Re: [PATCH] perf lock: Fix state machine to recognize lock sequence

[Frederic Weisbecker]

On Fri, Apr 16, 2010 at 05:44:06PM +0900, Hitoshi Mitake wrote:
> Hi Ingo,
> 
> I'm developing the model to recognize the correct sequence of lock events.
> Previous state machine of perf lock was really broken.
> This patch improves it a little.
> 
> This patch prepares the array of state machine represents lock sequence for each threads.
> These state machines represent one of these sequence:
> 
>       1) acquire -> acquired -> release
>       2) acquire -> contended -> acquired -> release
>       3) acquire (w/ try) -> release
>       4) acquire (w/ read) -> release
> 
> The case of 4) is a little special.
> Double acquire of read lock is allowed, so state machine of sequence
> counts read lock number, and permit double acquire and release.
> 
> But, things are not so simple. Something of my model is still wrong.
> I counted the number of lock instances with bad sequence,
> and ratio is like this (case of tracing whoami): bad:122, total:1956



I just gave your patch a try and it's worse: almost every sequences
were reported bad (it wasn't working either before your patch :)

This is not the fault of your patch though. Actually your patch seems to
be a nice improvement.

In fact I just found two things:

1) We are working on tasks in pid basis. We should work on a task by using
its tid.
In fact we are processing the sequences of several threads in a process as
if we were dealing with a single task.

If A and B are two threads belonging to a same process, and if we have:

A: acquire lock 1, release lock 1
B: acquire lock 2, release lock 2

...then we are dealing with a random mess of sequences:

AB: acquire lock 1, acquire lock 2, release lock 1, and any kind of random
things like this.

2) I can't get lock_acquired traces. Not sure why yet...


> 
> There is another new bad thing.
> The size of array of state machine is equal to max depth lockdep defines.
> If perf lock record tries to record lock events of the programs with lots of
> system call like "perf bench sched messaging", the array will be exhausted :(



Yeah, I suggest you use a list for that in fact. The max lockdep depth may
change in the future, or become variable, so we can't relay on that.

But that's still a cool improvement.

I'm queuing this patch.

Thanks.

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/