race condition with drivers/char/vt.c (bug in vt_ioctl.c)
PATCH 1/1 race condition with drivers/char/vt.c (bug in vt_ioctl.c)
While debugging Ingo's RT patch, I came accross this race condition. The mainline seems to be susceptible to this bug, although it may be 1 in a 1,000,000 to happen. But those are the nastiest races.
With debugging information in the RT patch, I was able to reproduce this race several times. Enough to be able to debug it.
The race is with the tty->driver_data, tty->count and vt.c
Here's the scoop:
Process P1 opens a tty:
tty_open
--> init_dev
sets tty->count to 1P1 does what it needs to, and closes the tty.
tty_release (now showing locks)
(grabs BKL)
--> release_dev
--> tty->driver->close ==> con_close (vt.c)
(down tty_sem)
(aquire console_sem)
tty->driver_data = NULLNow process P2 opens the console:
tty_open
(block on tty_sem)
back to P1
(release console_sem)
(up tty_sem)
back to P2
(down tty_sem)
--> init_dev
tty->count++ (tty->count now == 2)
(up tty_sem)
--> tty->driver->open ==> con_open (vt.c)
(aquire console_sem)
if (tty->count == 1) (which it does not)
allocate tty->driver_data
(which doesn't happen)
(release console_sem)
And P2 goes happily along with driver_data == NULL.Now in something like vt_ioctl (which I first saw the bug)
int vt_ioctl(struct tty_struct *tty, struct file * file,
unsigned int cmd, unsigned long arg)
{
struct vc_data *vc = (struct vc_data *)tty->driver_data;
struct console_font_op op; /* used in multiple places here */
struct kbd_struct * kbd;
unsigned int console;
unsigned char ucval;
void __user *up = (void __user *)arg;
int i, perm;
console = vc->vc_num;Where here vc->vc_num could very well be (0)->vc_num.
I googled a little and found where this may have already happened in the main line kernel:
http://seclists.org/lists/linux-kernel/2005/Aug/1603.html
So here's my proposal:
Instead of checking for tty->count == 1 in con_open, which we see is
not reliable. Just check for tty->driver_data == NULL.
This should work since it should always be NULL when we need to assign it. If we switch the events of the race, so that the init_dev went first, the driver_data would not be NULL and would not need to be allocated, because after init_dev tty->count would be greater than 1 (this is assuming the case that it is already allocated) and the con_close would not deallocate it. The tty_sem and console_sem and order of events protect the tty->driver_data but not the tty->count.
Without the patch, I was able to get the system to BUG on bootup every other time. With the patch applied, I was able to bootup 6 out of 6 times without a single crash.
-- Steve
drivers/char/vt.c renamed from drivers/char/vt.c.orig +1 -1
@@ -2433,7 +2433,7 @@ static int con_open(struct tty_struct *tint ret =0 ;acquire_console_sem (); -if (tty->count ==1 ) { +if (tty->driver_data ==NULL ) { ret =vc_allocate (currcons);if (ret ==0 ) {struct vc_data *vc = vc_cons[currcons].d;
Quoted 18 lines · Steven Rostedt
I googled a little and found where this may have already happened in the main line kernel:
http://seclists.org/lists/linux-kernel/2005/Aug/1603.html
So here's my proposal:
Instead of checking for
tty->count== 1 incon_open, which we see is
not reliable. Just check fortty->driver_data== NULL.This should work since it should always be NULL when we need to assign it. If we switch the events of the race, so that the
init_devwent first, thedriver_datawould not be NULL and would not need to be allocated, because afterinit_devtty->countwould be greater than 1 (this is assuming the case that it is already allocated) and thecon_closewould not deallocate it. Thetty_semandconsole_semand order of events protect thetty->driver_databut not thetty->count.
Quoted 3 lines
Without the patch, I was able to get the system to BUG on bootup every other time. With the patch applied, I was able to bootup 6 out of 6 times without a single crash.
cool fix. I'm wondering, there's a whole lot of other 'tty->count == 1' checks in drivers/char/*.c, could some of those be racy too?
Ingo
Quoted 3 lines · Ingo Molnar
cool fix. I'm wondering, there's a whole lot of other '
tty->count== 1' checks indrivers/char/*.c, could some of those be racy too?
I checked them out. The main problem is that tty->count == 1 is not reliable in the open function call. Of the 22 files that use tty->count == 1, only 4 of them (not including vt.c) have a potential race.
Three of the 4, have the race only if an error occurred.
synclink.csynclink_cs.csynclinkmp.c
The code in these three have something like this:
static int open(struct tty_struct *tty, struct file *filp) {
...
if (/* some error */) {
retval = /* something bad */
goto cleanup;
}...
cleanup:
if (retval) {
if (tty->count == 1)
info->tty = NULL; /* tty layer will release tty struct */
if(info->count)
info->count--;
}
return retval;
}So, the info->tty has a chance of not being set to NULL. I don't know how bad that is, but this is probably a bug.
The fourth file "ip2main.c" is a little more of a problem.
It's open has the following code:
noblock:
/* first open - Assign termios structure to port */
if ( tty->count == 1 ) {
i2QueueCommands(PTYPE_INLINE, pCh, 0, 2, CMD_CTSFL_DSAB, CMD_RTSFL_DSAB);
/* Now we must send the termios settings to the loadware */
set_params( pCh, NULL );
}So, there's a chance that tty->count does not equal 1 here when it should. The way to fix this would probably be to set driver_data to NULL with some locking around it, and check that instead. Like what was done for vt.c. Since I don't have an ip2 device that I know of, I won't be fixing this code.
-- Steve
There's a similar report in Kernel Bugzilla
http://bugzilla.kernel.org/show_bug.cgi?id=4812
I was wondering what's the likelihood of tty->driver_data being NULL in vt_ioctl but never had the time to do further exploration. Your patch should fix that bug too.
Tony
Quoted 109 lines · Steven Rostedt
While debugging Ingo's RT patch, I came accross this race condition. The mainline seems to be susceptible to this bug, although it may be 1 in a 1,000,000 to happen. But those are the nastiest races.
With debugging information in the RT patch, I was able to reproduce this race several times. Enough to be able to debug it.
The race is with the
tty->driver_data,tty->countandvt.cHere's the scoop:
Process P1 opens a tty:
tty_open--> init_dev sets tty->count to 1P1 does what it needs to, and closes the tty.
tty_release(now showing locks)(grabs BKL) --> release_dev --> tty->driver->close ==> con_close (vt.c) (down tty_sem) (aquire console_sem) tty->driver_data = NULLNow process P2 opens the console:
tty_open
(block ontty_sem)back to P1
(releaseconsole_sem)
(uptty_sem)back to P2
(down tty_sem) --> init_dev tty->count++ (tty->count now == 2) (up tty_sem) --> tty->driver->open ==> con_open (vt.c) (aquire console_sem) if (tty->count == 1) (which it does not) allocate tty->driver_data (which doesn't happen) (release console_sem) And P2 goes happily along with driver_data == NULL.Now in something like
vt_ioctl(which I first saw the bug)int vt_ioctl(struct
tty_struct*tty, struct file * file,
unsigned int cmd, unsigned long arg){ struct vc_data *vc = (struct vc_data *)tty->driver_data; struct console_font_op op; /* used in multiple places here */ struct kbd_struct * kbd; unsigned int console; unsigned char ucval; void __user *up = (void __user *)arg; int i, perm; console = vc->vc_num;Where here
vc->vc_numcould very well be (0)->vc_num.I googled a little and found where this may have already happened in the main line kernel:
http://seclists.org/lists/linux-kernel/2005/Aug/1603.html
So here's my proposal:
Instead of checking for
tty->count== 1 incon_open, which we see is
not reliable. Just check fortty->driver_data== NULL.This should work since it should always be NULL when we need to assign it. If we switch the events of the race, so that the
init_devwent first, thedriver_datawould not be NULL and would not need to be allocated, because afterinit_devtty->countwould be greater than 1 (this is assuming the case that it is already allocated) and thecon_closewould not deallocate it. Thetty_semandconsole_semand order of events protect thetty->driver_databut not thetty->count.Without the patch, I was able to get the system to BUG on bootup every other time. With the patch applied, I was able to bootup 6 out of 6 times without a single crash.
-- Steve
Signed-off-by Steven Rostedt <[email protected]>drivers/char/vt.c renamed from drivers/char/vt.c.orig +1
-2@@ -2433,7 +2433,7 @@ static int con_open(struct tty_struct *tint ret =0 ;acquire_console_sem (); -if (tty->count ==1 ) { +if (tty->driver_data ==NULL ) { ret =vc_allocate (currcons);if (ret ==0 ) {struct vc_data *vc = vc_cons[currcons].d; -To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
Quoted 8 lines · Antonino A. Daplas
There's a similar report in Kernel Bugzilla
http://bugzilla.kernel.org/show_bug.cgi?id=4812
I was wondering what's the likelihood of
tty->driver_databeing NULL invt_ioctlbut never had the time to do further exploration. Your patch should fix that bug too.
Yep, that's the bug that is fixed with this patch, so you can close it out. Well, not until it's added to the mainline (akpm already added it to -mm). I wouldn't check for tty->driver_data == NULL, since that would have been hiding the problem, and not solving it.
-- Steve