sched_4bsd.c revision 1.36
1 /* $NetBSD: sched_4bsd.c,v 1.36 2019/11/23 19:42:52 ad Exp $ */ 2 3 /* 4 * Copyright (c) 1999, 2000, 2004, 2006, 2007, 2008, 2019 5 * The NetBSD Foundation, Inc. 6 * All rights reserved. 7 * 8 * This code is derived from software contributed to The NetBSD Foundation 9 * by Jason R. Thorpe of the Numerical Aerospace Simulation Facility, 10 * NASA Ames Research Center, by Charles M. Hannum, Andrew Doran, and 11 * Daniel Sieger. 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, 1990, 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_synch.c 8.9 (Berkeley) 5/19/95 69 */ 70 71 #include <sys/cdefs.h> 72 __KERNEL_RCSID(0, "$NetBSD: sched_4bsd.c,v 1.36 2019/11/23 19:42:52 ad Exp $"); 73 74 #include "opt_ddb.h" 75 #include "opt_lockdebug.h" 76 77 #include <sys/param.h> 78 #include <sys/systm.h> 79 #include <sys/callout.h> 80 #include <sys/cpu.h> 81 #include <sys/proc.h> 82 #include <sys/kernel.h> 83 #include <sys/resourcevar.h> 84 #include <sys/sched.h> 85 #include <sys/sysctl.h> 86 #include <sys/lockdebug.h> 87 #include <sys/intr.h> 88 89 static void updatepri(struct lwp *); 90 static void resetpriority(struct lwp *); 91 92 extern unsigned int sched_pstats_ticks; /* defined in kern_synch.c */ 93 94 /* Number of hardclock ticks per sched_tick() */ 95 static int rrticks __read_mostly; 96 97 /* 98 * Force switch among equal priority processes every 100ms. 99 * Called from hardclock every hz/10 == rrticks hardclock ticks. 100 */ 101 /* ARGSUSED */ 102 void 103 sched_tick(struct cpu_info *ci) 104 { 105 struct schedstate_percpu *spc = &ci->ci_schedstate; 106 lwp_t *l; 107 108 spc->spc_ticks = rrticks; 109 110 if (CURCPU_IDLE_P()) { 111 atomic_or_uint(&ci->ci_want_resched, 112 RESCHED_IDLE | RESCHED_UPREEMPT); 113 return; 114 } 115 l = ci->ci_data.cpu_onproc; 116 if (l == NULL) { 117 return; 118 } 119 /* 120 * Can only be spc_lwplock or a turnstile lock at this point 121 * (if we interrupted priority inheritance trylock dance). 122 */ 123 KASSERT(l->l_mutex != spc->spc_mutex); 124 switch (l->l_class) { 125 case SCHED_FIFO: 126 /* No timeslicing for FIFO jobs. */ 127 break; 128 case SCHED_RR: 129 /* Force it into mi_switch() to look for other jobs to run. */ 130 atomic_or_uint(&l->l_dopreempt, DOPREEMPT_ACTIVE); 131 cpu_need_resched(ci, l, RESCHED_KPREEMPT); 132 break; 133 default: 134 if (spc->spc_flags & SPCF_SHOULDYIELD) { 135 /* 136 * Process is stuck in kernel somewhere, probably 137 * due to buggy or inefficient code. Force a 138 * kernel preemption. 139 */ 140 atomic_or_uint(&l->l_dopreempt, DOPREEMPT_ACTIVE); 141 cpu_need_resched(ci, l, RESCHED_KPREEMPT); 142 } else if (spc->spc_flags & SPCF_SEENRR) { 143 /* 144 * The process has already been through a roundrobin 145 * without switching and may be hogging the CPU. 146 * Indicate that the process should yield. 147 */ 148 spc->spc_flags |= SPCF_SHOULDYIELD; 149 cpu_need_resched(ci, l, RESCHED_UPREEMPT); 150 } else { 151 spc->spc_flags |= SPCF_SEENRR; 152 } 153 break; 154 } 155 } 156 157 /* 158 * Why PRIO_MAX - 2? From setpriority(2): 159 * 160 * prio is a value in the range -20 to 20. The default priority is 161 * 0; lower priorities cause more favorable scheduling. A value of 162 * 19 or 20 will schedule a process only when nothing at priority <= 163 * 0 is runnable. 164 * 165 * This gives estcpu influence over 18 priority levels, and leaves nice 166 * with 40 levels. One way to think about it is that nice has 20 levels 167 * either side of estcpu's 18. 168 */ 169 #define ESTCPU_SHIFT 11 170 #define ESTCPU_MAX ((PRIO_MAX - 2) << ESTCPU_SHIFT) 171 #define ESTCPU_ACCUM (1 << (ESTCPU_SHIFT - 1)) 172 #define ESTCPULIM(e) uimin((e), ESTCPU_MAX) 173 174 /* 175 * The main parameter used by this algorithm is 'l_estcpu'. It is an estimate 176 * of the recent CPU utilization of the thread. 177 * 178 * l_estcpu is: 179 * - increased each time the hardclock ticks and the thread is found to 180 * be executing, in sched_schedclock() called from hardclock() 181 * - decreased (filtered) on each sched tick, in sched_pstats_hook() 182 * If the lwp is sleeping for more than a second, we don't touch l_estcpu: it 183 * will be updated in sched_setrunnable() when the lwp wakes up, in burst mode 184 * (ie, we decrease it n times). 185 * 186 * Note that hardclock updates l_estcpu and l_cpticks independently. 187 * 188 * ----------------------------------------------------------------------------- 189 * 190 * Here we describe how l_estcpu is decreased. 191 * 192 * Constants for digital decay (filter): 193 * 90% of l_estcpu usage in (5 * loadavg) seconds 194 * 195 * We wish to decay away 90% of l_estcpu in (5 * loadavg) seconds. That is, we 196 * want to compute a value of decay such that the following loop: 197 * for (i = 0; i < (5 * loadavg); i++) 198 * l_estcpu *= decay; 199 * will result in 200 * l_estcpu *= 0.1; 201 * for all values of loadavg. 202 * 203 * Mathematically this loop can be expressed by saying: 204 * decay ** (5 * loadavg) ~= .1 205 * 206 * And finally, the corresponding value of decay we're using is: 207 * decay = (2 * loadavg) / (2 * loadavg + 1) 208 * 209 * ----------------------------------------------------------------------------- 210 * 211 * Now, let's prove that the value of decay stated above will always fulfill 212 * the equation: 213 * decay ** (5 * loadavg) ~= .1 214 * 215 * If we compute b as: 216 * b = 2 * loadavg 217 * then 218 * decay = b / (b + 1) 219 * 220 * We now need to prove two things: 221 * 1) Given [factor ** (5 * loadavg) =~ .1], prove [factor == b/(b+1)]. 222 * 2) Given [b/(b+1) ** power =~ .1], prove [power == (5 * loadavg)]. 223 * 224 * Facts: 225 * * For x real: exp(x) = 0! + x**1/1! + x**2/2! + ... 226 * Therefore, for x close to zero, exp(x) =~ 1 + x. 227 * In turn, for b large enough, exp(-1/b) =~ 1 - (1/b) = (b-1)/b. 228 * 229 * * For b large enough, (b-1)/b =~ b/(b+1). 230 * 231 * * For x belonging to [-1;1[, ln(1-x) = - x - x**2/2 - x**3/3 - ... 232 * Therefore ln(b/(b+1)) = ln(1 - 1/(b+1)) =~ -1/(b+1). 233 * 234 * * ln(0.1) =~ -2.30 235 * 236 * Proof of (1): 237 * factor ** (5 * loadavg) =~ 0.1 238 * => ln(factor) =~ -2.30 / (5 * loadavg) 239 * => factor =~ exp(-1 / ((5 / 2.30) * loadavg)) 240 * =~ exp(-1 / (2 * loadavg)) 241 * =~ exp(-1 / b) 242 * =~ (b - 1) / b 243 * =~ b / (b + 1) 244 * =~ (2 * loadavg) / ((2 * loadavg) + 1) 245 * 246 * Proof of (2): 247 * (b / (b + 1)) ** power =~ .1 248 * => power * ln(b / (b + 1)) =~ -2.30 249 * => power * (-1 / (b + 1)) =~ -2.30 250 * => power =~ 2.30 * (b + 1) 251 * => power =~ 4.60 * loadavg + 2.30 252 * => power =~ 5 * loadavg 253 * 254 * Conclusion: decay = (2 * loadavg) / (2 * loadavg + 1) 255 */ 256 257 /* See calculations above */ 258 #define loadfactor(loadavg) (2 * (loadavg)) 259 260 static fixpt_t 261 decay_cpu(fixpt_t loadfac, fixpt_t estcpu) 262 { 263 264 if (estcpu == 0) { 265 return 0; 266 } 267 268 #if !defined(_LP64) 269 /* avoid 64bit arithmetics. */ 270 #define FIXPT_MAX ((fixpt_t)((UINTMAX_C(1) << sizeof(fixpt_t) * CHAR_BIT) - 1)) 271 if (__predict_true(loadfac <= FIXPT_MAX / ESTCPU_MAX)) { 272 return estcpu * loadfac / (loadfac + FSCALE); 273 } 274 #endif 275 276 return (uint64_t)estcpu * loadfac / (loadfac + FSCALE); 277 } 278 279 static fixpt_t 280 decay_cpu_batch(fixpt_t loadfac, fixpt_t estcpu, unsigned int n) 281 { 282 283 /* 284 * For all load averages >= 1 and max l_estcpu of (255 << ESTCPU_SHIFT), 285 * if we slept for at least seven times the loadfactor, we will decay 286 * l_estcpu to less than (1 << ESTCPU_SHIFT), and therefore we can 287 * return zero directly. 288 * 289 * Note that our ESTCPU_MAX is actually much smaller than 290 * (255 << ESTCPU_SHIFT). 291 */ 292 if ((n << FSHIFT) >= 7 * loadfac) { 293 return 0; 294 } 295 296 while (estcpu != 0 && n > 1) { 297 estcpu = decay_cpu(loadfac, estcpu); 298 n--; 299 } 300 301 return estcpu; 302 } 303 304 /* 305 * sched_pstats_hook: 306 * 307 * Periodically called from sched_pstats(); used to recalculate priorities. 308 */ 309 void 310 sched_pstats_hook(struct lwp *l, int batch) 311 { 312 fixpt_t loadfac; 313 314 /* 315 * If the LWP has slept an entire second, stop recalculating 316 * its priority until it wakes up. 317 */ 318 KASSERT(lwp_locked(l, NULL)); 319 if (l->l_stat == LSSLEEP || l->l_stat == LSSTOP || 320 l->l_stat == LSSUSPENDED) { 321 if (l->l_slptime > 1) { 322 return; 323 } 324 } 325 326 loadfac = loadfactor(averunnable.ldavg[0]); 327 l->l_estcpu = decay_cpu(loadfac, l->l_estcpu); 328 resetpriority(l); 329 } 330 331 /* 332 * Recalculate the priority of an LWP after it has slept for a while. 333 */ 334 static void 335 updatepri(struct lwp *l) 336 { 337 fixpt_t loadfac; 338 339 KASSERT(lwp_locked(l, NULL)); 340 KASSERT(l->l_slptime > 1); 341 342 loadfac = loadfactor(averunnable.ldavg[0]); 343 344 l->l_slptime--; /* the first time was done in sched_pstats */ 345 l->l_estcpu = decay_cpu_batch(loadfac, l->l_estcpu, l->l_slptime); 346 resetpriority(l); 347 } 348 349 void 350 sched_rqinit(void) 351 { 352 353 } 354 355 void 356 sched_setrunnable(struct lwp *l) 357 { 358 359 if (l->l_slptime > 1) 360 updatepri(l); 361 } 362 363 void 364 sched_nice(struct proc *p, int n) 365 { 366 struct lwp *l; 367 368 KASSERT(mutex_owned(p->p_lock)); 369 370 p->p_nice = n; 371 LIST_FOREACH(l, &p->p_lwps, l_sibling) { 372 lwp_lock(l); 373 resetpriority(l); 374 lwp_unlock(l); 375 } 376 } 377 378 /* 379 * Recompute the priority of an LWP. Arrange to reschedule if 380 * the resulting priority is better than that of the current LWP. 381 */ 382 static void 383 resetpriority(struct lwp *l) 384 { 385 pri_t pri; 386 struct proc *p = l->l_proc; 387 388 KASSERT(lwp_locked(l, NULL)); 389 390 if (l->l_class != SCHED_OTHER) 391 return; 392 393 /* See comments above ESTCPU_SHIFT definition. */ 394 pri = (PRI_KERNEL - 1) - (l->l_estcpu >> ESTCPU_SHIFT) - p->p_nice; 395 pri = imax(pri, 0); 396 if (pri != l->l_priority) 397 lwp_changepri(l, pri); 398 } 399 400 /* 401 * We adjust the priority of the current LWP. The priority of a LWP 402 * gets worse as it accumulates CPU time. The CPU usage estimator (l_estcpu) 403 * is increased here. The formula for computing priorities will compute a 404 * different value each time l_estcpu increases. This can cause a switch, 405 * but unless the priority crosses a PPQ boundary the actual queue will not 406 * change. The CPU usage estimator ramps up quite quickly when the process 407 * is running (linearly), and decays away exponentially, at a rate which is 408 * proportionally slower when the system is busy. The basic principle is 409 * that the system will 90% forget that the process used a lot of CPU time 410 * in (5 * loadavg) seconds. This causes the system to favor processes which 411 * haven't run much recently, and to round-robin among other processes. 412 */ 413 void 414 sched_schedclock(struct lwp *l) 415 { 416 417 if (l->l_class != SCHED_OTHER) 418 return; 419 420 KASSERT(!CURCPU_IDLE_P()); 421 l->l_estcpu = ESTCPULIM(l->l_estcpu + ESTCPU_ACCUM); 422 lwp_lock(l); 423 resetpriority(l); 424 lwp_unlock(l); 425 } 426 427 /* 428 * sched_proc_fork: 429 * 430 * Inherit the parent's scheduler history. 431 */ 432 void 433 sched_proc_fork(struct proc *parent, struct proc *child) 434 { 435 lwp_t *pl; 436 437 KASSERT(mutex_owned(parent->p_lock)); 438 439 pl = LIST_FIRST(&parent->p_lwps); 440 child->p_estcpu_inherited = pl->l_estcpu; 441 child->p_forktime = sched_pstats_ticks; 442 } 443 444 /* 445 * sched_proc_exit: 446 * 447 * Chargeback parents for the sins of their children. 448 */ 449 void 450 sched_proc_exit(struct proc *parent, struct proc *child) 451 { 452 fixpt_t loadfac = loadfactor(averunnable.ldavg[0]); 453 fixpt_t estcpu; 454 lwp_t *pl, *cl; 455 456 /* XXX Only if parent != init?? */ 457 458 mutex_enter(parent->p_lock); 459 pl = LIST_FIRST(&parent->p_lwps); 460 cl = LIST_FIRST(&child->p_lwps); 461 estcpu = decay_cpu_batch(loadfac, child->p_estcpu_inherited, 462 sched_pstats_ticks - child->p_forktime); 463 if (cl->l_estcpu > estcpu) { 464 lwp_lock(pl); 465 pl->l_estcpu = ESTCPULIM(pl->l_estcpu + cl->l_estcpu - estcpu); 466 lwp_unlock(pl); 467 } 468 mutex_exit(parent->p_lock); 469 } 470 471 void 472 sched_wakeup(struct lwp *l) 473 { 474 475 } 476 477 void 478 sched_slept(struct lwp *l) 479 { 480 481 } 482 483 void 484 sched_lwp_fork(struct lwp *l1, struct lwp *l2) 485 { 486 487 l2->l_estcpu = l1->l_estcpu; 488 } 489 490 void 491 sched_lwp_collect(struct lwp *t) 492 { 493 lwp_t *l; 494 495 /* Absorb estcpu value of collected LWP. */ 496 l = curlwp; 497 lwp_lock(l); 498 l->l_estcpu += t->l_estcpu; 499 lwp_unlock(l); 500 } 501 502 void 503 sched_oncpu(lwp_t *l) 504 { 505 506 } 507 508 void 509 sched_newts(lwp_t *l) 510 { 511 512 } 513 514 /* 515 * Sysctl nodes and initialization. 516 */ 517 518 static int 519 sysctl_sched_rtts(SYSCTLFN_ARGS) 520 { 521 struct sysctlnode node; 522 int rttsms = hztoms(rrticks); 523 524 node = *rnode; 525 node.sysctl_data = &rttsms; 526 return sysctl_lookup(SYSCTLFN_CALL(&node)); 527 } 528 529 SYSCTL_SETUP(sysctl_sched_4bsd_setup, "sysctl sched setup") 530 { 531 const struct sysctlnode *node = NULL; 532 533 sysctl_createv(clog, 0, NULL, &node, 534 CTLFLAG_PERMANENT, 535 CTLTYPE_NODE, "sched", 536 SYSCTL_DESCR("Scheduler options"), 537 NULL, 0, NULL, 0, 538 CTL_KERN, CTL_CREATE, CTL_EOL); 539 540 if (node == NULL) 541 return; 542 543 rrticks = hz / 10; 544 545 sysctl_createv(NULL, 0, &node, NULL, 546 CTLFLAG_PERMANENT, 547 CTLTYPE_STRING, "name", NULL, 548 NULL, 0, __UNCONST("4.4BSD"), 0, 549 CTL_CREATE, CTL_EOL); 550 sysctl_createv(NULL, 0, &node, NULL, 551 CTLFLAG_PERMANENT, 552 CTLTYPE_INT, "rtts", 553 SYSCTL_DESCR("Round-robin time quantum (in milliseconds)"), 554 sysctl_sched_rtts, 0, NULL, 0, 555 CTL_CREATE, CTL_EOL); 556 } 557
Indexes created Thu Sep 25 08:09:54 GMT 2025