2.4.20: scheduler issue: bad scheduling latency case

[Hiroshi Inoue]

Hi,

I have found a case which may introduce bad scheduling latency up to
10 msec (or 1/HZ sec) in task scheduler of kernel 2.4.20 at SMP machine.

In schedule(), if other CPU in the system set "need_resched" flag
of task struct within the section showed below in order to request 
rescheduling, this reschedule request can be neglected.

		case TASK_RUNNING:;
	}
*****	prev->need_resched = 0;  ***************   // begin section 

	/*
	 * this is the scheduler proper:
	 */

(Omission)

	/*
	 * from this point on nothing can prevent us from
	 * switching to the next task, save this fact in
	 * sched_data.
	 */
*****	sched_data->curr = next;   *************  // end section
	task_set_cpu(next, this_cpu);

This case seems to be very rare, but it was observed that this occurred 
several times while I compiled a Linux kernel in my environment (machine 
with 2 logical CPUs by Hyper-Threading enabled processor). 

A simple patch for this issue is attached.
Does it make sense?

diff -Nru linux-2.4.20-orig/kernel/sched.c linux-2.4.20/kernel/sched.c
--- linux-2.4.20-orig/kernel/sched.c	Fri Nov 29 08:53:15 2002
+++ linux-2.4.20/kernel/sched.c	Fri Apr 11 16:04:34 2003
@@ -625,6 +625,11 @@
 		goto repeat_schedule;
 	}
 
+	if (unlikely(prev->need_resched)) {
+		prev->need_resched = 0;
+		goto repeat_schedule;
+	}
+
 	/*
 	 * from this point on nothing can prevent us from
 	 * switching to the next task, save this fact in

Regards,
Hiroshi Inoue <[e-mail]>