Re: [PATCH][BZ #16145] Reduce lock contention in __tz_convert()

[Kevin Easton <kevin at guarana dot org>]

On Tue, Feb 24, 2015 at 12:01:58AM -0500, Mike Frysinger wrote:
> On 12 Feb 2015 00:37, Kevin Easton wrote:
> > This patch is an "easy win" partial fix for BZ #16145, which notes
> > the heavy contention on tzset_lock when multiple threads are converting
> > times with localtime_r().
> > 
> > In __tz_convert(), the lock does not need to be held after
> > __tzfile_compute() / __tz_compute() have been called, so we can move the
> > unlock up.  At this point there is still significant work to be done in
> > __offtime(), so we see some improvement (in my testing with 8 cores
> > banging on localtime_r(), ~20% improvement in throughput).
> 
> my reading of __offtime is that it only operates on its arguments (or const 
> global data like __mon_yday).  it also looks expensive, so maybe we should 
> hoist the remaining call outside of holding the lock ?  can you see if that'd 
> have any measurable improvement ?
> 
>   ...
>   int offtime = 0;
>   if (!__use_tzfile)
>     offtime = __offtime (...);
> 
>   __libc_lock_lock (tzset_lock);

I don't think it's that easy, because __use_tzfile is protected by
tzset_lock.

We could call __offtime() outside the lock unconditionally, but that
would slow down the common case where __use_tzfile is set.

I'm intending to look at deeper changes to allow the entire 
__tzfile_compute() / tz_compute() to be done outside the lock, by taking
a reference on the parsed tz data.  (I think I will probably need to sort
the copyright assignment stuff for that though).

    - Kevin