ktime_t overflow from sys_nanosleep

On systems where longs are 64 bits, nanosleep can pass a timespec with tv_sec larger than 18446744073 (approx 2^34), which is then multiplied by NSEC_PER_SEC and stuffed into a single s64 by ktime_set in hrtimer_nanosleep. This overflows of course, and causes the sleep to return early (instantly if you set the timespec to its maximum).

(this gets triggered by doing a "sleep inf" with the sleep from the gnu coreutils).

I can think of two solutions, either increase the size of ktime_t to something larger than 64 bits, which would probably bring along a host of issues and and lose the entire point of having a single scalar, or we can limit the range of sys_nanosleep, which is simple, but will have to be documented and sleep(1) will still need fixing.

I've put the check in timespec_valid in the attached patch. This will limit the range of timespecs in general, but since most of them seem to end up as ktime_t's anyway, it seems like a reasonable precaution.

diff -urpN linux-2.6.17.11/include/linux/time.h linux-2.6.17.11-fvw/include/linux/time.h
@@ -68,11 +68,13 @@ extern unsigned long mktime(const unsign
 extern void set_normalized_timespec(struct timespec *ts, time_t sec, long nsec);
 
 /*
- * Returns true if the timespec is norm, false if denorm:
+ * Returns true iff the timespec nanoseconds is less than one second 
+ * ("normalised") and the seconds is in the range 0..2**31:
  */
 #define timespec_valid(ts) \
-	(((ts)->tv_sec >= 0) && (((unsigned long) (ts)->tv_nsec) < NSEC_PER_SEC))
-
+        (((ts)->tv_sec >= 0) && (((ts)->tv_sec) <= (~(1<<31))) && \
+        (((unsigned long) (ts)->tv_nsec) < NSEC_PER_SEC))
+          
 extern struct timespec xtime;
 extern struct timespec wall_to_monotonic;
 extern seqlock_t xtime_lock;

Frank v Waveren                                  Key fingerprint: BDD7 D61E
[email protected]                                              5D39 CF05 4BFC F57A
Public key: hkp://wwwkeys.pgp.net/468D62C8              FA00 7D51 468D 62C8