Re: 2.6.14-rc4-rt7

[john stultz <johnstul@us.ibm.com>]

On Wed, 2005-10-26 at 21:07 -0400, Steven Rostedt wrote:
> On Wed, 2005-10-26 at 17:45 -0700, john stultz wrote:
> 
> > Ok, I've reproduced the issue. 
> > 
> > However, running a clock_gettime(CLOCK_MONOTONIC) inconsistency check
> > results in no failures, but triggers this code in the kernel.
> > 
> > Looking at the code, these may be false positives. The bit that is
> > complaining I believe Ingo added to get_monotonic_clock_ts() in
> > kernel/time/timeofday.c.  However I don't see any locking that
> > serializes the writes the prev in the same order as the
> > get_monotonic_clock_ts is called.
> > 
> > I'm still digging and will send out some mail when I figure out whats
> > wrong.
> 
> Hmm, I'm wondering if these are a false positive. Being a fully
> preemptible kernel, we could be preempted between taking now and getting
> prev, so the prev could be updated to a time after now and show a warp.
> 
> William and Rui, could you try this patch and see if you still get the
> warnings.  Although I doubt this is really the problem, since I can't
> see how it would cause clusters of these messages.

I don't know if that would really fix it, because ideally you want to
read the prev_mono_time at the same point you calculate the time inside
the read lock'ed critical section.

The difficulty is then even if you do read the prev value in a
serialized manner, you have to serialize the writes properly as well. So
really you want to do all four operations (read prev, calculate time, do
comparision, write prev) under a write lock.

Again, I'm not totally sure about this, but even removing the part where
it overwrites the ts pointer w/ the prev time, I do not see
inconsistencies from userland. 

Also I'm not seeing clusters of messages. If you call
clock_gettime(CLOCK_MONOTONIC,..) in a loop, you'll see tons of these
message.

The only odd part is the regularness of the errors. I'm seeing the ~5000
ns deltas ~every ms. So there may be something going wrong, but I don't
see it yet.

thanks
-john