From mboxrd@z Thu Jan  1 00:00:00 1970
Return-Path: <linux-kernel-owner+w=401wt.eu-S933567AbXCPLhu@vger.kernel.org>
Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand
	id S933567AbXCPLhu (ORCPT <rfc822;w@1wt.eu>);
	Fri, 16 Mar 2007 07:37:50 -0400
Received: (majordomo@vger.kernel.org) by vger.kernel.org id S933580AbXCPLht
	(ORCPT <rfc822;linux-kernel-outgoing>);
	Fri, 16 Mar 2007 07:37:49 -0400
Received: from pfx2.jmh.fr ([194.153.89.55]:53209 "EHLO pfx2.jmh.fr"
	rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP
	id S933567AbXCPLht (ORCPT <rfc822;linux-kernel@vger.kernel.org>);
	Fri, 16 Mar 2007 07:37:49 -0400
From: Eric Dumazet <dada1@cosmosbay.com>
To: Pavel Emelianov <xemul@sw.ru>
Subject: Re: [RFC] kernel/pid.c pid allocation wierdness
Date: Fri, 16 Mar 2007 12:37:47 +0100
User-Agent: KMail/1.9.5
Cc: Oleg Nesterov <oleg@tv-sign.ru>,
       "Eric W. Biederman" <ebiederm@xmission.com>,
       Sukadev Bhattiprolu <sukadev@us.ibm.com>,
       Serge Hallyn <serue@us.ibm.com>,
       Linux Kernel Mailing List <linux-kernel@vger.kernel.org>,
       Linux Containers <containers@lists.osdl.org>
References: <45F7A4B3.5040005@sw.ru> <20070314153341.GA770@tv-sign.ru> <45FA7823.2040104@sw.ru>
In-Reply-To: <45FA7823.2040104@sw.ru>
MIME-Version: 1.0
Content-Type: text/plain;
  charset="iso-8859-1"
Content-Transfer-Encoding: 7bit
Content-Disposition: inline
Message-Id: <200703161237.48014.dada1@cosmosbay.com>
Sender: linux-kernel-owner@vger.kernel.org
X-Mailing-List: linux-kernel@vger.kernel.org

On Friday 16 March 2007 11:57, Pavel Emelianov wrote:
> Oleg Nesterov wrote:
> > On 03/14, Eric W. Biederman wrote:
> >> Pavel Emelianov <xemul@sw.ru> writes:
> >>> Hi.
> >>>
> >>> I'm looking at how alloc_pid() works and can't understand
> >>> one (simple/stupid) thing.
> >>>
> >>> It first kmem_cache_alloc()-s a strct pid, then calls
> >>> alloc_pidmap() and at the end it taks a global pidmap_lock()
> >>> to add new pid to hash.
> >
> > We need some global lock. pidmap_lock is already here, and it is
> > only used to protect pidmap->page allocation. Iow, it is almost
> > unused. So it was very natural to re-use it while implementing
> > pidrefs.
> >
> >>> The question is - why does alloc_pidmap() use at least
> >>> two atomic ops and potentially loop to find a zero bit
> >>> in pidmap? Why not call alloc_pidmap() under pidmap_lock
> >>> and find zero pid in pidmap w/o any loops and atomics?
> >
> > Currently we search for zero bit lockless, why do you want
> > to do it under spin_lock ?
>
> Search isn't lockless. Look:
>
> while (1) {
>    if (!test_and_set_bit(...)) {
>        atomic_dec(&nr_free);
>        return pid;
>    }
>    ...
> }
>
> we use two atomic operations to find and set a bit in a map.

The finding of the zero bit is done without lock. (Search/lookup)

Then , the reservation of the found bit (test_and_set_bit) is done, and 
decrement of nr_free. It may fail because the search was done lockless.

Finding a zero bit in a 4096 bytes array may consume about 6000 cycles on 
modern hardware. Much more on SMP/NUMA machines, or on machines where 
PAGE_SIZE is 64K instead of 4K :)

You don't want to hold pidmad_lock for so long period.