From mboxrd@z Thu Jan 1 00:00:00 1970 From: Sebastian Andrzej Siewior Subject: Re: [ANNOUNCE] 3.14.3-rt5 Date: Tue, 13 May 2014 17:40:02 +0200 Message-ID: <20140513154002.GD15049@linutronix.de> References: <20140509181214.GK29014@linutronix.de> <1399695303.5146.21.camel@marge.simpson.net> Mime-Version: 1.0 Content-Type: text/plain; charset=utf-8 Cc: linux-rt-users , LKML , Thomas Gleixner , rostedt@goodmis.org, John Kacur To: Mike Galbraith Return-path: Content-Disposition: inline In-Reply-To: <1399695303.5146.21.camel@marge.simpson.net> Sender: linux-kernel-owner@vger.kernel.org List-Id: linux-rt-users.vger.kernel.org * Mike Galbraith | 2014-05-10 06:15:03 [+0200]: >On Fri, 2014-05-09 at 20:12 +0200, Sebastian Andrzej Siewior wrote: > >> Known issues: >> >> - bcache is disabled. >> >> - lazy preempt on x86_64 leads to a crash with some load. > >That is only with NO_HZ_FUL enabled here. Box blows the stack during >task exit, eyeballing hasn't spotted the why. Even if I disable NO_HZ_FULL it explodes as soon as hackbench starts. >> - CPU hotplug works in general. Steven's test script however >> deadlocks usually on the second invocation. > >My 64 core box runs for up to 14 hours, and never deadlocks.. it >explodes in what looks like it should be an impossible manner instead. It deadlocks here and I haven't figured the exact root cause. From what it looks like is that the irq thread blocks on something during startup (migrate_disable() or so). One of the blocked irq thrad is disk driver. The userland tasks then block on ext4 in order to complete the requests. I also noticed that the frequent cpu up/down fails at some point and my kvm guest has just 7 out 8 CPUs. That one CPU remains dead and can't get back online. Once that happens, the deadlock is comming in a few minutes :) >-Mike Sebastian