On Wed, 09 Aug 2023, Greg Kroah-Hartman wrote:
From: Duoming Zhou duoming@zju.edu.cn
[ Upstream commit 1e7417c188d0a83fb385ba2dbe35fd2563f2b6f3 ]
The timer dev->stat_monitor can schedule the delayed work dev->wq and the delayed work dev->wq can also arm the dev->stat_monitor timer.
When the device is detaching, the net_device will be deallocated. but the net_device private data could still be dereferenced in delayed work or timer handler. As a result, the UAF bugs will happen.
One racy situation is shown below:
(Thread 1) | (Thread 2)lan78xx_stat_monitor() | ... | lan78xx_disconnect() lan78xx_defer_kevent() | ... ... | cancel_delayed_work_sync(&dev->wq); schedule_delayed_work() | ... (wait some time) | free_netdev(net); //free net_device lan78xx_delayedwork() | //use net_device private data | dev-> //use |
Although we use cancel_delayed_work_sync() to cancel the delayed work in lan78xx_disconnect(), it could still be scheduled in timer handler lan78xx_stat_monitor().
Another racy situation is shown below:
(Thread 1) | (Thread 2)lan78xx_delayedwork | mod_timer() | lan78xx_disconnect() | cancel_delayed_work_sync() (wait some time) | if (timer_pending(&dev->stat_monitor)) | del_timer_sync(&dev->stat_monitor); lan78xx_stat_monitor() | ... lan78xx_defer_kevent() | free_netdev(net); //free //use net_device private data| dev-> //use |
Although we use del_timer_sync() to delete the timer, the function timer_pending() returns 0 when the timer is activated. As a result, the del_timer_sync() will not be executed and the timer could be re-armed.
In order to mitigate this bug, We use timer_shutdown_sync() to shutdown the timer and then use cancel_delayed_work_sync() to cancel the delayed work. As a result, the net_device could be deallocated safely.
What's more, the dev->flags is set to EVENT_DEV_DISCONNECT in lan78xx_disconnect(). But it could still be set to EVENT_STAT_UPDATE in lan78xx_stat_monitor(). So this patch put the set_bit() behind timer_shutdown_sync().
Fixes: 77dfff5bb7e2 ("lan78xx: Fix race condition in disconnect handling")
Any idea why this stopped at linux-6.4.y? The aforementioned Fixes: commit also exists in linux-6.1.y and linux-5.15.y. I don't see any earlier backport attempts or failure reports that would otherwise explain this.
Signed-off-by: Duoming Zhou duoming@zju.edu.cn Signed-off-by: David S. Miller davem@davemloft.net Signed-off-by: Sasha Levin sashal@kernel.org
drivers/net/usb/lan78xx.c | 7 ++----- 1 file changed, 2 insertions(+), 5 deletions(-)