Re: [PATCH][RFC] perf lock: Distribute numerical IDs for each lock instances

[Frederic Weisbecker <fweisbec@gmail.com>]

On Mon, Dec 14, 2009 at 11:44:53PM +0900, Hitoshi Mitake wrote:
> On Mon, Dec 14, 2009 at 22:30, Frederic Weisbecker <fweisbec@gmail.com> wrote:
> > So if I understand well, this maps each lockdep_map
> > into a unique index, right?
> 
> There's a slightly difference. This patch maps each lock instances
> (spinlock_t, rwlock_t, etc) into a unique index.



Yeah.


 
> The usecase I assumed is (for example) that
> dividing copying name of lock instances to userspace from lock trace events.
> 
> I think that copying name of lock at every trace event time is not efficient.
> For example, ID <-> name table can be made in my way.
> So each lock events only have to output it's ID.
> Then, perf lock reads the table from file on debugfs.
> Finally perf lock can refer the table and obtain name of each lock.
> This may reduce the data transfer between kernel and userspace.
> 
> But... you are right. This effect can be also obtained by hashlist.
> There's no requirement of implementing array.
> And optimization should be done after implementation.
> I'll back to coding of perf lock, sorry..
> 
> # But I think that this is useful to measure the overhead of hashlist! :)
> 


Ah I understand better. Indeed if we have such index:lock_name mapping
available from debugfs, the tracing path would be more efficient because
we'd only need to trace the index, no need to copy the name.

Actually that looks like a good idea.