sched: allow unsetting SCHED_RESET_ON_FORK when new policy is not realtime

When implementing SIGXCPU handling for a thread boosted with rtkit to realtime priority, I noticed that sched_setscheduler(somepid, SCHED_OTHER) would return -EPERM because the thread was previously SCHED_RR | SCHED_RESET_ON_FORK, and looking from the code that flag cannot be reset without sufficient rights.

sched_setscheduler(somepid, SCHED_OTHER | SCHED_RESET_ON_FORK) which worked, and seems to be the only way to remove realtime priority set by rtkit on all kernels since the patch was introduced.

This patch allows SCHED_RESET_ON_FORK to be removed when the user sets a new policy on itself without realtime.

index 2d8927f..4c0712a 100644
@@ -3757,6 +3757,12 @@ recheck:
 			if (param->sched_priority > p->rt_priority &&
 			    param->sched_priority > rlim_rtprio)
 				return -EPERM;
+
+			/* Normal users shall not reset the
+			 * sched_reset_on_fork flag
+			 */
+			if (p->sched_reset_on_fork && !reset_on_fork)
+				return -EPERM;
 		}
 
 		/*

@@ -3771,10 +3777,6 @@ recheck:
 		/* can't change other user's priorities */
 		if (!check_same_owner(p))
 			return -EPERM;
-
-		/* Normal users shall not reset the sched_reset_on_fork flag */
-		if (p->sched_reset_on_fork && !reset_on_fork)
-			return -EPERM;
 	}
 
 	if (user) {

Op 17-10-12 23:18, Maarten Lankhorst schreef:

Argh must not send patches right before going to bed, this leaves open a hole if current policy is sched_idle and when nice level is at a privileged level still..

Not that I've ever seen nice level do anything. Will resend after I've had some sleep.

~Maarten