Re: [PATCH v5 0/7] multi-process and VF hotplug fixes

[Stephen Hemminger <stephen@networkplumber.org>]

On Thu, 26 Feb 2026 17:59:20 -0800
Long Li <longli@microsoft.com> wrote:

> Fix several issues with VF hotplug and multi-process support in
> netvsc, mana, mlx5, and mlx4 drivers:
> 
> - Fix race conditions between VSP notifications and DPDK device events
>   during VF add/remove, with proper locking of VF-related fields
> - Add multi-process communication infrastructure for coordinating VF
>   removal across primary and secondary processes
> - Fix Protection Domain resource leak on device close in mana
> - Fix devargs memory leak during VF hotplug in netvsc
> - Fix fast-path ops (rte_eth_fp_ops) setup in secondary processes for
>   mana, mlx5, and mlx4, ensuring burst function pointers are restored
>   after STOP->START cycles
> 
> v5:
> - Patches 5,6,7: Also restore rte_eth_fp_ops burst function pointers
>   (rx_pkt_burst, tx_pkt_burst) in START_RXTX handler, not just queue
>   data pointers. Without this, after a STOP->START cycle the secondary
>   process burst pointers remain set to dummy functions.
> 
> v4:
> - Patch 1: Check hn_vf_add() return value in netvsc_hotplug_retry
> - Patch 1: Track fresh_attach to avoid tearing down original VF
>   attachment when configure/start fails on an -EEXIST path
> - Patch 2: Move counter decrement and netvsc_uninit_once() after device
>   cleanup in eth_hn_remove() to prevent use-after-free of shared data
> - Patch 2: Clear netvsc_shared_data on init failure paths to prevent
>   dangling pointer
> 
> v3:
> - Fix review comments from v2
> 
> v2:
> - Initial rework of VF add/remove locking
> 
> Long Li (7):
>   net/netvsc: fix race conditions on VF add/remove events
>   net/netvsc: add multi-process VF device removal support
>   net/mana: fix PD resource leak on device close
>   net/netvsc: fix devargs memory leak on hotplug
>   net/mana: fix fast-path ops setup in secondary process
>   net/mlx5: fix fast-path ops setup in secondary process
>   net/mlx4: fix fast-path ops setup in secondary process
> 
>  drivers/net/mana/mana.c             |  14 ++
>  drivers/net/mana/mp.c               |   8 +
>  drivers/net/mlx4/mlx4_mp.c          |   6 +
>  drivers/net/mlx5/linux/mlx5_mp_os.c |   6 +
>  drivers/net/netvsc/hn_ethdev.c      | 300 +++++++++++++++++++++++++++-
>  drivers/net/netvsc/hn_nvs.h         |   6 +
>  drivers/net/netvsc/hn_rxtx.c        |  40 ++--
>  drivers/net/netvsc/hn_var.h         |   1 +
>  drivers/net/netvsc/hn_vf.c          | 148 ++++++++------
>  9 files changed, 437 insertions(+), 92 deletions(-)
> 

Looks okay to me, the AI review feedback raised a couple of questions.
If it is ok will take it as is for this release.

The AI summary was:

Patch 1: hn_vf_add_unlocked() — when hn_nvs_set_datapath() fails at switch_data_path: after a fresh attach, the VF is not detached (no goto detach). This leaves inconsistent state.

Patch 2: netvsc_uninit_once() — primary can free the shared memzone while secondaries still reference netvsc_shared_data, causing a dangling pointer. The local-only secondary_cnt check doesn't reflect remote secondary processes.

Warnings (should consider)

Patch 1: Potential deadlock in hn_vf_close() — holding write lock while calling rte_eth_dev_callback_unregister() which synchronously waits for in-progress callbacks that may themselves try to acquire the write lock via hn_remove_delayed().

There were a couple more things but these were just AI being overly paranoid.