ossrs
issue #3978: improve st_utime's default impl.
try to fix #3978
Background check #3978
Research
I referred the Android platform's solution, because I have android background, and there is a loop to handle message inside android.
https://github.com/aosp-mirror/platform_frameworks_base/blob/ff007a03c01bf936d1e961a13adff9f266d5189c/core/java/android/os/Handler.java#L701-L706C6
public final boolean sendMessageDelayed(@NonNull Message msg, long delayMillis) {
if (delayMillis < 0) {
delayMillis = 0;
}
return sendMessageAtTime(msg, SystemClock.uptimeMillis() + delayMillis);
}
https://github.com/aosp-mirror/platform_system_core/blob/59d9dc1f50b1ae8630ec11a431858a3cb66487b7/libutils/SystemClock.cpp#L37-L51
/*
* native public static long uptimeMillis();
*/
int64_t uptimeMillis()
{
return nanoseconds_to_milliseconds(uptimeNanos());
}
/*
* public static native long uptimeNanos();
*/
int64_t uptimeNanos()
{
return systemTime(SYSTEM_TIME_MONOTONIC);
}
https://github.com/aosp-mirror/platform_system_core/blob/59d9dc1f50b1ae8630ec11a431858a3cb66487b7/libutils/Timers.cpp#L32-L55
#if defined(__linux__)
nsecs_t systemTime(int clock) {
checkClockId(clock);
static constexpr clockid_t clocks[] = {CLOCK_REALTIME, CLOCK_MONOTONIC,
CLOCK_PROCESS_CPUTIME_ID, CLOCK_THREAD_CPUTIME_ID,
CLOCK_BOOTTIME};
static_assert(clock_id_max == arraysize(clocks));
timespec t = {};
clock_gettime(clocks[clock], &t);
return nsecs_t(t.tv_sec)*1000000000LL + t.tv_nsec;
}
#else
nsecs_t systemTime(int clock) {
// TODO: is this ever called with anything but REALTIME on mac/windows?
checkClockId(clock);
// Clock support varies widely across hosts. Mac OS doesn't support
// CLOCK_BOOTTIME (and doesn't even have clock_gettime until 10.12).
// Windows is windows.
timeval t = {};
gettimeofday(&t, nullptr);
return nsecs_t(t.tv_sec)*1000000000LL + nsecs_t(t.tv_usec)*1000LL;
}
#endif
For Linux system, we can use clock_gettime api, but it's first appeared in Mac OSX 10.12.
man clock_gettime
The requirement is to find an alternative way to get the timestamp in microsecond unit, but the clock_gettime get nanoseconds, the math formula is the nanoseconds / 1000 = microsecond. Then I check the performance of this api + math division.
I used those code to check the clock_gettime performance.
#include <sys/time.h>
#include <time.h>
#include <stdio.h>
#include <unistd.h>
int main() {
struct timeval tv;
struct timespec ts;
clock_t start;
clock_t end;
long t;
while (1) {
start = clock();
gettimeofday(&tv, NULL);
end = clock();
printf("gettimeofday clock is %lu\n", end - start);
printf("gettimeofday is %lld\n", (tv.tv_sec * 1000000LL + tv.tv_usec));
start = clock();
clock_gettime(CLOCK_MONOTONIC, &ts);
t = ts.tv_sec * 1000000L + ts.tv_nsec / 1000L;
end = clock();
printf("clock_monotonic clock is %lu\n", end - start);
printf("clock_monotonic: seconds is %ld, nanoseconds is %ld, sum is %ld\n", ts.tv_sec, ts.tv_nsec, t);
start = clock();
clock_gettime(CLOCK_MONOTONIC_RAW, &ts);
t = ts.tv_sec * 1000000L + ts.tv_nsec / 1000L;
end = clock();
printf("clock_monotonic_raw clock is %lu\n", end - start);
printf("clock_monotonic_raw: nanoseconds is %ld, sum is %ld\n", ts.tv_nsec, t);
sleep(3);
}
return 0;
}
Here is output:
env: Mac OS M2 chip.
gettimeofday clock is 11
gettimeofday is 1709775727153949
clock_monotonic clock is 2
clock_monotonic: seconds is 1525204, nanoseconds is 409453000, sum is 1525204409453
clock_monotonic_raw clock is 2
clock_monotonic_raw: nanoseconds is 770493000, sum is 1525222770493
We can see the clock_gettime is faster than gettimeofday, so there are no performance risks.
MacOS solution
clock_gettime api only available until mac os 10.12, for the mac os older than 10.12, just keep the gettimeofday.
check osx version in auto/options.sh, then add MACRO in auto/depends.sh, the MACRO is MD_OSX_HAS_NO_CLOCK_GETTIME.
CYGWIN
According to google search, it seems the clock_gettime(CLOCK_MONOTONIC) is not support well at least 10 years ago, but I didn't own an windows machine, so can't verify it. so keep win's solution.
Thank you for your excellent work. Please add new tests to validate your updates in the pull request. The utest is crucial and a necessary condition for merging the pull request. Without it and automated testing, we risk introducing bugs sooner or later due to many changes over time.
