Re: [Ksummit-discuss] [TECH TOPIC] Fix devm_kzalloc, its users, or both

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

On 08/02/2015 07:30 AM, Russell King - ARM Linux wrote:
> On Sat, Aug 01, 2015 at 05:48:30PM -0700, Guenter Roeck wrote:
>> On 08/01/2015 08:18 AM, Tejun Heo wrote:
>>> So, all in all, if we actually want to fix this issue, well at least
>>> most of it, I think the only viable way would be implementing revoke
>>> semantics and drain and sever all operations on removal.  Maybe it'll
>>> end up another devm interface.
>>
>> Maybe that is completely nonsense, but how about something like
>> 	devm_get(dev);
>> to be called whenever a file used by a device is opened, and
>> 	devm_put(dev);
>> whenever it is closed ?
>
> Absolutely not.  There's two problems there:
>
> 1) You'd also need to tie the lifetime of 'dev' to the lifetime of the
>     character device, which is getting to be an absurd level of lifetime
>     complexity - we can't have 'dev' vanishing before the last user of
>     the file operations has gone away, otherwise 'dev' becomes a stale
>     pointer.
>
Wonder how many drivers access the device data structure in file operations.

> 2) Remember that devm encompasses not only driver allocations, but also
>     resources as well, like interrupts, iomem mappings and iomem resource
>     claiming.  As has already been stated by others, we don't want to delay
>     the release of all these resources.
>
Again, I wonder how many drivers access io mappings and resources in file
operations.

I understand that my idea would tie the lifetime of 'dev' to the lifetime
of a character device; I guess that was the idea. But what is the
alternative ? Also, in practice, doesn't that dependency already exist
in many cases ?

Thanks,
Guenter