From a91755ae5e6333bd3f690820f086645cd122ec7f Mon Sep 17 00:00:00 2001
From: David du Colombier <0intro@gmail.com>
Date: Fri, 28 Feb 2003 00:00:00 +0000
Subject: [PATCH] Plan 9 from Bell Labs 2003-02-28

---
 pc/trap.c        |  21 +----
 port/devproc.c   |  29 ++-----
 port/portclock.c |  24 +-----
 port/portdat.h   |   5 +-
 port/portfns.h   |   3 +
 port/proc.c      | 215 +++++++++++++++++++++++++++--------------------
 6 files changed, 145 insertions(+), 152 deletions(-)

diff --git a/pc/trap.c b/pc/trap.c
index d8645ac90793f9ec560045ac683d6e9ad08120d6..0b0ef1aaae52be0533a2de0eb6b6ba8214a84993 100644
--- a/pc/trap.c
+++ b/pc/trap.c
@@ -310,26 +310,11 @@ trap(Ureg* ureg)
 		if(ctl->eoi)
 			ctl->eoi(vno);
 
-		/* 
-		 *  preemptive scheduling.  to limit stack depth,
-		 *  make sure process has a chance to return from
-		 *  the current interrupt before being preempted a
-		 *  second time.
-		 */
-		if(ctl->isintr && ctl->irq != IrqTIMER && ctl->irq != IrqCLOCK)
-		if(up && up->state == Running)
-		if(anyhigher())
-		if(up->preempted == 0)
-		if(!active.exiting){
-			up->preempted = 1;
+		if(ctl->isintr){
 			intrtime(m, vno);
-			sched();
-			splhi();
-			up->preempted = 0;
-			return;
+			if(up && ctl->irq != IrqTIMER && ctl->irq != IrqCLOCK)
+				preempted();
 		}
-		if(ctl->isintr)
-			intrtime(m, vno);
 	}
 	else if(vno <= nelem(excname) && user){
 		spllo();
diff --git a/port/devproc.c b/port/devproc.c
index 124d3be2be78388c8ca1ade7700020a577fe13a0..7db88b29ad0c6f8057f68955f6cbac5b4aca8167 100644
--- a/port/devproc.c
+++ b/port/devproc.c
@@ -1094,7 +1094,7 @@ void
 procctlreq(Proc *p, char *va, int n)
 {
 	Segment *s;
-	int i, npc;
+	int npc, pri;
 	Cmdbuf *cb;
 	Cmdtab *ct;
 
@@ -1116,17 +1116,6 @@ procctlreq(Proc *p, char *va, int n)
 	case CMclosefiles:
 		procctlclosefiles(p, 1, 0);
 		break;
-	case CMfixedpri:
-		i = atoi(cb->f[1]);
-		if(i < 0)
-			i = 0;
-		if(i >= Nrq)
-			i = Nrq - 1;
-		if(i > p->basepri && !iseve())
-			error(Eperm);
-		p->basepri = i;
-		p->fixedpri = 1;
-		break;
 	case CMhang:
 		p->hang = 1;
 		break;
@@ -1152,15 +1141,15 @@ procctlreq(Proc *p, char *va, int n)
 		p->noswap = 1;
 		break;
 	case CMpri:
-		i = atoi(cb->f[1]);
-		if(i < 0)
-			i = 0;
-		if(i >= Nrq)
-			i = Nrq - 1;
-		if(i > p->basepri && !iseve())
+		pri = atoi(cb->f[1]);
+		if(pri > PriNormal && !iseve())
+			error(Eperm);
+		procpriority(p, pri, 0);
+		break;
+	case CMfixedpri:
+		if(!iseve())
 			error(Eperm);
-		p->basepri = i;
-		p->fixedpri = 0;
+		procpriority(p, atoi(cb->f[1]), 1);
 		break;
 	case CMprivate:
 		p->privatemem = 1;
diff --git a/port/portclock.c b/port/portclock.c
index 9da9eeda5c3c9f56bd07d0bb318f2cdf617dc4a7..8ab90955cbf219020adcc3aeec83ebd5d412ec17 100644
--- a/port/portclock.c
+++ b/port/portclock.c
@@ -95,8 +95,6 @@ timerdel(Timer *dt)
 void
 hzclock(Ureg *ur)
 {
-	int callsched;
-
 	m->ticks++;
 	if(m->proc)
 		m->proc->pc = ur->pc;
@@ -126,29 +124,13 @@ hzclock(Ureg *ur)
 	if(up == 0 || up->state != Running)
 		return;
 
-	if(up->fixedpri){
-		/*  fixed priority processes are only preempted by
-		 *  higher piority processes.
-		 */
-		if(anyhigher())
-			callsched = 1;
-	} else {
-		/*  floating priority processes are are preempted
-		 *  by all sorts of things.
-		 */
-		if(anyready() && !edf->isedf(up))
-			callsched = 1;
-	}
-	if(callsched){
-		sched();
-		splhi();
-	}
-
 	/* user profiling clock */
-	if(userureg(ur)) {
+	if(userureg(ur)){
 		(*(ulong*)(USTKTOP-BY2WD)) += TK2MS(1);
 		segclock(ur->pc);
 	}
+
+	hzsched();	/* in proc.c */
 }
 
 void
diff --git a/port/portdat.h b/port/portdat.h
index e4440ed61cf2e33890899c43a8db7e74faf1c3fe..3d0fba91213d389c54b204e809163fca241c6380 100644
--- a/port/portdat.h
+++ b/port/portdat.h
@@ -657,10 +657,9 @@ struct Proc
 	ulong	priority;	/* priority level */
 	ulong	basepri;	/* base priority level */
 	uchar	fixedpri;	/* priority level deson't change */
-	ulong	rt;		/* # ticks used since last blocked */
-	ulong	art;		/* avg # ticks used since last blocked */
+	int	quanta;		/* quanta left */
+	ulong	readytime;	/* time process came ready */
 	ulong	movetime;	/* last time process switched processors */
-	ulong	readytime;	/* time process went ready */
 	int	preempted;	/* true if this process hasn't finished the interrupt
 				 *  that last preempted it
 				 */
diff --git a/port/portfns.h b/port/portfns.h
index e141c8aa0ddbce1ce9915121529afd35eaeee4be..1d4cc1ccfe9b2b46aec268ce6520a0191f75917b 100644
--- a/port/portfns.h
+++ b/port/portfns.h
@@ -119,6 +119,7 @@ char*		getconfenv(void);
 int		haswaitq(void*);
 long		hostdomainwrite(char*, int);
 long		hostownerwrite(char*, int);
+void		hzsched(void);
 void		iallocinit(void);
 Block*		iallocb(int);
 void		iallocsummary(void);
@@ -204,6 +205,7 @@ void		panic(char*, ...);
 Cmdbuf*		parsecmd(char *a, int n);
 ulong		perfticks(void);
 void		pexit(char*, int);
+int		preempted(void);
 void		printinit(void);
 int		procindex(ulong);
 void		pgrpcpy(Pgrp*, Pgrp*);
@@ -222,6 +224,7 @@ void		procdump(void);
 int		procfdprint(Chan*, int, int, char*, int);
 void		procinit0(void);
 void		procflushseg(Segment*);
+void		procpriority(Proc*, int, int);
 Proc*		proctab(int);
 void		procwired(Proc*, int);
 Pte*		ptealloc(void);
diff --git a/port/proc.c b/port/proc.c
index 27e3322a5b776311f66119c488e0685957375393..4a9646dffaa9df00f2f98f4910bc18d6435eaf3c 100644
--- a/port/proc.c
+++ b/port/proc.c
@@ -20,7 +20,22 @@ static struct Procalloc
 	Proc*	free;
 } procalloc;
 
+enum
+{
+	Q=(HZ/20)*4,
+	DQ=((HZ-Q)/40)*2,
+};
+
 static Schedq	runq[Nrq];
+static ulong	runvec;
+static int	quanta[Nrq] =
+{
+	Q+19*DQ, Q+18*DQ, Q+17*DQ, Q+16*DQ,
+	Q+15*DQ, Q+14*DQ, Q+13*DQ, Q+12*DQ,
+	Q+11*DQ, Q+10*DQ, Q+ 9*DQ, Q+ 8*DQ,
+	Q+ 7*DQ, Q+ 6*DQ, Q+ 5*DQ, Q+ 4*DQ,
+	Q+ 3*DQ, Q+ 2*DQ, Q+ 1*DQ, Q+ 0*DQ,
+};
 
 extern Edfinterface	nulledf;
 
@@ -128,28 +143,54 @@ sched(void)
 int
 anyready(void)
 {
-	return nrdy || edf->edfanyready();
+	return runvec || edf->edfanyready();
 }
 
 int
 anyhigher(void)
 {
-	Schedq *rq;
+	return runvec & ~((1<<(up->priority+1))-1);
+}
 
-	if(nrdy == 0)
-		return 0;
+/*
+ *  here once per clock tick to see if we should resched
+ */
+void
+hzsched(void)
+{
+	/* another cycle, another quantum */
+	up->quanta--;
 
-	for(rq = &runq[Nrq-1]; rq > &runq[up->priority]; rq--)
-		if(rq->head != nil)
-			return 1;
+	/* edf scheduler always gets first chance */
+	if(edf->isedf(up))
+		return;
 
-	return 0;
+	/* don't bother unless someone is elegible */
+	if(anyhigher() || (!up->fixedpri && anyready())){
+		sched();
+		splhi();
+	}
 }
 
-enum
+/*
+ *  here at the end of non-clock interrupts to see if we should preempt the
+ *  current process.  Returns 1 if preempted, 0 otherwise.
+ */
+int
+preempted(void)
 {
-	Squantum = 1,
-};
+	if(up && up->state == Running)
+	if(up->preempted == 0)
+	if(anyhigher())
+	if(!active.exiting){
+		up->preempted = 1;
+		sched();
+		splhi();
+		up->preempted = 0;
+		return 1;
+	}
+	return 0;
+}
 
 void
 ready(Proc *p)
@@ -164,34 +205,37 @@ ready(Proc *p)
 		splx(s);
 		return;
 	}
+
+	pri = p->priority;
+
 	if(p->fixedpri){
 		pri = p->basepri;
+		p->quanta = HZ;
+	} else if(p->state == Running){
+		if(p->quanta <= 0){
+			/* degrade priority of anyone that used their whole quanta */
+
+			/* processes > PriNormal and those below don't mix */
+			if(p->basepri > PriNormal){
+				if(pri > PriNormal)
+					pri--;
+			} else {
+				if(pri > 0)
+					pri--;
+			}
+			p->quanta = quanta[pri];
+		}
 	} else {
-		/* history counts */
-		if(p->state == Running){
-			p->rt++;
-			pri = (p->art + p->rt)/2;
-		} else {
-			p->art = (p->art + p->rt + 2)/2;
-			pri = (p->art + p->rt)/2;
-			p->rt = 0;
+		if(p->quanta > quanta[pri]/2){
+			/* blocked before using half its quanta, go up */
+			if(++pri > p->basepri)
+				pri = p->basepri;
 		}
-		pri = p->basepri - (pri/Squantum);
-		if(pri < 0)
-			pri = 0;
-
-		/* the only intersection between the classes is at PriNormal */
-		if(pri < PriNormal && p->basepri > PriNormal)
-			pri = PriNormal;
-
-		/* stick at low priority any process waiting for a lock */
-		if(p->lockwait)
-			pri = PriLock;
+		p->quanta = quanta[pri];
 	}
 
+	rq = &runq[pri];
 	p->priority = pri;
-	rq = &runq[p->priority];
-
 	lock(runq);
 	p->rnext = 0;
 	if(rq->tail)
@@ -201,6 +245,7 @@ ready(Proc *p)
 	rq->tail = p;
 	rq->n++;
 	nrdy++;
+	runvec |= 1<<pri;
 	p->readytime = m->ticks;
 	p->state = Ready;
 	unlock(runq);
@@ -210,10 +255,10 @@ ready(Proc *p)
 Proc*
 runproc(void)
 {
-	Schedq *rq, *xrq;
+	Schedq *rq;
 	Proc *p, *l;
-	ulong rt;
 	ulong start, now;
+	int i;
 
 	start = perfticks();
 
@@ -221,55 +266,30 @@ runproc(void)
 		return p;
 
 loop:
-
 	/*
 	 *  find a process that last ran on this processor (affinity),
-	 *  or one that hasn't moved in a while (load balancing).
+	 *  or one that hasn't moved in a while (load balancing).  Every
+	 *  time around the loop affinity goes down.
 	 */
 	spllo();
-	for(;;){
-		if((++(m->fairness) & 0x3) == 0){
-			/*
-			 *  once in a while, run process that's been waiting longest
-			 *  regardless of movetime
-			 */
-			rt = 0xffffffff;
-			xrq = nil;
-			for(rq = runq; rq < &runq[Nrq]; rq++){
-				p = rq->head;
-				if(p == 0)
-					continue;
-				if(p->readytime < rt){
-					xrq = rq;
-					rt = p->readytime;
-				}
-			}
-			if(xrq != nil){
-				rq = xrq;
-				p = rq->head;
-				if(p != nil && p->wired == nil)
-					p->movetime = 0;
-				goto found;
-			}
-		} else {
-			/*
-			 *  get highest priority process that this
-			 *  processor can run given affinity constraints
-			 */
-			for(rq = &runq[Nrq-1]; rq >= runq; rq--){
-				p = rq->head;
-				if(p == 0)
-					continue;
-				for(; p; p = p->rnext){
-					if(p->mp == MACHP(m->machno)
-					|| p->movetime < MACHP(0)->ticks)
-						goto found;
-				}
+	for(i = 0;; i++){
+		/*
+		 *  get highest priority process that this
+		 *  processor can run given affinity constraints
+		 */
+		for(rq = &runq[Nrq-1]; rq >= runq; rq--){
+			p = rq->head;
+			if(p == 0)
+				continue;
+			for(; p; p = p->rnext){
+				if(p->mp == MACHP(m->machno) || (!p->wired && i > 0))
+					goto found;
 			}
 		}
 
-		/* remember how much time we're here */
 		idlehands();
+
+		/* remember how much time we're here */
 		now = perfticks();
 		m->perf.inidle += now-start;
 		start = now;
@@ -282,7 +302,7 @@ found:
 
 	l = 0;
 	for(p = rq->head; p; p = p->rnext){
-		if(p->mp == MACHP(m->machno) || p->movetime <= MACHP(0)->ticks)
+		if(p->mp == MACHP(m->machno) || (!p->wired && i > 0))
 			break;
 		l = p;
 	}
@@ -300,6 +320,8 @@ found:
 		l->rnext = p->rnext;
 	else
 		rq->head = p->rnext;
+	if(rq->head == nil)
+		runvec &= ~(1<<(rq-runq));
 	rq->n--;
 	nrdy--;
 	if(p->state != Ready)
@@ -307,8 +329,6 @@ found:
 	unlock(runq);
 
 	p->state = Scheding;
-	if(p->mp != MACHP(m->machno))
-		p->movetime = MACHP(0)->ticks + HZ/10;
 	p->mp = MACHP(m->machno);
 
 	return p;
@@ -366,20 +386,15 @@ newproc(void)
 	p->kp = 0;
 	p->procctl = 0;
 	p->notepending = 0;
-	p->mp = 0;
-	p->movetime = 0;
-	p->wired = 0;
 	p->ureg = 0;
 	p->privatemem = 0;
 	p->noswap = 0;
-	p->lockwait = nil;
 	p->errstr = p->errbuf0;
 	p->syserrstr = p->errbuf1;
 	p->errbuf0[0] = '\0';
 	p->errbuf1[0] = '\0';
 	p->nlocks = 0;
 	p->delaysched = 0;
-	p->movetime = 0;
 	kstrdup(&p->user, "*nouser");
 	kstrdup(&p->text, "*notext");
 	kstrdup(&p->args, "");
@@ -394,6 +409,11 @@ newproc(void)
 	if(p->kstack == 0)
 		p->kstack = smalloc(KSTACK);
 
+	/* sched params */
+	p->mp = 0;
+	p->wired = 0;
+	procpriority(p, PriNormal, 0);
+
 	p->task = nil;
 
 	return p;
@@ -430,10 +450,27 @@ procwired(Proc *p, int bm)
 	}
 
 	p->wired = MACHP(bm);
-	p->movetime = 0xffffffff;
 	p->mp = p->wired;
 }
 
+void
+procpriority(Proc *p, int pri, int fixed)
+{
+	if(pri >= Nrq)
+		pri = Nrq - 1;
+	else if(pri < 0)
+		pri = 0;
+	p->basepri = pri;
+	p->priority = pri;
+	if(fixed){
+		p->quanta = 0xfffff;
+		p->fixedpri = 1;
+	} else {
+		p->quanta = quanta[pri];
+		p->fixedpri = 0;
+	}
+}
+
 void
 procinit0(void)		/* bad planning - clashes with devproc.c */
 {
@@ -967,9 +1004,9 @@ dumpaproc(Proc *p)
 	s = p->psstate;
 	if(s == 0)
 		s = statename[p->state];
-	print("%3lud:%10s pc %8lux dbgpc %8lux  %8s (%s) ut %ld st %ld bss %lux qpc %lux nl %lud nd %lud lpc %lux\n",
+	print("%3lud:%10s pc %8lux dbgpc %8lux  %8s (%s) ut %ld st %ld bss %lux qpc %lux nl %lud nd %lud lpc %lux pri %lud\n",
 		p->pid, p->text, p->pc, dbgpc(p),  s, statename[p->state],
-		p->time[0], p->time[1], bss, p->qpc, p->nlocks, p->delaysched, p->lastlock ? p->lastlock->pc : 0);
+		p->time[0], p->time[1], bss, p->qpc, p->nlocks, p->delaysched, p->lastlock ? p->lastlock->pc : 0, p->priority);
 }
 
 void
@@ -1047,8 +1084,7 @@ scheddump(void)
 			continue;
 		print("rq%ld:", rq-runq);
 		for(p = rq->head; p; p = p->rnext)
-			print(" %lud(%lud, %lud)", p->pid, m->ticks - p->readytime,
-				MACHP(0)->ticks - p->movetime);
+			print(" %lud(%lud)", p->pid, m->ticks - p->readytime);
 		print("\n");
 		delay(150);
 	}
@@ -1083,8 +1119,7 @@ kproc(char *name, void (*func)(void *), void *arg)
 	p->ureg = 0;
 	p->dbgreg = 0;
 
-	p->basepri = PriKproc;
-	p->priority = p->basepri;
+	procpriority(p, PriKproc, 0);
 
 	kprocchild(p, func, arg);