redo the performance throttling in the kernel.

introduce a new sysctl, hw.perfpolicy, that governs the policy.
when set to anything other than manual, hw.setperf then becomes read only.
phessler was heading in this direction, but this is slightly different. :)

5 files changed, 178 insertions, 25 deletions

diff --git a/sys/kern/init_main.c b/sys/kern/init_main.c
index e9e5827fb4b..4d5b672a619 100644
--- a/sys/kern/init_main.c
+++ b/sys/kern/init_main.c
@@ -1,4 +1,4 @@
-/*	$OpenBSD: init_main.c,v 1.221 2014/10/13 22:42:22 deraadt Exp $	*/
+/*	$OpenBSD: init_main.c,v 1.222 2014/10/17 01:51:39 tedu Exp $	*/
 /*	$NetBSD: init_main.c,v 1.84.4.1 1996/06/02 09:08:06 mrg Exp $	*/
 
 /*
@@ -116,6 +116,11 @@ struct	sigacts sigacts0;
 struct	process *initprocess;
 struct	proc *reaperproc;
 
+#ifndef SMALL_KERNEL
+extern struct timeout setperf_to;
+void setperf_auto(void *);
+#endif
+
 int	cmask = CMASK;
 extern	struct user *proc0paddr;
 
@@ -545,6 +550,10 @@ main(void *framep)
 	start_init_exec = 1;
 	wakeup((void *)&start_init_exec);
 
+#ifndef SMALL_KERNEL
+	timeout_set(&setperf_to, setperf_auto, NULL);
+#endif
+
         /*
          * proc0: nothing to do, back to sleep
          */
diff --git a/sys/kern/kern_sysctl.c b/sys/kern/kern_sysctl.c
index 15c467c23ba..dcf6b861981 100644
--- a/sys/kern/kern_sysctl.c
+++ b/sys/kern/kern_sysctl.c
@@ -1,4 +1,4 @@
-/*	$OpenBSD: kern_sysctl.c,v 1.266 2014/10/11 17:12:30 deraadt Exp $	*/
+/*	$OpenBSD: kern_sysctl.c,v 1.267 2014/10/17 01:51:39 tedu Exp $	*/
 /*	$NetBSD: kern_sysctl.c,v 1.17 1996/05/20 17:49:05 mrg Exp $	*/
 
 /*-
@@ -127,8 +127,6 @@ void fill_file(struct kinfo_file *, struct file *, struct filedesc *,
 void fill_kproc(struct process *, struct kinfo_proc *, struct proc *, int);
 
 int (*cpu_cpuspeed)(int *);
-void (*cpu_setperf)(int);
-int perflevel = 100;
 
 /*
  * Lock to avoid too many processes vslocking a large amount of memory
@@ -649,11 +647,6 @@ hw_sysctl(int *name, u_int namelen, void *oldp, size_t *oldlenp, void *newp,
 		    disk_count * sizeof(struct diskstats)));
 	case HW_DISKCOUNT:
 		return (sysctl_rdint(oldp, oldlenp, newp, disk_count));
-#ifndef	SMALL_KERNEL
-	case HW_SENSORS:
-		return (sysctl_sensors(name + 1, namelen - 1, oldp, oldlenp,
-		    newp, newlen));
-#endif
 	case HW_CPUSPEED:
 		if (!cpu_cpuspeed)
 			return (EOPNOTSUPP);
@@ -661,19 +654,15 @@ hw_sysctl(int *name, u_int namelen, void *oldp, size_t *oldlenp, void *newp,
 		if (err)
 			return err;
 		return (sysctl_rdint(oldp, oldlenp, newp, cpuspeed));
+#ifndef	SMALL_KERNEL
+	case HW_SENSORS:
+		return (sysctl_sensors(name + 1, namelen - 1, oldp, oldlenp,
+		    newp, newlen));
 	case HW_SETPERF:
-		if (!cpu_setperf)
-			return (EOPNOTSUPP);
-		err = sysctl_int(oldp, oldlenp, newp, newlen, &perflevel);
-		if (err)
-			return err;
-		if (perflevel > 100)
-			perflevel = 100;
-		if (perflevel < 0)
-			perflevel = 0;
-		if (newp)
-			cpu_setperf(perflevel);
-		return (0);
+		return (sysctl_hwsetperf(oldp, oldlenp, newp, newlen));
+	case HW_PERFPOLICY:
+		return (sysctl_hwperfpolicy(oldp, oldlenp, newp, newlen));
+#endif /* !SMALL_KERNEL */
 	case HW_VENDOR:
 		if (hw_vendor)
 			return (sysctl_rdstring(oldp, oldlenp, newp,
diff --git a/sys/kern/sched_bsd.c b/sys/kern/sched_bsd.c
index 98bb478b8bc..67813da2131 100644
--- a/sys/kern/sched_bsd.c
+++ b/sys/kern/sched_bsd.c
@@ -1,4 +1,4 @@
-/*	$OpenBSD: sched_bsd.c,v 1.35 2014/07/04 05:58:31 guenther Exp $	*/
+/*	$OpenBSD: sched_bsd.c,v 1.36 2014/10/17 01:51:39 tedu Exp $	*/
 /*	$NetBSD: kern_synch.c,v 1.37 1996/04/22 01:38:37 christos Exp $	*/
 
 /*-
@@ -41,6 +41,7 @@
 #include <sys/systm.h>
 #include <sys/proc.h>
 #include <sys/kernel.h>
+#include <sys/malloc.h>
 #include <sys/buf.h>
 #include <sys/signalvar.h>
 #include <sys/resourcevar.h>
@@ -563,3 +564,152 @@ schedclock(struct proc *p)
 		p->p_priority = p->p_usrpri;
 	SCHED_UNLOCK(s);
 }
+
+#ifndef SMALL_KERNEL
+/*
+ * The code below handles CPU throttling.
+ */
+#include <sys/sysctl.h>
+
+#define PERFPOL_MANUAL 0
+#define PERFPOL_AUTO 1
+#define PERFPOL_HIGH 2
+int perflevel = 100;
+int perfpolicy = PERFPOL_MANUAL;
+
+void (*cpu_setperf)(int);
+
+struct timeout setperf_to;
+void setperf_auto(void *);
+
+void
+setperf_auto(void *v)
+{
+	static uint64_t *idleticks, *totalticks;
+
+	int i, j;
+	int speedup;
+	CPU_INFO_ITERATOR cii;
+	struct cpu_info *ci;
+	uint64_t idle, total, allidle, alltotal;
+
+	if (perfpolicy != PERFPOL_AUTO)
+		return;
+
+	if (!idleticks)
+		if (!(idleticks = malloc(sizeof(*idleticks) * ncpusfound,
+		    M_DEVBUF, M_NOWAIT | M_ZERO)))
+			return;
+	if (!totalticks)
+		if (!(totalticks = malloc(sizeof(*totalticks) * ncpusfound,
+		    M_DEVBUF, M_NOWAIT | M_ZERO))) {
+			free(idleticks, M_DEVBUF, 0);
+			return;
+		}
+
+	alltotal = allidle = 0;
+	j = 0;
+	speedup = 0;
+	CPU_INFO_FOREACH(cii, ci) {
+		total = 0;
+		for (i = 0; i < CPUSTATES; i++) {
+			total += ci->ci_schedstate.spc_cp_time[i];
+		}
+		total -= totalticks[j];
+		idle = ci->ci_schedstate.spc_cp_time[CP_IDLE] - idleticks[j];
+		if (idle < total / 3)
+			speedup = 1;
+		alltotal += total;
+		allidle += idle;
+		idleticks[j] += idle;
+		totalticks[j] += total;
+		j++;
+	}
+	if (allidle < alltotal / 2)
+		speedup = 1;
+
+	if (speedup && perflevel != 100) {
+		perflevel = 100;
+		cpu_setperf(perflevel);
+	} else if (!speedup && perflevel != 0) {
+		perflevel = 0;
+		cpu_setperf(perflevel);
+	}
+	
+	timeout_add_msec(&setperf_to, 100);
+}
+
+int
+sysctl_hwsetperf(void *oldp, size_t *oldlenp, void *newp, size_t newlen)
+{
+	int err, newperf;
+
+	if (!cpu_setperf)
+		return EOPNOTSUPP;
+
+	if (perfpolicy != PERFPOL_MANUAL)
+		return sysctl_rdint(oldp, oldlenp, newp, perflevel);
+	
+	newperf = perflevel;
+	err = sysctl_int(oldp, oldlenp, newp, newlen, &newperf);
+	if (err)
+		return err;
+	if (newperf > 100)
+		newperf = 100;
+	if (newperf < 0)
+		newperf = 0;
+	perflevel = newperf;
+	cpu_setperf(perflevel);
+
+	return 0;
+}
+
+int
+sysctl_hwperfpolicy(void *oldp, size_t *oldlenp, void *newp, size_t newlen)
+{
+	char policy[32];
+	int err;
+
+	if (!cpu_setperf)
+		return EOPNOTSUPP;
+
+	switch (perfpolicy) {
+	case PERFPOL_MANUAL:
+		strlcpy(policy, "manual", sizeof(policy));
+		break;
+	case PERFPOL_AUTO:
+		strlcpy(policy, "auto", sizeof(policy));
+		break;
+	case PERFPOL_HIGH:
+		strlcpy(policy, "high", sizeof(policy));
+		break;
+	default:
+		strlcpy(policy, "unknown", sizeof(policy));
+		break;
+	}
+
+	if (newp == NULL)
+		return sysctl_rdstring(oldp, oldlenp, newp, policy);
+
+	err = sysctl_string(oldp, oldlenp, newp, newlen, policy, sizeof(policy));
+	if (err)
+		return err;
+	if (strcmp(policy, "manual") == 0)
+		perfpolicy = PERFPOL_MANUAL;
+	else if (strcmp(policy, "auto") == 0)
+		perfpolicy = PERFPOL_AUTO;
+	else if (strcmp(policy, "high") == 0)
+		perfpolicy = PERFPOL_HIGH;
+	else
+		return EINVAL;
+
+	if (perfpolicy == PERFPOL_AUTO) {
+		timeout_add_msec(&setperf_to, 200);
+	} else if (perfpolicy == PERFPOL_HIGH) {
+		perflevel = 100;
+		cpu_setperf(perflevel);
+	}
+	return 0;
+}
+#endif
+
diff --git a/sys/sys/sched.h b/sys/sys/sched.h
index 731d9dcbfd1..032580f5fe1 100644
--- a/sys/sys/sched.h
+++ b/sys/sys/sched.h
@@ -1,4 +1,4 @@
-/*	$OpenBSD: sched.h,v 1.36 2014/09/09 07:07:39 blambert Exp $	*/
+/*	$OpenBSD: sched.h,v 1.37 2014/10/17 01:51:39 tedu Exp $	*/
 /* $NetBSD: sched.h,v 1.2 1999/02/28 18:14:58 ross Exp $ */
 
 /*-
@@ -153,6 +153,9 @@ void cpu_idle_cycle(void);
 void cpu_idle_leave(void);
 void sched_peg_curproc(struct cpu_info *ci);
 
+int sysctl_hwsetperf(void *, size_t *, void *, size_t);
+int sysctl_hwperfpolicy(void *, size_t *, void *, size_t);
+
 #ifdef MULTIPROCESSOR
 void sched_start_secondary_cpus(void);
 void sched_stop_secondary_cpus(void);
diff --git a/sys/sys/sysctl.h b/sys/sys/sysctl.h
index 8bb7f090cf3..665aae654c6 100644
--- a/sys/sys/sysctl.h
+++ b/sys/sys/sysctl.h
@@ -1,4 +1,4 @@
-/*	$OpenBSD: sysctl.h,v 1.149 2014/08/20 06:23:03 mikeb Exp $	*/
+/*	$OpenBSD: sysctl.h,v 1.150 2014/10/17 01:51:39 tedu Exp $	*/
 /*	$NetBSD: sysctl.h,v 1.16 1996/04/09 20:55:36 cgd Exp $	*/
 
 /*
@@ -798,7 +798,8 @@ struct kinfo_file {
 #define	HW_USERMEM64		20	/* quad: non-kernel memory */
 #define	HW_NCPUFOUND		21	/* int: number of cpus found*/
 #define	HW_ALLOWPOWERDOWN	22	/* allow power button shutdown */
-#define	HW_MAXID		23	/* number of valid hw ids */
+#define	HW_PERFPOLICY		23	/* set performance policy */
+#define	HW_MAXID		24	/* number of valid hw ids */
 
 #define	CTL_HW_NAMES { \
 	{ 0, 0 }, \
@@ -824,6 +825,7 @@ struct kinfo_file {
 	{ "usermem", CTLTYPE_QUAD }, \
 	{ "ncpufound", CTLTYPE_INT }, \
 	{ "allowpowerdown", CTLTYPE_INT }, \
+	{ "perfpolicy", CTLTYPE_STRING }, \
 }
 
 /*