Re: [RFC 1/4] lockdep: additional lock specific information when dumping locks

[Sasha Levin <sasha.levin@oracle.com>]

On 01/12/2015 10:37 AM, Peter Zijlstra wrote:
> On Mon, Jan 12, 2015 at 10:12:38AM -0500, Sasha Levin wrote:
>> The reason for my patch is simple: 
> 
> That might have maybe been good changelog material?
> 
>> I'm fuzzing with hundreds of worker threads
>> which at some point trigger a complete system lockup for some reason.
>>
>> When lockdep dumps the list of held locks it shows that pretty much every one
>> of those threads is holding the lock which caused the lockup, which is incorrect
>> because it considers locks in the process of getting acquired as "held".
>>
>> This is my solution to that issue. I wanted to know which one of the threads is
>> really holding the lock rather than just waiting on it.
>>
>> Is there a better way to solve that problem?
> 
> Sure, think moar, if the accompanying stack trace is in the middle
> of the blocking primitive, ignore the top held lock ;-)

Tried that, it's a pain.

Consider this scenario:

Process A	|	Process B	| Process C-[...]
----------------|-----------------------|----------------
mutex_lock(x)	|			|
[busy working]	|			|
		|	mutex_lock(z)	|
		|	mutex_lock(x)	|
		|	[waiting on x]	|
		|			|	mutex_lock(z)
		|			|	[waiting on z]

So at the end of all of that I have 1000 processes waiting on 'z', while
the process that has 'z' is waiting on 'x'. So if I look at which processes
are not stuck inside a blocking primitive I'll miss on process B., and it's
link between process A and process B.

> Alternatively, make better/more use of lock_acquired() and track the
> acquire vs acquired information in the held_lock (1 bit) and look at it
> when printing.

We could do that, but then we'd lose the ability to get information out of
locks, what's the benefit of doing that?


Thanks,
Sasha