erl_time_sup.c

OTP是开放电信平台的简称

/* ``The contents of this file are subject to the Erlang Public License,
 * Version 1.1, (the "License"); you may not use this file except in
 * compliance with the License. You should have received a copy of the
 * Erlang Public License along with this software. If not, it can be
 * retrieved via the world wide web at http://www.erlang.org/.
 * 
 * Software distributed under the License is distributed on an "AS IS"
 * basis, WITHOUT WARRANTY OF ANY KIND, either express or implied. See
 * the License for the specific language governing rights and limitations
 * under the License.
 * 
 * The Initial Developer of the Original Code is Ericsson Utvecklings AB.
 * Portions created by Ericsson are Copyright 1999, Ericsson Utvecklings
 * AB. All Rights Reserved.''
 * 
 *     $Id$
 */
/*
** Support routines for the timer wheel
**
** This code contains two strategies for dealing with 
** date/time changes in the system. 
** If the system has some kind of high resolution timer (HAVE_GETHRTIME),
** the high resolution timer is used to correct the time-of-day and the
** timeouts, the base source is the hrtimer, but at certain intervals the 
** OS time-of-day is checked and if it is not within certain bounds, the 
** delivered time gets slowly adjusted for each call until
** it corresponds to the system time (built-in adjtime...). 
** The call gethrtime() is detected by autoconf on Unix, but other 
** platforms may define it in erl_*_sys.h and implement 
** their own high resolution timer. The high resolution timer
** strategy is (probably) best on all systems where the timer have 
** a resolution higher or equal to gettimeofday (or what's implemented
** is sys_gettimeofday()). The actual resolution is the interesting thing,
** not the unit's thats used (i.e. on VxWorks, nanoseconds can be
** retrieved in terms of units, but the actual resolution is the same as 
** for the clock ticks).
** If the systems best timer routine is kernel ticks returned from 
** sys_times(), and the actual resolution of sys_gettimeofday() is
** better (like most unixes that does not have any realtime extensions), 
** another strategy is used. The tolerant gettimeofday() corrects 
** the value with respect to uptime (sys_times() return value) and checks 
** for correction both when delivering timeticks and delivering nowtime.
** this strategy is slower, but accurate on systems without better timer 
** routines. The kernel tick resolution is not enough to implement
** a gethrtime routine. On Linux and other non solaris unix-boxes the second 
** strategy is used, on all other platforms we use the first.
** 
** The following is expected (from sys.[ch] and erl_*_sys.h):
**
** 64 bit integers. So it is, and so it will be.
**
** sys_init_time(), will return the clock resolution in MS and
** that's about it. More could be added of course
** If the clock-rate is constant (i.e. 1 ms) one can define 
** SYS_CLOCK_RESOLUTION (to 1),
** which makes erts_deliver_time/erts_time_remaining a bit faster.
**
** if HAVE_GETHRTIME is defined:
**    sys_gethrtime() will return a SysHrTime (long long) representing 
**    nanoseconds, sys_init_hrtime() will do any initialization.
** else
**    a long (64bit) integer type called Sint64 should be defined.
**
** sys_times() will return clock_ticks since start and 
**    fill in a SysTimes structure (struct tms). Instead of CLK_TCK, 
**    SYS_CLK_TCK is used to determine the resolution of kernel ticks.
**
** sys_gettimeofday() will take a SysTimeval (a struct timeval) as parameter
**    and fill it in as gettimeofday(X,NULL).
**
*/

#ifdef HAVE_CONFIG_H
#  include "config.h"
#endif

#include "sys.h"
#include "erl_vm.h"
#include "global.h"

static erts_smp_mtx_t erts_timeofday_mtx;

static SysTimeval inittv; /* Used everywhere, the initial time-of-day */

static SysTimes t_start; /* Used in elapsed_time_both */
static SysTimeval gtv; /* Used in wall_clock_elapsed_time_both */
static SysTimeval then; /* Used in get_now */

#ifdef HAVE_GETHRTIME

int erts_disable_tolerant_timeofday;

static SysHrTime hr_init_time, hr_last_correction_check, 
    hr_correction, hr_last_time;

static void init_tolerant_timeofday(void)
{
    /* Should be in sys.c */
#if defined(HAVE_SYSCONF) && defined(_SC_NPROCESSORS_CONF)
    if (sysconf(_SC_NPROCESSORS_CONF) > 1) {
	char b[1024];
	int maj,min,build;
	os_flavor(b,1024);
	os_version(&maj,&min,&build);
	if (!strcmp(b,"sunos") && maj <= 5 && min <= 7) {
	    erts_disable_tolerant_timeofday = 1;
	}
    }
#endif
    hr_init_time = sys_gethrtime();
    hr_last_correction_check = hr_last_time = hr_init_time;
    hr_correction = 0;
}

static void get_tolerant_timeofday(SysTimeval *tv)
{
    SysHrTime diff_time, curr;

    if (erts_disable_tolerant_timeofday) {
	sys_gettimeofday(tv);
	return;
    }
    *tv = inittv;
    diff_time = ((curr = sys_gethrtime()) + hr_correction - hr_init_time) / 1000; 

    if (curr < hr_init_time) {
	erl_exit(1,"Unexpected behaviour from operating system high "
		 "resolution timer");
    }

    if ((curr - hr_last_correction_check) / 1000 > 1000000) {
	/* Check the correction need */
	SysHrTime tv_diff, diffdiff;
	SysTimeval tmp;
	int done = 0;

	sys_gettimeofday(&tmp);
	tv_diff = ((SysHrTime) tmp.tv_sec) * 1000000 + tmp.tv_usec;
	tv_diff -= ((SysHrTime) inittv.tv_sec) * 1000000 + inittv.tv_usec;
	diffdiff = diff_time - tv_diff;
	if (diffdiff > 10000) {
	    SysHrTime corr = (curr - hr_last_time) / 100;
	    if (corr / 1000 >= diffdiff) {
		++done;
		hr_correction -= ((SysHrTime)diffdiff) * 1000;
	    } else {
		hr_correction -= corr;
	    }
	    diff_time = (curr + hr_correction - hr_init_time) / 1000; 
	} else if (diffdiff < -10000) {
	    SysHrTime corr = (curr - hr_last_time) / 100;
	    if (corr / 1000 >= -diffdiff) {
		++done;
		hr_correction -= ((SysHrTime)diffdiff) * 1000;
	    } else {
		hr_correction += corr;
	    }
	    diff_time = (curr + hr_correction - hr_init_time) / 1000; 
	} else {
	    ++done;
	}
	if (done) {
	    hr_last_correction_check = curr;
	}
    }
    tv->tv_sec += (int) (diff_time / ((SysHrTime) 1000000));
    tv->tv_usec += (int) (diff_time % ((SysHrTime) 1000000));
    if (tv->tv_usec >= 1000000) {
	tv->tv_usec -= 1000000;
	tv->tv_sec += 1;
    }
    hr_last_time = curr;
}

#define correction (hr_correction/1000000)

#else /* !HAVE_GETHRTIME */
#if !defined(CORRECT_USING_TIMES) 
#define init_tolerant_timeofday() 
#define get_tolerant_timeofday(tvp) sys_gettimeofday(tvp)
#else

typedef Sint64 Milli;

static clock_t init_ct;
static Sint64 ct_wrap;
static Milli init_tv_m;
static Milli correction_supress; 
static Milli last_ct_diff;
static Milli last_cc; 
static clock_t last_ct;

/* sys_times() might need to be wrapped and the values shifted (right)
   a bit to cope with newer linux (2.5.*) kernels, this has to be taken care 
   of dynamically to start with, a special version that uses
   the times() return value as a high resolution timer can be made
   to fully utilize the faster ticks, like on windows, but for now, we'll
   settle with this silly workaround */
#ifdef ERTS_WRAP_SYS_TIMES 
#define KERNEL_TICKS() (sys_times_wrap() &  \
			((1UL << ((sizeof(clock_t) * 8) - 1)) - 1)) 
#else
SysTimes dummy_tms;

#define KERNEL_TICKS() (sys_times(&dummy_tms) &  \
			((1UL << ((sizeof(clock_t) * 8) - 1)) - 1)) 

#endif

static void init_tolerant_timeofday(void)
{
    last_ct = init_ct = KERNEL_TICKS();
    last_cc = 0;
    init_tv_m = (((Milli) inittv.tv_sec) * 1000) + 
	(inittv.tv_usec / 1000);
    ct_wrap = 0;
    correction_supress = 0;
}


static void get_tolerant_timeofday(SysTimeval *tvp)
{
    clock_t current_ct;
    SysTimeval current_tv;
    Milli ct_diff;
    Milli tv_diff;
    Milli current_correction;
    Milli act_correction;	/* long shown to be too small */
    Milli max_adjust;

    if (erts_disable_tolerant_timeofday) {
	sys_gettimeofday(tvp);
	return;
    }

#ifdef ERTS_WRAP_SYS_TIMES 
#define TICK_MS (1000 / SYS_CLK_TCK_WRAP)
#else
#define TICK_MS (1000 / SYS_CLK_TCK)
#endif
    current_ct = KERNEL_TICKS();
    sys_gettimeofday(&current_tv);

    /* I dont know if uptime can move some units backwards
       on some systems, but I allow for small backward 
       jumps to avoid such problems if they exist...*/
    if (last_ct > 100 && current_ct < (last_ct - 100)) {
	ct_wrap += ((Sint64) 1) << ((sizeof(clock_t) * 8) - 1);
    }
    last_ct = current_ct;
    ct_diff = ((ct_wrap + current_ct) - init_ct) * TICK_MS;

    /*
     * We will adjust the time in milliseconds and we allow for 1% 
     * adjustments, but if this function is called more often then every 100 
     * millisecond (which is obviously possible), we will never adjust, so 
     * we accumulate small times by setting last_ct_diff iff max_adjust > 0
     */
    if ((max_adjust = (ct_diff - last_ct_diff)/100) > 0)
	last_ct_diff = ct_diff;

    tv_diff = ((((Milli) current_tv.tv_sec) * 1000) + 
	       (current_tv.tv_usec / 1000)) - init_tv_m;

    current_correction = ((ct_diff - tv_diff) / TICK_MS) * TICK_MS; /* trunc */

    /* 
     * We allow the current_correction value to wobble a little, as it
     * suffers from the low resolution of the kernel ticks. 
     * if it hasn't changed more than one tick in either direction, 
     * we will keep the old value.
     */
    if ((last_cc > current_correction + TICK_MS) ||
	(last_cc < current_correction - TICK_MS)) {
	last_cc = current_correction;
    } else {
	current_correction = last_cc;
    }
    
    /*
     * As time goes, we try to get the actual correction to 0, 
     * that is, make erlangs time correspond to the systems dito.
     * The act correction is what we seem to need (current_correction)
     * minus the correction suppression. The correction supression
     * will change slowly (max 1% of elapsed time) but in millisecond steps.
     */
    act_correction = current_correction - correction_supress;
    if (max_adjust > 0) {
	/*
	 * Here we slowly adjust erlangs time to correspond with the 
	 * system time by changing the correction_supress variable.
	 * It can change max_adjust milliseconds which is 1% of elapsed time
	 */
	if (act_correction > 0) {
	    if (current_correction - correction_supress > max_adjust) {
		correction_supress += max_adjust;
	    } else {
		correction_supress = current_correction;
	    }
	    act_correction = current_correction - correction_supress;
	} else if (act_correction < 0) {
	    if (correction_supress - current_correction > max_adjust) {
		correction_supress -= max_adjust;
	    } else {
		correction_supress = current_correction;
	    }
	    act_correction = current_correction - correction_supress;
	}
    }
    /*
     * The actual correction will correct the timeval so that system 
     * time warps gets smothed down.
     */
    current_tv.tv_sec += act_correction / 1000;
    current_tv.tv_usec += (act_correction % 1000) * 1000;

    if (current_tv.tv_usec >= 1000000) {
	++current_tv.tv_sec ;
	current_tv.tv_usec -= 1000000;
    } else if (current_tv.tv_usec < 0) {
	--current_tv.tv_sec;
	current_tv.tv_usec += 1000000;
    }
    *tvp = current_tv;
#undef TICK_MS
}

#endif /* CORRECT_USING_TIMES */
#endif /* !HAVE_GETHRTIME */

/*
** Why this? Well, most platforms have a constant clock resolution of 1,
** we dont want the deliver_time/time_remaining routines to waste 
** time dividing and multiplying by/with a variable that's always one.
** so the return value of sys_init_time is ignored on those platforms.
*/
 
#ifndef SYS_CLOCK_RESOLUTION
static int clock_resolution;
#define CLOCK_RESOLUTION clock_resolution
#else
#define CLOCK_RESOLUTION SYS_CLOCK_RESOLUTION
#endif

/*
** The clock resolution should really be the resolution of the 
** time function in use, which on most platforms 
** is 1. On VxWorks the resolution shold be 
** the number of ticks per second (or 1, which would work nicely to).
**
** Setting lower resolutions is mostly interesting when timers are used
** instead of something like select.
*/

#if defined(ERTS_TIMER_THREAD)
static ERTS_INLINE void init_erts_deliver_time(const SysTimeval *inittv) { }
static ERTS_INLINE void do_erts_deliver_time(const SysTimeval *current) { }
#else
static SysTimeval last_delivered; 

static void init_erts_deliver_time(const SysTimeval *inittv)
{
    /* We set the initial values for deliver_time here */
    last_delivered = *inittv;
    last_delivered.tv_usec = 1000 * (last_delivered.tv_usec / 1000); 
                                                   /* ms resolution */
}

static void do_erts_deliver_time(const SysTimeval *current)
{
    SysTimeval cur_time;
    long elapsed;
    
    /* calculate and deliver appropriate number of ticks */
    cur_time = *current;
    cur_time.tv_usec = 1000 * (cur_time.tv_usec / 1000); /* ms resolution */
    elapsed = (1000 * (cur_time.tv_sec - last_delivered.tv_sec) +
	       (cur_time.tv_usec - last_delivered.tv_usec) / 1000) / 
	CLOCK_RESOLUTION;

    /* Sometimes the time jump backwards,
       resulting in a negative elapsed time. We compensate for
       this by simply pretend as if the time stood still. :) */

    if (elapsed > 0) {
	do_time_add(elapsed);
	last_delivered = cur_time;
    }
}
#endif

int 
erts_init_time_sup(void)
{
    erts_smp_mtx_init(&erts_timeofday_mtx, "timeofday");

#ifndef SYS_CLOCK_RESOLUTION
    clock_resolution = sys_init_time();
#else
    (void) sys_init_time();
#endif
    sys_gettimeofday(&inittv);
    
#ifdef HAVE_GETHRTIME
    sys_init_hrtime();
#endif
    init_tolerant_timeofday();

    init_erts_deliver_time(&inittv);
    gtv = inittv;
    then.tv_sec = then.tv_usec = 0;