sched_4bsd.c revision 1.18
1 /* $NetBSD: sched_4bsd.c,v 1.18 2008/04/14 09:40:43 yamt Exp $ */ 2 3 /*- 4 * Copyright (c) 1999, 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, by Charles M. Hannum, Andrew Doran, and 10 * Daniel Sieger. 11 * 12 * Redistribution and use in source and binary forms, with or without 13 * modification, are permitted provided that the following conditions 14 * are met: 15 * 1. Redistributions of source code must retain the above copyright 16 * notice, this list of conditions and the following disclaimer. 17 * 2. Redistributions in binary form must reproduce the above copyright 18 * notice, this list of conditions and the following disclaimer in the 19 * documentation and/or other materials provided with the distribution. 20 * 3. All advertising materials mentioning features or use of this software 21 * must display the following acknowledgement: 22 * This product includes software developed by the NetBSD 23 * Foundation, Inc. and its contributors. 24 * 4. Neither the name of The NetBSD Foundation nor the names of its 25 * contributors may be used to endorse or promote products derived 26 * from this software without specific prior written permission. 27 * 28 * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS 29 * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED 30 * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR 31 * PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS 32 * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR 33 * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF 34 * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS 35 * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN 36 * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) 37 * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE 38 * POSSIBILITY OF SUCH DAMAGE. 39 */ 40 41 /*- 42 * Copyright (c) 1982, 1986, 1990, 1991, 1993 43 * The Regents of the University of California. All rights reserved. 44 * (c) UNIX System Laboratories, Inc. 45 * All or some portions of this file are derived from material licensed 46 * to the University of California by American Telephone and Telegraph 47 * Co. or Unix System Laboratories, Inc. and are reproduced herein with 48 * the permission of UNIX System Laboratories, Inc. 49 * 50 * Redistribution and use in source and binary forms, with or without 51 * modification, are permitted provided that the following conditions 52 * are met: 53 * 1. Redistributions of source code must retain the above copyright 54 * notice, this list of conditions and the following disclaimer. 55 * 2. Redistributions in binary form must reproduce the above copyright 56 * notice, this list of conditions and the following disclaimer in the 57 * documentation and/or other materials provided with the distribution. 58 * 3. Neither the name of the University nor the names of its contributors 59 * may be used to endorse or promote products derived from this software 60 * without specific prior written permission. 61 * 62 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 63 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 64 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 65 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 66 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 67 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 68 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 69 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 70 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 71 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 72 * SUCH DAMAGE. 73 * 74 * @(#)kern_synch.c 8.9 (Berkeley) 5/19/95 75 */ 76 77 #include <sys/cdefs.h> 78 __KERNEL_RCSID(0, "$NetBSD: sched_4bsd.c,v 1.18 2008/04/14 09:40:43 yamt Exp $"); 79 80 #include "opt_ddb.h" 81 #include "opt_lockdebug.h" 82 #include "opt_perfctrs.h" 83 84 #include <sys/param.h> 85 #include <sys/systm.h> 86 #include <sys/callout.h> 87 #include <sys/cpu.h> 88 #include <sys/proc.h> 89 #include <sys/kernel.h> 90 #include <sys/signalvar.h> 91 #include <sys/resourcevar.h> 92 #include <sys/sched.h> 93 #include <sys/sysctl.h> 94 #include <sys/kauth.h> 95 #include <sys/lockdebug.h> 96 #include <sys/kmem.h> 97 #include <sys/intr.h> 98 99 #include <uvm/uvm_extern.h> 100 101 static void updatepri(struct lwp *); 102 static void resetpriority(struct lwp *); 103 104 extern unsigned int sched_pstats_ticks; /* defined in kern_synch.c */ 105 106 /* Number of hardclock ticks per sched_tick() */ 107 static int rrticks; 108 109 /* 110 * Force switch among equal priority processes every 100ms. 111 * Called from hardclock every hz/10 == rrticks hardclock ticks. 112 * 113 * There's no need to lock anywhere in this routine, as it's 114 * CPU-local and runs at IPL_SCHED (called from clock interrupt). 115 */ 116 /* ARGSUSED */ 117 void 118 sched_tick(struct cpu_info *ci) 119 { 120 struct schedstate_percpu *spc = &ci->ci_schedstate; 121 122 spc->spc_ticks = rrticks; 123 124 if (CURCPU_IDLE_P()) { 125 cpu_need_resched(ci, 0); 126 return; 127 } 128 129 if (spc->spc_flags & SPCF_SEENRR) { 130 /* 131 * The process has already been through a roundrobin 132 * without switching and may be hogging the CPU. 133 * Indicate that the process should yield. 134 */ 135 spc->spc_flags |= SPCF_SHOULDYIELD; 136 cpu_need_resched(ci, 0); 137 } else 138 spc->spc_flags |= SPCF_SEENRR; 139 } 140 141 /* 142 * Why PRIO_MAX - 2? From setpriority(2): 143 * 144 * prio is a value in the range -20 to 20. The default priority is 145 * 0; lower priorities cause more favorable scheduling. A value of 146 * 19 or 20 will schedule a process only when nothing at priority <= 147 * 0 is runnable. 148 * 149 * This gives estcpu influence over 18 priority levels, and leaves nice 150 * with 40 levels. One way to think about it is that nice has 20 levels 151 * either side of estcpu's 18. 152 */ 153 #define ESTCPU_SHIFT 11 154 #define ESTCPU_MAX ((PRIO_MAX - 2) << ESTCPU_SHIFT) 155 #define ESTCPU_ACCUM (1 << (ESTCPU_SHIFT - 1)) 156 #define ESTCPULIM(e) min((e), ESTCPU_MAX) 157 158 /* 159 * Constants for digital decay and forget: 160 * 90% of (l_estcpu) usage in 5 * loadav time 161 * 95% of (l_pctcpu) usage in 60 seconds (load insensitive) 162 * Note that, as ps(1) mentions, this can let percentages 163 * total over 100% (I've seen 137.9% for 3 processes). 164 * 165 * Note that hardclock updates l_estcpu and l_cpticks independently. 166 * 167 * We wish to decay away 90% of l_estcpu in (5 * loadavg) seconds. 168 * That is, the system wants to compute a value of decay such 169 * that the following for loop: 170 * for (i = 0; i < (5 * loadavg); i++) 171 * l_estcpu *= decay; 172 * will compute 173 * l_estcpu *= 0.1; 174 * for all values of loadavg: 175 * 176 * Mathematically this loop can be expressed by saying: 177 * decay ** (5 * loadavg) ~= .1 178 * 179 * The system computes decay as: 180 * decay = (2 * loadavg) / (2 * loadavg + 1) 181 * 182 * We wish to prove that the system's computation of decay 183 * will always fulfill the equation: 184 * decay ** (5 * loadavg) ~= .1 185 * 186 * If we compute b as: 187 * b = 2 * loadavg 188 * then 189 * decay = b / (b + 1) 190 * 191 * We now need to prove two things: 192 * 1) Given factor ** (5 * loadavg) ~= .1, prove factor == b/(b+1) 193 * 2) Given b/(b+1) ** power ~= .1, prove power == (5 * loadavg) 194 * 195 * Facts: 196 * For x close to zero, exp(x) =~ 1 + x, since 197 * exp(x) = 0! + x**1/1! + x**2/2! + ... . 198 * therefore exp(-1/b) =~ 1 - (1/b) = (b-1)/b. 199 * For x close to zero, ln(1+x) =~ x, since 200 * ln(1+x) = x - x**2/2 + x**3/3 - ... -1 < x < 1 201 * therefore ln(b/(b+1)) = ln(1 - 1/(b+1)) =~ -1/(b+1). 202 * ln(.1) =~ -2.30 203 * 204 * Proof of (1): 205 * Solve (factor)**(power) =~ .1 given power (5*loadav): 206 * solving for factor, 207 * ln(factor) =~ (-2.30/5*loadav), or 208 * factor =~ exp(-1/((5/2.30)*loadav)) =~ exp(-1/(2*loadav)) = 209 * exp(-1/b) =~ (b-1)/b =~ b/(b+1). QED 210 * 211 * Proof of (2): 212 * Solve (factor)**(power) =~ .1 given factor == (b/(b+1)): 213 * solving for power, 214 * power*ln(b/(b+1)) =~ -2.30, or 215 * power =~ 2.3 * (b + 1) = 4.6*loadav + 2.3 =~ 5*loadav. QED 216 * 217 * Actual power values for the implemented algorithm are as follows: 218 * loadav: 1 2 3 4 219 * power: 5.68 10.32 14.94 19.55 220 */ 221 222 /* calculations for digital decay to forget 90% of usage in 5*loadav sec */ 223 #define loadfactor(loadav) (2 * (loadav)) 224 225 static fixpt_t 226 decay_cpu(fixpt_t loadfac, fixpt_t estcpu) 227 { 228 229 if (estcpu == 0) { 230 return 0; 231 } 232 233 #if !defined(_LP64) 234 /* avoid 64bit arithmetics. */ 235 #define FIXPT_MAX ((fixpt_t)((UINTMAX_C(1) << sizeof(fixpt_t) * CHAR_BIT) - 1)) 236 if (__predict_true(loadfac <= FIXPT_MAX / ESTCPU_MAX)) { 237 return estcpu * loadfac / (loadfac + FSCALE); 238 } 239 #endif /* !defined(_LP64) */ 240 241 return (uint64_t)estcpu * loadfac / (loadfac + FSCALE); 242 } 243 244 /* 245 * For all load averages >= 1 and max l_estcpu of (255 << ESTCPU_SHIFT), 246 * sleeping for at least seven times the loadfactor will decay l_estcpu to 247 * less than (1 << ESTCPU_SHIFT). 248 * 249 * note that our ESTCPU_MAX is actually much smaller than (255 << ESTCPU_SHIFT). 250 */ 251 static fixpt_t 252 decay_cpu_batch(fixpt_t loadfac, fixpt_t estcpu, unsigned int n) 253 { 254 255 if ((n << FSHIFT) >= 7 * loadfac) { 256 return 0; 257 } 258 259 while (estcpu != 0 && n > 1) { 260 estcpu = decay_cpu(loadfac, estcpu); 261 n--; 262 } 263 264 return estcpu; 265 } 266 267 /* 268 * sched_pstats_hook: 269 * 270 * Periodically called from sched_pstats(); used to recalculate priorities. 271 */ 272 void 273 sched_pstats_hook(struct lwp *l) 274 { 275 fixpt_t loadfac; 276 int sleeptm; 277 278 /* 279 * If the LWP has slept an entire second, stop recalculating 280 * its priority until it wakes up. 281 */ 282 if (l->l_stat == LSSLEEP || l->l_stat == LSSTOP || 283 l->l_stat == LSSUSPENDED) { 284 l->l_slptime++; 285 sleeptm = 1; 286 } else { 287 sleeptm = 0x7fffffff; 288 } 289 290 if (l->l_slptime <= sleeptm) { 291 loadfac = 2 * (averunnable.ldavg[0]); 292 l->l_estcpu = decay_cpu(loadfac, l->l_estcpu); 293 resetpriority(l); 294 } 295 } 296 297 /* 298 * Recalculate the priority of a process after it has slept for a while. 299 */ 300 static void 301 updatepri(struct lwp *l) 302 { 303 fixpt_t loadfac; 304 305 KASSERT(lwp_locked(l, NULL)); 306 KASSERT(l->l_slptime > 1); 307 308 loadfac = loadfactor(averunnable.ldavg[0]); 309 310 l->l_slptime--; /* the first time was done in sched_pstats */ 311 l->l_estcpu = decay_cpu_batch(loadfac, l->l_estcpu, l->l_slptime); 312 resetpriority(l); 313 } 314 315 void 316 sched_rqinit(void) 317 { 318 319 } 320 321 void 322 sched_setrunnable(struct lwp *l) 323 { 324 325 if (l->l_slptime > 1) 326 updatepri(l); 327 } 328 329 void 330 sched_nice(struct proc *p, int n) 331 { 332 struct lwp *l; 333 334 KASSERT(mutex_owned(&p->p_smutex)); 335 336 p->p_nice = n; 337 LIST_FOREACH(l, &p->p_lwps, l_sibling) { 338 lwp_lock(l); 339 resetpriority(l); 340 lwp_unlock(l); 341 } 342 } 343 344 /* 345 * Recompute the priority of an LWP. Arrange to reschedule if 346 * the resulting priority is better than that of the current LWP. 347 */ 348 static void 349 resetpriority(struct lwp *l) 350 { 351 pri_t pri; 352 struct proc *p = l->l_proc; 353 354 KASSERT(lwp_locked(l, NULL)); 355 356 if (l->l_class != SCHED_OTHER) 357 return; 358 359 /* See comments above ESTCPU_SHIFT definition. */ 360 pri = (PRI_KERNEL - 1) - (l->l_estcpu >> ESTCPU_SHIFT) - p->p_nice; 361 pri = imax(pri, 0); 362 if (pri != l->l_priority) 363 lwp_changepri(l, pri); 364 } 365 366 /* 367 * We adjust the priority of the current process. The priority of a process 368 * gets worse as it accumulates CPU time. The CPU usage estimator (l_estcpu) 369 * is increased here. The formula for computing priorities (in kern_synch.c) 370 * will compute a different value each time l_estcpu increases. This can 371 * cause a switch, but unless the priority crosses a PPQ boundary the actual 372 * queue will not change. The CPU usage estimator ramps up quite quickly 373 * when the process is running (linearly), and decays away exponentially, at 374 * a rate which is proportionally slower when the system is busy. The basic 375 * principle is that the system will 90% forget that the process used a lot 376 * of CPU time in 5 * loadav seconds. This causes the system to favor 377 * processes which haven't run much recently, and to round-robin among other 378 * processes. 379 */ 380 381 void 382 sched_schedclock(struct lwp *l) 383 { 384 385 if (l->l_class != SCHED_OTHER) 386 return; 387 388 KASSERT(!CURCPU_IDLE_P()); 389 l->l_estcpu = ESTCPULIM(l->l_estcpu + ESTCPU_ACCUM); 390 lwp_lock(l); 391 resetpriority(l); 392 lwp_unlock(l); 393 } 394 395 /* 396 * sched_proc_fork: 397 * 398 * Inherit the parent's scheduler history. 399 */ 400 void 401 sched_proc_fork(struct proc *parent, struct proc *child) 402 { 403 lwp_t *pl; 404 405 KASSERT(mutex_owned(&parent->p_smutex)); 406 407 pl = LIST_FIRST(&parent->p_lwps); 408 child->p_estcpu_inherited = pl->l_estcpu; 409 child->p_forktime = sched_pstats_ticks; 410 } 411 412 /* 413 * sched_proc_exit: 414 * 415 * Chargeback parents for the sins of their children. 416 */ 417 void 418 sched_proc_exit(struct proc *parent, struct proc *child) 419 { 420 fixpt_t loadfac = loadfactor(averunnable.ldavg[0]); 421 fixpt_t estcpu; 422 lwp_t *pl, *cl; 423 424 /* XXX Only if parent != init?? */ 425 426 mutex_enter(&parent->p_smutex); 427 pl = LIST_FIRST(&parent->p_lwps); 428 cl = LIST_FIRST(&child->p_lwps); 429 estcpu = decay_cpu_batch(loadfac, child->p_estcpu_inherited, 430 sched_pstats_ticks - child->p_forktime); 431 if (cl->l_estcpu > estcpu) { 432 lwp_lock(pl); 433 pl->l_estcpu = ESTCPULIM(pl->l_estcpu + cl->l_estcpu - estcpu); 434 lwp_unlock(pl); 435 } 436 mutex_exit(&parent->p_smutex); 437 } 438 439 void 440 sched_wakeup(struct lwp *l) 441 { 442 443 l->l_cpu = sched_takecpu(l); 444 } 445 446 void 447 sched_slept(struct lwp *l) 448 { 449 450 } 451 452 void 453 sched_lwp_fork(struct lwp *l1, struct lwp *l2) 454 { 455 456 l2->l_estcpu = l1->l_estcpu; 457 } 458 459 void 460 sched_lwp_exit(struct lwp *l) 461 { 462 463 } 464 465 void 466 sched_lwp_collect(struct lwp *t) 467 { 468 lwp_t *l; 469 470 /* Absorb estcpu value of collected LWP. */ 471 l = curlwp; 472 lwp_lock(l); 473 l->l_estcpu += t->l_estcpu; 474 lwp_unlock(l); 475 } 476 477 void 478 sched_oncpu(lwp_t *l) 479 { 480 481 } 482 483 void 484 sched_newts(lwp_t *l) 485 { 486 487 } 488 489 /* 490 * Sysctl nodes and initialization. 491 */ 492 493 static int 494 sysctl_sched_rtts(SYSCTLFN_ARGS) 495 { 496 struct sysctlnode node; 497 int rttsms = hztoms(rrticks); 498 499 node = *rnode; 500 node.sysctl_data = &rttsms; 501 return sysctl_lookup(SYSCTLFN_CALL(&node)); 502 } 503 504 SYSCTL_SETUP(sysctl_sched_4bsd_setup, "sysctl sched setup") 505 { 506 const struct sysctlnode *node = NULL; 507 508 sysctl_createv(clog, 0, NULL, NULL, 509 CTLFLAG_PERMANENT, 510 CTLTYPE_NODE, "kern", NULL, 511 NULL, 0, NULL, 0, 512 CTL_KERN, CTL_EOL); 513 sysctl_createv(clog, 0, NULL, &node, 514 CTLFLAG_PERMANENT, 515 CTLTYPE_NODE, "sched", 516 SYSCTL_DESCR("Scheduler options"), 517 NULL, 0, NULL, 0, 518 CTL_KERN, CTL_CREATE, CTL_EOL); 519 520 if (node == NULL) 521 return; 522 523 rrticks = hz / 10; 524 525 sysctl_createv(NULL, 0, &node, NULL, 526 CTLFLAG_PERMANENT, 527 CTLTYPE_STRING, "name", NULL, 528 NULL, 0, __UNCONST("4.4BSD"), 0, 529 CTL_CREATE, CTL_EOL); 530 sysctl_createv(NULL, 0, &node, NULL, 531 CTLFLAG_PERMANENT, 532 CTLTYPE_INT, "rtts", 533 SYSCTL_DESCR("Round-robin time quantum (in miliseconds)"), 534 sysctl_sched_rtts, 0, NULL, 0, 535 CTL_CREATE, CTL_EOL); 536 } 537
Indexes created Tue Sep 30 06:09:59 GMT 2025