(We found this (after a customer complained) and it is in the kernel.org
kernel.  Seems that for CLOCK_MONOTONIC absolute timers and clock_nanosleep
calls both the request time and wall_to_monotonic are subtracted prior to
the normalize resulting in an overflow in the existing normalize test.
This causes the result to be shifted ~4 seconds ahead instead of ~2 seconds
back in time.)
The normalize code in posix-timers.c fails when the tv_nsec member is ~1.2
seconds negative.  This can happen on absolute timers (and
clock_nanosleeps) requested on CLOCK_MONOTONIC (both the request time and
wall_to_monotonic are subtracted resulting in the possibility of a number
close to -2 seconds.)
This fix uses the set_normalized_timespec() (which does not have an
overflow problem) to fix the problem and as a side effect makes the code
cleaner.
Signed-off-by: George Anzinger <george@mvista.com>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>
                        jiffies_64_f = get_jiffies_64();
                }
                /*
-                * Take away now to get delta
+                * Take away now to get delta and normalize
                 */
-               oc.tv_sec -= now.tv_sec;
-               oc.tv_nsec -= now.tv_nsec;
-               /*
-                * Normalize...
-                */
-               while ((oc.tv_nsec - NSEC_PER_SEC) >= 0) {
-                       oc.tv_nsec -= NSEC_PER_SEC;
-                       oc.tv_sec++;
-               }
-               while ((oc.tv_nsec) < 0) {
-                       oc.tv_nsec += NSEC_PER_SEC;
-                       oc.tv_sec--;
-               }
+               set_normalized_timespec(&oc, oc.tv_sec - now.tv_sec,
+                                       oc.tv_nsec - now.tv_nsec);
        }else{
                jiffies_64_f = get_jiffies_64();
        }