled.
MTU=1500.
Single queue for vhost interfaces.

For completeness the QEMU argument vsperf uses for launching the VM is below. I didn’t see anything out of place myself.

/bin/bash', '-c', 'sudo -E taskset -c 6,7 /opt/istokes/qemu/x86_64-softmmu/qemu-system-x86_64 -m 2048 -smp 2 -cpu host,migratable=off -drive if=scsi,file=/opt/istokes/vm_images/vsperf_img1.qcow2 -boot c --enable-kvm -monitor unix:/tmp/vm0monitor,server,nowait -object memory-backend-file,id=mem,size=2048M,mem-path=/dev/hugepages,share=on -numa node,memdev=mem -mem-prealloc -nographic -vnc :0 -name Client0 -snapshot -net none -no-reboot -drive if=scsi,format=raw,file=fat:rw:/tmp/qemu0_share,snapshot=off -chardev socket,id=char0,path=/usr/local/var/run/openvswitch/dpdkvhostuserclient0,server -netdev type=vhost-user,id=net1,chardev=char0,vhostforce -device virtio-net-pci,mac=00:00:00:00:00:01,netdev=net1,csum=off,mrg_rxbuf=off,gso=off,guest_tso4=off,guest_tso6=off,guest_ecn=off -chardev socket,id=char1,path=/usr/local/var/run/openvswitch/dpdkvhostuserclient1,server -netdev type=vhost-user,id=net2,chardev=char1,vhostforce -device virtio-net-pci,mac=00:00:00:00:00:02,netdev=net2,csum=off,mrg_rxbuf=off,gso=off,guest_tso4=off,guest_tso6=off,guest_ecn=off'

In testing I swapped to QEMU 2.7.1 and in that case the test passes without issue using 17.11.2 (same setup as above).

Ian
> 
> >> Doing a git bisect on DPDK 17.11.2 the issue occurs with commit
> >>
> >> vhost: fix indirect descriptors table translation size (a0d990b19)
> >
> > I would be surprised the problem comes from this patch.
> > I guess your are testing without vIOMMU enabled> If this is the case,
> > then the size parameter isn't used at the time of this patch:
> >
> > static __rte_always_inline uint64_t
> > vhost_iova_to_vva(struct virtio_net *dev, struct vhost_virtqueue *vq,
> >             uint64_t iova, uint64_t size, uint8_t perm) {
> >     if (!(dev->features & (1ULL << VIRTIO_F_IOMMU_PLATFORM)))
> >         return rte_vhost_gpa_to_vva(dev->mem, iova);
> >
> >     return __vhost_iova_to_vva(dev, vq, iova, size, perm); }
> >
> > So I fail to see how it could cause any regression.
> >
> > Regards,
> > Maxime
> >
> >> Just wondering if you see similar behavior in testing?
> >>
> >> Regards
> >> Ian
> >>
> >>