Re: [PATCH] [RFC] profiler updates

On 8/06/2012 2:46 a.m., Alex Rousskov wrote:
> On 06/06/2012 07:25 PM, Amos Jeffries wrote:
>>>> + if (delta< head->best)
>>>> + head->best = delta;
>>>> + if (delta> head->worst)
>>>> + head->worst = delta;
>>>> + head->summ += delta;
>>>> + head->count++;
>>>> + head->overheads += get_tick() - stopped_;
>>> The above XProfilerNode::stop() code is not thread-safe because multiple
>>> threads may update the same "head" and you are not using atomics for
>>> *head members.
>> Ok, for now this is the second major TODO.
>> I'm torn between GCC __sync_* and C++11 atomic_* types.
>>
>> There is maybe a wrapper compat layer needed temporarily I think.
> Since SMP Squid already uses GCC __sync_* primitives and has the
> corresponding wrappers, consider reusing them for your project. See
> ipc/AtomicWord.h for a starting point.
>
> The existing wrappers can be adjusted to use C++11 atomic_* types when
> those are provided by the build environment, but that sounds like a very
> different project to me.

The __sync seem to needs a lot of individual data entry locks and
un-locks for this.
For a simpler patch I've gone with one spinlocks on the global
variables, and one on each linked-list entry of the stats accuumulation
tree. Making it atomic change to each bunch of details as a small group
rather than multiple sync and unlock operations. They are always
accessed for write as a group. Just means GCC-4.4 or later to build the
profiler.

Amos
Received on Sat Jun 09 2012 - 01:02:22 MDT