Skip to contentSkip to conversation

net/sched: cls_route: fix fastmap use-after-free and preserve netns teardown ordering

PATCH In review Jamal Hadi Salim 1 participant 2 messages Jul 24 – Jul 27

Subsystems (4)
lore
Share & tools
lore

Jamal Hadi Salim

author maintainer
#0 lore raw

PATCH v2 1/1 net/sched: cls_route: fix fastmap use-after-free and preserve netns teardown ordering

The route4 classifier maintains a 16-slot fastmap cache that stores raw struct route4_filter pointers indexed by (id, iif). The reader (route4_classify) populates this cache via route4_set_fastmap() for every classified packet that hits a filter. The writer (route4_delete, route4_change) clears the cache via route4_reset_fastmap() before RCU-deferred kfree of the filter.

This creates a UAF race:
1. Reader walks the RCU-protected bucket chain, finds filter f
2. Writer unlinks f, calls route4_reset_fastmap(), then tcf_queue_work()
3. Reader calls route4_set_fastmap() and writes f into the cache *after* the writer's reset, caching a pointer about to be freed
4. After the RCU grace period, kfree(f) executes
5. Next classified packet on the same (id, iif) tuple hits the stale

    fastmap entry and reads f->res from freed memory

Reproduced with an mdelay(100) accelerator in route4_set_fastmap() and a concurrent add/delete stress test (provided by both zdi and Santosh). Both triggered KASAN slab-use-after-free reports in the route4 fastmap paths.

Fix:
Move the fastmap flush to the writer side under RTNL, ordered after synchronize_rcu(). This ensures stale fastmap entries written by in-flight readers are visible (readers have finished) and no new readers can find the unlinked filter, so the subsequent route4_reset_fastmap(head) flushes them safely while head is valid under RTNL.

In route4_destroy() the synchronous __route4_delete_filter() fallback is retained for the case where tcf_exts_get_net() returns false (netns is being torn down in cleanup_net()). The trailing synchronize_rcu() before kfree_rcu(head) drains any in-flight readers first, so the synchronous free is safe. Removing the fallback and unconditionally deferring via route4_queue_work() would race with tc_action_net_exit(): cleanup_net() does not flush the tc_filter_wq workqueue, so the deferred work could run after the action idrinfo has been freed, causing a use-after-free on idrinfo->lock in __tcf_action_put().

Note: This fix will slow down deletes and filter add/replace, since synchronize_rcu() is called under RTNL in route4_delete() and route4_change(). An alternative fix would have been to introduce refcnt to the head struct but that is a much bigger patch. We justify this as a reasonable fix.

---
V1->V2
1. Fix based on feedback from the Sashikos
https://sashiko.dev/#/patchset/20260723105210.817079-1-jhs%40mojatatu.com
https://netdev-ai.bots.linux.dev/sashiko/#/patchset/20260723105210.817079-1-jhs%40mojatatu.com

Restore the "if (tcf_exts_get_net(&f->exts)) route4_queue_work(f); else __route4_delete_filter(f);"
preserving the tcf_exts_get_net() contract for netns teardown.

---
 net/sched/cls_route.c | 34 +++++++++++++++++++++++++++++++---
 1 file changed, 31 insertions(+), 3 deletions(-)
net/sched/cls_route.c +31 -3
index bd6f945bd388..c7d0164c1e9f 100644
@@ -307,6 +307,19 @@ static void route4_destroy(struct tcf_proto *tp, bool rtnl_held,
 			kfree_rcu(b, rcu);
 		}
 	}
+
+	/* All filters are unlinked; no new reader can find them on the
+	 * chain.  Wait for in-flight readers that may still hold a filter
+	 * pointer and have published it into the fastmap after we unlinked.
+	 * Then flush the stale entries while head is still valid under
+	 * RTNL, matching the route4_delete() pattern.  This also protects
+	 * the synchronous __route4_delete_filter() path above: when the
+	 * netns is dying, tcf_exts_get_net() returns false and the filter
+	 * is freed directly, but only after all readers have drained.
+	 */
+	synchronize_rcu();
+	route4_reset_fastmap(head);
+
 	kfree_rcu(head, rcu);
 }
 
@@ -334,10 +347,16 @@ static int route4_delete(struct tcf_proto *tp, void *arg, bool *last,
 			/* unlink it */
 			RCU_INIT_POINTER(*fp, rtnl_dereference(f->next));
 
-			/* Remove any fastmap lookups that might ref filter
-			 * notice we unlink'd the filter so we can't get it
-			 * back in the fastmap.
+			/* This code path assumes we have the RTNL lock.
+			 * We wait for in-flight readers that may still hold
+			 * the filter pointer and publish it into the fastmap
+			 * after we unlinked it.
+			 * Once done, no new reader can find the filter on the
+			 * chain, so the only stale fastmap entries are the
+			 * ones those readers just wrote. Flush them now
+			 * while head is still valid.
 			 */
+			synchronize_rcu();
 			route4_reset_fastmap(head);
 
 			/* Delete it */
@@ -558,6 +577,15 @@ static int route4_change(struct net *net, struct sk_buff *in_skb,
 		}
 	}
 
+	/* Assuming RTNL lock. Wait for in-flight readers that may still
+	 * hold a filter pointer and publish it into the fastmap after we
+	 * inserted the new filter (or unlinked fold when replacing).
+	 * Once they are done, flush the stale entries while head is still
+	 * valid.  Needed on both the creation and replace paths: on
+	 * creation, readers may have cached a ROUTE4_FAILURE entry for the
+	 * (id, iif) tuple that the new filter now matches.
+	 */
+	synchronize_rcu();
 	route4_reset_fastmap(head);
 	*arg = f;
 	if (fold) {
Signature · 1 line
2.34.1
author maintainer
#1 lore raw
Quoted 29 lines · Jamal Hadi Salim

The route4 classifier maintains a 16-slot fastmap cache that stores raw struct route4_filter pointers indexed by (id, iif). The reader (route4_classify) populates this cache via route4_set_fastmap() for every classified packet that hits a filter. The writer (route4_delete, route4_change) clears the cache via route4_reset_fastmap() before RCU-deferred kfree of the filter.

This creates a UAF race:
1. Reader walks the RCU-protected bucket chain, finds filter f
2. Writer unlinks f, calls route4_reset_fastmap(), then tcf_queue_work()
3. Reader calls route4_set_fastmap() and writes f into the cache *after* the writer's reset, caching a pointer about to be freed
4. After the RCU grace period, kfree(f) executes
5. Next classified packet on the same (id, iif) tuple hits the stale

    fastmap entry and reads f->res from freed memory

Reproduced with an mdelay(100) accelerator in route4_set_fastmap() and a concurrent add/delete stress test (provided by both zdi and Santosh). Both triggered KASAN slab-use-after-free reports in the route4 fastmap paths.

Fix:
Move the fastmap flush to the writer side under RTNL, ordered after synchronize_rcu(). This ensures stale fastmap entries written by in-flight readers are visible (readers have finished) and no new readers can find the unlinked filter, so the subsequent route4_reset_fastmap(head) flushes them safely while head is valid under RTNL.

Quoted 10 lines

In route4_destroy() the synchronous __route4_delete_filter() fallback is retained for the case where tcf_exts_get_net() returns false (netns is being torn down in cleanup_net()). The trailing synchronize_rcu() before kfree_rcu(head) drains any in-flight readers first, so the synchronous free is safe. Removing the fallback and unconditionally deferring via route4_queue_work() would race with tc_action_net_exit(): cleanup_net() does not flush the tc_filter_wq workqueue, so the deferred work could run after the action idrinfo has been freed, causing a use-after-free on idrinfo->lock in __tcf_action_put().

Well... AI reads commits and comments and correlates to code ;-> Both Sashiko's (legitimately) dont like this commit log's accuracy.

Quoted 6 lines

Note: This fix will slow down deletes and filter add/replace, since synchronize_rcu() is called under RTNL in route4_delete() and route4_change(). An alternative fix would have been to introduce refcnt to the head struct but that is a much bigger patch. We justify this as a reasonable fix.

[..]

Quoted 8 lines
net/sched/cls_route.c +0 -0
index bd6f945bd388..c7d0164c1e9f 100644
@@ -307,6 +307,19 @@ static void route4_destroy(struct tcf_proto *tp, bool rtnl_held,
                        kfree_rcu(b, rcu);
                }
        }

net/sched/cls_route.c static void route4_destroy(struct tcf_proto *tp, bool rtnl_held, · line 319

Quoted 10 lines
net/sched/cls_route.c +10 -0
+
+       /* All filters are unlinked; no new reader can find them on the
+        * chain.  Wait for in-flight readers that may still hold a filter
+        * pointer and have published it into the fastmap after we unlinked.
+        * Then flush the stale entries while head is still valid under
+        * RTNL, matching the route4_delete() pattern.  This also protects
+        * the synchronous __route4_delete_filter() path above: when the
+        * netns is dying, tcf_exts_get_net() returns false and the filter
+        * is freed directly, but only after all readers have drained.
+        */

And the above related comment.

sashiko-claude is over-reaching with some manufactured UAF and performance comment (Which i address in the commit log).

Let me know if i should send v3 making the commit and comment more accurate as well as re-emphasize the performance issue. (with hope next review sashiko wont bring it up).
We can ignore the UAF claim made.

cheers,
jamal