On 03/18/10 00:39, Frederic Weisbecker wrote:
> On Wed, Mar 17, 2010 at 04:30:53PM +0900, Hitoshi Mitake wrote:
>> On 03/17/10 10:32, Frederic Weisbecker wrote:
>>> On Sun, Mar 14, 2010 at 07:13:55PM +0100, Peter Zijlstra wrote:
>>>> On Sun, 2010-03-14 at 19:38 +0900, Hitoshi Mitake wrote:
>>>>> Current lockdep is too complicated because,
>>>>> * dependency validation
>>>>> * statistics
>>>>> * event tracing
>>>>> are all implemented by it.
>>>>> This cause problem of overhead.
>>>>> If user enables one of them, overhead of rests part is not avoidable.
>>>>> (tracing is exception. If user enables validation or stat,
>>>>> overhead of tracing doesn't occur.)
>>>>>
>>>>> So I suggest new subsystem "lock monitor".
>>>>> This is a general purpose lock event hooking mechanism.
>>>>>
>>>>> lock monitor will be enable easy implementing and running
>>>>> these features related to lock.
>>>>>
>>>>> And I'm hoping that lock monitor will reduce overhead of perf lock.
>>>>> Because lock monitor separates dependency validation and event
>> tracing clearly,
>>>>> so calling of functions of lockdep (e.g. lock_acquire()) only for
>> validation
>>>>> will not occur lock events.
>>>>>
>>>>> I implemented it on the branch perf/inject of Frederic's
>> random-tracing tree.
>>>>> Because the branch is hottest place of lock and tracing :)
>>>>
>>>> OK, so I really don't like this much..
>>>>
>>>> Building a lockstat kernel (PROVE_LOCKING=n) should not have much more
>>>> overhead than the proposed solution, if the simple lock acquistion
>>>> tracking bothers you, you can do a patch to weaken that.
>>>>
>>>> I really really dislike how you add a monitor variable between
>>>> everything for no reason what so ever.
>>>>
>>>> You use a new rwlock_t, which is an instant fail, those things are
worse
>>>> than useless.
>>>>
>>>> You add chained indirect calls into all lock ops, that's got to hurt.
>>>
>>>
>>> Well, the idea was not bad at the first glance. It was separating
>>> lockdep and lock events codes.
>>>
>>> But indeed, the indirect calls plus the locking are not good
>>> for such a fast path.
>>>
>>> There is something else, it would be nice to keep the
>>> lockdep_map -> lockdep_class mapping so that we can
>>> do lock profiling based on classes too. So we actually
>>> need the lockdep code. What we don't need is the prove
>>> locking or the lock stats. So I guess we can have a new
>>> config to enable lock events and get rid of the prove
>>> locking / lock stat code if we don't need it.
>>>
>>>
>>
>> Thanks for your comments, Peter and Frederic.
>>
>> My main motivation of writing this patch series was that
>> some kernel codes uses lockdep functions (e.g. lock_acquire()) directly,
>> so perf lock gets a lot of trace events without actual locks (e.g.
>> might_lock_read()).
>> I think that these are confusable things for users.
>>
>> But I noticed that these events can be reduced by
>> turning off CONFIG_PROVE_LOCKING. Yeah, my patch series was
pointless... :)
>>
>> Should perf lock warn not to use with CONFIG_PROVE_LOCKING?
>
>
> Ah I see.
>
> might_lock_read() uses might_fault(), rcu, workqueues and probably
> yet some others use sequences of lock_acquire/lock_release to prove
> locking while there is actually no real lock operation involved, but
> this is to detect dependency/balance mistakes.
>
> I think that these cases are easily detectable in that they never have
> any lock_acquired in their scenario. So may be we can just ignore
> scenarios without lock_acquired and indeed advise users not to use
> PROVE_LOCKING.
Unfortunately, we cannot use this detection method.
Because trylock series (e.g. spin_trylock()) only issues
lock_acquire() like this,
static inline int __raw_spin_trylock(raw_spinlock_t *lock)
{
preempt_disable();
if (do_raw_spin_trylock(lock)) {
spin_acquire(&lock->monitor, 0, 1, _RET_IP_); <- spin_acquire() only
issues lock_acquire()
return 1;
}
preempt_enable();
return 0;
}
So distinguishing trylocks and lock_acquire()/lock_release() pairs from
might_lock_read(), might_fault() and etc is hard.
It seems that turning off PROVE_LOCKING must be required
for state machine of perf lock.
--
