mlx5_vdpa problem

[wenxu <wenxu@ucloud.cn>]

Hi mellanox team,


I test the mlx5 vdpa  in linux-5.9 and meet several problem.


# lspci | grep Ether | grep Dx
b3:00.0 Ethernet controller: Mellanox Technologies MT2892 Family [ConnectX-6 Dx]
b3:00.1 Ethernet controller: Mellanox Technologies MT2892 Family [ConnectX-6 Dx]

# ethtool -i net2
driver: mlx5e_rep
version: 5.9.0
firmware-version: 22.28.1002 (MT_0000000430)
expansion-rom-version:
bus-info: 0000:b3:00.0
supports-statistics: yes
supports-test: no
supports-eeprom-access: no
supports-register-dump: no
supports-priv-flags: no


init switchdev:


# echo 1 > /sys/class/net/net2/device/sriov_numvfs
# echo 0000:b3:00.2 > /sys/bus/pci/drivers/mlx5_core/unbind
# devlink dev eswitch set pci/0000:b3:00.0  mode switchdev encap enable

# modprobe vdpa vhost-vdpa mlx5_vdpa

# ip l set dev net2 vf 0 mac 52:90:01:00:02:13
# echo 0000:b3:00.2 > /sys/bus/pci/drivers/mlx5_core/bind


setup vm:

# qemu-system-x86_64 -name test  -enable-kvm -smp 16,sockets=2,cores=8,threads=1 -m 8192 -drive file=./centos7.qcow2,format=qcow2,if=none,id=drive-virtio-disk0 -device virtio-blk-pci,scsi=off,bus=pci.0,addr=0x7,drive=drive-virtio-disk0,id=virtio-disk0,bootindex=1 -netdev type=vhost-vdpa,vhostdev=/dev/vhost-vdpa-0,id=vhost-vdpa0 -device virtio-net-pci,netdev=vhost-vdpa0,page-per-vq=on,iommu_platform=on,id=net0,bus=pci.0,addr=0x3,disable-legacy=on -vnc 0.0.0.0:0


In the vm:  virtio net device  eth0 with ip address 10.0.0.75/24

On the host: VF0 rep device pf0vf0 with ip address 10.0.0.7/24


problem 1:

On the host:

# ping 10.0.0.75

Some times there will be loss packets.

And in the VM:

dmesg shows:

eth0: bad tso: type 100, size: 0

eth0: bad tso: type 10, size: 28


So I think maybe the vnet header is not init with 0?  And Then I clear the gso_type, gso_size and flags in the virtio_net driver.  There is no packets dropped.


problem 2:

In the vm: iperf -s

On the host: iperf -c 10.0.0.75 -t 100 -i 2.


The tcp connection can't established for the syn+ack with partail cum but not handle correct by hardware.

After I set the csum off  for eth0 in the vm, the problem is resolved. Although the mlx5_vnet support VIRTIO_NET_F_CSUM feature.



problem 3:


The iperf perofrmance not stable before I disable the pf0vf0 tso offload

#ethtool -K pf0vf0 tso off


I know the mlx5_vnet did not support feature VIRTIO_NET_F_GUEST_TSO4. But the hardware can't  cut the big tso packet to several small tcp packet and send to virtio  net device?




BR

wenxu