Re: [PATCH] Optimize sys_times for a single thread process

[Christoph Lameter <christoph@lameter.com>]

On Tue, 17 May 2005, Andrew Morton wrote:

> Well, hrm, maybe.  If this task has one sibling thread, and that thread is
> in the process of exitting then (current == next_thread(current)) may
> become true before that sibling thread has had a chance to dump its process
> accounting info into the signal structure.

The task is only "unhashed" after the counters have been added in 
__exit_signal. See release_task in kernel/exit.c

> If that dumping happens prior to the __detach_pid() call then things are
> probably OK (modulo memory ordering issues).  Otherwise there's a little
> window where the accounting will go wrong.

__exit_signal takes various locks that will insure the proper sequencing.

> Have you audited that code to ensure that the desired sequencing occurs in
> all cases and that the appropriate barriers are in place?

AFAIK release task is always called for task removal.

> It all looks a bit fast-and-loose.  If there are significant performance
> benefits and these issues are loudly commented (they aren't at present)
> then maybe-OK, I guess.

There are significant performance benefits in particular for one standard 
NUMA benchmark that keeps calling sys_times over and over. I believe other 
programs may exhibit the same brain dead behavior.

-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/