Stopping counting and reporting CPU time spent spinning on a lock as

system time.

Introduce a new CP_SPIN "scheduler state" and modify userland tools
to display the % of timer a CPU spents spinning.

Based on a diff from jmatthew@, ok pirofti@, bluhm@, visa@, deraadt@

6 files changed, 44 insertions, 35 deletions

diff --git a/sys/kern/kern_clock.c b/sys/kern/kern_clock.c
index 2aef10f72c5..d77a4191d8c 100644
--- a/sys/kern/kern_clock.c
+++ b/sys/kern/kern_clock.c
@@ -1,4 +1,4 @@
-/*	$OpenBSD: kern_clock.c,v 1.93 2017/07/22 14:33:45 kettenis Exp $	*/
+/*	$OpenBSD: kern_clock.c,v 1.94 2018/05/14 12:31:21 mpi Exp $	*/
 /*	$NetBSD: kern_clock.c,v 1.34 1996/06/09 04:51:03 briggs Exp $	*/
 
 /*-
@@ -381,6 +381,7 @@ statclock(struct clockframe *frame)
 			return;
 		/*
 		 * Came from kernel mode, so we were:
+		 * - spinning on a lock
 		 * - handling an interrupt,
 		 * - doing syscall or trap work on behalf of the current
 		 *   user process, or
@@ -391,7 +392,9 @@ statclock(struct clockframe *frame)
 		 * so that we know how much of its real time was spent
 		 * in ``non-process'' (i.e., interrupt) work.
 		 */
-		if (CLKF_INTR(frame)) {
+		if (spc->spc_spinning)
+			spc->spc_cp_time[CP_SPIN]++;
+		else if (CLKF_INTR(frame)) {
 			if (p != NULL)
 				p->p_iticks++;
 			spc->spc_cp_time[CP_INTR]++;
diff --git a/sys/kern/kern_lock.c b/sys/kern/kern_lock.c
index e1a8adc59a7..657e028d4c5 100644
--- a/sys/kern/kern_lock.c
+++ b/sys/kern/kern_lock.c
@@ -1,4 +1,4 @@
-/*	$OpenBSD: kern_lock.c,v 1.63 2018/04/26 06:51:48 mpi Exp $	*/
+/*	$OpenBSD: kern_lock.c,v 1.64 2018/05/14 12:31:21 mpi Exp $	*/
 
 /*
  * Copyright (c) 2017 Visa Hankala
@@ -113,10 +113,12 @@ ___mp_lock_init(struct __mp_lock *mpl, struct lock_type *type)
 static __inline void
 __mp_lock_spin(struct __mp_lock *mpl, u_int me)
 {
+	struct schedstate_percpu *spc = &curcpu()->ci_schedstate;
 #ifdef MP_LOCKDEBUG
 	int nticks = __mp_lock_spinout;
 #endif
 
+	spc->spc_spinning++;
 	while (mpl->mpl_ticket != me) {
 		CPU_BUSY_CYCLE();
 
@@ -128,6 +130,7 @@ __mp_lock_spin(struct __mp_lock *mpl, u_int me)
 		}
 #endif
 	}
+	spc->spc_spinning--;
 }
 
 void
@@ -257,10 +260,12 @@ __mtx_init(struct mutex *mtx, int wantipl)
 void
 __mtx_enter(struct mutex *mtx)
 {
+	struct schedstate_percpu *spc = &curcpu()->ci_schedstate;
 #ifdef MP_LOCKDEBUG
 	int nticks = __mp_lock_spinout;
 #endif
 
+	spc->spc_spinning++;
 	while (__mtx_enter_try(mtx) == 0) {
 		CPU_BUSY_CYCLE();
 
@@ -272,6 +277,7 @@ __mtx_enter(struct mutex *mtx)
 		}
 #endif
 	}
+	spc->spc_spinning--;
 }
 
 int
diff --git a/sys/sys/sched.h b/sys/sys/sched.h
index 698eda820a0..a8b39ec11f0 100644
--- a/sys/sys/sched.h
+++ b/sys/sys/sched.h
@@ -1,4 +1,4 @@
-/*	$OpenBSD: sched.h,v 1.44 2017/12/14 23:21:04 dlg Exp $	*/
+/*	$OpenBSD: sched.h,v 1.45 2018/05/14 12:31:21 mpi Exp $	*/
 /* $NetBSD: sched.h,v 1.2 1999/02/28 18:14:58 ross Exp $ */
 
 /*-
@@ -83,17 +83,20 @@
 #define CP_USER		0
 #define CP_NICE		1
 #define CP_SYS		2
-#define CP_INTR		3
-#define CP_IDLE		4
-#define CPUSTATES	5
+#define CP_SPIN		3
+#define CP_INTR		4
+#define CP_IDLE		5
+#define CPUSTATES	6
 
 #define	SCHED_NQS	32			/* 32 run queues. */
 
 /*
  * Per-CPU scheduler state.
- * XXX - expose to userland for now.
  */
 struct schedstate_percpu {
+	struct proc *spc_idleproc;	/* idle proc for this cpu */
+	TAILQ_HEAD(prochead, proc) spc_qs[SCHED_NQS];
+	LIST_HEAD(,proc) spc_deadproc;
 	struct timespec spc_runtime;	/* time curproc started running */
 	volatile int spc_schedflags;	/* flags; see below */
 	u_int spc_schedticks;		/* ticks for schedclock() */
@@ -102,18 +105,12 @@ struct schedstate_percpu {
 	int spc_rrticks;		/* ticks until roundrobin() */
 	int spc_pscnt;			/* prof/stat counter */
 	int spc_psdiv;			/* prof/stat divisor */	
-	struct proc *spc_idleproc;	/* idle proc for this cpu */
 
 	u_int spc_nrun;			/* procs on the run queues */
 	fixpt_t spc_ldavg;		/* shortest load avg. for this cpu */
 
-	TAILQ_HEAD(prochead, proc) spc_qs[SCHED_NQS];
 	volatile uint32_t spc_whichqs;
-
-#ifdef notyet
-	struct proc *spc_reaper;	/* dead proc reaper */
-#endif
-	LIST_HEAD(,proc) spc_deadproc;
+	volatile u_int spc_spinning;	/* this cpu is currently spinning */
 };
 
 #ifdef	_KERNEL
diff --git a/usr.bin/systat/cpu.c b/usr.bin/systat/cpu.c
index 90cdeeed6c2..ac7e2ba000e 100644
--- a/usr.bin/systat/cpu.c
+++ b/usr.bin/systat/cpu.c
@@ -1,4 +1,4 @@
-/*	$OpenBSD: cpu.c,v 1.5 2016/01/02 20:02:40 benno Exp $	*/
+/*	$OpenBSD: cpu.c,v 1.6 2018/05/14 12:31:21 mpi Exp $	*/
 
 /*
  * Copyright (c) 2013 Reyk Floeter <reyk@openbsd.org>
@@ -68,21 +68,23 @@ field_def fields_cpu[] = {
 	{ "User", 10, 20, 1, FLD_ALIGN_RIGHT, -1, 0, 0, 0 },
 	{ "Nice", 10, 20, 1, FLD_ALIGN_RIGHT, -1, 0, 0, 0 },
 	{ "System", 10, 20, 1, FLD_ALIGN_RIGHT, -1, 0, 0, 0 },
+	{ "Spin", 10, 20, 1, FLD_ALIGN_RIGHT, -1, 0, 0, 0 },
 	{ "Interrupt", 10, 20, 1, FLD_ALIGN_RIGHT, -1, 0, 0, 0 },
 	{ "Idle", 10, 20, 1, FLD_ALIGN_RIGHT, -1, 0, 0, 0 },
 };
 
 #define FLD_CPU_CPU	FIELD_ADDR(fields_cpu, 0)
-#define FLD_CPU_INT	FIELD_ADDR(fields_cpu, 1)
-#define FLD_CPU_SYS	FIELD_ADDR(fields_cpu, 2)
-#define FLD_CPU_USR	FIELD_ADDR(fields_cpu, 3)
-#define FLD_CPU_NIC	FIELD_ADDR(fields_cpu, 4)
-#define FLD_CPU_IDLE	FIELD_ADDR(fields_cpu, 5)
+#define FLD_CPU_USR	FIELD_ADDR(fields_cpu, 1)
+#define FLD_CPU_NIC	FIELD_ADDR(fields_cpu, 2)
+#define FLD_CPU_SYS	FIELD_ADDR(fields_cpu, 3)
+#define FLD_CPU_SPIN	FIELD_ADDR(fields_cpu, 4)
+#define FLD_CPU_INT	FIELD_ADDR(fields_cpu, 5)
+#define FLD_CPU_IDLE	FIELD_ADDR(fields_cpu, 6)
 
 /* Define views */
 field_def *view_cpu_0[] = {
-	FLD_CPU_CPU,
-	FLD_CPU_INT, FLD_CPU_SYS, FLD_CPU_USR, FLD_CPU_NIC, FLD_CPU_IDLE, NULL
+	FLD_CPU_CPU, FLD_CPU_USR, FLD_CPU_NIC, FLD_CPU_SYS, FLD_CPU_SPIN,
+	FLD_CPU_INT, FLD_CPU_IDLE, NULL
 };
 
 /* Define view managers */
@@ -232,11 +234,12 @@ initcpu(void)
 	do {								\
 		if (cur >= dispstart && cur < end) { 			\
 			print_fld_size(FLD_CPU_CPU, (v));		\
-			print_fld_percentage(FLD_CPU_INT, (cs[0]));	\
-			print_fld_percentage(FLD_CPU_SYS, (cs[1]));	\
-			print_fld_percentage(FLD_CPU_USR, (cs[2]));	\
-			print_fld_percentage(FLD_CPU_NIC, (cs[3]));	\
-			print_fld_percentage(FLD_CPU_IDLE, (cs[4]));	\
+			print_fld_percentage(FLD_CPU_USR, (cs[CP_USER]));\
+			print_fld_percentage(FLD_CPU_NIC, (cs[CP_NICE]));\
+			print_fld_percentage(FLD_CPU_SYS, (cs[CP_SYS]));\
+			print_fld_percentage(FLD_CPU_SPIN, (cs[CP_SPIN]));\
+			print_fld_percentage(FLD_CPU_INT, (cs[CP_INTR]));\
+			print_fld_percentage(FLD_CPU_IDLE, (cs[CP_IDLE]));	\
 			end_line();					\
 		}							\
 		if (++cur >= end)					\
diff --git a/usr.bin/systat/vmstat.c b/usr.bin/systat/vmstat.c
index a515b73f958..13feb293b6e 100644
--- a/usr.bin/systat/vmstat.c
+++ b/usr.bin/systat/vmstat.c
@@ -1,4 +1,4 @@
-/*	$OpenBSD: vmstat.c,v 1.83 2018/05/03 07:49:18 otto Exp $	*/
+/*	$OpenBSD: vmstat.c,v 1.84 2018/05/14 12:31:21 mpi Exp $	*/
 /*	$NetBSD: vmstat.c,v 1.5 1996/05/10 23:16:40 thorpej Exp $	*/
 
 /*-
@@ -268,7 +268,7 @@ labelkre(void)
 	mvprintw(GENSTATROW, GENSTATCOL, "   Csw   Trp   Sys   Int   Sof  Flt");
 
 	mvprintw(GRAPHROW, GRAPHCOL,
-	    "    . %%Int    . %%Sys    . %%Usr    . %%Nic    . %%Idle");
+	    "    . %%Int    . %%Sys    . %%Usr    . %%Spn    . %%Idle");
 	mvprintw(PROCSROW, PROCSCOL, "Proc:r  d  s  w");
 	mvprintw(GRAPHROW + 1, GRAPHCOL,
 	    "|    |    |    |    |    |    |    |    |    |    |");
@@ -306,8 +306,8 @@ labelkre(void)
 	} while (0)
 #define MAXFAIL 5
 
-static	char cpuchar[CPUSTATES] = { '|', '=', '>', '-', ' ' };
-static	char cpuorder[CPUSTATES] = { CP_INTR, CP_SYS, CP_USER, CP_NICE, CP_IDLE };
+static	char cpuchar[] = { '|', '=', '>', '-', ' ' };
+static	char cpuorder[] = { CP_INTR, CP_SYS, CP_USER, CP_SPIN, CP_IDLE };
 
 void
 showkre(void)
@@ -367,7 +367,7 @@ showkre(void)
 	psiz = 0;
 	f2 = 0.0;
 
-	for (c = 0; c < CPUSTATES; c++) {
+	for (c = 0; c < nitems(cpuorder); c++) {
 		i = cpuorder[c];
 		f1 = cputime(i);
 		f2 += f1;
diff --git a/usr.bin/top/machine.c b/usr.bin/top/machine.c
index 0b4d5d7e748..4643778a15c 100644
--- a/usr.bin/top/machine.c
+++ b/usr.bin/top/machine.c
@@ -1,4 +1,4 @@
-/* $OpenBSD: machine.c,v 1.89 2017/05/30 06:01:30 tedu Exp $	 */
+/* $OpenBSD: machine.c,v 1.90 2018/05/14 12:31:21 mpi Exp $	 */
 
 /*-
  * Copyright (c) 1994 Thorsten Lockert <tholo@sigmasoft.com>
@@ -108,7 +108,7 @@ char *procstatenames[] = {
 /* these are for detailing the cpu states */
 int64_t *cpu_states;
 char *cpustatenames[] = {
-	"user", "nice", "system", "interrupt", "idle", NULL
+	"user", "nice", "sys", "spin", "intr", "idle", NULL
 };
 
 /* these are for detailing the memory statistics */