kern_clock.c revision 45245
159243Sobrien/*- 259243Sobrien * Copyright (c) 1997, 1998 Poul-Henning Kamp <phk@FreeBSD.org> 359243Sobrien * Copyright (c) 1982, 1986, 1991, 1993 459243Sobrien * The Regents of the University of California. All rights reserved. 559243Sobrien * (c) UNIX System Laboratories, Inc. 659243Sobrien * All or some portions of this file are derived from material licensed 759243Sobrien * to the University of California by American Telephone and Telegraph 859243Sobrien * Co. or Unix System Laboratories, Inc. and are reproduced herein with 959243Sobrien * the permission of UNIX System Laboratories, Inc. 1059243Sobrien * 1159243Sobrien * Redistribution and use in source and binary forms, with or without 1259243Sobrien * modification, are permitted provided that the following conditions 1359243Sobrien * are met: 1459243Sobrien * 1. Redistributions of source code must retain the above copyright 1559243Sobrien * notice, this list of conditions and the following disclaimer. 1659243Sobrien * 2. Redistributions in binary form must reproduce the above copyright 1759243Sobrien * notice, this list of conditions and the following disclaimer in the 1859243Sobrien * documentation and/or other materials provided with the distribution. 1959243Sobrien * 3. All advertising materials mentioning features or use of this software 2059243Sobrien * must display the following acknowledgement: 2159243Sobrien * This product includes software developed by the University of 2259243Sobrien * California, Berkeley and its contributors. 2359243Sobrien * 4. Neither the name of the University nor the names of its contributors 2459243Sobrien * may be used to endorse or promote products derived from this software 2559243Sobrien * without specific prior written permission. 2659243Sobrien * 2759243Sobrien * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 2859243Sobrien * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 2959243Sobrien * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 3059243Sobrien * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 3159243Sobrien * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 3259243Sobrien * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 3359243Sobrien * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 3459243Sobrien * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 3559243Sobrien * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 3659243Sobrien * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 3759243Sobrien * SUCH DAMAGE. 3859243Sobrien * 3959243Sobrien * @(#)kern_clock.c 8.5 (Berkeley) 1/21/94 4059243Sobrien * $Id: kern_clock.c,v 1.92 1999/03/12 21:58:54 phk Exp $ 4159243Sobrien */ 4259243Sobrien 4359243Sobrien#include "opt_ntp.h" 4459243Sobrien 4559243Sobrien#include <sys/param.h> 4659243Sobrien#include <sys/systm.h> 4759243Sobrien#include <sys/dkstat.h> 4859243Sobrien#include <sys/callout.h> 4959243Sobrien#include <sys/kernel.h> 5059243Sobrien#include <sys/proc.h> 5159243Sobrien#include <sys/malloc.h> 5259243Sobrien#include <sys/resourcevar.h> 5359243Sobrien#include <sys/signalvar.h> 5459243Sobrien#include <sys/timex.h> 5559243Sobrien#include <sys/timepps.h> 5659243Sobrien#include <vm/vm.h> 5759243Sobrien#include <sys/lock.h> 5859243Sobrien#include <vm/pmap.h> 5959243Sobrien#include <vm/vm_map.h> 6059243Sobrien#include <sys/sysctl.h> 6159243Sobrien 6259243Sobrien#include <machine/cpu.h> 6359243Sobrien#include <machine/limits.h> 6459243Sobrien 6559243Sobrien#ifdef GPROF 6659243Sobrien#include <sys/gmon.h> 6759243Sobrien#endif 6859243Sobrien 6959243Sobrien#if defined(SMP) && defined(BETTER_CLOCK) 7059243Sobrien#include <machine/smp.h> 7159243Sobrien#endif 7259243Sobrien 7359243Sobrien/* 7459243Sobrien * Number of timecounters used to implement stable storage 7559243Sobrien */ 7659243Sobrien#ifndef NTIMECOUNTER 7759243Sobrien#define NTIMECOUNTER 5 7859243Sobrien#endif 7959243Sobrien 8059243Sobrienstatic MALLOC_DEFINE(M_TIMECOUNTER, "timecounter", 8159243Sobrien "Timecounter stable storage"); 8259243Sobrien 8359243Sobrienstatic void initclocks __P((void *dummy)); 8459243SobrienSYSINIT(clocks, SI_SUB_CLOCKS, SI_ORDER_FIRST, initclocks, NULL) 85195609Smp 8659243Sobrienstatic void tco_forward __P((int force)); 8759243Sobrienstatic void tco_setscales __P((struct timecounter *tc)); 8859243Sobrienstatic __inline unsigned tco_delta __P((struct timecounter *tc)); 8959243Sobrien 9059243Sobrien/* Some of these don't belong here, but it's easiest to concentrate them. */ 9159243Sobrien#if defined(SMP) && defined(BETTER_CLOCK) 9259243Sobrienlong cp_time[CPUSTATES]; 9359243Sobrien#else 9459243Sobrienstatic long cp_time[CPUSTATES]; 9559243Sobrien#endif 9659243Sobrien 9759243Sobrienlong tk_cancc; 9859243Sobrienlong tk_nin; 9959243Sobrienlong tk_nout; 10059243Sobrienlong tk_rawcc; 10159243Sobrien 10259243Sobrientime_t time_second; 10359243Sobrien 10459243Sobrien/* 10559243Sobrien * Which update policy to use. 10659243Sobrien * 0 - every tick, bad hardware may fail with "calcru negative..." 10759243Sobrien * 1 - more resistent to the above hardware, but less efficient. 10859243Sobrien */ 10959243Sobrienstatic int tco_method; 11059243Sobrien 11159243Sobrien/* 11259243Sobrien * Implement a dummy timecounter which we can use until we get a real one 11359243Sobrien * in the air. This allows the console and other early stuff to use 11459243Sobrien * timeservices. 11559243Sobrien */ 11659243Sobrien 11759243Sobrienstatic unsigned 11859243Sobriendummy_get_timecount(struct timecounter *tc) 11959243Sobrien{ 12059243Sobrien static unsigned now; 12159243Sobrien return (++now); 12259243Sobrien} 12359243Sobrien 12459243Sobrienstatic struct timecounter dummy_timecounter = { 12559243Sobrien dummy_get_timecount, 12659243Sobrien 0, 12759243Sobrien ~0u, 12859243Sobrien 1000000, 12959243Sobrien "dummy" 13059243Sobrien}; 13159243Sobrien 13259243Sobrienstruct timecounter *timecounter = &dummy_timecounter; 13359243Sobrien 13459243Sobrien/* 135167465Smp * Clock handling routines. 136167465Smp * 137167465Smp * This code is written to operate with two timers that run independently of 138195609Smp * each other. 139167465Smp * 140167465Smp * The main timer, running hz times per second, is used to trigger interval 141167465Smp * timers, timeouts and rescheduling as needed. 142167465Smp * 143167465Smp * The second timer handles kernel and user profiling, 144167465Smp * and does resource use estimation. If the second timer is programmable, 145167465Smp * it is randomized to avoid aliasing between the two clocks. For example, 146167465Smp * the randomization prevents an adversary from always giving up the cpu 147167465Smp * just before its quantum expires. Otherwise, it would never accumulate 148167465Smp * cpu ticks. The mean frequency of the second timer is stathz. 149167465Smp * 150167465Smp * If no second timer exists, stathz will be zero; in this case we drive 15159243Sobrien * profiling and statistics off the main clock. This WILL NOT be accurate; 15259243Sobrien * do not do it unless absolutely necessary. 15359243Sobrien * 15459243Sobrien * The statistics clock may (or may not) be run at a higher rate while 155167465Smp * profiling. This profile clock runs at profhz. We require that profhz 156167465Smp * be an integral multiple of stathz. 15759243Sobrien * 158145479Smp * If the statistics clock is running fast, it must be divided by the ratio 159145479Smp * profhz/stathz for statistics. (For profiling, every tick counts.) 16059243Sobrien * 16159243Sobrien * Time-of-day is maintained using a "timecounter", which may or may 16259243Sobrien * not be related to the hardware generating the above mentioned 16359243Sobrien * interrupts. 16459243Sobrien */ 16559243Sobrien 16659243Sobrienint stathz; 16759243Sobrienint profhz; 16859243Sobrienstatic int profprocs; 16959243Sobrienint ticks; 17059243Sobrienstatic int psdiv, pscnt; /* prof => stat divider */ 17159243Sobrienint psratio; /* ratio: prof / stat */ 17259243Sobrien 17359243Sobrien/* 17459243Sobrien * Initialize clock frequencies and start both clocks running. 17559243Sobrien */ 176195609Smp/* ARGSUSED*/ 17759243Sobrienstatic void 17859243Sobrieninitclocks(dummy) 17959243Sobrien void *dummy; 18059243Sobrien{ 18159243Sobrien register int i; 18259243Sobrien 18359243Sobrien /* 18459243Sobrien * Set divisors to 1 (normal case) and let the machine-specific 18559243Sobrien * code do its bit. 18659243Sobrien */ 18759243Sobrien psdiv = pscnt = 1; 18859243Sobrien cpu_initclocks(); 18959243Sobrien 19059243Sobrien /* 19159243Sobrien * Compute profhz/stathz, and fix profhz if needed. 19259243Sobrien */ 19359243Sobrien i = stathz ? stathz : hz; 19459243Sobrien if (profhz == 0) 19559243Sobrien profhz = i; 19659243Sobrien psratio = profhz / i; 19759243Sobrien} 19859243Sobrien 19959243Sobrien/* 20059243Sobrien * The real-time timer, interrupting hz times per second. 20159243Sobrien */ 20259243Sobrienvoid 20359243Sobrienhardclock(frame) 20459243Sobrien register struct clockframe *frame; 20559243Sobrien{ 20659243Sobrien register struct proc *p; 20759243Sobrien 20859243Sobrien p = curproc; 20959243Sobrien if (p) { 21059243Sobrien register struct pstats *pstats; 21159243Sobrien 21259243Sobrien /* 21359243Sobrien * Run current process's virtual and profile time, as needed. 21459243Sobrien */ 21559243Sobrien pstats = p->p_stats; 21659243Sobrien if (CLKF_USERMODE(frame) && 21759243Sobrien timevalisset(&pstats->p_timer[ITIMER_VIRTUAL].it_value) && 21859243Sobrien itimerdecr(&pstats->p_timer[ITIMER_VIRTUAL], tick) == 0) 21959243Sobrien psignal(p, SIGVTALRM); 22059243Sobrien if (timevalisset(&pstats->p_timer[ITIMER_PROF].it_value) && 22159243Sobrien itimerdecr(&pstats->p_timer[ITIMER_PROF], tick) == 0) 22259243Sobrien psignal(p, SIGPROF); 22359243Sobrien } 22459243Sobrien 22559243Sobrien#if defined(SMP) && defined(BETTER_CLOCK) 22659243Sobrien forward_hardclock(pscnt); 22759243Sobrien#endif 22859243Sobrien 22959243Sobrien /* 23059243Sobrien * If no separate statistics clock is available, run it from here. 23159243Sobrien */ 232167465Smp if (stathz == 0) 23359243Sobrien statclock(frame); 234145479Smp 235145479Smp tco_forward(0); 236145479Smp ticks++; 23759243Sobrien 23859243Sobrien /* 23959243Sobrien * Process callouts at a very low cpu priority, so we don't keep the 24059243Sobrien * relatively high clock interrupt priority any longer than necessary. 24159243Sobrien */ 24259243Sobrien if (TAILQ_FIRST(&callwheel[ticks & callwheelmask]) != NULL) { 24359243Sobrien if (CLKF_BASEPRI(frame)) { 24459243Sobrien /* 24559243Sobrien * Save the overhead of a software interrupt; 24659243Sobrien * it will happen as soon as we return, so do it now. 24759243Sobrien */ 24859243Sobrien (void)splsoftclock(); 249167465Smp softclock(); 250167465Smp } else 25159243Sobrien setsoftclock(); 25259243Sobrien } else if (softticks + 1 == ticks) 25359243Sobrien ++softticks; 25459243Sobrien} 25559243Sobrien 25659243Sobrien/* 25759243Sobrien * Compute number of ticks in the specified amount of time. 25859243Sobrien */ 25959243Sobrienint 26059243Sobrientvtohz(tv) 26159243Sobrien struct timeval *tv; 26259243Sobrien{ 26359243Sobrien register unsigned long ticks; 26459243Sobrien register long sec, usec; 26559243Sobrien 26659243Sobrien /* 26759243Sobrien * If the number of usecs in the whole seconds part of the time 26859243Sobrien * difference fits in a long, then the total number of usecs will 26959243Sobrien * fit in an unsigned long. Compute the total and convert it to 27059243Sobrien * ticks, rounding up and adding 1 to allow for the current tick 27159243Sobrien * to expire. Rounding also depends on unsigned long arithmetic 27259243Sobrien * to avoid overflow. 27359243Sobrien * 27459243Sobrien * Otherwise, if the number of ticks in the whole seconds part of 27559243Sobrien * the time difference fits in a long, then convert the parts to 27659243Sobrien * ticks separately and add, using similar rounding methods and 27759243Sobrien * overflow avoidance. This method would work in the previous 27859243Sobrien * case but it is slightly slower and assumes that hz is integral. 27959243Sobrien * 28059243Sobrien * Otherwise, round the time difference down to the maximum 28159243Sobrien * representable value. 28259243Sobrien * 28359243Sobrien * If ints have 32 bits, then the maximum value for any timeout in 284167465Smp * 10ms ticks is 248 days. 28559243Sobrien */ 286145479Smp sec = tv->tv_sec; 287145479Smp usec = tv->tv_usec; 288145479Smp if (usec < 0) { 28959243Sobrien sec--; 29059243Sobrien usec += 1000000; 29159243Sobrien } 29259243Sobrien if (sec < 0) { 29359243Sobrien#ifdef DIAGNOSTIC 29459243Sobrien if (usec > 0) { 29559243Sobrien sec++; 29659243Sobrien usec -= 1000000; 29759243Sobrien } 29859243Sobrien printf("tvotohz: negative time difference %ld sec %ld usec\n", 29959243Sobrien sec, usec); 30059243Sobrien#endif 30159243Sobrien ticks = 1; 30259243Sobrien } else if (sec <= LONG_MAX / 1000000) 30359243Sobrien ticks = (sec * 1000000 + (unsigned long)usec + (tick - 1)) 30459243Sobrien / tick + 1; 30559243Sobrien else if (sec <= LONG_MAX / hz) 30659243Sobrien ticks = sec * hz 30759243Sobrien + ((unsigned long)usec + (tick - 1)) / tick + 1; 30859243Sobrien else 30959243Sobrien ticks = LONG_MAX; 31059243Sobrien if (ticks > INT_MAX) 31159243Sobrien ticks = INT_MAX; 312167465Smp return ((int)ticks); 31359243Sobrien} 314145479Smp 315145479Smp/* 316145479Smp * Start profiling on a process. 31759243Sobrien * 31859243Sobrien * Kernel profiling passes proc0 which never exits and hence 31959243Sobrien * keeps the profile clock running constantly. 32059243Sobrien */ 32159243Sobrienvoid 32259243Sobrienstartprofclock(p) 32359243Sobrien register struct proc *p; 32459243Sobrien{ 32559243Sobrien int s; 32659243Sobrien 32759243Sobrien if ((p->p_flag & P_PROFIL) == 0) { 32859243Sobrien p->p_flag |= P_PROFIL; 32959243Sobrien if (++profprocs == 1 && stathz != 0) { 33059243Sobrien s = splstatclock(); 33159243Sobrien psdiv = pscnt = psratio; 33259243Sobrien setstatclockrate(profhz); 33359243Sobrien splx(s); 33459243Sobrien } 33559243Sobrien } 33659243Sobrien} 33759243Sobrien 33859243Sobrien/* 33959243Sobrien * Stop profiling on a process. 34059243Sobrien */ 34159243Sobrienvoid 34259243Sobrienstopprofclock(p) 34359243Sobrien register struct proc *p; 34459243Sobrien{ 345195609Smp int s; 34659243Sobrien 347145479Smp if (p->p_flag & P_PROFIL) { 348145479Smp p->p_flag &= ~P_PROFIL; 349145479Smp if (--profprocs == 0 && stathz != 0) { 350167465Smp s = splstatclock(); 35159243Sobrien psdiv = pscnt = 1; 35259243Sobrien setstatclockrate(stathz); 35359243Sobrien splx(s); 35459243Sobrien } 35559243Sobrien } 35659243Sobrien} 35759243Sobrien 35859243Sobrien/* 35959243Sobrien * Statistics clock. Grab profile sample, and if divider reaches 0, 36059243Sobrien * do process and kernel statistics. 36159243Sobrien */ 36259243Sobrienvoid 36359243Sobrienstatclock(frame) 36459243Sobrien register struct clockframe *frame; 36559243Sobrien{ 36659243Sobrien#ifdef GPROF 36759243Sobrien register struct gmonparam *g; 36859243Sobrien int i; 36959243Sobrien#endif 37059243Sobrien register struct proc *p; 37159243Sobrien struct pstats *pstats; 37259243Sobrien long rss; 37359243Sobrien struct rusage *ru; 37459243Sobrien struct vmspace *vm; 37559243Sobrien 37659243Sobrien if (curproc != NULL && CLKF_USERMODE(frame)) { 37759243Sobrien p = curproc; 37859243Sobrien if (p->p_flag & P_PROFIL) 37959243Sobrien addupc_intr(p, CLKF_PC(frame), 1); 38059243Sobrien#if defined(SMP) && defined(BETTER_CLOCK) 38159243Sobrien if (stathz != 0) 38259243Sobrien forward_statclock(pscnt); 38359243Sobrien#endif 38459243Sobrien if (--pscnt > 0) 38559243Sobrien return; 38659243Sobrien /* 38759243Sobrien * Came from user mode; CPU was in user state. 388167465Smp * If this process is being profiled record the tick. 38959243Sobrien */ 390145479Smp p->p_uticks++; 391145479Smp if (p->p_nice > NZERO) 392145479Smp cp_time[CP_NICE]++; 393167465Smp else 39459243Sobrien cp_time[CP_USER]++; 39559243Sobrien } else { 39659243Sobrien#ifdef GPROF 39759243Sobrien /* 39859243Sobrien * Kernel statistics are just like addupc_intr, only easier. 39959243Sobrien */ 40059243Sobrien g = &_gmonparam; 40159243Sobrien if (g->state == GMON_PROF_ON) { 40259243Sobrien i = CLKF_PC(frame) - g->lowpc; 40359243Sobrien if (i < g->textsize) { 40459243Sobrien i /= HISTFRACTION * sizeof(*g->kcount); 40559243Sobrien g->kcount[i]++; 40659243Sobrien } 40759243Sobrien } 40859243Sobrien#endif 40959243Sobrien#if defined(SMP) && defined(BETTER_CLOCK) 41059243Sobrien if (stathz != 0) 41159243Sobrien forward_statclock(pscnt); 412167465Smp#endif 41359243Sobrien if (--pscnt > 0) 414145479Smp return; 415145479Smp /* 41659243Sobrien * Came from kernel mode, so we were: 41759243Sobrien * - handling an interrupt, 41859243Sobrien * - doing syscall or trap work on behalf of the current 41959243Sobrien * user process, or 42059243Sobrien * - spinning in the idle loop. 42159243Sobrien * Whichever it is, charge the time as appropriate. 42259243Sobrien * Note that we charge interrupts to the current process, 42359243Sobrien * regardless of whether they are ``for'' that process, 42459243Sobrien * so that we know how much of its real time was spent 42559243Sobrien * in ``non-process'' (i.e., interrupt) work. 42659243Sobrien */ 42759243Sobrien p = curproc; 428167465Smp if (CLKF_INTR(frame)) { 42959243Sobrien if (p != NULL) 430145479Smp p->p_iticks++; 431145479Smp cp_time[CP_INTR]++; 43259243Sobrien } else if (p != NULL) { 43359243Sobrien p->p_sticks++; 43459243Sobrien cp_time[CP_SYS]++; 43559243Sobrien } else 43659243Sobrien cp_time[CP_IDLE]++; 43759243Sobrien } 438131962Smp pscnt = psdiv; 439131962Smp 44059243Sobrien /* 44159243Sobrien * We maintain statistics shown by user-level statistics 44259243Sobrien * programs: the amount of time in each cpu state. 44359243Sobrien */ 444167465Smp 44559243Sobrien /* 446145479Smp * We adjust the priority of the current process. The priority of 44759243Sobrien * a process gets worse as it accumulates CPU time. The cpu usage 44859243Sobrien * estimator (p_estcpu) is increased here. The formula for computing 44959243Sobrien * priorities (in kern_synch.c) will compute a different value each 45059243Sobrien * time p_estcpu increases by 4. The cpu usage estimator ramps up 45159243Sobrien * quite quickly when the process is running (linearly), and decays 45259243Sobrien * away exponentially, at a rate which is proportionally slower when 453167465Smp * the system is busy. The basic principal is that the system will 45459243Sobrien * 90% forget that the process used a lot of CPU time in 5 * loadav 455145479Smp * seconds. This causes the system to favor processes which haven't 456145479Smp * run much recently, and to round-robin among other processes. 457167465Smp */ 45859243Sobrien if (p != NULL) { 45959243Sobrien p->p_cpticks++; 46059243Sobrien if (++p->p_estcpu == 0) 46159243Sobrien p->p_estcpu--; 46259243Sobrien if ((p->p_estcpu & 3) == 0) { 46359243Sobrien resetpriority(p); 46459243Sobrien if (p->p_priority >= PUSER) 46559243Sobrien p->p_priority = p->p_usrpri; 46659243Sobrien } 46759243Sobrien 46859243Sobrien /* Update resource usage integrals and maximums. */ 46959243Sobrien if ((pstats = p->p_stats) != NULL && 47059243Sobrien (ru = &pstats->p_ru) != NULL && 47159243Sobrien (vm = p->p_vmspace) != NULL) { 47259243Sobrien ru->ru_ixrss += pgtok(vm->vm_tsize); 47359243Sobrien ru->ru_idrss += pgtok(vm->vm_dsize); 47459243Sobrien ru->ru_isrss += pgtok(vm->vm_ssize); 47559243Sobrien rss = pgtok(vmspace_resident_count(vm)); 47659243Sobrien if (ru->ru_maxrss < rss) 47759243Sobrien ru->ru_maxrss = rss; 47859243Sobrien } 479167465Smp } 48059243Sobrien} 481145479Smp 482145479Smp/* 48359243Sobrien * Return information about system clocks. 48459243Sobrien */ 48559243Sobrienstatic int 48659243Sobriensysctl_kern_clockrate SYSCTL_HANDLER_ARGS 48759243Sobrien{ 48859243Sobrien struct clockinfo clkinfo; 48959243Sobrien /* 49059243Sobrien * Construct clockinfo structure. 49159243Sobrien */ 49259243Sobrien clkinfo.hz = hz; 49359243Sobrien clkinfo.tick = tick; 49459243Sobrien clkinfo.tickadj = tickadj; 495167465Smp clkinfo.profhz = profhz; 49659243Sobrien clkinfo.stathz = stathz ? stathz : hz; 497145479Smp return (sysctl_handle_opaque(oidp, &clkinfo, sizeof clkinfo, req)); 49859243Sobrien} 49959243Sobrien 50059243SobrienSYSCTL_PROC(_kern, KERN_CLOCKRATE, clockrate, CTLTYPE_STRUCT|CTLFLAG_RD, 50159243Sobrien 0, 0, sysctl_kern_clockrate, "S,clockinfo",""); 50259243Sobrien 50359243Sobrienstatic __inline unsigned 50459243Sobrientco_delta(struct timecounter *tc) 50559243Sobrien{ 50659243Sobrien 50759243Sobrien return ((tc->tc_get_timecount(tc) - tc->tc_offset_count) & 50859243Sobrien tc->tc_counter_mask); 50959243Sobrien} 510167465Smp 51159243Sobrien/* 512145479Smp * We have eight functions for looking at the clock, four for 51359243Sobrien * microseconds and four for nanoseconds. For each there is fast 51459243Sobrien * but less precise version "get{nano|micro}[up]time" which will 51559243Sobrien * return a time which is up to 1/HZ previous to the call, whereas 51659243Sobrien * the raw version "{nano|micro}[up]time" will return a timestamp 51759243Sobrien * which is as precise as possible. The "up" variants return the 51859243Sobrien * time relative to system boot, these are well suited for time 51959243Sobrien * interval measurements. 52059243Sobrien */ 52159243Sobrien 52259243Sobrienvoid 52359243Sobriengetmicrotime(struct timeval *tvp) 52459243Sobrien{ 525167465Smp struct timecounter *tc; 52659243Sobrien 527145479Smp if (!tco_method) { 528145479Smp tc = timecounter; 52959243Sobrien *tvp = tc->tc_microtime; 53059243Sobrien } else { 53159243Sobrien microtime(tvp); 53259243Sobrien } 53359243Sobrien} 53459243Sobrien 53559243Sobrienvoid 53659243Sobriengetnanotime(struct timespec *tsp) 53759243Sobrien{ 53859243Sobrien struct timecounter *tc; 53959243Sobrien 54059243Sobrien if (!tco_method) { 541167465Smp tc = timecounter; 54259243Sobrien *tsp = tc->tc_nanotime; 543145479Smp } else { 54459243Sobrien nanotime(tsp); 54559243Sobrien } 54659243Sobrien} 54759243Sobrien 54859243Sobrienvoid 54959243Sobrienmicrotime(struct timeval *tv) 55059243Sobrien{ 55159243Sobrien struct timecounter *tc; 55259243Sobrien 55359243Sobrien tc = (struct timecounter *)timecounter; 55459243Sobrien tv->tv_sec = tc->tc_offset_sec; 55559243Sobrien tv->tv_usec = tc->tc_offset_micro; 556167465Smp tv->tv_usec += ((u_int64_t)tco_delta(tc) * tc->tc_scale_micro) >> 32; 55759243Sobrien tv->tv_usec += boottime.tv_usec; 558145479Smp tv->tv_sec += boottime.tv_sec; 559167465Smp while (tv->tv_usec >= 1000000) { 56059243Sobrien tv->tv_usec -= 1000000; 56159243Sobrien tv->tv_sec++; 56259243Sobrien } 56359243Sobrien} 56459243Sobrien 56559243Sobrienvoid 56659243Sobriennanotime(struct timespec *ts) 56759243Sobrien{ 56859243Sobrien unsigned count; 56959243Sobrien u_int64_t delta; 57059243Sobrien struct timecounter *tc; 57159243Sobrien 57259243Sobrien tc = (struct timecounter *)timecounter; 57359243Sobrien ts->tv_sec = tc->tc_offset_sec; 57459243Sobrien count = tco_delta(tc); 57559243Sobrien delta = tc->tc_offset_nano; 57659243Sobrien delta += ((u_int64_t)count * tc->tc_scale_nano_f); 577167465Smp delta >>= 32; 57859243Sobrien delta += ((u_int64_t)count * tc->tc_scale_nano_i); 579145479Smp delta += boottime.tv_usec * 1000; 580145479Smp ts->tv_sec += boottime.tv_sec; 581167465Smp while (delta >= 1000000000) { 58259243Sobrien delta -= 1000000000; 58359243Sobrien ts->tv_sec++; 58459243Sobrien } 58559243Sobrien ts->tv_nsec = delta; 58659243Sobrien} 58759243Sobrien 58859243Sobrienvoid 58959243Sobriengetmicrouptime(struct timeval *tvp) 59059243Sobrien{ 59159243Sobrien struct timecounter *tc; 59259243Sobrien 59359243Sobrien if (!tco_method) { 59459243Sobrien tc = timecounter; 595 tvp->tv_sec = tc->tc_offset_sec; 596 tvp->tv_usec = tc->tc_offset_micro; 597 } else { 598 microuptime(tvp); 599 } 600} 601 602void 603getnanouptime(struct timespec *tsp) 604{ 605 struct timecounter *tc; 606 607 if (!tco_method) { 608 tc = timecounter; 609 tsp->tv_sec = tc->tc_offset_sec; 610 tsp->tv_nsec = tc->tc_offset_nano >> 32; 611 } else { 612 nanouptime(tsp); 613 } 614} 615 616void 617microuptime(struct timeval *tv) 618{ 619 struct timecounter *tc; 620 621 tc = (struct timecounter *)timecounter; 622 tv->tv_sec = tc->tc_offset_sec; 623 tv->tv_usec = tc->tc_offset_micro; 624 tv->tv_usec += ((u_int64_t)tco_delta(tc) * tc->tc_scale_micro) >> 32; 625 if (tv->tv_usec >= 1000000) { 626 tv->tv_usec -= 1000000; 627 tv->tv_sec++; 628 } 629} 630 631void 632nanouptime(struct timespec *ts) 633{ 634 unsigned count; 635 u_int64_t delta; 636 struct timecounter *tc; 637 638 tc = (struct timecounter *)timecounter; 639 ts->tv_sec = tc->tc_offset_sec; 640 count = tco_delta(tc); 641 delta = tc->tc_offset_nano; 642 delta += ((u_int64_t)count * tc->tc_scale_nano_f); 643 delta >>= 32; 644 delta += ((u_int64_t)count * tc->tc_scale_nano_i); 645 if (delta >= 1000000000) { 646 delta -= 1000000000; 647 ts->tv_sec++; 648 } 649 ts->tv_nsec = delta; 650} 651 652static void 653tco_setscales(struct timecounter *tc) 654{ 655 u_int64_t scale; 656 657 scale = 1000000000LL << 32; 658 scale += tc->tc_adjustment; 659 scale /= tc->tc_frequency; 660 tc->tc_scale_micro = scale / 1000; 661 tc->tc_scale_nano_f = scale & 0xffffffff; 662 tc->tc_scale_nano_i = scale >> 32; 663} 664 665void 666init_timecounter(struct timecounter *tc) 667{ 668 struct timespec ts1; 669 struct timecounter *t1, *t2, *t3; 670 int i; 671 672 tc->tc_adjustment = 0; 673 tco_setscales(tc); 674 tc->tc_offset_count = tc->tc_get_timecount(tc); 675 tc->tc_tweak = tc; 676 MALLOC(t1, struct timecounter *, sizeof *t1, M_TIMECOUNTER, M_WAITOK); 677 *t1 = *tc; 678 t2 = t1; 679 for (i = 1; i < NTIMECOUNTER; i++) { 680 MALLOC(t3, struct timecounter *, sizeof *t3, 681 M_TIMECOUNTER, M_WAITOK); 682 *t3 = *tc; 683 t3->tc_other = t2; 684 t2 = t3; 685 } 686 t1->tc_other = t3; 687 tc = t1; 688 689 printf("Timecounter \"%s\" frequency %lu Hz\n", 690 tc->tc_name, (u_long)tc->tc_frequency); 691 692 /* XXX: For now always start using the counter. */ 693 tc->tc_offset_count = tc->tc_get_timecount(tc); 694 nanouptime(&ts1); 695 tc->tc_offset_nano = (u_int64_t)ts1.tv_nsec << 32; 696 tc->tc_offset_micro = ts1.tv_nsec / 1000; 697 tc->tc_offset_sec = ts1.tv_sec; 698 timecounter = tc; 699} 700 701void 702set_timecounter(struct timespec *ts) 703{ 704 struct timespec ts2; 705 706 nanouptime(&ts2); 707 boottime.tv_sec = ts->tv_sec - ts2.tv_sec; 708 boottime.tv_usec = (ts->tv_nsec - ts2.tv_nsec) / 1000; 709 if (boottime.tv_usec < 0) { 710 boottime.tv_usec += 1000000; 711 boottime.tv_sec--; 712 } 713 /* fiddle all the little crinkly bits around the fiords... */ 714 tco_forward(1); 715} 716 717 718#if 0 /* Currently unused */ 719void 720switch_timecounter(struct timecounter *newtc) 721{ 722 int s; 723 struct timecounter *tc; 724 struct timespec ts; 725 726 s = splclock(); 727 tc = timecounter; 728 if (newtc == tc || newtc == tc->tc_other) { 729 splx(s); 730 return; 731 } 732 nanouptime(&ts); 733 newtc->tc_offset_sec = ts.tv_sec; 734 newtc->tc_offset_nano = (u_int64_t)ts.tv_nsec << 32; 735 newtc->tc_offset_micro = ts.tv_nsec / 1000; 736 newtc->tc_offset_count = newtc->tc_get_timecount(newtc); 737 timecounter = newtc; 738 splx(s); 739} 740#endif 741 742static struct timecounter * 743sync_other_counter(void) 744{ 745 struct timecounter *tc, *tcn, *tco; 746 unsigned delta; 747 748 tco = timecounter; 749 tc = tco->tc_other; 750 tcn = tc->tc_other; 751 *tc = *tco; 752 tc->tc_other = tcn; 753 delta = tco_delta(tc); 754 tc->tc_offset_count += delta; 755 tc->tc_offset_count &= tc->tc_counter_mask; 756 tc->tc_offset_nano += (u_int64_t)delta * tc->tc_scale_nano_f; 757 tc->tc_offset_nano += (u_int64_t)delta * tc->tc_scale_nano_i << 32; 758 return (tc); 759} 760 761static void 762tco_forward(int force) 763{ 764 struct timecounter *tc, *tco; 765 766 tco = timecounter; 767 tc = sync_other_counter(); 768 /* 769 * We may be inducing a tiny error here, the tc_poll_pps() may 770 * process a latched count which happens after the tco_delta() 771 * in sync_other_counter(), which would extend the previous 772 * counters parameters into the domain of this new one. 773 * Since the timewindow is very small for this, the error is 774 * going to be only a few weenieseconds (as Dave Mills would 775 * say), so lets just not talk more about it, OK ? 776 */ 777 if (tco->tc_poll_pps) 778 tco->tc_poll_pps(tco); 779 if (timedelta != 0) { 780 tc->tc_offset_nano += (u_int64_t)(tickdelta * 1000) << 32; 781 timedelta -= tickdelta; 782 force++; 783 } 784 785 while (tc->tc_offset_nano >= 1000000000ULL << 32) { 786 tc->tc_offset_nano -= 1000000000ULL << 32; 787 tc->tc_offset_sec++; 788 ntp_update_second(tc); /* XXX only needed if xntpd runs */ 789 tco_setscales(tc); 790 force++; 791 } 792 793 if (tco_method && !force) 794 return; 795 796 tc->tc_offset_micro = (tc->tc_offset_nano / 1000) >> 32; 797 798 /* Figure out the wall-clock time */ 799 tc->tc_nanotime.tv_sec = tc->tc_offset_sec + boottime.tv_sec; 800 tc->tc_nanotime.tv_nsec = 801 (tc->tc_offset_nano >> 32) + boottime.tv_usec * 1000; 802 tc->tc_microtime.tv_usec = tc->tc_offset_micro + boottime.tv_usec; 803 if (tc->tc_nanotime.tv_nsec >= 1000000000) { 804 tc->tc_nanotime.tv_nsec -= 1000000000; 805 tc->tc_microtime.tv_usec -= 1000000; 806 tc->tc_nanotime.tv_sec++; 807 } 808 time_second = tc->tc_microtime.tv_sec = tc->tc_nanotime.tv_sec; 809 810 timecounter = tc; 811} 812 813SYSCTL_NODE(_kern, OID_AUTO, timecounter, CTLFLAG_RW, 0, ""); 814 815SYSCTL_INT(_kern_timecounter, OID_AUTO, method, CTLFLAG_RW, &tco_method, 0, 816 "This variable determines the method used for updating timecounters. " 817 "If the default algorithm (0) fails with \"calcru negative...\" messages " 818 "try the alternate algorithm (1) which handles bad hardware better." 819 820); 821 822 823int 824pps_ioctl(u_long cmd, caddr_t data, struct pps_state *pps) 825{ 826 pps_params_t *app; 827 pps_info_t *api; 828 829 switch (cmd) { 830 case PPS_IOC_CREATE: 831 return (0); 832 case PPS_IOC_DESTROY: 833 return (0); 834 case PPS_IOC_SETPARAMS: 835 app = (pps_params_t *)data; 836 if (app->mode & ~pps->ppscap) 837 return (EINVAL); 838 pps->ppsparam = *app; 839 return (0); 840 case PPS_IOC_GETPARAMS: 841 app = (pps_params_t *)data; 842 *app = pps->ppsparam; 843 return (0); 844 case PPS_IOC_GETCAP: 845 *(int*)data = pps->ppscap; 846 return (0); 847 case PPS_IOC_FETCH: 848 api = (pps_info_t *)data; 849 pps->ppsinfo.current_mode = pps->ppsparam.mode; 850 *api = pps->ppsinfo; 851 return (0); 852 case PPS_IOC_WAIT: 853 return (EOPNOTSUPP); 854 default: 855 return (ENOTTY); 856 } 857} 858 859void 860pps_init(struct pps_state *pps) 861{ 862 pps->ppscap |= PPS_TSFMT_TSPEC; 863 if (pps->ppscap & PPS_CAPTUREASSERT) 864 pps->ppscap |= PPS_OFFSETASSERT; 865 if (pps->ppscap & PPS_CAPTURECLEAR) 866 pps->ppscap |= PPS_OFFSETCLEAR; 867#ifdef PPS_SYNC 868 if (pps->ppscap & PPS_CAPTUREASSERT) 869 pps->ppscap |= PPS_HARDPPSONASSERT; 870 if (pps->ppscap & PPS_CAPTURECLEAR) 871 pps->ppscap |= PPS_HARDPPSONCLEAR; 872#endif 873} 874 875void 876pps_event(struct pps_state *pps, struct timecounter *tc, unsigned count, int event) 877{ 878 struct timespec ts, *tsp, *osp; 879 u_int64_t delta; 880 unsigned tcount, *pcount; 881 int foff, fhard; 882 pps_seq_t *pseq; 883 884 /* Things would be easier with arrays... */ 885 if (event == PPS_CAPTUREASSERT) { 886 tsp = &pps->ppsinfo.assert_timestamp; 887 osp = &pps->ppsparam.assert_offset; 888 foff = pps->ppsparam.mode & PPS_OFFSETASSERT; 889 fhard = pps->ppsparam.mode & PPS_HARDPPSONASSERT; 890 pcount = &pps->ppscount[0]; 891 pseq = &pps->ppsinfo.assert_sequence; 892 } else { 893 tsp = &pps->ppsinfo.clear_timestamp; 894 osp = &pps->ppsparam.clear_offset; 895 foff = pps->ppsparam.mode & PPS_OFFSETCLEAR; 896 fhard = pps->ppsparam.mode & PPS_HARDPPSONCLEAR; 897 pcount = &pps->ppscount[1]; 898 pseq = &pps->ppsinfo.clear_sequence; 899 } 900 901 /* The timecounter changed: bail */ 902 if (!pps->ppstc || 903 pps->ppstc->tc_name != tc->tc_name || 904 tc->tc_name != timecounter->tc_name) { 905 pps->ppstc = tc; 906 *pcount = count; 907 return; 908 } 909 910 /* Nothing really happened */ 911 if (*pcount == count) 912 return; 913 914 *pcount = count; 915 916 /* Convert the count to timespec */ 917 ts.tv_sec = tc->tc_offset_sec; 918 tcount = count - tc->tc_offset_count; 919 tcount &= tc->tc_counter_mask; 920 delta = tc->tc_offset_nano; 921 delta += ((u_int64_t)tcount * tc->tc_scale_nano_f); 922 delta >>= 32; 923 delta += ((u_int64_t)tcount * tc->tc_scale_nano_i); 924 delta += boottime.tv_usec * 1000; 925 ts.tv_sec += boottime.tv_sec; 926 while (delta >= 1000000000) { 927 delta -= 1000000000; 928 ts.tv_sec++; 929 } 930 ts.tv_nsec = delta; 931 932 (*pseq)++; 933 *tsp = ts; 934 935 if (foff) { 936 timespecadd(tsp, osp); 937 if (tsp->tv_nsec < 0) { 938 tsp->tv_nsec += 1000000000; 939 tsp->tv_sec -= 1; 940 } 941 } 942#ifdef PPS_SYNC 943 if (fhard) { 944 /* magic, at its best... */ 945 tcount = count - pps->ppscount[2]; 946 pps->ppscount[2] = count; 947 tcount &= tc->tc_counter_mask; 948 delta = ((u_int64_t)tcount * tc->tc_tweak->tc_scale_nano_f); 949 delta >>= 32; 950 delta += ((u_int64_t)tcount * tc->tc_tweak->tc_scale_nano_i); 951 hardpps(tsp, delta); 952 } 953#endif 954} 955 956