usbnet uses the work usbnet_deferred_kevent() to perform tasks which may sleep. On disconnect, completion of the work was originally awaited in ->ndo_stop(). But in 2003, that was moved to ->disconnect() by historic commit "[PATCH] USB: usbnet, prevent exotic rtnl deadlock": https://git.kernel.org/tglx/history/c/0f138bbfd83c The change was made because back then, the kernel's workqueue implementation did not allow waiting for a single work. One had to wait for completion of *all* work by calling flush_scheduled_work(), and that could deadlock when waiting for usbnet_deferred_kevent() with rtnl_mutex held in ->ndo_stop(). The commit solved one problem but created another: It causes a use-after-free in USB Ethernet drivers aqc111.c, asix_devices.c, ax88179_178a.c, ch9200.c and smsc75xx.c: * If the drivers receive a link change interrupt immediately before disconnect, they raise EVENT_LINK_RESET in their (non-sleepable) ->status() callback and schedule usbnet_deferred_kevent(). * usbnet_deferred_kevent() invokes the driver's ->link_reset() callback, which calls netif_carrier_{on,off}(). * That in turn schedules the work linkwatch_event(). Because usbnet_deferred_kevent() is awaited after unregister_netdev(), netif_carrier_{on,off}() may operate on an unregistered netdev and linkwatch_event() may run after free_netdev(), causing a use-after-free. In 2010, usbnet was changed to only wait for a single instance of usbnet_deferred_kevent() instead of *all* work by commit 23f333a2bfaf ("drivers/net: don't use flush_scheduled_work()"). Unfortunately the commit neglected to move the wait back to ->ndo_stop(). Rectify that omission at long last. Reported-by: Jann Horn <jannh@xxxxxxxxxx> Link: https://lore.kernel.org/netdev/CAG48ez0MHBbENX5gCdHAUXZ7h7s20LnepBF-pa5M=7Bi-jZrEA@xxxxxxxxxxxxxx/ Reported-by: Oleksij Rempel <o.rempel@xxxxxxxxxxxxxx> Link: https://lore.kernel.org/netdev/20220315113841.GA22337@xxxxxxxxxxxxxx/ Signed-off-by: Lukas Wunner <lukas@xxxxxxxxx> Cc: stable@xxxxxxxxxxxxxxx Cc: Oliver Neukum <oneukum@xxxxxxxx> --- This seems to be Jakub's preferred way to solve the use-after-free. I've tagged the patch for net-next so that it can bake in linux-next for a few weeks, but it alternatively applies cleanly to net (in case maintainers are eager to get it into mainline). The patch supersedes the following prior attempts: usbnet: Fix use-after-free on disconnect https://lore.kernel.org/netdev/127121d9d933ebe3fc13f9f91cc33363d6a8a8ac.1649859147.git.lukas@xxxxxxxxx/ net: linkwatch: ignore events for unregistered netdevs https://lore.kernel.org/netdev/18b3541e5372bc9b9fc733d422f4e698c089077c.1650177997.git.lukas@xxxxxxxxx/ net: linkwatch: ignore events for unregistered netdevs (v2) https://lore.kernel.org/netdev/cover.1655024266.git.lukas@xxxxxxxxx/ drivers/net/usb/usbnet.c | 8 ++------ 1 file changed, 2 insertions(+), 6 deletions(-) diff --git a/drivers/net/usb/usbnet.c b/drivers/net/usb/usbnet.c index dc79811535c2..63868c1fbea4 100644 --- a/drivers/net/usb/usbnet.c +++ b/drivers/net/usb/usbnet.c @@ -849,13 +849,11 @@ int usbnet_stop (struct net_device *net) mpn = !test_and_clear_bit(EVENT_NO_RUNTIME_PM, &dev->flags); - /* deferred work (task, timer, softirq) must also stop. - * can't flush_scheduled_work() until we drop rtnl (later), - * else workers could deadlock; so make workers a NOP. - */ + /* deferred work (timer, softirq, task) must also stop */ dev->flags = 0; del_timer_sync (&dev->delay); tasklet_kill (&dev->bh); + cancel_work_sync(&dev->kevent); if (!pm) usb_autopm_put_interface(dev->intf); @@ -1619,8 +1617,6 @@ void usbnet_disconnect (struct usb_interface *intf) net = dev->net; unregister_netdev (net); - cancel_work_sync(&dev->kevent); - usb_scuttle_anchored_urbs(&dev->deferred); if (dev->driver_info->unbind) -- 2.35.2
