mac802154: hold an interface reference across the scan worker

mac802154_scan_worker() captures the scanning sub-interface once under RCU:

	sdata = IEEE802154_WPAN_DEV_TO_SUB_IF(scan_req->wpan_dev);

and then, after rcu_read_unlock() and outside the rtnl, keeps dereferencing sdata->dev: in the channel-change and restart failure traces, in mac802154_transmit_beacon_req() (skb->dev = sdata->dev) for active scans, in the final dev_dbg(), and in the end_scan mac802154_scan_cleanup_locked() path. Nothing keeps that netdev alive for the duration of the worker iteration.

A concurrent teardown of the scanning interface -- userspace issuing NL802154_CMD_DEL_INTERFACE (ieee802154_if_remove() -> unregister_netdevice()), or a full PHY removal via ieee802154_unregister_hw() -> ieee802154_remove_interfaces() -- can run as soon as the worker drops the rtnl between its two short drv_set_channel()/drv_start() sections. The netdev is not freed synchronously by unregister_netdevice(): it is queued to net_todo_list and freed later from netdev_run_todo(), which drops the rtnl mutex (__rtnl_unlock()) *before* netdev_wait_allrefs_any()/free_netdev(). The freeing therefore runs with the rtnl not held, on whichever task next drains net_todo_list. Holding the rtnl in the worker does not prevent it, and the per-PHY IEEE802154_IS_SCANNING flag does not identify the specific interface: a subsequent NEW_INTERFACE + TRIGGER_SCAN re-arms the flag, so a stale worker iteration sails past an is-scanning recheck and dereferences the already-freed netdev.

Triggering the race requires CAP_NET_ADMIN: both NL802154_CMD_TRIGGER_SCAN and NL802154_CMD_DEL_INTERFACE are GENL_ADMIN_PERM, reachable only from the initial user namespace, so the attacker is a locally privileged (CAP_NET_ADMIN) user, not an unprivileged local user or a remote peer.

KASAN slab-use-after-free, kworker reading the freed net_device/ieee802154_sub_if_data (kmalloc-cg-4k) allocated and freed by the racing NEW_INTERFACE/DEL_INTERFACE task:

  BUG: KASAN: slab-use-after-free in mac802154_scan_worker+0x... [mac802154]
  Read of size 8 ... by task kworker/u8:N
  Workqueue: phy0-mac-cmds mac802154_scan_worker [mac802154]
   mac802154_scan_worker
   process_one_work
  (also hit via ieee802154_mlme_tx_locked() from the beacon-request path
   and via _dev_err()/__dev_printk() formatting sdata->dev->dev)

Fix it by taking a reference on the interface while the RCU read lock is still held -- so the netdev cannot be freed before the refcount is raised -- and releasing it at every exit of the worker past that point. This keeps sdata->dev valid for the whole iteration.

The reference does not defer the free indefinitely and does not deadlock the single-threaded mac_wq. Each worker iteration drops the reference before it requeues itself, and a teardown started while the reference is held simply blocks in netdev_run_todo() until the current iteration returns. It cannot be the worker's own rtnl_unlock() that blocks on the reference either: the unregistering task removes the netdev from net_todo_list under the rtnl (list_replace_init() in netdev_run_todo() runs before __rtnl_unlock()), so it always owns the todo entry, and the worker acquires the rtnl only afterwards -- the blocking netdev_wait_allrefs_any() therefore always runs on the teardown task, never on the worker.

Verified on a v6.19 KASAN build: racing DEL_INTERFACE against an in-flight TRIGGER_SCAN reliably tripped a slab-use-after-free KASAN report inside mac802154_scan_worker() before this patch, and the same reproducer no longer triggers it with the fix applied.

---
 net/mac802154/scan.c | 22 ++++++++++++++++++++++
 1 file changed, 22 insertions(+)

index 0a31ac8d8415..48e661031cc0 100644
@@ -209,6 +209,26 @@ void mac802154_scan_worker(struct work_struct *work)
 		return;
 	}
 
+	/* From here on sdata->dev is dereferenced after rcu_read_unlock() and
+	 * outside the rtnl: in the dev_err()/dev_dbg() traces below, in
+	 * mac802154_transmit_beacon_req() (skb->dev = sdata->dev) and in the
+	 * end_scan mac802154_scan_cleanup_locked() call. A concurrent teardown
+	 * of that interface (NL802154_CMD_DEL_INTERFACE ->
+	 * ieee802154_if_remove(), or a full PHY removal via
+	 * ieee802154_unregister_hw()) can unregister the netdev; the actual
+	 * free then runs asynchronously from netdev_run_todo() with the rtnl
+	 * already dropped, so neither holding the rtnl nor the per-PHY
+	 * IEEE802154_IS_SCANNING flag keeps sdata->dev alive here. Pin it with
+	 * a reference taken while we still hold the RCU read lock (so the
+	 * netdev cannot be freed before we bump the refcount) and drop it at
+	 * every exit below. This blocks the teardown's netdev_run_todo() until
+	 * this worker iteration is done; it cannot self-deadlock because the
+	 * unregistering task claims the net_todo_list entry under the rtnl, so
+	 * the blocking netdev_wait_allrefs_any() always runs on that task, not
+	 * on this single-threaded worker.
+	 */
+	dev_hold(sdata->dev);
+
 	wpan_phy = scan_req->wpan_phy;
 	scan_req_type = scan_req->type;
 	scan_req_duration = scan_req->duration;

@@ -262,12 +282,14 @@ void mac802154_scan_worker(struct work_struct *work)
 		"Scan page %u channel %u for %ums\n",
 		page, channel, jiffies_to_msecs(scan_duration));
 	queue_delayed_work(local->mac_wq, &local->scan_work, scan_duration);
+	dev_put(sdata->dev);
 	return;
 
 end_scan:
 	rtnl_lock();
 	mac802154_scan_cleanup_locked(local, sdata, false);
 	rtnl_unlock();
+	dev_put(sdata->dev);
 }
 
 int mac802154_trigger_scan_locked(struct ieee802154_sub_if_data *sdata,

2.50.1 (Apple Git-155)

Hi Ibrahim,

mac802154_scan_worker() captures the scanning sub-interface once under RCU:

	sdata = IEEE802154_WPAN_DEV_TO_SUB_IF(scan_req->wpan_dev);

and then, after rcu_read_unlock() and outside the rtnl, keeps dereferencing sdata->dev: in the channel-change and restart failure traces, in mac802154_transmit_beacon_req() (skb->dev = sdata->dev) for active scans, in the final dev_dbg(), and in the end_scan mac802154_scan_cleanup_locked() path. Nothing keeps that netdev alive for the duration of the worker iteration.

A concurrent teardown of the scanning interface -- userspace issuing NL802154_CMD_DEL_INTERFACE (ieee802154_if_remove() -> unregister_netdevice()), or a full PHY removal via ieee802154_unregister_hw() -> ieee802154_remove_interfaces() -- can run as soon as the worker drops the rtnl between its two short drv_set_channel()/drv_start() sections. The netdev is not freed synchronously by unregister_netdevice(): it is queued to net_todo_list and freed later from netdev_run_todo(), which drops the rtnl mutex (__rtnl_unlock()) *before* netdev_wait_allrefs_any()/free_netdev(). The freeing therefore runs with the rtnl not held, on whichever task next drains net_todo_list. Holding the rtnl in the worker does not prevent it, and the per-PHY IEEE802154_IS_SCANNING flag does not identify the specific interface: a subsequent NEW_INTERFACE + TRIGGER_SCAN re-arms the flag, so a stale worker iteration sails past an is-scanning recheck and dereferences the already-freed netdev.

Triggering the race requires CAP_NET_ADMIN: both NL802154_CMD_TRIGGER_SCAN and NL802154_CMD_DEL_INTERFACE are GENL_ADMIN_PERM, reachable only from the initial user namespace, so the attacker is a locally privileged (CAP_NET_ADMIN) user, not an unprivileged local user or a remote peer.

KASAN slab-use-after-free, kworker reading the freed net_device/ieee802154_sub_if_data (kmalloc-cg-4k) allocated and freed by the racing NEW_INTERFACE/DEL_INTERFACE task:

  BUG: KASAN: slab-use-after-free in mac802154_scan_worker+0x... [mac802154]
  Read of size 8 ... by task kworker/u8:N
  Workqueue: phy0-mac-cmds mac802154_scan_worker [mac802154]
   mac802154_scan_worker
   process_one_work

Fix it by taking a reference on the interface while the RCU read lock is still held -- so the netdev cannot be freed before the refcount is raised -- and releasing it at every exit of the worker past that point. This keeps sdata->dev valid for the whole iteration. The reference does not defer the free indefinitely: a teardown started while it is held simply blocks in netdev_run_todo() until the current iteration returns, and it cannot self-deadlock the single-threaded mac_wq because the unregistering task claims the net_todo_list entry under the rtnl, so the blocking netdev_wait_allrefs_any() always runs on that task, not on the worker.

Verified on a v6.19 KASAN build: racing DEL_INTERFACE against an in-flight TRIGGER_SCAN reliably tripped a slab-use-after-free KASAN report inside mac802154_scan_worker() before this patch, and the same reproducer no longer triggers it with the fix applied.

---
v2: trim the in-worker comment down to the essentials, as requested by
Miquel Raynal. No functional change.

 net/mac802154/scan.c | 10 ++++++++++
 1 file changed, 10 insertions(+)

index 300d4584533e..5b4ea2a895cc 100644
@@ -209,6 +209,14 @@ void mac802154_scan_worker(struct work_struct *work)
 		return;
 	}
 
+	/*
+	 * sdata->dev is dereferenced below after rcu_read_unlock() and outside
+	 * the rtnl, and a concurrent DEL_INTERFACE / PHY teardown can free it
+	 * asynchronously from netdev_run_todo(). Pin it with a reference taken
+	 * while the RCU read lock is still held, and drop it at every exit.
+	 */
+	dev_hold(sdata->dev);
+
 	wpan_phy = scan_req->wpan_phy;
 	scan_req_type = scan_req->type;
 	scan_req_duration = scan_req->duration;

@@ -262,12 +270,14 @@ void mac802154_scan_worker(struct work_struct *work)
 		"Scan page %u channel %u for %ums\n",
 		page, channel, jiffies_to_msecs(scan_duration));
 	queue_delayed_work(local->mac_wq, &local->scan_work, scan_duration);
+	dev_put(sdata->dev);
 	return;
 
 end_scan:
 	rtnl_lock();
 	mac802154_scan_cleanup_locked(local, sdata, false);
 	rtnl_unlock();
+	dev_put(sdata->dev);
 }
 
 int mac802154_trigger_scan_locked(struct ieee802154_sub_if_data *sdata,

2.50.1 (Apple Git-155)

Hello,

Thanks,
Miquèl