Re: [PATCH 4/5] locking/lockdep: Make class->ops a percpu counter

[Waiman Long <longman@redhat.com>]

On 10/02/2018 05:55 AM, Ingo Molnar wrote:
> * Peter Zijlstra <peterz@infradead.org> wrote:
>
>> On Fri, Sep 28, 2018 at 01:53:20PM -0400, Waiman Long wrote:
>>> diff --git a/kernel/locking/lockdep.c b/kernel/locking/lockdep.c
>>> index ca002c0..7a0ed1d 100644
>>> --- a/kernel/locking/lockdep.c
>>> +++ b/kernel/locking/lockdep.c
>>> @@ -139,6 +139,7 @@ static inline int debug_locks_off_graph_unlock(void)
>>>   */
>>>  unsigned long nr_lock_classes;
>>>  static struct lock_class lock_classes[MAX_LOCKDEP_KEYS];
>>> +static DEFINE_PER_CPU(unsigned long [MAX_LOCKDEP_KEYS], lock_class_ops);
>>> @@ -1387,11 +1391,15 @@ static inline int usage_match(struct lock_list *entry, void *bit)
>>>  
>>>  static void print_lock_class_header(struct lock_class *class, int depth)
>>>  {
>>> -	int bit;
>>> +	int bit, cpu;
>>> +	unsigned long ops = 0UL;
>>> +
>>> +	for_each_possible_cpu(cpu)
>>> +		ops += *per_cpu(class->pops, cpu);
>>>  
>>>  	printk("%*s->", depth, "");
>>>  	print_lock_name(class);
>>> -	printk(KERN_CONT " ops: %lu", class->ops);
>>> +	printk(KERN_CONT " ops: %lu", ops);
>>>  	printk(KERN_CONT " {\n");
>>>  
>>>  	for (bit = 0; bit < LOCK_USAGE_STATES; bit++) {
>> That is an aweful lot of storage for a stupid number. Some archs
>> (sparc64) are bzImage size constrained and this will hurt them.
>>
>> Ingo, do you happen to remember what that number was good for?
> Just a spur of the moment statistics to satisfy curiousity, and it's useful to see how 'busy' a 
> particular class is, right?
>
>> Can't we simply ditch it?
> We certainly could. Do we have roughly equivalent metrics to arrive at this number via other 
> methods?
>
> Thanks,
>
> 	Ingo


One alternative is to group it under CONFIG_DEBUG_LOCKDEP again. This
metric was originally under CONFIG_DEBUG_LOCKDEP, but was moved to
CONFIG_LOCKDEP when trying to make other lock debugging statistics
per-cpu counters. It was probably because this metric is per lock class
while the rests are global.

By doing so, you incur the memory cost only when CONFIG_DEBUG_LOCKDEP is
defined.

What do you think?

Cheers,
Longman