[V2 0/7] hinic3 change for support new SPx NIC

[Stephen Hemminger]

On Mon, 16 Mar 2026 21:43:22 +0800
Feifei Wang <wff_light at vip.163.com> wrote:

> From: chenyi221 <chenyi221 at huawei.com>
> 
> Change hinic3 driver to support Huawei new SPx series NIC.
> 
> v2:
> --fix build issues
> 
> Feifei Wang (7):
>   net/hinic3: add support for new SPx series NIC
>   net/hinic3: add enhance cmdq support for new SPx series NIC
>   net/hinic3: use different callback func to split new/old cmdq
>     operations
>   net/hinic3: add fun init ops to support Compact CQE
>   net/hinic3: add rx ops to support Compact CQE
>   net/hinic3: add tx ops to support Compact CQE
>   net/hinic3: use different callback func to support htn fdir
> 
>  drivers/net/hinic3/base/hinic3_cmd.h          |  80 ++-
>  drivers/net/hinic3/base/hinic3_cmdq.c         | 370 ++++------
>  drivers/net/hinic3/base/hinic3_cmdq.h         | 112 ++-
>  drivers/net/hinic3/base/hinic3_cmdq_enhance.c | 111 +++
>  drivers/net/hinic3/base/hinic3_cmdq_enhance.h | 125 ++++
>  drivers/net/hinic3/base/hinic3_csr.h          |  18 +-
>  drivers/net/hinic3/base/hinic3_hw_comm.c      |  15 +-
>  drivers/net/hinic3/base/hinic3_hw_comm.h      |  31 +-
>  drivers/net/hinic3/base/hinic3_hwdev.c        |  13 +-
>  drivers/net/hinic3/base/hinic3_hwdev.h        |  18 +
>  drivers/net/hinic3/base/hinic3_hwif.c         |  10 +-
>  drivers/net/hinic3/base/hinic3_mgmt.c         |   5 +-
>  drivers/net/hinic3/base/hinic3_mgmt.h         |   2 +
>  drivers/net/hinic3/base/hinic3_nic_cfg.c      | 182 ++---
>  drivers/net/hinic3/base/hinic3_nic_cfg.h      |  98 ++-
>  drivers/net/hinic3/base/meson.build           |   1 +
>  drivers/net/hinic3/hinic3_ethdev.c            | 280 ++++++--
>  drivers/net/hinic3/hinic3_ethdev.h            | 120 ++--
>  drivers/net/hinic3/hinic3_fdir.c              | 657 +++++++++++++-----
>  drivers/net/hinic3/hinic3_fdir.h              | 361 ++++++++--
>  drivers/net/hinic3/hinic3_nic_io.c            | 525 ++++++--------
>  drivers/net/hinic3/hinic3_nic_io.h            | 163 ++++-
>  drivers/net/hinic3/hinic3_rx.c                | 265 +++++--
>  drivers/net/hinic3/hinic3_rx.h                | 182 ++++-
>  drivers/net/hinic3/hinic3_tx.c                | 458 ++++++------
>  drivers/net/hinic3/hinic3_tx.h                | 154 +++-
>  .../net/hinic3/htn_adapt/hinic3_htn_cmdq.c    | 167 +++++
>  .../net/hinic3/htn_adapt/hinic3_htn_cmdq.h    |  63 ++
>  drivers/net/hinic3/htn_adapt/meson.build      |   7 +
>  drivers/net/hinic3/meson.build                |   8 +-
>  .../net/hinic3/stn_adapt/hinic3_stn_cmdq.c    | 151 ++++
>  .../net/hinic3/stn_adapt/hinic3_stn_cmdq.h    |  46 ++
>  drivers/net/hinic3/stn_adapt/meson.build      |   7 +
>  33 files changed, 3362 insertions(+), 1443 deletions(-)
>  create mode 100644 drivers/net/hinic3/base/hinic3_cmdq_enhance.c
>  create mode 100644 drivers/net/hinic3/base/hinic3_cmdq_enhance.h
>  create mode 100644 drivers/net/hinic3/htn_adapt/hinic3_htn_cmdq.c
>  create mode 100644 drivers/net/hinic3/htn_adapt/hinic3_htn_cmdq.h
>  create mode 100644 drivers/net/hinic3/htn_adapt/meson.build
>  create mode 100644 drivers/net/hinic3/stn_adapt/hinic3_stn_cmdq.c
>  create mode 100644 drivers/net/hinic3/stn_adapt/hinic3_stn_cmdq.h
>  create mode 100644 drivers/net/hinic3/stn_adapt/meson.build
> 
> --
> 2.45.1.windows.1
> 

Still lots of things that need to be addressed here.
See the following AI patch review.


Patch 1/7: net/hinic3: add support for new SPx series NIC

HINIC3_DEV_ID_SP230 is defined as 0X0229 with an uppercase X.
Every other hex constant in the file uses lowercase 0x. Should be
0x0229.

HINIC3_DEV_ID_SP920 has no VF counterpart and no entry in
hinic3_is_vf_dev(). Is that intentional?

Patch 2/7: net/hinic3: add enhance cmdq support

The unified cmdq_sync_cmd() retains rte_smp_rmb() which is a
deprecated barrier. Since this function is being rewritten, please
convert to rte_atomic_thread_fence(rte_memory_order_acquire).

HINIC3_CMDQ_BUF_SIZE changed from 2048 to 1024 without any
explanation in the commit message. If existing normal-cmdq commands
relied on the larger buffer this could silently truncate them.

Patch 3/7: net/hinic3: use different callback func to split new/old cmdq operations

The call sites in hinic3_ethdev.c use hinic3_cmdq_get_stn_ops()
and hinic3_cmdq_get_htn_ops() but the actual definitions are
hinic3_nic_cmdq_get_stn_ops() and hinic3_nic_cmdq_get_htn_ops().
This will not link. Patch 4 fixes the call site but each commit
must compile independently.

In htn_adapt/hinic3_htn_cmdq.c the static function
prepare_rss_indir_table_cmd_header() is called before it is
defined, with no forward declaration. This will fail with
-Werror=implicit-function-declaration.

Patch 4/7: net/hinic3: add fun init ops to support Compact CQE

In hinic3_rx_queue_setup(), when ci_mz allocation fails the code
calls hinic3_memzone_free(ci_mz) on a NULL pointer before jumping
to the error label. This call is either a NULL dereference or dead
code and should be removed.

The error-path goto labels in hinic3_func_init() are misordered.
The allocation order is mac_addrs, cmdq_ops, rx_ops, tx_ops,
mc_list but the cleanup labels do not reverse this properly. For
example if cmdq_ops allocation fails, mac_addrs is never freed.
If tx_ops fails, neither cmdq_ops nor mac_addrs are freed. This
leaks memory on every init failure path.

The ternary in hinic3_pf_get_default_cos() assigns
HINIC3_COS_NUM_MAX (8) when NIC_F_HTN_CMDQ is set and
HINIC3_COS_NUM_MAX_HTN (4) when it is not. The naming suggests
this is backwards. Please verify.

Patch 5/7: net/hinic3: add rx ops to support Compact CQE

hinic3_poll_integrated_cqe_rq_empty() uses __atomic_load_n() with
__ATOMIC_ACQUIRE. New code must use rte_atomic_load_explicit()
with rte_memory_order_acquire instead of the GCC built-in.

Patch 7/7: net/hinic3: use different callback func to support htn fdir

Copy-paste bug in hinic3_rss_hash_update() and hinic3_rss_conf_get().
In both functions the else branch has:

    rss_type.ipv6_ext = 0;
    rss_type.ipv6_ext = 0;

The second line should be rss_type.tcp_ipv6_ext = 0. As written
tcp_ipv6_ext is never cleared on non-HTN hardware which will
produce incorrect RSS configuration.