Content-Type: multipart/mixed; boundary="===============3581076980276315976==" MIME-Version: 1.0 From: Matt Fleming To: lkp@lists.01.org Subject: Re: [lkp-robot] [sched/fair] fbd5188493: WARNING:inconsistent_lock_state Date: Thu, 05 Jul 2018 15:43:54 +0100 Message-ID: <20180705144354.GC3864@codeblueprint.co.uk> In-Reply-To: <20180705132458.GA3864@codeblueprint.co.uk> List-Id: --===============3581076980276315976== Content-Type: text/plain; charset="utf-8" MIME-Version: 1.0 Content-Transfer-Encoding: quoted-printable On Thu, 05 Jul, at 02:24:58PM, Matt Fleming wrote: > = > Hmm.. it still looks to me like we should be saving and restoring IRQs > since this can be called from IRQ context, no? > = > The patch was a forward-port from one of our SLE kernels, and I messed > up the IRQ flag balancing for the v4.18-rc3 code :-( = Something like this? ---->8---- >From 9b152d8dadec04ac631300d86a92552e57e81db5 Mon Sep 17 00:00:00 2001 From: Matt Fleming Date: Wed, 4 Jul 2018 14:22:51 +0100 Subject: [PATCH v2] sched/fair: Avoid divide by zero when rebalancing domai= ns It's possible that the CPU doing nohz idle balance hasn't had its own load updated for many seconds. This can lead to huge deltas between rq->avg_stamp and rq->clock when rebalancing, and has been seen to cause the following crash: divide error: 0000 [#1] SMP Call Trace: [] update_sd_lb_stats+0xe8/0x560 [] find_busiest_group+0x2d/0x4b0 [] load_balance+0x170/0x950 [] rebalance_domains+0x13f/0x290 [] __do_softirq+0xec/0x300 [] irq_exit+0xfa/0x110 [] reschedule_interrupt+0xc9/0xd0 Make sure we update the rq clock and load before balancing. Cc: Ingo Molnar Cc: Mike Galbraith Cc: Peter Zijlstra Cc: Dietmar Eggemann Cc: Valentin Schneider Signed-off-by: Matt Fleming --- kernel/sched/fair.c | 11 +++++++++++ 1 file changed, 11 insertions(+) Changes in v2: Balance IRQ flags properly. diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c index 2f0a0be4d344..150b92c7c9d1 100644 --- a/kernel/sched/fair.c +++ b/kernel/sched/fair.c @@ -9676,6 +9676,7 @@ static bool nohz_idle_balance(struct rq *this_rq, enu= m cpu_idle_type idle) { int this_cpu =3D this_rq->cpu; unsigned int flags; + struct rq_flags rf; = if (!(atomic_read(nohz_flags(this_cpu)) & NOHZ_KICK_MASK)) return false; @@ -9692,6 +9693,16 @@ static bool nohz_idle_balance(struct rq *this_rq, en= um cpu_idle_type idle) if (!(flags & NOHZ_KICK_MASK)) return false; = + /* + * Ensure this_rq's clock and load are up-to-date before we + * rebalance since it's possible that they haven't been + * updated for multiple schedule periods, i.e. many seconds. + */ + rq_lock_irqsave(this_rq, &rf); + update_rq_clock(this_rq); + cpu_load_update_idle(this_rq); + rq_unlock_irqrestore(this_rq, &rf); + _nohz_idle_balance(this_rq, flags, idle); = return true; -- = 2.13.6 --===============3581076980276315976==--