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

[Guenter Roeck <linux@roeck-us.net>]

On 12/22/2015 08:09 AM, Damien Riegel wrote:
> 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.
>
Yes, that will happen. Problem here is that the driver is buggy -
pretty much all drivers which dynamically allocate struct watchdog_device
have this problem.

This is the ultimate reason for coming up with this patch.

>>
>> [ ... ]
>>
>>>>
>>>>   /*
>>>> + * 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*?
>

This is to hide the data type, since the structure is not exported
to drivers.

I could pre-declare the structure with 'struct watchdog_data;', but
then I'd have to use a different structure name (watchdog_cdev_data,
maybe, or watchdog_core_data) to make it less generic. Any opinion ?
Would that be better / preferred ? I am 50/50 about it.

Thanks,
Guenter