diff options
Diffstat (limited to 'compat/delay.c')
| -rw-r--r-- | compat/delay.c | 407 |
1 files changed, 407 insertions, 0 deletions
diff --git a/compat/delay.c b/compat/delay.c new file mode 100644 index 0000000..95bde3e --- /dev/null +++ b/compat/delay.c @@ -0,0 +1,407 @@ +/*--------------------------------------------------------------- + * Copyright (c) 1999,2000,2001,2002,2003 + * The Board of Trustees of the University of Illinois + * All Rights Reserved. + *--------------------------------------------------------------- + * Permission is hereby granted, free of charge, to any person + * obtaining a copy of this software (Iperf) and associated + * documentation files (the "Software"), to deal in the Software + * without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, + * sublicense, and/or sell copies of the Software, and to permit + * persons to whom the Software is furnished to do + * so, subject to the following conditions: + * + * + * Redistributions of source code must retain the above + * copyright notice, this list of conditions and + * the following disclaimers. + * + * + * Redistributions in binary form must reproduce the above + * copyright notice, this list of conditions and the following + * disclaimers in the documentation and/or other materials + * provided with the distribution. + * + * + * Neither the names of the University of Illinois, NCSA, + * nor the names of its contributors may be used to endorse + * or promote products derived from this Software without + * specific prior written permission. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, + * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES + * OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND + * NONINFRINGEMENT. IN NO EVENT SHALL THE CONTIBUTORS OR COPYRIGHT + * HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, + * WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, + * ARISING FROM, OUT OF OR IN CONNECTION WITH THE + * SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE. + * ________________________________________________________________ + * National Laboratory for Applied Network Research + * National Center for Supercomputing Applications + * University of Illinois at Urbana-Champaign + * http://www.ncsa.uiuc.edu + * ________________________________________________________________ + * + * delay.c + * by Mark Gates <mgates@nlanr.net> + * updates + * by Robert J. McMahon <rmcmahon@broadcom.com> <rjmcmahon@rjmcmahon.com> + * ------------------------------------------------------------------- + * attempts at accurate microsecond delays + * ------------------------------------------------------------------- */ +#include "headers.h" +#include "util.h" +#include "delay.h" +#include "Thread.h" +#include <math.h> + +#define MILLION 1000000 +#define BILLION 1000000000 + +/* ------------------------------------------------------------------- + * A micro-second delay function + * o Use a busy loop or nanosleep + * + * Some notes: + * o clock nanosleep with a relative is preferred (see man page for why) + * o clock_gettime() (if available) is preferred over gettimeofday() + * as it give nanosecond resolution and should be more efficient. + * It also supports CLOCK_MONOTONIC and CLOCK_MONOTONIC_RAW + * though CLOCK_REALTIME is being used by the code. + * o This code does not use Timestamp object, as the goal of these + * functions is accurate delays (vs accurate timestamps.) + * o The syscalls such as nanosleep guarantee at least the request time + * and can and will delay longer, particularly due to things like context + * switching, causing the delay to lose accuracy + * o Kalman filtering is used to predict delay error which in turn + * is used to adjust the delay, hopefully mitigating the above. + * Note: This can cause the delay to return faster than the request, + * i.e. the *at least* guarantee is not preserved for the kalman + * adjusted delay calls. + * o Remember, the Client is keeping a running average delay for the + * thread so errors in delay will also be adjusted there. (Assuming + * it's possible. It's not really possible at top line link rates + * because lost time can't be made up for by speeding up the transmits. + * Hence, don't lose time with delay calls which error on the side of + * taking too long. Kalman should help much here.) + * + * POSIX nanosleep(). This allows a higher timing resolution + * (under Linux e.g. it uses hrtimers), does not affect any signals, + * and will use up remaining time when interrupted. + * ------------------------------------------------------------------- */ + +void delay_loop(unsigned long usec) +{ +#ifdef HAVE_CLOCK_NANOSLEEP + { + struct timespec res; + res.tv_sec = usec/MILLION; + res.tv_nsec = (usec * 1000) % BILLION; + #ifndef WIN32 + clock_nanosleep(CLOCK_MONOTONIC, 0, &res, NULL); + #else + clock_nanosleep(0, 0, &res, NULL); + #endif + } +#else + #ifdef HAVE_KALMAN + delay_kalman(usec); + #else + #ifdef HAVE_NANOSLEEP + delay_nanosleep(usec); + #else + delay_busyloop(usec); + #endif + #endif +#endif +} + +int clock_usleep (struct timeval *request) { + int rc = 0; +#if HAVE_THREAD_DEBUG + thread_debug("Thread called clock_usleep() until %ld.%ld", request->tv_sec, request->tv_usec); +#endif +#ifdef HAVE_CLOCK_NANOSLEEP + struct timespec tmp; + tmp.tv_sec = request->tv_sec; + tmp.tv_nsec = request->tv_usec * 1000; + +// Cygwin systems have an issue with CLOCK_MONOTONIC +#if defined(CLOCK_MONOTONIC) && !defined(WIN32) + rc = clock_nanosleep(CLOCK_MONOTONIC, 0, &tmp, NULL); +#else + rc = clock_nanosleep(0, 0, &tmp, NULL); +#endif + if (rc) { + fprintf(stderr, "failed clock_nanosleep()=%d\n", rc); + } +#else + struct timeval now; + struct timeval next = *request; +#ifdef HAVE_CLOCK_GETTIME + struct timespec t1; + clock_gettime(CLOCK_REALTIME, &t1); + now.tv_sec = t1.tv_sec; + now.tv_usec = t1.tv_nsec / 1000; +#else + gettimeofday(&now, NULL); +#endif + double delta_usecs; + if ((delta_usecs = TimeDifference(next, now)) > 0.0) { + delay_loop(delta_usecs); + } +#endif + return rc; +} + +int clock_usleep_abstime (struct timeval *request) { + int rc = 0; +#if defined(HAVE_CLOCK_NANOSLEEP) && defined(TIMER_ABSTIME) && !defined(WIN32) + struct timespec tmp; + tmp.tv_sec = request->tv_sec; + tmp.tv_nsec = request->tv_usec * 1000; + rc = clock_nanosleep(CLOCK_REALTIME, TIMER_ABSTIME, &tmp, NULL); + if (rc) { + fprintf(stderr, "failed clock_nanosleep()=%d\n", rc); + } +#else + struct timeval now; + struct timeval next = *request; +#ifdef HAVE_CLOCK_GETTIME + struct timespec t1; + clock_gettime(CLOCK_REALTIME, &t1); + now.tv_sec = t1.tv_sec; + now.tv_usec = t1.tv_nsec / 1000; +#else + gettimeofday(&now, NULL); +#endif + double delta_usecs; + if ((delta_usecs = (1e6 * TimeDifference(next, now))) > 0.0) { + delay_loop(delta_usecs); + } +#endif + return rc; +} + +#ifdef HAVE_NANOSLEEP +// Can use the nanosleep syscall suspending the thread +void delay_nanosleep (unsigned long usec) { + struct timespec requested, remaining; + requested.tv_sec = 0; + requested.tv_nsec = usec * 1000L; + // Note, signals will cause the nanosleep + // to return early. That's fine. + nanosleep(&requested, &remaining); +} +#endif + +#if defined (HAVE_NANOSLEEP) || defined (HAVE_CLOCK_GETTIME) +static void timespec_add_ulong (struct timespec *tv0, unsigned long value) { + tv0->tv_sec += (value / BILLION); + tv0->tv_nsec += (value % BILLION); + if (tv0->tv_nsec >= BILLION) { + tv0->tv_sec++; + tv0->tv_nsec -= BILLION; + } +} +#endif + +#ifdef HAVE_KALMAN +// Kalman versions attempt to support delay request +// accuracy over a minimum guaranteed delay by +// prediciting the delay error. This is +// the basic recursive algorithm. +static void kalman_update (struct kalman_state *state, double measurement) { + //prediction update + state->p = state->p + state->q; + //measurement update + state->k = state->p / (state->p + state->r); + state->x = state->x + (state->k * (measurement - state->x)); + state->p = (1 - state->k) * state->p; +} +#endif + +#ifdef HAVE_CLOCK_GETTIME +// Delay calls for systems with clock_gettime +// Working units are nanoseconds and structures are timespec +static void timespec_add_double (struct timespec *tv0, double value) { + tv0->tv_nsec += (unsigned long) value; + if (tv0->tv_nsec >= BILLION) { + tv0->tv_sec++; + tv0->tv_nsec -= BILLION; + } +} +// tv1 assumed greater than tv0 +static double timespec_diff (struct timespec tv1, struct timespec tv0) { + double result; + if (tv1.tv_nsec < tv0.tv_nsec) { + tv1.tv_nsec += BILLION; + tv1.tv_sec--; + } + result = (double) (((tv1.tv_sec - tv0.tv_sec) * BILLION) + (tv1.tv_nsec - tv0.tv_nsec)); + return result; +} +static void timespec_add( struct timespec *tv0, struct timespec *tv1) +{ + tv0->tv_sec += tv1->tv_sec; + tv0->tv_nsec += tv1->tv_nsec; + if ( tv0->tv_nsec >= BILLION ) { + tv0->tv_nsec -= BILLION; + tv0->tv_sec++; + } +} +static inline +int timespec_greaterthan(struct timespec tv1, struct timespec tv0) { + if (tv1.tv_sec > tv0.tv_sec || \ + ((tv0.tv_sec == tv1.tv_sec) && (tv1.tv_nsec > tv0.tv_nsec))) { + return 1; + } else { + return 0; + } +} +// A cpu busy loop for systems with clock_gettime +void delay_busyloop (unsigned long usec) { + struct timespec t1, t2; + clock_gettime(CLOCK_REALTIME, &t1); + timespec_add_ulong(&t1, (usec * 1000L)); + while (1) { + clock_gettime(CLOCK_REALTIME, &t2); + if (timespec_greaterthan(t2, t1)) + break; + } +} +// Kalman routines for systems with clock_gettime +#ifdef HAVE_KALMAN +// Request units is microseconds +// Adjust units is nanoseconds +void delay_kalman (unsigned long usec) { + struct timespec t1, t2, finishtime, requested={0,0}, remaining; + double nsec_adjusted, err; + static struct kalman_state kalmanerr={ + 0.00001, //q process noise covariance + 0.1, //r measurement noise covariance + 0.0, //x value, error predictio (units nanoseconds) + 1, //p estimation error covariance + 0.75 //k kalman gain + }; + // Get the current clock + clock_gettime(CLOCK_REALTIME, &t1); + // Perform the kalman adjust per the predicted delay error + nsec_adjusted = (usec * 1000.0) - kalmanerr.x; + // Set a timespec to be used by the nanosleep + // as well as for the finished time calculation + timespec_add_double(&requested, nsec_adjusted); + // Set the finish time in timespec format + finishtime = t1; + timespec_add(&finishtime, &requested); +# ifdef HAVE_NANOSLEEP + // Don't call nanosleep for values less than 10 microseconds + // as the syscall is too expensive. Let the busy loop + // provide the delay for times under that. + if (nsec_adjusted > 10000) { + nanosleep(&requested, &remaining); + } +# endif + while (1) { + clock_gettime(CLOCK_REALTIME, &t2); + if (timespec_greaterthan(t2, finishtime)) + break; + } + // Compute the delay error in units of nanoseconds + // and cast to type double + err = (timespec_diff(t2, t1) - (usec * 1000)); + // printf("req: %ld adj: %f err: %.5f (ns)\n", usec, nsec_adjusted, kalmanerr.x); + kalman_update(&kalmanerr, err); +} +#endif // HAVE_KALMAN +#else +// Sadly, these systems must use the not so efficient gettimeofday() +// and working units are microseconds, struct is timeval +static void timeval_add_ulong (struct timeval *tv0, unsigned long value) { + tv0->tv_usec += value; + if (tv0->tv_usec >= MILLION) { + tv0->tv_sec++; + tv0->tv_usec -= MILLION; + } +} +static inline +int timeval_greaterthan(struct timeval tv1, struct timeval tv0) { + if (tv1.tv_sec > tv0.tv_sec || \ + ((tv0.tv_sec == tv1.tv_sec) && (tv1.tv_usec > tv0.tv_usec))) { + return 1; + } else { + return 0; + } +} +// tv1 assumed greater than tv0 +static double timeval_diff (struct timeval tv1, struct timeval tv0) { + double result; + if (tv1.tv_usec < tv0.tv_usec) { + tv1.tv_usec += MILLION; + tv1.tv_sec--; + } + result = (double) (((tv1.tv_sec - tv0.tv_sec) * MILLION) + (tv1.tv_usec - tv0.tv_usec)); + return result; +} +void delay_busyloop (unsigned long usec) { + struct timeval t1, t2; + gettimeofday( &t1, NULL ); + timeval_add_ulong(&t1, usec); + while (1) { + gettimeofday( &t2, NULL ); + if (timeval_greaterthan(t2, t1)) + break; + } +} +#ifdef HAVE_KALMAN +// Request units is microseconds +// Adjust units is microseconds +void delay_kalman (unsigned long usec) { + struct timeval t1, t2, finishtime; + long usec_adjusted; + double err; + static struct kalman_state kalmanerr={ + 0.00001, //q process noise covariance + 0.1, //r measurement noise covariance + 0.0, //x value, error predictio (units nanoseconds) + 1, //p estimation error covariance + 0.25 //k kalman gain + }; + // Get the current clock + gettimeofday( &t1, NULL ); + // Perform the kalman adjust per the predicted delay error + if (kalmanerr.x > 0) { + usec_adjusted = usec - (long) floor(kalmanerr.x); + if (usec_adjusted < 0) + usec_adjusted = 0; + } + else + usec_adjusted = usec + (long) floor(kalmanerr.x); + // Set the finishtime + finishtime = t1; + timeval_add_ulong(&finishtime, usec_adjusted); +# ifdef HAVE_NANOSLEEP + // Don't call nanosleep for values less than 10 microseconds + // as the syscall is too expensive. Let the busy loop + // provide the delay for times under that. + if (usec_adjusted > 10) { + struct timespec requested={0,0}, remaining; + timespec_add_ulong(&requested, (usec_adjusted * 1000)); + nanosleep(&requested, &remaining); + } +# endif + while (1) { + gettimeofday(&t2, NULL ); + if (timeval_greaterthan(t2, finishtime)) + break; + } + // Compute the delay error in units of microseconds + // and cast to type double + err = (double)(timeval_diff(t2, t1) - usec); + // printf("req: %ld adj: %ld err: %.5f (us)\n", usec, usec_adjusted, kalmanerr.x); + kalman_update(&kalmanerr, err); +} +#endif // Kalman +#endif |