blob: 3b9e9ff64a4977cff597d7d8ca7b43fc5ca9042b [file] [log] [blame]
Willy Tarreau55542642021-10-08 09:33:24 +02001/*
2 * General time-keeping code and variables
3 *
4 * Copyright 2000-2021 Willy Tarreau <w@1wt.eu>
5 *
6 * This program is free software; you can redistribute it and/or
7 * modify it under the terms of the GNU General Public License
8 * as published by the Free Software Foundation; either version
9 * 2 of the License, or (at your option) any later version.
10 *
11 */
12
13#include <sys/time.h>
Willy Tarreau6cb0c392021-10-08 14:48:30 +020014#include <signal.h>
Willy Tarreau55542642021-10-08 09:33:24 +020015#include <time.h>
16
Willy Tarreau44c58da2021-10-08 12:27:54 +020017#ifdef USE_THREAD
18#include <pthread.h>
19#endif
20
Willy Tarreau55542642021-10-08 09:33:24 +020021#include <haproxy/api.h>
Willy Tarreauf9d5e102021-10-08 10:43:59 +020022#include <haproxy/activity.h>
Willy Tarreau55542642021-10-08 09:33:24 +020023#include <haproxy/clock.h>
Willy Tarreau6cb0c392021-10-08 14:48:30 +020024#include <haproxy/signal-t.h>
Willy Tarreau55542642021-10-08 09:33:24 +020025#include <haproxy/time.h>
26#include <haproxy/tinfo-t.h>
27#include <haproxy/tools.h>
28
29struct timeval start_date; /* the process's start date in wall-clock time */
30volatile ullong global_now; /* common monotonic date between all threads (32:32) */
31volatile uint global_now_ms; /* common monotonic date in milliseconds (may wrap) */
32
33THREAD_ALIGNED(64) static ullong now_offset; /* global offset between system time and global time */
34
35THREAD_LOCAL uint now_ms; /* internal monotonic date in milliseconds (may wrap) */
36THREAD_LOCAL struct timeval now; /* internal monotonic date derived from real clock */
37THREAD_LOCAL struct timeval date; /* the real current date (wall-clock time) */
Willy Tarreau55542642021-10-08 09:33:24 +020038
Willy Tarreau2c6a9982021-10-08 11:38:30 +020039static THREAD_LOCAL struct timeval before_poll; /* system date before calling poll() */
40static THREAD_LOCAL struct timeval after_poll; /* system date after leaving poll() */
Willy Tarreauf9d5e102021-10-08 10:43:59 +020041static THREAD_LOCAL unsigned int samp_time; /* total elapsed time over current sample */
42static THREAD_LOCAL unsigned int idle_time; /* total idle time over current sample */
Willy Tarreau55542642021-10-08 09:33:24 +020043static THREAD_LOCAL unsigned int iso_time_sec; /* last iso time value for this thread */
44static THREAD_LOCAL char iso_time_str[34]; /* ISO time representation of gettimeofday() */
45
Willy Tarreau21694982021-10-08 15:09:17 +020046#if defined(_POSIX_TIMERS) && (_POSIX_TIMERS > 0) && defined(_POSIX_THREAD_CPUTIME)
47static clockid_t per_thread_clock_id[MAX_THREADS];
48#endif
49
Willy Tarreau55542642021-10-08 09:33:24 +020050/* returns the system's monotonic time in nanoseconds if supported, otherwise zero */
51uint64_t now_mono_time(void)
52{
53 uint64_t ret = 0;
Willy Tarreau6cb0c392021-10-08 14:48:30 +020054#if defined(_POSIX_TIMERS) && defined(_POSIX_TIMERS) && (_POSIX_TIMERS > 0) && defined(_POSIX_MONOTONIC_CLOCK)
Willy Tarreau55542642021-10-08 09:33:24 +020055 struct timespec ts;
56 clock_gettime(CLOCK_MONOTONIC, &ts);
57 ret = ts.tv_sec * 1000000000ULL + ts.tv_nsec;
58#endif
59 return ret;
60}
61
62/* returns the current thread's cumulated CPU time in nanoseconds if supported, otherwise zero */
63uint64_t now_cpu_time(void)
64{
65 uint64_t ret = 0;
66#if defined(_POSIX_TIMERS) && (_POSIX_TIMERS > 0) && defined(_POSIX_THREAD_CPUTIME)
67 struct timespec ts;
68 clock_gettime(CLOCK_THREAD_CPUTIME_ID, &ts);
69 ret = ts.tv_sec * 1000000000ULL + ts.tv_nsec;
70#endif
71 return ret;
72}
73
74/* returns another thread's cumulated CPU time in nanoseconds if supported, otherwise zero */
Willy Tarreau21694982021-10-08 15:09:17 +020075uint64_t now_cpu_time_thread(int thr)
Willy Tarreau55542642021-10-08 09:33:24 +020076{
77 uint64_t ret = 0;
78#if defined(_POSIX_TIMERS) && (_POSIX_TIMERS > 0) && defined(_POSIX_THREAD_CPUTIME)
79 struct timespec ts;
Willy Tarreau21694982021-10-08 15:09:17 +020080 clock_gettime(per_thread_clock_id[thr], &ts);
Willy Tarreau55542642021-10-08 09:33:24 +020081 ret = ts.tv_sec * 1000000000ULL + ts.tv_nsec;
82#endif
83 return ret;
84}
85
Willy Tarreau44c58da2021-10-08 12:27:54 +020086/* set the clock source for the local thread */
87void clock_set_local_source(void)
88{
89#if defined(_POSIX_TIMERS) && (_POSIX_TIMERS > 0) && defined(_POSIX_THREAD_CPUTIME)
90#ifdef USE_THREAD
Willy Tarreau21694982021-10-08 15:09:17 +020091 pthread_getcpuclockid(pthread_self(), &per_thread_clock_id[tid]);
Willy Tarreau44c58da2021-10-08 12:27:54 +020092#else
Willy Tarreau21694982021-10-08 15:09:17 +020093 per_thread_clock_id[tid] = CLOCK_THREAD_CPUTIME_ID;
Willy Tarreau44c58da2021-10-08 12:27:54 +020094#endif
95#endif
96}
97
Willy Tarreau6cb0c392021-10-08 14:48:30 +020098/* registers a timer <tmr> of type timer_t delivering signal <sig> with value
99 * <val>. It tries on the current thread's clock ID first and falls back to
100 * CLOCK_REALTIME. Returns non-zero on success, 1 on failure.
101 */
102int clock_setup_signal_timer(void *tmr, int sig, int val)
103{
104 int ret = 0;
105
106#if defined(USE_RT) && (_POSIX_TIMERS > 0) && defined(_POSIX_THREAD_CPUTIME)
107 struct sigevent sev = { };
108 timer_t *timer = tmr;
109 sigset_t set;
110
111 /* unblock the WDTSIG signal we intend to use */
112 sigemptyset(&set);
113 sigaddset(&set, WDTSIG);
114 ha_sigmask(SIG_UNBLOCK, &set, NULL);
115
116 /* this timer will signal WDTSIG when it fires, with tid in the si_int
117 * field (important since any thread will receive the signal).
118 */
119 sev.sigev_notify = SIGEV_SIGNAL;
120 sev.sigev_signo = sig;
121 sev.sigev_value.sival_int = val;
Willy Tarreau21694982021-10-08 15:09:17 +0200122 if (timer_create(per_thread_clock_id[tid], &sev, timer) != -1 ||
Willy Tarreau6cb0c392021-10-08 14:48:30 +0200123 timer_create(CLOCK_REALTIME, &sev, timer) != -1)
124 ret = 1;
125#endif
126 return ret;
127}
128
Willy Tarreau55542642021-10-08 09:33:24 +0200129/* clock_update_date: sets <date> to system time, and sets <now> to something as
130 * close as possible to real time, following a monotonic function. The main
131 * principle consists in detecting backwards and forwards time jumps and adjust
132 * an offset to correct them. This function should be called once after each
133 * poll, and never farther apart than MAX_DELAY_MS*2. The poll's timeout should
134 * be passed in <max_wait>, and the return value in <interrupted> (a non-zero
135 * value means that we have not expired the timeout).
136 *
137 * clock_init_process_date() must have been called once first, and
138 * clock_init_thread_date() must also have been called once for each thread.
139 *
140 * An offset is used to adjust the current time (date), to figure a monotonic
141 * local time (now). The offset is not critical, as it is only updated after a
142 * clock jump is detected. From this point all threads will apply it to their
143 * locally measured time, and will then agree around a common monotonic
144 * global_now value that serves to further refine their local time. As it is
145 * not possible to atomically update a timeval, both global_now and the
146 * now_offset values are instead stored as 64-bit integers made of two 32 bit
147 * values for the tv_sec and tv_usec parts. The offset is made of two signed
148 * ints so that the clock can be adjusted in the two directions.
149 */
150void clock_update_date(int max_wait, int interrupted)
151{
152 struct timeval min_deadline, max_deadline, tmp_now;
153 uint old_now_ms;
154 ullong old_now;
155 ullong new_now;
156 ullong ofs, ofs_new;
157 uint sec_ofs, usec_ofs;
158
159 gettimeofday(&date, NULL);
160
161 /* compute the minimum and maximum local date we may have reached based
162 * on our past date and the associated timeout. There are three possible
163 * extremities:
164 * - the new date cannot be older than before_poll
165 * - if not interrupted, the new date cannot be older than
166 * before_poll+max_wait
167 * - in any case the new date cannot be newer than
168 * before_poll+max_wait+some margin (100ms used here).
169 * In case of violation, we'll ignore the current date and instead
170 * restart from the last date we knew.
171 */
172 _tv_ms_add(&min_deadline, &before_poll, max_wait);
173 _tv_ms_add(&max_deadline, &before_poll, max_wait + 100);
174
175 ofs = HA_ATOMIC_LOAD(&now_offset);
176
177 if (unlikely(__tv_islt(&date, &before_poll) || // big jump backwards
178 (!interrupted && __tv_islt(&date, &min_deadline)) || // small jump backwards
179 __tv_islt(&max_deadline, &date))) { // big jump forwards
180 if (!interrupted)
181 _tv_ms_add(&now, &now, max_wait);
182 } else {
183 /* The date is still within expectations. Let's apply the
184 * now_offset to the system date. Note: ofs if made of two
185 * independent signed ints.
186 */
187 now.tv_sec = date.tv_sec + (int)(ofs >> 32); // note: may be positive or negative
188 now.tv_usec = date.tv_usec + (int)ofs; // note: may be positive or negative
189 if ((int)now.tv_usec < 0) {
190 now.tv_usec += 1000000;
191 now.tv_sec -= 1;
192 } else if (now.tv_usec >= 1000000) {
193 now.tv_usec -= 1000000;
194 now.tv_sec += 1;
195 }
196 }
197
198 /* now that we have bounded the local time, let's check if it's
199 * realistic regarding the global date, which only moves forward,
200 * otherwise catch up.
201 */
202 old_now = global_now;
203 old_now_ms = global_now_ms;
204
205 do {
206 tmp_now.tv_sec = (unsigned int)(old_now >> 32);
207 tmp_now.tv_usec = old_now & 0xFFFFFFFFU;
208
209 if (__tv_islt(&now, &tmp_now))
210 now = tmp_now;
211
212 /* now <now> is expected to be the most accurate date,
213 * equal to <global_now> or newer.
214 */
215 new_now = ((ullong)now.tv_sec << 32) + (uint)now.tv_usec;
216 now_ms = __tv_to_ms(&now);
217
218 /* let's try to update the global <now> (both in timeval
219 * and ms forms) or loop again.
220 */
221 } while (((new_now != old_now && !_HA_ATOMIC_CAS(&global_now, &old_now, new_now)) ||
222 (now_ms != old_now_ms && !_HA_ATOMIC_CAS(&global_now_ms, &old_now_ms, now_ms))) &&
223 __ha_cpu_relax());
224
225 /* <now> and <now_ms> are now updated to the last value of global_now
226 * and global_now_ms, which were also monotonically updated. We can
227 * compute the latest offset, we don't care who writes it last, the
228 * variations will not break the monotonic property.
229 */
230
231 sec_ofs = now.tv_sec - date.tv_sec;
232 usec_ofs = now.tv_usec - date.tv_usec;
233 if ((int)usec_ofs < 0) {
234 usec_ofs += 1000000;
235 sec_ofs -= 1;
236 }
237 ofs_new = ((ullong)sec_ofs << 32) + usec_ofs;
238 if (ofs_new != ofs)
239 HA_ATOMIC_STORE(&now_offset, ofs_new);
240}
241
242/* must be called once at boot to initialize some global variables */
243void clock_init_process_date(void)
244{
245 now_offset = 0;
246 gettimeofday(&date, NULL);
247 now = after_poll = before_poll = date;
248 global_now = ((ullong)date.tv_sec << 32) + (uint)date.tv_usec;
249 global_now_ms = now.tv_sec * 1000 + now.tv_usec / 1000;
Willy Tarreau45c38e22021-09-30 18:28:49 +0200250 th_ctx->idle_pct = 100;
Willy Tarreau55542642021-10-08 09:33:24 +0200251 clock_update_date(0, 1);
252}
253
254/* must be called once per thread to initialize their thread-local variables.
255 * Note that other threads might also be initializing and running in parallel.
256 */
257void clock_init_thread_date(void)
258{
259 ullong old_now;
260
261 gettimeofday(&date, NULL);
262 after_poll = before_poll = date;
263
264 old_now = _HA_ATOMIC_LOAD(&global_now);
265 now.tv_sec = old_now >> 32;
266 now.tv_usec = (uint)old_now;
Willy Tarreau45c38e22021-09-30 18:28:49 +0200267 th_ctx->idle_pct = 100;
Willy Tarreau55542642021-10-08 09:33:24 +0200268 clock_update_date(0, 1);
269}
270
Willy Tarreauf9d5e102021-10-08 10:43:59 +0200271/* report the average CPU idle percentage over all running threads, between 0 and 100 */
272uint clock_report_idle(void)
273{
274 uint total = 0;
275 uint rthr = 0;
276 uint thr;
277
278 for (thr = 0; thr < MAX_THREADS; thr++) {
279 if (!(all_threads_mask & (1UL << thr)))
280 continue;
Willy Tarreau45c38e22021-09-30 18:28:49 +0200281 total += HA_ATOMIC_LOAD(&ha_thread_ctx[thr].idle_pct);
Willy Tarreauf9d5e102021-10-08 10:43:59 +0200282 rthr++;
283 }
284 return rthr ? total / rthr : 0;
285}
286
287/* Update the idle time value twice a second, to be called after
288 * clock_update_date() when called after poll(), and currently called only by
289 * clock_leaving_poll() below. It relies on <before_poll> to be updated to
290 * the system time before calling poll().
291 */
292static inline void clock_measure_idle(void)
293{
294 /* Let's compute the idle to work ratio. We worked between after_poll
295 * and before_poll, and slept between before_poll and date. The idle_pct
296 * is updated at most twice every second. Note that the current second
297 * rarely changes so we avoid a multiply when not needed.
298 */
299 int delta;
300
301 if ((delta = date.tv_sec - before_poll.tv_sec))
302 delta *= 1000000;
303 idle_time += delta + (date.tv_usec - before_poll.tv_usec);
304
305 if ((delta = date.tv_sec - after_poll.tv_sec))
306 delta *= 1000000;
307 samp_time += delta + (date.tv_usec - after_poll.tv_usec);
308
309 after_poll.tv_sec = date.tv_sec; after_poll.tv_usec = date.tv_usec;
310 if (samp_time < 500000)
311 return;
312
Willy Tarreau45c38e22021-09-30 18:28:49 +0200313 HA_ATOMIC_STORE(&th_ctx->idle_pct, (100ULL * idle_time + samp_time / 2) / samp_time);
Willy Tarreauf9d5e102021-10-08 10:43:59 +0200314 idle_time = samp_time = 0;
315}
316
317/* Collect date and time information after leaving poll(). <timeout> must be
318 * set to the maximum sleep time passed to poll (in milliseconds), and
319 * <interrupted> must be zero if the poller reached the timeout or non-zero
320 * otherwise, which generally is provided by the poller's return value.
321 */
322void clock_leaving_poll(int timeout, int interrupted)
323{
324 clock_measure_idle();
Willy Tarreau45c38e22021-09-30 18:28:49 +0200325 th_ctx->prev_cpu_time = now_cpu_time();
326 th_ctx->prev_mono_time = now_mono_time();
Willy Tarreauf9d5e102021-10-08 10:43:59 +0200327}
328
329/* Collect date and time information before calling poll(). This will be used
330 * to count the run time of the past loop and the sleep time of the next poll.
331 * It also compares the elasped and cpu times during the activity period to
332 * estimate the amount of stolen time, which is reported if higher than half
333 * a millisecond.
334 */
335void clock_entering_poll(void)
336{
337 uint64_t new_mono_time;
338 uint64_t new_cpu_time;
Willy Tarreau20adfde2021-10-08 11:34:46 +0200339 uint32_t run_time;
Willy Tarreauf9d5e102021-10-08 10:43:59 +0200340 int64_t stolen;
341
342 gettimeofday(&before_poll, NULL);
343
Willy Tarreau20adfde2021-10-08 11:34:46 +0200344 run_time = (before_poll.tv_sec - after_poll.tv_sec) * 1000000U + (before_poll.tv_usec - after_poll.tv_usec);
345
Willy Tarreauf9d5e102021-10-08 10:43:59 +0200346 new_cpu_time = now_cpu_time();
347 new_mono_time = now_mono_time();
348
Willy Tarreau45c38e22021-09-30 18:28:49 +0200349 if (th_ctx->prev_cpu_time && th_ctx->prev_mono_time) {
350 new_cpu_time -= th_ctx->prev_cpu_time;
351 new_mono_time -= th_ctx->prev_mono_time;
Willy Tarreauf9d5e102021-10-08 10:43:59 +0200352 stolen = new_mono_time - new_cpu_time;
353 if (unlikely(stolen >= 500000)) {
354 stolen /= 500000;
355 /* more than half a millisecond difference might
356 * indicate an undesired preemption.
357 */
358 report_stolen_time(stolen);
359 }
360 }
Willy Tarreau20adfde2021-10-08 11:34:46 +0200361
362 /* update the average runtime */
363 activity_count_runtime(run_time);
Willy Tarreauf9d5e102021-10-08 10:43:59 +0200364}
365
Willy Tarreau55542642021-10-08 09:33:24 +0200366/* returns the current date as returned by gettimeofday() in ISO+microsecond
367 * format. It uses a thread-local static variable that the reader can consume
368 * for as long as it wants until next call. Thus, do not call it from a signal
369 * handler. If <pad> is non-0, a trailing space will be added. It will always
370 * return exactly 32 or 33 characters (depending on padding) and will always be
371 * zero-terminated, thus it will always fit into a 34 bytes buffer.
372 * This also always include the local timezone (in +/-HH:mm format) .
373 */
374char *timeofday_as_iso_us(int pad)
375{
376 struct timeval new_date;
377 struct tm tm;
378 const char *offset;
379 char c;
380
381 gettimeofday(&new_date, NULL);
382 if (new_date.tv_sec != iso_time_sec || !new_date.tv_sec) {
383 get_localtime(new_date.tv_sec, &tm);
384 offset = get_gmt_offset(new_date.tv_sec, &tm);
385 if (unlikely(strftime(iso_time_str, sizeof(iso_time_str), "%Y-%m-%dT%H:%M:%S.000000+00:00", &tm) != 32))
386 strcpy(iso_time_str, "YYYY-mm-ddTHH:MM:SS.000000-00:00"); // make the failure visible but respect format.
387 iso_time_str[26] = offset[0];
388 iso_time_str[27] = offset[1];
389 iso_time_str[28] = offset[2];
390 iso_time_str[30] = offset[3];
391 iso_time_str[31] = offset[4];
392 iso_time_sec = new_date.tv_sec;
393 }
394
395 /* utoa_pad adds a trailing 0 so we save the char for restore */
396 c = iso_time_str[26];
397 utoa_pad(new_date.tv_usec, iso_time_str + 20, 7);
398 iso_time_str[26] = c;
399 if (pad) {
400 iso_time_str[32] = ' ';
401 iso_time_str[33] = 0;
402 }
403 return iso_time_str;
404}