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("$FreeBSD$"); 39 40 #include "opt_kdb.h" 41 #include "opt_device_polling.h" 42 #include "opt_hwpmc_hooks.h" 43 #include "opt_ntp.h" 44 #include "opt_watchdog.h" 45 46 #include <sys/param.h> 47 #include <sys/systm.h> 48 #include <sys/callout.h> 49 #include <sys/kdb.h> 50 #include <sys/kernel.h> 51 #include <sys/lock.h> 52 #include <sys/ktr.h> 53 #include <sys/mutex.h> 54 #include <sys/proc.h> 55 #include <sys/resource.h> 56 #include <sys/resourcevar.h> 57 #include <sys/sched.h> 58 #include <sys/signalvar.h> 59 #include <sys/smp.h> 60 #include <vm/vm.h> 61 #include <vm/pmap.h> 62 #include <vm/vm_map.h> 63 #include <sys/sysctl.h> 64 #include <sys/bus.h> 65 #include <sys/interrupt.h> 66 #include <sys/limits.h> 67 #include <sys/timetc.h> 68 69 #ifdef GPROF 70 #include <sys/gmon.h> 71 #endif 72 73 #ifdef HWPMC_HOOKS 74 #include <sys/pmckern.h> 75 #endif 76 77 #ifdef DEVICE_POLLING 78 extern void hardclock_device_poll(void); 79 #endif /* DEVICE_POLLING */ 80 81 static void initclocks(void *dummy); 82 SYSINIT(clocks, SI_SUB_CLOCKS, SI_ORDER_FIRST, initclocks, NULL) 83 84 /* Spin-lock protecting profiling statistics. */ 85 static struct mtx time_lock; 86 87 static int 88 sysctl_kern_cp_time(SYSCTL_HANDLER_ARGS) 89 { 90 int error; 91 long cp_time[CPUSTATES]; 92 #ifdef SCTL_MASK32 93 int i; 94 unsigned int cp_time32[CPUSTATES]; 95 #endif 96 97 read_cpu_time(cp_time); 98 #ifdef SCTL_MASK32 99 if (req->flags & SCTL_MASK32) { 100 if (!req->oldptr) 101 return SYSCTL_OUT(req, 0, sizeof(cp_time32)); 102 for (i = 0; i < CPUSTATES; i++) 103 cp_time32[i] = (unsigned int)cp_time[i]; 104 error = SYSCTL_OUT(req, cp_time32, sizeof(cp_time32)); 105 } else 106 #endif 107 { 108 if (!req->oldptr) 109 return SYSCTL_OUT(req, 0, sizeof(cp_time)); 110 error = SYSCTL_OUT(req, cp_time, sizeof(cp_time)); 111 } 112 return error; 113 } 114 115 SYSCTL_PROC(_kern, OID_AUTO, cp_time, CTLTYPE_LONG|CTLFLAG_RD, 116 0,0, sysctl_kern_cp_time, "LU", "CPU time statistics"); 117 118 static long empty[CPUSTATES]; 119 120 static int 121 sysctl_kern_cp_times(SYSCTL_HANDLER_ARGS) 122 { 123 struct pcpu *pcpu; 124 int error; 125 int c; 126 long *cp_time; 127 #ifdef SCTL_MASK32 128 unsigned int cp_time32[CPUSTATES]; 129 int i; 130 #endif 131 132 if (!req->oldptr) { 133 #ifdef SCTL_MASK32 134 if (req->flags & SCTL_MASK32) 135 return SYSCTL_OUT(req, 0, sizeof(cp_time32) * (mp_maxid + 1)); 136 else 137 #endif 138 return SYSCTL_OUT(req, 0, sizeof(long) * CPUSTATES * (mp_maxid + 1)); 139 } 140 for (error = 0, c = 0; error == 0 && c <= mp_maxid; c++) { 141 if (!CPU_ABSENT(c)) { 142 pcpu = pcpu_find(c); 143 cp_time = pcpu->pc_cp_time; 144 } else { 145 cp_time = empty; 146 } 147 #ifdef SCTL_MASK32 148 if (req->flags & SCTL_MASK32) { 149 for (i = 0; i < CPUSTATES; i++) 150 cp_time32[i] = (unsigned int)cp_time[i]; 151 error = SYSCTL_OUT(req, cp_time32, sizeof(cp_time32)); 152 } else 153 #endif 154 error = SYSCTL_OUT(req, cp_time, sizeof(long) * CPUSTATES); 155 } 156 return error; 157 } 158 159 SYSCTL_PROC(_kern, OID_AUTO, cp_times, CTLTYPE_LONG|CTLFLAG_RD, 160 0,0, sysctl_kern_cp_times, "LU", "per-CPU time statistics"); 161 162 void 163 read_cpu_time(long *cp_time) 164 { 165 struct pcpu *pc; 166 int i, j; 167 168 /* Sum up global cp_time[]. */ 169 bzero(cp_time, sizeof(long) * CPUSTATES); 170 for (i = 0; i <= mp_maxid; i++) { 171 if (CPU_ABSENT(i)) 172 continue; 173 pc = pcpu_find(i); 174 for (j = 0; j < CPUSTATES; j++) 175 cp_time[j] += pc->pc_cp_time[j]; 176 } 177 } 178 179 #ifdef SW_WATCHDOG 180 #include <sys/watchdog.h> 181 182 static int watchdog_ticks; 183 static int watchdog_enabled; 184 static void watchdog_fire(void); 185 static void watchdog_config(void *, u_int, int *); 186 #endif /* SW_WATCHDOG */ 187 188 /* 189 * Clock handling routines. 190 * 191 * This code is written to operate with two timers that run independently of 192 * each other. 193 * 194 * The main timer, running hz times per second, is used to trigger interval 195 * timers, timeouts and rescheduling as needed. 196 * 197 * The second timer handles kernel and user profiling, 198 * and does resource use estimation. If the second timer is programmable, 199 * it is randomized to avoid aliasing between the two clocks. For example, 200 * the randomization prevents an adversary from always giving up the cpu 201 * just before its quantum expires. Otherwise, it would never accumulate 202 * cpu ticks. The mean frequency of the second timer is stathz. 203 * 204 * If no second timer exists, stathz will be zero; in this case we drive 205 * profiling and statistics off the main clock. This WILL NOT be accurate; 206 * do not do it unless absolutely necessary. 207 * 208 * The statistics clock may (or may not) be run at a higher rate while 209 * profiling. This profile clock runs at profhz. We require that profhz 210 * be an integral multiple of stathz. 211 * 212 * If the statistics clock is running fast, it must be divided by the ratio 213 * profhz/stathz for statistics. (For profiling, every tick counts.) 214 * 215 * Time-of-day is maintained using a "timecounter", which may or may 216 * not be related to the hardware generating the above mentioned 217 * interrupts. 218 */ 219 220 int stathz; 221 int profhz; 222 int profprocs; 223 int ticks; 224 int psratio; 225 226 /* 227 * Initialize clock frequencies and start both clocks running. 228 */ 229 /* ARGSUSED*/ 230 static void 231 initclocks(dummy) 232 void *dummy; 233 { 234 register int i; 235 236 /* 237 * Set divisors to 1 (normal case) and let the machine-specific 238 * code do its bit. 239 */ 240 mtx_init(&time_lock, "time lock", NULL, MTX_SPIN); 241 cpu_initclocks(); 242 243 /* 244 * Compute profhz/stathz, and fix profhz if needed. 245 */ 246 i = stathz ? stathz : hz; 247 if (profhz == 0) 248 profhz = i; 249 psratio = profhz / i; 250 #ifdef SW_WATCHDOG 251 EVENTHANDLER_REGISTER(watchdog_list, watchdog_config, NULL, 0); 252 #endif 253 } 254 255 /* 256 * Each time the real-time timer fires, this function is called on all CPUs. 257 * Note that hardclock() calls hardclock_cpu() for the boot CPU, so only 258 * the other CPUs in the system need to call this function. 259 */ 260 void 261 hardclock_cpu(int usermode) 262 { 263 struct pstats *pstats; 264 struct thread *td = curthread; 265 struct proc *p = td->td_proc; 266 int flags; 267 268 /* 269 * Run current process's virtual and profile time, as needed. 270 */ 271 pstats = p->p_stats; 272 flags = 0; 273 if (usermode && 274 timevalisset(&pstats->p_timer[ITIMER_VIRTUAL].it_value)) { 275 PROC_SLOCK(p); 276 if (itimerdecr(&pstats->p_timer[ITIMER_VIRTUAL], tick) == 0) 277 flags |= TDF_ALRMPEND | TDF_ASTPENDING; 278 PROC_SUNLOCK(p); 279 } 280 if (timevalisset(&pstats->p_timer[ITIMER_PROF].it_value)) { 281 PROC_SLOCK(p); 282 if (itimerdecr(&pstats->p_timer[ITIMER_PROF], tick) == 0) 283 flags |= TDF_PROFPEND | TDF_ASTPENDING; 284 PROC_SUNLOCK(p); 285 } 286 thread_lock(td); 287 sched_tick(); 288 td->td_flags |= flags; 289 thread_unlock(td); 290 291 #ifdef HWPMC_HOOKS 292 if (PMC_CPU_HAS_SAMPLES(PCPU_GET(cpuid))) 293 PMC_CALL_HOOK_UNLOCKED(curthread, PMC_FN_DO_SAMPLES, NULL); 294 #endif 295 } 296 297 /* 298 * The real-time timer, interrupting hz times per second. 299 */ 300 void 301 hardclock(int usermode, uintfptr_t pc) 302 { 303 int need_softclock = 0; 304 305 hardclock_cpu(usermode); 306 307 tc_ticktock(); 308 /* 309 * If no separate statistics clock is available, run it from here. 310 * 311 * XXX: this only works for UP 312 */ 313 if (stathz == 0) { 314 profclock(usermode, pc); 315 statclock(usermode); 316 } 317 318 #ifdef DEVICE_POLLING 319 hardclock_device_poll(); /* this is very short and quick */ 320 #endif /* DEVICE_POLLING */ 321 322 /* 323 * Process callouts at a very low cpu priority, so we don't keep the 324 * relatively high clock interrupt priority any longer than necessary. 325 */ 326 mtx_lock_spin_flags(&callout_lock, MTX_QUIET); 327 ticks++; 328 if (!TAILQ_EMPTY(&callwheel[ticks & callwheelmask])) { 329 need_softclock = 1; 330 } else if (softticks + 1 == ticks) 331 ++softticks; 332 mtx_unlock_spin_flags(&callout_lock, MTX_QUIET); 333 334 /* 335 * swi_sched acquires the thread lock, so we don't want to call it 336 * with callout_lock held; incorrect locking order. 337 */ 338 if (need_softclock) 339 swi_sched(softclock_ih, 0); 340 341 #ifdef SW_WATCHDOG 342 if (watchdog_enabled > 0 && --watchdog_ticks <= 0) 343 watchdog_fire(); 344 #endif /* SW_WATCHDOG */ 345 } 346 347 /* 348 * Compute number of ticks in the specified amount of time. 349 */ 350 int 351 tvtohz(tv) 352 struct timeval *tv; 353 { 354 register unsigned long ticks; 355 register long sec, usec; 356 357 /* 358 * If the number of usecs in the whole seconds part of the time 359 * difference fits in a long, then the total number of usecs will 360 * fit in an unsigned long. Compute the total and convert it to 361 * ticks, rounding up and adding 1 to allow for the current tick 362 * to expire. Rounding also depends on unsigned long arithmetic 363 * to avoid overflow. 364 * 365 * Otherwise, if the number of ticks in the whole seconds part of 366 * the time difference fits in a long, then convert the parts to 367 * ticks separately and add, using similar rounding methods and 368 * overflow avoidance. This method would work in the previous 369 * case but it is slightly slower and assumes that hz is integral. 370 * 371 * Otherwise, round the time difference down to the maximum 372 * representable value. 373 * 374 * If ints have 32 bits, then the maximum value for any timeout in 375 * 10ms ticks is 248 days. 376 */ 377 sec = tv->tv_sec; 378 usec = tv->tv_usec; 379 if (usec < 0) { 380 sec--; 381 usec += 1000000; 382 } 383 if (sec < 0) { 384 #ifdef DIAGNOSTIC 385 if (usec > 0) { 386 sec++; 387 usec -= 1000000; 388 } 389 printf("tvotohz: negative time difference %ld sec %ld usec\n", 390 sec, usec); 391 #endif 392 ticks = 1; 393 } else if (sec <= LONG_MAX / 1000000) 394 ticks = (sec * 1000000 + (unsigned long)usec + (tick - 1)) 395 / tick + 1; 396 else if (sec <= LONG_MAX / hz) 397 ticks = sec * hz 398 + ((unsigned long)usec + (tick - 1)) / tick + 1; 399 else 400 ticks = LONG_MAX; 401 if (ticks > INT_MAX) 402 ticks = INT_MAX; 403 return ((int)ticks); 404 } 405 406 /* 407 * Start profiling on a process. 408 * 409 * Kernel profiling passes proc0 which never exits and hence 410 * keeps the profile clock running constantly. 411 */ 412 void 413 startprofclock(p) 414 register struct proc *p; 415 { 416 417 PROC_LOCK_ASSERT(p, MA_OWNED); 418 if (p->p_flag & P_STOPPROF) 419 return; 420 if ((p->p_flag & P_PROFIL) == 0) { 421 p->p_flag |= P_PROFIL; 422 mtx_lock_spin(&time_lock); 423 if (++profprocs == 1) 424 cpu_startprofclock(); 425 mtx_unlock_spin(&time_lock); 426 } 427 } 428 429 /* 430 * Stop profiling on a process. 431 */ 432 void 433 stopprofclock(p) 434 register struct proc *p; 435 { 436 437 PROC_LOCK_ASSERT(p, MA_OWNED); 438 if (p->p_flag & P_PROFIL) { 439 if (p->p_profthreads != 0) { 440 p->p_flag |= P_STOPPROF; 441 while (p->p_profthreads != 0) 442 msleep(&p->p_profthreads, &p->p_mtx, PPAUSE, 443 "stopprof", 0); 444 p->p_flag &= ~P_STOPPROF; 445 } 446 if ((p->p_flag & P_PROFIL) == 0) 447 return; 448 p->p_flag &= ~P_PROFIL; 449 mtx_lock_spin(&time_lock); 450 if (--profprocs == 0) 451 cpu_stopprofclock(); 452 mtx_unlock_spin(&time_lock); 453 } 454 } 455 456 /* 457 * Statistics clock. Updates rusage information and calls the scheduler 458 * to adjust priorities of the active thread. 459 * 460 * This should be called by all active processors. 461 */ 462 void 463 statclock(int usermode) 464 { 465 struct rusage *ru; 466 struct vmspace *vm; 467 struct thread *td; 468 struct proc *p; 469 long rss; 470 long *cp_time; 471 472 td = curthread; 473 p = td->td_proc; 474 475 cp_time = (long *)PCPU_PTR(cp_time); 476 if (usermode) { 477 /* 478 * Charge the time as appropriate. 479 */ 480 #ifdef KSE 481 if (p->p_flag & P_SA) 482 thread_statclock(1); 483 #endif 484 td->td_uticks++; 485 if (p->p_nice > NZERO) 486 cp_time[CP_NICE]++; 487 else 488 cp_time[CP_USER]++; 489 } else { 490 /* 491 * Came from kernel mode, so we were: 492 * - handling an interrupt, 493 * - doing syscall or trap work on behalf of the current 494 * user process, or 495 * - spinning in the idle loop. 496 * Whichever it is, charge the time as appropriate. 497 * Note that we charge interrupts to the current process, 498 * regardless of whether they are ``for'' that process, 499 * so that we know how much of its real time was spent 500 * in ``non-process'' (i.e., interrupt) work. 501 */ 502 if ((td->td_pflags & TDP_ITHREAD) || 503 td->td_intr_nesting_level >= 2) { 504 td->td_iticks++; 505 cp_time[CP_INTR]++; 506 } else { 507 #ifdef KSE 508 if (p->p_flag & P_SA) 509 thread_statclock(0); 510 #endif 511 td->td_pticks++; 512 td->td_sticks++; 513 if (!TD_IS_IDLETHREAD(td)) 514 cp_time[CP_SYS]++; 515 else 516 cp_time[CP_IDLE]++; 517 } 518 } 519 520 /* Update resource usage integrals and maximums. */ 521 MPASS(p->p_vmspace != NULL); 522 vm = p->p_vmspace; 523 ru = &td->td_ru; 524 ru->ru_ixrss += pgtok(vm->vm_tsize); 525 ru->ru_idrss += pgtok(vm->vm_dsize); 526 ru->ru_isrss += pgtok(vm->vm_ssize); 527 rss = pgtok(vmspace_resident_count(vm)); 528 if (ru->ru_maxrss < rss) 529 ru->ru_maxrss = rss; 530 CTR4(KTR_SCHED, "statclock: %p(%s) prio %d stathz %d", 531 td, td->td_name, td->td_priority, (stathz)?stathz:hz); 532 thread_lock_flags(td, MTX_QUIET); 533 sched_clock(td); 534 thread_unlock(td); 535 } 536 537 void 538 profclock(int usermode, uintfptr_t pc) 539 { 540 struct thread *td; 541 #ifdef GPROF 542 struct gmonparam *g; 543 uintfptr_t i; 544 #endif 545 546 td = curthread; 547 if (usermode) { 548 /* 549 * Came from user mode; CPU was in user state. 550 * If this process is being profiled, record the tick. 551 * if there is no related user location yet, don't 552 * bother trying to count it. 553 */ 554 if (td->td_proc->p_flag & P_PROFIL) 555 addupc_intr(td, pc, 1); 556 } 557 #ifdef GPROF 558 else { 559 /* 560 * Kernel statistics are just like addupc_intr, only easier. 561 */ 562 g = &_gmonparam; 563 if (g->state == GMON_PROF_ON && pc >= g->lowpc) { 564 i = PC_TO_I(g, pc); 565 if (i < g->textsize) { 566 KCOUNT(g, i)++; 567 } 568 } 569 } 570 #endif 571 } 572 573 /* 574 * Return information about system clocks. 575 */ 576 static int 577 sysctl_kern_clockrate(SYSCTL_HANDLER_ARGS) 578 { 579 struct clockinfo clkinfo; 580 /* 581 * Construct clockinfo structure. 582 */ 583 bzero(&clkinfo, sizeof(clkinfo)); 584 clkinfo.hz = hz; 585 clkinfo.tick = tick; 586 clkinfo.profhz = profhz; 587 clkinfo.stathz = stathz ? stathz : hz; 588 return (sysctl_handle_opaque(oidp, &clkinfo, sizeof clkinfo, req)); 589 } 590 591 SYSCTL_PROC(_kern, KERN_CLOCKRATE, clockrate, CTLTYPE_STRUCT|CTLFLAG_RD, 592 0, 0, sysctl_kern_clockrate, "S,clockinfo", 593 "Rate and period of various kernel clocks"); 594 595 #ifdef SW_WATCHDOG 596 597 static void 598 watchdog_config(void *unused __unused, u_int cmd, int *error) 599 { 600 u_int u; 601 602 u = cmd & WD_INTERVAL; 603 if (u >= WD_TO_1SEC) { 604 watchdog_ticks = (1 << (u - WD_TO_1SEC)) * hz; 605 watchdog_enabled = 1; 606 *error = 0; 607 } else { 608 watchdog_enabled = 0; 609 } 610 } 611 612 /* 613 * Handle a watchdog timeout by dumping interrupt information and 614 * then either dropping to DDB or panicking. 615 */ 616 static void 617 watchdog_fire(void) 618 { 619 int nintr; 620 u_int64_t inttotal; 621 u_long *curintr; 622 char *curname; 623 624 curintr = intrcnt; 625 curname = intrnames; 626 inttotal = 0; 627 nintr = eintrcnt - intrcnt; 628 629 printf("interrupt total\n"); 630 while (--nintr >= 0) { 631 if (*curintr) 632 printf("%-12s %20lu\n", curname, *curintr); 633 curname += strlen(curname) + 1; 634 inttotal += *curintr++; 635 } 636 printf("Total %20ju\n", (uintmax_t)inttotal); 637 638 #if defined(KDB) && !defined(KDB_UNATTENDED) 639 kdb_backtrace(); 640 kdb_enter(KDB_WHY_WATCHDOG, "watchdog timeout"); 641 #else 642 panic("watchdog timeout"); 643 #endif 644 } 645 646 #endif /* SW_WATCHDOG */ 647