Re: [patch] "HT scheduler", sched-2.5.63-B3

Re: [patch] "HT scheduler", sched-2.5.63-B3

[Ingo Molnar]

ok, forget the -B0 patch, it's broken - we cannot change the current
task's priority without unqueueing it first. -B1 will have the right code.

	Ingo

Re: [patch] "HT scheduler", sched-2.5.63-B3

[Mike Galbraith]

At 10:03 AM 3/7/2003 +0100, Ingo Molnar wrote:

>here's sched-2.5.64-B0:

KABOOM!  Major explosions on boot.  First time I booted, I got an instant 
reboot.  Second time, I got a couple of double faults and then explosion 
during fsck (uhoh;).  PID is 1684368482 :-/

Time to reboot into .virgin for (hopefully) a couple of fretful 
moments.  (rebooting into .virgin after first explosion would have been a 
bit smarter;)

         -Mike  

[patch] "interactivity changes", sched-2.5.64-B2

[Ingo Molnar]

i've attached the -B2 patch (against vanilla 2.5.64), which, with its
default settings, should be equivalent to what -B0 was supposed to be.
(ie. the only change is that the priority backboost is immediately
propagated into current's priority.)

-B2 also has all scheduler tunables exported into /proc/sys/kernel/, for
testing/development purposes only. The following values are now tunable:  
MIN_TIMESLICE, MAX_TIMESLICE, CHILD_PENALTY, PARENT_PENALTY, EXIT_WEIGHT,
PRIO_BONUS_RATIO, PRIO_BONUS_RATIO, INTERACTIVE_DELTA, WAKER_BONUS_RATIO,
MAX_SLEEP_AVG, STARVATION_LIMIT, NODE_THRESHOLD.

NOTE: you really have to know which value means what, some combinations
might cause a broken scheduler & cause crashes. Eg. dont even attempt to
set a smaller min timeslice than max timeslice, and dont set any of them
to 0.

NOTE2: when setting these tunables, always make sure to reach an 'idle'
scheduling state first. While it's safe to change these values on the fly,
it makes little sense to eg. set them while a kernel compile is ongoing -
the make processes might have accumulated dynamic priority already. The
recommended way to set these tunables is to stop everything in the system,
(X can stay around, it should just be mostly idle) and wait 10 seconds (or
the timeout you use for MAX_SLEEP_AVG), so that all the sleep-averages get
back to an 'idle' state. Then start whatever workload you do to test
interactivity, from scratch.

i've also implemented a new tunable, WAKER_BONUS_RATIO, which implements
Linus' idea of sharing the boost between waker and wakee in a more
immediate way. It defaults to 0 currently, but you might want to try to
set it to 50%, or 25%, or even 75%.

The -B2 patch has the following changes over -A6 (BK-curr):

 - fix a (now-) bug in kernel/softirq.c, it did a wakeup outside any
   atomic regions, which falsely identified random processes as a
   non-atomic wakeup, and which causes random priority boost to be
   distributed.

 - reset the initial idle thread's priority back to PRIO_MAX after doing
   the wakeup_forked_process() - correct preemption relies on this.

 - introduce WAKER_BONUS_RATIO

 - update current->prio immediately after a backboost.

 - clean up effective_prio() & sleep_avg calculations so that there are
   fewer RT-task special cases. This, besides resulting in a much cleaner
   WAKER_BONUS_RATIO code, also has the advantage of the sleep_avg being
   maintained even for RT tasks - this could be advantegous for tasks that
   briefly enter/exit RT mode.

	Ingo

--- linux/include/linux/sched.h.orig	
+++ linux/include/linux/sched.h	
@@ -328,7 +328,7 @@ struct task_struct {
 	prio_array_t *array;
 
 	unsigned long sleep_avg;
-	unsigned long sleep_timestamp;
+	unsigned long last_run;
 
 	unsigned long policy;
 	unsigned long cpus_allowed;
--- linux/kernel/sched.c.orig	
+++ linux/kernel/sched.c	
@@ -54,20 +54,21 @@
 /*
  * These are the 'tuning knobs' of the scheduler:
  *
- * Minimum timeslice is 10 msecs, default timeslice is 150 msecs,
- * maximum timeslice is 300 msecs. Timeslices get refilled after
+ * Minimum timeslice is 10 msecs, default timeslice is 100 msecs,
+ * maximum timeslice is 200 msecs. Timeslices get refilled after
  * they expire.
  */
-#define MIN_TIMESLICE		( 10 * HZ / 1000)
-#define MAX_TIMESLICE		(300 * HZ / 1000)
-#define CHILD_PENALTY		95
-#define PARENT_PENALTY		100
-#define EXIT_WEIGHT		3
-#define PRIO_BONUS_RATIO	25
-#define INTERACTIVE_DELTA	2
-#define MAX_SLEEP_AVG		(2*HZ)
-#define STARVATION_LIMIT	(2*HZ)
-#define NODE_THRESHOLD          125
+unsigned int MIN_TIMESLICE =		( 10 * HZ / 1000);
+unsigned int MAX_TIMESLICE =		(200 * HZ / 1000);
+unsigned int CHILD_PENALTY =		50;
+unsigned int PARENT_PENALTY =		100;
+unsigned int EXIT_WEIGHT =		3;
+unsigned int PRIO_BONUS_RATIO =		25;
+unsigned int INTERACTIVE_DELTA =	2;
+unsigned int WAKER_BONUS_RATIO =	0;
+unsigned int MAX_SLEEP_AVG =		(10*HZ);
+unsigned int STARVATION_LIMIT =		(10*HZ);
+unsigned int NODE_THRESHOLD =		125;
 
 /*
  * If a task is 'interactive' then we reinsert it in the active
@@ -302,10 +303,13 @@ static inline void enqueue_task(struct t
  *
  * Both properties are important to certain workloads.
  */
-static inline int effective_prio(task_t *p)
+static int effective_prio(task_t *p)
 {
 	int bonus, prio;
 
+	if (rt_task(p))
+		return p->prio;
+
 	bonus = MAX_USER_PRIO*PRIO_BONUS_RATIO*p->sleep_avg/MAX_SLEEP_AVG/100 -
 			MAX_USER_PRIO*PRIO_BONUS_RATIO/100/2;
 
@@ -323,26 +327,61 @@ static inline int effective_prio(task_t 
  * Also update all the scheduling statistics stuff. (sleep average
  * calculation, priority modifiers, etc.)
  */
+static inline void __activate_task(task_t *p, runqueue_t *rq)
+{
+	enqueue_task(p, rq->active);
+	rq->nr_running++;
+}
+
 static inline void activate_task(task_t *p, runqueue_t *rq)
 {
-	unsigned long sleep_time = jiffies - p->sleep_timestamp;
-	prio_array_t *array = rq->active;
+	unsigned long sleep_time = jiffies - p->last_run;
 
-	if (!rt_task(p) && sleep_time) {
+	if (sleep_time) {
+		int waker_bonus = 0;
 		/*
-		 * This code gives a bonus to interactive tasks. We update
-		 * an 'average sleep time' value here, based on
-		 * sleep_timestamp. The more time a task spends sleeping,
-		 * the higher the average gets - and the higher the priority
-		 * boost gets as well.
+		 * This code gives a bonus to interactive tasks. For
+		 * asynchronous wakeups all the bonus goes to the woken up
+		 * task. For non-atomic-context wakeups, the bonus is
+		 * shared between the waker and the woken up task. Via
+		 * this we recognize the waker as being related to the
+		 * 'interactiveness' of the woken up task.
+		 *
+		 * The boost works by updating the 'average sleep time'
+		 * value here, based on ->last_run. The more time a task
+		 * spends sleeping, the higher the average gets - and the
+		 * higher the priority boost gets as well.
 		 */
-		p->sleep_avg += sleep_time;
-		if (p->sleep_avg > MAX_SLEEP_AVG)
+		if (!in_interrupt())
+			waker_bonus = sleep_time * WAKER_BONUS_RATIO / 100;
+
+		p->sleep_avg += sleep_time - waker_bonus;
+
+		/*
+		 * 'Overflow' bonus ticks go to the waker as well, so the
+		 * ticks are not lost. This has the effect of further
+		 * boosting tasks that are related to maximum-interactive
+		 * tasks.
+		 */
+		if (p->sleep_avg > MAX_SLEEP_AVG) {
+			waker_bonus += p->sleep_avg - MAX_SLEEP_AVG;
 			p->sleep_avg = MAX_SLEEP_AVG;
-		p->prio = effective_prio(p);
+		}
+ 		p->prio = effective_prio(p);
+
+		if (!in_interrupt()) {
+			prio_array_t *array = current->array;
+			BUG_ON(!array);
+
+			current->sleep_avg += waker_bonus;
+			if (current->sleep_avg > MAX_SLEEP_AVG)
+				current->sleep_avg = MAX_SLEEP_AVG;
+			dequeue_task(current, array);
+			current->prio = effective_prio(current);
+			enqueue_task(current, array);
+		}
 	}
-	enqueue_task(p, array);
-	nr_running_inc(rq);
+	__activate_task(p, rq);
 }
 
 /*
@@ -479,10 +518,13 @@ repeat_lock_task:
 			}
 			if (old_state == TASK_UNINTERRUPTIBLE)
 				rq->nr_uninterruptible--;
-			activate_task(p, rq);
-	
-			if (p->prio < rq->curr->prio)
-				resched_task(rq->curr);
+			if (sync)
+				__activate_task(p, rq);
+			else {
+				activate_task(p, rq);
+				if (p->prio < rq->curr->prio)
+					resched_task(rq->curr);
+			}
 			success = 1;
 		}
 		p->state = TASK_RUNNING;
@@ -514,19 +556,25 @@ void wake_up_forked_process(task_t * p)
 	runqueue_t *rq = task_rq_lock(current, &flags);
 
 	p->state = TASK_RUNNING;
-	if (!rt_task(p)) {
-		/*
-		 * We decrease the sleep average of forking parents
-		 * and children as well, to keep max-interactive tasks
-		 * from forking tasks that are max-interactive.
-		 */
-		current->sleep_avg = current->sleep_avg * PARENT_PENALTY / 100;
-		p->sleep_avg = p->sleep_avg * CHILD_PENALTY / 100;
-		p->prio = effective_prio(p);
-	}
+	/*
+	 * We decrease the sleep average of forking parents
+	 * and children as well, to keep max-interactive tasks
+	 * from forking tasks that are max-interactive.
+	 */
+	current->sleep_avg = current->sleep_avg * PARENT_PENALTY / 100;
+	p->sleep_avg = p->sleep_avg * CHILD_PENALTY / 100;
+	p->prio = effective_prio(p);
 	set_task_cpu(p, smp_processor_id());
-	activate_task(p, rq);
 
+	if (unlikely(!current->array))
+		__activate_task(p, rq);
+	else {
+		p->prio = current->prio;
+		list_add_tail(&p->run_list, &current->run_list);
+		p->array = current->array;
+		p->array->nr_active++;
+		rq->nr_running++;
+	}
 	task_rq_unlock(rq, &flags);
 }
 
@@ -953,6 +1001,11 @@ static inline void pull_task(runqueue_t 
 	 */
 	if (p->prio < this_rq->curr->prio)
 		set_need_resched();
+	else {
+		if (p->prio == this_rq->curr->prio &&
+				p->time_slice > this_rq->curr->time_slice)
+			set_need_resched();
+	}
 }
 
 /*
@@ -1016,7 +1069,7 @@ skip_queue:
 	 */
 
 #define CAN_MIGRATE_TASK(p,rq,this_cpu)					\
-	((jiffies - (p)->sleep_timestamp > cache_decay_ticks) &&	\
+	((jiffies - (p)->last_run > cache_decay_ticks) &&	\
 		!task_running(rq, p) &&					\
 			((p)->cpus_allowed & (1UL << (this_cpu))))
 
@@ -1076,9 +1129,9 @@ DEFINE_PER_CPU(struct kernel_stat, kstat
  * increasing number of running tasks:
  */
 #define EXPIRED_STARVING(rq) \
-		((rq)->expired_timestamp && \
+		(STARVATION_LIMIT && ((rq)->expired_timestamp && \
 		(jiffies - (rq)->expired_timestamp >= \
-			STARVATION_LIMIT * ((rq)->nr_running) + 1))
+			STARVATION_LIMIT * ((rq)->nr_running) + 1)))
 
 /*
  * This function gets called by the timer code, with HZ frequency.
@@ -1121,6 +1174,16 @@ void scheduler_tick(int user_ticks, int 
 		return;
 	}
 	spin_lock(&rq->lock);
+	/*
+	 * The task was running during this tick - update the
+	 * time slice counter and the sleep average. Note: we
+	 * do not update a thread's priority until it either
+	 * goes to sleep or uses up its timeslice. This makes
+	 * it possible for interactive tasks to use up their
+	 * timeslices at their highest priority levels.
+	 */
+	if (p->sleep_avg)
+		p->sleep_avg--;
 	if (unlikely(rt_task(p))) {
 		/*
 		 * RR tasks need a special form of timeslice management.
@@ -1137,16 +1200,6 @@ void scheduler_tick(int user_ticks, int 
 		}
 		goto out;
 	}
-	/*
-	 * The task was running during this tick - update the
-	 * time slice counter and the sleep average. Note: we
-	 * do not update a thread's priority until it either
-	 * goes to sleep or uses up its timeslice. This makes
-	 * it possible for interactive tasks to use up their
-	 * timeslices at their highest priority levels.
-	 */
-	if (p->sleep_avg)
-		p->sleep_avg--;
 	if (!--p->time_slice) {
 		dequeue_task(p, rq->active);
 		set_tsk_need_resched(p);
@@ -1201,7 +1254,7 @@ need_resched:
 	rq = this_rq();
 
 	release_kernel_lock(prev);
-	prev->sleep_timestamp = jiffies;
+	prev->last_run = jiffies;
 	spin_lock_irq(&rq->lock);
 
 	/*
@@ -1701,7 +1754,7 @@ static int setscheduler(pid_t pid, int p
 	else
 		p->prio = p->static_prio;
 	if (array)
-		activate_task(p, task_rq(p));
+		__activate_task(p, task_rq(p));
 
 out_unlock:
 	task_rq_unlock(rq, &flags);
@@ -2442,6 +2495,7 @@ void __init sched_init(void)
 	rq->idle = current;
 	set_task_cpu(current, smp_processor_id());
 	wake_up_forked_process(current);
+	current->prio = MAX_PRIO;
 
 	init_timers();
 
--- linux/kernel/fork.c.orig	
+++ linux/kernel/fork.c	
@@ -916,7 +916,7 @@ static struct task_struct *copy_process(
 	 */
 	p->first_time_slice = 1;
 	current->time_slice >>= 1;
-	p->sleep_timestamp = jiffies;
+	p->last_run = jiffies;
 	if (!current->time_slice) {
 		/*
 	 	 * This case is rare, it happens when the parent has only
--- linux/kernel/sysctl.c.orig	
+++ linux/kernel/sysctl.c	
@@ -43,6 +43,7 @@
 
 /* External variables not in a header file. */
 extern int panic_timeout;
+extern unsigned int MIN_TIMESLICE, MAX_TIMESLICE, CHILD_PENALTY, PARENT_PENALTY, EXIT_WEIGHT, PRIO_BONUS_RATIO, PRIO_BONUS_RATIO, INTERACTIVE_DELTA, WAKER_BONUS_RATIO, MAX_SLEEP_AVG, STARVATION_LIMIT, NODE_THRESHOLD;
 extern int C_A_D;
 extern int sysctl_overcommit_memory;
 extern int sysctl_overcommit_ratio;
@@ -172,6 +173,28 @@ static ctl_table kern_table[] = {
 	 0644, NULL, &proc_doutsstring, &sysctl_string},
 	{KERN_PANIC, "panic", &panic_timeout, sizeof(int),
 	 0644, NULL, &proc_dointvec},
+	{KERN_PANIC, "MIN_TIMESLICE", &MIN_TIMESLICE, sizeof(int),
+	 0644, NULL, &proc_dointvec},
+	{KERN_PANIC, "MAX_TIMESLICE", &MAX_TIMESLICE, sizeof(int),
+	 0644, NULL, &proc_dointvec},
+	{KERN_PANIC, "CHILD_PENALTY", &CHILD_PENALTY, sizeof(int),
+	 0644, NULL, &proc_dointvec},
+	{KERN_PANIC, "PARENT_PENALTY", &PARENT_PENALTY, sizeof(int),
+	 0644, NULL, &proc_dointvec},
+	{KERN_PANIC, "EXIT_WEIGHT", &EXIT_WEIGHT, sizeof(int),
+	 0644, NULL, &proc_dointvec},
+	{KERN_PANIC, "PRIO_BONUS_RATIO", &PRIO_BONUS_RATIO, sizeof(int),
+	 0644, NULL, &proc_dointvec},
+	{KERN_PANIC, "INTERACTIVE_DELTA", &INTERACTIVE_DELTA, sizeof(int),
+	 0644, NULL, &proc_dointvec},
+	{KERN_PANIC, "WAKER_BONUS_RATIO", &WAKER_BONUS_RATIO, sizeof(int),
+	 0644, NULL, &proc_dointvec},
+	{KERN_PANIC, "MAX_SLEEP_AVG", &MAX_SLEEP_AVG, sizeof(int),
+	 0644, NULL, &proc_dointvec},
+	{KERN_PANIC, "STARVATION_LIMIT", &STARVATION_LIMIT, sizeof(int),
+	 0644, NULL, &proc_dointvec},
+	{KERN_PANIC, "NODE_THRESHOLD", &NODE_THRESHOLD, sizeof(int),
+	 0644, NULL, &proc_dointvec},
 	{KERN_CORE_USES_PID, "core_uses_pid", &core_uses_pid, sizeof(int),
 	 0644, NULL, &proc_dointvec},
 	{KERN_CORE_PATTERN, "core_pattern", core_pattern, 64,
--- linux/kernel/softirq.c.orig	
+++ linux/kernel/softirq.c	
@@ -92,10 +92,9 @@ restart:
 			mask &= ~pending;
 			goto restart;
 		}
-		__local_bh_enable();
-
 		if (pending)
 			wakeup_softirqd(cpu);
+		__local_bh_enable();
 	}
 
 	local_irq_restore(flags);

Re: [patch] "interactivity changes", sched-2.5.64-B2

[Mike Galbraith]

At 11:03 AM 3/7/2003 +0100, Ingo Molnar wrote:

>i've attached the -B2 patch (against vanilla 2.5.64).....

(I'll test and report back in an hour or so.  RL honey-do's are SCHED_RR... 
if a guy knows what's good for him;)

Re: [patch] "interactivity changes", sched-2.5.64-B2

[Mike Galbraith]

At 11:23 AM 3/7/2003 +0100, Mike Galbraith wrote:
>At 11:03 AM 3/7/2003 +0100, Ingo Molnar wrote:
>
>>i've attached the -B2 patch (against vanilla 2.5.64).....
>
>(I'll test and report back in an hour or so...

B2 appears to fix my interactivity woes w. X + make -j5 bzImage.  I can now 
waggle windows to my heart's content, browse around and click on RFC's and 
such and they pop up with nice response times.  Spiffy :)  /Me thinks 
desktop users will like this patch a bunch.

         -Mike

(I'd really like to see a multi-client db load time-to-query comparison 
though.  that make -j30 weirdness might not be a bad thing, but it makes me 
nervous [worry wart])

Re: [patch] "interactivity changes", sched-2.5.64-B2

[Mike Galbraith]

At 01:54 PM 3/7/2003 +0100, Mike Galbraith wrote:
>Spiffy :)  /Me thinks desktop users will like this patch a bunch.

(I can even play asteroids [fly little rocket ship around, shoot at and ram 
space rocks] with make -j25 bzImage and some swapping [sucks when you hit 
heavy swap of course, but quite playable as long as swap is light]) 

Re: [patch] "interactivity changes", sched-2.5.64-B2

[Ingo Molnar]

On Fri, 7 Mar 2003, Mike Galbraith wrote:

> >Spiffy :)  /Me thinks desktop users will like this patch a bunch.
> 
> (I can even play asteroids [fly little rocket ship around, shoot at and
> ram space rocks] with make -j25 bzImage and some swapping [sucks when
> you hit heavy swap of course, but quite playable as long as swap is
> light])

cool! Could you please also play a bit with various WAKER_BONUS_RATIO
values? If the default '0' value is the best-performing variant then i'll
nuke it from the patch altogether.

	Ingo

Re: [patch] "interactivity changes", sched-2.5.64-B2

[Ingo Molnar]

btw., could you please revert the kernel/softirq.c change, and re-test
-j25 interactivity with that patch? That way we'll know exactly which
component of -B2 caused the improvement on your box.

	Ingo

Re: [patch] "interactivity changes", sched-2.5.64-B2

[Mike Galbraith]

At 03:45 PM 3/7/2003 +0100, Ingo Molnar wrote:

>btw., could you please revert the kernel/softirq.c change, and re-test
>-j25 interactivity with that patch? That way we'll know exactly which
>component of -B2 caused the improvement on your box.

Done.  If either the softirq.c change or changing WAKER_BONUS_RATIO value 
(25:50:75) make any difference at all with what I'm doing, it's too close 
for me to tell.

         -Mike

Re: [patch] "interactivity changes", sched-2.5.64-B2

[Ingo Molnar]

On Fri, 7 Mar 2003, Mike Galbraith wrote:

> Done.  If either the softirq.c change or changing WAKER_BONUS_RATIO
> value (25:50:75) make any difference at all with what I'm doing, it's
> too close for me to tell.

hm. Could you then please just re-test the -A6 patch with the same base
tree again? I'd like to make sure that it _is_ the -A6 => -B0 delta that
causes this considerable level of improvement for you.

	Ingo

Re: [patch] "interactivity changes", sched-2.5.64-B2

[Linus Torvalds]

On Fri, 7 Mar 2003, Ingo Molnar wrote:
> 
> hm. Could you then please just re-test the -A6 patch with the same base
> tree again? I'd like to make sure that it _is_ the -A6 => -B0 delta that
> causes this considerable level of improvement for you.

Side note: current BK has the -B2 thing without the non-core changes (ie
no WAKER_BONUS_RATIO, and no sysctl stuff, and some reorg of the code due
to the dropping of the WAKER_BONUS_RATIO).

Ingo, you might want to double-check that you agree that my reorg is
equivalent.

		Linus

Re: [patch] "interactivity changes", sched-2.5.64-B2

[Linus Torvalds]

Ingo, 
 I already merged and pushed out the core changes in -B2, but now that I'm 
looking more closely at it I'm almost certain that it's buggy.

In particular, you do this:

	..
	dequeue_task(current, array);
	current->prio = effective_prio(current);
	enqueue_task(current, array);
	..

without actually holding the rq lock on the array!

Yes, we hold _a_ rq_lock - we hold the rq lock for the array that "p" is 
on, but that's necessarily the same as this_rq().

So you can end up with the priority bitmap and the runqueue being 
corrupted by another wakeup on another CPU that wakes up something that is 
on the runqueue for the current CPU. At which point the machine will be 
basically dead (runqueue corruption is not likely to be a survivable 
event).

Or am I missing something?

		Linus

Re: [patch] "interactivity changes", sched-2.5.64-B2

[Ingo Molnar]

you are right - i'm working on a fix.

	Ingo

Re: [patch] "interactivity changes", sched-2.5.64-B2

[Ingo Molnar]

the attached patch (against your tree) fixes the SMP locking bug. The
patch also includes some other stuff i already added to my tree by that
time:

 - only update the priority and do a requeueing if the sleep average has
   changed. (this does not happen for pure CPU hogs or pure interactive
   tasks, so no need to requeue/recalc-prio in that case.) [All the 
   necessary values are available at that point already, so gcc should
   have an easy job making this branch really cheap.]

 - do not do a full task activation in the migration-thread path - that is
   supposed to be near-atomic anyway.

 - fix up comments

i solved the SMP locking bug by moving the requeueing outside of
try_to_wake_up(). It does not matter that the priority update is not
atomically done now, since the current process wont do anything inbetween.  
(well, it could get preempted in a preemptible kernel, but even that wont
do any harm.)

did a quick testcompile & testboot on an SMP box, appears to work.

	Ingo

--- linux/kernel/sched.c.orig	
+++ linux/kernel/sched.c	
@@ -321,10 +321,7 @@ static int effective_prio(task_t *p)
 }
 
 /*
- * activate_task - move a task to the runqueue.
-
- * Also update all the scheduling statistics stuff. (sleep average
- * calculation, priority modifiers, etc.)
+ * __activate_task - move a task to the runqueue.
  */
 static inline void __activate_task(task_t *p, runqueue_t *rq)
 {
@@ -332,9 +329,16 @@ static inline void __activate_task(task_
 	nr_running_inc(rq);
 }
 
-static inline void activate_task(task_t *p, runqueue_t *rq)
+/*
+ * activate_task - move a task to the runqueue and do priority recalculation
+ *
+ * Update all the scheduling statistics stuff. (sleep average
+ * calculation, priority modifiers, etc.)
+ */
+static inline int activate_task(task_t *p, runqueue_t *rq)
 {
 	unsigned long sleep_time = jiffies - p->last_run;
+	int requeue_waker = 0;
 
 	if (sleep_time) {
 		int sleep_avg;
@@ -357,23 +361,25 @@ static inline void activate_task(task_t 
 		 */
 		if (sleep_avg > MAX_SLEEP_AVG) {
 			if (!in_interrupt()) {
-				prio_array_t *array = current->array;
-				BUG_ON(!array);
 				sleep_avg += current->sleep_avg - MAX_SLEEP_AVG;
 				if (sleep_avg > MAX_SLEEP_AVG)
 					sleep_avg = MAX_SLEEP_AVG;
 
-				current->sleep_avg = sleep_avg;
-				dequeue_task(current, array);
-				current->prio = effective_prio(current);
-				enqueue_task(current, array);
+				if (current->sleep_avg != sleep_avg) {
+					current->sleep_avg = sleep_avg;
+					requeue_waker = 1;
+				}
 			}
 			sleep_avg = MAX_SLEEP_AVG;
 		}
-		p->sleep_avg = sleep_avg;
-		p->prio = effective_prio(p);
+		if (p->sleep_avg != sleep_avg) {
+			p->sleep_avg = sleep_avg;
+			p->prio = effective_prio(p);
+		}
 	}
 	__activate_task(p, rq);
+
+	return requeue_waker;
 }
 
 /*
@@ -486,8 +492,8 @@ void kick_if_running(task_t * p)
  */
 static int try_to_wake_up(task_t * p, unsigned int state, int sync)
 {
+	int success = 0, requeue_waker = 0;
 	unsigned long flags;
-	int success = 0;
 	long old_state;
 	runqueue_t *rq;
 
@@ -513,7 +519,7 @@ repeat_lock_task:
 			if (sync)
 				__activate_task(p, rq);
 			else {
-				activate_task(p, rq);
+				requeue_waker = activate_task(p, rq);
 				if (p->prio < rq->curr->prio)
 					resched_task(rq->curr);
 			}
@@ -523,6 +529,21 @@ repeat_lock_task:
 	}
 	task_rq_unlock(rq, &flags);
 
+	/*
+	 * We have to do this outside the other spinlock, the two
+	 * runqueues might be different:
+	 */
+	if (requeue_waker) {
+		prio_array_t *array;
+
+		rq = task_rq_lock(current, &flags);
+		array = current->array;
+		dequeue_task(current, array);
+		current->prio = effective_prio(current);
+		enqueue_task(current, array);
+		task_rq_unlock(rq, &flags);
+	}
+
 	return success;
 }
 
@@ -2360,7 +2381,7 @@ repeat:
 			set_task_cpu(p, cpu_dest);
 			if (p->array) {
 				deactivate_task(p, rq_src);
-				activate_task(p, rq_dest);
+				__activate_task(p, rq_dest);
 				if (p->prio < rq_dest->curr->prio)
 					resched_task(rq_dest->curr);
 			}

Re: [patch] "interactivity changes", sched-2.5.64-B2

[Martin Josefsson]

On Fri, 2003-03-07 at 15:43, Ingo Molnar wrote:
> On Fri, 7 Mar 2003, Mike Galbraith wrote:
> 
> > >Spiffy :)  /Me thinks desktop users will like this patch a bunch.
> > 
> > (I can even play asteroids [fly little rocket ship around, shoot at and
> > ram space rocks] with make -j25 bzImage and some swapping [sucks when
> > you hit heavy swap of course, but quite playable as long as swap is
> > light])
> 
> cool! Could you please also play a bit with various WAKER_BONUS_RATIO
> values? If the default '0' value is the best-performing variant then i'll
> nuke it from the patch altogether.

Hi,

I've tested 2.5.64 + sched-2.5.64-B2 on an UP pIII 700 with 704MB ram,
here's my results...

General impression is that interactivity is good but it seems there's a
few cases that gets worse, sometimes a lot worse.

First the things that got better...

Xmms doesn't skip when switching tabs in galeon anymore during general
desktopusage (more on that below)
Earlier I had to change some schedulersettings to get it to stop
skipping.

And it was possible to watch a movie using mplayer while running a make
-j10 on a kerneltree, the playback wasn't 100% perfect but I'd say it
was ~90% , the jerkyness I saw wasn't very irritating. I tried the
window wiggle-test while the compile and movie was playing and it was
smooth and only added a little more jerkyness to the playback.


Some negative things:

Xmms could skip during the compile/movie playback, most notably one
rather long skip right after a song-change but sometimes in the middle
of a song as well, those skips were quite long as well, 2-3 seconds.

Another xmms related problem is that even without any load on the
machine the mousepointer gets jerky for 0.1-0.5 seconds right after
having used the mouse to search forward or backwards in a song. It gets
jerky at the moment you release the mousebutton.
Basically no change in the jerkyness even with a make -j10 and movie
playing.
 
The biggest problem is the fact that my windowmanager gets unresponsive
for ~30 seconds when I perform something that works fine with vanilla
2.4 and 2.5. I have two keybindings that execute 'aumix -v+3' and 
'aumix -v-3'. And if I hold one of those buttons down for say 2 seconds
then at first it seems like nothing is happening but then the volume
starts to inc-/decrease sloowly and the windowmanager (sawfish) is
unresponsive for ~30 seconds. With 2.4 or vanilla 2.5 the change in
volume is instant and I've never noticed that sawfish got unresponsive.
Sure holding the button down for 2 seconds at a repeatrate of 30/s will
cause a lot of aumix's to be executed but 30 seconds for 60 aumix's is
quite a long time. A few times it has been enough to just press one of
those keys a few times to get the exact same effect. It's as if sawfish
goes into a weird loop. All this might be a bug in sawfish...
This behaviour is the same with or without the compile/movie running.

Another problem is that the composer-window in evolution gets
unresponsive for 1-2 seconds if you hold down a key, and when it gets
responsive again it gets all the keypresses...

I've tried changing WAKER_BONUS_RATIO but I didn't see any diffrence.
At least it didn't help with the Xmms mouse-jerkyness or the
sawfish_goes_out_for_coffee_and_a_cigarette problem.

There, that was my impressions of 2 hours usage of the patch.

-- 
/Martin

Never argue with an idiot. They drag you down to their level, then beat you with experience.

Re: [patch] "interactivity changes", sched-2.5.64-B2

[Ingo Molnar]

On 7 Mar 2003, Martin Josefsson wrote:

> Some negative things:

if you have time/interest, could you re-test the negative things with X
reniced to -10, to further isolate the problem? Another thing to try is to
renice xmms to -10 (or RT priority). Which one makes the larger
difference?

	Ingo

Re: [patch] "interactivity changes", sched-2.5.64-B2

[David Lang]

sounds like the code in X that detects a key being held down is running
into problems. any chance it's doing a busy loop or something silly like
that that's just not running enough? (probably not, but since you have
problems in a couple applications that happen when you hold a key down I
would look there rather then at the scheduling code itself)

David Lang

 On 7 Mar 2003, Martin Josefsson wrote:

> Date: 07 Mar 2003 21:34:38 +0100
> From: Martin Josefsson <gandalf@wlug.westbo.se>
> To: Ingo Molnar <mingo@elte.hu>
> Cc: Mike Galbraith <efault@gmx.de>, Andrew Morton <akpm@digeo.com>,
>      Linus Torvalds <torvalds@transmeta.com>, Robert Love <rml@tech9.net>,
>      linux-kernel@vger.kernel.org
> Subject: Re: [patch] "interactivity changes", sched-2.5.64-B2
>
> On Fri, 2003-03-07 at 15:43, Ingo Molnar wrote:
> > On Fri, 7 Mar 2003, Mike Galbraith wrote:
> >
> > > >Spiffy :)  /Me thinks desktop users will like this patch a bunch.
> > >
> > > (I can even play asteroids [fly little rocket ship around, shoot at and
> > > ram space rocks] with make -j25 bzImage and some swapping [sucks when
> > > you hit heavy swap of course, but quite playable as long as swap is
> > > light])
> >
> > cool! Could you please also play a bit with various WAKER_BONUS_RATIO
> > values? If the default '0' value is the best-performing variant then i'll
> > nuke it from the patch altogether.
>
> Hi,
>
> I've tested 2.5.64 + sched-2.5.64-B2 on an UP pIII 700 with 704MB ram,
> here's my results...
>
> General impression is that interactivity is good but it seems there's a
> few cases that gets worse, sometimes a lot worse.
>
> First the things that got better...
>
> Xmms doesn't skip when switching tabs in galeon anymore during general
> desktopusage (more on that below)
> Earlier I had to change some schedulersettings to get it to stop
> skipping.
>
> And it was possible to watch a movie using mplayer while running a make
> -j10 on a kerneltree, the playback wasn't 100% perfect but I'd say it
> was ~90% , the jerkyness I saw wasn't very irritating. I tried the
> window wiggle-test while the compile and movie was playing and it was
> smooth and only added a little more jerkyness to the playback.
>
>
> Some negative things:
>
> Xmms could skip during the compile/movie playback, most notably one
> rather long skip right after a song-change but sometimes in the middle
> of a song as well, those skips were quite long as well, 2-3 seconds.
>
> Another xmms related problem is that even without any load on the
> machine the mousepointer gets jerky for 0.1-0.5 seconds right after
> having used the mouse to search forward or backwards in a song. It gets
> jerky at the moment you release the mousebutton.
> Basically no change in the jerkyness even with a make -j10 and movie
> playing.
>
> The biggest problem is the fact that my windowmanager gets unresponsive
> for ~30 seconds when I perform something that works fine with vanilla
> 2.4 and 2.5. I have two keybindings that execute 'aumix -v+3' and
> 'aumix -v-3'. And if I hold one of those buttons down for say 2 seconds
> then at first it seems like nothing is happening but then the volume
> starts to inc-/decrease sloowly and the windowmanager (sawfish) is
> unresponsive for ~30 seconds. With 2.4 or vanilla 2.5 the change in
> volume is instant and I've never noticed that sawfish got unresponsive.
> Sure holding the button down for 2 seconds at a repeatrate of 30/s will
> cause a lot of aumix's to be executed but 30 seconds for 60 aumix's is
> quite a long time. A few times it has been enough to just press one of
> those keys a few times to get the exact same effect. It's as if sawfish
> goes into a weird loop. All this might be a bug in sawfish...
> This behaviour is the same with or without the compile/movie running.
>
> Another problem is that the composer-window in evolution gets
> unresponsive for 1-2 seconds if you hold down a key, and when it gets
> responsive again it gets all the keypresses...
>
> I've tried changing WAKER_BONUS_RATIO but I didn't see any diffrence.
> At least it didn't help with the Xmms mouse-jerkyness or the
> sawfish_goes_out_for_coffee_and_a_cigarette problem.
>
> There, that was my impressions of 2 hours usage of the patch.
>
> --
> /Martin
>
> Never argue with an idiot. They drag you down to their level, then beat you with experience.
> -
> To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
> the body of a message to majordomo@vger.kernel.org
> More majordomo info at  http://vger.kernel.org/majordomo-info.html
> Please read the FAQ at  http://www.tux.org/lkml/
>

Re: [patch] "interactivity changes", sched-2.5.64-B2

[Martin Josefsson]

On Fri, 2003-03-07 at 21:39, Ingo Molnar wrote:
> On 7 Mar 2003, Martin Josefsson wrote:
> 
> > Some negative things:
> 
> if you have time/interest, could you re-test the negative things with X
> reniced to -10, to further isolate the problem? Another thing to try is to
> renice xmms to -10 (or RT priority). Which one makes the larger
> difference?

X niced to -10 makes the xmms mouse-jerkyness disappear. but it can
still skip right after a songchange. If I renice xmms to -10 as well I
get the mouse-jerkyness again, it I renice it to -5 it seems to behave
quite well, no mouse-jerkyness and so far I havn't heard any skips, but
it can take some time until a skip occurs.

With X reniced to -10 it feels like the wiggle_a_window while a
'make -j10' and a movie is playing gets a bit more sluggish.

And sawfish still takes 30 second naps when executing a bunch of
aumix's.

-- 
/Martin

Never argue with an idiot. They drag you down to their level, then beat you with experience.

Re: [patch] "interactivity changes", sched-2.5.64-B2

[Martin Josefsson]

On Fri, 2003-03-07 at 22:03, David Lang wrote:
> sounds like the code in X that detects a key being held down is running
> into problems. any chance it's doing a busy loop or something silly like
> that that's just not running enough? (probably not, but since you have
> problems in a couple applications that happen when you hold a key down I
> would look there rather then at the scheduling code itself)

Wouldn't surprise me if it's an X problem... I can't say that I feel
like going digging into X sources...

I can get the same problem in sawfish if I press the key a few times
quite rapidly as well, without any background load at all. This problem
has never occured before on 2.4 or 2.5, with or without load. It could
be that the scheduler changes exposes some bug in X but I'm not really
sure how to start investigating...

-- 
/Martin

Never argue with an idiot. They drag you down to their level, then beat you with experience.

Re: [patch] "interactivity changes", sched-2.5.64-B2

[Martin Josefsson]

On Fri, 2003-03-07 at 22:05, Martin Josefsson wrote:

> And sawfish still takes 30 second naps when executing a bunch of
> aumix's.

For information: this is 2.5.64 + sched-2.5.64-B2, no preempt or other
patches.

I changed the command that sawfish executes to only append the current
date to a file and then I made it take a nap again.

I got a file containing 57 lines, one for each time the command was
executed by sawfish.

it took 41 seconds to execute the command 57 times and the machine was
very idle during that time.

According to top sawfish's priority is between 15 and 20 the whole time.
And as I said the machine is very idle during the test.

The command was a very small bashscript:

#!/bin/bash

date >> /tmp/result


So no aumix executed in the script, didn't want another error factor in
the test.

I've reproduced the problem three times and this is what I got:

executions   time
----------   ----
    57        41 seconds
    52        39 seconds
    28        25 seconds


results from the third run

Fri Mar  7 22:31:00 CET 2003
Fri Mar  7 22:31:02 CET 2003
Fri Mar  7 22:31:03 CET 2003
Fri Mar  7 22:31:04 CET 2003
Fri Mar  7 22:31:05 CET 2003
Fri Mar  7 22:31:06 CET 2003
Fri Mar  7 22:31:07 CET 2003
Fri Mar  7 22:31:07 CET 2003
Fri Mar  7 22:31:07 CET 2003
Fri Mar  7 22:31:09 CET 2003
Fri Mar  7 22:31:10 CET 2003
Fri Mar  7 22:31:11 CET 2003
Fri Mar  7 22:31:12 CET 2003
Fri Mar  7 22:31:14 CET 2003
Fri Mar  7 22:31:14 CET 2003
Fri Mar  7 22:31:14 CET 2003
Fri Mar  7 22:31:15 CET 2003
Fri Mar  7 22:31:16 CET 2003
Fri Mar  7 22:31:17 CET 2003
Fri Mar  7 22:31:18 CET 2003
Fri Mar  7 22:31:19 CET 2003
Fri Mar  7 22:31:19 CET 2003
Fri Mar  7 22:31:19 CET 2003
Fri Mar  7 22:31:20 CET 2003
Fri Mar  7 22:31:22 CET 2003
Fri Mar  7 22:31:23 CET 2003
Fri Mar  7 22:31:24 CET 2003
Fri Mar  7 22:31:25 CET 2003

Do you have any clues to what might be happening?

-- 
/Martin

Never argue with an idiot. They drag you down to their level, then beat you with experience.

Re: [patch] "interactivity changes", sched-2.5.64-B2

[Helge Hafting]

Martin Josefsson wrote:

> Wouldn't surprise me if it's an X problem... I can't say that I feel
> like going digging into X sources...
> 
> I can get the same problem in sawfish if I press the key a few times
> quite rapidly as well, without any background load at all. This problem
> has never occured before on 2.4 or 2.5, with or without load. It could
> be that the scheduler changes exposes some bug in X but I'm not really
> sure how to start investigating...

You could perhaps rule out sawfish bugs by testing this with
another window manager?

Helge Hafting