Attachment 221999 Details for Bug 246940 – Patch for 12-STABLE powerd(8) to weight (scale) nice and user idle CPU load

[patch] Patch for 12-STABLE powerd(8) to weight (scale) nice and user idle CPU load

cp_time_ext.powerd.diff (text/plain), 13.75 KB, created by Walter von Entferndt on 2021-01-28 17:43:54 UTC

(hide)

Description:

Filename:

MIME Type:

Creator: Walter von Entferndt

Created: 2021-01-28 17:43:54 UTC

Size: 13.75 KB

patch

obsolete

>--- usr.sbin/powerd/powerd.c
>+++ usr.sbin/powerd/powerd.c
>@@ -24,15 +24,15 @@
>  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT,
>  * STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING
>  * IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
>  * POSSIBILITY OF SUCH DAMAGE.
>  */
> 
> #include <sys/cdefs.h>
>-__FBSDID("$FreeBSD: stable/12/usr.sbin/powerd/powerd.c 358990 2020-03-14 06:57:50Z cperciva $");
>+__FBSDID("$FreeBSD: head/usr.sbin/powerd/powerd.c 358478 2020-02-29 22:31:23Z cperciva $");
> 
> #include <sys/param.h>
> #include <sys/ioctl.h>
> #include <sys/sysctl.h>
> #include <sys/resource.h>
> #include <sys/socket.h>
> #include <sys/time.h>
>@@ -82,28 +82,29 @@
> 
> #define ACPIAC		"hw.acpi.acline"
> #define PMUAC		"dev.pmu.0.acline"
> #define APMDEV		"/dev/apm"
> #define DEVDPIPE	"/var/run/devd.pipe"
> #define DEVCTL_MAXBUF	1024
> 
>-static int	read_usage_times(int *load, int nonice);
>+static int	read_usage_times(int *load, int nonice, int noidle);
> static int	read_freqs(int *numfreqs, int **freqs, int **power,
> 		    int minfreq, int maxfreq);
> static int	set_freq(int freq);
> static void	acline_init(void);
> static void	acline_read(void);
> static int	devd_init(void);
> static void	devd_close(void);
> static void	handle_sigs(int sig);
> static void	parse_mode(char *arg, int *mode, int ch);
> static void	usage(void);
> 
> /* Sysctl data structures. */
> static int	cp_times_mib[2];
>+static int	cpx_times_mib[2];
> static int	freq_mib[4];
> static int	levels_mib[4];
> static int	acline_mib[4];
> static size_t	acline_mib_len;
> 
> /* Configuration */
> static int	cpu_running_mark;
>@@ -132,23 +133,26 @@
> static struct timeval tried_devd;
> 
> /*
>  * This function returns summary load of all CPUs.  It was made so
>  * intentionally to not reduce performance in scenarios when several
>  * threads are processing requests as a pipeline -- running one at
>  * a time on different CPUs and waiting for each other.  If nonice
>- * is nonzero, only user+sys+intr time will be counted as load; any
>- * nice time will be treated as if idle.
>+ * is 100, only user+sys+intr time will be counted as load; any
>+ * nice time will be treated as if idle, likewise noidle.  I.e. these
>+ * nice and/or idle loads are weighted by nonice and noidle, resp.
>+ * NOTE nonice and noidle are the "reverse" from the the command line.
>  */
> static int
>-read_usage_times(int *load, int nonice)
>+read_usage_times(int *load, int nonice, int noidle)
> {
> 	static long *cp_times = NULL, *cp_times_old = NULL;
>+	static long *cpx_times = NULL, *cpx_times_old = NULL;
> 	static int ncpus = 0;
>-	size_t cp_times_len;
>+	size_t cp_times_len, cpx_times_len;
> 	int error, cpu, i, total, excl;
> 
> 	if (cp_times == NULL) {
> 		cp_times_len = 0;
> 		error = sysctl(cp_times_mib, 2, NULL, &cp_times_len, NULL, 0);
> 		if (error)
> 			return (error);
>@@ -157,40 +161,71 @@
> 		if ((cp_times_old = malloc(cp_times_len)) == NULL) {
> 			free(cp_times);
> 			cp_times = NULL;
> 			return (errno);
> 		}
> 		ncpus = cp_times_len / (sizeof(long) * CPUSTATES);
> 	}
>-
> 	cp_times_len = sizeof(long) * CPUSTATES * ncpus;
> 	error = sysctl(cp_times_mib, 2, cp_times, &cp_times_len, NULL, 0);
> 	if (error)
> 		return (error);
> 
>+	if (noidle) {
>+		if (cpx_times == NULL) {
>+			cpx_times_len = 0;
>+			error = sysctl(cpx_times_mib, 2, NULL, &cpx_times_len, NULL, 0);
>+			if (error)
>+				return (error);
>+			if ((cpx_times = malloc(cpx_times_len)) == NULL)
>+				return (errno);
>+			if ((cpx_times_old = malloc(cpx_times_len)) == NULL) {
>+				free(cpx_times);
>+				cpx_times = NULL;
>+				return (errno);
>+			}
>+		}
>+		/* FIXME This could be as assert (...);				*/
>+		/* ncpus = cpx_times_len / (sizeof(long) * CPUSTATES_EXT);	*/
>+		cpx_times_len = sizeof(long) * CPUSTATES_EXT * ncpus;
>+		error = sysctl(cpx_times_mib, 2, cpx_times, &cpx_times_len, NULL, 0);
>+		if (error)
>+			return (error);
>+	}
>+
> 	if (load) {
> 		*load = 0;
> 		for (cpu = 0; cpu < ncpus; cpu++) {
> 			total = 0;
> 			for (i = 0; i < CPUSTATES; i++) {
> 			    total += cp_times[cpu * CPUSTATES + i] -
> 				cp_times_old[cpu * CPUSTATES + i];
> 			}
> 			if (total == 0)
> 				continue;
> 			excl = cp_times[cpu * CPUSTATES + CP_IDLE] -
> 			    cp_times_old[cpu * CPUSTATES + CP_IDLE];
> 			if (nonice)
>-				excl += cp_times[cpu * CPUSTATES + CP_NICE] -
>-				    cp_times_old[cpu * CPUSTATES + CP_NICE];
>+				excl += nonice*(cp_times[cpu * CPUSTATES + CP_NICE] -
>+				    cp_times_old[cpu * CPUSTATES + CP_NICE])/100;
>+			/* These two can overlap: e.g. a command started "nice idprio 5 ..." */
>+			else if (noidle)
>+				excl += noidle*(cpx_times[cpu * CPUSTATES_EXT + CPX_NICE_IDLE] -
>+				    cpx_times_old[cpu * CPUSTATES_EXT + CPX_NICE_IDLE])/100;
>+			if (noidle)
>+				excl += noidle*(cpx_times[cpu * CPUSTATES_EXT + CPX_USER_IDLE] -
>+				    cpx_times_old[cpu * CPUSTATES_EXT + CPX_USER_IDLE])/100;
>+
> 			*load += 100 - excl * 100 / total;
> 		}
> 	}
> 
> 	memcpy(cp_times_old, cp_times, cp_times_len);
>+	if (noidle)
>+		memcpy(cpx_times_old, cpx_times, cpx_times_len);
> 
> 	return (0);
> }
> 
> static int
> read_freqs(int *numfreqs, int **freqs, int **power, int minfreq, int maxfreq)
> {
>@@ -475,15 +510,15 @@
> }
> 
> static void
> usage(void)
> {
> 
> 	fprintf(stderr,
>-"usage: powerd [-v] [-a mode] [-b mode] [-i %%] [-m freq] [-M freq] [-N] [-n mode] [-p ival] [-r %%] [-s source] [-P pidfile]\n");
>+"usage: powerd [-v] [-a mode] [-b mode] [-n mode] [-i %%] [-m freq] [-M freq] [-N %%] [-I %%] [-p ival] [-r %%] [-s source] [-P pidfile]\n");
> 	exit(1);
> }
> 
> int
> main(int argc, char * argv[])
> {
> 	struct timeval timeout;
>@@ -492,31 +527,31 @@
> 	struct pidfh *pfh = NULL;
> 	const char *pidfile = NULL;
> 	int freq, curfreq, initfreq, *freqs, i, j, *mwatts, numfreqs, load;
> 	int minfreq = -1, maxfreq = -1;
> 	int ch, mode, mode_ac, mode_battery, mode_none, idle, to;
> 	uint64_t mjoules_used;
> 	size_t len;
>-	int nonice;
>+	int nonice, noidle;
> 
> 	/* Default mode for all AC states is adaptive. */
> 	mode_ac = mode_none = MODE_HIADAPTIVE;
> 	mode_battery = MODE_ADAPTIVE;
> 	cpu_running_mark = DEFAULT_ACTIVE_PERCENT;
> 	cpu_idle_mark = DEFAULT_IDLE_PERCENT;
> 	poll_ival = DEFAULT_POLL_INTERVAL;
> 	mjoules_used = 0;
> 	vflag = 0;
>-	nonice = 0;
>+	nonice = noidle = 0;
> 
> 	/* User must be root to control frequencies. */
> 	if (geteuid() != 0)
> 		errx(1, "must be root to run");
> 
>-	while ((ch = getopt(argc, argv, "a:b:i:m:M:Nn:p:P:r:s:v")) != -1)
>+	while ((ch = getopt(argc, argv, "a:b:i:m:M:N:I:n:p:P:r:s:v")) != -1)
> 		switch (ch) {
> 		case 'a':
> 			parse_mode(optarg, &mode_ac, ch);
> 			break;
> 		case 'b':
> 			parse_mode(optarg, &mode_battery, ch);
> 			break;
>@@ -544,16 +579,29 @@
> 			if (maxfreq < 0) {
> 				warnx("%d is not a valid CPU frequency",
> 				    maxfreq);
> 				usage();
> 			}
> 			break;
> 		case 'N':
>-			nonice = 1;
>+			nonice = atoi(optarg);
>+			if (nonice < 0 || nonice > 100) {
>+				warnx("%d is not a valid percent", nonice);
>+				usage();
>+			} else
>+				nonice = 100 - nonice;
> 			break;
>+		case 'I':
>+			noidle = atoi(optarg);
>+			if (noidle < 0 || noidle > 100) {
>+				warnx("%d is not a valid percent", noidle);
>+				usage();
>+			} else
>+				noidle = 100 - noidle;
>+			break;
> 		case 'n':
> 			parse_mode(optarg, &mode_none, ch);
> 			break;
> 		case 'p':
> 			poll_ival = atoi(optarg);
> 			if (poll_ival < 5) {
> 				warnx("poll interval is in units of ms");
>@@ -583,23 +631,28 @@
> 	/* Poll interval is in units of ms. */
> 	poll_ival *= 1000;
> 
> 	/* Look up various sysctl MIBs. */
> 	len = 2;
> 	if (sysctlnametomib("kern.cp_times", cp_times_mib, &len))
> 		err(1, "lookup kern.cp_times");
>+	if (noidle) {
>+		len = 2;
>+		if (sysctlnametomib("kern.cp_times_ext", cpx_times_mib, &len))
>+			err(1, "lookup kern.cp_times_ext");
>+	}
> 	len = 4;
> 	if (sysctlnametomib("dev.cpu.0.freq", freq_mib, &len))
> 		err(EX_UNAVAILABLE, "no cpufreq(4) support -- aborting");
> 	len = 4;
> 	if (sysctlnametomib("dev.cpu.0.freq_levels", levels_mib, &len))
> 		err(1, "lookup freq_levels");
> 
> 	/* Check if we can read the load and supported freqs. */
>-	if (read_usage_times(NULL, nonice))
>+	if (read_usage_times(NULL, nonice, noidle))
> 		err(1, "read_usage_times");
> 	if (read_freqs(&numfreqs, &freqs, &mwatts, minfreq, maxfreq))
> 		err(1, "error reading supported CPU frequencies");
> 	if (numfreqs == 0)
> 		errx(1, "no CPU frequencies in user-specified range");
> 
> 	/* Run in the background unless in verbose mode. */
>@@ -773,15 +826,15 @@
> 					continue;
> 				}
> 			}
> 			continue;
> 		}
> 
> 		/* Adaptive mode; get the current CPU usage times. */
>-		if (read_usage_times(&load, nonice)) {
>+		if (read_usage_times(&load, nonice, noidle)) {
> 			if (vflag)
> 				warn("read_usage_times() failed");
> 			continue;
> 		}
> 
> 		if (mode == MODE_ADAPTIVE) {
> 			if (load > cpu_running_mark) {
>--- usr.sbin/powerd/powerd.8
>+++ usr.sbin/powerd/powerd.8
>@@ -18,133 +18,147 @@
> .\" DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
> .\" OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
> .\" HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
> .\" LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
> .\" OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
> .\" SUCH DAMAGE.
> .\"
>-.\" $FreeBSD: stable/12/usr.sbin/powerd/powerd.8 358990 2020-03-14 06:57:50Z cperciva $
>+.\" $FreeBSD: head/usr.sbin/powerd/powerd.8 358574 2020-03-03 13:25:08Z 0mp $
> .\"
>-.Dd January 13, 2019
>+.Dd Jan 27, 2021
> .Dt POWERD 8
> .Os
> .Sh NAME
> .Nm powerd
> .Nd "system power control utility"
> .Sh SYNOPSIS
> .Nm
> .Op Fl a Ar mode
> .Op Fl b Ar mode
>-.Op Fl i Ar percent
>-.Op Fl m Ar freq
>-.Op Fl M Ar freq
>-.Op Fl N
> .Op Fl n Ar mode
>-.Op Fl p Ar ival
>-.Op Fl P Ar pidfile
>+.Op Fl M Ar freq
>+.Op Fl m Ar freq
>+.Op Fl N Ar percent
>+.Op Fl I Ar percent
>+.Op Fl i Ar percent
> .Op Fl r Ar percent
>+.Op Fl p Ar ival
> .Op Fl s Ar source
>+.Op Fl P Ar pidfile
> .Op Fl v
> .Sh DESCRIPTION
> The
> .Nm
> utility monitors the system state and sets various power control options
> accordingly.
> It offers power-saving modes that can be
> individually selected for operation on AC power or batteries.
>-.Bl -tag -width ".Ar hiadaptive"
>-.It Ar maximum
>+.Bl -tag -width "hiadaptive"
>+.It Cm maximum
> Choose the highest performance values.
> May be abbreviated as
>-.Ar max .
>-.It Ar minimum
>+.Cm max .
>+.It Cm minimum
> Choose the lowest performance values to get the most power savings.
> May be abbreviated as
>-.Ar min .
>-.It Ar adaptive
>+.Cm min .
>+.It Cm adaptive
> Attempt to strike a balance by degrading performance when the system
> appears idle and increasing it when the system is busy.
> It offers a good balance between a small performance loss for greatly
> increased power savings.
> May be abbreviated as
>-.Ar adp .
>-.It Ar hiadaptive
>+.Cm adp .
>+.It Cm hiadaptive
> Like
>-.Ar adaptive
>+.Cm adaptive
> mode, but tuned for systems where performance and interactivity are
> more important than power consumption.
> It increases frequency faster, reduces frequency less aggressively, and
> will maintain full frequency for longer.
> May be abbreviated as
>-.Ar hadp .
>+.Cm hadp .
> .El
> .Pp
> The default mode is
>-.Ar adaptive
>+.Cm adaptive
> for battery power and
>-.Ar hiadaptive
>+.Cm hiadaptive
> for the rest.
> .Pp
> .Nm
> recognizes these runtime options:
>-.Bl -tag -width ".Fl r Ar percent"
>+.Bl -tag -width "-r percent"
> .It Fl a Ar mode
> Selects the
> .Ar mode
> to use while on AC power.
> .It Fl b Ar mode
> Selects the
> .Ar mode
> to use while on battery power.
>-.It Fl i Ar percent
>-Specifies the CPU load percent level when adaptive
>-mode should begin to degrade performance to save power.
>-The default is 50% or lower.
>-.It Fl m Ar freq
>-Specifies the minimum frequency to throttle down to.
>-.It Fl M Ar freq
>-Specifies the maximum frequency to throttle up to.
>-.It Fl N
>-Treat "nice" time as idle for the purpose of load calculation;
>-i.e. do not increase the CPU frequency if the CPU is only busy
>-with "nice" processes.
> .It Fl n Ar mode
> Selects the
> .Ar mode
> to use normally when the AC line state is unknown.
>-.It Fl p Ar ival
>-Specifies a different polling interval (in milliseconds) for AC line state
>-and system idle levels.
>-The default is 250 ms.
>-.It Fl P Ar pidfile
>-Specifies an alternative file in which the process ID should be stored.
>-The default is
>-.Pa /var/run/powerd.pid .
>+.It Fl M Ar freq
>+Specifies the maximum frequency to throttle up to.
>+.It Fl m Ar freq
>+Specifies the minimum frequency to throttle down to.
>+.It Fl N Ar percent
>+Weight "nice" CPU time by the given scale.
>+.It Fl I Ar percent
>+Weight "user idle" CPU time by the given scale.
>+.Pp
>+In both cases, 0 counts like "system idle" for the purpose of
>+.Nm
>+\'s internal load calculation; i.e., do not increase the CPU frequency
>+if the CPU is only busy with "nice" or "user idle" processes.
>+However, the system and interrupt load caused by these processes
>+can not be scaled out.
>+The implied default is 100, to fully count "nice" and "user idle" load.
>+.It Fl i Ar percent
>+Specifies the CPU load percent level when adaptive
>+mode should begin to degrade performance to save power.
>+The default is 50% or lower.
> .It Fl r Ar percent
> Specifies the CPU load percent level where adaptive
> mode should consider the CPU running and increase performance.
> The default is 75% or higher.
>+.It Fl p Ar ival
>+Specifies a different polling interval (in milliseconds) for AC line state
>+and system idle levels.
>+The default is 250 ms.
> .It Fl s Ar source
> Enforces method for AC line state refresh; by default, it is chosen
> automatically.
> The set of valid methods is
> .Cm sysctl , devd
> and
> .Cm apm
> (i386 only).
>+.It Fl P Ar pidfile
>+Specifies an alternative file in which the process ID should be stored.
> .It Fl v
> Verbose mode.
> Messages about power changes will be printed to stdout and
> .Nm
> will operate in the foreground.
> .El
>+.Sh FILES
>+.Bl -tag -width "/var/run/powerd.pid"
>+.It Pa /var/run/powerd.pid
>+The default PID file.
>+.El
> .Sh SEE ALSO
> .Xr acpi 4 ,
> .Xr apm 4 ,
>-.Xr cpufreq 4
>+.Xr cpufreq 4 ,
>+.Xr nice 1 ,
>+.Xr idprio 1 .
> .Sh HISTORY
> The
> .Nm
> utility first appeared in
> .Fx 6.0 .
> .Sh AUTHORS
> .An -nosplit

Actions: View | Diff

Attachments on bug 246940: 215172 | 215173 | 215241 | 215284 | 215285 | 221995 | 221997 | 221998 | 221999