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

[Eric Biggers <ebiggers@kernel.org>]

On Mon, Feb 24, 2020 at 09:04:33AM +0100, Jiri Slaby wrote:
> > 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@vger.kernel.org> # v3.4+
> > Reported-by: syzbot+522643ab5729b0421998@syzkaller.appspotmail.com
> > Signed-off-by: Eric Biggers <ebiggers@google.com>
> > ---
> >  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?

It's the same locking order used during release_tty().

> 
> 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*.)
> 

How would that work?  We could make struct vc_data refcounted such that
VT_DISALLOCATE doesn't free it right away but rather it's freed in the next
con_shutdown().  But release_tty() still accesses tty->port afterwards, which is
part of the 'struct vc_data' that would have just been freed.

- Eric