Re: CONFIG_NO_HZ + CONFIG_CPU_IDLE freeze the system (Was Re: [PATCH]acpi : remove power from acpi_processor_cx structure)

From: John Stultz
Date: Fri Sep 07 2012 - 13:26:20 EST


On 09/07/2012 07:20 AM, Daniel Lezcano wrote:
On 09/06/2012 11:18 PM, Rafael J. Wysocki wrote:
On Thursday, September 06, 2012, Daniel Lezcano wrote:
On 09/06/2012 10:04 PM, Rafael J. Wysocki wrote:
On Thursday, September 06, 2012, Daniel Lezcano wrote:
On 09/06/2012 09:54 AM, Daniel Lezcano wrote:
I fall into this issue because NETCONSOLE is set, disabling it allowed
me to go further.

Unfortunately I am facing to some random freeze on the system which
seems to be related to CONFIG_NO_HZ=y and CONFIG_CPU_IDLE=y.

Disabling one of them, make the freezes to disappear.

Is it a known issue ?
Well, there are systems having problems with this configuration, but they
should be exceptional. What system is that?
It is a laptop T61p with a Core 2 Duo T9500. Nothing exceptional I
believe. Maybe someone got the same issue ?
Is it a regression for you?
Yes, I think so. The issue appears between v3.5 and v3.6-rc1.

It is not easy to reproduce but after taking some time to dig, it seems
to appear with this commit:

1e75fa8be9fb61e1af46b5b3b176347a4c958ca1 is the first bad commit
commit 1e75fa8be9fb61e1af46b5b3b176347a4c958ca1
Author: John Stultz <john.stultz@xxxxxxxxxx>
Date: Fri Jul 13 01:21:53 2012 -0400

time: Condense timekeeper.xtime into xtime_sec

The timekeeper struct has a xtime_nsec, which keeps the
sub-nanosecond remainder. This ends up being somewhat
duplicative of the timekeeper.xtime.tv_nsec value, and we
have to do extra work to keep them apart, copying the full
nsec portion out and back in over and over.

This patch simplifies some of the logic by taking the timekeeper
xtime value and splitting it into timekeeper.xtime_sec and
reuses the timekeeper.xtime_nsec for the sub-second portion
(stored in higher res shifted nanoseconds).

This simplifies some of the accumulation logic. And will
allow for more accurate timekeeping once the vsyscall code
is updated to use the shifted nanosecond remainder.

Signed-off-by: John Stultz <john.stultz@xxxxxxxxxx>
Reviewed-by: Ingo Molnar <mingo@xxxxxxxxxx>
Cc: Peter Zijlstra <a.p.zijlstra@xxxxxxxxx>
Cc: Richard Cochran <richardcochran@xxxxxxxxx>
Cc: Prarit Bhargava <prarit@xxxxxxxxxx>
Link:
http://lkml.kernel.org/r/1342156917-25092-5-git-send-email-john.stultz@xxxxxxxxxx
Signed-off-by: Thomas Gleixner <tglx@xxxxxxxxxxxxx>

:040000 040000 4d6541ac1f6075d7adee1eef494b31a0cbda0934
dc5708bc738af695f092bf822809b13a1da104b6 M kernel

How to reproduce: with a laptop T61p, with a Core 2 Duo. I boot the
kernel in busybox and wait some minutes before writing something in the
console. At this moment, nothing appears to the console but the
characters are echo'ed several seconds later (could be 1, 5, or 10 secs
or more).

That happens when CONFIG_CPU_IDLE and CONFIG_NO_HZ are set. Disabling
one of them, the issue does not appear.

Thanks for bisecting this down and the heads up!

Right off I can't see what might be causing this. Bunch of questions:

Is this a 32 or 64 bit kernel?

By your description above, it sounds like the system is still functioning, but there's just a high latency for key-input. Is that right?

Are other things on the system happening slowly?

Does generating interrupts by hitting/holding down the ctrl key make the system respond faster?

Is there any dmesg output near when it occurs?

If you don't wait that minute after boot before typing anything, does it still trigger later? (or is it tied to early boot?)

On a whim, does the patch below avoid the problem?

thanks
-john

diff --git a/kernel/time/timekeeping.c b/kernel/time/timekeeping.c
index 34e5eac..2fa0e52 100644
--- a/kernel/time/timekeeping.c
+++ b/kernel/time/timekeeping.c
@@ -1179,6 +1179,7 @@ static void update_wall_time(void)
timekeeping_adjust(tk, offset);
+#if 0
/*
* Store only full nanoseconds into xtime_nsec after rounding
* it up and add the remainder to the error difference.
@@ -1192,6 +1193,7 @@ static void update_wall_time(void)
tk->xtime_nsec -= remainder;
tk->xtime_nsec += 1ULL << tk->shift;
tk->ntp_error += remainder << tk->ntp_error_shift;
+#endif
/*
* Finally, make sure that after the rounding

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/