indydog driver race

indydog driver race

[Dave Hansen]

[-- Attachment #1: Type: text/plain, Size: 659 bytes --]

indydog_open has a race

indydog_open(struct inode *inode, struct file *file)
{
         if( indydog_alive )
		return -EBUSY;	
	...
	 indydog_alive = 1;
         return 0;
}

If 2 opens happen simultaneously, there can be 2 tasks which both see 
indydog_alive as 0 and both set it to 1, successfully opening the 
device.  Block and char devices hold the BKL before calling the driver's 
open function, but misc devices don't do the same.

The BKL is not needed in the release function, as far as I can see. 
Alan's softdog.c driver uses atomic ops without BKL to do the same thing 
as your driver.  Patch to fix attached.

-- 
Dave Hansen
haveblue@us.ibm.com

[-- Attachment #2: indydog.bklfix.2.4.19-pre4.patch --]
[-- Type: text/plain, Size: 1035 bytes --]

--- linux-2.4.19-pre4-clean/drivers/char/indydog.c	Tue Mar  5 15:19:28 2002
+++ linux/drivers/char/indydog.c	Tue Mar 26 09:38:32 2002
@@ -24,7 +24,7 @@
 #include <asm/uaccess.h>
 #include <asm/sgi/sgimc.h>
 
-static int indydog_alive;
+static unsigned long indydog_alive;
 static struct sgimc_misc_ctrl *mcmisc_regs; 
 
 static void indydog_ping()
@@ -41,7 +41,7 @@
 {
 	u32 mc_ctrl0;
 	
-	if(indydog_alive)
+	if( test_and_set_bit(0,&indydog_alive) )
 		return -EBUSY;
 #ifdef CONFIG_WATCHDOG_NOWAYOUT
 	MOD_INC_USE_COUNT;
@@ -55,7 +55,6 @@
 	mcmisc_regs->cpuctrl0 = mc_ctrl0;
 	indydog_ping();
 			
-	indydog_alive=1;
 	printk("Started watchdog timer.\n");
 	return 0;
 }
@@ -66,7 +65,6 @@
 	 *	Shut off the timer.
 	 * 	Lock it in if it's a module and we defined ...NOWAYOUT
 	 */
-	 lock_kernel();
 #ifndef CONFIG_WATCHDOG_NOWAYOUT
 	{
 	u32 mc_ctrl0 = mcmisc_regs->cpuctrl0; 
@@ -75,8 +73,7 @@
 	printk("Stopped watchdog timer.\n");
 	}
 #endif
-	indydog_alive=0;
-	unlock_kernel();
+	clear_bit(0,&indydog_alive);
 	return 0;
 }
 

Re: indydog driver race

[Guido Guenther]

On Tue, Mar 26, 2002 at 10:01:40AM -0800, Dave Hansen wrote:
> indydog_open has a race
> 
> indydog_open(struct inode *inode, struct file *file)
> {
>          if( indydog_alive )
> 		return -EBUSY;	
> 	...
> 	 indydog_alive = 1;
>          return 0;
> }
> 
> If 2 opens happen simultaneously, there can be 2 tasks which both see 
> indydog_alive as 0 and both set it to 1, successfully opening the 
> device.  Block and char devices hold the BKL before calling the driver's 
> open function, but misc devices don't do the same.
As Dave and me discussed in private mail this can't actually happen on
non-preemtible kernels since IP22 is non-SMP(which is why I didn't
bother when I wrote it). Anyway, a race is a race, no excuses.

> 
> The BKL is not needed in the release function, as far as I can see. 
> Alan's softdog.c driver uses atomic ops without BKL to do the same thing 
> as your driver.  Patch to fix attached.
Tested and works on Mips/IP22. Thanks again Dave,
 -- Guido