// Copyright 2015 Google Inc. All rights reserved. // // Licensed under the Apache License, Version 2.0 (the "License"); // you may not use this file except in compliance with the License. // You may obtain a copy of the License at // // http://www.apache.org/licenses/LICENSE-2.0 // // Unless required by applicable law or agreed to in writing, software // distributed under the License is distributed on an "AS IS" BASIS, // WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. // See the License for the specific language governing permissions and // limitations under the License. #include "sysinfo.h" #include "internal_macros.h" #ifdef BENCHMARK_OS_WINDOWS #include #include #include #else #include #include #include // this header must be included before 'sys/sysctl.h' to avoid compilation error on FreeBSD #include #include #if defined BENCHMARK_OS_FREEBSD || defined BENCHMARK_OS_MACOSX #include #endif #endif #include #include #include #include #include #include #include #include #include "arraysize.h" #include "check.h" #include "cycleclock.h" #include "internal_macros.h" #include "log.h" #include "sleep.h" #include "string_util.h" namespace benchmark { namespace { std::once_flag cpuinfo_init; double cpuinfo_cycles_per_second = 1.0; int cpuinfo_num_cpus = 1; // Conservative guess std::mutex cputimens_mutex; #if !defined BENCHMARK_OS_MACOSX const int64_t estimate_time_ms = 1000; // Helper function estimates cycles/sec by observing cycles elapsed during // sleep(). Using small sleep time decreases accuracy significantly. int64_t EstimateCyclesPerSecond() { const int64_t start_ticks = cycleclock::Now(); SleepForMilliseconds(estimate_time_ms); return cycleclock::Now() - start_ticks; } #endif #if defined BENCHMARK_OS_LINUX || defined BENCHMARK_OS_CYGWIN // Helper function for reading an int from a file. Returns true if successful // and the memory location pointed to by value is set to the value read. bool ReadIntFromFile(const char* file, long* value) { bool ret = false; int fd = open(file, O_RDONLY); if (fd != -1) { char line[1024]; char* err; memset(line, '\0', sizeof(line)); CHECK(read(fd, line, sizeof(line) - 1)); const long temp_value = strtol(line, &err, 10); if (line[0] != '\0' && (*err == '\n' || *err == '\0')) { *value = temp_value; ret = true; } close(fd); } return ret; } #endif void InitializeSystemInfo() { #if defined BENCHMARK_OS_LINUX || defined BENCHMARK_OS_CYGWIN char line[1024]; char* err; long freq; bool saw_mhz = false; // If the kernel is exporting the tsc frequency use that. There are issues // where cpuinfo_max_freq cannot be relied on because the BIOS may be // exporintg an invalid p-state (on x86) or p-states may be used to put the // processor in a new mode (turbo mode). Essentially, those frequencies // cannot always be relied upon. The same reasons apply to /proc/cpuinfo as // well. if (!saw_mhz && ReadIntFromFile("/sys/devices/system/cpu/cpu0/tsc_freq_khz", &freq)) { // The value is in kHz (as the file name suggests). For example, on a // 2GHz warpstation, the file contains the value "2000000". cpuinfo_cycles_per_second = freq * 1000.0; saw_mhz = true; } // If CPU scaling is in effect, we want to use the *maximum* frequency, // not whatever CPU speed some random processor happens to be using now. if (!saw_mhz && ReadIntFromFile("/sys/devices/system/cpu/cpu0/cpufreq/cpuinfo_max_freq", &freq)) { // The value is in kHz. For example, on a 2GHz warpstation, the file // contains the value "2000000". cpuinfo_cycles_per_second = freq * 1000.0; saw_mhz = true; } // Read /proc/cpuinfo for other values, and if there is no cpuinfo_max_freq. const char* pname = "/proc/cpuinfo"; int fd = open(pname, O_RDONLY); if (fd == -1) { perror(pname); if (!saw_mhz) { cpuinfo_cycles_per_second = static_cast(EstimateCyclesPerSecond()); } return; } double bogo_clock = 1.0; bool saw_bogo = false; long max_cpu_id = 0; int num_cpus = 0; line[0] = line[1] = '\0'; size_t chars_read = 0; do { // we'll exit when the last read didn't read anything // Move the next line to the beginning of the buffer const size_t oldlinelen = strlen(line); if (sizeof(line) == oldlinelen + 1) // oldlinelen took up entire line line[0] = '\0'; else // still other lines left to save memmove(line, line + oldlinelen + 1, sizeof(line) - (oldlinelen + 1)); // Terminate the new line, reading more if we can't find the newline char* newline = strchr(line, '\n'); if (newline == nullptr) { const size_t linelen = strlen(line); const size_t bytes_to_read = sizeof(line) - 1 - linelen; CHECK(bytes_to_read > 0); // because the memmove recovered >=1 bytes chars_read = read(fd, line + linelen, bytes_to_read); line[linelen + chars_read] = '\0'; newline = strchr(line, '\n'); } if (newline != nullptr) *newline = '\0'; // When parsing the "cpu MHz" and "bogomips" (fallback) entries, we only // accept postive values. Some environments (virtual machines) report zero, // which would cause infinite looping in WallTime_Init. if (!saw_mhz && strncasecmp(line, "cpu MHz", sizeof("cpu MHz") - 1) == 0) { const char* freqstr = strchr(line, ':'); if (freqstr) { cpuinfo_cycles_per_second = strtod(freqstr + 1, &err) * 1000000.0; if (freqstr[1] != '\0' && *err == '\0' && cpuinfo_cycles_per_second > 0) saw_mhz = true; } } else if (strncasecmp(line, "bogomips", sizeof("bogomips") - 1) == 0) { const char* freqstr = strchr(line, ':'); if (freqstr) { bogo_clock = strtod(freqstr + 1, &err) * 1000000.0; if (freqstr[1] != '\0' && *err == '\0' && bogo_clock > 0) saw_bogo = true; } } else if (strncmp(line, "processor", sizeof("processor") - 1) == 0) { // The above comparison is case-sensitive because ARM kernels often // include a "Processor" line that tells you about the CPU, distinct // from the usual "processor" lines that give you CPU ids. No current // Linux architecture is using "Processor" for CPU ids. num_cpus++; // count up every time we see an "processor :" entry const char* id_str = strchr(line, ':'); if (id_str) { const long cpu_id = strtol(id_str + 1, &err, 10); if (id_str[1] != '\0' && *err == '\0' && max_cpu_id < cpu_id) max_cpu_id = cpu_id; } } } while (chars_read > 0); close(fd); if (!saw_mhz) { if (saw_bogo) { // If we didn't find anything better, we'll use bogomips, but // we're not happy about it. cpuinfo_cycles_per_second = bogo_clock; } else { // If we don't even have bogomips, we'll use the slow estimation. cpuinfo_cycles_per_second = static_cast(EstimateCyclesPerSecond()); } } if (num_cpus == 0) { fprintf(stderr, "Failed to read num. CPUs correctly from /proc/cpuinfo\n"); } else { if ((max_cpu_id + 1) != num_cpus) { fprintf(stderr, "CPU ID assignments in /proc/cpuinfo seem messed up." " This is usually caused by a bad BIOS.\n"); } cpuinfo_num_cpus = num_cpus; } #elif defined BENCHMARK_OS_FREEBSD // For this sysctl to work, the machine must be configured without // SMP, APIC, or APM support. hz should be 64-bit in freebsd 7.0 // and later. Before that, it's a 32-bit quantity (and gives the // wrong answer on machines faster than 2^32 Hz). See // http://lists.freebsd.org/pipermail/freebsd-i386/2004-November/001846.html // But also compare FreeBSD 7.0: // http://fxr.watson.org/fxr/source/i386/i386/tsc.c?v=RELENG70#L223 // 231 error = sysctl_handle_quad(oidp, &freq, 0, req); // To FreeBSD 6.3 (it's the same in 6-STABLE): // http://fxr.watson.org/fxr/source/i386/i386/tsc.c?v=RELENG6#L131 // 139 error = sysctl_handle_int(oidp, &freq, sizeof(freq), req); #if __FreeBSD__ >= 7 uint64_t hz = 0; #else unsigned int hz = 0; #endif size_t sz = sizeof(hz); const char* sysctl_path = "machdep.tsc_freq"; if (sysctlbyname(sysctl_path, &hz, &sz, nullptr, 0) != 0) { fprintf(stderr, "Unable to determine clock rate from sysctl: %s: %s\n", sysctl_path, strerror(errno)); cpuinfo_cycles_per_second = static_cast(EstimateCyclesPerSecond()); } else { cpuinfo_cycles_per_second = hz; } // TODO: also figure out cpuinfo_num_cpus #elif defined BENCHMARK_OS_WINDOWS // In NT, read MHz from the registry. If we fail to do so or we're in win9x // then make a crude estimate. DWORD data, data_size = sizeof(data); if (IsWindowsXPOrGreater() && SUCCEEDED( SHGetValueA(HKEY_LOCAL_MACHINE, "HARDWARE\\DESCRIPTION\\System\\CentralProcessor\\0", "~MHz", nullptr, &data, &data_size))) cpuinfo_cycles_per_second = static_cast((int64_t)data * (int64_t)(1000 * 1000)); // was mhz else cpuinfo_cycles_per_second = static_cast(EstimateCyclesPerSecond()); // TODO: also figure out cpuinfo_num_cpus #elif defined BENCHMARK_OS_MACOSX // returning "mach time units" per second. the current number of elapsed // mach time units can be found by calling uint64 mach_absolute_time(); // while not as precise as actual CPU cycles, it is accurate in the face // of CPU frequency scaling and multi-cpu/core machines. // Our mac users have these types of machines, and accuracy // (i.e. correctness) trumps precision. // See cycleclock.h: CycleClock::Now(), which returns number of mach time // units on Mac OS X. mach_timebase_info_data_t timebase_info; mach_timebase_info(&timebase_info); double mach_time_units_per_nanosecond = static_cast(timebase_info.denom) / static_cast(timebase_info.numer); cpuinfo_cycles_per_second = mach_time_units_per_nanosecond * 1e9; int num_cpus = 0; size_t size = sizeof(num_cpus); int numcpus_name[] = {CTL_HW, HW_NCPU}; if (::sysctl(numcpus_name, arraysize(numcpus_name), &num_cpus, &size, nullptr, 0) == 0 && (size == sizeof(num_cpus))) cpuinfo_num_cpus = num_cpus; #else // Generic cycles per second counter cpuinfo_cycles_per_second = static_cast(EstimateCyclesPerSecond()); #endif } } // end namespace // getrusage() based implementation of MyCPUUsage static double MyCPUUsageRUsage() { #ifndef BENCHMARK_OS_WINDOWS struct rusage ru; if (getrusage(RUSAGE_SELF, &ru) == 0) { return (static_cast(ru.ru_utime.tv_sec) + static_cast(ru.ru_utime.tv_usec) * 1e-6 + static_cast(ru.ru_stime.tv_sec) + static_cast(ru.ru_stime.tv_usec) * 1e-6); } else { return 0.0; } #else HANDLE proc = GetCurrentProcess(); FILETIME creation_time; FILETIME exit_time; FILETIME kernel_time; FILETIME user_time; ULARGE_INTEGER kernel; ULARGE_INTEGER user; GetProcessTimes(proc, &creation_time, &exit_time, &kernel_time, &user_time); kernel.HighPart = kernel_time.dwHighDateTime; kernel.LowPart = kernel_time.dwLowDateTime; user.HighPart = user_time.dwHighDateTime; user.LowPart = user_time.dwLowDateTime; return (static_cast(kernel.QuadPart) + static_cast(user.QuadPart)) * 1e-7; #endif // OS_WINDOWS } #ifndef BENCHMARK_OS_WINDOWS static bool MyCPUUsageCPUTimeNsLocked(double* cputime) { static int cputime_fd = -1; if (cputime_fd == -1) { cputime_fd = open("/proc/self/cputime_ns", O_RDONLY); if (cputime_fd < 0) { cputime_fd = -1; return false; } } char buff[64]; memset(buff, 0, sizeof(buff)); if (pread(cputime_fd, buff, sizeof(buff) - 1, 0) <= 0) { close(cputime_fd); cputime_fd = -1; return false; } unsigned long long result = strtoull(buff, nullptr, 0); if (result == (std::numeric_limits::max)()) { close(cputime_fd); cputime_fd = -1; return false; } *cputime = static_cast(result) / 1e9; return true; } #endif // OS_WINDOWS double MyCPUUsage() { #ifndef BENCHMARK_OS_WINDOWS { std::lock_guard l(cputimens_mutex); static bool use_cputime_ns = true; if (use_cputime_ns) { double value; if (MyCPUUsageCPUTimeNsLocked(&value)) { return value; } // Once MyCPUUsageCPUTimeNsLocked fails once fall back to getrusage(). VLOG(1) << "Reading /proc/self/cputime_ns failed. Using getrusage().\n"; use_cputime_ns = false; } } #endif // OS_WINDOWS return MyCPUUsageRUsage(); } double ChildrenCPUUsage() { #ifndef BENCHMARK_OS_WINDOWS struct rusage ru; if (getrusage(RUSAGE_CHILDREN, &ru) == 0) { return (static_cast(ru.ru_utime.tv_sec) + static_cast(ru.ru_utime.tv_usec) * 1e-6 + static_cast(ru.ru_stime.tv_sec) + static_cast(ru.ru_stime.tv_usec) * 1e-6); } else { return 0.0; } #else // TODO: Not sure what this even means on Windows return 0.0; #endif // OS_WINDOWS } double CyclesPerSecond(void) { std::call_once(cpuinfo_init, InitializeSystemInfo); return cpuinfo_cycles_per_second; } int NumCPUs(void) { std::call_once(cpuinfo_init, InitializeSystemInfo); return cpuinfo_num_cpus; } // The ""'s catch people who don't pass in a literal for "str" #define strliterallen(str) (sizeof("" str "") - 1) // Must use a string literal for prefix. #define memprefix(str, len, prefix) \ ((((len) >= strliterallen(prefix)) && \ std::memcmp(str, prefix, strliterallen(prefix)) == 0) \ ? str + strliterallen(prefix) \ : nullptr) bool CpuScalingEnabled() { #ifndef BENCHMARK_OS_WINDOWS // On Linux, the CPUfreq subsystem exposes CPU information as files on the // local file system. If reading the exported files fails, then we may not be // running on Linux, so we silently ignore all the read errors. for (int cpu = 0, num_cpus = NumCPUs(); cpu < num_cpus; ++cpu) { std::string governor_file = StrCat("/sys/devices/system/cpu/cpu", cpu, "/cpufreq/scaling_governor"); FILE* file = fopen(governor_file.c_str(), "r"); if (!file) break; char buff[16]; size_t bytes_read = fread(buff, 1, sizeof(buff), file); fclose(file); if (memprefix(buff, bytes_read, "performance") == nullptr) return true; } #endif return false; } } // end namespace benchmark