2023-10-05 16:42:50

by Alexey Dobriyan

[permalink] [raw]
Subject: stack leak via uart_get_info() ?

If this check ever triggers

static int uart_get_info(struct tty_port *port, struct serial_struct *retinfo)
{

uport = uart_port_check(state);
if (!uport)
goto out;

then all those sysfs users will print stack contents to userspace.

Can it trigger while sysfs read is executing?


Signed-off-by: Alexey Dobriyan <[email protected]>
---

--- a/drivers/tty/serial/serial_core.c
+++ b/drivers/tty/serial/serial_core.c
@@ -775,6 +775,8 @@ static int uart_get_info(struct tty_port *port, struct serial_struct *retinfo)
struct uart_port *uport;
int ret = -ENODEV;

+ *retinfo = (struct serial_struct){};
+
/*
* Ensure the state we copy is consistent and no hardware changes
* occur as we go


2023-10-05 17:56:41

by Greg Kroah-Hartman

[permalink] [raw]
Subject: Re: stack leak via uart_get_info() ?

On Thu, Oct 05, 2023 at 07:34:21PM +0300, Alexey Dobriyan wrote:
> If this check ever triggers
>
> static int uart_get_info(struct tty_port *port, struct serial_struct *retinfo)
> {
>
> uport = uart_port_check(state);
> if (!uport)
> goto out;
>
> then all those sysfs users will print stack contents to userspace.
>
> Can it trigger while sysfs read is executing?

I don't think it can ever fail, we don't even check the result in other
places, so it should all be fine.

> Signed-off-by: Alexey Dobriyan <[email protected]>
> ---
>
> --- a/drivers/tty/serial/serial_core.c
> +++ b/drivers/tty/serial/serial_core.c
> @@ -775,6 +775,8 @@ static int uart_get_info(struct tty_port *port, struct serial_struct *retinfo)
> struct uart_port *uport;
> int ret = -ENODEV;
>
> + *retinfo = (struct serial_struct){};

This is good (although I hate the implied memcpy), a real memset would
be best to ensure that any holes are also filled. Want to do that, or
want me to?

thanks,

greg k-h

2023-10-05 18:45:09

by Alexey Dobriyan

[permalink] [raw]
Subject: Re: stack leak via uart_get_info() ?

On Thu, Oct 05, 2023 at 07:55:34PM +0200, Greg Kroah-Hartman wrote:
> On Thu, Oct 05, 2023 at 07:34:21PM +0300, Alexey Dobriyan wrote:
> > If this check ever triggers
> >
> > static int uart_get_info(struct tty_port *port, struct serial_struct *retinfo)
> > {
> >
> > uport = uart_port_check(state);
> > if (!uport)
> > goto out;
> >
> > then all those sysfs users will print stack contents to userspace.
> >
> > Can it trigger while sysfs read is executing?
>
> I don't think it can ever fail, we don't even check the result in other
> places, so it should all be fine.
>
> > Signed-off-by: Alexey Dobriyan <[email protected]>
> > ---
> >
> > --- a/drivers/tty/serial/serial_core.c
> > +++ b/drivers/tty/serial/serial_core.c
> > @@ -775,6 +775,8 @@ static int uart_get_info(struct tty_port *port, struct serial_struct *retinfo)
> > struct uart_port *uport;
> > int ret = -ENODEV;
> >
> > + *retinfo = (struct serial_struct){};
>
> This is good (although I hate the implied memcpy), a real memset would
> be best to ensure that any holes are also filled. Want to do that, or
> want me to?

I don't mind memset(), but "struct serial_struct" structure has kernel
pointers:

unsigned char* iomem_base;

so it is not shipped to userspace, so padding isn't an issue.