Re: [patch] lockless, scalable get_pid(), for_each_process() elimination, 2.5.35-BK

[William Lee Irwin III <wli@holomorphy.com>]

On Wed, 18 Sep 2002, William Lee Irwin III wrote:
>> The lockups I see range from hours to "it spun over the weekend, time to
>> pull the plug".

On Wed, Sep 18, 2002 at 07:36:00PM +0200, Ingo Molnar wrote:
> this can happen if there's a genuine PID space squeeze wrt. nr_threads -
> that is solved by adding Linus' suggestion to the PID allocator. I believe
> you saw that problem, not any inherent get_pid() algorithmic inefficiency.
> nevertheless we do lock up for 32 seconds if there are 32K PIDs allocated
> in a row and last_pid hits that range - regardless of pid_max. (Depending
> on the cache architecture it could take significantly more.)

There were only 10K tasks, with likely consecutively-allocated PID's,
and some minor background fork()/exit() activity, but there are more
offenders on the read side than get_pid() itself.

There is no question of PID space: the full 2^30 was configured in
the tests done after the PID space expansion. 


Bill