Re: [PATCH] vt: vt_ioctl: fix VT_DISALLOCATE freeing in-use virtual console

[Jiri Slaby]

On 24. 02. 20, 8:12, Eric Biggers wrote:
> From: Eric Biggers <ebiggers@xxxxxxxxxx>
> 
> The VT_DISALLOCATE ioctl can free a virtual console while tty_release()
> is still running, causing a use-after-free in con_shutdown().  This
> occurs because VT_DISALLOCATE only considers a virtual console to be
> in-use if it has a tty_struct with count > 0.  But actually when
> count == 0, the tty is still in the process of being closed.
> 
> Fix this by treating a virtual console as in-use if it has a tty_struct
> at all, even with zero count; and by making VT_DISALLOCATE take the
> tty_mutex in order to provide synchronization with release_tty().
> 
> Reproducer:
> 	#include <fcntl.h>
> 	#include <linux/vt.h>
> 	#include <sys/ioctl.h>
> 	#include <unistd.h>
> 
> 	int main()
> 	{
> 		if (fork()) {
> 			for (;;)
> 				close(open("/dev/tty5", O_RDWR));
> 		} else {
> 			int fd = open("/dev/tty10", O_RDWR);
> 
> 			for (;;)
> 				ioctl(fd, VT_DISALLOCATE, 5);
> 		}
> 	}
> 
> KASAN report:
> 	BUG: KASAN: use-after-free in con_shutdown+0x76/0x80 drivers/tty/vt/vt.c:3278
> 	Write of size 8 at addr ffff88806a4ec108 by task syz_vt/129
> 
> 	CPU: 0 PID: 129 Comm: syz_vt Not tainted 5.6.0-rc2 #11
> 	Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS ?-20191223_100556-anatol 04/01/2014
> 	Call Trace:
> 	 [...]
> 	 con_shutdown+0x76/0x80 drivers/tty/vt/vt.c:3278
> 	 release_tty+0xa8/0x410 drivers/tty/tty_io.c:1514
> 	 tty_release_struct+0x34/0x50 drivers/tty/tty_io.c:1629
> 	 tty_release+0x984/0xed0 drivers/tty/tty_io.c:1789
> 	 [...]
> 
> 	Allocated by task 129:
> 	 [...]
> 	 kzalloc include/linux/slab.h:669 [inline]
> 	 vc_allocate drivers/tty/vt/vt.c:1085 [inline]
> 	 vc_allocate+0x1ac/0x680 drivers/tty/vt/vt.c:1066
> 	 con_install+0x4d/0x3f0 drivers/tty/vt/vt.c:3229
> 	 tty_driver_install_tty drivers/tty/tty_io.c:1228 [inline]
> 	 tty_init_dev+0x94/0x350 drivers/tty/tty_io.c:1341
> 	 tty_open_by_driver drivers/tty/tty_io.c:1987 [inline]
> 	 tty_open+0x3ca/0xb30 drivers/tty/tty_io.c:2035
> 	 [...]
> 
> 	Freed by task 130:
> 	 [...]
> 	 kfree+0xbf/0x1e0 mm/slab.c:3757
> 	 vt_disallocate drivers/tty/vt/vt_ioctl.c:300 [inline]
> 	 vt_ioctl+0x16dc/0x1e30 drivers/tty/vt/vt_ioctl.c:818
> 	 tty_ioctl+0x9db/0x11b0 drivers/tty/tty_io.c:2660

That means the associated tty_port is destroyed while the tty layer
still has a tty on the top of it. That is a BUG anyway.

> Fixes: 4001d7b7fc27 ("vt: push down the tty lock so we can see what is left to tackle")
> Cc: <stable@xxxxxxxxxxxxxxx> # v3.4+
> Reported-by: syzbot+522643ab5729b0421998@xxxxxxxxxxxxxxxxxxxxxxxxx
> Signed-off-by: Eric Biggers <ebiggers@xxxxxxxxxx>
> ---
>  drivers/tty/vt/vt_ioctl.c | 6 +++++-
>  1 file changed, 5 insertions(+), 1 deletion(-)
> 
> diff --git a/drivers/tty/vt/vt_ioctl.c b/drivers/tty/vt/vt_ioctl.c
> index ee6c91ef1f6cf..57d681706fa85 100644
> --- a/drivers/tty/vt/vt_ioctl.c
> +++ b/drivers/tty/vt/vt_ioctl.c
> @@ -42,7 +42,7 @@
>  char vt_dont_switch;
>  extern struct tty_driver *console_driver;
>  
> -#define VT_IS_IN_USE(i)	(console_driver->ttys[i] && console_driver->ttys[i]->count)
> +#define VT_IS_IN_USE(i)	(console_driver->ttys[i] != NULL)
>  #define VT_BUSY(i)	(VT_IS_IN_USE(i) || i == fg_console || vc_cons[i].d == sel_cons)
>  
>  /*
> @@ -288,12 +288,14 @@ static int vt_disallocate(unsigned int vc_num)
>  	struct vc_data *vc = NULL;
>  	int ret = 0;
>  
> +	mutex_lock(&tty_mutex); /* synchronize with release_tty() */
>  	console_lock();

Is this lock dependency new or pre-existing?

Locking tty_mutex here does not sound quite right. What about switching
vc_data to proper refcounting based on tty_port? (Instead of doing
tty_port_destroy and kfree in vt_disallocate*.)

>  	if (VT_BUSY(vc_num))
>  		ret = -EBUSY;
>  	else if (vc_num)
>  		vc = vc_deallocate(vc_num);
>  	console_unlock();
> +	mutex_unlock(&tty_mutex);
>  
>  	if (vc && vc_num >= MIN_NR_CONSOLES) {
>  		tty_port_destroy(&vc->port);

thanks,
-- 
js
suse labs