linux-kernel - Re: [tip:sched/core] sched/core: Add debugging code to catch missing update_rq

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite: Windows password security audit tool. GUI, reports in PDF.

[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]

Date:   Mon, 6 Feb 2017 07:14:06 -0800
From:   "Paul E. McKenney" <paulmck@...ux.vnet.ibm.com>
To:     Sachin Sant <sachinp@...ux.vnet.ibm.com>
Cc:     Peter Zijlstra <peterz@...radead.org>,
        Matt Fleming <matt@...eblueprint.co.uk>,
        "linuxppc-dev@...ts.ozlabs.org" <linuxppc-dev@...ts.ozlabs.org>,
        Mike Galbraith <efault@....de>,
        LKML <linux-kernel@...r.kernel.org>,
        "linux-next@...r.kernel.org" <linux-next@...r.kernel.org>,
        Ross Zwisler <zwisler@...il.com>
Subject: Re: [tip:sched/core] sched/core: Add debugging code to catch missing
 update_rq_clock() calls

On Mon, Feb 06, 2017 at 07:10:48AM -0800, Paul E. McKenney wrote:
> On Mon, Feb 06, 2017 at 11:53:10AM +0530, Sachin Sant wrote:
> > 
> > >>> I've seen it on tip. It looks like hot unplug goes really slow when
> > >>> there's running tasks on the CPU being taken down.
> > >>> 
> > >>> What I did was something like:
> > >>> 
> > >>>  taskset -p $((1<<1)) $$
> > >>>  for ((i=0; i<20; i++)) do while :; do :; done & done
> > >>> 
> > >>>  taskset -p $((1<<0)) $$
> > >>>  echo 0 > /sys/devices/system/cpu/cpu1/online
> > >>> 
> > >>> And with those 20 tasks stuck sucking cycles on CPU1, the unplug goes
> > >>> _really_ slow and the RCU stall triggers. What I suspect happens is that
> > >>> hotplug stops participating in the RCU state machine early, but only
> > >>> tells RCU about it really late, and in between it gets suspicious it
> > >>> takes too long.
> > >>> 
> > >>> I've yet to dig through the RCU code to figure out the exact sequence of
> > >>> events, but found the above to be fairly reliable in triggering the
> > >>> issue.
> > > 
> > >> If you send me the full splat from the dmesg and the RCU portions of
> > >> .config, I will take a look.  Is this new behavior, or a new test?
> > > 
> > 
> > I have sent the required files to you via separate email.
> > 
> > > If new behavior, I would be most suspicious of these commits in -rcu which
> > > recently entered -tip:
> > > 
> > > 19e4d983cda1 rcu: Place guard on rcu_all_qs() and rcu_note_context_switch() actions
> > > 913324b1364f rcu: Eliminate flavor scan in rcu_momentary_dyntick_idle()
> > > fcdcfefafa45 rcu: Pull rcu_qs_ctr into rcu_dynticks structure
> > > 0919a0b7e7a5 rcu: Pull rcu_sched_qs_mask into rcu_dynticks structure
> > > caa7c8e34293 rcu: Make rcu_note_context_switch() do deferred NOCB wakeups
> > > 41e4b159d516 rcu: Make rcu_all_qs() do deferred NOCB wakeups
> > > b457a3356a68 rcu: Make call_rcu() do deferred NOCB wakeups
> > > 
> > > Does reverting any of these help?
> > 
> > I tried reverting the above commits. That does not help. I can still recreate the issue.
> 
> Thank you for testing, Sachin!
> 
> Could you please try building and testing with CONFIG_RCU_BOOST=y?
> You will need to enable CONFIG_RCU_EXPERT=y to see this Kconfig option.

Ah, but looking ahead to your .config file, you have CONFIG_PREEMPT=n,
which means boosting would not help and is not available in any case.

So it looks like there is a very long loop within an RCU read-side
critical section, and that this critical section needs to be broken
up a bit -- 21 seconds in pretty much any kind of critical section is
a bit excessive, after all.

							Thanx, Paul