Add perf counters based metrics for intel platforms
Unlike AMD or Nvidia intel historically only exposes gpu telemetry with linux perf counters. As with all perf counters on modern linux these require root or CAP_PERFMON. This adds a small helper program that can be setuid/setcap'ed with the appropriate permissions to export the telemetry that mangohud is interested in. Most of the gpu metrics were informed by the existing intel_gpu_top tool we are using, however this also exports power metrics that the intel_gpu_top tool does not support for dgpu configurations and ensures that we select the appropriate card for all metrics in a multi-gpu system. Also this tool exports CPU power metrics from the perf counters. This allows power monitoring for Intel and zen+ AMD cpus. But this isnt hooked in this PR since it seems cpu and gpu metrics collection are separated in the rest of the mangohud. Mostly this is just a PoC to show what avoiding taking a third-party dependency on the rarely packaged intel_gpu_top might require. The meson setup in particular is just something thrown together that builds.pull/877/head
Kurt Kartaltepe
1 year ago
parent
38604927d7
commit
96b605ce88
@ -0,0 +1,230 @@
|
||||
#include <nlohmann/json.hpp>
|
||||
|
||||
#include <dirent.h>
|
||||
#include <iostream>
|
||||
#include <libdrm/i915_drm.h>
|
||||
#include <linux/perf_event.h>
|
||||
#include <stdio.h>
|
||||
#include <sys/stat.h>
|
||||
#include <sys/syscall.h>
|
||||
#include <thread>
|
||||
#include <unistd.h>
|
||||
#include <cinttypes>
|
||||
|
||||
// cut down from file_utils.cpp
|
||||
static std::vector<std::string> lsdir(const char *root, const char *prefix) {
|
||||
std::vector<std::string> list;
|
||||
struct dirent *dp;
|
||||
|
||||
DIR *dirp = opendir(root);
|
||||
if (!dirp) {
|
||||
return list;
|
||||
}
|
||||
|
||||
while ((dp = readdir(dirp))) {
|
||||
if ((prefix && !std::string(dp->d_name).find(prefix) == 0) ||
|
||||
!strcmp(dp->d_name, ".") || !strcmp(dp->d_name, ".."))
|
||||
continue;
|
||||
|
||||
switch (dp->d_type) {
|
||||
case DT_DIR:
|
||||
list.push_back(dp->d_name);
|
||||
break;
|
||||
}
|
||||
}
|
||||
|
||||
closedir(dirp);
|
||||
return list;
|
||||
}
|
||||
|
||||
static inline int perf_event_open(struct perf_event_attr *attr, pid_t pid,
|
||||
int cpu, int group_fd, unsigned long flags) {
|
||||
attr->size = sizeof(*attr);
|
||||
return syscall(SYS_perf_event_open, attr, pid, cpu, group_fd, flags);
|
||||
}
|
||||
|
||||
static uint64_t i915_perf_device_type(const char *bus_id) {
|
||||
FILE *fd;
|
||||
char buf[80] = {0};
|
||||
uint64_t type;
|
||||
|
||||
bool is_igpu = strcmp(bus_id, "0000_00_02.0") == 0; // Ref: igt-gpu-tools for magic.
|
||||
const char *dgpu_path = "/sys/bus/event_source/devices/i915_%s/type";
|
||||
const char *igpu_path = "/sys/bus/event_source/devices/i915/type";
|
||||
|
||||
snprintf(buf, sizeof(buf), is_igpu ? igpu_path : dgpu_path, bus_id);
|
||||
if ((fd = fopen(buf, "r")) == nullptr)
|
||||
return 0;
|
||||
|
||||
if (fscanf(fd, "%" PRIu64, &type) != 1)
|
||||
type = 0;
|
||||
|
||||
fclose(fd);
|
||||
return type;
|
||||
}
|
||||
|
||||
static inline int i915_perf_event_open(uint64_t type, uint64_t config) {
|
||||
uint64_t format = PERF_FORMAT_TOTAL_TIME_ENABLED;
|
||||
uint64_t group = -1;
|
||||
struct perf_event_attr attr = {0};
|
||||
|
||||
attr.type = type;
|
||||
attr.read_format = format;
|
||||
attr.config = config;
|
||||
attr.use_clockid = 1;
|
||||
attr.clockid = CLOCK_MONOTONIC;
|
||||
|
||||
int cpu = 0, ncpu = 16; // hopefully one of these works.
|
||||
int ret;
|
||||
do {
|
||||
ret = perf_event_open(&attr, -1, cpu++, group, 0);
|
||||
} while ((ret < 0 && errno == EINVAL) &&
|
||||
(cpu < ncpu)); // find a cpu to open on.
|
||||
|
||||
return ret;
|
||||
}
|
||||
|
||||
static int power_perf_event_open(const char *metric) {
|
||||
FILE *fd;
|
||||
char buf[80] = {0};
|
||||
uint64_t type, config;
|
||||
|
||||
snprintf(buf, sizeof(buf), "/sys/devices/power/type");
|
||||
if ((fd = fopen(buf, "r")) == nullptr)
|
||||
return -1;
|
||||
|
||||
int scanned = fscanf(fd, "%" PRIu64, &type);
|
||||
fclose(fd);
|
||||
if (scanned != 1)
|
||||
return -1;
|
||||
|
||||
snprintf(buf, sizeof(buf), "/sys/devices/power/events/%s", metric);
|
||||
if ((fd = fopen(buf, "r")) == nullptr)
|
||||
return -1;
|
||||
|
||||
scanned = fscanf(fd, "event=%" PRIx64, &config);
|
||||
fclose(fd);
|
||||
if (scanned != 1)
|
||||
return -1;
|
||||
|
||||
return i915_perf_event_open(type, config);
|
||||
}
|
||||
|
||||
struct counter {
|
||||
union {
|
||||
int fd;
|
||||
FILE *fp;
|
||||
};
|
||||
uint64_t val;
|
||||
uint64_t ts;
|
||||
uint64_t val_prev;
|
||||
uint64_t ts_prev;
|
||||
};
|
||||
|
||||
static void counter_update(struct counter *c, uint64_t in[4]) {
|
||||
c->val_prev = c->val;
|
||||
c->ts_prev = c->ts;
|
||||
c->val = in[0];
|
||||
c->ts = in[1];
|
||||
}
|
||||
|
||||
static float counter_value(struct counter c, float scale) {
|
||||
if (c.val < c.val_prev || c.ts_prev == 0) { // just try again next time.
|
||||
return 0.0;
|
||||
}
|
||||
float t = c.ts - c.ts_prev;
|
||||
float d = c.val - c.val_prev;
|
||||
return d / t * scale;
|
||||
}
|
||||
|
||||
static std::string find_pci_busid(const char *drm_dev) {
|
||||
char drm_link_c[PATH_MAX] = {0};
|
||||
ssize_t count = readlink((std::string("/sys/class/drm/") + drm_dev).c_str(),
|
||||
drm_link_c, PATH_MAX);
|
||||
if (count <= 0) {
|
||||
fprintf(stderr, "Invalid drm device: %s\n", drm_dev);
|
||||
exit(1);
|
||||
}
|
||||
std::string drm_link(drm_link_c);
|
||||
// Example link format:
|
||||
// ../../devices/pci0000:00/0000:00:03.1/0000:07:00.0/0000:08:01.0/0000:09:00.0/drm/card0
|
||||
ssize_t pci_start = drm_link.rfind("/", count - 14) + 1;
|
||||
ssize_t pci_end = drm_link.find("/", pci_start);
|
||||
return drm_link.substr(pci_start, pci_end - pci_start);
|
||||
}
|
||||
|
||||
int main(int argc, char **argv) {
|
||||
if (argc < 2) {
|
||||
fprintf(stderr, "Not enough arguments. usage: mango_intel_stats \"card0\"\n");
|
||||
exit(1);
|
||||
}
|
||||
char *drm_dev = argv[1];
|
||||
std::string pci_path = find_pci_busid(drm_dev);
|
||||
std::replace(pci_path.begin(), pci_path.end(), ':',
|
||||
'_'); // just sysfs things.
|
||||
uint64_t type = i915_perf_device_type(pci_path.c_str());
|
||||
|
||||
// Counters for this tool.
|
||||
struct counter busy_gpu, freq_gpu, energy_dgpu, energy_igpu, energy_cpu;
|
||||
|
||||
freq_gpu.fd = i915_perf_event_open(type, I915_PMU_ACTUAL_FREQUENCY);
|
||||
if (errno == EACCES) {
|
||||
fprintf(
|
||||
stderr,
|
||||
"Permission denied on perf events.\nThis binary is meant to have "
|
||||
"CAP_PERFMON\n via: sudo setcap cap_perfmon=+ep mango_intel_stats\n");
|
||||
exit(1);
|
||||
}
|
||||
busy_gpu.fd = i915_perf_event_open(
|
||||
type, I915_PMU_ENGINE_BUSY(I915_ENGINE_CLASS_RENDER, 0));
|
||||
energy_cpu.fd = power_perf_event_open("energy-pkg");
|
||||
energy_igpu.fd = power_perf_event_open("energy-gpu");
|
||||
|
||||
energy_dgpu = {0};
|
||||
std::string hwmon_path =
|
||||
std::string("/sys/class/drm/") + drm_dev + "/device/hwmon/";
|
||||
const auto dirs = lsdir(hwmon_path.c_str(), "hwmon");
|
||||
for (const auto &dir : dirs) {
|
||||
FILE *fp = fopen((hwmon_path + dir + "/energy1_input").c_str(), "r");
|
||||
if (fp) {
|
||||
energy_dgpu.fp = fp;
|
||||
break;
|
||||
}
|
||||
}
|
||||
|
||||
while (true) {
|
||||
nlohmann::json j;
|
||||
uint64_t res[4] = {0};
|
||||
if (read(busy_gpu.fd, res, sizeof(res)) > 0) {
|
||||
counter_update(&busy_gpu, res);
|
||||
j["gpu_busy_pct"] = counter_value(busy_gpu, 100.0);
|
||||
}
|
||||
if (read(freq_gpu.fd, res, sizeof(res)) > 0) {
|
||||
counter_update(&freq_gpu, res);
|
||||
j["gpu_clock_mhz"] = counter_value(freq_gpu, 1e9);
|
||||
}
|
||||
if (read(energy_cpu.fd, res, sizeof(res)) > 0) {
|
||||
counter_update(&energy_cpu, res);
|
||||
// https://lwn.net/Articles/573602/, in practice intel and amd use 1/2^32J
|
||||
j["cpu_power_w"] = counter_value(energy_cpu, 2.328e-1);
|
||||
}
|
||||
if (read(energy_igpu.fd, res, sizeof(res)) > 0) {
|
||||
counter_update(&energy_igpu, res);
|
||||
j["igpu_power_w"] = counter_value(energy_igpu, 2.328e-1);
|
||||
}
|
||||
if (energy_dgpu.fp && fscanf(energy_dgpu.fp, "%" PRIu64, res) == 1) {
|
||||
rewind(energy_dgpu.fp);
|
||||
fflush(energy_dgpu.fp);
|
||||
// We can just depend on old clock readings in res[1]
|
||||
/*
|
||||
struct timespec t = {0};
|
||||
clock_gettime(CLOCK_MONOTONIC, &t);
|
||||
res[1] = (t.tv_sec % (3600 * 24)) * 1e9 + t.tv_nsec;
|
||||
*/
|
||||
counter_update(&energy_dgpu, res);
|
||||
j["dgpu_power_w"] = counter_value(energy_dgpu, 1e3);
|
||||
}
|
||||
std::cout << j << std::endl;
|
||||
std::this_thread::sleep_for(std::chrono::milliseconds(500));
|
||||
}
|
||||
}
|
||||
Loading…
Reference in New Issue