Re: [PATCH 2/5] watchdog: Separate and maintain variables based on variable lifetime

[Damien Riegel]

On Mon, Dec 21, 2015 at 05:10:58PM -0800, Guenter Roeck wrote:
> On 12/21/2015 09:28 AM, Damien Riegel wrote:
> >On Sun, Dec 20, 2015 at 01:05:00PM -0800, Guenter Roeck wrote:
> >>All variables required by the watchdog core to manage a watchdog are
> >>currently stored in struct watchdog_device. The lifetime of those
> >>variables is determined by the watchdog driver. However, the lifetime
> >>of variables used by the watchdog core differs from the lifetime of
> >>struct watchdog_device. To remedy this situation, watchdog drivers
> >>can implement ref and unref callbacks, to be used by the watchdog
> >>core to lock struct watchdog_device in memory.
> >>
> >>While this solves the immediate problem, it depends on watchdog drivers
> >>to actually implement the ref/unref callbacks. This is error prone,
> >>often not implemented in the first place, or not implemented correctly.
> >>
> >>To solve the problem without requiring driver support, split the variables
> >>in struct watchdog_device into two data structures - one for variables
> >>associated with the watchdog driver, one for variables associated with
> >>the watchdog core. With this approach, the watchdog core can keep track
> >>of its variable lifetime and no longer depends on ref/unref callbacks
> >>in the driver. As a side effect, some of the variables originally in
> >>struct watchdog_driver are now private to the watchdog core and no longer
> >>visible in watchdog drivers.
> >>
> >>The 'ref' and 'unref' callbacks in struct watchdog_driver are no longer
> >>used and marked as deprecated.
> >
> >Two comments below. It's great to see that unbinding a driver no longer
> >triggers a kernel panic.
> >
> It should not have caused a panic to start with, but the ref/unref functions
> for the most part were either not or wrongly implemented. Not really
> surprising - it took me a while to understand the problem.

I tested on a driver which did not implement ref/unref. When ping is
called, it tries to dereference a freed 'struct watchdog_device' in
watchdog_get_drvdata, leading to a panic. 

> 
> [ ... ]
> 
> >>
> >>  /*
> >>+ * struct _watchdog_device - watchdog core internal data
> >
> >Think it should be /**. Anyway, I find it confusing to have both
> >_watchdog_device and watchdog_device, but I can't think of a better
> >name right now.
> 
> I renamed the data structure to watchdog_data and moved it into watchdog_dev.c
> since it is only used there. No '**', though, because it is not a published
> API, but just an internal data structure.
> 
> I also renamed the matching variable name to 'wd_data' (from '_wdd').

Okay. Also, why didn't you use the explicit type for 'wdd_data' in
'struct watchdog_device' instead of a void*?

> 
> >>
> >>  static void watchdog_cdev_unregister(struct watchdog_device *wdd)
> >>  {
> >>-	mutex_lock(&wdd->lock);
> >>-	set_bit(WDOG_UNREGISTERED, &wdd->status);
> >>-	mutex_unlock(&wdd->lock);
> >>+	struct _watchdog_device *_wdd = wdd->wdd_data;
> >>
> >>-	cdev_del(&wdd->cdev);
> >>+	cdev_del(&_wdd->cdev);
> >>  	if (wdd->id == 0) {
> >>  		misc_deregister(&watchdog_miscdev);
> >>-		old_wdd = NULL;
> >>+		_old_wdd = NULL;
> >>  	}
> >>+
> >>+	if (watchdog_active(wdd))
> >>+		pr_crit("watchdog%d: watchdog still running!\n", wdd->id);
> >
> >As it is now safe to unbind and rebind a driver, it means that a
> >watchdog driver probe function can now be called with a running
> >watchdog. Some drivers handle this situation, but I think that most of
> >them expect the watchdog to be off at this point.
> >
> No semantics change, though, and no change in behavior. Drivers _should_
> handle that situation today. Sure, many don't, but that is a different issue.

All right, that's what confused me. It was, and still will be, driver
responsiblity to handle this situation.

> 
> I'll address handling an already-running watchdog by the watchdog core until
> the character device is opened in a separate patch set, but we'll have to have
> this series accepted before I re-introduce that. Even with that, it will still
> be the driver's responsibility to detect and report that/if a watchdog is
> already running.
> 
> Thanks,
> Guenter
> 

Thanks,
Damien
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/