- short-running processes really should be short-running, like 1s not 30s.
- when there are more runnable processes than processors, be migrate
processes even when there is an equally good runnable without affinity.
this prevents vast unfairness.
- use monmwit to take it easy on cachelines.

performance improvements are pretty impressive.  on my ivy bridge machine,
kernel builds have gone from 5.4s to 3.9s.

Reference: /n/atom/patch/applied/fairsched
Date: Sat May 24 07:04:04 CES 2014
Signed-off-by: quanstro@quanstro.net

--- /sys/src/nix/port/proc.c	Sat May 24 07:03:26 2014
+++ /sys/src/nix/port/proc.c	Sat May 24 07:03:28 2014
@@ -11,13 +11,14 @@
 
 enum
 {
-	Scaling=2,
+	Scaling		= 2,
+	Shortburst	= HZ,		/* 1 second */
 
 	/*
 	 * number of schedulers used.
 	 * 1 uses just one, which is the behavior of Plan 9.
 	 */
-	Nsched = 16,
+	Nsched		= 16,
 };
 
 Ref	noteidalloc;
@@ -90,19 +91,6 @@
 }
 
 /*
- * bad planning, once more.
- */
-void
-procinit0(void)
-{
-	int i;
-
-	for(i = 0; i < Nsched; i++)
-		run[i].schedgain = 30;
-
-}
-
-/*
  * Always splhi()'ed.
  */
 void
@@ -231,7 +219,7 @@
 	}
 	if(p != m->readied)
 		m->schedticks = m->ticks + HZ/10;
-	m->readied = 0;
+	m->readied = nil;
 	up = p;
 	up->nqtrap = 0;
 	up->nqsyscall = 0;
@@ -268,7 +256,7 @@
 
 	/* unless preempted, get to run for at least 100ms */
 	if(anyhigher()
-	|| (!m->proc->fixedpri && (long)(m->ticks - m->schedticks) > 0 && anyready())){
+	|| (!m->proc->fixedpri && tickscmp(m->ticks, m->schedticks) > 0 && anyready())){
 		m->readied = nil;	/* avoid cooperative scheduling */
 		m->proc->delaysched++;
 	}
@@ -352,7 +340,7 @@
 		return;
 	if(m->sch == nil)	/* may happen during boot */
 		return;
-	D = m->sch->schedgain*HZ*Scaling;
+	D = Shortburst*Scaling;
 	if(n > D)
 		n = D;
 
@@ -365,7 +353,7 @@
 		p->cpu = 1000 - t;
 	}
 
-//iprint("pid %d %s for %d cpu %d -> %d\n", p->pid,p==up?"active":"inactive",n, ocpu,p->cpu);
+//iprint("pid %d %s for %d cpu %d -> %d\n", p->pid, p==up? "active": "inactive", n, ocpu, p->cpu);
 }
 
 /*
@@ -494,6 +482,7 @@
 	pri = reprioritize(p);
 	p->priority = pri;
 	rq = &sch->runq[pri];
+	p->readytime = fastticks(nil);
 	p->state = Ready;
 	queueproc(sch, rq, p);
 	if(p->trace)
@@ -558,7 +547,7 @@
 		if(npri != pri){
 			pl = splhi();
 			p = dequeueproc(sch, rq, p);
-			if(p)
+			if(p != nil)
 				queueproc(sch, &sch->runq[npri], p);
 			splx(pl);
 			goto another;
@@ -624,6 +613,16 @@
 	return nil;
 }
 
+enum {
+	/*
+	 * if there are more runnable processes than mach, and nmach>1,
+	 * and nproc%nmach != 0, we need to migrate processes around to
+	 * avoid serious unfairness.  this is architecture sensitive.
+	 */
+//	Migratedelay	= 250*1000,	/* 250 µs */
+	Migratedelay	= 10*1000*1000,	/* 10 ms */
+};
+
 /*
  *  pick a process to run
  */
@@ -660,12 +659,11 @@
 		/*
 		 *  find the highest priority target process that this
 		 *  processor can run given affinity constraints.
-		 *
 		 */
 		for(rq = &sch->runq[Nrq-1]; rq >= sch->runq; rq--){
-			for(p = rq->head; p; p = p->rnext){
+			for(p = rq->head; p != nil; p = p->rnext){
 				if(p->mp == nil || p->mp == m
-				|| (p->wired == nil && i > 0))
+				|| p->wired == nil && (i>0 || fastticks2ns(fastticks(nil) - p->readytime) >= Migratedelay))
 					goto found;
 			}
 		}
@@ -675,8 +673,10 @@
 		if(p != nil)
 			goto stolen;
 		spllo();
-		/* waste time or halt the CPU */
-		idlehands();
+
+		while(sch->runvec == 0)
+			monmwait((int*)&sch->runvec, 0);
+
 		/* remember how much time we're here */
 		now = perfticks();
 		m->perf.inidle += now-start;
@@ -1452,16 +1452,17 @@
 	Schedq *rq;
 
 	for(sch = run; sch < &run[Nsched]; sch++){
+		if(sch->nmach == 0)
+			continue;
 		for(rq = &sch->runq[Nrq-1]; rq >= sch->runq; rq--){
 			if(rq->head == nil)
 				continue;
-			print("sch%ld rq%ld:", sch - run, rq-sch->runq);
+			iprint("sch%ld rq%ld:", sch - run, rq-sch->runq);
 			for(p = rq->head; p; p = p->rnext)
-				print(" %d(%lud)", p->pid, m->ticks - p->readytime);
-			print("\n");
-			delay(150);
+				iprint(" %d(%lludµs)", p->pid, fastticks2us(fastticks(nil) - p->readytime));
+			iprint("\n");
 		}
-		print("sch%ld: nrdy %d\n", sch - run, sch->nrdy);
+		iprint("sch%ld: nrdy %d\n", sch - run, sch->nrdy);
 	}
 }
 
@@ -1671,7 +1672,7 @@
 
 	sch = m->sch;
 	p = m->proc;
-	if(p) {
+	if(p != nil){
 		sch->nrun++;
 		p->time[p->insyscall]++;
 	}
--- /sys/src/nix/port/portdat.h	Sat May 24 07:03:30 2014
+++ /sys/src/nix/port/portdat.h	Sat May 24 07:03:32 2014
@@ -703,9 +703,8 @@
 	Lock;			/* runq */
 	int	nrdy;
 	uint	delayedscheds;	/* statistics */
-	int	skipscheds;
-	int	preempts;
-	int	schedgain;
+	uint	skipscheds;
+	uint	preempts;
 	uint	balancetime;
 	Schedq	runq[Nrq];
 	uint	runvec;
@@ -835,7 +834,7 @@
 	int	fixedpri;	/* priority level does not change */
 	uint	cpu;		/* cpu average */
 	ulong	lastupdate;
-	ulong	readytime;	/* time process came ready */
+	uvlong	readytime;	/* time process came ready, in fastticks*/
 	ulong	movetime;	/* last time process switched processors */
 	int	preempted;	/* true if this process hasn't finished the interrupt
 				 *  that last preempted it
@@ -1083,7 +1082,6 @@
 extern	char*	statename[];
 extern	Lockstats lockstats;
 extern	QLockstats qlockstats;
-extern	Waitstats waitstats;
 extern	struct {
 	char*	n;
 	void (*f)(Ar0*, va_list);
--- /sys/src/nix/port/sysproc.c	Sat May 24 07:03:34 2014
+++ /sys/src/nix/port/sysproc.c	Sat May 24 07:03:36 2014
@@ -847,7 +847,7 @@
 			val = p->rendval;
 			p->rendval = PTR2UINT(va_arg(list, void*));
 
-			while(p->mach != 0)
+			while(!procsaved(p))
 				pause();
 			ready(p);
 			unlock(up->rgrp);
--- /sys/src/nix/port/edf.c	Sat May 24 07:03:38 2014
+++ /sys/src/nix/port/edf.c	Sat May 24 07:03:39 2014
@@ -575,7 +575,7 @@
 	sch->nrdy++;
 	sch->runvec |= 1 << PriEdf;
 	p->priority = PriEdf;
-	p->readytime = m->ticks;
+	p->readytime = fastticks(nil);
 	p->state = Ready;
 	unlock(sch);
 	if(p->trace)
--- /sys/src/nix/port/portfns.h	Sat May 24 07:03:41 2014
+++ /sys/src/nix/port/portfns.h	Sat May 24 07:03:42 2014
@@ -248,7 +248,6 @@
 void		procdump(void);
 int		procfdprint(Chan*, int, int, char*, int);
 void		procflushseg(Segment*);
-void		procinit0(void);
 void		procpriority(Proc*, int, int);
 void		procrestore(Proc*);
 void		procsave(Proc*);
--- /sys/src/nix/k10/main.c	Sat May 24 07:03:43 2014
+++ /sys/src/nix/k10/main.c	Sat May 24 07:03:44 2014
@@ -166,7 +166,6 @@
 	trapinit();
 	printinit();
 
-	procinit0();
 	lapiconline();
 	ioapiconline();
 	sipi();
--- /sys/src/nix/port/edf.c	Sat May 24 07:03:46 2014
+++ /sys/src/nix/port/edf.c	Sat May 24 07:03:47 2014
@@ -575,7 +575,7 @@
 	sch->nrdy++;
 	sch->runvec |= 1 << PriEdf;
 	p->priority = PriEdf;
-	p->readytime = m->ticks;
+	p->readytime = fastticks(nil);
 	p->state = Ready;
 	unlock(sch);
 	if(p->trace)
--- /sys/src/nix/port/portfns.h	Sat May 24 07:03:49 2014
+++ /sys/src/nix/port/portfns.h	Sat May 24 07:03:50 2014
@@ -248,7 +248,6 @@
 void		procdump(void);
 int		procfdprint(Chan*, int, int, char*, int);
 void		procflushseg(Segment*);
-void		procinit0(void);
 void		procpriority(Proc*, int, int);
 void		procrestore(Proc*);
 void		procsave(Proc*);