summaryrefslogtreecommitdiff
path: root/tools/power/cpupower/utils/idle_monitor/mperf_monitor.c
diff options
context:
space:
mode:
Diffstat (limited to 'tools/power/cpupower/utils/idle_monitor/mperf_monitor.c')
-rw-r--r--tools/power/cpupower/utils/idle_monitor/mperf_monitor.c338
1 files changed, 338 insertions, 0 deletions
diff --git a/tools/power/cpupower/utils/idle_monitor/mperf_monitor.c b/tools/power/cpupower/utils/idle_monitor/mperf_monitor.c
new file mode 100644
index 000000000..90a8c4f07
--- /dev/null
+++ b/tools/power/cpupower/utils/idle_monitor/mperf_monitor.c
@@ -0,0 +1,338 @@
+/*
+ * (C) 2010,2011 Thomas Renninger <trenn@suse.de>, Novell Inc.
+ *
+ * Licensed under the terms of the GNU GPL License version 2.
+ */
+
+#if defined(__i386__) || defined(__x86_64__)
+
+#include <stdio.h>
+#include <stdint.h>
+#include <stdlib.h>
+#include <string.h>
+#include <limits.h>
+
+#include <cpufreq.h>
+
+#include "helpers/helpers.h"
+#include "idle_monitor/cpupower-monitor.h"
+
+#define MSR_APERF 0xE8
+#define MSR_MPERF 0xE7
+
+#define MSR_TSC 0x10
+
+#define MSR_AMD_HWCR 0xc0010015
+
+enum mperf_id { C0 = 0, Cx, AVG_FREQ, MPERF_CSTATE_COUNT };
+
+static int mperf_get_count_percent(unsigned int self_id, double *percent,
+ unsigned int cpu);
+static int mperf_get_count_freq(unsigned int id, unsigned long long *count,
+ unsigned int cpu);
+static struct timespec time_start, time_end;
+
+static cstate_t mperf_cstates[MPERF_CSTATE_COUNT] = {
+ {
+ .name = "C0",
+ .desc = N_("Processor Core not idle"),
+ .id = C0,
+ .range = RANGE_THREAD,
+ .get_count_percent = mperf_get_count_percent,
+ },
+ {
+ .name = "Cx",
+ .desc = N_("Processor Core in an idle state"),
+ .id = Cx,
+ .range = RANGE_THREAD,
+ .get_count_percent = mperf_get_count_percent,
+ },
+
+ {
+ .name = "Freq",
+ .desc = N_("Average Frequency (including boost) in MHz"),
+ .id = AVG_FREQ,
+ .range = RANGE_THREAD,
+ .get_count = mperf_get_count_freq,
+ },
+};
+
+enum MAX_FREQ_MODE { MAX_FREQ_SYSFS, MAX_FREQ_TSC_REF };
+static int max_freq_mode;
+/*
+ * The max frequency mperf is ticking at (in C0), either retrieved via:
+ * 1) calculated after measurements if we know TSC ticks at mperf/P0 frequency
+ * 2) cpufreq /sys/devices/.../cpu0/cpufreq/cpuinfo_max_freq at init time
+ * 1. Is preferred as it also works without cpufreq subsystem (e.g. on Xen)
+ */
+static unsigned long max_frequency;
+
+static unsigned long long tsc_at_measure_start;
+static unsigned long long tsc_at_measure_end;
+static unsigned long long *mperf_previous_count;
+static unsigned long long *aperf_previous_count;
+static unsigned long long *mperf_current_count;
+static unsigned long long *aperf_current_count;
+
+/* valid flag for all CPUs. If a MSR read failed it will be zero */
+static int *is_valid;
+
+static int mperf_get_tsc(unsigned long long *tsc)
+{
+ int ret;
+ ret = read_msr(0, MSR_TSC, tsc);
+ if (ret)
+ dprint("Reading TSC MSR failed, returning %llu\n", *tsc);
+ return ret;
+}
+
+static int mperf_init_stats(unsigned int cpu)
+{
+ unsigned long long val;
+ int ret;
+
+ ret = read_msr(cpu, MSR_APERF, &val);
+ aperf_previous_count[cpu] = val;
+ ret |= read_msr(cpu, MSR_MPERF, &val);
+ mperf_previous_count[cpu] = val;
+ is_valid[cpu] = !ret;
+
+ return 0;
+}
+
+static int mperf_measure_stats(unsigned int cpu)
+{
+ unsigned long long val;
+ int ret;
+
+ ret = read_msr(cpu, MSR_APERF, &val);
+ aperf_current_count[cpu] = val;
+ ret |= read_msr(cpu, MSR_MPERF, &val);
+ mperf_current_count[cpu] = val;
+ is_valid[cpu] = !ret;
+
+ return 0;
+}
+
+static int mperf_get_count_percent(unsigned int id, double *percent,
+ unsigned int cpu)
+{
+ unsigned long long aperf_diff, mperf_diff, tsc_diff;
+ unsigned long long timediff;
+
+ if (!is_valid[cpu])
+ return -1;
+
+ if (id != C0 && id != Cx)
+ return -1;
+
+ mperf_diff = mperf_current_count[cpu] - mperf_previous_count[cpu];
+ aperf_diff = aperf_current_count[cpu] - aperf_previous_count[cpu];
+
+ if (max_freq_mode == MAX_FREQ_TSC_REF) {
+ tsc_diff = tsc_at_measure_end - tsc_at_measure_start;
+ *percent = 100.0 * mperf_diff / tsc_diff;
+ dprint("%s: TSC Ref - mperf_diff: %llu, tsc_diff: %llu\n",
+ mperf_cstates[id].name, mperf_diff, tsc_diff);
+ } else if (max_freq_mode == MAX_FREQ_SYSFS) {
+ timediff = timespec_diff_us(time_start, time_end);
+ *percent = 100.0 * mperf_diff / timediff;
+ dprint("%s: MAXFREQ - mperf_diff: %llu, time_diff: %llu\n",
+ mperf_cstates[id].name, mperf_diff, timediff);
+ } else
+ return -1;
+
+ if (id == Cx)
+ *percent = 100.0 - *percent;
+
+ dprint("%s: previous: %llu - current: %llu - (%u)\n",
+ mperf_cstates[id].name, mperf_diff, aperf_diff, cpu);
+ dprint("%s: %f\n", mperf_cstates[id].name, *percent);
+ return 0;
+}
+
+static int mperf_get_count_freq(unsigned int id, unsigned long long *count,
+ unsigned int cpu)
+{
+ unsigned long long aperf_diff, mperf_diff, time_diff, tsc_diff;
+
+ if (id != AVG_FREQ)
+ return 1;
+
+ if (!is_valid[cpu])
+ return -1;
+
+ mperf_diff = mperf_current_count[cpu] - mperf_previous_count[cpu];
+ aperf_diff = aperf_current_count[cpu] - aperf_previous_count[cpu];
+
+ if (max_freq_mode == MAX_FREQ_TSC_REF) {
+ /* Calculate max_freq from TSC count */
+ tsc_diff = tsc_at_measure_end - tsc_at_measure_start;
+ time_diff = timespec_diff_us(time_start, time_end);
+ max_frequency = tsc_diff / time_diff;
+ }
+
+ *count = max_frequency * ((double)aperf_diff / mperf_diff);
+ dprint("%s: Average freq based on %s maximum frequency:\n",
+ mperf_cstates[id].name,
+ (max_freq_mode == MAX_FREQ_TSC_REF) ? "TSC calculated" : "sysfs read");
+ dprint("%max_frequency: %lu", max_frequency);
+ dprint("aperf_diff: %llu\n", aperf_diff);
+ dprint("mperf_diff: %llu\n", mperf_diff);
+ dprint("avg freq: %llu\n", *count);
+ return 0;
+}
+
+static int mperf_start(void)
+{
+ int cpu;
+ unsigned long long dbg;
+
+ clock_gettime(CLOCK_REALTIME, &time_start);
+ mperf_get_tsc(&tsc_at_measure_start);
+
+ for (cpu = 0; cpu < cpu_count; cpu++)
+ mperf_init_stats(cpu);
+
+ mperf_get_tsc(&dbg);
+ dprint("TSC diff: %llu\n", dbg - tsc_at_measure_start);
+ return 0;
+}
+
+static int mperf_stop(void)
+{
+ unsigned long long dbg;
+ int cpu;
+
+ for (cpu = 0; cpu < cpu_count; cpu++)
+ mperf_measure_stats(cpu);
+
+ mperf_get_tsc(&tsc_at_measure_end);
+ clock_gettime(CLOCK_REALTIME, &time_end);
+
+ mperf_get_tsc(&dbg);
+ dprint("TSC diff: %llu\n", dbg - tsc_at_measure_end);
+
+ return 0;
+}
+
+/*
+ * Mperf register is defined to tick at P0 (maximum) frequency
+ *
+ * Instead of reading out P0 which can be tricky to read out from HW,
+ * we use TSC counter if it reliably ticks at P0/mperf frequency.
+ *
+ * Still try to fall back to:
+ * /sys/devices/system/cpu/cpu0/cpufreq/cpuinfo_max_freq
+ * on older Intel HW without invariant TSC feature.
+ * Or on AMD machines where TSC does not tick at P0 (do not exist yet, but
+ * it's still double checked (MSR_AMD_HWCR)).
+ *
+ * On these machines the user would still get useful mperf
+ * stats when acpi-cpufreq driver is loaded.
+ */
+static int init_maxfreq_mode(void)
+{
+ int ret;
+ unsigned long long hwcr;
+ unsigned long min;
+
+ if (!(cpupower_cpu_info.caps & CPUPOWER_CAP_INV_TSC))
+ goto use_sysfs;
+
+ if (cpupower_cpu_info.vendor == X86_VENDOR_AMD) {
+ /* MSR_AMD_HWCR tells us whether TSC runs at P0/mperf
+ * freq.
+ * A test whether hwcr is accessable/available would be:
+ * (cpupower_cpu_info.family > 0x10 ||
+ * cpupower_cpu_info.family == 0x10 &&
+ * cpupower_cpu_info.model >= 0x2))
+ * This should be the case for all aperf/mperf
+ * capable AMD machines and is therefore safe to test here.
+ * Compare with Linus kernel git commit: acf01734b1747b1ec4
+ */
+ ret = read_msr(0, MSR_AMD_HWCR, &hwcr);
+ /*
+ * If the MSR read failed, assume a Xen system that did
+ * not explicitly provide access to it and assume TSC works
+ */
+ if (ret != 0) {
+ dprint("TSC read 0x%x failed - assume TSC working\n",
+ MSR_AMD_HWCR);
+ return 0;
+ } else if (1 & (hwcr >> 24)) {
+ max_freq_mode = MAX_FREQ_TSC_REF;
+ return 0;
+ } else { /* Use sysfs max frequency if available */ }
+ } else if (cpupower_cpu_info.vendor == X86_VENDOR_INTEL) {
+ /*
+ * On Intel we assume mperf (in C0) is ticking at same
+ * rate than TSC
+ */
+ max_freq_mode = MAX_FREQ_TSC_REF;
+ return 0;
+ }
+use_sysfs:
+ if (cpufreq_get_hardware_limits(0, &min, &max_frequency)) {
+ dprint("Cannot retrieve max freq from cpufreq kernel "
+ "subsystem\n");
+ return -1;
+ }
+ max_freq_mode = MAX_FREQ_SYSFS;
+ return 0;
+}
+
+/*
+ * This monitor provides:
+ *
+ * 1) Average frequency a CPU resided in
+ * This always works if the CPU has aperf/mperf capabilities
+ *
+ * 2) C0 and Cx (any sleep state) time a CPU resided in
+ * Works if mperf timer stops ticking in sleep states which
+ * seem to be the case on all current HW.
+ * Both is directly retrieved from HW registers and is independent
+ * from kernel statistics.
+ */
+struct cpuidle_monitor mperf_monitor;
+struct cpuidle_monitor *mperf_register(void)
+{
+ if (!(cpupower_cpu_info.caps & CPUPOWER_CAP_APERF))
+ return NULL;
+
+ if (init_maxfreq_mode())
+ return NULL;
+
+ /* Free this at program termination */
+ is_valid = calloc(cpu_count, sizeof(int));
+ mperf_previous_count = calloc(cpu_count, sizeof(unsigned long long));
+ aperf_previous_count = calloc(cpu_count, sizeof(unsigned long long));
+ mperf_current_count = calloc(cpu_count, sizeof(unsigned long long));
+ aperf_current_count = calloc(cpu_count, sizeof(unsigned long long));
+
+ mperf_monitor.name_len = strlen(mperf_monitor.name);
+ return &mperf_monitor;
+}
+
+void mperf_unregister(void)
+{
+ free(mperf_previous_count);
+ free(aperf_previous_count);
+ free(mperf_current_count);
+ free(aperf_current_count);
+ free(is_valid);
+}
+
+struct cpuidle_monitor mperf_monitor = {
+ .name = "Mperf",
+ .hw_states_num = MPERF_CSTATE_COUNT,
+ .hw_states = mperf_cstates,
+ .start = mperf_start,
+ .stop = mperf_stop,
+ .do_register = mperf_register,
+ .unregister = mperf_unregister,
+ .needs_root = 1,
+ .overflow_s = 922000000 /* 922337203 seconds TSC overflow
+ at 20GHz */
+};
+#endif /* #if defined(__i386__) || defined(__x86_64__) */