Skip to contentSkip to conversation

Please revert mlx IOVA VA mode patch

dpdk-dev Stephen Hemminger 2 participants 2 messages Jun 7, 2019

lore
Share & tools
lore

Stephen Hemminger

author Acked-by Shahaf Shuler
#0 lore raw

The Netvsc PMD got broken on DPDK 19.08 by a bad patch in Mellanox driver. On a simple setup with SRIOV enabled and using netvsc PMD with the testpmd. It worked in 19.05 and does not work with current master.

The probe code gets stuck because netvsc PMD sends a request to host and never sees a response.

# ./build/app/testpmd --log-level='pmd.net.netvsc.*:debug' -l 0-3 -n 4  -- --rxq=4 --txq=4 -i
EAL: Detected 4 lcore(s)
EAL: Detected 1 NUMA nodes
EAL: Multi-process socket /var/run/dpdk/rte/mp_socket
EAL: Probing VFIO support...
EAL: WARNING: cpu flags constant_tsc=yes nonstop_tsc=no -> using unreliable clock cycles !
EAL: PCI device 935f:00:02.0 on NUMA socket 0
EAL:   probe driver: 15b3:1014 net_mlx5
net_mlx5: can not get IB device "mlx5_0" ports number
eth_hn_probe():  >>
eth_hn_dev_init():  >>
hn_nvs_init(): NVS version 0x60001, NDIS version 6.30
hn_nvs_conn_rxbuf(): connect rxbuff va=0x2200402000 gpad=0xe1e2d
hn_nvs_conn_rxbuf(): receive buffer size 1728 count 18811
hn_nvs_conn_chim(): connect send buf va=0x2202302000 gpad=0xe1e2e
hn_nvs_conn_chim(): send buffer 16777216 section size:6144, count:2730
(hung)

The problem does not occur without SRIOV (or if MLX driver is not compiled in).

Doing bisect the problem is caused by:

commit 69c06d0e357ed0064b498d510d169603cf7308cd

Author: Yongseok Koh <[email protected]>
Date:   Thu May 2 02:07:54 2019 -0700

net/mlx: support IOVA VA mode

Set RTE_PCI_DRV_IOVA_AS_VA to driver's drv_flags as device's IOMMU takes virtual address.

Please either revert or fix this patch ASAP.

#1 lore raw

We'll have to look into the issue.
It doesn't seem simple though.
Thomas, can you please revert the patch from the master?

Thanks,
Yongseok

Quoted 48 lines · Stephen Hemminger

The Netvsc PMD got broken on DPDK 19.08 by a bad patch in Mellanox driver. On a simple setup with SRIOV enabled and using netvsc PMD with the testpmd. It worked in 19.05 and does not work with current master.

The probe code gets stuck because netvsc PMD sends a request to host and never sees a response.

# ./build/app/testpmd --log-level='pmd.net.netvsc.*:debug' -l 0-3 -n 4  -- --rxq=4 --txq=4 -i
EAL: Detected 4 lcore(s)
EAL: Detected 1 NUMA nodes
EAL: Multi-process socket /var/run/dpdk/rte/mp_socket
EAL: Probing VFIO support...
EAL: WARNING: cpu flags constant_tsc=yes nonstop_tsc=no -> using unreliable clock cycles !
EAL: PCI device 935f:00:02.0 on NUMA socket 0
EAL:   probe driver: 15b3:1014 net_mlx5
net_mlx5: can not get IB device "mlx5_0" ports number
eth_hn_probe():  >>
eth_hn_dev_init():  >>
hn_nvs_init(): NVS version 0x60001, NDIS version 6.30
hn_nvs_conn_rxbuf(): connect rxbuff va=0x2200402000 gpad=0xe1e2d
hn_nvs_conn_rxbuf(): receive buffer size 1728 count 18811
hn_nvs_conn_chim(): connect send buf va=0x2202302000 gpad=0xe1e2e
hn_nvs_conn_chim(): send buffer 16777216 section size:6144, count:2730
(hung)

The problem does not occur without SRIOV (or if MLX driver is not compiled in).

Doing bisect the problem is caused by:

commit 69c06d0e357ed0064b498d510d169603cf7308cd

Author: Yongseok Koh <[email protected]>
Date:   Thu May 2 02:07:54 2019 -0700

net/mlx: support IOVA VA mode

Set RTE_PCI_DRV_IOVA_AS_VA to driver's drv_flags as device's IOMMU takes virtual address.

Please either revert or fix this patch ASAP.