[PATCH net] net/sched: cls_api: reclaim an empty proto on the error path

Two racing tc filter add requests on the same chain/prio of an
unlocked classifier both run change() on the shared proto and both
can fail: the winner's tcf_chain_tp_delete_empty() attempt gives up
because the loser's handle is still in the idr, and the loser's
error path drops only its own reference without a second reclamation
attempt. The empty proto stays linked in the chain, holding the
chain reference, a block reference and the classifier module
reference until the chain or block is torn down.

Reclaim the proto on the error path of any failed request that holds
a proto reference. The reclamation is emptiness-gated:
tcf_chain_tp_delete_empty() unlinks the proto only when
delete_empty() admits it is empty, so a live shared proto is never
unlinked. A proto the request created is reclaimed unconditionally -
it is the only owner, so marking it for deletion is safe even without
a delete_empty callback. Classifiers without one (the check marks the
proto unconditionally) are rtnl-serialized, so the raced window this
guard closes cannot arise for them.

This is a follow-up to commit d4e359b3608a ("net/sched: cls_api: fix
teardown of an adopted proto on insert-race loss"), which stopped the
loser of the insert race from unlinking the winner's live proto but
left the empty-proto residual in place.

Conditions to recreate:
- CONFIG_NET_CLS_FLOWER=y; veth pair
- tc qdisc add dev veth0 ingress
- two concurrent `tc filter add dev veth0 ingress protocol ip pref 1
  flower skip_sw ... action drop` (both fail in fl_hw_replace_filter
  after publishing their handle in the idr); repeat in a loop
- an empty flower tp stays linked after both requests fail; visible
  as a bare `filter protocol ip pref 1 flower chain 0` header in
  `tc filter show` with no filter entries
- CAP_NET_ADMIN (namespace-local via unshare -Urn suffices)

Fixes: 8b64678e0af8 ("net: sched: refactor tp insert/delete for concurrent execution")
Reported-by: Sashiko (nipa) <sashiko-bot@kernel.org>
Closes: https://netdev-ai.bots.linux.dev/sashiko/#/patchset/20260805134049.927864-1-victor@mojatatu.com
Tested-by: hybris <redacted>
Signed-off-by: Jamal Hadi Salim <jhs@mojatatu.com>
---
 net/sched/cls_api.c | 15 ++++++++++++++-
 1 file changed, 14 insertions(+), 1 deletion(-)

diff --git a/net/sched/cls_api.c b/net/sched/cls_api.c
index c47d2ee13641..a9f54988561f 100644
--- a/net/sched/cls_api.c
+++ b/net/sched/cls_api.c

@@ -2463,7 +2463,20 @@ static int tc_new_tfilter(struct sk_buff *skb, struct nlmsghdr *n,
 	}
 
 errout:
-	if (err && tp_state == TP_CREATED)
+	if (err && !IS_ERR_OR_NULL(tp) &&
+	    (tp_state == TP_CREATED || tp->ops->delete_empty))
+		/*
+		 * The request is dropping its reference to tp. If it was
+		 * the last user (the idr is empty now), reclaim the proto.
+		 * A tp this request created is reclaimed unconditionally:
+		 * it is the only owner, so marking it for deletion is
+		 * safe. Otherwise only classifiers with a delete_empty
+		 * callback are reclaimed -- the callback admits an empty
+		 * proto only, so a live shared proto is never unlinked.
+		 * Classifiers without one (deleting is set
+		 * unconditionally) are rtnl-serialized, so the raced
+		 * window this guard closes cannot arise for them.
+		 */
 		tcf_chain_tp_delete_empty(chain, tp, rtnl_held, NULL);
 errout_tp:
 	if (chain) {

-- 
2.43.0