[PATCH] net/txgbe: store flow rules without wrappers
Stephen Hemminger
stephen at networkplumber.org
Tue Nov 10 16:04:58 CET 2026
On Tue, 10 Nov 2026 21:22:00 +0800
Zhang Tengfei <zhtfdev at gmail.com> wrote:
> The flow rule wrappers each held one filter structure after their
> list links were removed. Store the filter structure directly in
> rte_flow. RSS still goes through txgbe_rss_conf_init() so the saved
> key and queue pointers refer to that stored copy.
>
> Signed-off-by: Zhang Tengfei <zhtfdev at gmail.com>
> ---
Looks good, AI review did have some observations to consider for later releases
[PATCH] net/txgbe: store flow rules without wrappers
Mechanical change; the conversion is correct. Allocation sizes match
the copied types, every flow->rule consumer is updated, and the RSS
path still goes through txgbe_rss_conf_init() so conf.key and
conf.queue point into the stored copy rather than the parse-time
stack buffer. txgbe_config_rss_filter(..., FALSE) and
txgbe_action_rss_same() therefore still see valid pointers on
destroy.
Info
flow->rule is void *, so the casts on the lines being touched in
txgbe_flow_destroy() are unnecessary. Since those lines are being
rewritten anyway, plain struct assignment is simpler than
rte_memcpy() in the control path:
case RTE_ETH_FILTER_NTUPLE:
ntuple_filter_ptr = pmd_flow->rule;
ntuple_filter = *ntuple_filter_ptr;
ret = txgbe_add_del_ntuple_filter(dev, &ntuple_filter, FALSE);
Same for the ethertype, syn, fdir and l2_tunnel cases.
Not introduced here, but worth a follow-up: these rule copies are
CPU-only bookkeeping and do not need hugepage memory; malloc()/free()
would do.
More information about the dev
mailing list