Linux kernel -stable discussions
 help / color / mirror / Atom feed
* [PATCH net v2] net/sched: act_ife: return -EAGAIN after dropping rtnl_lock in load_metaops_and_vet()
@ 2026-11-22 12:45 YanXin Li via B4 Relay
  2026-11-22 17:18 ` Stephen Hemminger
  0 siblings, 1 reply; 2+ messages in thread
From: YanXin Li via B4 Relay @ 2026-11-22 12:45 UTC (permalink / raw)
  To: Jamal Hadi Salim, Jiri Pirko, David S. Miller, Eric Dumazet,
	Jakub Kicinski, Paolo Abeni
  Cc: Simon Horman, netdev, linux-kernel, stable

From: YanXin Li <fadouse@pm.me>

load_metaops_and_vet() unlocks rtnl_lock() to invoke request_module()
when an IFE metadata provider module (such as act_meta_mark) is not
yet loaded, re-acquires rtnl_lock(), validates the metadata attribute,
and returns 0.

Returning 0 after dropping rtnl_lock() violates the serialization
contract of RTNL-locked classifiers (cls_basic, cls_flow, cls_fw,
cls_route, and cls_u32), which hold an unrefcounted filter pointer
(fold) across tcf_exts_validate() -> tcf_ife_init().  A concurrent
RTM_DELTFILTER can delete and RCU-free fold during the unlocked
request_module() window, after which basic_change() executes
list_replace_rcu(&fold->link, &fnew->link) and queues a second
deletion on the freed filter.  Generic KASAN reports:

  BUG: KASAN: slab-use-after-free in basic_change+0xf4c/0x1220
  Read of size 8 at addr ffff88800e043c58 by task test_act_ife_rt/101
  ...
  Allocated by task 101:
   basic_change+0x1e0/0x1220
  Freed by task 57:
   ...
  Oops: general protection fault, probably for non-canonical address
  KASAN: maybe wild-memory-access in range [0xdead000000000120-0xdead000000000127]

Every other TC module autoloader that drops rtnl_lock() during filter
updates (tc_action_load_ops(), tcf_em_validate(), and em_text_change())
returns -EAGAIN so tc_new_tfilter() and tcf_action_add() unwind and
replay under a fresh RTNL critical section.

Manipulating TC filters requires CAP_NET_ADMIN in the target network
namespace, which is reachable by an unprivileged local user inside a
user and network namespace when CONFIG_USER_NS=y, CONFIG_MODULES=y,
CONFIG_NET_ACT_IFE is enabled, and an IFE metadata module is not yet
loaded.

When load_metaops_and_vet() drops rtnl_lock() and loads the requested
metadata module, drop the module reference and return -EAGAIN so the
caller replays the request with fresh lookups.  With this change, the
same concurrent RTM_NEWTFILTER and RTM_DELTFILTER sequence cleanly
loads act_meta_mark, replays without touching the deleted filter, and
completes with rc=0 and no KASAN or lockdep reports.  A finite
two-thread state-space interleaving model found the old use-after-free
and no use-after-free, double-free, or deadlock in its reachable states
after the change.

The vulnerable code is present in upstream mainline v7.4-rc3.
Runtime verification and fix validation were performed in an x86-64
QEMU/KVM guest running mainline after v7.4-rc3 with Generic KASAN,
UBSAN, and lockdep.

This issue was discovered through LLM-assisted code audit and manual
code inspection of net/sched/act_ife.c.

Fixes: ef6980b6becb ("introduce IFE action")
Cc: stable@vger.kernel.org
Signed-off-by: YanXin Li <fadouse@pm.me>
---
Changes in v2:
- Resend via b4 because the archived v1 email has a corrupted diff and
  cannot be applied.
- State how the issue was discovered and clarify the finite model scope.
- Link to v1: https://lore.kernel.org/r/20261122030728.9368-1-fadouse@pm.me
---
 net/sched/act_ife.c | 8 ++++++++
 1 file changed, 8 insertions(+)

diff --git a/net/sched/act_ife.c b/net/sched/act_ife.c
index 5a527da0056..5d6ca124e00 100644
--- a/net/sched/act_ife.c
+++ b/net/sched/act_ife.c
@@ -275,6 +275,14 @@ static int load_metaops_and_vet(u32 metaid, void *val, int len, bool rtnl_held)
 		if (rtnl_held)
 			rtnl_lock();
 		ops = find_ife_oplist(metaid);
+		if (ops && rtnl_held) {
+			/* We dropped the RTNL mutex in order to perform the
+			 * module load. Tell the caller to replay the request
+			 * under a fresh RTNL critical section.
+			 */
+			module_put(ops->owner);
+			return -EAGAIN;
+		}
 #endif
 	}
 

---
base-commit: c44104f3642da720164042623e4832836aefdd8e
change-id: 20261122-b4-02-act-ife-rtnl-uaf-v2-102eba468698-d819f7ac7006

Best regards,
--  
YanXin Li <fadouse@pm.me>



^ permalink raw reply related	[flat|nested] 2+ messages in thread

* Re: [PATCH net v2] net/sched: act_ife: return -EAGAIN after dropping rtnl_lock in load_metaops_and_vet()
  2026-11-22 12:45 [PATCH net v2] net/sched: act_ife: return -EAGAIN after dropping rtnl_lock in load_metaops_and_vet() YanXin Li via B4 Relay
@ 2026-11-22 17:18 ` Stephen Hemminger
  0 siblings, 0 replies; 2+ messages in thread
From: Stephen Hemminger @ 2026-11-22 17:18 UTC (permalink / raw)
  To: YanXin Li via B4 Relay
  Cc: fadouse, Jamal Hadi Salim, Jiri Pirko, David S. Miller,
	Eric Dumazet, Jakub Kicinski, Paolo Abeni, Simon Horman, netdev,
	linux-kernel, stable

On Sun, 22 Nov 2026 20:45:52 +0800
YanXin Li via B4 Relay <devnull+fadouse.pm.me@kernel.org> wrote:

> +		if (ops && rtnl_held) {
> +			/* We dropped the RTNL mutex in order to perform the
> +			 * module load. Tell the caller to replay the request
> +			 * under a fresh RTNL critical section.
> +			 */

I prefer that comments stay in "short and sweet" style.
AI and new developers seem to like long comments.
This comment has the classic /* add one to i */ style...
In other words the comment doesn't tell human more than is already
obvious from the code.

^ permalink raw reply	[flat|nested] 2+ messages in thread

end of thread, other threads:[~2026-11-22 17:18 UTC | newest]

Thread overview: 2+ messages (download: mbox.gz follow: Atom feed
-- links below jump to the message on this page --
2026-11-22 12:45 [PATCH net v2] net/sched: act_ife: return -EAGAIN after dropping rtnl_lock in load_metaops_and_vet() YanXin Li via B4 Relay
2026-11-22 17:18 ` Stephen Hemminger

This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox