Re: [PATCH] tty: Serialize saved termios access with device registration

[Johan Hovold <johan@kernel.org>]

On Sun, Sep 27, 2026 at 02:41:54AM +0800, Chengfeng Ye wrote:
> tty_register_device_attr() frees saved termios data when reusing a minor
> number without serializing with tty_init_termios() or tty_save_termios().
> The tty_mutex held by the open and close paths does not protect against
> this registration-side free.
> 
> For example, an n_gsm mux reconfiguration can register a device while its
> previous tty is being released. tty_save_termios() loads the saved pointer,
> registration clears the array entry and frees the object, and
> tty_save_termios() then writes through its stale pointer.

Thanks for reporting this. The intention here was that the TTY drivers
would not be reusing a minor number while it is still in use, but that
is clearly not adhered to everywhere.

> The saved-termios
> copy in tty_init_termios() is vulnerable to the same lifetime race.

How could this happen? TTY drivers should hang up their ports during
deregistration so this shouldn't be an issue unless we have a driver bug
(which we do in n_gsm apparently).

> KASAN reported:
> 
>   BUG: KASAN: slab-use-after-free in tty_save_termios+0x39a/0x3d0
>   Write of size 44 at addr ffff8881012e8180 by task poc/114
> 
>   Call Trace:
>    tty_save_termios+0x39a/0x3d0
>    release_tty+0xb3/0x7a0
>    tty_release_struct+0xa0/0xd0
>    tty_release+0xc1b/0x11b0
>    __fput+0x2f8/0x9e0
>    fput_close_sync+0xe2/0x190
>    __x64_sys_close+0x78/0xd0
> 
>   Allocated by task 110:
>    tty_save_termios+0x2c1/0x3d0
>    release_tty+0xb3/0x7a0
>    tty_release_struct+0xa0/0xd0
>    tty_release+0xc1b/0x11b0
> 
>   Freed by task 113:
>    kfree+0x131/0x3c0
>    tty_register_device_attr+0x498/0x8b0
>    gsm_activate_mux+0xfb/0x210
>    gsmld_ioctl+0x92f/0x14d0
>    tty_ioctl+0x915/0x1240
>    __x64_sys_ioctl+0x134/0x1c0
> 
> Serialize the saved-termios copies and reset with a private mutex. Taking
> tty_mutex during registration would invert existing driver lock ordering:
> UART registration holds port->mutex, while tty_find_polling_driver() holds
> tty_mutex when calling uart_poll_init(), which takes port->mutex. Keep the
> new lock confined to the saved-termios operations, with no driver callbacks
> inside the critical sections.

I'm a bit torn about this, but I think we should avoid adding another
mutex in favouring of making the saved termios reset opt-in.

Clearing the saved termios state on device registration doesn't make
much sense if a minor number may still be in use as an open TTY would
prevent the termios from being reset.

I've just posted a patch for this here:

	https://lore.kernel.org/r/20260930091938.1715754-1-johan@kernel.org

Johan