kern_clock.c revision 163709
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 * 4. Neither the name of the University nor the names of its contributors 19 * may be used to endorse or promote products derived from this software 20 * without specific prior written permission. 21 * 22 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 23 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 24 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 25 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 26 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 27 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 28 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 29 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 30 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 31 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 32 * SUCH DAMAGE. 33 * 34 * @(#)kern_clock.c 8.5 (Berkeley) 1/21/94 35 */ 36 37#include <sys/cdefs.h> 38__FBSDID("$FreeBSD: head/sys/kern/kern_clock.c 163709 2006-10-26 21:42:22Z jb $"); 39 40#include "opt_device_polling.h" 41#include "opt_hwpmc_hooks.h" 42#include "opt_ntp.h" 43#include "opt_watchdog.h" 44 45#include <sys/param.h> 46#include <sys/systm.h> 47#include <sys/callout.h> 48#include <sys/kdb.h> 49#include <sys/kernel.h> 50#include <sys/lock.h> 51#include <sys/ktr.h> 52#include <sys/mutex.h> 53#include <sys/proc.h> 54#include <sys/resource.h> 55#include <sys/resourcevar.h> 56#include <sys/sched.h> 57#include <sys/signalvar.h> 58#include <sys/smp.h> 59#include <vm/vm.h> 60#include <vm/pmap.h> 61#include <vm/vm_map.h> 62#include <sys/sysctl.h> 63#include <sys/bus.h> 64#include <sys/interrupt.h> 65#include <sys/limits.h> 66#include <sys/timetc.h> 67 68#ifdef GPROF 69#include <sys/gmon.h> 70#endif 71 72#ifdef HWPMC_HOOKS 73#include <sys/pmckern.h> 74#endif 75 76#ifdef DEVICE_POLLING 77extern void hardclock_device_poll(void); 78#endif /* DEVICE_POLLING */ 79 80static void initclocks(void *dummy); 81SYSINIT(clocks, SI_SUB_CLOCKS, SI_ORDER_FIRST, initclocks, NULL) 82 83/* Some of these don't belong here, but it's easiest to concentrate them. */ 84long cp_time[CPUSTATES]; 85 86static int 87sysctl_kern_cp_time(SYSCTL_HANDLER_ARGS) 88{ 89 int error; 90#ifdef SCTL_MASK32 91 int i; 92 unsigned int cp_time32[CPUSTATES]; 93 94 if (req->flags & SCTL_MASK32) { 95 if (!req->oldptr) 96 return SYSCTL_OUT(req, 0, sizeof(cp_time32)); 97 for (i = 0; i < CPUSTATES; i++) 98 cp_time32[i] = (unsigned int)cp_time[i]; 99 error = SYSCTL_OUT(req, cp_time32, sizeof(cp_time32)); 100 } else 101#endif 102 { 103 if (!req->oldptr) 104 return SYSCTL_OUT(req, 0, sizeof(cp_time)); 105 error = SYSCTL_OUT(req, cp_time, sizeof(cp_time)); 106 } 107 return error; 108} 109 110SYSCTL_PROC(_kern, OID_AUTO, cp_time, CTLTYPE_LONG|CTLFLAG_RD, 111 0,0, sysctl_kern_cp_time, "LU", "CPU time statistics"); 112 113#ifdef SW_WATCHDOG 114#include <sys/watchdog.h> 115 116static int watchdog_ticks; 117static int watchdog_enabled; 118static void watchdog_fire(void); 119static void watchdog_config(void *, u_int, int *); 120#endif /* SW_WATCHDOG */ 121 122/* 123 * Clock handling routines. 124 * 125 * This code is written to operate with two timers that run independently of 126 * each other. 127 * 128 * The main timer, running hz times per second, is used to trigger interval 129 * timers, timeouts and rescheduling as needed. 130 * 131 * The second timer handles kernel and user profiling, 132 * and does resource use estimation. If the second timer is programmable, 133 * it is randomized to avoid aliasing between the two clocks. For example, 134 * the randomization prevents an adversary from always giving up the cpu 135 * just before its quantum expires. Otherwise, it would never accumulate 136 * cpu ticks. The mean frequency of the second timer is stathz. 137 * 138 * If no second timer exists, stathz will be zero; in this case we drive 139 * profiling and statistics off the main clock. This WILL NOT be accurate; 140 * do not do it unless absolutely necessary. 141 * 142 * The statistics clock may (or may not) be run at a higher rate while 143 * profiling. This profile clock runs at profhz. We require that profhz 144 * be an integral multiple of stathz. 145 * 146 * If the statistics clock is running fast, it must be divided by the ratio 147 * profhz/stathz for statistics. (For profiling, every tick counts.) 148 * 149 * Time-of-day is maintained using a "timecounter", which may or may 150 * not be related to the hardware generating the above mentioned 151 * interrupts. 152 */ 153 154int stathz; 155int profhz; 156int profprocs; 157int ticks; 158int psratio; 159 160/* 161 * Initialize clock frequencies and start both clocks running. 162 */ 163/* ARGSUSED*/ 164static void 165initclocks(dummy) 166 void *dummy; 167{ 168 register int i; 169 170 /* 171 * Set divisors to 1 (normal case) and let the machine-specific 172 * code do its bit. 173 */ 174 cpu_initclocks(); 175 176 /* 177 * Compute profhz/stathz, and fix profhz if needed. 178 */ 179 i = stathz ? stathz : hz; 180 if (profhz == 0) 181 profhz = i; 182 psratio = profhz / i; 183#ifdef SW_WATCHDOG 184 EVENTHANDLER_REGISTER(watchdog_list, watchdog_config, NULL, 0); 185#endif 186} 187 188/* 189 * Each time the real-time timer fires, this function is called on all CPUs. 190 * Note that hardclock() calls hardclock_cpu() for the boot CPU, so only 191 * the other CPUs in the system need to call this function. 192 */ 193void 194hardclock_cpu(int usermode) 195{ 196 struct pstats *pstats; 197 struct thread *td = curthread; 198 struct proc *p = td->td_proc; 199 200 /* 201 * Run current process's virtual and profile time, as needed. 202 */ 203 mtx_lock_spin_flags(&sched_lock, MTX_QUIET); 204 sched_tick(); 205#ifdef KSE 206 if (p->p_flag & P_SA) { 207 /* XXXKSE What to do? */ 208 } else { 209 pstats = p->p_stats; 210 if (usermode && 211 timevalisset(&pstats->p_timer[ITIMER_VIRTUAL].it_value) && 212 itimerdecr(&pstats->p_timer[ITIMER_VIRTUAL], tick) == 0) { 213 p->p_sflag |= PS_ALRMPEND; 214 td->td_flags |= TDF_ASTPENDING; 215 } 216 if (timevalisset(&pstats->p_timer[ITIMER_PROF].it_value) && 217 itimerdecr(&pstats->p_timer[ITIMER_PROF], tick) == 0) { 218 p->p_sflag |= PS_PROFPEND; 219 td->td_flags |= TDF_ASTPENDING; 220 } 221 } 222#else 223 pstats = p->p_stats; 224 if (usermode && 225 timevalisset(&pstats->p_timer[ITIMER_VIRTUAL].it_value) && 226 itimerdecr(&pstats->p_timer[ITIMER_VIRTUAL], tick) == 0) { 227 p->p_sflag |= PS_ALRMPEND; 228 td->td_flags |= TDF_ASTPENDING; 229 } 230 if (timevalisset(&pstats->p_timer[ITIMER_PROF].it_value) && 231 itimerdecr(&pstats->p_timer[ITIMER_PROF], tick) == 0) { 232 p->p_sflag |= PS_PROFPEND; 233 td->td_flags |= TDF_ASTPENDING; 234 } 235#endif 236 mtx_unlock_spin_flags(&sched_lock, MTX_QUIET); 237 238#ifdef HWPMC_HOOKS 239 if (PMC_CPU_HAS_SAMPLES(PCPU_GET(cpuid))) 240 PMC_CALL_HOOK_UNLOCKED(curthread, PMC_FN_DO_SAMPLES, NULL); 241#endif 242} 243 244/* 245 * The real-time timer, interrupting hz times per second. 246 */ 247void 248hardclock(int usermode, uintfptr_t pc) 249{ 250 int need_softclock = 0; 251 252 hardclock_cpu(usermode); 253 254 tc_ticktock(); 255 /* 256 * If no separate statistics clock is available, run it from here. 257 * 258 * XXX: this only works for UP 259 */ 260 if (stathz == 0) { 261 profclock(usermode, pc); 262 statclock(usermode); 263 } 264 265#ifdef DEVICE_POLLING 266 hardclock_device_poll(); /* this is very short and quick */ 267#endif /* DEVICE_POLLING */ 268 269 /* 270 * Process callouts at a very low cpu priority, so we don't keep the 271 * relatively high clock interrupt priority any longer than necessary. 272 */ 273 mtx_lock_spin_flags(&callout_lock, MTX_QUIET); 274 ticks++; 275 if (!TAILQ_EMPTY(&callwheel[ticks & callwheelmask])) { 276 need_softclock = 1; 277 } else if (softticks + 1 == ticks) 278 ++softticks; 279 mtx_unlock_spin_flags(&callout_lock, MTX_QUIET); 280 281 /* 282 * swi_sched acquires sched_lock, so we don't want to call it with 283 * callout_lock held; incorrect locking order. 284 */ 285 if (need_softclock) 286 swi_sched(softclock_ih, 0); 287 288#ifdef SW_WATCHDOG 289 if (watchdog_enabled > 0 && --watchdog_ticks <= 0) 290 watchdog_fire(); 291#endif /* SW_WATCHDOG */ 292} 293 294/* 295 * Compute number of ticks in the specified amount of time. 296 */ 297int 298tvtohz(tv) 299 struct timeval *tv; 300{ 301 register unsigned long ticks; 302 register long sec, usec; 303 304 /* 305 * If the number of usecs in the whole seconds part of the time 306 * difference fits in a long, then the total number of usecs will 307 * fit in an unsigned long. Compute the total and convert it to 308 * ticks, rounding up and adding 1 to allow for the current tick 309 * to expire. Rounding also depends on unsigned long arithmetic 310 * to avoid overflow. 311 * 312 * Otherwise, if the number of ticks in the whole seconds part of 313 * the time difference fits in a long, then convert the parts to 314 * ticks separately and add, using similar rounding methods and 315 * overflow avoidance. This method would work in the previous 316 * case but it is slightly slower and assumes that hz is integral. 317 * 318 * Otherwise, round the time difference down to the maximum 319 * representable value. 320 * 321 * If ints have 32 bits, then the maximum value for any timeout in 322 * 10ms ticks is 248 days. 323 */ 324 sec = tv->tv_sec; 325 usec = tv->tv_usec; 326 if (usec < 0) { 327 sec--; 328 usec += 1000000; 329 } 330 if (sec < 0) { 331#ifdef DIAGNOSTIC 332 if (usec > 0) { 333 sec++; 334 usec -= 1000000; 335 } 336 printf("tvotohz: negative time difference %ld sec %ld usec\n", 337 sec, usec); 338#endif 339 ticks = 1; 340 } else if (sec <= LONG_MAX / 1000000) 341 ticks = (sec * 1000000 + (unsigned long)usec + (tick - 1)) 342 / tick + 1; 343 else if (sec <= LONG_MAX / hz) 344 ticks = sec * hz 345 + ((unsigned long)usec + (tick - 1)) / tick + 1; 346 else 347 ticks = LONG_MAX; 348 if (ticks > INT_MAX) 349 ticks = INT_MAX; 350 return ((int)ticks); 351} 352 353/* 354 * Start profiling on a process. 355 * 356 * Kernel profiling passes proc0 which never exits and hence 357 * keeps the profile clock running constantly. 358 */ 359void 360startprofclock(p) 361 register struct proc *p; 362{ 363 364 /* 365 * XXX; Right now sched_lock protects statclock(), but perhaps 366 * it should be protected later on by a time_lock, which would 367 * cover psdiv, etc. as well. 368 */ 369 PROC_LOCK_ASSERT(p, MA_OWNED); 370 if (p->p_flag & P_STOPPROF) 371 return; 372 if ((p->p_flag & P_PROFIL) == 0) { 373 mtx_lock_spin(&sched_lock); 374 p->p_flag |= P_PROFIL; 375 if (++profprocs == 1) 376 cpu_startprofclock(); 377 mtx_unlock_spin(&sched_lock); 378 } 379} 380 381/* 382 * Stop profiling on a process. 383 */ 384void 385stopprofclock(p) 386 register struct proc *p; 387{ 388 389 PROC_LOCK_ASSERT(p, MA_OWNED); 390 if (p->p_flag & P_PROFIL) { 391 if (p->p_profthreads != 0) { 392 p->p_flag |= P_STOPPROF; 393 while (p->p_profthreads != 0) 394 msleep(&p->p_profthreads, &p->p_mtx, PPAUSE, 395 "stopprof", 0); 396 p->p_flag &= ~P_STOPPROF; 397 } 398 if ((p->p_flag & P_PROFIL) == 0) 399 return; 400 mtx_lock_spin(&sched_lock); 401 p->p_flag &= ~P_PROFIL; 402 if (--profprocs == 0) 403 cpu_stopprofclock(); 404 mtx_unlock_spin(&sched_lock); 405 } 406} 407 408/* 409 * Statistics clock. Grab profile sample, and if divider reaches 0, 410 * do process and kernel statistics. Most of the statistics are only 411 * used by user-level statistics programs. The main exceptions are 412 * ke->ke_uticks, p->p_rux.rux_sticks, p->p_rux.rux_iticks, and p->p_estcpu. 413 * This should be called by all active processors. 414 */ 415void 416statclock(int usermode) 417{ 418 struct rusage *ru; 419 struct vmspace *vm; 420 struct thread *td; 421 struct proc *p; 422 long rss; 423 424 td = curthread; 425 p = td->td_proc; 426 427 mtx_lock_spin_flags(&sched_lock, MTX_QUIET); 428 if (usermode) { 429 /* 430 * Charge the time as appropriate. 431 */ 432#ifdef KSE 433 if (p->p_flag & P_SA) 434 thread_statclock(1); 435#endif 436 td->td_uticks++; 437 if (p->p_nice > NZERO) 438 cp_time[CP_NICE]++; 439 else 440 cp_time[CP_USER]++; 441 } else { 442 /* 443 * Came from kernel mode, so we were: 444 * - handling an interrupt, 445 * - doing syscall or trap work on behalf of the current 446 * user process, or 447 * - spinning in the idle loop. 448 * Whichever it is, charge the time as appropriate. 449 * Note that we charge interrupts to the current process, 450 * regardless of whether they are ``for'' that process, 451 * so that we know how much of its real time was spent 452 * in ``non-process'' (i.e., interrupt) work. 453 */ 454 if ((td->td_pflags & TDP_ITHREAD) || 455 td->td_intr_nesting_level >= 2) { 456 td->td_iticks++; 457 cp_time[CP_INTR]++; 458 } else { 459#ifdef KSE 460 if (p->p_flag & P_SA) 461 thread_statclock(0); 462#endif 463 td->td_pticks++; 464 td->td_sticks++; 465 if (td != PCPU_GET(idlethread)) 466 cp_time[CP_SYS]++; 467 else 468 cp_time[CP_IDLE]++; 469 } 470 } 471 CTR4(KTR_SCHED, "statclock: %p(%s) prio %d stathz %d", 472 td, td->td_proc->p_comm, td->td_priority, (stathz)?stathz:hz); 473 474 sched_clock(td); 475 476 /* Update resource usage integrals and maximums. */ 477 MPASS(p->p_stats != NULL); 478 MPASS(p->p_vmspace != NULL); 479 vm = p->p_vmspace; 480 ru = &p->p_stats->p_ru; 481 ru->ru_ixrss += pgtok(vm->vm_tsize); 482 ru->ru_idrss += pgtok(vm->vm_dsize); 483 ru->ru_isrss += pgtok(vm->vm_ssize); 484 rss = pgtok(vmspace_resident_count(vm)); 485 if (ru->ru_maxrss < rss) 486 ru->ru_maxrss = rss; 487 mtx_unlock_spin_flags(&sched_lock, MTX_QUIET); 488} 489 490void 491profclock(int usermode, uintfptr_t pc) 492{ 493 struct thread *td; 494#ifdef GPROF 495 struct gmonparam *g; 496 uintfptr_t i; 497#endif 498 499 td = curthread; 500 if (usermode) { 501 /* 502 * Came from user mode; CPU was in user state. 503 * If this process is being profiled, record the tick. 504 * if there is no related user location yet, don't 505 * bother trying to count it. 506 */ 507 if (td->td_proc->p_flag & P_PROFIL) 508 addupc_intr(td, pc, 1); 509 } 510#ifdef GPROF 511 else { 512 /* 513 * Kernel statistics are just like addupc_intr, only easier. 514 */ 515 g = &_gmonparam; 516 if (g->state == GMON_PROF_ON && pc >= g->lowpc) { 517 i = PC_TO_I(g, pc); 518 if (i < g->textsize) { 519 KCOUNT(g, i)++; 520 } 521 } 522 } 523#endif 524} 525 526/* 527 * Return information about system clocks. 528 */ 529static int 530sysctl_kern_clockrate(SYSCTL_HANDLER_ARGS) 531{ 532 struct clockinfo clkinfo; 533 /* 534 * Construct clockinfo structure. 535 */ 536 bzero(&clkinfo, sizeof(clkinfo)); 537 clkinfo.hz = hz; 538 clkinfo.tick = tick; 539 clkinfo.profhz = profhz; 540 clkinfo.stathz = stathz ? stathz : hz; 541 return (sysctl_handle_opaque(oidp, &clkinfo, sizeof clkinfo, req)); 542} 543 544SYSCTL_PROC(_kern, KERN_CLOCKRATE, clockrate, CTLTYPE_STRUCT|CTLFLAG_RD, 545 0, 0, sysctl_kern_clockrate, "S,clockinfo", 546 "Rate and period of various kernel clocks"); 547 548#ifdef SW_WATCHDOG 549 550static void 551watchdog_config(void *unused __unused, u_int cmd, int *err) 552{ 553 u_int u; 554 555 u = cmd & WD_INTERVAL; 556 if ((cmd & WD_ACTIVE) && u >= WD_TO_1SEC) { 557 watchdog_ticks = (1 << (u - WD_TO_1SEC)) * hz; 558 watchdog_enabled = 1; 559 *err = 0; 560 } else { 561 watchdog_enabled = 0; 562 } 563} 564 565/* 566 * Handle a watchdog timeout by dumping interrupt information and 567 * then either dropping to DDB or panicing. 568 */ 569static void 570watchdog_fire(void) 571{ 572 int nintr; 573 u_int64_t inttotal; 574 u_long *curintr; 575 char *curname; 576 577 curintr = intrcnt; 578 curname = intrnames; 579 inttotal = 0; 580 nintr = eintrcnt - intrcnt; 581 582 printf("interrupt total\n"); 583 while (--nintr >= 0) { 584 if (*curintr) 585 printf("%-12s %20lu\n", curname, *curintr); 586 curname += strlen(curname) + 1; 587 inttotal += *curintr++; 588 } 589 printf("Total %20ju\n", (uintmax_t)inttotal); 590 591#ifdef KDB 592 kdb_backtrace(); 593 kdb_enter("watchdog timeout"); 594#else 595 panic("watchdog timeout"); 596#endif /* KDB */ 597} 598 599#endif /* SW_WATCHDOG */ 600