From mboxrd@z Thu Jan 1 00:00:00 1970 From: Linus Torvalds Subject: Re: [Bug #11342] Linux 2.6.27-rc3: kernel BUG at mm/vmalloc.c - bisected Date: Thu, 25 Sep 2008 08:42:13 -0700 (PDT) Message-ID: References: <86802c440808260136t3a33a9c8if53b6f70ab9df9e2@mail.gmail.com> <200809251150.26760.rusty@rustcorp.com.au> Mime-Version: 1.0 Return-path: In-Reply-To: <200809251150.26760.rusty-8n+1lVoiYb80n/F98K4Iww@public.gmane.org> Sender: kernel-testers-owner-u79uwXL29TY76Z2rM5mHXA@public.gmane.org List-ID: Content-Type: TEXT/PLAIN; charset="us-ascii" Content-Transfer-Encoding: 7bit To: Rusty Russell Cc: Yinghai Lu , Ingo Molnar , David Miller , Alan.Brunelle-VXdhtT5mjnY@public.gmane.org, travis-sJ/iWh9BUns@public.gmane.org, tglx-hfZtesqFncYOwBW4kG4KsQ@public.gmane.org, rjw-KKrjLPT3xs0@public.gmane.org, Linux Kernel Mailing List , kernel-testers-u79uwXL29TY76Z2rM5mHXA@public.gmane.org, Andrew Morton , arjan-VuQAYsv1563Yd54FQh9/CA@public.gmane.org, Jack Steiner On Thu, 25 Sep 2008, Rusty Russell wrote: > > This turns out to be awful in practice, mainly due to const. Consider: > > #ifdef CONFIG_CPUMASK_OFFSTACK > typedef unsigned long *cpumask_t; > #else > typedef unsigned long cpumask_t[1]; > #endif > > cpumask_t returns_cpumask(void); No. That's already broken. You cannot return a cpumask_t, regardless of interface. We must not do it regardless of how we pass those things around, since it generates _yet_ another temporary on the stack for the return slot for any kind of structure. So all cpumask functions should always return pointers and/or take pointers to be filled in. That's true *regardless* of how we actually are to then allocate them. So forget returning cpumasks. It's irrelevant. What _is_ relevant is how we allocate them when we need temporary CPU masks. And _that_ is where my suggestion comes in. For small NR_CPUS, we really do want to allocate them on the stack, because calling kmalloc for a 4- or 8-byte allocation is just _stupid_. So all your arguments are invalid, because you're looking at the wrong thing. The thing that I was talking about is converting current code that has random_function(..) { cpumask_t mask; .. do something with mask ... } which has to be converted some way. And I think it needs to be converted in a way that does *not* force us to call kmalloc() for idiotically small values. Linus