1 /* $NetBSD: kern_clock.c,v 1.151 2023/09/02 17:44:59 riastradh Exp $ */ 2 3 /*- 4 * Copyright (c) 2000, 2004, 2006, 2007, 2008 The NetBSD Foundation, Inc. 5 * All rights reserved. 6 * 7 * This code is derived from software contributed to The NetBSD Foundation 8 * by Jason R. Thorpe of the Numerical Aerospace Simulation Facility, 9 * NASA Ames Research Center. 10 * This code is derived from software contributed to The NetBSD Foundation 11 * by Charles M. Hannum. 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 * 22 * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS 23 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED 24 * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR 25 * PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS 26 * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR 27 * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF 28 * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS 29 * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN 30 * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) 31 * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE 32 * POSSIBILITY OF SUCH DAMAGE. 33 */ 34 35 /*- 36 * Copyright (c) 1982, 1986, 1991, 1993 37 * The Regents of the University of California. All rights reserved. 38 * (c) UNIX System Laboratories, Inc. 39 * All or some portions of this file are derived from material licensed 40 * to the University of California by American Telephone and Telegraph 41 * Co. or Unix System Laboratories, Inc. and are reproduced herein with 42 * the permission of UNIX System Laboratories, Inc. 43 * 44 * Redistribution and use in source and binary forms, with or without 45 * modification, are permitted provided that the following conditions 46 * are met: 47 * 1. Redistributions of source code must retain the above copyright 48 * notice, this list of conditions and the following disclaimer. 49 * 2. Redistributions in binary form must reproduce the above copyright 50 * notice, this list of conditions and the following disclaimer in the 51 * documentation and/or other materials provided with the distribution. 52 * 3. Neither the name of the University nor the names of its contributors 53 * may be used to endorse or promote products derived from this software 54 * without specific prior written permission. 55 * 56 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 57 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 58 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 59 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 60 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 61 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 62 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 63 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 64 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 65 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 66 * SUCH DAMAGE. 67 * 68 * @(#)kern_clock.c 8.5 (Berkeley) 1/21/94 69 */ 70 71 #include <sys/cdefs.h> 72 __KERNEL_RCSID(0, "$NetBSD: kern_clock.c,v 1.151 2023/09/02 17:44:59 riastradh Exp $"); 73 74 #ifdef _KERNEL_OPT 75 #include "opt_dtrace.h" 76 #include "opt_gprof.h" 77 #include "opt_multiprocessor.h" 78 #endif 79 80 #include <sys/param.h> 81 #include <sys/systm.h> 82 #include <sys/callout.h> 83 #include <sys/kernel.h> 84 #include <sys/proc.h> 85 #include <sys/resourcevar.h> 86 #include <sys/signalvar.h> 87 #include <sys/sysctl.h> 88 #include <sys/timex.h> 89 #include <sys/sched.h> 90 #include <sys/time.h> 91 #include <sys/timetc.h> 92 #include <sys/cpu.h> 93 #include <sys/atomic.h> 94 #include <sys/rndsource.h> 95 #include <sys/heartbeat.h> 96 97 #ifdef GPROF 98 #include <sys/gmon.h> 99 #endif 100 101 #ifdef KDTRACE_HOOKS 102 #include <sys/dtrace_bsd.h> 103 #include <sys/cpu.h> 104 105 cyclic_clock_func_t cyclic_clock_func[MAXCPUS]; 106 #endif 107 108 static int sysctl_kern_clockrate(SYSCTLFN_PROTO); 109 110 /* 111 * Clock handling routines. 112 * 113 * This code is written to operate with two timers that run independently of 114 * each other. The main clock, running hz times per second, is used to keep 115 * track of real time. The second timer handles kernel and user profiling, 116 * and does resource use estimation. If the second timer is programmable, 117 * it is randomized to avoid aliasing between the two clocks. For example, 118 * the randomization prevents an adversary from always giving up the CPU 119 * just before its quantum expires. Otherwise, it would never accumulate 120 * CPU ticks. The mean frequency of the second timer is stathz. 121 * 122 * If no second timer exists, stathz will be zero; in this case we drive 123 * profiling and statistics off the main clock. This WILL NOT be accurate; 124 * do not do it unless absolutely necessary. 125 * 126 * The statistics clock may (or may not) be run at a higher rate while 127 * profiling. This profile clock runs at profhz. We require that profhz 128 * be an integral multiple of stathz. 129 * 130 * If the statistics clock is running fast, it must be divided by the ratio 131 * profhz/stathz for statistics. (For profiling, every tick counts.) 132 */ 133 134 int stathz; 135 int profhz; 136 int profsrc; 137 int schedhz; 138 int profprocs; 139 static int hardclock_ticks; 140 static int hardscheddiv; /* hard => sched divider (used if schedhz == 0) */ 141 static int psdiv; /* prof => stat divider */ 142 int psratio; /* ratio: prof / stat */ 143 144 struct clockrnd { 145 struct krndsource source; 146 unsigned needed; 147 }; 148 149 static struct clockrnd hardclockrnd __aligned(COHERENCY_UNIT); 150 static struct clockrnd statclockrnd __aligned(COHERENCY_UNIT); 151 152 static void 153 clockrnd_get(size_t needed, void *cookie) 154 { 155 struct clockrnd *C = cookie; 156 157 /* Start sampling. */ 158 atomic_store_relaxed(&C->needed, 2*NBBY*needed); 159 } 160 161 static void 162 clockrnd_sample(struct clockrnd *C) 163 { 164 struct cpu_info *ci = curcpu(); 165 166 /* If there's nothing needed right now, stop here. */ 167 if (__predict_true(atomic_load_relaxed(&C->needed) == 0)) 168 return; 169 170 /* 171 * If we're not the primary core of a package, we're probably 172 * driven by the same clock as the primary core, so don't 173 * bother. 174 */ 175 if (ci != ci->ci_package1st) 176 return; 177 178 /* Take a sample and enter it into the pool. */ 179 rnd_add_uint32(&C->source, 0); 180 181 /* 182 * On the primary CPU, count down. Using an atomic decrement 183 * here isn't really necessary -- on every platform we care 184 * about, stores to unsigned int are atomic, and the only other 185 * memory operation that could happen here is for another CPU 186 * to store a higher value for needed. But using an atomic 187 * decrement avoids giving the impression of data races, and is 188 * unlikely to hurt because only one CPU will ever be writing 189 * to the location. 190 */ 191 if (CPU_IS_PRIMARY(curcpu())) { 192 unsigned needed __diagused; 193 194 needed = atomic_dec_uint_nv(&C->needed); 195 KASSERT(needed != UINT_MAX); 196 } 197 } 198 199 static u_int get_intr_timecount(struct timecounter *); 200 201 static struct timecounter intr_timecounter = { 202 .tc_get_timecount = get_intr_timecount, 203 .tc_poll_pps = NULL, 204 .tc_counter_mask = ~0u, 205 .tc_frequency = 0, 206 .tc_name = "clockinterrupt", 207 /* quality - minimum implementation level for a clock */ 208 .tc_quality = 0, 209 .tc_priv = NULL, 210 }; 211 212 static u_int 213 get_intr_timecount(struct timecounter *tc) 214 { 215 216 return (u_int)getticks(); 217 } 218 219 int 220 getticks(void) 221 { 222 return atomic_load_relaxed(&hardclock_ticks); 223 } 224 225 /* 226 * Initialize clock frequencies and start both clocks running. 227 */ 228 void 229 initclocks(void) 230 { 231 static struct sysctllog *clog; 232 int i; 233 234 /* 235 * Set divisors to 1 (normal case) and let the machine-specific 236 * code do its bit. 237 */ 238 psdiv = 1; 239 240 /* 241 * Call cpu_initclocks() before registering the default 242 * timecounter, in case it needs to adjust hz. 243 */ 244 const int old_hz = hz; 245 cpu_initclocks(); 246 if (old_hz != hz) { 247 tick = 1000000 / hz; 248 tickadj = (240000 / (60 * hz)) ? (240000 / (60 * hz)) : 1; 249 } 250 251 /* 252 * provide minimum default time counter 253 * will only run at interrupt resolution 254 */ 255 intr_timecounter.tc_frequency = hz; 256 tc_init(&intr_timecounter); 257 258 /* 259 * Compute profhz and stathz, fix profhz if needed. 260 */ 261 i = stathz ? stathz : hz; 262 if (profhz == 0) 263 profhz = i; 264 psratio = profhz / i; 265 if (schedhz == 0) { 266 /* 16Hz is best */ 267 hardscheddiv = hz / 16; 268 if (hardscheddiv <= 0) 269 panic("hardscheddiv"); 270 } 271 272 sysctl_createv(&clog, 0, NULL, NULL, 273 CTLFLAG_PERMANENT, 274 CTLTYPE_STRUCT, "clockrate", 275 SYSCTL_DESCR("Kernel clock rates"), 276 sysctl_kern_clockrate, 0, NULL, 277 sizeof(struct clockinfo), 278 CTL_KERN, KERN_CLOCKRATE, CTL_EOL); 279 sysctl_createv(&clog, 0, NULL, NULL, 280 CTLFLAG_PERMANENT, 281 CTLTYPE_INT, "hardclock_ticks", 282 SYSCTL_DESCR("Number of hardclock ticks"), 283 NULL, 0, &hardclock_ticks, sizeof(hardclock_ticks), 284 CTL_KERN, KERN_HARDCLOCK_TICKS, CTL_EOL); 285 286 rndsource_setcb(&hardclockrnd.source, clockrnd_get, &hardclockrnd); 287 rnd_attach_source(&hardclockrnd.source, "hardclock", RND_TYPE_SKEW, 288 RND_FLAG_COLLECT_TIME|RND_FLAG_ESTIMATE_TIME|RND_FLAG_HASCB); 289 if (stathz) { 290 rndsource_setcb(&statclockrnd.source, clockrnd_get, 291 &statclockrnd); 292 rnd_attach_source(&statclockrnd.source, "statclock", 293 RND_TYPE_SKEW, 294 (RND_FLAG_COLLECT_TIME|RND_FLAG_ESTIMATE_TIME| 295 RND_FLAG_HASCB)); 296 } 297 } 298 299 /* 300 * The real-time timer, interrupting hz times per second. 301 */ 302 void 303 hardclock(struct clockframe *frame) 304 { 305 struct lwp *l; 306 struct cpu_info *ci; 307 308 clockrnd_sample(&hardclockrnd); 309 310 ci = curcpu(); 311 l = ci->ci_onproc; 312 313 ptimer_tick(l, CLKF_USERMODE(frame)); 314 315 /* 316 * If no separate statistics clock is available, run it from here. 317 */ 318 if (stathz == 0) 319 statclock(frame); 320 /* 321 * If no separate schedclock is provided, call it here 322 * at about 16 Hz. 323 */ 324 if (schedhz == 0) { 325 if ((int)(--ci->ci_schedstate.spc_schedticks) <= 0) { 326 schedclock(l); 327 ci->ci_schedstate.spc_schedticks = hardscheddiv; 328 } 329 } 330 if ((--ci->ci_schedstate.spc_ticks) <= 0) 331 sched_tick(ci); 332 333 if (CPU_IS_PRIMARY(ci)) { 334 atomic_store_relaxed(&hardclock_ticks, 335 atomic_load_relaxed(&hardclock_ticks) + 1); 336 tc_ticktock(); 337 } 338 339 /* 340 * Make sure the CPUs and timecounter are making progress. 341 */ 342 heartbeat(); 343 344 /* 345 * Update real-time timeout queue. 346 */ 347 callout_hardclock(); 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(struct proc *p) 358 { 359 360 KASSERT(mutex_owned(&p->p_stmutex)); 361 362 if ((p->p_stflag & PST_PROFIL) == 0) { 363 p->p_stflag |= PST_PROFIL; 364 /* 365 * This is only necessary if using the clock as the 366 * profiling source. 367 */ 368 if (++profprocs == 1 && stathz != 0) 369 psdiv = psratio; 370 } 371 } 372 373 /* 374 * Stop profiling on a process. 375 */ 376 void 377 stopprofclock(struct proc *p) 378 { 379 380 KASSERT(mutex_owned(&p->p_stmutex)); 381 382 if (p->p_stflag & PST_PROFIL) { 383 p->p_stflag &= ~PST_PROFIL; 384 /* 385 * This is only necessary if using the clock as the 386 * profiling source. 387 */ 388 if (--profprocs == 0 && stathz != 0) 389 psdiv = 1; 390 } 391 } 392 393 void 394 schedclock(struct lwp *l) 395 { 396 if ((l->l_flag & LW_IDLE) != 0) 397 return; 398 399 sched_schedclock(l); 400 } 401 402 /* 403 * Statistics clock. Grab profile sample, and if divider reaches 0, 404 * do process and kernel statistics. 405 */ 406 void 407 statclock(struct clockframe *frame) 408 { 409 #ifdef GPROF 410 struct gmonparam *g; 411 intptr_t i; 412 #endif 413 struct cpu_info *ci = curcpu(); 414 struct schedstate_percpu *spc = &ci->ci_schedstate; 415 struct proc *p; 416 struct lwp *l; 417 418 if (stathz) 419 clockrnd_sample(&statclockrnd); 420 421 /* 422 * Notice changes in divisor frequency, and adjust clock 423 * frequency accordingly. 424 */ 425 if (spc->spc_psdiv != psdiv) { 426 spc->spc_psdiv = psdiv; 427 spc->spc_pscnt = psdiv; 428 if (psdiv == 1) { 429 setstatclockrate(stathz); 430 } else { 431 setstatclockrate(profhz); 432 } 433 } 434 l = ci->ci_onproc; 435 if ((l->l_flag & LW_IDLE) != 0) { 436 /* 437 * don't account idle lwps as swapper. 438 */ 439 p = NULL; 440 } else { 441 p = l->l_proc; 442 mutex_spin_enter(&p->p_stmutex); 443 } 444 445 if (CLKF_USERMODE(frame)) { 446 KASSERT(p != NULL); 447 if ((p->p_stflag & PST_PROFIL) && profsrc == PROFSRC_CLOCK) 448 addupc_intr(l, CLKF_PC(frame)); 449 if (--spc->spc_pscnt > 0) { 450 mutex_spin_exit(&p->p_stmutex); 451 return; 452 } 453 454 /* 455 * Came from user mode; CPU was in user state. 456 * If this process is being profiled record the tick. 457 */ 458 p->p_uticks++; 459 if (p->p_nice > NZERO) 460 spc->spc_cp_time[CP_NICE]++; 461 else 462 spc->spc_cp_time[CP_USER]++; 463 } else { 464 #ifdef GPROF 465 /* 466 * Kernel statistics are just like addupc_intr, only easier. 467 */ 468 #if defined(MULTIPROCESSOR) && !defined(_RUMPKERNEL) 469 g = curcpu()->ci_gmon; 470 if (g != NULL && 471 profsrc == PROFSRC_CLOCK && g->state == GMON_PROF_ON) { 472 #else 473 g = &_gmonparam; 474 if (profsrc == PROFSRC_CLOCK && g->state == GMON_PROF_ON) { 475 #endif 476 i = CLKF_PC(frame) - g->lowpc; 477 if (i < g->textsize) { 478 i /= HISTFRACTION * sizeof(*g->kcount); 479 g->kcount[i]++; 480 } 481 } 482 #endif 483 #ifdef LWP_PC 484 if (p != NULL && profsrc == PROFSRC_CLOCK && 485 (p->p_stflag & PST_PROFIL)) { 486 addupc_intr(l, LWP_PC(l)); 487 } 488 #endif 489 if (--spc->spc_pscnt > 0) { 490 if (p != NULL) 491 mutex_spin_exit(&p->p_stmutex); 492 return; 493 } 494 /* 495 * Came from kernel mode, so we were: 496 * - handling an interrupt, 497 * - doing syscall or trap work on behalf of the current 498 * user process, or 499 * - spinning in the idle loop. 500 * Whichever it is, charge the time as appropriate. 501 * Note that we charge interrupts to the current process, 502 * regardless of whether they are ``for'' that process, 503 * so that we know how much of its real time was spent 504 * in ``non-process'' (i.e., interrupt) work. 505 */ 506 if (CLKF_INTR(frame) || (curlwp->l_pflag & LP_INTR) != 0) { 507 if (p != NULL) { 508 p->p_iticks++; 509 } 510 spc->spc_cp_time[CP_INTR]++; 511 } else if (p != NULL) { 512 p->p_sticks++; 513 spc->spc_cp_time[CP_SYS]++; 514 } else { 515 spc->spc_cp_time[CP_IDLE]++; 516 } 517 } 518 spc->spc_pscnt = psdiv; 519 520 if (p != NULL) { 521 atomic_inc_uint(&l->l_cpticks); 522 mutex_spin_exit(&p->p_stmutex); 523 } 524 525 #ifdef KDTRACE_HOOKS 526 cyclic_clock_func_t func = cyclic_clock_func[cpu_index(ci)]; 527 if (func) { 528 (*func)((struct clockframe *)frame); 529 } 530 #endif 531 } 532 533 /* 534 * sysctl helper routine for kern.clockrate. Assembles a struct on 535 * the fly to be returned to the caller. 536 */ 537 static int 538 sysctl_kern_clockrate(SYSCTLFN_ARGS) 539 { 540 struct clockinfo clkinfo; 541 struct sysctlnode node; 542 543 clkinfo.tick = tick; 544 clkinfo.tickadj = tickadj; 545 clkinfo.hz = hz; 546 clkinfo.profhz = profhz; 547 clkinfo.stathz = stathz ? stathz : hz; 548 549 node = *rnode; 550 node.sysctl_data = &clkinfo; 551 return (sysctl_lookup(SYSCTLFN_CALL(&node))); 552 } 553