Re: boot panic regression introduced in 3.5-rc7

[John Stultz <johnstul@us.ibm.com>]

On 07/29/2012 08:51 PM, CAI Qian wrote:
> The bisecting pointed out this patch caused one of dell servers boot panic.
>
>    5baefd6d84163443215f4a99f6a20f054ef11236
>    hrtimer: Update hrtimer base offsets each hrtimer_interrupt
>
> [    2.971092] WARNING: at kernel/time/clockevents.c:209 clockevents_program_event+0x10a/0x120()
> [    2.971092] Hardware name: PowerEdge M605
> [    2.971092] Modules linked in:
Looking at the dmesg:

[    0.000000] Extended CMOS year: 8200


I'm working with Prarit to try to debug the issue on the affected 
machine. He noticed part of the problem is that the offs_real was set 
to  0x7FFFFFFFFFFFFFFF, which is the same as KTIME_MAX.

I suspect from the dmesg above we're getting bad data from the CMOS 
clock, and that's then causing an overflow converting to a ktime_t 
(64bits of nanoseconds can only hold ~584 years).

I've still not quite narrowed down why this hasn't bit you earlier, 
since the same wall_to_monotonic -> ktime conversion was done in 
retrigger_next_event before the change.  Maybe something called 
settimeofday(), fixing crazy time value before you switched to highres mode?

Once I sort out this last question,  I'll try to see where we can add 
some sanity checking for this sort of thing.

thanks
-john