Return-Path: Received: (majordomo@vger.kernel.org) by vger.kernel.org via listexpand id S1756000Ab3EVQJk (ORCPT ); Wed, 22 May 2013 12:09:40 -0400 Received: from mail-lb0-f180.google.com ([209.85.217.180]:62792 "EHLO mail-lb0-f180.google.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1750956Ab3EVQJj (ORCPT ); Wed, 22 May 2013 12:09:39 -0400 MIME-Version: 1.0 In-Reply-To: <1369181381.6828.159.camel@gandalf.local.home> References: <1369158390.6828.148.camel@gandalf.local.home> <1369181381.6828.159.camel@gandalf.local.home> Date: Wed, 22 May 2013 18:09:37 +0200 Message-ID: Subject: Re: [PATCH][3.10] nohz: Fix lockup on restart from wrong error code From: Frederic Weisbecker To: Steven Rostedt Cc: LKML , Ingo Molnar , Peter Zijlstra , Thomas Gleixner , "Paul E. McKenney" , Andrew Morton , Paul Gortmaker , Tejun Heo Content-Type: text/plain; charset=ISO-8859-1 Sender: linux-kernel-owner@vger.kernel.org List-ID: X-Mailing-List: linux-kernel@vger.kernel.org Content-Length: 3093 Lines: 66 2013/5/22, Steven Rostedt : > On Tue, 2013-05-21 at 22:14 +0200, Frederic Weisbecker wrote: >> 2013/5/21 Steven Rostedt : >> > commit a382bf934449 "nohz: Assign timekeeping duty to a CPU outside the >> > full dynticks range" added a cpu notifier callback that would prevent >> > the time keeping CPU from going offline if the have_nohz_full_mask was >> > set. >> > >> > This also prevents the CPU from going offline on system reboot. >> > >> > Worse yet, the return code was -EINVAL, but the notifier does not >> > recognize error codes, and it must be wrapped by a >> > notifier_from_errno() >> > function. This means that even though the CPU would fail to go down, >> > the >> > notifier would think it succeeded, and the cpu down process would >> > continue. >> > >> > This caused two different problems. One, the migration thread after >> > moving tasks from the CPU would park itself and then a task, namely the >> > reboot task, could migrate onto that CPU. Then the reboot task spins >> > waiting for the cpu to go idle. But because the reboot task happens to >> > be spinning on the cpu its waiting for, the system hangs. >> >> Can that happen if that CPU is the boot CPU? Note this is the only >> possible timekeeper with the upstream code. > > Yep it can happen in upstream (that's all I'm using). In > tick_broadcast_setup_oneshot(), it sets the tick_do_timer_cpu to the > current CPU, which can be something other than the boot CPU. Now that > CPU wont be able to be hot plugged. Ah indeed it can happen on broadcast timer initialization. A secondary CPU then steal the duty from the boot CPU. Hmm this reminds me of this patch: https://patchwork.kernel.org/patch/2302951/ I thought it deserved some attention due its code simplification but it could also solve the issue. >> >> > >> > The other error that happened was that the sched_domain re-setup would >> > get confused, and in get_group() the cpu = cpumask_first() would >> > process >> > a mask that had nothing set, and return cpu > nr_cpu_ids. Later it >> > would >> > reference the per_cpu sg with this CPU and get a bogus pointer and >> > crash. >> >> Ouch, when are we doing this domain re-setup? I remember we >> repartition the domains after cpu down/up but I don't understand how >> that can interfere with this issue. > > I haven't looked hard enough yet, but this problem only appeared when > this bug triggered. By telling the system a CPU is offline, but still > having tasks schedule to it, causes all sorts of weird side effects. I > haven't figured out in detail how this affected the sched domains, but I > don't get the sched domain corruption after fixing this bug. Weird. but the CPU refuses to offline so how could it see itself online? Anyway if that happen again I'll have a look when I'm fully back next week. -- To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to majordomo@vger.kernel.org More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/