1 /*- 2 * Copyright (c) 1982, 1986, 1991, 1993 3 * The Regents of the University of California. All rights reserved. 4 * (c) UNIX System Laboratories, Inc. 5 * All or some portions of this file are derived from material licensed 6 * to the University of California by American Telephone and Telegraph 7 * Co. or Unix System Laboratories, Inc. and are reproduced herein with 8 * the permission of UNIX System Laboratories, Inc. 9 * 10 * Redistribution and use in source and binary forms, with or without 11 * modification, are permitted provided that the following conditions 12 * are met: 13 * 1. Redistributions of source code must retain the above copyright 14 * notice, this list of conditions and the following disclaimer. 15 * 2. Redistributions in binary form must reproduce the above copyright 16 * notice, this list of conditions and the following disclaimer in the 17 * documentation and/or other materials provided with the distribution. 18 * 3. All advertising materials mentioning features or use of this software 19 * must display the following acknowledgement: 20 * This product includes software developed by the University of 21 * California, Berkeley and its contributors. 22 * 4. Neither the name of the University nor the names of its contributors 23 * may be used to endorse or promote products derived from this software 24 * without specific prior written permission. 25 * 26 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 27 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 28 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 29 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 30 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 31 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 32 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 33 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 34 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 35 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 36 * SUCH DAMAGE. 37 * 38 * @(#)kern_clock.c 8.5 (Berkeley) 1/21/94 39 */ 40 41 #include <sys/cdefs.h> 42 __FBSDID("$FreeBSD$"); 43 44 #include "opt_ntp.h" 45 #include "opt_ddb.h" 46 #include "opt_watchdog.h" 47 48 #include <sys/param.h> 49 #include <sys/systm.h> 50 #include <sys/callout.h> 51 #include <sys/kernel.h> 52 #include <sys/lock.h> 53 #include <sys/ktr.h> 54 #include <sys/mutex.h> 55 #include <sys/proc.h> 56 #include <sys/resource.h> 57 #include <sys/resourcevar.h> 58 #include <sys/sched.h> 59 #include <sys/signalvar.h> 60 #include <sys/smp.h> 61 #include <vm/vm.h> 62 #include <vm/pmap.h> 63 #include <vm/vm_map.h> 64 #include <sys/sysctl.h> 65 #include <sys/bus.h> 66 #include <sys/interrupt.h> 67 #include <sys/limits.h> 68 #include <sys/timetc.h> 69 70 #include <machine/cpu.h> 71 72 #ifdef GPROF 73 #include <sys/gmon.h> 74 #endif 75 76 #ifdef DDB 77 #include <ddb/ddb.h> 78 #endif 79 80 #ifdef DEVICE_POLLING 81 extern void hardclock_device_poll(void); 82 #endif /* DEVICE_POLLING */ 83 84 static void initclocks(void *dummy); 85 SYSINIT(clocks, SI_SUB_CLOCKS, SI_ORDER_FIRST, initclocks, NULL) 86 87 /* Some of these don't belong here, but it's easiest to concentrate them. */ 88 long cp_time[CPUSTATES]; 89 90 SYSCTL_OPAQUE(_kern, OID_AUTO, cp_time, CTLFLAG_RD, &cp_time, sizeof(cp_time), 91 "LU", "CPU time statistics"); 92 93 #ifdef SW_WATCHDOG 94 #include <sys/watchdog.h> 95 96 static int watchdog_ticks; 97 static int watchdog_enabled; 98 static void watchdog_fire(void); 99 static void watchdog_config(void *, u_int, int *); 100 #endif /* SW_WATCHDOG */ 101 102 /* 103 * Clock handling routines. 104 * 105 * This code is written to operate with two timers that run independently of 106 * each other. 107 * 108 * The main timer, running hz times per second, is used to trigger interval 109 * timers, timeouts and rescheduling as needed. 110 * 111 * The second timer handles kernel and user profiling, 112 * and does resource use estimation. If the second timer is programmable, 113 * it is randomized to avoid aliasing between the two clocks. For example, 114 * the randomization prevents an adversary from always giving up the cpu 115 * just before its quantum expires. Otherwise, it would never accumulate 116 * cpu ticks. The mean frequency of the second timer is stathz. 117 * 118 * If no second timer exists, stathz will be zero; in this case we drive 119 * profiling and statistics off the main clock. This WILL NOT be accurate; 120 * do not do it unless absolutely necessary. 121 * 122 * The statistics clock may (or may not) be run at a higher rate while 123 * profiling. This profile clock runs at profhz. We require that profhz 124 * be an integral multiple of stathz. 125 * 126 * If the statistics clock is running fast, it must be divided by the ratio 127 * profhz/stathz for statistics. (For profiling, every tick counts.) 128 * 129 * Time-of-day is maintained using a "timecounter", which may or may 130 * not be related to the hardware generating the above mentioned 131 * interrupts. 132 */ 133 134 int stathz; 135 int profhz; 136 int profprocs; 137 int ticks; 138 int psratio; 139 140 /* 141 * Initialize clock frequencies and start both clocks running. 142 */ 143 /* ARGSUSED*/ 144 static void 145 initclocks(dummy) 146 void *dummy; 147 { 148 register int i; 149 150 /* 151 * Set divisors to 1 (normal case) and let the machine-specific 152 * code do its bit. 153 */ 154 cpu_initclocks(); 155 156 /* 157 * Compute profhz/stathz, and fix profhz if needed. 158 */ 159 i = stathz ? stathz : hz; 160 if (profhz == 0) 161 profhz = i; 162 psratio = profhz / i; 163 #ifdef SW_WATCHDOG 164 EVENTHANDLER_REGISTER(watchdog_list, watchdog_config, NULL, 0); 165 #endif 166 } 167 168 /* 169 * Each time the real-time timer fires, this function is called on all CPUs. 170 * Note that hardclock() calls hardclock_process() for the boot CPU, so only 171 * the other CPUs in the system need to call this function. 172 */ 173 void 174 hardclock_process(frame) 175 register struct clockframe *frame; 176 { 177 struct pstats *pstats; 178 struct thread *td = curthread; 179 struct proc *p = td->td_proc; 180 181 /* 182 * Run current process's virtual and profile time, as needed. 183 */ 184 mtx_lock_spin_flags(&sched_lock, MTX_QUIET); 185 if (p->p_flag & P_SA) { 186 /* XXXKSE What to do? */ 187 } else { 188 pstats = p->p_stats; 189 if (CLKF_USERMODE(frame) && 190 timevalisset(&pstats->p_timer[ITIMER_VIRTUAL].it_value) && 191 itimerdecr(&pstats->p_timer[ITIMER_VIRTUAL], tick) == 0) { 192 p->p_sflag |= PS_ALRMPEND; 193 td->td_flags |= TDF_ASTPENDING; 194 } 195 if (timevalisset(&pstats->p_timer[ITIMER_PROF].it_value) && 196 itimerdecr(&pstats->p_timer[ITIMER_PROF], tick) == 0) { 197 p->p_sflag |= PS_PROFPEND; 198 td->td_flags |= TDF_ASTPENDING; 199 } 200 } 201 mtx_unlock_spin_flags(&sched_lock, MTX_QUIET); 202 } 203 204 /* 205 * The real-time timer, interrupting hz times per second. 206 */ 207 void 208 hardclock(frame) 209 register struct clockframe *frame; 210 { 211 int need_softclock = 0; 212 213 CTR0(KTR_CLK, "hardclock fired"); 214 hardclock_process(frame); 215 216 tc_ticktock(); 217 /* 218 * If no separate statistics clock is available, run it from here. 219 * 220 * XXX: this only works for UP 221 */ 222 if (stathz == 0) { 223 profclock(frame); 224 statclock(frame); 225 } 226 227 #ifdef DEVICE_POLLING 228 hardclock_device_poll(); /* this is very short and quick */ 229 #endif /* DEVICE_POLLING */ 230 231 /* 232 * Process callouts at a very low cpu priority, so we don't keep the 233 * relatively high clock interrupt priority any longer than necessary. 234 */ 235 mtx_lock_spin_flags(&callout_lock, MTX_QUIET); 236 ticks++; 237 if (TAILQ_FIRST(&callwheel[ticks & callwheelmask]) != NULL) { 238 need_softclock = 1; 239 } else if (softticks + 1 == ticks) 240 ++softticks; 241 mtx_unlock_spin_flags(&callout_lock, MTX_QUIET); 242 243 /* 244 * swi_sched acquires sched_lock, so we don't want to call it with 245 * callout_lock held; incorrect locking order. 246 */ 247 if (need_softclock) 248 swi_sched(softclock_ih, 0); 249 250 #ifdef SW_WATCHDOG 251 if (watchdog_enabled > 0 && --watchdog_ticks <= 0) 252 watchdog_fire(); 253 #endif /* SW_WATCHDOG */ 254 } 255 256 /* 257 * Compute number of ticks in the specified amount of time. 258 */ 259 int 260 tvtohz(tv) 261 struct timeval *tv; 262 { 263 register unsigned long ticks; 264 register long sec, usec; 265 266 /* 267 * If the number of usecs in the whole seconds part of the time 268 * difference fits in a long, then the total number of usecs will 269 * fit in an unsigned long. Compute the total and convert it to 270 * ticks, rounding up and adding 1 to allow for the current tick 271 * to expire. Rounding also depends on unsigned long arithmetic 272 * to avoid overflow. 273 * 274 * Otherwise, if the number of ticks in the whole seconds part of 275 * the time difference fits in a long, then convert the parts to 276 * ticks separately and add, using similar rounding methods and 277 * overflow avoidance. This method would work in the previous 278 * case but it is slightly slower and assumes that hz is integral. 279 * 280 * Otherwise, round the time difference down to the maximum 281 * representable value. 282 * 283 * If ints have 32 bits, then the maximum value for any timeout in 284 * 10ms ticks is 248 days. 285 */ 286 sec = tv->tv_sec; 287 usec = tv->tv_usec; 288 if (usec < 0) { 289 sec--; 290 usec += 1000000; 291 } 292 if (sec < 0) { 293 #ifdef DIAGNOSTIC 294 if (usec > 0) { 295 sec++; 296 usec -= 1000000; 297 } 298 printf("tvotohz: negative time difference %ld sec %ld usec\n", 299 sec, usec); 300 #endif 301 ticks = 1; 302 } else if (sec <= LONG_MAX / 1000000) 303 ticks = (sec * 1000000 + (unsigned long)usec + (tick - 1)) 304 / tick + 1; 305 else if (sec <= LONG_MAX / hz) 306 ticks = sec * hz 307 + ((unsigned long)usec + (tick - 1)) / tick + 1; 308 else 309 ticks = LONG_MAX; 310 if (ticks > INT_MAX) 311 ticks = INT_MAX; 312 return ((int)ticks); 313 } 314 315 /* 316 * Start profiling on a process. 317 * 318 * Kernel profiling passes proc0 which never exits and hence 319 * keeps the profile clock running constantly. 320 */ 321 void 322 startprofclock(p) 323 register struct proc *p; 324 { 325 326 /* 327 * XXX; Right now sched_lock protects statclock(), but perhaps 328 * it should be protected later on by a time_lock, which would 329 * cover psdiv, etc. as well. 330 */ 331 PROC_LOCK_ASSERT(p, MA_OWNED); 332 if (p->p_flag & P_STOPPROF) 333 return; 334 if ((p->p_flag & P_PROFIL) == 0) { 335 mtx_lock_spin(&sched_lock); 336 p->p_flag |= P_PROFIL; 337 if (++profprocs == 1) 338 cpu_startprofclock(); 339 mtx_unlock_spin(&sched_lock); 340 } 341 } 342 343 /* 344 * Stop profiling on a process. 345 */ 346 void 347 stopprofclock(p) 348 register struct proc *p; 349 { 350 351 PROC_LOCK_ASSERT(p, MA_OWNED); 352 if (p->p_flag & P_PROFIL) { 353 if (p->p_profthreads != 0) { 354 p->p_flag |= P_STOPPROF; 355 while (p->p_profthreads != 0) 356 msleep(&p->p_profthreads, &p->p_mtx, PPAUSE, 357 "stopprof", 0); 358 p->p_flag &= ~P_STOPPROF; 359 } 360 mtx_lock_spin(&sched_lock); 361 p->p_flag &= ~P_PROFIL; 362 if (--profprocs == 0) 363 cpu_stopprofclock(); 364 mtx_unlock_spin(&sched_lock); 365 } 366 } 367 368 /* 369 * Statistics clock. Grab profile sample, and if divider reaches 0, 370 * do process and kernel statistics. Most of the statistics are only 371 * used by user-level statistics programs. The main exceptions are 372 * ke->ke_uticks, p->p_sticks, p->p_iticks, and p->p_estcpu. 373 * This should be called by all active processors. 374 */ 375 void 376 statclock(frame) 377 register struct clockframe *frame; 378 { 379 struct pstats *pstats; 380 struct rusage *ru; 381 struct vmspace *vm; 382 struct thread *td; 383 struct proc *p; 384 long rss; 385 386 td = curthread; 387 p = td->td_proc; 388 389 mtx_lock_spin_flags(&sched_lock, MTX_QUIET); 390 if (CLKF_USERMODE(frame)) { 391 /* 392 * Charge the time as appropriate. 393 */ 394 if (p->p_flag & P_SA) 395 thread_statclock(1); 396 p->p_uticks++; 397 if (td->td_ksegrp->kg_nice > NZERO) 398 cp_time[CP_NICE]++; 399 else 400 cp_time[CP_USER]++; 401 } else { 402 /* 403 * Came from kernel mode, so we were: 404 * - handling an interrupt, 405 * - doing syscall or trap work on behalf of the current 406 * user process, or 407 * - spinning in the idle loop. 408 * Whichever it is, charge the time as appropriate. 409 * Note that we charge interrupts to the current process, 410 * regardless of whether they are ``for'' that process, 411 * so that we know how much of its real time was spent 412 * in ``non-process'' (i.e., interrupt) work. 413 */ 414 if ((td->td_ithd != NULL) || td->td_intr_nesting_level >= 2) { 415 p->p_iticks++; 416 cp_time[CP_INTR]++; 417 } else { 418 if (p->p_flag & P_SA) 419 thread_statclock(0); 420 td->td_sticks++; 421 p->p_sticks++; 422 if (p != PCPU_GET(idlethread)->td_proc) 423 cp_time[CP_SYS]++; 424 else 425 cp_time[CP_IDLE]++; 426 } 427 } 428 429 sched_clock(td); 430 431 /* Update resource usage integrals and maximums. */ 432 if ((pstats = p->p_stats) != NULL && 433 (ru = &pstats->p_ru) != NULL && 434 (vm = p->p_vmspace) != NULL) { 435 ru->ru_ixrss += pgtok(vm->vm_tsize); 436 ru->ru_idrss += pgtok(vm->vm_dsize); 437 ru->ru_isrss += pgtok(vm->vm_ssize); 438 rss = pgtok(vmspace_resident_count(vm)); 439 if (ru->ru_maxrss < rss) 440 ru->ru_maxrss = rss; 441 } 442 mtx_unlock_spin_flags(&sched_lock, MTX_QUIET); 443 } 444 445 void 446 profclock(frame) 447 register struct clockframe *frame; 448 { 449 struct thread *td; 450 #ifdef GPROF 451 struct gmonparam *g; 452 int i; 453 #endif 454 455 td = curthread; 456 if (CLKF_USERMODE(frame)) { 457 /* 458 * Came from user mode; CPU was in user state. 459 * If this process is being profiled, record the tick. 460 * if there is no related user location yet, don't 461 * bother trying to count it. 462 */ 463 td = curthread; 464 if (td->td_proc->p_flag & P_PROFIL) 465 addupc_intr(td, CLKF_PC(frame), 1); 466 } 467 #ifdef GPROF 468 else { 469 /* 470 * Kernel statistics are just like addupc_intr, only easier. 471 */ 472 g = &_gmonparam; 473 if (g->state == GMON_PROF_ON) { 474 i = CLKF_PC(frame) - g->lowpc; 475 if (i < g->textsize) { 476 i /= HISTFRACTION * sizeof(*g->kcount); 477 g->kcount[i]++; 478 } 479 } 480 } 481 #endif 482 } 483 484 /* 485 * Return information about system clocks. 486 */ 487 static int 488 sysctl_kern_clockrate(SYSCTL_HANDLER_ARGS) 489 { 490 struct clockinfo clkinfo; 491 /* 492 * Construct clockinfo structure. 493 */ 494 bzero(&clkinfo, sizeof(clkinfo)); 495 clkinfo.hz = hz; 496 clkinfo.tick = tick; 497 clkinfo.profhz = profhz; 498 clkinfo.stathz = stathz ? stathz : hz; 499 return (sysctl_handle_opaque(oidp, &clkinfo, sizeof clkinfo, req)); 500 } 501 502 SYSCTL_PROC(_kern, KERN_CLOCKRATE, clockrate, CTLTYPE_STRUCT|CTLFLAG_RD, 503 0, 0, sysctl_kern_clockrate, "S,clockinfo", 504 "Rate and period of various kernel clocks"); 505 506 #ifdef SW_WATCHDOG 507 508 static void 509 watchdog_config(void *unused __unused, u_int cmd, int *err) 510 { 511 u_int u; 512 513 if (cmd) { 514 u = cmd & WD_INTERVAL; 515 if (u < WD_TO_1SEC) 516 return; 517 watchdog_ticks = (1 << (u - WD_TO_1SEC)) * hz; 518 watchdog_enabled = 1; 519 *err = 0; 520 } else { 521 watchdog_enabled = 0; 522 } 523 } 524 525 /* 526 * Handle a watchdog timeout by dumping interrupt information and 527 * then either dropping to DDB or panicing. 528 */ 529 static void 530 watchdog_fire(void) 531 { 532 int nintr; 533 u_int64_t inttotal; 534 u_long *curintr; 535 char *curname; 536 537 curintr = intrcnt; 538 curname = intrnames; 539 inttotal = 0; 540 nintr = eintrcnt - intrcnt; 541 542 printf("interrupt total\n"); 543 while (--nintr >= 0) { 544 if (*curintr) 545 printf("%-12s %20lu\n", curname, *curintr); 546 curname += strlen(curname) + 1; 547 inttotal += *curintr++; 548 } 549 printf("Total %20ju\n", (uintmax_t)inttotal); 550 551 #ifdef DDB 552 db_print_backtrace(); 553 Debugger("watchdog timeout"); 554 #else /* !DDB */ 555 panic("watchdog timeout"); 556 #endif /* DDB */ 557 } 558 559 #endif /* SW_WATCHDOG */ 560