Introduce cpufreq-aperf

What does this program do?

On latest processors there exist two MSR registers referred to as:
     - MPERF increasing with maxium (P0) frequency in C0
     - APERF increasing with current/actual frequency in C0

From this information the average frequency over a time period can be
calculated and this is what this tool does.

A nice falloff feature beside the average frequency is the time
a processor core remained in C0 (working state) or any CX (sleep state)
processor sleep state during the measured time period. This information
can be determined from the fact that MPERF only increases in C0 state.

Note:	There were kernels which reset MPERF/APERF registers to 0.
	This got reverted by git commit
		18b2646fe3babeb40b34a0c1751e0bf5adfdc64c
	which was commited to 2.6.30-rcX mainline kernels
	For kernels where the kernel rests MPERF/APERF registers to 0,
	this tool will not work. It cannot be detected whether this happened.

[linux@dominikbrodowski.net: Makefile and .gitignore integration]
Signed-off-by: Thomas Renninger <trenn@suse.de>
Signed-off-by: Dominik Brodowski <linux@dominikbrodowski.net>

3 files changed, 471 insertions, 2 deletions

diff --git a/.gitignore b/.gitignore
index 9ec7cac..fc3454b 100644
--- a/.gitignore
+++ b/.gitignore
@@ -2,6 +2,7 @@
 build/ccdv
 cpufreq-info
 cpufreq-set
+cpufreq-aperf
 lib/.libs
 lib/cpufreq.lo
 lib/cpufreq.o
@@ -14,3 +15,4 @@ po/cpufrequtils.pot
 po/*.gmo
 utils/cpufreq-info.o
 utils/cpufreq-set.o
+utils/cpufreq-aperf.o
\ No newline at end of file
diff --git a/Makefile b/Makefile
index 8d769e6..ace985a 100644
--- a/Makefile
+++ b/Makefile
@@ -205,7 +205,7 @@ cpufreq-%: libcpufreq.la $(UTIL_OBJS)
 	$(QUIET) $(CC) $(CFLAGDEF) $(CFLAGS) -g -I./lib/ -L. -L./.libs/ -lcpufreq -o $@ utils/$@.o
 	$(QUIET) $(STRIPCMD) $@
 
-utils: cpufreq-info cpufreq-set
+utils: cpufreq-info cpufreq-set cpufreq-aperf
 
 po/$(PACKAGE).pot: $(UTIL_OBJS)
 	@xgettext --default-domain=$(PACKAGE) --add-comments \
@@ -234,7 +234,7 @@ clean:
 	 | xargs rm -f
 	-rm -rf lib/.libs
 	-rm -rf .libs
-	-rm -f cpufreq-info cpufreq-set
+	-rm -f cpufreq-info cpufreq-set cpufreq-aperf
 	-rm -f build/ccdv
 	-rm -rf po/*.gmo po/*.pot
 	make -C bench clean
@@ -250,6 +250,7 @@ install-tools:
 	$(INSTALL) -d $(DESTDIR)${bindir}
 	$(INSTALL_PROGRAM) cpufreq-set $(DESTDIR)${bindir}/cpufreq-set
 	$(INSTALL_PROGRAM) cpufreq-info $(DESTDIR)${bindir}/cpufreq-info
+	$(INSTALL_PROGRAM) cpufreq-aperf $(DESTDIR)${bindir}/cpufreq-aperf
 
 install-man:
 	$(INSTALL_DATA) -D man/cpufreq-set.1 $(DESTDIR)${mandir}/man1/cpufreq-set.1
@@ -273,6 +274,7 @@ uninstall:
 	- rm -f $(DESTDIR)${includedir}/cpufreq.h
 	- rm -f $(DESTDIR)${bindir}/cpufreq-set
 	- rm -f $(DESTDIR)${bindir}/cpufreq-info
+	- rm -f $(DESTDIR)${bindir}/cpufreq-aperf
 	- rm -f $(DESTDIR)${mandir}/man1/cpufreq-set.1
 	- rm -f $(DESTDIR)${mandir}/man1/cpufreq-info.1
 	- for HLANG in $(LANGUAGES); do \
diff --git a/utils/aperf.c b/utils/aperf.c
new file mode 100644
index 0000000..b3cd8ac
--- /dev/null
+++ b/utils/aperf.c
@@ -0,0 +1,465 @@
+/*
+ *  Copyright (C) 2009  Thomas Renninger <trenn@suse.de>, Novell Inc.
+ *
+ *  Inspired by these projects:
+ *    cpuid (by Todd Allen)
+ *    msr-tools (by H. Peter Anvin <hpa@zytor.com>)
+ *
+ *  Licensed under the terms of the GNU GPL License version 2.
+ *
+ *
+ *  What does this program do:
+ *
+ *  On latest processors exist two MSR registers refered to as:
+ *    - MPERF increasing with maxium (P0) frequency in C0
+ *    - APERF increasing with current/actual frequency in C0
+ *
+ *  From this information the average frequency over a time period can be
+ *  calculated and this is what this tool does.
+ *
+ *  A nice falloff feature beside the average frequency is the time
+ *  a processor core remained in C0 (working state) or any CX (sleep state)
+ *  processor sleep state during the measured time period. This information
+ *  can be determined from the fact that MPERF only increases in C0 state.
+ *
+ *  Note: There were kernels which reset MPERF/APERF registers to 0.
+ *        This got reverted by git commit
+ *                  18b2646fe3babeb40b34a0c1751e0bf5adfdc64c
+ *        which was commited to 2.6.30-rcX mainline kernels
+ *        For kernels where the kernel rests MPERF/APERF registers to 0,
+ *        this tool will not work. It cannot be detected whether this happened.
+ *
+ * Possible ToDos/Enhancments:
+ *
+ *  - Use cpuid assmbler command to retrieve cpuid
+ *    -No dependency to the cpuid driver
+ *  - Refresh the screen when mulitple cpus are poked and display results
+ *    on one screen
+ *    -This would introduce a lot more complexity, not sure whether it's
+ *       wanted/needed. I'd vote to better not do that.
+ *  - Manpage
+ *  - Translations
+ *  - ...
+ */
+
+
+#include <stdio.h>
+#include <stdlib.h>
+#include <unistd.h>
+#include <getopt.h>
+#include <sys/time.h>
+#include <stdint.h>
+#include <errno.h>
+#include <fcntl.h>
+#include <string.h>
+
+#include "cpufreq.h"
+
+#define MSR_IA32_APERF 0x000000E8
+#define MSR_IA32_MPERF 0x000000E7
+
+#define LINE_LEN 10
+
+struct avg_perf_cpu_info
+{
+	unsigned long max_freq;
+	uint64_t saved_aperf;
+	uint64_t saved_mperf;
+	uint32_t is_valid:1;
+};
+
+static unsigned int count_cpus(void)
+{
+	FILE *fp;
+	char value[LINE_LEN];
+	unsigned int ret = 0;
+	unsigned int cpunr = 0;
+
+	fp = fopen("/proc/stat", "r");
+	if(!fp) {
+		printf("Couldn't count the number of CPUs (%s: %s), "
+		       "assuming 1\n", "/proc/stat", strerror(errno));
+		return 1;
+	}
+
+	while (!feof(fp)) {
+		fgets(value, LINE_LEN, fp);
+		value[LINE_LEN - 1] = '\0';
+		if (strlen(value) < (LINE_LEN - 2))
+			continue;
+		if (strstr(value, "cpu "))
+			continue;
+		if (sscanf(value, "cpu%d ", &cpunr) != 1)
+			continue;
+		if (cpunr > ret)
+			ret = cpunr;
+	}
+	fclose(fp);
+
+	/* cpu count starts from 0, on error return 1 (UP) */
+	return (ret+1);
+}
+
+static int has_mperf_aperf_support(int cpu)
+{
+	int fd;
+	char msr_file_name[64];
+	uint32_t words[4];
+	int support;
+
+	sprintf(msr_file_name, "/dev/cpu/%d/cpuid", cpu);
+	fd = open(msr_file_name, O_RDONLY);
+	if (fd < 0)
+		return -1;
+	if (lseek(fd, 6, SEEK_CUR) == -1)
+		goto err;
+	if (read(fd, words, 16) != 16)
+		goto err;
+	support = words[2] & 0x1;
+	close(fd);
+	return support;
+ err:
+	close(fd);
+	return -1;
+}
+
+/*
+ * read_msr
+ *
+ * Will return 0 on success and -1 on failure.
+ * Possible errno values could be:
+ * EFAULT -If the read/write did not fully complete
+ * EIO    -If the CPU does not support MSRs
+ * ENXIO  -If the CPU does not exist
+ */
+
+static int read_msr(int cpu, unsigned int idx, unsigned long long *val)
+{
+	int fd;
+	char msr_file_name[64];
+
+	sprintf(msr_file_name, "/dev/cpu/%d/msr", cpu);
+	fd = open(msr_file_name, O_RDONLY);
+	if (fd < 0)
+		return -1;
+	if (lseek(fd, idx, SEEK_CUR) == -1)
+		goto err;
+	if (read(fd, val, sizeof val) != sizeof *val)
+		goto err;
+	close(fd);
+	return 0;
+ err:
+	close(fd);
+	return -1;
+}
+
+/*
+ * get_aperf_mperf()
+ *
+ * Returns the current aperf/mperf MSR values of cpu
+ */
+static int get_aperf_mperf(unsigned int cpu, uint64_t *aperf, uint64_t *mperf)
+{
+	int retval;
+
+	retval = read_msr(cpu, MSR_IA32_APERF, (unsigned long long*)aperf);
+	if (retval < 0)
+		return retval;
+	
+	retval = read_msr(cpu, MSR_IA32_MPERF, (unsigned long long*)mperf);
+	if (retval < 0)
+		return retval;
+	return 0;
+}
+
+/*
+ * get_average_perf()
+ *
+ * Returns the average performance (also considers boosted frequencies)
+ * 
+ * Input:
+ *   aperf_diff: Difference of the aperf register over a time period
+ *   mperf_diff: Difference of the mperf register over the same time period
+ *   max_freq:   Maximum frequency (P0)
+ *
+ * Returns:
+ *   Average performance over the time period
+ */
+static unsigned long get_average_perf(unsigned long max_freq,
+				      uint64_t aperf_diff,
+				      uint64_t mperf_diff)
+{
+	unsigned int perf_percent = 0;
+	if (((unsigned long)(-1) / 100) < aperf_diff) {
+		int shift_count = 7;
+		aperf_diff >>= shift_count;
+		mperf_diff >>= shift_count;
+	}
+	perf_percent = (aperf_diff * 100) / mperf_diff;
+	return (max_freq * perf_percent) / 100;
+}
+
+/*
+ * get_C_state_time()
+ *
+ * Calculates the time the processor was in C0 and Cx processor sleep states
+ *
+ * As mperf does only tick in C0 at maximum frequency, this is a nice "falloff"
+ * functionality and more accurate than powertop or other kernel timer based
+ * C-state measurings (and can be used to verify whether they are correct.
+ *
+ * Input:
+ *   time_diff:  The time passed for which the mperf_diff was calulcated on
+ *   mperf_diff: The value the mperf register increased during time_diff
+ *   max_freq:   Maximum frequency of the processor (P0) in kHz
+ *
+ * Output:
+ *   C0_time:    The time the processor was in C0
+ *   CX_time:    The time the processor was in CX
+ *   percent:    Percentage the processor stayed in C0
+ */
+static int get_C_state_time(struct timeval time_diff, uint64_t mperf_diff,
+		     unsigned long max_freq,
+		     struct timeval *C0_time, struct timeval *CX_time,
+		     unsigned int *percent)
+{
+	unsigned long long overall_msecs, expected_ticks, c0_time, cx_time;
+
+	overall_msecs = (time_diff.tv_sec * 1000 * 1000  + time_diff.tv_usec)
+		/ 1000;
+
+	expected_ticks = max_freq * overall_msecs;
+	*percent = (mperf_diff * 100) / expected_ticks;
+
+	cx_time = (expected_ticks - mperf_diff) / max_freq;
+	c0_time = mperf_diff / max_freq;
+
+	CX_time->tv_sec  = cx_time / 1000;
+	CX_time->tv_usec = cx_time % 1000;
+	C0_time->tv_sec  = c0_time / 1000;
+	C0_time->tv_usec = c0_time % 1000;
+	return 0;
+}
+
+static int get_measure_start_info(unsigned int cpu,
+				  struct avg_perf_cpu_info *cpu_info)
+{
+	unsigned long min, max;
+	uint64_t aperf, mperf;
+	int ret;
+
+	cpu_info->is_valid = 0;
+		
+	ret = has_mperf_aperf_support(cpu);
+	if (ret < 0) {
+		fprintf(stderr, "Could not read cpuid, is the cpuid "
+			"driver loaded or compiled into the kernel?\n");
+		return ret;
+	} else if (ret == 0) {
+		fprintf(stderr, "CPU(s) do not support mperf/aperf MSR "
+			"registers\n");
+		return EXIT_FAILURE;
+	}
+	
+	if (cpufreq_get_hardware_limits(cpu, &min, &max)) {
+		fprintf(stderr, "Could not get max frequency (P0), a "
+			"cpufreq driver must be loaded?\n");
+		return EXIT_FAILURE;
+	} else
+		cpu_info->max_freq = max;
+	
+	ret = get_aperf_mperf(cpu, &aperf, &mperf);
+	if (ret < 0) {
+		fprintf(stderr, "Could not read MSRs, is the msr driver loaded"
+			" or compiled into the kernel?\n");
+		return EXIT_FAILURE;
+	} else {
+		cpu_info->saved_aperf = aperf;
+		cpu_info->saved_mperf = mperf;
+	}
+	cpu_info->is_valid = 1;
+	return 0;
+}
+
+static void print_cpu_stats(unsigned int cpu, unsigned long average,
+			    struct timeval c0_time, struct timeval cx_time,
+			    unsigned int c0_percent)
+{
+	printf("%.3u\t", cpu);
+	printf("%.7lu\t\t\t", average);
+	printf("%.2lu sec %.3lu ms\t", c0_time.tv_sec, c0_time.tv_usec);
+	printf("%.2lu sec %.3lu ms\t", cx_time.tv_sec, cx_time.tv_usec);
+	printf("%.2u", c0_percent);
+}
+
+
+static int do_measuring_on_cpu(int sleep_time, int once, int cpu)
+{
+	int ret;
+	unsigned long average;
+	unsigned int c0_percent;
+	struct timeval start_time, current_time, diff_time, C0_time, CX_time;
+	uint64_t current_aperf, current_mperf, mperf_diff, aperf_diff;
+	struct avg_perf_cpu_info cpu_info;
+
+	ret = get_measure_start_info(cpu, &cpu_info);
+	if (ret)
+		return ret;
+
+	while(1) {
+		gettimeofday(&start_time, NULL);
+		sleep(sleep_time);
+		/* ToDo: just add a second on the timeval struct? */
+		gettimeofday(&current_time, NULL);
+		timersub(&current_time, &start_time, &diff_time);
+		memcpy(&start_time, &current_time,
+		       sizeof(struct timeval));
+
+		if (!cpu_info.is_valid)
+			continue;
+		
+		ret = get_aperf_mperf(cpu, &current_aperf, &current_mperf);
+		if (ret < 0)
+			return EXIT_FAILURE;
+		mperf_diff = current_mperf - cpu_info.saved_mperf;
+		aperf_diff = current_aperf - cpu_info.saved_aperf;
+		
+		get_C_state_time(diff_time, mperf_diff,
+				 cpu_info.max_freq,
+				 &C0_time, &CX_time,
+				 &c0_percent);
+		average = get_average_perf(cpu_info.max_freq,
+					   aperf_diff, mperf_diff);
+		cpu_info.saved_mperf = current_mperf;
+		cpu_info.saved_aperf = current_aperf;
+		print_cpu_stats(cpu, average, C0_time, CX_time,
+				c0_percent);
+
+		if (once) {
+			printf("\n");
+			break;
+		} else {
+			printf("\r");
+			fflush(stdout);
+		}
+	}
+	return 0;
+}
+
+static int do_measure_all_cpus(int sleep_time, int once)
+{
+	int ret;
+	unsigned long average;
+	unsigned int c0_percent, cpus, cpu;
+	struct timeval start_time, current_time, diff_time, C0_time, CX_time;
+	uint64_t current_aperf, current_mperf, mperf_diff, aperf_diff;
+	struct avg_perf_cpu_info *cpu_list;
+
+	cpus = count_cpus();
+
+	cpu_list = (struct avg_perf_cpu_info*)
+		malloc(cpus * sizeof (struct avg_perf_cpu_info));
+	
+	for (cpu = 0; cpu < cpus; cpu++) {
+		ret = get_measure_start_info(cpu, &cpu_list[cpu]);
+		if (ret)
+			return ret;
+	}	
+
+	while(1) {
+		gettimeofday(&start_time, NULL);
+		sleep(sleep_time);
+		/* ToDo: Just add a second on the timeval struct?
+		         Would save one gettimeofday, but would not
+			 be that accurate anymore
+		*/
+		gettimeofday(&current_time, NULL);
+		timersub(&current_time, &start_time, &diff_time);
+		memcpy(&start_time, &current_time,
+		       sizeof(struct timeval));
+
+		for (cpu = 0; cpu < cpus; cpu++) {
+			if (!cpu_list[cpu].is_valid)
+				continue;
+
+			ret = get_aperf_mperf(cpu, &current_aperf,
+					      &current_mperf);
+			if (ret < 0)
+				return EXIT_FAILURE;
+			mperf_diff = current_mperf - cpu_list[cpu].saved_mperf;
+			aperf_diff = current_aperf - cpu_list[cpu].saved_aperf;
+
+			get_C_state_time(diff_time, mperf_diff,
+					 cpu_list[cpu].max_freq,
+					 &C0_time, &CX_time,
+					 &c0_percent);
+			average = get_average_perf(cpu_list[cpu].max_freq,
+						   aperf_diff, mperf_diff);
+			cpu_list[cpu].saved_mperf = current_mperf;
+			cpu_list[cpu].saved_aperf = current_aperf;
+			print_cpu_stats(cpu, average, C0_time, CX_time,
+					c0_percent);
+			printf("\n");
+		}
+		if (once)
+			break;
+		printf("\n");
+	}
+	return 0;
+}
+
+
+/******* Options parsing, main ********/
+
+static struct option long_options[] = {
+  { "help",		0, 0, 'h' },
+  { "intervall",	1, 0, 'i' },
+  { "cpu",		1, 0, 'c' },
+  { "once",		0, 0, 'o' },
+  { 0, 0, 0, 0 }
+};
+
+static void usage(void) {
+	printf("cpufreq-aperf [OPTIONS]\n\n"
+	       "-c [ --cpu ] CPU               "
+	       "The CPU core to measure - default all cores\n"
+	       "-i [ --intervall ] seconds     "
+	       "Refresh rate - default 1 second\n"
+	       "-o [ --once ]                  "
+	       "Exit after one intervall\n"
+	       "-h [ --help ]                  "
+	       "This help text\n"
+	       "The msr driver must be loaded for this command to work\n");
+}
+
+int main(int argc, char *argv[])
+{
+	int c, ret, cpu = -1;
+	int sleep_time = 1, once = 0;
+	while ( (c = getopt_long(argc,argv,"c:ohi:",long_options,
+				 NULL)) != -1 ) {
+		switch ( c ) {
+		case 'o':
+			once = 1;
+			break;
+		case 'c':
+			cpu = atoi(optarg);
+			break;
+		case 'h':
+			usage();
+			exit(0);
+		case 'i':
+			sleep_time = atoi(optarg);
+			break;
+		}
+	}
+	printf("CPU\tAverage freq(KHz)\tTime in C0\tTime in"
+	       " Cx\tC0 percentage\n");
+
+	if (cpu == -1)
+		ret = do_measure_all_cpus(sleep_time, once);
+	else
+		ret = do_measuring_on_cpu(sleep_time, once, cpu);
+	return ret;		
+}
+/******* Options parsing, main ********/