Re: [PATCH] serial: digicolor: fix use-after-free on driver unbind

On Mon, Dec 22, 2025 at 07:22:14AM +0100, Greg Kroah-Hartman wrote:

On Mon, Dec 22, 2025 at 12:55:02PM +0800, Junrui Luo wrote:

quoted

The digicolor_uart_console_write() function accesses the global
digicolor_ports[] array to retrieve the uart port pointer, which
can lead to a use-after-free if the console write occurs after
the port has been removed via unbind.

digicolor_uart_remove() leaves a dangling pointer in the array.

Fix by clearing the array entry in digicolor_uart_remove() and
adding a NULL check in digicolor_uart_console_write().

Reported-by: Yuhao Jiang <redacted>
Reported-by: Junrui Luo <redacted>
Fixes: 5930cb3511df ("serial: driver for Conexant Digicolor USART")
Signed-off-by: Junrui Luo <redacted>
---
 drivers/tty/serial/digicolor-usart.c | 4 ++++
 1 file changed, 4 insertions(+)

diff --git a/drivers/tty/serial/digicolor-usart.c b/drivers/tty/serial/digicolor-usart.c
index d2482df5cb9b..5861be2072c4 100644
--- a/drivers/tty/serial/digicolor-usart.c
+++ b/drivers/tty/serial/digicolor-usart.c

@@ -397,6 +397,9 @@ static void digicolor_uart_console_write(struct console *co, const char *c,
 	unsigned long flags;
 	int locked = 1;
 
+	if (!port)
+		return;
+

What prevents port from changing right after you tested this?

Thanks for the review. You're right that there's a potentially race window
between the NULL check and port usage. I found that several drivers handle
this the same way:

- meson_uart: has NULL check in console_write, clears array in remove
- sprd_serial: Fixed similar issue in commit 99038fe75afa

This check does not fully eliminate the race window, and adding stronger
synchronization in the console_write() path may have non-trivial cost. I
am willing to adopt an alternative if one is preferred.

And who is calling unbind on a port?  Why?  That's a debuggging thing
that a developer could do, it should not be part of any normal system
operation.

Although this may most commonly occur in development or testing
environments, it is still a legitimate bug that should be fixed.

In addition, unbind may also occur in:

- Some embedded systems using runtime device tree overlays
- Certain virtualization scenarios with device passthrough

Though these are uncommon.

thanks,
Junrui Luo