9/port: allow kiloprocs -- allocate procs lazily

Treallocate the small data structures around procs eagerly,
but use malloc to allocate the large proc data structures
when we need them, which allows us to scale to many more procs.

There are still many scalability bottlenecks, so we only crank
up the nproc limit by a little bit this time around, and crank
it up more as we optimize more.

15 files changed, 80 insertions, 75 deletions

diff --git a/sys/src/9/bcm/main.c b/sys/src/9/bcm/main.c
index d6c4acaa6..13754d7d8 100644
--- a/sys/src/9/bcm/main.c
+++ b/sys/src/9/bcm/main.c
@@ -222,8 +222,8 @@ confinit(void)
 	conf.nproc = 100 + ((conf.npage*BY2PG)/MB)*5;
 	if(cpuserver)
 		conf.nproc *= 3;
-	if(conf.nproc > 2000)
-		conf.nproc = 2000;
+	if(conf.nproc > 4000)
+		conf.nproc = 4000;
 	conf.nswap = conf.npage*3;
 	conf.nswppo = 4096;
 	conf.nimage = 200;
diff --git a/sys/src/9/bcm64/main.c b/sys/src/9/bcm64/main.c
index 6679d75d7..4d818408a 100644
--- a/sys/src/9/bcm64/main.c
+++ b/sys/src/9/bcm64/main.c
@@ -93,8 +93,8 @@ confinit(void)
 	conf.nproc = 100 + ((conf.npage*BY2PG)/MB)*5;
 	if(cpuserver)
 		conf.nproc *= 3;
-	if(conf.nproc > 2000)
-		conf.nproc = 2000;
+	if(conf.nproc > 4000)
+		conf.nproc = 4000;
 	conf.nswap = conf.npage*3;
 	conf.nswppo = 4096;
 	conf.nimage = 200;
diff --git a/sys/src/9/imx8/main.c b/sys/src/9/imx8/main.c
index 6196e3aa8..7180267dd 100644
--- a/sys/src/9/imx8/main.c
+++ b/sys/src/9/imx8/main.c
@@ -191,8 +191,8 @@ confinit(void)
 	conf.nproc = 100 + ((conf.npage*BY2PG)/MB)*5;
 	if(cpuserver)
 		conf.nproc *= 3;
-	if(conf.nproc > 2000)
-		conf.nproc = 2000;
+	if(conf.nproc > 4000)
+		conf.nproc = 4000;
 	conf.nswap = conf.npage*3;
 	conf.nswppo = 4096;
 	conf.nimage = 200;
diff --git a/sys/src/9/kw/main.c b/sys/src/9/kw/main.c
index 060fe8b32..ac42b762f 100644
--- a/sys/src/9/kw/main.c
+++ b/sys/src/9/kw/main.c
@@ -433,8 +433,8 @@ confinit(void)
 	conf.nproc = 100 + ((conf.npage*BY2PG)/MB)*5;
 	if(cpuserver)
 		conf.nproc *= 3;
-	if(conf.nproc > 2000)
-		conf.nproc = 2000;
+	if(conf.nproc > 4000)
+		conf.nproc = 4000;
 	conf.nswap = conf.npage*3;
 	conf.nswppo = 4096;
 	conf.nimage = 200;
diff --git a/sys/src/9/mtx/main.c b/sys/src/9/mtx/main.c
index 71c31ce91..3255783e2 100644
--- a/sys/src/9/mtx/main.c
+++ b/sys/src/9/mtx/main.c
@@ -183,8 +183,8 @@ confinit(void)
 	conf.nproc = 100 + ((conf.npage*BY2PG)/MB)*5;
 	if(cpuserver)
 		conf.nproc *= 3;
-	if(conf.nproc > 2000)
-		conf.nproc = 2000;
+	if(conf.nproc > 4000)
+		conf.nproc = 4000;
 	conf.nimage = 200;
 	conf.nswap = conf.nproc*80;
 	conf.nswppo = 4096;
diff --git a/sys/src/9/omap/main.c b/sys/src/9/omap/main.c
index 839a6795e..32a83c0d7 100644
--- a/sys/src/9/omap/main.c
+++ b/sys/src/9/omap/main.c
@@ -452,8 +452,8 @@ confinit(void)
 	conf.nproc = 100 + ((conf.npage*BY2PG)/MB)*5;
 	if(cpuserver)
 		conf.nproc *= 3;
-	if(conf.nproc > 2000)
-		conf.nproc = 2000;
+	if(conf.nproc > 4000)
+		conf.nproc = 4000;
 	conf.nswap = conf.npage*3;
 	conf.nswppo = 4096;
 	conf.nimage = 200;
diff --git a/sys/src/9/pc/main.c b/sys/src/9/pc/main.c
index d59c200a1..73f8d5787 100644
--- a/sys/src/9/pc/main.c
+++ b/sys/src/9/pc/main.c
@@ -158,8 +158,8 @@ confinit(void)
 	conf.nproc = 100 + ((conf.npage*BY2PG)/MB)*5;
 	if(cpuserver)
 		conf.nproc *= 3;
-	if(conf.nproc > 2000)
-		conf.nproc = 2000;
+	if(conf.nproc > 4000)
+		conf.nproc = 4000;
 	conf.nimage = 200;
 	conf.nswap = conf.nproc*80;
 	conf.nswppo = 4096;
diff --git a/sys/src/9/pc64/main.c b/sys/src/9/pc64/main.c
index 78cd8ef4c..7da5d11fb 100644
--- a/sys/src/9/pc64/main.c
+++ b/sys/src/9/pc64/main.c
@@ -43,8 +43,8 @@ confinit(void)
 	conf.nproc = 100 + ((conf.npage*BY2PG)/MB)*5;
 	if(cpuserver)
 		conf.nproc *= 3;
-	if(conf.nproc > 2000)
-		conf.nproc = 2000;
+	if(conf.nproc > 4000)
+		conf.nproc = 4000;
 	conf.nimage = 200;
 	conf.nswap = conf.nproc*80;
 	conf.nswppo = 4096;
diff --git a/sys/src/9/port/devproc.c b/sys/src/9/port/devproc.c
index d19d4a9a0..503b00922 100644
--- a/sys/src/9/port/devproc.c
+++ b/sys/src/9/port/devproc.c
@@ -149,7 +149,8 @@ static char *sname[]={ "Text", "Data", "Bss", "Stack", "Shared", "Phys", "Fixed"
 #define	QSHIFT	5	/* location in qid of proc slot # */
 
 #define	QID(q)		((((ulong)(q).path)&0x0000001F)>>0)
-#define	SLOT(q)		(((((ulong)(q).path)&0x07FFFFFE0)>>QSHIFT)-1)
+#define	SLOTMAX		0x4000000
+#define	SLOT(q)		(((((ulong)(q).path)>>QSHIFT)&(SLOTMAX-1))-1)
 #define	PID(q)		((q).vers)
 #define	NOTEID(q)	((q).vers)
 
@@ -202,6 +203,8 @@ procgen(Chan *c, char *name, Dirtab *tab, int, int s, Dir *dp)
 			return -1;
 
 		p = proctab(s);
+		if(p == nil)
+			return 0;
 		pid = p->pid;
 		if(pid == 0)
 			return 0;
@@ -227,7 +230,7 @@ procgen(Chan *c, char *name, Dirtab *tab, int, int s, Dir *dp)
 		panic("procgen");
 
 	tab = &procdir[s];
-	path = c->qid.path&~(((1<<QSHIFT)-1));	/* slot component */
+	path = c->qid.path&~((1<<QSHIFT)-1);	/* slot component */
 
 	/* p->procmode determines default mode for files in /proc */
 	p = proctab(SLOT(c->qid));
@@ -281,8 +284,9 @@ _proctrace(Proc* p, Tevent etype, vlong ts)
 static void
 procinit(void)
 {
-	if(conf.nproc >= (1<<(16-QSHIFT))-1)
-		print("warning: too many procs for devproc\n");
+	/* slot masks: lets see how big we can go */
+	if(conf.nproc > SLOTMAX)
+		panic("warning: too many procs for devproc\n");
 }
 
 static Chan*
@@ -335,8 +339,7 @@ changenoteid(Proc *p, ulong noteid)
 		setnoteid(p, noteid);
 		return;
 	}
-	for(i = 0; i < conf.nproc; i++){
-		pp = proctab(i);
+	for(i = 0; (pp = proctab(i)) != nil; i++){
 		if(pp->noteid != noteid || pp->kp)
 			continue;
 		if(strcmp(pp->user, p->user) == 0){
@@ -354,8 +357,7 @@ postnotepg(ulong noteid, char *n, int flag)
 	Proc *p;
 	int i;
 
-	for(i = 0; i < conf.nproc; i++){
-		p = proctab(i);
+	for(i = 0; (p = proctab(i)) != nil; i++){
 		if(p == up)
 			continue;
 		if(p->noteid != noteid || p->kp)
@@ -414,7 +416,7 @@ procopen(Chan *c, int omode0)
 		nexterror();
 	}
 	pid = PID(c->qid);
-	if(p->pid != pid)
+	if(p == nil || p->pid != pid)
 		error(Eprocdied);
 
 	omode = openmode(omode0);
diff --git a/sys/src/9/port/devswap.c b/sys/src/9/port/devswap.c
index 261ecf7c1..a30560c90 100644
--- a/sys/src/9/port/devswap.c
+++ b/sys/src/9/port/devswap.c
@@ -162,7 +162,7 @@ pager(void*)
 	while(waserror())
 		;
 
-	x = -1;
+	x = 0;
 	for(;;){
 		up->psstate = "Reclaim";
 		if(reclaim()){
@@ -183,12 +183,13 @@ pager(void*)
 
 		i = ageclock;
 		do {
-			if(++x >= conf.nproc){
+			p = proctab(x++);
+			if(p == nil){
 				if(++ageclock == i)
 					goto Killbig;
 				x = 0;
+				continue;
 			}
-			p = proctab(x);
 		} while(p->state == Dead || p->noswap || !canqlock(&p->seglock));
 		up->psstate = "Pageout";
 		for(i = 0; i < NSEG; i++) {
@@ -274,8 +275,7 @@ canflush(Proc *p, Segment *s)
 	 * Now we must do hardwork to ensure all processes which have tlb
 	 * entries for this segment will be flushed if we succeed in paging it out
 	 */
-	for(x = 0; x < conf.nproc; x++){
-		p = proctab(x);
+	for(x = 0; (p = proctab(x)) != nil; x++){
 		if(p->state == Dead)
 			continue;
 		for(i = 0; i < NSEG; i++){
diff --git a/sys/src/9/port/edf.c b/sys/src/9/port/edf.c
index 99bbf1f61..213567023 100644
--- a/sys/src/9/port/edf.c
+++ b/sys/src/9/port/edf.c
@@ -373,17 +373,15 @@ edfadmit(Proc *p)
 		pt(p, SAdmit, 0);
 
 	/* Look for another proc with the same period to synchronize to */
-	SET(r);
-	for(i=0; i<conf.nproc; i++) {
-		r = proctab(i);
+	for(i=0; (r = proctab(i)) != nil; i++) {
 		if(r->state == Dead || r == p)
 			continue;
-		if (r->edf == nil || (r->edf->flags & Admitted) == 0)
+		if(r->edf == nil || (r->edf->flags & Admitted) == 0)
 			continue;
-		if (r->edf->T == e->T)
-				break;
+		if(r->edf->T == e->T)
+			break;
 	}
-	if (i == conf.nproc){
+	if(r == nil){
 		/* Can't synchronize to another proc, release now */
 		e->t = now;
 		e->d = 0;
@@ -627,11 +625,10 @@ testschedulability(Proc *theproc)
 	/* initialize */
 	DPRINT("schedulability test %lud\n", theproc->pid);
 	qschedulability = nil;
-	for(i=0; i<conf.nproc; i++) {
-		p = proctab(i);
+	for(i=0; (p = proctab(i)) != nil; i++) {
 		if(p->state == Dead)
 			continue;
-		if ((p->edf == nil || (p->edf->flags & Admitted) == 0) && p != theproc)
+		if((p->edf == nil || (p->edf->flags & Admitted) == 0) && p != theproc)
 			continue;
 		p->edf->testtype = Rl;
 		p->edf->testtime = 0;
diff --git a/sys/src/9/port/portdat.h b/sys/src/9/port/portdat.h
index e2aa2aa07..8f3904441 100644
--- a/sys/src/9/port/portdat.h
+++ b/sys/src/9/port/portdat.h
@@ -671,6 +671,7 @@ struct Proc
 	ulong	pid;
 	ulong	noteid;		/* Equivalent of note group */
 	ulong	parentpid;
+	ulong	index;
 
 	Proc	*parent;	/* Process to send wait record on exit */
 	Lock	exl;		/* Lock count and waitq */
@@ -760,7 +761,7 @@ struct Proc
 	ulong	delaysched;
 	ulong	priority;	/* priority level */
 	ulong	basepri;	/* base priority level */
-	uchar	fixedpri;	/* priority level deson't change */
+	uchar	fixedpri;	/* priority level doesn't change */
 	ulong	cpu;		/* cpu average */
 	ulong	lastupdate;
 	uchar	yield;		/* non-zero if the process just did a sleep(0) */
diff --git a/sys/src/9/port/proc.c b/sys/src/9/port/proc.c
index c8deaa401..0af37022c 100644
--- a/sys/src/9/port/proc.c
+++ b/sys/src/9/port/proc.c
@@ -23,8 +23,9 @@ ulong	load;
 static struct Procalloc
 {
 	Lock;
-	Proc*	arena;
-	Proc*	free;
+	Proc	**tab;
+	Proc	*free;
+	int	nextindex;
 } procalloc;
 
 enum
@@ -630,13 +631,25 @@ canpage(Proc *p)
 Proc*
 newproc(void)
 {
+	char *b;
 	Proc *p;
 
 	lock(&procalloc);
 	p = procalloc.free;
-	if(p == nil || (p->kstack == nil && (p->kstack = malloc(KSTACK)) == nil)){
-		unlock(&procalloc);
-		return nil;
+	if(p == nil){
+		if(procalloc.nextindex >= conf.nproc){
+			unlock(&procalloc);
+			return nil;
+		}
+		b = malloc(KSTACK+sizeof(Proc));
+		if(b == nil){
+			unlock(&procalloc);
+			return nil;
+		}
+		p = (Proc*)(b + KSTACK);
+		p->index = procalloc.nextindex++;
+		p->kstack = b;
+		procalloc.tab[p->index] = p;
 	}
 	procalloc.free = p->qnext;
 	p->qnext = nil;
@@ -682,8 +695,7 @@ procwired(Proc *p, int bm)
 		/* pick a machine to wire to */
 		memset(nwired, 0, sizeof(nwired));
 		p->wired = nil;
-		for(i=0; i<conf.nproc; i++){
-			pp = proctab(i);
+		for(i=0; (pp = proctab(i)) != nil; i++){
 			wm = pp->wired;
 			if(wm != nil && pp->pid)
 				nwired[wm->machno]++;
@@ -720,20 +732,14 @@ procpriority(Proc *p, int pri, int fixed)
 void
 procinit0(void)		/* bad planning - clashes with devproc.c */
 {
-	Proc *p;
-	int i;
-
-	p = xalloc(conf.nproc*sizeof(Proc));
-	if(p == nil){
+	procalloc.free = nil;
+	/* allocate 1 extra for a nil terminator */
+	procalloc.tab = xalloc((conf.nproc+1)*sizeof(Proc*));
+	if(procalloc.tab == nil){
 		xsummary();
 		panic("cannot allocate %lud procs (%ludMB)", conf.nproc, conf.nproc*sizeof(Proc)/(1024*1024));
 	}
-	procalloc.arena = p;
-	procalloc.free = p;
-	for(i=0; i<conf.nproc-1; i++, p++)
-		p->qnext = p+1;
-	p->qnext = nil;
-
+	memset(procalloc.tab, 0, (conf.nproc+1)*sizeof(Proc*));
 	pidinit();
 }
 
@@ -1263,7 +1269,7 @@ pwait(Waitmsg *w)
 Proc*
 proctab(int i)
 {
-#define proctab(x) (&procalloc.arena[(x)])
+#define proctab(x) (procalloc.tab[(x)])
 	return proctab(i);
 }
 
@@ -1304,8 +1310,7 @@ procflushmmu(int (*match)(Proc*, void*), void *a)
 	 */
 	memset(await, 0, conf.nmach*sizeof(await[0]));
 	nwait = 0;
-	for(i = 0; i < conf.nproc; i++){
-		p = proctab(i);
+	for(i = 0; (p = proctab(i)) != nil; i++){
 		if(p->state != Dead && (*match)(p, a)){
 			p->newtlb = 1;
 			for(nm = 0; nm < conf.nmach; nm++){
@@ -1572,8 +1577,7 @@ killbig(char *why)
 
 	max = 0;
 	kp = nil;
-	for(i = 0; i < conf.nproc; i++) {
-		p = proctab(i);
+	for(i = 0; (p = proctab(i)) != nil; i++) {
 		if(p->state == Dead || p->kp || p->parentpid == 0)
 			continue;
 		if((p->noswap || (p->procmode & 0222) == 0) && strcmp(eve, p->user) == 0)
@@ -1588,8 +1592,7 @@ killbig(char *why)
 		return;
 	print("%lud: %s killed: %s\n", kp->pid, kp->text, why);
 	qlock(&kp->seglock);
-	for(i = 0; i < conf.nproc; i++) {
-		p = proctab(i);
+	for(i = 0; (p = proctab(i)) != nil; i++) {
 		if(p->state == Dead || p->kp)
 			continue;
 		if(p != kp && p->seg[BSEG] != nil && p->seg[BSEG] == kp->seg[BSEG])
@@ -1624,8 +1627,7 @@ renameuser(char *old, char *new)
 	Proc *p;
 	int i;
 
-	for(i = 0; i < conf.nproc; i++){
-		p = proctab(i);
+	for(i = 0; (p = proctab(i)) != nil; i++){
 		qlock(&p->debug);
 		if(p->user != nil && strcmp(old, p->user) == 0)
 			kstrdup(&p->user, new);
@@ -1809,12 +1811,15 @@ piddel(Pid *i)
 int
 procindex(ulong pid)
 {
+	Proc *p;
 	Pid *i;
+	int x;
 
 	i = pidlookup(pid);
 	if(i != nil){
-		int x = i->procindex;
-		if(proctab(x)->pid == pid)
+		x = i->procindex;
+		p = proctab(x);
+		if(p != nil && p->pid == pid)
 			return x;
 	}
 	return -1;
@@ -1869,7 +1874,7 @@ pidalloc(Proc *p)
 		p->parentpid = 0;
 
 	i = pidadd(0);
-	i->procindex = (int)(p - procalloc.arena);
+	i->procindex = p->index;
 
 	if(p->noteid == 0){
 		incref(i);
diff --git a/sys/src/9/ppc/main.c b/sys/src/9/ppc/main.c
index 1dcd9ab22..059469a4f 100644
--- a/sys/src/9/ppc/main.c
+++ b/sys/src/9/ppc/main.c
@@ -246,8 +246,8 @@ confinit(void)
 	conf.nproc = 100 + ((conf.npage*BY2PG)/MB)*5;
 	if(cpuserver)
 		conf.nproc *= 3;
-	if(conf.nproc > 2000)
-		conf.nproc = 2000;
+	if(conf.nproc > 4000)
+		conf.nproc = 4000;
 	conf.nimage = 200;
 	conf.nswap = conf.nproc*80;
 	conf.nswppo = 4096;
diff --git a/sys/src/9/xen/main.c b/sys/src/9/xen/main.c
index 32fac0679..91eae7923 100644
--- a/sys/src/9/xen/main.c
+++ b/sys/src/9/xen/main.c
@@ -252,8 +252,8 @@ confinit(void)
 	conf.nproc = 100 + ((conf.npage*BY2PG)/MB)*5;
 	if(cpuserver)
 		conf.nproc *= 3;
-	if(conf.nproc > 2000)
-		conf.nproc = 2000;
+	if(conf.nproc > 4000)
+		conf.nproc = 4000;
 	conf.nimage = 200;
 	conf.nswap = conf.nproc*80;
 	conf.nswppo = 4096;