From mboxrd@z Thu Jan 1 00:00:00 1970 From: Al Viro Subject: Re: [PATCH] VFS: br_write_lock locks on possible CPUs other than online CPUs Date: Mon, 19 Dec 2011 04:11:42 +0000 Message-ID: <20111219041142.GH2203@ZenIV.linux.org.uk> References: <1324265775.25089.20.camel@mengcong> Mime-Version: 1.0 Content-Type: text/plain; charset=us-ascii Cc: linux-kernel@vger.kernel.org, linux-fsdevel@vger.kernel.org, Nick Piggin To: mengcong Return-path: Received: from zeniv.linux.org.uk ([195.92.253.2]:50779 "EHLO ZenIV.linux.org.uk" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1752190Ab1LSELo (ORCPT ); Sun, 18 Dec 2011 23:11:44 -0500 Content-Disposition: inline In-Reply-To: <1324265775.25089.20.camel@mengcong> Sender: linux-fsdevel-owner@vger.kernel.org List-ID: On Mon, Dec 19, 2011 at 11:36:15AM +0800, mengcong wrote: > In a heavily loaded system, when frequently turning on and off CPUs, the > kernel will detect soft-lockups on multiple CPUs. The detailed bug report > is at https://lkml.org/lkml/2011/8/24/185. > > The root cause is that brlock functions, i.e. br_write_lock() and > br_write_unlock(), only locks/unlocks the per-CPU spinlock of CPUs that > are online, which means, if one online CPU is locked and then goes > offline, any later unlocking operation happens during its offline state > will not touch it; and when it goes online again, it has the incorrect > brlock state. This has been verified in current kernel. > > I can reproduce this bug on the intact 3.1 kernel. After my patch applied, > I've ran an 8-hours long test(test script provided by the bug reporter), > and no soft lockup happened again. Argh... OK, that's seriously nasty. I agree that this is broken, but your patch makes br_write_lock() very costly on kernels build with huge number of possible CPUs, even when it's run on a box with few CPUs ;-/ That sucks. Worse, AFAICS, the only way to prevent on-/off-line status changes is blocking (and both directions are bad - if the thing goes online between br_write_lock() and br_write_unlock(), we'll get spin_unlock without spin_lock). And I really don't want to make vfsmount_lock writers blocking - we *probably* could get away with that, but it'll suck very badly. Especially since we'll have that nested inside namespace_sem... Alternative is to do get_online_cpus/put_online_cpus around the stuff in fs/namespace.c, putting it *outside* everything but actual IO. We can do that (since right now vfsmount_lock is non-blocking and the only potentially blocking operations under namespace_sem is kmalloc()), but I'm not particulary comfortable doing that - I never played with the code in kernel/cpu.c and I don't know if there's anything subtle to watch out for. The same issue exists for lg_global_lock_online(), but that beast is never used (and the only remaining user of lg_global_lock() is hardly time-critical - with Miklos' patches it's only done on mount -o remount,force,ro). Nick, any comments? That's your code...