Re: [Bug] WARNING in static_key_disable_cpuslocked

[Jason Baron <jbaron@akamai.com>]

On 3/6/24 5:16 PM, Josh Poimboeuf wrote:
> On Wed, Mar 06, 2024 at 03:12:07PM -0500, Jason Baron wrote:
>>
>>
>> On 3/6/24 2:31 PM, Josh Poimboeuf wrote:
>>> On Wed, Mar 06, 2024 at 10:54:20AM -0500, Steven Rostedt wrote:
>>>> Now I guess the question is, why is something trying to disable something
>>>> that is not enabled? Is the above scenario OK? Or should the users of
>>>> static_key also prevent this?
>>>
>>> Apparently that's an allowed scenario, as the jump label code seems to
>>> be actively trying to support it.  Basically the last one "wins".
>>>
>>> See for example:
>>>
>>>     1dbb6704de91 ("jump_label: Fix concurrent static_key_enable/disable()")
>>>
>>> Also the purpose of the first atomic_read() is to do a quick test before
>>> grabbing the jump lock.  So instead of grabbing the jump lock earlier,
>>> it should actually do the first test atomically:
>>
>> Makes sense but the enable path can also set key->enabled to -1.
> 
> Ah, this code is really subtle :-/
> 
>> So I think a concurrent disable could then see the -1 in tmp and still
>> trigger the WARN.
> 
> I think this shouldn't be possible, for the same reason that
> static_key_slow_try_dec() warns on -1:  key->enabled can only be -1
> during the first enable.  And disable should never be called before
> then.

hmm, right but I think in this case the reproducer is writing to a sysfs 
file to enable/disable randomly so i'm not sure if there is anything 
that would enforce that ordering. I guess you could try the reproducer, 
I haven't really looked at it in any detail.

The code in question here is in mm/vmscan.c which actually already takes 
the local 'state_mutex' for some cases. So that could be extended I 
think easily to avoid this warning.

> 
>> So I think we could change the WARN to be:
>> WARN_ON_ONCE(tmp != 0 && tmp != -1). And also add a similar check
>> for enable if we have enable vs enable racing?
> 
> My patch subtly changed the "key->enabled > 0" to "key->enabled != 0".
> If I change that back then it should be fine.
> 
>> Although it seems like the set key->enabled to -1 while used in the inc/dec
>> API isn't really doing anything in the enable/disable part here?
>> But then the key->enabled I think has to move in front of the
>> jump_label_update() to make that part work right...
> 
> Yeah, this code needs better comments.  Let me turn it into a proper
> patch.
>