kern_clock.c revision 170468
150476Speter/*- 232236Shelbig * Copyright (c) 1982, 1986, 1991, 1993 332236Shelbig * The Regents of the University of California. All rights reserved. 432236Shelbig * (c) UNIX System Laboratories, Inc. 532236Shelbig * All or some portions of this file are derived from material licensed 632236Shelbig * to the University of California by American Telephone and Telegraph 732236Shelbig * Co. or Unix System Laboratories, Inc. and are reproduced herein with 832236Shelbig * the permission of UNIX System Laboratories, Inc. 932236Shelbig * 1032236Shelbig * Redistribution and use in source and binary forms, with or without 1132236Shelbig * modification, are permitted provided that the following conditions 1232236Shelbig * are met: 1332236Shelbig * 1. Redistributions of source code must retain the above copyright 1432236Shelbig * notice, this list of conditions and the following disclaimer. 1532236Shelbig * 2. Redistributions in binary form must reproduce the above copyright 1632236Shelbig * notice, this list of conditions and the following disclaimer in the 1732236Shelbig * documentation and/or other materials provided with the distribution. 1832236Shelbig * 4. Neither the name of the University nor the names of its contributors 1932236Shelbig * may be used to endorse or promote products derived from this software 2032236Shelbig * without specific prior written permission. 2132236Shelbig * 2232236Shelbig * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 2332236Shelbig * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 2432236Shelbig * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 2532236Shelbig * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 2632236Shelbig * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 2732236Shelbig * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 2832236Shelbig * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 2932236Shelbig * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 3032236Shelbig * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 3132236Shelbig * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 3232236Shelbig * SUCH DAMAGE. 3332236Shelbig * 3432236Shelbig * @(#)kern_clock.c 8.5 (Berkeley) 1/21/94 3532236Shelbig */ 3632236Shelbig 3732236Shelbig#include <sys/cdefs.h> 3832236Shelbig__FBSDID("$FreeBSD: head/sys/kern/kern_clock.c 170468 2007-06-09 19:41:14Z attilio $"); 3932236Shelbig 4032236Shelbig#include "opt_kdb.h" 4132236Shelbig#include "opt_device_polling.h" 4232236Shelbig#include "opt_hwpmc_hooks.h" 4332236Shelbig#include "opt_ntp.h" 4432236Shelbig#include "opt_watchdog.h" 4532236Shelbig 4632236Shelbig#include <sys/param.h> 4732236Shelbig#include <sys/systm.h> 4832236Shelbig#include <sys/callout.h> 4932236Shelbig#include <sys/kdb.h> 5032236Shelbig#include <sys/kernel.h> 5132236Shelbig#include <sys/lock.h> 5232236Shelbig#include <sys/ktr.h> 5332236Shelbig#include <sys/mutex.h> 5432236Shelbig#include <sys/proc.h> 5532236Shelbig#include <sys/resource.h> 5632236Shelbig#include <sys/resourcevar.h> 5732236Shelbig#include <sys/sched.h> 5832236Shelbig#include <sys/signalvar.h> 5932236Shelbig#include <sys/smp.h> 6032236Shelbig#include <vm/vm.h> 6132236Shelbig#include <vm/pmap.h> 6232236Shelbig#include <vm/vm_map.h> 6332236Shelbig#include <sys/sysctl.h> 6432236Shelbig#include <sys/bus.h> 6532236Shelbig#include <sys/interrupt.h> 6653943Sache#include <sys/limits.h> 6732236Shelbig#include <sys/timetc.h> 6832236Shelbig 6932236Shelbig#ifdef GPROF 7032236Shelbig#include <sys/gmon.h> 7132236Shelbig#endif 7232236Shelbig 7332236Shelbig#ifdef HWPMC_HOOKS 7432236Shelbig#include <sys/pmckern.h> 7532236Shelbig#endif 7632236Shelbig 7732236Shelbig#ifdef DEVICE_POLLING 7853943Sacheextern void hardclock_device_poll(void); 7953943Sache#endif /* DEVICE_POLLING */ 8053943Sache 8153943Sachestatic void initclocks(void *dummy); 8253943SacheSYSINIT(clocks, SI_SUB_CLOCKS, SI_ORDER_FIRST, initclocks, NULL) 8353943Sache 8453943Sache/* Some of these don't belong here, but it's easiest to concentrate them. */ 8553943Sachelong cp_time[CPUSTATES]; 8653943Sache 8753943Sache/* Spin-lock protecting profiling statistics. */ 8853943Sachestatic struct mtx time_lock; 8953943Sache 9053943Sachestatic int 9153943Sachesysctl_kern_cp_time(SYSCTL_HANDLER_ARGS) 9253943Sache{ 9353943Sache int error; 9453943Sache#ifdef SCTL_MASK32 9553943Sache int i; 9653943Sache unsigned int cp_time32[CPUSTATES]; 9753943Sache 98 if (req->flags & SCTL_MASK32) { 99 if (!req->oldptr) 100 return SYSCTL_OUT(req, 0, sizeof(cp_time32)); 101 for (i = 0; i < CPUSTATES; i++) 102 cp_time32[i] = (unsigned int)cp_time[i]; 103 error = SYSCTL_OUT(req, cp_time32, sizeof(cp_time32)); 104 } else 105#endif 106 { 107 if (!req->oldptr) 108 return SYSCTL_OUT(req, 0, sizeof(cp_time)); 109 error = SYSCTL_OUT(req, cp_time, sizeof(cp_time)); 110 } 111 return error; 112} 113 114SYSCTL_PROC(_kern, OID_AUTO, cp_time, CTLTYPE_LONG|CTLFLAG_RD, 115 0,0, sysctl_kern_cp_time, "LU", "CPU time statistics"); 116 117#ifdef SW_WATCHDOG 118#include <sys/watchdog.h> 119 120static int watchdog_ticks; 121static int watchdog_enabled; 122static void watchdog_fire(void); 123static void watchdog_config(void *, u_int, int *); 124#endif /* SW_WATCHDOG */ 125 126/* 127 * Clock handling routines. 128 * 129 * This code is written to operate with two timers that run independently of 130 * each other. 131 * 132 * The main timer, running hz times per second, is used to trigger interval 133 * timers, timeouts and rescheduling as needed. 134 * 135 * The second timer handles kernel and user profiling, 136 * and does resource use estimation. If the second timer is programmable, 137 * it is randomized to avoid aliasing between the two clocks. For example, 138 * the randomization prevents an adversary from always giving up the cpu 139 * just before its quantum expires. Otherwise, it would never accumulate 140 * cpu ticks. The mean frequency of the second timer is stathz. 141 * 142 * If no second timer exists, stathz will be zero; in this case we drive 143 * profiling and statistics off the main clock. This WILL NOT be accurate; 144 * do not do it unless absolutely necessary. 145 * 146 * The statistics clock may (or may not) be run at a higher rate while 147 * profiling. This profile clock runs at profhz. We require that profhz 148 * be an integral multiple of stathz. 149 * 150 * If the statistics clock is running fast, it must be divided by the ratio 151 * profhz/stathz for statistics. (For profiling, every tick counts.) 152 * 153 * Time-of-day is maintained using a "timecounter", which may or may 154 * not be related to the hardware generating the above mentioned 155 * interrupts. 156 */ 157 158int stathz; 159int profhz; 160int profprocs; 161int ticks; 162int psratio; 163 164/* 165 * Initialize clock frequencies and start both clocks running. 166 */ 167/* ARGSUSED*/ 168static void 169initclocks(dummy) 170 void *dummy; 171{ 172 register int i; 173 174 /* 175 * Set divisors to 1 (normal case) and let the machine-specific 176 * code do its bit. 177 */ 178 mtx_init(&time_lock, "time lock", NULL, MTX_SPIN); 179 cpu_initclocks(); 180 181 /* 182 * Compute profhz/stathz, and fix profhz if needed. 183 */ 184 i = stathz ? stathz : hz; 185 if (profhz == 0) 186 profhz = i; 187 psratio = profhz / i; 188#ifdef SW_WATCHDOG 189 EVENTHANDLER_REGISTER(watchdog_list, watchdog_config, NULL, 0); 190#endif 191} 192 193/* 194 * Each time the real-time timer fires, this function is called on all CPUs. 195 * Note that hardclock() calls hardclock_cpu() for the boot CPU, so only 196 * the other CPUs in the system need to call this function. 197 */ 198void 199hardclock_cpu(int usermode) 200{ 201 struct pstats *pstats; 202 struct thread *td = curthread; 203 struct proc *p = td->td_proc; 204 int ast; 205 206 /* 207 * Run current process's virtual and profile time, as needed. 208 */ 209 pstats = p->p_stats; 210 ast = 0; 211 if (usermode && 212 timevalisset(&pstats->p_timer[ITIMER_VIRTUAL].it_value)) { 213 PROC_SLOCK(p); 214 if (itimerdecr(&pstats->p_timer[ITIMER_VIRTUAL], tick) == 0) { 215 p->p_sflag |= PS_ALRMPEND; 216 ast = 1; 217 } 218 PROC_SUNLOCK(p); 219 } 220 if (timevalisset(&pstats->p_timer[ITIMER_PROF].it_value)) { 221 PROC_SLOCK(p); 222 if (itimerdecr(&pstats->p_timer[ITIMER_PROF], tick) == 0) { 223 p->p_sflag |= PS_PROFPEND; 224 ast = 1; 225 } 226 PROC_SUNLOCK(p); 227 } 228 thread_lock(td); 229 sched_tick(); 230 if (ast) 231 td->td_flags |= TDF_ASTPENDING; 232 thread_unlock(td); 233 234#ifdef HWPMC_HOOKS 235 if (PMC_CPU_HAS_SAMPLES(PCPU_GET(cpuid))) 236 PMC_CALL_HOOK_UNLOCKED(curthread, PMC_FN_DO_SAMPLES, NULL); 237#endif 238} 239 240/* 241 * The real-time timer, interrupting hz times per second. 242 */ 243void 244hardclock(int usermode, uintfptr_t pc) 245{ 246 int need_softclock = 0; 247 248 hardclock_cpu(usermode); 249 250 tc_ticktock(); 251 /* 252 * If no separate statistics clock is available, run it from here. 253 * 254 * XXX: this only works for UP 255 */ 256 if (stathz == 0) { 257 profclock(usermode, pc); 258 statclock(usermode); 259 } 260 261#ifdef DEVICE_POLLING 262 hardclock_device_poll(); /* this is very short and quick */ 263#endif /* DEVICE_POLLING */ 264 265 /* 266 * Process callouts at a very low cpu priority, so we don't keep the 267 * relatively high clock interrupt priority any longer than necessary. 268 */ 269 mtx_lock_spin_flags(&callout_lock, MTX_QUIET); 270 ticks++; 271 if (!TAILQ_EMPTY(&callwheel[ticks & callwheelmask])) { 272 need_softclock = 1; 273 } else if (softticks + 1 == ticks) 274 ++softticks; 275 mtx_unlock_spin_flags(&callout_lock, MTX_QUIET); 276 277 /* 278 * swi_sched acquires the thread lock, so we don't want to call it 279 * with callout_lock held; incorrect locking order. 280 */ 281 if (need_softclock) 282 swi_sched(softclock_ih, 0); 283 284#ifdef SW_WATCHDOG 285 if (watchdog_enabled > 0 && --watchdog_ticks <= 0) 286 watchdog_fire(); 287#endif /* SW_WATCHDOG */ 288} 289 290/* 291 * Compute number of ticks in the specified amount of time. 292 */ 293int 294tvtohz(tv) 295 struct timeval *tv; 296{ 297 register unsigned long ticks; 298 register long sec, usec; 299 300 /* 301 * If the number of usecs in the whole seconds part of the time 302 * difference fits in a long, then the total number of usecs will 303 * fit in an unsigned long. Compute the total and convert it to 304 * ticks, rounding up and adding 1 to allow for the current tick 305 * to expire. Rounding also depends on unsigned long arithmetic 306 * to avoid overflow. 307 * 308 * Otherwise, if the number of ticks in the whole seconds part of 309 * the time difference fits in a long, then convert the parts to 310 * ticks separately and add, using similar rounding methods and 311 * overflow avoidance. This method would work in the previous 312 * case but it is slightly slower and assumes that hz is integral. 313 * 314 * Otherwise, round the time difference down to the maximum 315 * representable value. 316 * 317 * If ints have 32 bits, then the maximum value for any timeout in 318 * 10ms ticks is 248 days. 319 */ 320 sec = tv->tv_sec; 321 usec = tv->tv_usec; 322 if (usec < 0) { 323 sec--; 324 usec += 1000000; 325 } 326 if (sec < 0) { 327#ifdef DIAGNOSTIC 328 if (usec > 0) { 329 sec++; 330 usec -= 1000000; 331 } 332 printf("tvotohz: negative time difference %ld sec %ld usec\n", 333 sec, usec); 334#endif 335 ticks = 1; 336 } else if (sec <= LONG_MAX / 1000000) 337 ticks = (sec * 1000000 + (unsigned long)usec + (tick - 1)) 338 / tick + 1; 339 else if (sec <= LONG_MAX / hz) 340 ticks = sec * hz 341 + ((unsigned long)usec + (tick - 1)) / tick + 1; 342 else 343 ticks = LONG_MAX; 344 if (ticks > INT_MAX) 345 ticks = INT_MAX; 346 return ((int)ticks); 347} 348 349/* 350 * Start profiling on a process. 351 * 352 * Kernel profiling passes proc0 which never exits and hence 353 * keeps the profile clock running constantly. 354 */ 355void 356startprofclock(p) 357 register struct proc *p; 358{ 359 360 PROC_LOCK_ASSERT(p, MA_OWNED); 361 if (p->p_flag & P_STOPPROF) 362 return; 363 if ((p->p_flag & P_PROFIL) == 0) { 364 p->p_flag |= P_PROFIL; 365 mtx_lock_spin(&time_lock); 366 if (++profprocs == 1) 367 cpu_startprofclock(); 368 mtx_unlock_spin(&time_lock); 369 } 370} 371 372/* 373 * Stop profiling on a process. 374 */ 375void 376stopprofclock(p) 377 register struct proc *p; 378{ 379 380 PROC_LOCK_ASSERT(p, MA_OWNED); 381 if (p->p_flag & P_PROFIL) { 382 if (p->p_profthreads != 0) { 383 p->p_flag |= P_STOPPROF; 384 while (p->p_profthreads != 0) 385 msleep(&p->p_profthreads, &p->p_mtx, PPAUSE, 386 "stopprof", 0); 387 p->p_flag &= ~P_STOPPROF; 388 } 389 if ((p->p_flag & P_PROFIL) == 0) 390 return; 391 p->p_flag &= ~P_PROFIL; 392 mtx_lock_spin(&time_lock); 393 if (--profprocs == 0) 394 cpu_stopprofclock(); 395 mtx_unlock_spin(&time_lock); 396 } 397} 398 399/* 400 * Statistics clock. Updates rusage information and calls the scheduler 401 * to adjust priorities of the active thread. 402 * 403 * This should be called by all active processors. 404 */ 405void 406statclock(int usermode) 407{ 408 struct rusage *ru; 409 struct vmspace *vm; 410 struct thread *td; 411 struct proc *p; 412 long rss; 413 414 td = curthread; 415 p = td->td_proc; 416 417 thread_lock_flags(td, MTX_QUIET); 418 if (usermode) { 419 /* 420 * Charge the time as appropriate. 421 */ 422#ifdef KSE 423 if (p->p_flag & P_SA) 424 thread_statclock(1); 425#endif 426 td->td_uticks++; 427 if (p->p_nice > NZERO) 428 atomic_add_long(&cp_time[CP_NICE], 1); 429 else 430 atomic_add_long(&cp_time[CP_USER], 1); 431 } else { 432 /* 433 * Came from kernel mode, so we were: 434 * - handling an interrupt, 435 * - doing syscall or trap work on behalf of the current 436 * user process, or 437 * - spinning in the idle loop. 438 * Whichever it is, charge the time as appropriate. 439 * Note that we charge interrupts to the current process, 440 * regardless of whether they are ``for'' that process, 441 * so that we know how much of its real time was spent 442 * in ``non-process'' (i.e., interrupt) work. 443 */ 444 if ((td->td_pflags & TDP_ITHREAD) || 445 td->td_intr_nesting_level >= 2) { 446 td->td_iticks++; 447 atomic_add_long(&cp_time[CP_INTR], 1); 448 } else { 449#ifdef KSE 450 if (p->p_flag & P_SA) 451 thread_statclock(0); 452#endif 453 td->td_pticks++; 454 td->td_sticks++; 455 if (!TD_IS_IDLETHREAD(td)) 456 atomic_add_long(&cp_time[CP_SYS], 1); 457 else 458 atomic_add_long(&cp_time[CP_IDLE], 1); 459 } 460 } 461 462 /* Update resource usage integrals and maximums. */ 463 MPASS(p->p_vmspace != NULL); 464 vm = p->p_vmspace; 465 ru = &td->td_ru; 466 ru->ru_ixrss += pgtok(vm->vm_tsize); 467 ru->ru_idrss += pgtok(vm->vm_dsize); 468 ru->ru_isrss += pgtok(vm->vm_ssize); 469 rss = pgtok(vmspace_resident_count(vm)); 470 if (ru->ru_maxrss < rss) 471 ru->ru_maxrss = rss; 472 CTR4(KTR_SCHED, "statclock: %p(%s) prio %d stathz %d", 473 td, td->td_proc->p_comm, td->td_priority, (stathz)?stathz:hz); 474 sched_clock(td); 475 thread_unlock(td); 476} 477 478void 479profclock(int usermode, uintfptr_t pc) 480{ 481 struct thread *td; 482#ifdef GPROF 483 struct gmonparam *g; 484 uintfptr_t i; 485#endif 486 487 td = curthread; 488 if (usermode) { 489 /* 490 * Came from user mode; CPU was in user state. 491 * If this process is being profiled, record the tick. 492 * if there is no related user location yet, don't 493 * bother trying to count it. 494 */ 495 if (td->td_proc->p_flag & P_PROFIL) 496 addupc_intr(td, pc, 1); 497 } 498#ifdef GPROF 499 else { 500 /* 501 * Kernel statistics are just like addupc_intr, only easier. 502 */ 503 g = &_gmonparam; 504 if (g->state == GMON_PROF_ON && pc >= g->lowpc) { 505 i = PC_TO_I(g, pc); 506 if (i < g->textsize) { 507 KCOUNT(g, i)++; 508 } 509 } 510 } 511#endif 512} 513 514/* 515 * Return information about system clocks. 516 */ 517static int 518sysctl_kern_clockrate(SYSCTL_HANDLER_ARGS) 519{ 520 struct clockinfo clkinfo; 521 /* 522 * Construct clockinfo structure. 523 */ 524 bzero(&clkinfo, sizeof(clkinfo)); 525 clkinfo.hz = hz; 526 clkinfo.tick = tick; 527 clkinfo.profhz = profhz; 528 clkinfo.stathz = stathz ? stathz : hz; 529 return (sysctl_handle_opaque(oidp, &clkinfo, sizeof clkinfo, req)); 530} 531 532SYSCTL_PROC(_kern, KERN_CLOCKRATE, clockrate, CTLTYPE_STRUCT|CTLFLAG_RD, 533 0, 0, sysctl_kern_clockrate, "S,clockinfo", 534 "Rate and period of various kernel clocks"); 535 536#ifdef SW_WATCHDOG 537 538static void 539watchdog_config(void *unused __unused, u_int cmd, int *error) 540{ 541 u_int u; 542 543 u = cmd & WD_INTERVAL; 544 if (u >= WD_TO_1SEC) { 545 watchdog_ticks = (1 << (u - WD_TO_1SEC)) * hz; 546 watchdog_enabled = 1; 547 *error = 0; 548 } else { 549 watchdog_enabled = 0; 550 } 551} 552 553/* 554 * Handle a watchdog timeout by dumping interrupt information and 555 * then either dropping to DDB or panicking. 556 */ 557static void 558watchdog_fire(void) 559{ 560 int nintr; 561 u_int64_t inttotal; 562 u_long *curintr; 563 char *curname; 564 565 curintr = intrcnt; 566 curname = intrnames; 567 inttotal = 0; 568 nintr = eintrcnt - intrcnt; 569 570 printf("interrupt total\n"); 571 while (--nintr >= 0) { 572 if (*curintr) 573 printf("%-12s %20lu\n", curname, *curintr); 574 curname += strlen(curname) + 1; 575 inttotal += *curintr++; 576 } 577 printf("Total %20ju\n", (uintmax_t)inttotal); 578 579#if defined(KDB) && !defined(KDB_UNATTENDED) 580 kdb_backtrace(); 581 kdb_enter("watchdog timeout"); 582#else 583 panic("watchdog timeout"); 584#endif 585} 586 587#endif /* SW_WATCHDOG */ 588