Re: [PATCH 9/9] x86/UV: Add ability to disable UV NMI handler

[Mike Travis <travis@sgi.com>]

On 9/12/2013 10:27 AM, Paul E. McKenney wrote:
> On Tue, Sep 10, 2013 at 11:03:49AM +0200, Peter Zijlstra wrote:
>> On Mon, Sep 09, 2013 at 10:07:03AM -0700, Mike Travis wrote:
>>> On 9/9/2013 5:43 AM, Peter Zijlstra wrote:
>>>> On Thu, Sep 05, 2013 at 05:50:41PM -0500, Mike Travis wrote:
>>>>> For performance reasons, the NMI handler may be disabled to lessen the
>>>>> performance impact caused by the multiple perf tools running concurently.
>>>>> If the system nmi command is issued when the UV NMI handler is disabled,
>>>>> the "Dazed and Confused" messages occur for all cpus.  The NMI handler is
>>>>> disabled by setting the nmi disabled variable to '1'.  Setting it back to
>>>>> '0' will re-enable the NMI handler.
>>>>
>>>> I'm not entirely sure why this is still needed now that you've moved all
>>>> really expensive bits into the UNKNOWN handler.
>>>>
>>>
>>> Yes, it could be considered optional.  My primary use was to isolate
>>> new bugs I found to see if my NMI changes were causing them.  But it
>>> appears that they are not since the problems occur with or without
>>> using the NMI entry into KDB.  So it can be safely removed.
>>
>> OK, as a debug option it might make sense, but removing it is (of course)
>> fine with me ;-)
>>
>>> (The basic problem is that if you hang out in KDB too long the machine
>>> locks up.  
>>
>> Yeah, known issue. Not much you can do about it either I suspect. The
>> system generally isn't build for things like that.
>>
>>> Other problems like the rcu stall detector does not have a
>>> means to be "touched" like the nmi_watchdog_timer so it fires off a
>>> few to many, many messages.  
>>
>> That however might be easily cured if you ask Paul nicely ;-)
> 
> RCU's grace-period mechanism is supposed to be what touches it.  ;-)
> 
> But what is it that you are looking for?  If you want to silence it
> completely, the rcu_cpu_stall_suppress boot/sysfs parameter is what
> you want to use.

We have by default rcutree.rcu_cpu_stall_suppress=1 on the kernel
cmdline.  I'll double check if it was set during my testing.

> 
>>> Another, any network connections will time
>>> out if you are in KDB more than say 20 or 30 seconds.)
> 
> Ah, you are looking for RCU to refrain from complaining about grace
> periods that have been delayed by breakpoints in the kernel?  Is there
> some way that RCU can learn that a breakpoint has happened?  If so,
> this should not be hard.

Yes, exactly.  After a UV NMI event which might or might not call KDB,
but definitely can consume some time with the system stopped, I have
these notifications:

static void uv_nmi_touch_watchdogs(void)
{
        touch_softlockup_watchdog_sync();
        clocksource_touch_watchdog();
        rcu_cpu_stall_reset();
        touch_nmi_watchdog();
}


In all the cases I checked, I had all the cpus in the NMI event so
I don't think it was a straggler who triggered the problem.  One
question though, the above is called by all cpus exiting the NMI
event.  Should I limit that to only one cpu?

Note btw, that this also happens when KGDB/KDB is entered via the
sysrq-trigger 'g' event.

Perhaps there is some other timer that is going off?

> If not, I must fall back on the rcu_cpu_stall_suppress that I mentioned
> earlier.
> 
>>> One other problem is with the perf tool.  It seems running more than
>>> about 2 or 3 perf top instances on a medium (1k cpu threads) sized
>>> system, they start behaving badly with a bunch of NMI stackdumps
>>> appearing on the console.  Eventually the system become unusable.
>>
>> Yuck.. I haven't seen anything like that on the 'tiny' systems I have :/
> 
> Indeed, with that definition of "medium", large must be truly impressive!

I say medium because it's only one rack w/~4TB of memory (and quite
popular).  Large would be 4k cpus/64TB.  Not sure yet what is "huge",
at least in terms of an SSI system.

> 
> 							Thanx, Paul
> 
>>> On a large system (4k), the perf tools get an error message (sorry
>>> don't have it handy at the moment) the basically implies that the
>>> perf config option is not set.  Again, I wanted to remove the new
>>> NMI handler to insure that it wasn't doing something weird, and
>>> it wasn't.
>>
>> Cute.. 
>> --
>> 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/
>>
>