Re: Linux 3.16-rc6

[Waiman Long <waiman.long@hp.com>]

On 07/25/2014 12:10 PM, Peter Zijlstra wrote:
> On Thu, Jul 24, 2014 at 04:38:28PM -0400, Waiman Long wrote:
>> Yes, I think I may have a solution for that.
>>
>> Borislav, can you apply the following patch on top of the lockdep patch to
>> see if it can fix the problem?
>>
>> diff --git a/kernel/locking/lockdep.c b/kernel/locking/lockdep.c
>> index d24e433..507a8ce 100644
>> --- a/kernel/locking/lockdep.c
>> +++ b/kernel/locking/lockdep.c
>> @@ -3595,6 +3595,12 @@ void lock_acquire(struct lockdep_map *lock, unsigned int
>>          raw_local_irq_save(flags);
>>          check_flags(flags);
>>
>> +       /*
>> +        * An interrupt recursive read in interrupt context can be considered
>> +        * to be the same as a recursive read from checking perspective.
>> +        */
>> +       if ((read == 3)&&  in_interrupt())
>> +               read = 2;
>>          current->lockdep_recursion = 1;
>>          trace_lock_acquire(lock, subclass, trylock, read, check, nest_lock, ip);
>>          __lock_acquire(lock, subclass, trylock, read, check,
> Just had another look at the initial patch and it cannot be right, even
> with the above.
>
> The problem is you cannot use in_interrupt() in check_deadlock().
> Check_deadlock() must be context invariant, it should only test the
> chain state and not rely on where or when its called.
>
>

I am planning to take out the check in check_deadlock and only have the 
test in lock_acquire which change a 3 to 2 when in interrupt context. 
Now my question is whether to do it as a new patch on top of the 
existing one in tip or a total replacement. I also intend to use 
symbolic names for the read states for better readability as suggested 
by John.

-Longman