From mboxrd@z Thu Jan 1 00:00:00 1970 Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S1755477Ab1JDJ0i (ORCPT ); Tue, 4 Oct 2011 05:26:38 -0400 Received: from cam-admin0.cambridge.arm.com ([217.140.96.50]:58088 "EHLO cam-admin0.cambridge.arm.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1755215Ab1JDJ0h (ORCPT ); Tue, 4 Oct 2011 05:26:37 -0400 Date: Tue, 4 Oct 2011 10:26:24 +0100 From: Catalin Marinas To: Tejun Heo Cc: "linux-kernel@vger.kernel.org" , Huajun Li , Christoph Lameter Subject: Re: [PATCH 2/4] kmemleak: Handle percpu memory allocation Message-ID: <20111004092624.GB10071@e102109-lin.cambridge.arm.com> References: <20110929105940.10660.76130.stgit@e102109-lin.cambridge.arm.com> <20110929110228.10660.87600.stgit@e102109-lin.cambridge.arm.com> <20110929192818.GB10425@mtj.dyndns.org> <20111003152126.GM18195@e102109-lin.cambridge.arm.com> <20111004075915.GE15637@dhcp-172-17-108-109.mtv.corp.google.com> <20111004090446.GA10071@e102109-lin.cambridge.arm.com> <20111004091314.GF15637@dhcp-172-17-108-109.mtv.corp.google.com> MIME-Version: 1.0 Content-Type: text/plain; charset=us-ascii Content-Disposition: inline In-Reply-To: <20111004091314.GF15637@dhcp-172-17-108-109.mtv.corp.google.com> User-Agent: Mutt/1.5.20 (2009-06-14) Sender: linux-kernel-owner@vger.kernel.org List-ID: X-Mailing-List: linux-kernel@vger.kernel.org On Tue, Oct 04, 2011 at 10:13:14AM +0100, Tejun Heo wrote: > On Tue, Oct 04, 2011 at 10:04:46AM +0100, Catalin Marinas wrote: > > > The percpu part looks fine to me but I don't know how kmemleak works > > > to judge whether the kmemleak part is okay or not. This just avoids > > > false positives from slab and would still require bumping up the early > > > log memory as # of cpus increases, right? > > > > No, there is only one kmemleak call for each __percpu pointer (to the > > specific kmemleak_*_percpu function). The kmemleak expands the percpu > > pointer into corresponding blocks for each cpu but the early log only > > stores a single call. > > Hmmm... but the following definitely seems O(#PCPU_ALLOCS * #CPUS)? > What am I missing? > > +/* > + * Log an early allocated block and populate the stack trace. > + */ > +static void early_alloc_percpu(struct early_log *log) > +{ > + unsigned int cpu; > + const void __percpu *ptr = log->ptr; > + > + for_each_possible_cpu(cpu) { > + log->ptr = per_cpu_ptr(ptr, cpu); > + early_alloc(log); > + } > +} Before kmemleak is initialised we still get memory allocations that kmemleak stores in an early_log buffer (via the log_early() function called from kmemleak_alloc_percpu). Later when kmemleak has all the data structures in place, the kmemleak_init() function goes through the early_log array and replays the previously recorded requests. The early_alloc_percpu() function is used during early_log replaying and it indeed registers every percpu memory block but the early_log is always O(#PCPU_ALLOCS). The reason we don't call kmemleak_alloc_percpu() directly during replaying is that early_alloc() also copies the previously recorded stack trace into the newly created object (otherwise all early allocations would be shown as done by kmemleak_init). Thanks. -- Catalin