chromium/third_party/abseil-cpp/absl/base/internal/sysinfo.cc

// Copyright 2017 The Abseil Authors.
//
// 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
//
//      https://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 "absl/base/internal/sysinfo.h"

#include "absl/base/attributes.h"

#ifdef _WIN32
#include <windows.h>
#else
#include <fcntl.h>
#include <pthread.h>
#include <sys/stat.h>
#include <sys/types.h>
#include <unistd.h>
#endif

#ifdef __linux__
#include <sys/syscall.h>
#endif

#if defined(__APPLE__) || defined(__FreeBSD__)
#include <sys/sysctl.h>
#endif

#ifdef __FreeBSD__
#include <pthread_np.h>
#endif

#ifdef __NetBSD__
#include <lwp.h>
#endif

#if defined(__myriad2__)
#include <rtems.h>
#endif

#if defined(__Fuchsia__)
#include <zircon/process.h>
#endif

#include <string.h>

#include <cassert>
#include <cerrno>
#include <cstdint>
#include <cstdio>
#include <cstdlib>
#include <ctime>
#include <limits>
#include <thread>  // NOLINT(build/c++11)
#include <utility>
#include <vector>

#include "absl/base/call_once.h"
#include "absl/base/config.h"
#include "absl/base/internal/raw_logging.h"
#include "absl/base/internal/spinlock.h"
#include "absl/base/internal/unscaledcycleclock.h"
#include "absl/base/thread_annotations.h"

namespace absl {
ABSL_NAMESPACE_BEGIN
namespace base_internal {

namespace {

#if defined(_WIN32)

// Returns number of bits set in `bitMask`
DWORD Win32CountSetBits(ULONG_PTR bitMask) {
  for (DWORD bitSetCount = 0; ; ++bitSetCount) {
    if (bitMask == 0) return bitSetCount;
    bitMask &= bitMask - 1;
  }
}

// Returns the number of logical CPUs using GetLogicalProcessorInformation(), or
// 0 if the number of processors is not available or can not be computed.
// https://docs.microsoft.com/en-us/windows/win32/api/sysinfoapi/nf-sysinfoapi-getlogicalprocessorinformation
int Win32NumCPUs() {
#pragma comment(lib, "kernel32.lib")
  using Info = SYSTEM_LOGICAL_PROCESSOR_INFORMATION;

  DWORD info_size = sizeof(Info);
  Info* info(static_cast<Info*>(malloc(info_size)));
  if (info == nullptr) return 0;

  bool success = GetLogicalProcessorInformation(info, &info_size);
  if (!success && GetLastError() == ERROR_INSUFFICIENT_BUFFER) {
    free(info);
    info = static_cast<Info*>(malloc(info_size));
    if (info == nullptr) return 0;
    success = GetLogicalProcessorInformation(info, &info_size);
  }

  DWORD logicalProcessorCount = 0;
  if (success) {
    Info* ptr = info;
    DWORD byteOffset = 0;
    while (byteOffset + sizeof(Info) <= info_size) {
      switch (ptr->Relationship) {
        case RelationProcessorCore:
          logicalProcessorCount += Win32CountSetBits(ptr->ProcessorMask);
          break;

        case RelationNumaNode:
        case RelationCache:
        case RelationProcessorPackage:
          // Ignore other entries
          break;

        default:
          // Ignore unknown entries
          break;
      }
      byteOffset += sizeof(Info);
      ptr++;
    }
  }
  free(info);
  return static_cast<int>(logicalProcessorCount);
}

#endif

}  // namespace

static int GetNumCPUs() { … }

#if defined(_WIN32)

static double GetNominalCPUFrequency() {
#if WINAPI_FAMILY_PARTITION(WINAPI_PARTITION_APP) && \
    !WINAPI_FAMILY_PARTITION(WINAPI_PARTITION_DESKTOP)
  // UWP apps don't have access to the registry and currently don't provide an
  // API informing about CPU nominal frequency.
  return 1.0;
#else
#pragma comment(lib, "advapi32.lib")  // For Reg* functions.
  HKEY key;
  // Use the Reg* functions rather than the SH functions because shlwapi.dll
  // pulls in gdi32.dll which makes process destruction much more costly.
  if (RegOpenKeyExA(HKEY_LOCAL_MACHINE,
                    "HARDWARE\\DESCRIPTION\\System\\CentralProcessor\\0", 0,
                    KEY_READ, &key) == ERROR_SUCCESS) {
    DWORD type = 0;
    DWORD data = 0;
    DWORD data_size = sizeof(data);
    auto result = RegQueryValueExA(key, "~MHz", nullptr, &type,
                                   reinterpret_cast<LPBYTE>(&data), &data_size);
    RegCloseKey(key);
    if (result == ERROR_SUCCESS && type == REG_DWORD &&
        data_size == sizeof(data)) {
      return data * 1e6;  // Value is MHz.
    }
  }
  return 1.0;
#endif  // WINAPI_PARTITION_APP && !WINAPI_PARTITION_DESKTOP
}

#elif defined(CTL_HW) && defined(HW_CPU_FREQ)

static double GetNominalCPUFrequency() {
  unsigned freq;
  size_t size = sizeof(freq);
  int mib[2] = {CTL_HW, HW_CPU_FREQ};
  if (sysctl(mib, 2, &freq, &size, nullptr, 0) == 0) {
    return static_cast<double>(freq);
  }
  return 1.0;
}

#else

// Helper function for reading a long from a file. Returns true if successful
// and the memory location pointed to by value is set to the value read.
static bool ReadLongFromFile(const char *file, long *value) { … }

#if defined(ABSL_INTERNAL_UNSCALED_CYCLECLOCK_FREQUENCY_IS_CPU_FREQUENCY)

// Reads a monotonic time source and returns a value in
// nanoseconds. The returned value uses an arbitrary epoch, not the
// Unix epoch.
static int64_t ReadMonotonicClockNanos() { … }

class UnscaledCycleClockWrapperForInitializeFrequency { … };

struct TimeTscPair { … };

// Returns a pair of values (monotonic kernel time, TSC ticks) that
// approximately correspond to each other.  This is accomplished by
// doing several reads and picking the reading with the lowest
// latency.  This approach is used to minimize the probability that
// our thread was preempted between clock reads.
static TimeTscPair GetTimeTscPair() { … }

// Measures and returns the TSC frequency by taking a pair of
// measurements approximately `sleep_nanoseconds` apart.
static double MeasureTscFrequencyWithSleep(int sleep_nanoseconds) { … }

// Measures and returns the TSC frequency by calling
// MeasureTscFrequencyWithSleep(), doubling the sleep interval until the
// frequency measurement stabilizes.
static double MeasureTscFrequency() { … }

#endif  // ABSL_INTERNAL_UNSCALED_CYCLECLOCK_FREQUENCY_IS_CPU_FREQUENCY

static double GetNominalCPUFrequency() { … }

#endif

ABSL_CONST_INIT static once_flag init_num_cpus_once;
ABSL_CONST_INIT static int num_cpus = …;

// NumCPUs() may be called before main() and before malloc is properly
// initialized, therefore this must not allocate memory.
int NumCPUs() { … }

// A default frequency of 0.0 might be dangerous if it is used in division.
ABSL_CONST_INIT static once_flag init_nominal_cpu_frequency_once;
ABSL_CONST_INIT static double nominal_cpu_frequency = …;

// NominalCPUFrequency() may be called before main() and before malloc is
// properly initialized, therefore this must not allocate memory.
double NominalCPUFrequency() { … }

#if defined(_WIN32)

pid_t GetTID() {
  return pid_t{GetCurrentThreadId()};
}

#elif defined(__linux__)

#ifndef SYS_gettid
#define SYS_gettid …
#endif

pid_t GetTID() { … }

#elif defined(__akaros__)

pid_t GetTID() {
  // Akaros has a concept of "vcore context", which is the state the program
  // is forced into when we need to make a user-level scheduling decision, or
  // run a signal handler.  This is analogous to the interrupt context that a
  // CPU might enter if it encounters some kind of exception.
  //
  // There is no current thread context in vcore context, but we need to give
  // a reasonable answer if asked for a thread ID (e.g., in a signal handler).
  // Thread 0 always exists, so if we are in vcore context, we return that.
  //
  // Otherwise, we know (since we are using pthreads) that the uthread struct
  // current_uthread is pointing to is the first element of a
  // struct pthread_tcb, so we extract and return the thread ID from that.
  //
  // TODO(dcross): Akaros anticipates moving the thread ID to the uthread
  // structure at some point. We should modify this code to remove the cast
  // when that happens.
  if (in_vcore_context())
    return 0;
  return reinterpret_cast<struct pthread_tcb *>(current_uthread)->id;
}

#elif defined(__myriad2__)

pid_t GetTID() {
  uint32_t tid;
  rtems_task_ident(RTEMS_SELF, 0, &tid);
  return tid;
}

#elif defined(__APPLE__)

pid_t GetTID() {
  uint64_t tid;
  // `nullptr` here implies this thread.  This only fails if the specified
  // thread is invalid or the pointer-to-tid is null, so we needn't worry about
  // it.
  pthread_threadid_np(nullptr, &tid);
  return static_cast<pid_t>(tid);
}

#elif defined(__FreeBSD__)

pid_t GetTID() { return static_cast<pid_t>(pthread_getthreadid_np()); }

#elif defined(__OpenBSD__)

pid_t GetTID() { return getthrid(); }

#elif defined(__NetBSD__)

pid_t GetTID() { return static_cast<pid_t>(_lwp_self()); }

#elif defined(__native_client__)

pid_t GetTID() {
  auto* thread = pthread_self();
  static_assert(sizeof(pid_t) == sizeof(thread),
                "In NaCL int expected to be the same size as a pointer");
  return reinterpret_cast<pid_t>(thread);
}

#elif defined(__Fuchsia__)

pid_t GetTID() {
  // Use our thread handle as the TID, which should be unique within this
  // process (but may not be globally unique). The handle value was chosen over
  // a kernel object ID (KOID) because zx_handle_t (32-bits) can be cast to a
  // pid_t type without loss of precision, but a zx_koid_t (64-bits) cannot.
  return static_cast<pid_t>(zx_thread_self());
}

#else

// Fallback implementation of `GetTID` using `pthread_self`.
pid_t GetTID() {
  // `pthread_t` need not be arithmetic per POSIX; platforms where it isn't
  // should be handled above.
  return static_cast<pid_t>(pthread_self());
}

#endif

// GetCachedTID() caches the thread ID in thread-local storage (which is a
// userspace construct) to avoid unnecessary system calls. Without this caching,
// it can take roughly 98ns, while it takes roughly 1ns with this caching.
pid_t GetCachedTID() { … }

}  // namespace base_internal
ABSL_NAMESPACE_END
}  // namespace absl