On Wed, Nov 4, 2026 at 4:46 PM Alexandre Cassen [off-list ref] wrote:
quoted hunk ↗ jump to hunk
A skip_sw filter gives its skip_sw flag to its actions only after their
->init() returns. ->init() then cannot know that the action never runs
in software unless the user also sets skip_sw on the action. Since
commit 34d35b4edbbe ("net/sched: act_api: deny mismatched
skip_sw/skip_hw flags for actions created by classifiers"), this action
flag also needs the filter one. The Sashiko AI review [1] of the
tunnel_key dst cache patch reports this limit.
Pass the filter offload flags to ->init() so that an action can skip the
state only the software path needs, like the tunnel_key dst cache.
tcf_idr_create() stores them in the new action, which replaces the
inheritance after ->init(). Since commit 9463e8d8c692 ("net/sched:
act_api: inherit filter offload flags only on new actions"), this
inheritance only applies to new actions, so the final action flags do
not change.
No action reads skip_hw at init and only act_skbedit reads skip_sw.
Since commit 4a6a676f8c16 ("act_skbedit: skbedit queue mapping for
receive queue"), it requires skip_sw for a receive queue_mapping. The
filter flag is now enough for this check.
Link: https://netdev-ai.bots.linux.dev/sashiko/#/patchset/20261009051930.3619697-1-acassen%40corp.free.fr [1]
Signed-off-by: Alexandre Cassen <redacted>
---
net/sched/act_api.c | 17 +++++++----------
1 file changed, 7 insertions(+), 10 deletions(-)
diff --git a/net/sched/act_api.c b/net/sched/act_api.c
index e142d33dc5ef..13317d1abde1 100644
--- a/net/sched/act_api.c
+++ b/net/sched/act_api.c
@@ -1531,6 +1531,12 @@ int tcf_action_init(struct net *net, struct tcf_proto *tp, struct nlattr *nla,
ops[i - 1] = a_o;
}
+ /* New actions inherit these flags and init can skip software state */
+ if (tc_skip_sw(fl_flags))
+ flags |= TCA_ACT_FLAGS_SKIP_SW;
+ if (tc_skip_hw(fl_flags))
+ flags |= TCA_ACT_FLAGS_SKIP_HW;
+
for (i = 1; i <= TCA_ACT_MAX_PRIO && tb[i]; i++) {
act = tcf_action_init_1(net, tp, tb[i], est, ops[i - 1],
&init_res[i - 1], flags, extack);@@ -1549,9 +1555,7 @@ int tcf_action_init(struct net *net, struct tcf_proto *tp, struct nlattr *nla,
/* Action is created by classifier and is not
* standalone. Check that the user did not set
* any action flags different than the
- * classifier flags, and inherit the flags from
- * the classifier for the compatibility case
- * where no flags were specified at all.
+ * classifier flags.
*/
if ((tc_act_skip_sw(act->tcfa_flags) && !skip_sw) ||
(tc_act_skip_hw(act->tcfa_flags) && !skip_hw)) {@@ -1560,13 +1564,6 @@ int tcf_action_init(struct net *net, struct tcf_proto *tp, struct nlattr *nla,
err = -EINVAL;
goto err;
}
- /* Other filters share an existing action */
- if (init_res[i - 1] != ACT_P_CREATED)
- continue;
- if (skip_sw)
- act->tcfa_flags |= TCA_ACT_FLAGS_SKIP_SW;
- if (skip_hw)
- act->tcfa_flags |= TCA_ACT_FLAGS_SKIP_HW;
continue;
}
Reviewed-by: Jamal Hadi Salim <jhs@mojatatu.com>
BTW - you will have to resend this because net-next is closed (If i am
not mistaken it opens monday).
Re-looking if you are going to make an update: act_bpf, act_ife and
act_skbmod still hardcode the per-cpu stats; their fast paths use raw
per-cpu derefs so they need the helper conversion first. Sorry
Alexandre ;->
cheers,
jamal