[PATCH net v3] xfrm: hold state while flushing device GC list

From: Hu EnZe <redacted>

The device GC list does not hold a reference to the states it
contains. During netdevice teardown, xfrm_dev_state_flush() drops
xfrm_state_dev_gc_lock before calling xfrm_dev_state_free(), so the
regular GC worker can concurrently destroy and free the entire
xfrm_state. Subsequent access through the stale pointer can corrupt
memory and cause a kernel panic.

Pin live states with xfrm_state_hold_rcu() while holding the device GC
lock, and skip states whose reference count has already reached zero.
Drop the acquired reference after the device state cleanup completes.
Also use hlist_del_init() when removing dev_gclist so the node is
marked as unhashed after removal.

Fixes: 07b87f9eea0c ("xfrm: Fix unregister netdevice hang on hardware offload.")
Reported-by: TencentOS Corvus AI <redacted>
Assisted-by: Codex:GPT-5.6-Sol
Signed-off-by: Hu EnZe <redacted>
---
v3:
  - fix the outer email From name from "paulinsider" to "Hu EnZe"
  - no code changes

v2: https://lore.kernel.org/netdev/20260809081151.56774-1-paulinsider@gmail.com/ (local)
  - rebase onto the latest ipsec tree
  - use Hu EnZe [off-list ref] for authorship and Signed-off-by
  - add the Assisted-by trailer for Codex
  - clarify the netdevice teardown trigger in the commit message
  - no code changes

v1: https://lore.kernel.org/netdev/20260801041340.92117-1-paulinsider@gmail.com/ (local)

 net/xfrm/xfrm_state.c | 7 ++++++-
 1 file changed, 6 insertions(+), 1 deletion(-)

diff --git a/net/xfrm/xfrm_state.c b/net/xfrm/xfrm_state.c
index 36a4f6793ede..01138d3d1067 100644
--- a/net/xfrm/xfrm_state.c
+++ b/net/xfrm/xfrm_state.c

@@ -785,7 +785,7 @@ void xfrm_dev_state_free(struct xfrm_state *x)
 	if (dev && dev->xfrmdev_ops) {
 		spin_lock_bh(&xfrm_state_dev_gc_lock);
 		if (!hlist_unhashed(&x->dev_gclist))
-			hlist_del(&x->dev_gclist);
+			hlist_del_init(&x->dev_gclist);
 		spin_unlock_bh(&xfrm_state_dev_gc_lock);
 
 		if (dev->xfrmdev_ops->xdo_dev_state_free)

@@ -1006,8 +1006,13 @@ int xfrm_dev_state_flush(struct net *net, struct net_device *dev, bool task_vali
 		xso = &x->xso;
 
 		if (xso->dev == dev) {
+			/* The device GC list does not hold a reference to x. */
+			if (!xfrm_state_hold_rcu(x))
+				continue;
+
 			spin_unlock_bh(&xfrm_state_dev_gc_lock);
 			xfrm_dev_state_free(x);
+			xfrm_state_put(x);
 			spin_lock_bh(&xfrm_state_dev_gc_lock);
 			goto restart_gc;
 		}

base-commit: c12cbf56320fb633484ee0ca1fb7d68d6b64b213
-- 
2.55.0