1 /* $OpenBSD: kern_clock.c,v 1.43 2004/06/09 20:18:28 art Exp $ */ 2 /* $NetBSD: kern_clock.c,v 1.34 1996/06/09 04:51:03 briggs Exp $ */ 3 4 /*- 5 * Copyright (c) 1982, 1986, 1991, 1993 6 * The Regents of the University of California. All rights reserved. 7 * (c) UNIX System Laboratories, Inc. 8 * All or some portions of this file are derived from material licensed 9 * to the University of California by American Telephone and Telegraph 10 * Co. or Unix System Laboratories, Inc. and are reproduced herein with 11 * the permission of UNIX System Laboratories, Inc. 12 * 13 * Redistribution and use in source and binary forms, with or without 14 * modification, are permitted provided that the following conditions 15 * are met: 16 * 1. Redistributions of source code must retain the above copyright 17 * notice, this list of conditions and the following disclaimer. 18 * 2. Redistributions in binary form must reproduce the above copyright 19 * notice, this list of conditions and the following disclaimer in the 20 * documentation and/or other materials provided with the distribution. 21 * 3. Neither the name of the University nor the names of its contributors 22 * may be used to endorse or promote products derived from this software 23 * without specific prior written permission. 24 * 25 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 26 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 27 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 28 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 29 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 30 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 31 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 32 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 33 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 34 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 35 * SUCH DAMAGE. 36 * 37 * @(#)kern_clock.c 8.5 (Berkeley) 1/21/94 38 */ 39 40 #include <sys/param.h> 41 #include <sys/systm.h> 42 #include <sys/dkstat.h> 43 #include <sys/timeout.h> 44 #include <sys/kernel.h> 45 #include <sys/limits.h> 46 #include <sys/proc.h> 47 #include <sys/resourcevar.h> 48 #include <sys/signalvar.h> 49 #include <uvm/uvm_extern.h> 50 #include <sys/sysctl.h> 51 #include <sys/sched.h> 52 53 #include <machine/cpu.h> 54 55 #ifdef GPROF 56 #include <sys/gmon.h> 57 #endif 58 59 /* 60 * Clock handling routines. 61 * 62 * This code is written to operate with two timers that run independently of 63 * each other. The main clock, running hz times per second, is used to keep 64 * track of real time. The second timer handles kernel and user profiling, 65 * and does resource use estimation. If the second timer is programmable, 66 * it is randomized to avoid aliasing between the two clocks. For example, 67 * the randomization prevents an adversary from always giving up the cpu 68 * just before its quantum expires. Otherwise, it would never accumulate 69 * cpu ticks. The mean frequency of the second timer is stathz. 70 * 71 * If no second timer exists, stathz will be zero; in this case we drive 72 * profiling and statistics off the main clock. This WILL NOT be accurate; 73 * do not do it unless absolutely necessary. 74 * 75 * The statistics clock may (or may not) be run at a higher rate while 76 * profiling. This profile clock runs at profhz. We require that profhz 77 * be an integral multiple of stathz. 78 * 79 * If the statistics clock is running fast, it must be divided by the ratio 80 * profhz/stathz for statistics. (For profiling, every tick counts.) 81 */ 82 83 /* 84 * Bump a timeval by a small number of usec's. 85 */ 86 #define BUMPTIME(t, usec) { \ 87 register volatile struct timeval *tp = (t); \ 88 register long us; \ 89 \ 90 tp->tv_usec = us = tp->tv_usec + (usec); \ 91 if (us >= 1000000) { \ 92 tp->tv_usec = us - 1000000; \ 93 tp->tv_sec++; \ 94 } \ 95 } 96 97 int stathz; 98 int schedhz; 99 int profhz; 100 int profprocs; 101 int ticks; 102 static int psdiv, pscnt; /* prof => stat divider */ 103 int psratio; /* ratio: prof / stat */ 104 int tickfix, tickfixinterval; /* used if tick not really integral */ 105 static int tickfixcnt; /* accumulated fractional error */ 106 107 long cp_time[CPUSTATES]; 108 109 volatile struct timeval time 110 __attribute__((__aligned__(__alignof__(quad_t)))); 111 volatile struct timeval mono_time; 112 113 #ifdef __HAVE_GENERIC_SOFT_INTERRUPTS 114 void *softclock_si; 115 void generic_softclock(void *); 116 117 void 118 generic_softclock(void *ignore) 119 { 120 /* 121 * XXX - dont' commit, just a dummy wrapper until we learn everyone 122 * deal with a changed proto for softclock(). 123 */ 124 softclock(); 125 } 126 #endif 127 128 /* 129 * Initialize clock frequencies and start both clocks running. 130 */ 131 void 132 initclocks() 133 { 134 int i; 135 136 #ifdef __HAVE_GENERIC_SOFT_INTERRUPTS 137 softclock_si = softintr_establish(IPL_SOFTCLOCK, generic_softclock, NULL); 138 if (softclock_si == NULL) 139 panic("initclocks: unable to register softclock intr"); 140 #endif 141 142 /* 143 * Set divisors to 1 (normal case) and let the machine-specific 144 * code do its bit. 145 */ 146 psdiv = pscnt = 1; 147 cpu_initclocks(); 148 149 /* 150 * Compute profhz/stathz, and fix profhz if needed. 151 */ 152 i = stathz ? stathz : hz; 153 if (profhz == 0) 154 profhz = i; 155 psratio = profhz / i; 156 } 157 158 /* 159 * The real-time timer, interrupting hz times per second. 160 */ 161 void 162 hardclock(struct clockframe *frame) 163 { 164 struct proc *p; 165 int delta; 166 extern int tickdelta; 167 extern long timedelta; 168 #ifdef __HAVE_CPUINFO 169 struct cpu_info *ci = curcpu(); 170 #endif 171 172 p = curproc; 173 if (p) { 174 register struct pstats *pstats; 175 176 /* 177 * Run current process's virtual and profile time, as needed. 178 */ 179 pstats = p->p_stats; 180 if (CLKF_USERMODE(frame) && 181 timerisset(&pstats->p_timer[ITIMER_VIRTUAL].it_value) && 182 itimerdecr(&pstats->p_timer[ITIMER_VIRTUAL], tick) == 0) 183 psignal(p, SIGVTALRM); 184 if (timerisset(&pstats->p_timer[ITIMER_PROF].it_value) && 185 itimerdecr(&pstats->p_timer[ITIMER_PROF], tick) == 0) 186 psignal(p, SIGPROF); 187 } 188 189 /* 190 * If no separate statistics clock is available, run it from here. 191 */ 192 if (stathz == 0) 193 statclock(frame); 194 195 #ifdef __HAVE_CPUINFO 196 if (--ci->ci_schedstate.spc_rrticks <= 0) 197 roundrobin(ci); 198 #endif 199 200 /* 201 * Increment the time-of-day. The increment is normally just 202 * ``tick''. If the machine is one which has a clock frequency 203 * such that ``hz'' would not divide the second evenly into 204 * milliseconds, a periodic adjustment must be applied. Finally, 205 * if we are still adjusting the time (see adjtime()), 206 * ``tickdelta'' may also be added in. 207 */ 208 ticks++; 209 delta = tick; 210 211 if (tickfix) { 212 tickfixcnt += tickfix; 213 if (tickfixcnt >= tickfixinterval) { 214 delta++; 215 tickfixcnt -= tickfixinterval; 216 } 217 } 218 /* Imprecise 4bsd adjtime() handling */ 219 if (timedelta != 0) { 220 delta += tickdelta; 221 timedelta -= tickdelta; 222 } 223 224 #ifdef notyet 225 microset(); 226 #endif 227 228 BUMPTIME(&time, delta); 229 BUMPTIME(&mono_time, delta); 230 231 #ifdef CPU_CLOCKUPDATE 232 CPU_CLOCKUPDATE(); 233 #endif 234 235 /* 236 * Update real-time timeout queue. 237 * Process callouts at a very low cpu priority, so we don't keep the 238 * relatively high clock interrupt priority any longer than necessary. 239 */ 240 if (timeout_hardclock_update()) { 241 #ifdef __HAVE_GENERIC_SOFT_INTERRUPTS 242 softintr_schedule(softclock_si); 243 #else 244 setsoftclock(); 245 #endif 246 } 247 } 248 249 /* 250 * Compute number of hz until specified time. Used to 251 * compute the second argument to timeout_add() from an absolute time. 252 */ 253 int 254 hzto(tv) 255 struct timeval *tv; 256 { 257 unsigned long ticks; 258 long sec, usec; 259 int s; 260 261 /* 262 * If the number of usecs in the whole seconds part of the time 263 * difference fits in a long, then the total number of usecs will 264 * fit in an unsigned long. Compute the total and convert it to 265 * ticks, rounding up and adding 1 to allow for the current tick 266 * to expire. Rounding also depends on unsigned long arithmetic 267 * to avoid overflow. 268 * 269 * Otherwise, if the number of ticks in the whole seconds part of 270 * the time difference fits in a long, then convert the parts to 271 * ticks separately and add, using similar rounding methods and 272 * overflow avoidance. This method would work in the previous 273 * case but it is slightly slower and assumes that hz is integral. 274 * 275 * Otherwise, round the time difference down to the maximum 276 * representable value. 277 * 278 * If ints have 32 bits, then the maximum value for any timeout in 279 * 10ms ticks is 248 days. 280 */ 281 s = splhigh(); 282 sec = tv->tv_sec - time.tv_sec; 283 usec = tv->tv_usec - time.tv_usec; 284 splx(s); 285 if (usec < 0) { 286 sec--; 287 usec += 1000000; 288 } 289 if (sec < 0 || (sec == 0 && usec <= 0)) { 290 ticks = 0; 291 } else if (sec <= LONG_MAX / 1000000) 292 ticks = (sec * 1000000 + (unsigned long)usec + (tick - 1)) 293 / tick + 1; 294 else if (sec <= LONG_MAX / hz) 295 ticks = sec * hz 296 + ((unsigned long)usec + (tick - 1)) / tick + 1; 297 else 298 ticks = LONG_MAX; 299 if (ticks > INT_MAX) 300 ticks = INT_MAX; 301 return ((int)ticks); 302 } 303 304 /* 305 * Compute number of hz in the specified amount of time. 306 */ 307 int 308 tvtohz(struct timeval *tv) 309 { 310 unsigned long ticks; 311 long sec, usec; 312 313 /* 314 * If the number of usecs in the whole seconds part of the time 315 * fits in a long, then the total number of usecs will 316 * fit in an unsigned long. Compute the total and convert it to 317 * ticks, rounding up and adding 1 to allow for the current tick 318 * to expire. Rounding also depends on unsigned long arithmetic 319 * to avoid overflow. 320 * 321 * Otherwise, if the number of ticks in the whole seconds part of 322 * the time fits in a long, then convert the parts to 323 * ticks separately and add, using similar rounding methods and 324 * overflow avoidance. This method would work in the previous 325 * case but it is slightly slower and assumes that hz is integral. 326 * 327 * Otherwise, round the time down to the maximum 328 * representable value. 329 * 330 * If ints have 32 bits, then the maximum value for any timeout in 331 * 10ms ticks is 248 days. 332 */ 333 sec = tv->tv_sec; 334 usec = tv->tv_usec; 335 if (sec < 0 || (sec == 0 && usec <= 0)) 336 ticks = 0; 337 else if (sec <= LONG_MAX / 1000000) 338 ticks = (sec * 1000000 + (unsigned long)usec + (tick - 1)) 339 / tick + 1; 340 else if (sec <= LONG_MAX / hz) 341 ticks = sec * hz 342 + ((unsigned long)usec + (tick - 1)) / tick + 1; 343 else 344 ticks = LONG_MAX; 345 if (ticks > INT_MAX) 346 ticks = INT_MAX; 347 return ((int)ticks); 348 } 349 350 /* 351 * Start profiling on a process. 352 * 353 * Kernel profiling passes proc0 which never exits and hence 354 * keeps the profile clock running constantly. 355 */ 356 void 357 startprofclock(p) 358 register struct proc *p; 359 { 360 int s; 361 362 if ((p->p_flag & P_PROFIL) == 0) { 363 p->p_flag |= P_PROFIL; 364 if (++profprocs == 1 && stathz != 0) { 365 s = splstatclock(); 366 psdiv = pscnt = psratio; 367 setstatclockrate(profhz); 368 splx(s); 369 } 370 } 371 } 372 373 /* 374 * Stop profiling on a process. 375 */ 376 void 377 stopprofclock(p) 378 register struct proc *p; 379 { 380 int s; 381 382 if (p->p_flag & P_PROFIL) { 383 p->p_flag &= ~P_PROFIL; 384 if (--profprocs == 0 && stathz != 0) { 385 s = splstatclock(); 386 psdiv = pscnt = 1; 387 setstatclockrate(stathz); 388 splx(s); 389 } 390 } 391 } 392 393 /* 394 * Statistics clock. Grab profile sample, and if divider reaches 0, 395 * do process and kernel statistics. 396 */ 397 void 398 statclock(struct clockframe *frame) 399 { 400 #ifdef GPROF 401 struct gmonparam *g; 402 int i; 403 #endif 404 #ifdef __HAVE_CPUINFO 405 struct cpu_info *ci = curcpu(); 406 struct schedstate_percpu *spc = &ci->ci_schedstate; 407 #else 408 static int schedclk; 409 #endif 410 struct proc *p = curproc; 411 412 #ifdef __HAVE_CPUINFO 413 /* 414 * Notice changes in divisor frequency, and adjust clock 415 * frequency accordingly. 416 */ 417 if (spc->spc_psdiv != psdiv) { 418 spc->spc_psdiv = psdiv; 419 spc->spc_pscnt = psdiv; 420 if (psdiv == 1) { 421 setstatclockrate(stathz); 422 } else { 423 setstatclockrate(profhz); 424 } 425 } 426 /* XXX Kludgey */ 427 #define pscnt spc->spc_pscnt 428 #define cp_time spc->spc_cp_time 429 #endif 430 431 if (CLKF_USERMODE(frame)) { 432 if (p->p_flag & P_PROFIL) 433 addupc_intr(p, CLKF_PC(frame)); 434 if (--pscnt > 0) 435 return; 436 /* 437 * Came from user mode; CPU was in user state. 438 * If this process is being profiled record the tick. 439 */ 440 p->p_uticks++; 441 if (p->p_nice > NZERO) 442 cp_time[CP_NICE]++; 443 else 444 cp_time[CP_USER]++; 445 } else { 446 #ifdef GPROF 447 /* 448 * Kernel statistics are just like addupc_intr, only easier. 449 */ 450 g = &_gmonparam; 451 if (g->state == GMON_PROF_ON) { 452 i = CLKF_PC(frame) - g->lowpc; 453 if (i < g->textsize) { 454 i /= HISTFRACTION * sizeof(*g->kcount); 455 g->kcount[i]++; 456 } 457 } 458 #endif 459 if (--pscnt > 0) 460 return; 461 /* 462 * Came from kernel mode, so we were: 463 * - handling an interrupt, 464 * - doing syscall or trap work on behalf of the current 465 * user process, or 466 * - spinning in the idle loop. 467 * Whichever it is, charge the time as appropriate. 468 * Note that we charge interrupts to the current process, 469 * regardless of whether they are ``for'' that process, 470 * so that we know how much of its real time was spent 471 * in ``non-process'' (i.e., interrupt) work. 472 */ 473 if (CLKF_INTR(frame)) { 474 if (p != NULL) 475 p->p_iticks++; 476 cp_time[CP_INTR]++; 477 } else if (p != NULL) { 478 p->p_sticks++; 479 cp_time[CP_SYS]++; 480 } else 481 cp_time[CP_IDLE]++; 482 } 483 pscnt = psdiv; 484 485 #ifdef __HAVE_CPUINFO 486 #undef pscnt 487 #undef cp_time 488 #endif 489 490 if (p != NULL) { 491 p->p_cpticks++; 492 /* 493 * If no schedclock is provided, call it here at ~~12-25 Hz; 494 * ~~16 Hz is best 495 */ 496 if (schedhz == 0) { 497 #ifdef __HAVE_CPUINFO 498 if ((++curcpu()->ci_schedstate.spc_schedticks & 3) == 0) 499 schedclock(p); 500 #else 501 if ((++schedclk & 3) == 0) 502 schedclock(p); 503 #endif 504 } 505 } 506 } 507 508 /* 509 * Return information about system clocks. 510 */ 511 int 512 sysctl_clockrate(where, sizep) 513 register char *where; 514 size_t *sizep; 515 { 516 struct clockinfo clkinfo; 517 518 /* 519 * Construct clockinfo structure. 520 */ 521 clkinfo.tick = tick; 522 clkinfo.tickadj = tickadj; 523 clkinfo.hz = hz; 524 clkinfo.profhz = profhz; 525 clkinfo.stathz = stathz ? stathz : hz; 526 return (sysctl_rdstruct(where, sizep, NULL, &clkinfo, sizeof(clkinfo))); 527 } 528