Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S1752624AbdFNUba (ORCPT ); Wed, 14 Jun 2017 16:31:30 -0400 Received: from mail-qt0-f170.google.com ([209.85.216.170]:35542 "EHLO mail-qt0-f170.google.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1751760AbdFNUb2 (ORCPT ); Wed, 14 Jun 2017 16:31:28 -0400 Subject: Re: [RFC][PATCH] slub: Introduce 'alternate' per cpu partial lists To: Joonsoo Kim Cc: Christoph Lameter , Pekka Enberg , David Rientjes , Andrew Morton , linux-mm@kvack.org, linux-kernel@vger.kernel.org, Kees Cook References: <1496965984-21962-1-git-send-email-labbott@redhat.com> <20170614044528.GA5924@js1304-desktop> From: Laura Abbott Message-ID: Date: Wed, 14 Jun 2017 13:31:03 -0700 User-Agent: Mozilla/5.0 (X11; Linux x86_64; rv:52.0) Gecko/20100101 Thunderbird/52.1.0 MIME-Version: 1.0 In-Reply-To: <20170614044528.GA5924@js1304-desktop> Content-Type: text/plain; charset=utf-8 Content-Language: en-US Content-Transfer-Encoding: 8bit Sender: linux-kernel-owner@vger.kernel.org List-ID: X-Mailing-List: linux-kernel@vger.kernel.org Content-Length: 2605 Lines: 53 On 06/13/2017 09:45 PM, Joonsoo Kim wrote: > On Thu, Jun 08, 2017 at 04:53:04PM -0700, Laura Abbott wrote: >> SLUB debugging features (poisoning, red zoning etc.) skip the fast path >> completely. This ensures there is a single place to do all checks and >> take any locks that may be necessary for debugging. The overhead of some >> of the debugging features (e.g. poisoning) ends up being comparatively >> small vs the overhead of not using the fast path. >> >> We don't want to impose any kind of overhead on the fast path so >> introduce the notion of an alternate fast path. This is essentially the >> same idea as the existing fast path (store partially used pages on the >> per-cpu list) but it happens after the real fast path. Debugging that >> doesn't require locks (poisoning/red zoning) can happen on this path to >> avoid the penalty of always needing to go for the slow path. >> >> Signed-off-by: Laura Abbott >> --- >> This is a follow up to my previous proposal to speed up slub_debug=P >> https://marc.info/?l=linux-mm&m=145920558822958&w=2 . The current approach >> is hopelessly slow and can't really be used outside of limited debugging. >> The goal is to make slub_debug=P more usable for general use. >> >> Joonsoo Kim pointed out that my previous attempt still wouldn't scale >> as it still involved taking the list_lock for every allocation. He suggested >> adding per-cpu support, as did Christoph Lameter in a separate thread. This >> proposal adds a separate per-cpu list for use when poisoning is enabled. >> For this version, I'm mostly looking for general feedback about how reasonable >> this approach is before trying to clean it up more. >> >> - Some of this code is redundant and can probably be combined. >> - The fast path is very sensitive and it was suggested I leave it alone. The >> approach I took means the fastpath cmpxchg always fails before trying the >> alternate cmpxchg. From some of my profiling, the cmpxchg seemed to be fairly >> expensive. > > It looks better to modify the fastpath for non-debuging poisoning. If > we use the jump label, it doesn't cause any overhead to the fastpath > for the user who doesn't use this feature. It really makes thing > simpler. Only a few more lines will be needed in the fastpath. > When I initially tried something like that with the first version I still saw an increase in the fast path even with slub_debug=-. I might not have been testing with jump labels though and I didn't spend that much time trying to narrow down the issue. Thanks, Laura > Christoph, any opinion? > > Thanks. >