diff options
Diffstat (limited to '')
-rw-r--r-- | libnetdata/clocks/clocks.c | 301 | ||||
-rw-r--r-- | libnetdata/clocks/clocks.h | 41 |
2 files changed, 232 insertions, 110 deletions
diff --git a/libnetdata/clocks/clocks.c b/libnetdata/clocks/clocks.c index 4ec5fa98b..f0e17b232 100644 --- a/libnetdata/clocks/clocks.c +++ b/libnetdata/clocks/clocks.c @@ -2,8 +2,13 @@ #include "../libnetdata.h" -static int clock_boottime_valid = 1; -static int clock_monotonic_coarse_valid = 1; +// defaults are for compatibility +// call clocks_init() once, to optimize these default settings +static clockid_t clock_boottime_to_use = CLOCK_MONOTONIC; +static clockid_t clock_monotonic_to_use = CLOCK_MONOTONIC; + +usec_t clock_monotonic_resolution = 1000; +usec_t clock_realtime_resolution = 1000; #ifndef HAVE_CLOCK_GETTIME inline int clock_gettime(clockid_t clk_id, struct timespec *ts) { @@ -18,19 +23,60 @@ inline int clock_gettime(clockid_t clk_id, struct timespec *ts) { } #endif -void test_clock_boottime(void) { +// Similar to CLOCK_MONOTONIC, but provides access to a raw hardware-based time that is not subject to NTP adjustments +// or the incremental adjustments performed by adjtime(3). This clock does not count time that the system is suspended + +static void test_clock_monotonic_raw(void) { +#ifdef CLOCK_MONOTONIC_RAW + struct timespec ts; + if(clock_gettime(CLOCK_MONOTONIC_RAW, &ts) == -1 && errno == EINVAL) + clock_monotonic_to_use = CLOCK_MONOTONIC; + else + clock_monotonic_to_use = CLOCK_MONOTONIC_RAW; +#else + clock_monotonic_to_use = CLOCK_MONOTONIC; +#endif +} + +// When running a binary with CLOCK_BOOTTIME defined on a system with a linux kernel older than Linux 2.6.39 the +// clock_gettime(2) system call fails with EINVAL. In that case it must fall-back to CLOCK_MONOTONIC. + +static void test_clock_boottime(void) { struct timespec ts; if(clock_gettime(CLOCK_BOOTTIME, &ts) == -1 && errno == EINVAL) - clock_boottime_valid = 0; + clock_boottime_to_use = clock_monotonic_to_use; + else + clock_boottime_to_use = CLOCK_BOOTTIME; } -void test_clock_monotonic_coarse(void) { +static usec_t get_clock_resolution(clockid_t clock) { struct timespec ts; - if(clock_gettime(CLOCK_MONOTONIC_COARSE, &ts) == -1 && errno == EINVAL) - clock_monotonic_coarse_valid = 0; + clock_getres(clock, &ts); + return ts.tv_sec * USEC_PER_SEC + ts.tv_nsec * NSEC_PER_USEC; +} + +// perform any initializations required for clocks + +void clocks_init(void) { + // monotonic raw has to be tested before boottime + test_clock_monotonic_raw(); + + // boottime has to be tested after monotonic coarse + test_clock_boottime(); + + clock_monotonic_resolution = get_clock_resolution(clock_monotonic_to_use); + clock_realtime_resolution = get_clock_resolution(CLOCK_REALTIME); + + // if for any reason these are zero, netdata will crash + // since we use them as modulo to calculations + if(!clock_realtime_resolution) + clock_realtime_resolution = 1000; + + if(!clock_monotonic_resolution) + clock_monotonic_resolution = 1000; } -static inline time_t now_sec(clockid_t clk_id) { +inline time_t now_sec(clockid_t clk_id) { struct timespec ts; if(unlikely(clock_gettime(clk_id, &ts) == -1)) { error("clock_gettime(%d, ×pec) failed.", clk_id); @@ -39,7 +85,7 @@ static inline time_t now_sec(clockid_t clk_id) { return ts.tv_sec; } -static inline usec_t now_usec(clockid_t clk_id) { +inline usec_t now_usec(clockid_t clk_id) { struct timespec ts; if(unlikely(clock_gettime(clk_id, &ts) == -1)) { error("clock_gettime(%d, ×pec) failed.", clk_id); @@ -48,7 +94,7 @@ static inline usec_t now_usec(clockid_t clk_id) { return (usec_t)ts.tv_sec * USEC_PER_SEC + (ts.tv_nsec % NSEC_PER_SEC) / NSEC_PER_USEC; } -static inline int now_timeval(clockid_t clk_id, struct timeval *tv) { +inline int now_timeval(clockid_t clk_id, struct timeval *tv) { struct timespec ts; if(unlikely(clock_gettime(clk_id, &ts) == -1)) { @@ -76,15 +122,15 @@ inline int now_realtime_timeval(struct timeval *tv) { } inline time_t now_monotonic_sec(void) { - return now_sec(likely(clock_monotonic_coarse_valid) ? CLOCK_MONOTONIC_COARSE : CLOCK_MONOTONIC); + return now_sec(clock_monotonic_to_use); } inline usec_t now_monotonic_usec(void) { - return now_usec(likely(clock_monotonic_coarse_valid) ? CLOCK_MONOTONIC_COARSE : CLOCK_MONOTONIC); + return now_usec(clock_monotonic_to_use); } inline int now_monotonic_timeval(struct timeval *tv) { - return now_timeval(likely(clock_monotonic_coarse_valid) ? CLOCK_MONOTONIC_COARSE : CLOCK_MONOTONIC, tv); + return now_timeval(clock_monotonic_to_use, tv); } inline time_t now_monotonic_high_precision_sec(void) { @@ -100,19 +146,15 @@ inline int now_monotonic_high_precision_timeval(struct timeval *tv) { } inline time_t now_boottime_sec(void) { - return now_sec(likely(clock_boottime_valid) ? CLOCK_BOOTTIME : - likely(clock_monotonic_coarse_valid) ? CLOCK_MONOTONIC_COARSE : CLOCK_MONOTONIC); + return now_sec(clock_boottime_to_use); } inline usec_t now_boottime_usec(void) { - return now_usec(likely(clock_boottime_valid) ? CLOCK_BOOTTIME : - likely(clock_monotonic_coarse_valid) ? CLOCK_MONOTONIC_COARSE : CLOCK_MONOTONIC); + return now_usec(clock_boottime_to_use); } inline int now_boottime_timeval(struct timeval *tv) { - return now_timeval(likely(clock_boottime_valid) ? CLOCK_BOOTTIME : - likely(clock_monotonic_coarse_valid) ? CLOCK_MONOTONIC_COARSE : CLOCK_MONOTONIC, - tv); + return now_timeval(clock_boottime_to_use, tv); } inline usec_t timeval_usec(struct timeval *tv) { @@ -137,9 +179,113 @@ inline usec_t dt_usec(struct timeval *now, struct timeval *old) { return (ts1 > ts2) ? (ts1 - ts2) : (ts2 - ts1); } -inline void heartbeat_init(heartbeat_t *hb) -{ - hb->monotonic = hb->realtime = 0ULL; +#ifdef __linux__ +void sleep_to_absolute_time(usec_t usec) { + static int einval_printed = 0, enotsup_printed = 0, eunknown_printed = 0; + clockid_t clock = CLOCK_REALTIME; + + struct timespec req = { + .tv_sec = (time_t)(usec / USEC_PER_SEC), + .tv_nsec = (suseconds_t)((usec % USEC_PER_SEC) * NSEC_PER_USEC) + }; + + int ret = 0; + while( (ret = clock_nanosleep(clock, TIMER_ABSTIME, &req, NULL)) != 0 ) { + if(ret == EINTR) continue; + else { + if (ret == EINVAL) { + if (!einval_printed) { + einval_printed++; + error( + "Invalid time given to clock_nanosleep(): clockid = %d, tv_sec = %ld, tv_nsec = %ld", + clock, + req.tv_sec, + req.tv_nsec); + } + } else if (ret == ENOTSUP) { + if (!enotsup_printed) { + enotsup_printed++; + error( + "Invalid clock id given to clock_nanosleep(): clockid = %d, tv_sec = %ld, tv_nsec = %ld", + clock, + req.tv_sec, + req.tv_nsec); + } + } else { + if (!eunknown_printed) { + eunknown_printed++; + error( + "Unknown return value %d from clock_nanosleep(): clockid = %d, tv_sec = %ld, tv_nsec = %ld", + ret, + clock, + req.tv_sec, + req.tv_nsec); + } + } + sleep_usec(usec); + } + } +}; +#endif + +#define HEARTBEAT_ALIGNMENT_STATISTICS_SIZE 10 +netdata_mutex_t heartbeat_alignment_mutex = NETDATA_MUTEX_INITIALIZER; +static size_t heartbeat_alignment_id = 0; + +struct heartbeat_thread_statistics { + size_t sequence; + usec_t dt; +}; +static struct heartbeat_thread_statistics heartbeat_alignment_values[HEARTBEAT_ALIGNMENT_STATISTICS_SIZE] = { 0 }; + +void heartbeat_statistics(usec_t *min_ptr, usec_t *max_ptr, usec_t *average_ptr, size_t *count_ptr) { + struct heartbeat_thread_statistics current[HEARTBEAT_ALIGNMENT_STATISTICS_SIZE]; + static struct heartbeat_thread_statistics old[HEARTBEAT_ALIGNMENT_STATISTICS_SIZE] = { 0 }; + + memcpy(current, heartbeat_alignment_values, sizeof(struct heartbeat_thread_statistics) * HEARTBEAT_ALIGNMENT_STATISTICS_SIZE); + + usec_t min = 0, max = 0, total = 0, average = 0; + size_t i, count = 0; + for(i = 0; i < HEARTBEAT_ALIGNMENT_STATISTICS_SIZE ;i++) { + if(current[i].sequence == old[i].sequence) continue; + usec_t value = current[i].dt - old[i].dt; + + if(!count) { + min = max = total = value; + count = 1; + } + else { + total += value; + if(value < min) min = value; + if(value > max) max = value; + count++; + } + } + if(count) + average = total / count; + + if(min_ptr) *min_ptr = min; + if(max_ptr) *max_ptr = max; + if(average_ptr) *average_ptr = average; + if(count_ptr) *count_ptr = count; + + memcpy(old, current, sizeof(struct heartbeat_thread_statistics) * HEARTBEAT_ALIGNMENT_STATISTICS_SIZE); +} + +inline void heartbeat_init(heartbeat_t *hb) { + hb->realtime = 0ULL; + hb->randomness = 250 * USEC_PER_MS + ((now_realtime_usec() * clock_realtime_resolution) % (250 * USEC_PER_MS)); + hb->randomness -= (hb->randomness % clock_realtime_resolution); + + netdata_mutex_lock(&heartbeat_alignment_mutex); + hb->statistics_id = heartbeat_alignment_id; + heartbeat_alignment_id++; + netdata_mutex_unlock(&heartbeat_alignment_mutex); + + if(hb->statistics_id < HEARTBEAT_ALIGNMENT_STATISTICS_SIZE) { + heartbeat_alignment_values[hb->statistics_id].dt = 0; + heartbeat_alignment_values[hb->statistics_id].sequence = 0; + } } // waits for the next heartbeat @@ -147,96 +293,81 @@ inline void heartbeat_init(heartbeat_t *hb) // it returns the dt using the realtime clock usec_t heartbeat_next(heartbeat_t *hb, usec_t tick) { - heartbeat_t now; - now.monotonic = now_monotonic_usec(); - now.realtime = now_realtime_usec(); - - usec_t next_monotonic = now.monotonic - (now.monotonic % tick) + tick; - - while(now.monotonic < next_monotonic) { - sleep_usec(next_monotonic - now.monotonic); - now.monotonic = now_monotonic_usec(); - now.realtime = now_realtime_usec(); + if(unlikely(hb->randomness > tick / 2)) { + // TODO: The heartbeat tick should be specified at the heartbeat_init() function + usec_t tmp = (now_realtime_usec() * clock_realtime_resolution) % (tick / 2); + info("heartbeat randomness of %llu is too big for a tick of %llu - setting it to %llu", hb->randomness, tick, tmp); + hb->randomness = tmp; } - if(likely(hb->realtime != 0ULL)) { - usec_t dt_monotonic = now.monotonic - hb->monotonic; - usec_t dt_realtime = now.realtime - hb->realtime; + usec_t dt; + usec_t now = now_realtime_usec(); + usec_t next = now - (now % tick) + tick + hb->randomness; - hb->monotonic = now.monotonic; - hb->realtime = now.realtime; + // align the next time we want to the clock resolution + if(next % clock_realtime_resolution) + next = next - (next % clock_realtime_resolution) + clock_realtime_resolution; - if(unlikely(dt_monotonic >= tick + tick / 2)) { - errno = 0; - error("heartbeat missed %llu monotonic microseconds", dt_monotonic - tick); - } + // sleep_usec() has a loop to guarantee we will sleep for at least the requested time. + // According the specs, when we sleep for a relative time, clock adjustments should not affect the duration + // we sleep. + sleep_usec(next - now); + now = now_realtime_usec(); + dt = now - hb->realtime; - return dt_realtime; + if(hb->statistics_id < HEARTBEAT_ALIGNMENT_STATISTICS_SIZE) { + heartbeat_alignment_values[hb->statistics_id].dt += now - next; + heartbeat_alignment_values[hb->statistics_id].sequence++; } - else { - hb->monotonic = now.monotonic; - hb->realtime = now.realtime; - return 0ULL; + + if(unlikely(now < next)) { + errno = 0; + error("heartbeat clock: woke up %llu microseconds earlier than expected (can be due to the CLOCK_REALTIME set to the past).", next - now); + } + else if(unlikely(now - next > tick / 2)) { + errno = 0; + error("heartbeat clock: woke up %llu microseconds later than expected (can be due to system load or the CLOCK_REALTIME set to the future).", now - next); } -} -// returned the elapsed time, since the last heartbeat -// using the monotonic clock + if(unlikely(!hb->realtime)) { + // the first time return zero + dt = 0; + } -inline usec_t heartbeat_monotonic_dt_to_now_usec(heartbeat_t *hb) { - if(!hb || !hb->monotonic) return 0ULL; - return now_monotonic_usec() - hb->monotonic; + hb->realtime = now; + return dt; } -int sleep_usec(usec_t usec) { - -#ifndef NETDATA_WITH_USLEEP +void sleep_usec(usec_t usec) { // we expect microseconds (1.000.000 per second) // but timespec is nanoseconds (1.000.000.000 per second) struct timespec rem, req = { - .tv_sec = (time_t) (usec / 1000000), - .tv_nsec = (suseconds_t) ((usec % 1000000) * 1000) + .tv_sec = (time_t) (usec / USEC_PER_SEC), + .tv_nsec = (suseconds_t) ((usec % USEC_PER_SEC) * NSEC_PER_USEC) }; - while (nanosleep(&req, &rem) == -1) { +#ifdef __linux__ + while ((errno = clock_nanosleep(CLOCK_REALTIME, 0, &req, &rem)) != 0) { +#else + while ((errno = nanosleep(&req, &rem)) != 0) { +#endif if (likely(errno == EINTR)) { - debug(D_SYSTEM, "nanosleep() interrupted (while sleeping for %llu microseconds).", usec); req.tv_sec = rem.tv_sec; req.tv_nsec = rem.tv_nsec; } else { +#ifdef __linux__ + error("Cannot clock_nanosleep(CLOCK_REALTIME) for %llu microseconds.", usec); +#else error("Cannot nanosleep() for %llu microseconds.", usec); +#endif break; } } - - return 0; -#else - int ret = usleep(usec); - if(unlikely(ret == -1 && errno == EINVAL)) { - // on certain systems, usec has to be up to 999999 - if(usec > 999999) { - int counter = usec / 999999; - while(counter--) - usleep(999999); - - usleep(usec % 999999); - } - else { - error("Cannot usleep() for %llu microseconds.", usec); - return ret; - } - } - - if(ret != 0) - error("usleep() failed for %llu microseconds.", usec); - - return ret; -#endif } static inline collected_number uptime_from_boottime(void) { #ifdef CLOCK_BOOTTIME_IS_AVAILABLE - return now_boottime_usec() / 1000; + return (collected_number)(now_boottime_usec() / USEC_PER_MS); #else error("uptime cannot be read from CLOCK_BOOTTIME on this system."); return 0; diff --git a/libnetdata/clocks/clocks.h b/libnetdata/clocks/clocks.h index d3475df6a..53c036ece 100644 --- a/libnetdata/clocks/clocks.h +++ b/libnetdata/clocks/clocks.h @@ -22,8 +22,9 @@ typedef unsigned long long usec_t; typedef long long susec_t; typedef struct heartbeat { - usec_t monotonic; usec_t realtime; + usec_t randomness; + size_t statistics_id; } heartbeat_t; /* Linux value is as good as any other */ @@ -36,20 +37,14 @@ typedef struct heartbeat { #define CLOCK_MONOTONIC CLOCK_REALTIME #endif -/* Prefer CLOCK_MONOTONIC_COARSE where available to reduce overhead. It has the same semantics as CLOCK_MONOTONIC */ -#ifndef CLOCK_MONOTONIC_COARSE -/* fallback to CLOCK_MONOTONIC if not available */ -#define CLOCK_MONOTONIC_COARSE CLOCK_MONOTONIC -#endif - #ifndef CLOCK_BOOTTIME #ifdef CLOCK_UPTIME /* CLOCK_BOOTTIME falls back to CLOCK_UPTIME on FreeBSD */ #define CLOCK_BOOTTIME CLOCK_UPTIME #else // CLOCK_UPTIME -/* CLOCK_BOOTTIME falls back to CLOCK_MONOTONIC */ -#define CLOCK_BOOTTIME CLOCK_MONOTONIC_COARSE +/* CLOCK_BOOTTIME falls back to CLOCK_REALTIME */ +#define CLOCK_BOOTTIME CLOCK_REALTIME #endif // CLOCK_UPTIME #else // CLOCK_BOOTTIME @@ -115,8 +110,6 @@ extern int clock_gettime(clockid_t clk_id, struct timespec *ts); * All now_*_sec() functions return the time in seconds from the appropriate clock, or 0 on error. * All now_*_usec() functions return the time in microseconds from the appropriate clock, or 0 on error. * - * Most functions will attempt to use CLOCK_MONOTONIC_COARSE if available to reduce contention overhead and improve - * performance scaling. If high precision is required please use one of the available now_*_high_precision_* functions. */ extern int now_realtime_timeval(struct timeval *tv); extern time_t now_realtime_sec(void); @@ -133,7 +126,6 @@ extern int now_boottime_timeval(struct timeval *tv); extern time_t now_boottime_sec(void); extern usec_t now_boottime_usec(void); - extern usec_t timeval_usec(struct timeval *tv); extern msec_t timeval_msec(struct timeval *tv); @@ -147,23 +139,22 @@ extern void heartbeat_init(heartbeat_t *hb); */ extern usec_t heartbeat_next(heartbeat_t *hb, usec_t tick); -/* Returns elapsed time in microseconds since last heartbeat */ -extern usec_t heartbeat_monotonic_dt_to_now_usec(heartbeat_t *hb); +extern void heartbeat_statistics(usec_t *min_ptr, usec_t *max_ptr, usec_t *average_ptr, size_t *count_ptr); -extern int sleep_usec(usec_t usec); +extern void sleep_usec(usec_t usec); -/* - * When running a binary with CLOCK_BOOTTIME defined on a system with a linux kernel older than Linux 2.6.39 the - * clock_gettime(2) system call fails with EINVAL. In that case it must fall-back to CLOCK_MONOTONIC. - */ -void test_clock_boottime(void); +extern void clocks_init(void); -/* - * When running a binary with CLOCK_MONOTONIC_COARSE defined on a system with a linux kernel older than Linux 2.6.32 the - * clock_gettime(2) system call fails with EINVAL. In that case it must fall-back to CLOCK_MONOTONIC. - */ -void test_clock_monotonic_coarse(void); +// lower level functions - avoid using directly +extern time_t now_sec(clockid_t clk_id); +extern usec_t now_usec(clockid_t clk_id); +extern int now_timeval(clockid_t clk_id, struct timeval *tv); extern collected_number uptime_msec(char *filename); +extern usec_t clock_monotonic_resolution; +extern usec_t clock_realtime_resolution; + +extern void sleep_to_absolute_time(usec_t usec); + #endif /* NETDATA_CLOCKS_H */ |