#include <asm/unistd.h>
#include <stdbool.h>
#include <assert.h>
+#include <linux/kernel.h>
#define UNUSED(x) (void)(x)
enum counter_type { COUNTER_ITEMS, COUNTER_CYCLES, COUNTER_SECONDS, COUNTER_USEC };
enum counter_format { FORMAT_RAW, FORMAT_DELTA, FORMAT_PERCENT };
enum amperf_source { AMPERF_SOURCE_PERF, AMPERF_SOURCE_MSR };
+enum rapl_source { RAPL_SOURCE_NONE, RAPL_SOURCE_PERF, RAPL_SOURCE_MSR };
struct msr_counter {
unsigned int msr_num;
#define MAX_ADDED_THREAD_COUNTERS 24
#define BITMASK_SIZE 32
+/* Indexes used to map data read from perf and MSRs into global variables */
+enum rapl_rci_index {
+ RAPL_RCI_INDEX_ENERGY_PKG = 0,
+ RAPL_RCI_INDEX_ENERGY_CORES = 1,
+ RAPL_RCI_INDEX_DRAM = 2,
+ RAPL_RCI_INDEX_GFX = 3,
+ RAPL_RCI_INDEX_PKG_PERF_STATUS = 4,
+ RAPL_RCI_INDEX_DRAM_PERF_STATUS = 5,
+ RAPL_RCI_INDEX_CORE_ENERGY = 6,
+ NUM_RAPL_COUNTERS,
+};
+
+enum rapl_unit {
+ RAPL_UNIT_INVALID,
+ RAPL_UNIT_JOULES,
+ RAPL_UNIT_WATTS,
+};
+
+struct rapl_counter_info_t {
+ unsigned long long data[NUM_RAPL_COUNTERS];
+ enum rapl_source source[NUM_RAPL_COUNTERS];
+ unsigned long long flags[NUM_RAPL_COUNTERS];
+ double scale[NUM_RAPL_COUNTERS];
+ enum rapl_unit unit[NUM_RAPL_COUNTERS];
+
+ union {
+ /* Active when source == RAPL_SOURCE_MSR */
+ struct {
+ unsigned long long msr[NUM_RAPL_COUNTERS];
+ unsigned long long msr_mask[NUM_RAPL_COUNTERS];
+ int msr_shift[NUM_RAPL_COUNTERS];
+ };
+ };
+
+ int fd_perf;
+};
+
+/* struct rapl_counter_info_t for each RAPL domain */
+struct rapl_counter_info_t *rapl_counter_info_perdomain;
+
+#define RAPL_COUNTER_FLAG_USE_MSR_SUM (1u << 1)
+
+struct rapl_counter_arch_info {
+ int feature_mask; /* Mask for testing if the counter is supported on host */
+ const char *perf_subsys;
+ const char *perf_name;
+ unsigned long long msr;
+ unsigned long long msr_mask;
+ int msr_shift; /* Positive mean shift right, negative mean shift left */
+ double *platform_rapl_msr_scale; /* Scale applied to values read by MSR (platform dependent, filled at runtime) */
+ unsigned int rci_index; /* Maps data from perf counters to global variables */
+ unsigned long long bic;
+ double compat_scale; /* Some counters require constant scaling to be in the same range as other, similar ones */
+ unsigned long long flags;
+};
+
+static const struct rapl_counter_arch_info rapl_counter_arch_infos[] = {
+ {
+ .feature_mask = RAPL_PKG,
+ .perf_subsys = "power",
+ .perf_name = "energy-pkg",
+ .msr = MSR_PKG_ENERGY_STATUS,
+ .msr_mask = 0xFFFFFFFFFFFFFFFF,
+ .msr_shift = 0,
+ .platform_rapl_msr_scale = &rapl_energy_units,
+ .rci_index = RAPL_RCI_INDEX_ENERGY_PKG,
+ .bic = BIC_PkgWatt | BIC_Pkg_J,
+ .compat_scale = 1.0,
+ .flags = RAPL_COUNTER_FLAG_USE_MSR_SUM,
+ },
+ {
+ .feature_mask = RAPL_AMD_F17H,
+ .perf_subsys = "power",
+ .perf_name = "energy-pkg",
+ .msr = MSR_PKG_ENERGY_STAT,
+ .msr_mask = 0xFFFFFFFFFFFFFFFF,
+ .msr_shift = 0,
+ .platform_rapl_msr_scale = &rapl_energy_units,
+ .rci_index = RAPL_RCI_INDEX_ENERGY_PKG,
+ .bic = BIC_PkgWatt | BIC_Pkg_J,
+ .compat_scale = 1.0,
+ .flags = RAPL_COUNTER_FLAG_USE_MSR_SUM,
+ },
+ {
+ .feature_mask = RAPL_CORE_ENERGY_STATUS,
+ .perf_subsys = "power",
+ .perf_name = "energy-cores",
+ .msr = MSR_PP0_ENERGY_STATUS,
+ .msr_mask = 0xFFFFFFFFFFFFFFFF,
+ .msr_shift = 0,
+ .platform_rapl_msr_scale = &rapl_energy_units,
+ .rci_index = RAPL_RCI_INDEX_ENERGY_CORES,
+ .bic = BIC_CorWatt | BIC_Cor_J,
+ .compat_scale = 1.0,
+ .flags = RAPL_COUNTER_FLAG_USE_MSR_SUM,
+ },
+ {
+ .feature_mask = RAPL_DRAM,
+ .perf_subsys = "power",
+ .perf_name = "energy-ram",
+ .msr = MSR_DRAM_ENERGY_STATUS,
+ .msr_mask = 0xFFFFFFFFFFFFFFFF,
+ .msr_shift = 0,
+ .platform_rapl_msr_scale = &rapl_dram_energy_units,
+ .rci_index = RAPL_RCI_INDEX_DRAM,
+ .bic = BIC_RAMWatt | BIC_RAM_J,
+ .compat_scale = 1.0,
+ .flags = RAPL_COUNTER_FLAG_USE_MSR_SUM,
+ },
+ {
+ .feature_mask = RAPL_GFX,
+ .perf_subsys = "power",
+ .perf_name = "energy-gpu",
+ .msr = MSR_PP1_ENERGY_STATUS,
+ .msr_mask = 0xFFFFFFFFFFFFFFFF,
+ .msr_shift = 0,
+ .platform_rapl_msr_scale = &rapl_energy_units,
+ .rci_index = RAPL_RCI_INDEX_GFX,
+ .bic = BIC_GFXWatt | BIC_GFX_J,
+ .compat_scale = 1.0,
+ .flags = RAPL_COUNTER_FLAG_USE_MSR_SUM,
+ },
+ {
+ .feature_mask = RAPL_PKG_PERF_STATUS,
+ .perf_subsys = NULL,
+ .perf_name = NULL,
+ .msr = MSR_PKG_PERF_STATUS,
+ .msr_mask = 0xFFFFFFFFFFFFFFFF,
+ .msr_shift = 0,
+ .platform_rapl_msr_scale = &rapl_time_units,
+ .rci_index = RAPL_RCI_INDEX_PKG_PERF_STATUS,
+ .bic = BIC_PKG__,
+ .compat_scale = 100.0,
+ .flags = RAPL_COUNTER_FLAG_USE_MSR_SUM,
+ },
+ {
+ .feature_mask = RAPL_DRAM_PERF_STATUS,
+ .perf_subsys = NULL,
+ .perf_name = NULL,
+ .msr = MSR_DRAM_PERF_STATUS,
+ .msr_mask = 0xFFFFFFFFFFFFFFFF,
+ .msr_shift = 0,
+ .platform_rapl_msr_scale = &rapl_time_units,
+ .rci_index = RAPL_RCI_INDEX_DRAM_PERF_STATUS,
+ .bic = BIC_RAM__,
+ .compat_scale = 100.0,
+ .flags = RAPL_COUNTER_FLAG_USE_MSR_SUM,
+ },
+ {
+ .feature_mask = RAPL_AMD_F17H,
+ .perf_subsys = NULL,
+ .perf_name = NULL,
+ .msr = MSR_CORE_ENERGY_STAT,
+ .msr_mask = 0xFFFFFFFF,
+ .msr_shift = 0,
+ .platform_rapl_msr_scale = &rapl_energy_units,
+ .rci_index = RAPL_RCI_INDEX_CORE_ENERGY,
+ .bic = BIC_CorWatt | BIC_Cor_J,
+ .compat_scale = 1.0,
+ .flags = 0,
+ },
+};
+
+struct rapl_counter {
+ unsigned long long raw_value;
+ enum rapl_unit unit;
+ double scale;
+};
+
struct thread_data {
struct timeval tv_begin;
struct timeval tv_end;
unsigned long long c7;
unsigned long long mc6_us; /* duplicate as per-core for now, even though per module */
unsigned int core_temp_c;
- unsigned int core_energy; /* MSR_CORE_ENERGY_STAT */
+ struct rapl_counter core_energy; /* MSR_CORE_ENERGY_STAT */
unsigned int core_id;
unsigned long long core_throt_cnt;
unsigned long long counter[MAX_ADDED_COUNTERS];
unsigned int gfx_mhz;
unsigned int gfx_act_mhz;
unsigned int package_id;
- unsigned long long energy_pkg; /* MSR_PKG_ENERGY_STATUS */
- unsigned long long energy_dram; /* MSR_DRAM_ENERGY_STATUS */
- unsigned long long energy_cores; /* MSR_PP0_ENERGY_STATUS */
- unsigned long long energy_gfx; /* MSR_PP1_ENERGY_STATUS */
- unsigned long long rapl_pkg_perf_status; /* MSR_PKG_PERF_STATUS */
- unsigned long long rapl_dram_perf_status; /* MSR_DRAM_PERF_STATUS */
+ struct rapl_counter energy_pkg; /* MSR_PKG_ENERGY_STATUS */
+ struct rapl_counter energy_dram; /* MSR_DRAM_ENERGY_STATUS */
+ struct rapl_counter energy_cores; /* MSR_PP0_ENERGY_STATUS */
+ struct rapl_counter energy_gfx; /* MSR_PP1_ENERGY_STATUS */
+ struct rapl_counter rapl_pkg_perf_status; /* MSR_PKG_PERF_STATUS */
+ struct rapl_counter rapl_dram_perf_status; /* MSR_DRAM_PERF_STATUS */
unsigned int pkg_temp_c;
unsigned int uncore_mhz;
unsigned long long counter[MAX_ADDED_COUNTERS];
return 0;
}
+int probe_msr(int cpu, off_t offset)
+{
+ ssize_t retval;
+ unsigned long long dummy;
+
+ assert(!no_msr);
+
+ retval = pread(get_msr_fd(cpu), &dummy, sizeof(dummy), offset);
+
+ if (retval != sizeof(dummy))
+ return 1;
+
+ return 0;
+}
+
#define MAX_DEFERRED 16
char *deferred_add_names[MAX_DEFERRED];
char *deferred_skip_names[MAX_DEFERRED];
outp += sprintf(outp, "c7: %016llX\n", c->c7);
outp += sprintf(outp, "DTS: %dC\n", c->core_temp_c);
outp += sprintf(outp, "cpu_throt_count: %016llX\n", c->core_throt_cnt);
- outp += sprintf(outp, "Joules: %0X\n", c->core_energy);
+
+ const unsigned long long energy_value = c->core_energy.raw_value * c->core_energy.scale;
+ const double energy_scale = c->core_energy.scale;
+ if (c->core_energy.unit == RAPL_UNIT_JOULES)
+ outp += sprintf(outp, "Joules: %0llX (scale: %lf)\n", energy_value, energy_scale);
for (i = 0, mp = sys.cp; mp; i++, mp = mp->next) {
outp +=
outp += sprintf(outp, "pc10: %016llX\n", p->pc10);
outp += sprintf(outp, "cpu_lpi: %016llX\n", p->cpu_lpi);
outp += sprintf(outp, "sys_lpi: %016llX\n", p->sys_lpi);
- outp += sprintf(outp, "Joules PKG: %0llX\n", p->energy_pkg);
- outp += sprintf(outp, "Joules COR: %0llX\n", p->energy_cores);
- outp += sprintf(outp, "Joules GFX: %0llX\n", p->energy_gfx);
- outp += sprintf(outp, "Joules RAM: %0llX\n", p->energy_dram);
- outp += sprintf(outp, "Throttle PKG: %0llX\n", p->rapl_pkg_perf_status);
- outp += sprintf(outp, "Throttle RAM: %0llX\n", p->rapl_dram_perf_status);
+ outp += sprintf(outp, "Joules PKG: %0llX\n", p->energy_pkg.raw_value);
+ outp += sprintf(outp, "Joules COR: %0llX\n", p->energy_cores.raw_value);
+ outp += sprintf(outp, "Joules GFX: %0llX\n", p->energy_gfx.raw_value);
+ outp += sprintf(outp, "Joules RAM: %0llX\n", p->energy_dram.raw_value);
+ outp += sprintf(outp, "Throttle PKG: %0llX\n", p->rapl_pkg_perf_status.raw_value);
+ outp += sprintf(outp, "Throttle RAM: %0llX\n", p->rapl_dram_perf_status.raw_value);
outp += sprintf(outp, "PTM: %dC\n", p->pkg_temp_c);
for (i = 0, mp = sys.pp; mp; i++, mp = mp->next) {
return 0;
}
+double rapl_counter_get_value(const struct rapl_counter *c, enum rapl_unit desired_unit, double interval)
+{
+ assert(desired_unit != RAPL_UNIT_INVALID);
+
+ /*
+ * For now we don't expect anything other than joules,
+ * so just simplify the logic.
+ */
+ assert(c->unit == RAPL_UNIT_JOULES);
+
+ const double scaled = c->raw_value * c->scale;
+
+ if (desired_unit == RAPL_UNIT_WATTS)
+ return scaled / interval;
+ return scaled;
+}
+
/*
* column formatting convention & formats
*/
if (DO_BIC(BIC_CorWatt) && platform->has_per_core_rapl)
outp +=
- sprintf(outp, fmt8, (printed++ ? delim : ""), c->core_energy * rapl_energy_units / interval_float);
+ sprintf(outp, fmt8, (printed++ ? delim : ""),
+ rapl_counter_get_value(&c->core_energy, RAPL_UNIT_WATTS, interval_float));
if (DO_BIC(BIC_Cor_J) && platform->has_per_core_rapl)
- outp += sprintf(outp, fmt8, (printed++ ? delim : ""), c->core_energy * rapl_energy_units);
+ outp += sprintf(outp, fmt8, (printed++ ? delim : ""),
+ rapl_counter_get_value(&c->core_energy, RAPL_UNIT_JOULES, interval_float));
/* print per-package data only for 1st core in package */
if (!is_cpu_first_core_in_package(t, c, p))
if (DO_BIC(BIC_PkgWatt))
outp +=
- sprintf(outp, fmt8, (printed++ ? delim : ""), p->energy_pkg * rapl_energy_units / interval_float);
-
+ sprintf(outp, fmt8, (printed++ ? delim : ""),
+ rapl_counter_get_value(&p->energy_pkg, RAPL_UNIT_WATTS, interval_float));
if (DO_BIC(BIC_CorWatt) && !platform->has_per_core_rapl)
outp +=
- sprintf(outp, fmt8, (printed++ ? delim : ""), p->energy_cores * rapl_energy_units / interval_float);
+ sprintf(outp, fmt8, (printed++ ? delim : ""),
+ rapl_counter_get_value(&p->energy_cores, RAPL_UNIT_WATTS, interval_float));
if (DO_BIC(BIC_GFXWatt))
outp +=
- sprintf(outp, fmt8, (printed++ ? delim : ""), p->energy_gfx * rapl_energy_units / interval_float);
+ sprintf(outp, fmt8, (printed++ ? delim : ""),
+ rapl_counter_get_value(&p->energy_gfx, RAPL_UNIT_WATTS, interval_float));
if (DO_BIC(BIC_RAMWatt))
outp +=
sprintf(outp, fmt8, (printed++ ? delim : ""),
- p->energy_dram * rapl_dram_energy_units / interval_float);
+ rapl_counter_get_value(&p->energy_dram, RAPL_UNIT_WATTS, interval_float));
if (DO_BIC(BIC_Pkg_J))
- outp += sprintf(outp, fmt8, (printed++ ? delim : ""), p->energy_pkg * rapl_energy_units);
+ outp += sprintf(outp, fmt8, (printed++ ? delim : ""),
+ rapl_counter_get_value(&p->energy_pkg, RAPL_UNIT_JOULES, interval_float));
if (DO_BIC(BIC_Cor_J) && !platform->has_per_core_rapl)
- outp += sprintf(outp, fmt8, (printed++ ? delim : ""), p->energy_cores * rapl_energy_units);
+ outp += sprintf(outp, fmt8, (printed++ ? delim : ""),
+ rapl_counter_get_value(&p->energy_cores, RAPL_UNIT_JOULES, interval_float));
if (DO_BIC(BIC_GFX_J))
- outp += sprintf(outp, fmt8, (printed++ ? delim : ""), p->energy_gfx * rapl_energy_units);
+ outp += sprintf(outp, fmt8, (printed++ ? delim : ""),
+ rapl_counter_get_value(&p->energy_gfx, RAPL_UNIT_JOULES, interval_float));
if (DO_BIC(BIC_RAM_J))
- outp += sprintf(outp, fmt8, (printed++ ? delim : ""), p->energy_dram * rapl_dram_energy_units);
+ outp += sprintf(outp, fmt8, (printed++ ? delim : ""),
+ rapl_counter_get_value(&p->energy_dram, RAPL_UNIT_JOULES, interval_float));
if (DO_BIC(BIC_PKG__))
outp +=
sprintf(outp, fmt8, (printed++ ? delim : ""),
- 100.0 * p->rapl_pkg_perf_status * rapl_time_units / interval_float);
+ rapl_counter_get_value(&p->rapl_pkg_perf_status, RAPL_UNIT_WATTS, interval_float));
if (DO_BIC(BIC_RAM__))
outp +=
sprintf(outp, fmt8, (printed++ ? delim : ""),
- 100.0 * p->rapl_dram_perf_status * rapl_time_units / interval_float);
+ rapl_counter_get_value(&p->rapl_dram_perf_status, RAPL_UNIT_WATTS, interval_float));
/* UncMHz */
if (DO_BIC(BIC_UNCORE_MHZ))
outp += sprintf(outp, "%s%d", (printed++ ? delim : ""), p->uncore_mhz);
old->gfx_mhz = new->gfx_mhz;
old->gfx_act_mhz = new->gfx_act_mhz;
- old->energy_pkg = new->energy_pkg - old->energy_pkg;
- old->energy_cores = new->energy_cores - old->energy_cores;
- old->energy_gfx = new->energy_gfx - old->energy_gfx;
- old->energy_dram = new->energy_dram - old->energy_dram;
- old->rapl_pkg_perf_status = new->rapl_pkg_perf_status - old->rapl_pkg_perf_status;
- old->rapl_dram_perf_status = new->rapl_dram_perf_status - old->rapl_dram_perf_status;
+ old->energy_pkg.raw_value = new->energy_pkg.raw_value - old->energy_pkg.raw_value;
+ old->energy_cores.raw_value = new->energy_cores.raw_value - old->energy_cores.raw_value;
+ old->energy_gfx.raw_value = new->energy_gfx.raw_value - old->energy_gfx.raw_value;
+ old->energy_dram.raw_value = new->energy_dram.raw_value - old->energy_dram.raw_value;
+ old->rapl_pkg_perf_status.raw_value = new->rapl_pkg_perf_status.raw_value - old->rapl_pkg_perf_status.raw_value;
+ old->rapl_dram_perf_status.raw_value =
+ new->rapl_dram_perf_status.raw_value - old->rapl_dram_perf_status.raw_value;
for (i = 0, mp = sys.pp; mp; i++, mp = mp->next) {
if (mp->format == FORMAT_RAW)
old->core_throt_cnt = new->core_throt_cnt;
old->mc6_us = new->mc6_us - old->mc6_us;
- DELTA_WRAP32(new->core_energy, old->core_energy);
+ DELTA_WRAP32(new->core_energy.raw_value, old->core_energy.raw_value);
for (i = 0, mp = sys.cp; mp; i++, mp = mp->next) {
if (mp->format == FORMAT_RAW)
return retval;
}
+void rapl_counter_clear(struct rapl_counter *c)
+{
+ c->raw_value = 0;
+ c->scale = 0.0;
+ c->unit = RAPL_UNIT_INVALID;
+}
+
void clear_counters(struct thread_data *t, struct core_data *c, struct pkg_data *p)
{
int i;
c->c7 = 0;
c->mc6_us = 0;
c->core_temp_c = 0;
- c->core_energy = 0;
+ rapl_counter_clear(&c->core_energy);
c->core_throt_cnt = 0;
p->pkg_wtd_core_c0 = 0;
p->cpu_lpi = 0;
p->sys_lpi = 0;
- p->energy_pkg = 0;
- p->energy_dram = 0;
- p->energy_cores = 0;
- p->energy_gfx = 0;
- p->rapl_pkg_perf_status = 0;
- p->rapl_dram_perf_status = 0;
+ rapl_counter_clear(&p->energy_pkg);
+ rapl_counter_clear(&p->energy_dram);
+ rapl_counter_clear(&p->energy_cores);
+ rapl_counter_clear(&p->energy_gfx);
+ rapl_counter_clear(&p->rapl_pkg_perf_status);
+ rapl_counter_clear(&p->rapl_dram_perf_status);
p->pkg_temp_c = 0;
p->gfx_rc6_ms = 0;
p->counter[i] = 0;
}
+void rapl_counter_accumulate(struct rapl_counter *dst, const struct rapl_counter *src)
+{
+ /* Copy unit and scale from src if dst is not initialized */
+ if (dst->unit == RAPL_UNIT_INVALID) {
+ dst->unit = src->unit;
+ dst->scale = src->scale;
+ }
+
+ assert(dst->unit == src->unit);
+ assert(dst->scale == src->scale);
+
+ dst->raw_value += src->raw_value;
+}
+
int sum_counters(struct thread_data *t, struct core_data *c, struct pkg_data *p)
{
int i;
average.cores.core_temp_c = MAX(average.cores.core_temp_c, c->core_temp_c);
average.cores.core_throt_cnt = MAX(average.cores.core_throt_cnt, c->core_throt_cnt);
- average.cores.core_energy += c->core_energy;
+ rapl_counter_accumulate(&average.cores.core_energy, &c->core_energy);
for (i = 0, mp = sys.cp; mp; i++, mp = mp->next) {
if (mp->format == FORMAT_RAW)
average.packages.cpu_lpi = p->cpu_lpi;
average.packages.sys_lpi = p->sys_lpi;
- average.packages.energy_pkg += p->energy_pkg;
- average.packages.energy_dram += p->energy_dram;
- average.packages.energy_cores += p->energy_cores;
- average.packages.energy_gfx += p->energy_gfx;
+ rapl_counter_accumulate(&average.packages.energy_pkg, &p->energy_pkg);
+ rapl_counter_accumulate(&average.packages.energy_dram, &p->energy_dram);
+ rapl_counter_accumulate(&average.packages.energy_cores, &p->energy_cores);
+ rapl_counter_accumulate(&average.packages.energy_gfx, &p->energy_gfx);
average.packages.gfx_rc6_ms = p->gfx_rc6_ms;
average.packages.uncore_mhz = p->uncore_mhz;
average.packages.pkg_temp_c = MAX(average.packages.pkg_temp_c, p->pkg_temp_c);
- average.packages.rapl_pkg_perf_status += p->rapl_pkg_perf_status;
- average.packages.rapl_dram_perf_status += p->rapl_dram_perf_status;
+ rapl_counter_accumulate(&average.packages.rapl_pkg_perf_status, &p->rapl_pkg_perf_status);
+ rapl_counter_accumulate(&average.packages.rapl_dram_perf_status, &p->rapl_dram_perf_status);
for (i = 0, mp = sys.pp; mp; i++, mp = mp->next) {
if ((mp->format == FORMAT_RAW) && (topo.num_packages == 0))
int mperf;
};
-static unsigned int read_perf_counter_info(const char *const path, const char *const parse_format)
+static int read_perf_counter_info(const char *const path, const char *const parse_format, void *value_ptr)
{
int fdmt;
- char buf[16];
- unsigned int v;
+ int bytes_read;
+ char buf[64];
+ int ret = -1;
fdmt = open(path, O_RDONLY, 0);
- if (fdmt == -1)
- errx(1, "Failed to read perf counter info %s\n", path);
+ if (fdmt == -1) {
+ if (debug)
+ fprintf(stderr, "Failed to parse perf counter info %s\n", path);
+ ret = -1;
+ goto cleanup_and_exit;
+ }
- if (read(fdmt, buf, sizeof(buf)) <= 0)
- return 0;
+ bytes_read = read(fdmt, buf, sizeof(buf) - 1);
+ if (bytes_read <= 0 || bytes_read >= (int)sizeof(buf)) {
+ if (debug)
+ fprintf(stderr, "Failed to parse perf counter info %s\n", path);
+ ret = -1;
+ goto cleanup_and_exit;
+ }
- buf[sizeof(buf) - 1] = '\0';
+ buf[bytes_read] = '\0';
- if (sscanf(buf, parse_format, &v) != 1)
- errx(1, "Failed to parse perf counter info %s\n", path);
+ if (sscanf(buf, parse_format, value_ptr) != 1) {
+ if (debug)
+ fprintf(stderr, "Failed to parse perf counter info %s\n", path);
+ ret = -1;
+ goto cleanup_and_exit;
+ }
+ ret = 0;
+
+cleanup_and_exit:
close(fdmt);
+ return ret;
+}
+
+static unsigned int read_perf_counter_info_n(const char *const path, const char *const parse_format)
+{
+ unsigned int v;
+ int status;
+
+ status = read_perf_counter_info(path, parse_format, &v);
+ if (status)
+ v = -1;
return v;
}
const char *const path = "/sys/bus/event_source/devices/msr/type";
const char *const format = "%u";
- return read_perf_counter_info(path, format);
+ return read_perf_counter_info_n(path, format);
}
static unsigned read_aperf_config(void)
const char *const path = "/sys/bus/event_source/devices/msr/events/aperf";
const char *const format = "event=%x";
- return read_perf_counter_info(path, format);
+ return read_perf_counter_info_n(path, format);
}
static unsigned read_mperf_config(void)
const char *const path = "/sys/bus/event_source/devices/msr/events/mperf";
const char *const format = "event=%x";
- return read_perf_counter_info(path, format);
+ return read_perf_counter_info_n(path, format);
+}
+
+static unsigned read_perf_type(const char *subsys)
+{
+ const char *const path_format = "/sys/bus/event_source/devices/%s/type";
+ const char *const format = "%u";
+ char path[128];
+
+ snprintf(path, sizeof(path), path_format, subsys);
+
+ return read_perf_counter_info_n(path, format);
+}
+
+static unsigned read_rapl_config(const char *subsys, const char *event_name)
+{
+ const char *const path_format = "/sys/bus/event_source/devices/%s/events/%s";
+ const char *const format = "event=%x";
+ char path[128];
+
+ snprintf(path, sizeof(path), path_format, subsys, event_name);
+
+ return read_perf_counter_info_n(path, format);
+}
+
+static unsigned read_perf_rapl_unit(const char *subsys, const char *event_name)
+{
+ const char *const path_format = "/sys/bus/event_source/devices/%s/events/%s.unit";
+ const char *const format = "%s";
+ char path[128];
+ char unit_buffer[16];
+
+ snprintf(path, sizeof(path), path_format, subsys, event_name);
+
+ read_perf_counter_info(path, format, &unit_buffer);
+ if (strcmp("Joules", unit_buffer) == 0)
+ return RAPL_UNIT_JOULES;
+
+ return RAPL_UNIT_INVALID;
+}
+
+static double read_perf_rapl_scale(const char *subsys, const char *event_name)
+{
+ const char *const path_format = "/sys/bus/event_source/devices/%s/events/%s.scale";
+ const char *const format = "%lf";
+ char path[128];
+ double scale;
+
+ snprintf(path, sizeof(path), path_format, subsys, event_name);
+
+ if (read_perf_counter_info(path, format, &scale))
+ return 0.0;
+
+ return scale;
}
static struct amperf_group_fd open_amperf_fd(int cpu)
return 0;
}
+size_t rapl_counter_info_count_perf(const struct rapl_counter_info_t *rci)
+{
+ size_t ret = 0;
+
+ for (int i = 0; i < NUM_RAPL_COUNTERS; ++i)
+ if (rci->source[i] == RAPL_SOURCE_PERF)
+ ++ret;
+
+ return ret;
+}
+
+void write_rapl_counter(struct rapl_counter *rc, struct rapl_counter_info_t *rci, unsigned int idx)
+{
+ rc->raw_value = rci->data[idx];
+ rc->unit = rci->unit[idx];
+ rc->scale = rci->scale[idx];
+}
+
+int get_rapl_counters(int cpu, int domain, struct core_data *c, struct pkg_data *p)
+{
+ unsigned long long perf_data[NUM_RAPL_COUNTERS + 1];
+ struct rapl_counter_info_t *rci = &rapl_counter_info_perdomain[domain];
+
+ if (debug)
+ fprintf(stderr, "get_rapl_counters: cpu%d domain%d\n", cpu, domain);
+
+ assert(rapl_counter_info_perdomain);
+
+ /*
+ * If we have any perf counters to read, read them all now, in bulk
+ */
+ if (rci->fd_perf != -1) {
+ size_t num_perf_counters = rapl_counter_info_count_perf(rci);
+ const ssize_t expected_read_size = (num_perf_counters + 1) * sizeof(unsigned long long);
+ const ssize_t actual_read_size = read(rci->fd_perf, &perf_data[0], sizeof(perf_data));
+ if (actual_read_size != expected_read_size)
+ err(-1, "get_rapl_counters: failed to read perf_data (%zu %zu)", expected_read_size,
+ actual_read_size);
+ }
+
+ for (unsigned int i = 0, pi = 1; i < NUM_RAPL_COUNTERS; ++i) {
+ switch (rci->source[i]) {
+ case RAPL_SOURCE_NONE:
+ break;
+
+ case RAPL_SOURCE_PERF:
+ assert(pi < ARRAY_SIZE(perf_data));
+ assert(rci->fd_perf != -1);
+
+ if (debug)
+ fprintf(stderr, "Reading rapl counter via perf at %u (%llu %e %lf)\n",
+ i, perf_data[pi], rci->scale[i], perf_data[pi] * rci->scale[i]);
+
+ rci->data[i] = perf_data[pi];
+
+ ++pi;
+ break;
+
+ case RAPL_SOURCE_MSR:
+ if (debug)
+ fprintf(stderr, "Reading rapl counter via msr at %u\n", i);
+
+ assert(!no_msr);
+ if (rci->flags[i] & RAPL_COUNTER_FLAG_USE_MSR_SUM) {
+ if (get_msr_sum(cpu, rci->msr[i], &rci->data[i]))
+ return -13 - i;
+ } else {
+ if (get_msr(cpu, rci->msr[i], &rci->data[i]))
+ return -13 - i;
+ }
+
+ rci->data[i] &= rci->msr_mask[i];
+ if (rci->msr_shift[i] >= 0)
+ rci->data[i] >>= abs(rci->msr_shift[i]);
+ else
+ rci->data[i] <<= abs(rci->msr_shift[i]);
+
+ break;
+ }
+ }
+
+ _Static_assert(NUM_RAPL_COUNTERS == 7);
+ write_rapl_counter(&p->energy_pkg, rci, RAPL_RCI_INDEX_ENERGY_PKG);
+ write_rapl_counter(&p->energy_cores, rci, RAPL_RCI_INDEX_ENERGY_CORES);
+ write_rapl_counter(&p->energy_dram, rci, RAPL_RCI_INDEX_DRAM);
+ write_rapl_counter(&p->energy_gfx, rci, RAPL_RCI_INDEX_GFX);
+ write_rapl_counter(&p->rapl_pkg_perf_status, rci, RAPL_RCI_INDEX_PKG_PERF_STATUS);
+ write_rapl_counter(&p->rapl_dram_perf_status, rci, RAPL_RCI_INDEX_DRAM_PERF_STATUS);
+ write_rapl_counter(&c->core_energy, rci, RAPL_RCI_INDEX_CORE_ENERGY);
+
+ return 0;
+}
+
/*
* get_counters(...)
* migrate to cpu
unsigned long long msr;
struct msr_counter *mp;
int i;
+ int status;
if (cpu_migrate(cpu)) {
fprintf(outf, "get_counters: Could not migrate to CPU %d\n", cpu);
if (!is_cpu_first_thread_in_core(t, c, p))
goto done;
+ if (platform->has_per_core_rapl) {
+ status = get_rapl_counters(cpu, c->core_id, c, p);
+ if (status != 0)
+ return status;
+ }
+
if (DO_BIC(BIC_CPU_c3) || soft_c1_residency_display(BIC_CPU_c3)) {
if (get_msr(cpu, MSR_CORE_C3_RESIDENCY, &c->c3))
return -6;
if (DO_BIC(BIC_CORE_THROT_CNT))
get_core_throt_cnt(cpu, &c->core_throt_cnt);
- if ((platform->rapl_msrs & RAPL_AMD_F17H) && !no_msr) {
- if (get_msr(cpu, MSR_CORE_ENERGY_STAT, &msr))
- return -14;
- c->core_energy = msr & 0xFFFFFFFF;
- }
-
for (i = 0, mp = sys.cp; mp; i++, mp = mp->next) {
if (get_mp(cpu, mp, &c->counter[i]))
return -10;
if (DO_BIC(BIC_SYS_LPI))
p->sys_lpi = cpuidle_cur_sys_lpi_us;
- if (!no_msr) {
- if (platform->rapl_msrs & RAPL_PKG) {
- if (get_msr_sum(cpu, MSR_PKG_ENERGY_STATUS, &msr))
- return -13;
- p->energy_pkg = msr;
- }
- if (platform->rapl_msrs & RAPL_CORE_ENERGY_STATUS) {
- if (get_msr_sum(cpu, MSR_PP0_ENERGY_STATUS, &msr))
- return -14;
- p->energy_cores = msr;
- }
- if (platform->rapl_msrs & RAPL_DRAM) {
- if (get_msr_sum(cpu, MSR_DRAM_ENERGY_STATUS, &msr))
- return -15;
- p->energy_dram = msr;
- }
- if (platform->rapl_msrs & RAPL_GFX) {
- if (get_msr_sum(cpu, MSR_PP1_ENERGY_STATUS, &msr))
- return -16;
- p->energy_gfx = msr;
- }
- if (platform->rapl_msrs & RAPL_PKG_PERF_STATUS) {
- if (get_msr_sum(cpu, MSR_PKG_PERF_STATUS, &msr))
- return -16;
- p->rapl_pkg_perf_status = msr;
- }
- if (platform->rapl_msrs & RAPL_DRAM_PERF_STATUS) {
- if (get_msr_sum(cpu, MSR_DRAM_PERF_STATUS, &msr))
- return -16;
- p->rapl_dram_perf_status = msr;
- }
- if (platform->rapl_msrs & RAPL_AMD_F17H) {
- if (get_msr_sum(cpu, MSR_PKG_ENERGY_STAT, &msr))
- return -13;
- p->energy_pkg = msr;
- }
+ if (!platform->has_per_core_rapl) {
+ status = get_rapl_counters(cpu, p->package_id, c, p);
+ if (status != 0)
+ return status;
}
if (DO_BIC(BIC_PkgTmp)) {
fd_instr_count_percpu = NULL;
}
+void free_fd_rapl_percpu(void)
+{
+ if (!rapl_counter_info_perdomain)
+ return;
+
+ const int num_domains = platform->has_per_core_rapl ? topo.num_cores : topo.num_packages;
+
+ for (int domain_id = 0; domain_id < num_domains; ++domain_id) {
+ if (rapl_counter_info_perdomain[domain_id].fd_perf != -1)
+ close(rapl_counter_info_perdomain[domain_id].fd_perf);
+ }
+
+ free(rapl_counter_info_perdomain);
+}
+
void free_all_buffers(void)
{
int i;
free_fd_percpu();
free_fd_instr_count_percpu();
free_fd_amperf_percpu();
+ free_fd_rapl_percpu();
free(irq_column_2_cpu);
free(irqs_per_cpu);
}
void linux_perf_init(void);
+void rapl_perf_init(void);
void re_initialize(void)
{
free_all_buffers();
setup_all_buffers(false);
linux_perf_init();
+ rapl_perf_init();
fprintf(outf, "turbostat: re-initialized with num_cpus %d, allowed_cpus %d\n", topo.num_cpus,
topo.allowed_cpus);
}
unsigned long long msr;
unsigned int time_unit;
double tdp;
+ const unsigned long long bic_watt_bits = BIC_PkgWatt | BIC_CorWatt | BIC_RAMWatt | BIC_GFXWatt;
+ const unsigned long long bic_joules_bits = BIC_Pkg_J | BIC_Cor_J | BIC_RAM_J | BIC_GFX_J;
- if (rapl_joules) {
- if (platform->rapl_msrs & RAPL_PKG_ENERGY_STATUS)
- BIC_PRESENT(BIC_Pkg_J);
- if (platform->rapl_msrs & RAPL_CORE_ENERGY_STATUS)
- BIC_PRESENT(BIC_Cor_J);
- if (platform->rapl_msrs & RAPL_DRAM_ENERGY_STATUS)
- BIC_PRESENT(BIC_RAM_J);
- if (platform->rapl_msrs & RAPL_GFX_ENERGY_STATUS)
- BIC_PRESENT(BIC_GFX_J);
- } else {
- if (platform->rapl_msrs & RAPL_PKG_ENERGY_STATUS)
- BIC_PRESENT(BIC_PkgWatt);
- if (platform->rapl_msrs & RAPL_CORE_ENERGY_STATUS)
- BIC_PRESENT(BIC_CorWatt);
- if (platform->rapl_msrs & RAPL_DRAM_ENERGY_STATUS)
- BIC_PRESENT(BIC_RAMWatt);
- if (platform->rapl_msrs & RAPL_GFX_ENERGY_STATUS)
- BIC_PRESENT(BIC_GFXWatt);
- }
+ if (rapl_joules)
+ bic_enabled &= ~bic_watt_bits;
+ else
+ bic_enabled &= ~bic_joules_bits;
- if (platform->rapl_msrs & RAPL_PKG_PERF_STATUS)
- BIC_PRESENT(BIC_PKG__);
- if (platform->rapl_msrs & RAPL_DRAM_PERF_STATUS)
- BIC_PRESENT(BIC_RAM__);
+ if (!(platform->rapl_msrs & RAPL_PKG_PERF_STATUS))
+ bic_enabled &= ~BIC_PKG__;
+ if (!(platform->rapl_msrs & RAPL_DRAM_PERF_STATUS))
+ bic_enabled &= ~BIC_RAM__;
/* units on package 0, verify later other packages match */
if (get_msr(base_cpu, MSR_RAPL_POWER_UNIT, &msr))
{
unsigned long long msr;
double tdp;
+ const unsigned long long bic_watt_bits = BIC_PkgWatt | BIC_CorWatt;
+ const unsigned long long bic_joules_bits = BIC_Pkg_J | BIC_Cor_J;
- if (rapl_joules) {
- BIC_PRESENT(BIC_Pkg_J);
- BIC_PRESENT(BIC_Cor_J);
- } else {
- BIC_PRESENT(BIC_PkgWatt);
- BIC_PRESENT(BIC_CorWatt);
- }
+ if (rapl_joules)
+ bic_enabled &= ~bic_watt_bits;
+ else
+ bic_enabled &= ~bic_joules_bits;
if (get_msr(base_cpu, MSR_RAPL_PWR_UNIT, &msr))
return;
|| BIC_IS_ENABLED(BIC_IPC);
}
+int add_rapl_perf_counter_(int cpu, struct rapl_counter_info_t *rci, const struct rapl_counter_arch_info *cai,
+ double *scale_, enum rapl_unit *unit_)
+{
+ if (no_perf)
+ return -1;
+
+ const double scale = read_perf_rapl_scale(cai->perf_subsys, cai->perf_name);
+ if (scale == 0.0)
+ return -1;
+
+ const enum rapl_unit unit = read_perf_rapl_unit(cai->perf_subsys, cai->perf_name);
+ if (unit == RAPL_UNIT_INVALID)
+ return -1;
+
+ const unsigned rapl_type = read_perf_type(cai->perf_subsys);
+ const unsigned rapl_energy_pkg_config = read_rapl_config(cai->perf_subsys, cai->perf_name);
+
+ const int fd_counter =
+ open_perf_counter(cpu, rapl_type, rapl_energy_pkg_config, rci->fd_perf, PERF_FORMAT_GROUP);
+ if (fd_counter == -1)
+ return -1;
+
+ /* If it's the first counter opened, make it a group descriptor */
+ if (rci->fd_perf == -1)
+ rci->fd_perf = fd_counter;
+
+ *scale_ = scale;
+ *unit_ = unit;
+ return fd_counter;
+}
+
+int add_rapl_perf_counter(int cpu, struct rapl_counter_info_t *rci, const struct rapl_counter_arch_info *cai,
+ double *scale, enum rapl_unit *unit)
+{
+ int ret = add_rapl_perf_counter_(cpu, rci, cai, scale, unit);
+
+ if (debug)
+ fprintf(stderr, "add_rapl_perf_counter: %d (cpu: %d)\n", ret, cpu);
+
+ return ret;
+}
+
/*
* Linux-perf manages the HW instructions-retired counter
* by enabling when requested, and hiding rollover
}
}
-static int has_amperf_access_via_msr(void)
+void rapl_perf_init(void)
{
- unsigned long long dummy;
+ const int num_domains = platform->has_per_core_rapl ? topo.num_cores : topo.num_packages;
+ bool *domain_visited = calloc(num_domains, sizeof(bool));
+
+ rapl_counter_info_perdomain = calloc(num_domains, sizeof(*rapl_counter_info_perdomain));
+ if (rapl_counter_info_perdomain == NULL)
+ err(-1, "calloc rapl_counter_info_percpu");
+ /*
+ * Initialize rapl_counter_info_percpu
+ */
+ for (int domain_id = 0; domain_id < num_domains; ++domain_id) {
+ struct rapl_counter_info_t *rci = &rapl_counter_info_perdomain[domain_id];
+ rci->fd_perf = -1;
+ for (size_t i = 0; i < NUM_RAPL_COUNTERS; ++i) {
+ rci->data[i] = 0;
+ rci->source[i] = RAPL_SOURCE_NONE;
+ }
+ }
+
+ /*
+ * Open/probe the counters
+ * If can't get it via perf, fallback to MSR
+ */
+ for (size_t i = 0; i < ARRAY_SIZE(rapl_counter_arch_infos); ++i) {
+
+ const struct rapl_counter_arch_info *const cai = &rapl_counter_arch_infos[i];
+ bool has_counter = 0;
+ double scale;
+ enum rapl_unit unit;
+ int next_domain;
+
+ memset(domain_visited, 0, num_domains * sizeof(*domain_visited));
+
+ for (int cpu = 0; cpu < topo.max_cpu_num + 1; ++cpu) {
+
+ if (cpu_is_not_allowed(cpu))
+ continue;
+
+ /* Skip already seen and handled RAPL domains */
+ next_domain =
+ platform->has_per_core_rapl ? cpus[cpu].physical_core_id : cpus[cpu].physical_package_id;
+
+ if (domain_visited[next_domain])
+ continue;
+
+ domain_visited[next_domain] = 1;
+
+ struct rapl_counter_info_t *rci = &rapl_counter_info_perdomain[next_domain];
+
+ /* Check if the counter is enabled and accessible */
+ if (BIC_IS_ENABLED(cai->bic) && (platform->rapl_msrs & cai->feature_mask)) {
+
+ /* Use perf API for this counter */
+ if (!no_perf && cai->perf_name
+ && add_rapl_perf_counter(cpu, rci, cai, &scale, &unit) != -1) {
+ rci->source[cai->rci_index] = RAPL_SOURCE_PERF;
+ rci->scale[cai->rci_index] = scale * cai->compat_scale;
+ rci->unit[cai->rci_index] = unit;
+ rci->flags[cai->rci_index] = cai->flags;
+
+ /* Use MSR for this counter */
+ } else if (!no_msr && cai->msr && probe_msr(cpu, cai->msr) == 0) {
+ rci->source[cai->rci_index] = RAPL_SOURCE_MSR;
+ rci->msr[cai->rci_index] = cai->msr;
+ rci->msr_mask[cai->rci_index] = cai->msr_mask;
+ rci->msr_shift[cai->rci_index] = cai->msr_shift;
+ rci->unit[cai->rci_index] = RAPL_UNIT_JOULES;
+ rci->scale[cai->rci_index] = *cai->platform_rapl_msr_scale * cai->compat_scale;
+ rci->flags[cai->rci_index] = cai->flags;
+ }
+ }
+
+ if (rci->source[cai->rci_index] != RAPL_SOURCE_NONE)
+ has_counter = 1;
+ }
+
+ /* If any CPU has access to the counter, make it present */
+ if (has_counter)
+ BIC_PRESENT(cai->bic);
+ }
+
+ free(domain_visited);
+}
+
+static int has_amperf_access_via_msr(void)
+{
if (no_msr)
return 0;
- if (get_msr(base_cpu, MSR_IA32_APERF, &dummy))
+ if (probe_msr(base_cpu, MSR_IA32_APERF))
return 0;
- if (get_msr(base_cpu, MSR_IA32_MPERF, &dummy))
+ if (probe_msr(base_cpu, MSR_IA32_MPERF))
return 0;
return 1;
|| BIC_IS_ENABLED(BIC_Pkgpc8)
|| BIC_IS_ENABLED(BIC_Pkgpc9)
|| BIC_IS_ENABLED(BIC_Pkgpc10)
+ /* TODO: Multiplex access with perf */
|| BIC_IS_ENABLED(BIC_CorWatt)
|| BIC_IS_ENABLED(BIC_Cor_J)
|| BIC_IS_ENABLED(BIC_PkgWatt)
probe_pm_features();
set_amperf_source();
linux_perf_init();
+ rapl_perf_init();
for_all_cpus(get_cpu_type, ODD_COUNTERS);
for_all_cpus(get_cpu_type, EVEN_COUNTERS);
projects / linux.git / commitdiff