The original code relies on cancel_delayed_work() in tb_dp_dprx_stop(),
which does not ensure that the delayed work item tunnel->dprx_work has
fully completed if it was already running. This leads to use-after-free
scenarios where tb_tunnel is deallocated by tb_tunnel_put(), while
tunnel->dprx_work remains active and attempts to dereference tb_tunnel
in tb_dp_dprx_work().

A typical race condition is illustrated below:

CPU 0                            | CPU 1
tb_dp_tunnel_active()            |
  tb_deactivate_and_free_tunnel()| tb_dp_dprx_start()
    tb_tunnel_deactivate()       |   queue_delayed_work()
      tb_dp_activate()           |
        tb_dp_dprx_stop()        | tb_dp_dprx_work() //delayed worker
          cancel_delayed_work()  |
    tb_tunnel_put(tunnel);       |
                                 |   tunnel = container_of(...); //UAF
                                 |   tunnel-> //UAF

Replacing cancel_delayed_work() with cancel_delayed_work_sync() is
not feasible as it would introduce a deadlock: both tb_dp_dprx_work()
and the cleanup path acquire tb->lock, and cancel_delayed_work_sync()
would wait indefinitely for the work item that cannot proceed.

Instead, implement proper reference counting:
- If cancel_delayed_work() returns true (work is pending), we release
  the reference in the stop function.
- If it returns false (work is executing or already completed), the
  reference is released in delayed work function itself.

This ensures the tb_tunnel remains valid during work item execution
while preventing memory leaks.

This bug was found by static analysis.

Fixes: d6d458d42e1e ("thunderbolt: Handle DisplayPort tunnel activation asynchronously")
Signed-off-by: Duoming Zhou <duoming@zju.edu.cn>
---
Changes in v2:
  - Drop the local variable used to check the return value of cancel_delayed_work().

 drivers/thunderbolt/tunnel.c | 5 +++--
 1 file changed, 3 insertions(+), 2 deletions(-)

diff --git a/drivers/thunderbolt/tunnel.c b/drivers/thunderbolt/tunnel.c
index d52efe3f658c..8333fc7f3d55 100644
--- a/drivers/thunderbolt/tunnel.c
+++ b/drivers/thunderbolt/tunnel.c
@@ -1073,6 +1073,7 @@ static void tb_dp_dprx_work(struct work_struct *work)
 
 	if (tunnel->callback)
 		tunnel->callback(tunnel, tunnel->callback_data);
+	tb_tunnel_put(tunnel);
 }
 
 static int tb_dp_dprx_start(struct tb_tunnel *tunnel)
@@ -1100,8 +1101,8 @@ static void tb_dp_dprx_stop(struct tb_tunnel *tunnel)
 	if (tunnel->dprx_started) {
 		tunnel->dprx_started = false;
 		tunnel->dprx_canceled = true;
-		cancel_delayed_work(&tunnel->dprx_work);
-		tb_tunnel_put(tunnel);
+		if (cancel_delayed_work(&tunnel->dprx_work))
+			tb_tunnel_put(tunnel);
 	}
 }
 
-- 
2.34.1

Hi,

On Tue, Sep 23, 2025 at 01:13:57PM +0800, Duoming Zhou wrote:
> The original code relies on cancel_delayed_work() in tb_dp_dprx_stop(),
> which does not ensure that the delayed work item tunnel->dprx_work has
> fully completed if it was already running. This leads to use-after-free
> scenarios where tb_tunnel is deallocated by tb_tunnel_put(), while
> tunnel->dprx_work remains active and attempts to dereference tb_tunnel
> in tb_dp_dprx_work().
> 
> A typical race condition is illustrated below:
> 
> CPU 0                            | CPU 1
> tb_dp_tunnel_active()            |
>   tb_deactivate_and_free_tunnel()| tb_dp_dprx_start()
>     tb_tunnel_deactivate()       |   queue_delayed_work()
>       tb_dp_activate()           |
>         tb_dp_dprx_stop()        | tb_dp_dprx_work() //delayed worker
>           cancel_delayed_work()  |
>     tb_tunnel_put(tunnel);       |
>                                  |   tunnel = container_of(...); //UAF
>                                  |   tunnel-> //UAF
> 
> Replacing cancel_delayed_work() with cancel_delayed_work_sync() is
> not feasible as it would introduce a deadlock: both tb_dp_dprx_work()
> and the cleanup path acquire tb->lock, and cancel_delayed_work_sync()
> would wait indefinitely for the work item that cannot proceed.
> 
> Instead, implement proper reference counting:
> - If cancel_delayed_work() returns true (work is pending), we release
>   the reference in the stop function.
> - If it returns false (work is executing or already completed), the
>   reference is released in delayed work function itself.
> 
> This ensures the tb_tunnel remains valid during work item execution
> while preventing memory leaks.
> 
> This bug was found by static analysis.
> 
> Fixes: d6d458d42e1e ("thunderbolt: Handle DisplayPort tunnel activation asynchronously")
> Signed-off-by: Duoming Zhou <duoming@zju.edu.cn>

Because it is pretty late in the cycle, I would like to get some exposure
for this one before it goes to stable trees. I applied this to my
thunderbolt.git/next now. Once it is pulled into mainline this will be
picked up by the stable machinery too.

Thanks!