sched_4bsd.c revision 1.16
1 /* $NetBSD: sched_4bsd.c,v 1.16 2008/04/12 17:02:08 ad 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 1.16 2008/04/12 17:02:08 ad Exp $"); 79 80 #include "opt_ddb.h" 81 #include "opt_lockdebug.h" 82 #include "opt_perfctrs.h" 83 84 #define __MUTEX_PRIVATE 85 86 #include <sys/param.h> 87 #include <sys/systm.h> 88 #include <sys/callout.h> 89 #include <sys/cpu.h> 90 #include <sys/proc.h> 91 #include <sys/kernel.h> 92 #include <sys/signalvar.h> 93 #include <sys/resourcevar.h> 94 #include <sys/sched.h> 95 #include <sys/sysctl.h> 96 #include <sys/kauth.h> 97 #include <sys/lockdebug.h> 98 #include <sys/kmem.h> 99 #include <sys/intr.h> 100 101 #include <uvm/uvm_extern.h> 102 103 static void updatepri(struct lwp *); 104 static void resetpriority(struct lwp *); 105 106 fixpt_t decay_cpu(fixpt_t, fixpt_t); 107 108 extern unsigned int sched_pstats_ticks; /* defined in kern_synch.c */ 109 110 /* Number of hardclock ticks per sched_tick() */ 111 static int rrticks; 112 113 /* 114 * Force switch among equal priority processes every 100ms. 115 * Called from hardclock every hz/10 == rrticks hardclock ticks. 116 * 117 * There's no need to lock anywhere in this routine, as it's 118 * CPU-local and runs at IPL_SCHED (called from clock interrupt). 119 */ 120 /* ARGSUSED */ 121 void 122 sched_tick(struct cpu_info *ci) 123 { 124 struct schedstate_percpu *spc = &ci->ci_schedstate; 125 126 spc->spc_ticks = rrticks; 127 128 if (CURCPU_IDLE_P()) { 129 cpu_need_resched(ci, 0); 130 return; 131 } 132 133 if (spc->spc_flags & SPCF_SEENRR) { 134 /* 135 * The process has already been through a roundrobin 136 * without switching and may be hogging the CPU. 137 * Indicate that the process should yield. 138 */ 139 spc->spc_flags |= SPCF_SHOULDYIELD; 140 cpu_need_resched(ci, 0); 141 } else 142 spc->spc_flags |= SPCF_SEENRR; 143 } 144 145 /* 146 * Why PRIO_MAX - 2? From setpriority(2): 147 * 148 * prio is a value in the range -20 to 20. The default priority is 149 * 0; lower priorities cause more favorable scheduling. A value of 150 * 19 or 20 will schedule a process only when nothing at priority <= 151 * 0 is runnable. 152 * 153 * This gives estcpu influence over 18 priority levels, and leaves nice 154 * with 40 levels. One way to think about it is that nice has 20 levels 155 * either side of estcpu's 18. 156 */ 157 #define ESTCPU_SHIFT 11 158 #define ESTCPU_MAX ((PRIO_MAX - 2) << ESTCPU_SHIFT) 159 #define ESTCPU_ACCUM (1 << (ESTCPU_SHIFT - 1)) 160 #define ESTCPULIM(e) min((e), ESTCPU_MAX) 161 162 /* 163 * Constants for digital decay and forget: 164 * 90% of (l_estcpu) usage in 5 * loadav time 165 * 95% of (l_pctcpu) usage in 60 seconds (load insensitive) 166 * Note that, as ps(1) mentions, this can let percentages 167 * total over 100% (I've seen 137.9% for 3 processes). 168 * 169 * Note that hardclock updates l_estcpu and l_cpticks independently. 170 * 171 * We wish to decay away 90% of l_estcpu in (5 * loadavg) seconds. 172 * That is, the system wants to compute a value of decay such 173 * that the following for loop: 174 * for (i = 0; i < (5 * loadavg); i++) 175 * l_estcpu *= decay; 176 * will compute 177 * l_estcpu *= 0.1; 178 * for all values of loadavg: 179 * 180 * Mathematically this loop can be expressed by saying: 181 * decay ** (5 * loadavg) ~= .1 182 * 183 * The system computes decay as: 184 * decay = (2 * loadavg) / (2 * loadavg + 1) 185 * 186 * We wish to prove that the system's computation of decay 187 * will always fulfill the equation: 188 * decay ** (5 * loadavg) ~= .1 189 * 190 * If we compute b as: 191 * b = 2 * loadavg 192 * then 193 * decay = b / (b + 1) 194 * 195 * We now need to prove two things: 196 * 1) Given factor ** (5 * loadavg) ~= .1, prove factor == b/(b+1) 197 * 2) Given b/(b+1) ** power ~= .1, prove power == (5 * loadavg) 198 * 199 * Facts: 200 * For x close to zero, exp(x) =~ 1 + x, since 201 * exp(x) = 0! + x**1/1! + x**2/2! + ... . 202 * therefore exp(-1/b) =~ 1 - (1/b) = (b-1)/b. 203 * For x close to zero, ln(1+x) =~ x, since 204 * ln(1+x) = x - x**2/2 + x**3/3 - ... -1 < x < 1 205 * therefore ln(b/(b+1)) = ln(1 - 1/(b+1)) =~ -1/(b+1). 206 * ln(.1) =~ -2.30 207 * 208 * Proof of (1): 209 * Solve (factor)**(power) =~ .1 given power (5*loadav): 210 * solving for factor, 211 * ln(factor) =~ (-2.30/5*loadav), or 212 * factor =~ exp(-1/((5/2.30)*loadav)) =~ exp(-1/(2*loadav)) = 213 * exp(-1/b) =~ (b-1)/b =~ b/(b+1). QED 214 * 215 * Proof of (2): 216 * Solve (factor)**(power) =~ .1 given factor == (b/(b+1)): 217 * solving for power, 218 * power*ln(b/(b+1)) =~ -2.30, or 219 * power =~ 2.3 * (b + 1) = 4.6*loadav + 2.3 =~ 5*loadav. QED 220 * 221 * Actual power values for the implemented algorithm are as follows: 222 * loadav: 1 2 3 4 223 * power: 5.68 10.32 14.94 19.55 224 */ 225 226 /* calculations for digital decay to forget 90% of usage in 5*loadav sec */ 227 #define loadfactor(loadav) (2 * (loadav)) 228 229 fixpt_t 230 decay_cpu(fixpt_t loadfac, fixpt_t estcpu) 231 { 232 233 if (estcpu == 0) { 234 return 0; 235 } 236 237 #if !defined(_LP64) 238 /* avoid 64bit arithmetics. */ 239 #define FIXPT_MAX ((fixpt_t)((UINTMAX_C(1) << sizeof(fixpt_t) * CHAR_BIT) - 1)) 240 if (__predict_true(loadfac <= FIXPT_MAX / ESTCPU_MAX)) { 241 return estcpu * loadfac / (loadfac + FSCALE); 242 } 243 #endif /* !defined(_LP64) */ 244 245 return (uint64_t)estcpu * loadfac / (loadfac + FSCALE); 246 } 247 248 /* 249 * For all load averages >= 1 and max l_estcpu of (255 << ESTCPU_SHIFT), 250 * sleeping for at least seven times the loadfactor will decay l_estcpu to 251 * less than (1 << ESTCPU_SHIFT). 252 * 253 * note that our ESTCPU_MAX is actually much smaller than (255 << ESTCPU_SHIFT). 254 */ 255 static fixpt_t 256 decay_cpu_batch(fixpt_t loadfac, fixpt_t estcpu, unsigned int n) 257 { 258 259 if ((n << FSHIFT) >= 7 * loadfac) { 260 return 0; 261 } 262 263 while (estcpu != 0 && n > 1) { 264 estcpu = decay_cpu(loadfac, estcpu); 265 n--; 266 } 267 268 return estcpu; 269 } 270 271 /* 272 * sched_pstats_hook: 273 * 274 * Periodically called from sched_pstats(); used to recalculate priorities. 275 */ 276 void 277 sched_pstats_hook(struct lwp *l) 278 { 279 fixpt_t loadfac; 280 int sleeptm; 281 282 /* 283 * If the LWP has slept an entire second, stop recalculating 284 * its priority until it wakes up. 285 */ 286 if (l->l_stat == LSSLEEP || l->l_stat == LSSTOP || 287 l->l_stat == LSSUSPENDED) { 288 l->l_slptime++; 289 sleeptm = 1; 290 } else { 291 sleeptm = 0x7fffffff; 292 } 293 294 if (l->l_slptime <= sleeptm) { 295 loadfac = 2 * (averunnable.ldavg[0]); 296 l->l_estcpu = decay_cpu(loadfac, l->l_estcpu); 297 resetpriority(l); 298 } 299 } 300 301 /* 302 * Recalculate the priority of a process after it has slept for a while. 303 */ 304 static void 305 updatepri(struct lwp *l) 306 { 307 fixpt_t loadfac; 308 309 KASSERT(lwp_locked(l, NULL)); 310 KASSERT(l->l_slptime > 1); 311 312 loadfac = loadfactor(averunnable.ldavg[0]); 313 314 l->l_slptime--; /* the first time was done in sched_pstats */ 315 l->l_estcpu = decay_cpu_batch(loadfac, l->l_estcpu, l->l_slptime); 316 resetpriority(l); 317 } 318 319 void 320 sched_rqinit(void) 321 { 322 323 } 324 325 void 326 sched_setrunnable(struct lwp *l) 327 { 328 329 if (l->l_slptime > 1) 330 updatepri(l); 331 } 332 333 void 334 sched_nice(struct proc *p, int n) 335 { 336 struct lwp *l; 337 338 KASSERT(mutex_owned(&p->p_smutex)); 339 340 p->p_nice = n; 341 LIST_FOREACH(l, &p->p_lwps, l_sibling) { 342 lwp_lock(l); 343 resetpriority(l); 344 lwp_unlock(l); 345 } 346 } 347 348 /* 349 * Recompute the priority of an LWP. Arrange to reschedule if 350 * the resulting priority is better than that of the current LWP. 351 */ 352 static void 353 resetpriority(struct lwp *l) 354 { 355 pri_t pri; 356 struct proc *p = l->l_proc; 357 358 KASSERT(lwp_locked(l, NULL)); 359 360 if (l->l_class != SCHED_OTHER) 361 return; 362 363 /* See comments above ESTCPU_SHIFT definition. */ 364 pri = (PRI_KERNEL - 1) - (l->l_estcpu >> ESTCPU_SHIFT) - p->p_nice; 365 pri = imax(pri, 0); 366 if (pri != l->l_priority) 367 lwp_changepri(l, pri); 368 } 369 370 /* 371 * We adjust the priority of the current process. The priority of a process 372 * gets worse as it accumulates CPU time. The CPU usage estimator (l_estcpu) 373 * is increased here. The formula for computing priorities (in kern_synch.c) 374 * will compute a different value each time l_estcpu increases. This can 375 * cause a switch, but unless the priority crosses a PPQ boundary the actual 376 * queue will not change. The CPU usage estimator ramps up quite quickly 377 * when the process is running (linearly), and decays away exponentially, at 378 * a rate which is proportionally slower when the system is busy. The basic 379 * principle is that the system will 90% forget that the process used a lot 380 * of CPU time in 5 * loadav seconds. This causes the system to favor 381 * processes which haven't run much recently, and to round-robin among other 382 * processes. 383 */ 384 385 void 386 sched_schedclock(struct lwp *l) 387 { 388 389 if (l->l_class != SCHED_OTHER) 390 return; 391 392 KASSERT(!CURCPU_IDLE_P()); 393 l->l_estcpu = ESTCPULIM(l->l_estcpu + ESTCPU_ACCUM); 394 lwp_lock(l); 395 resetpriority(l); 396 lwp_unlock(l); 397 } 398 399 /* 400 * sched_proc_fork: 401 * 402 * Inherit the parent's scheduler history. 403 */ 404 void 405 sched_proc_fork(struct proc *parent, struct proc *child) 406 { 407 lwp_t *pl; 408 409 KASSERT(mutex_owned(&parent->p_smutex)); 410 411 pl = LIST_FIRST(&parent->p_lwps); 412 child->p_estcpu_inherited = pl->l_estcpu; 413 child->p_forktime = sched_pstats_ticks; 414 } 415 416 /* 417 * sched_proc_exit: 418 * 419 * Chargeback parents for the sins of their children. 420 */ 421 void 422 sched_proc_exit(struct proc *parent, struct proc *child) 423 { 424 fixpt_t loadfac = loadfactor(averunnable.ldavg[0]); 425 fixpt_t estcpu; 426 lwp_t *pl, *cl; 427 428 /* XXX Only if parent != init?? */ 429 430 mutex_enter(&parent->p_smutex); 431 pl = LIST_FIRST(&parent->p_lwps); 432 cl = LIST_FIRST(&child->p_lwps); 433 estcpu = decay_cpu_batch(loadfac, child->p_estcpu_inherited, 434 sched_pstats_ticks - child->p_forktime); 435 if (cl->l_estcpu > estcpu) { 436 lwp_lock(pl); 437 pl->l_estcpu = ESTCPULIM(pl->l_estcpu + cl->l_estcpu - estcpu); 438 lwp_unlock(pl); 439 } 440 mutex_exit(&parent->p_smutex); 441 } 442 443 void 444 sched_wakeup(struct lwp *l) 445 { 446 447 l->l_cpu = sched_takecpu(l); 448 } 449 450 void 451 sched_slept(struct lwp *l) 452 { 453 454 } 455 456 void 457 sched_lwp_fork(struct lwp *l1, struct lwp *l2) 458 { 459 460 l2->l_estcpu = l1->l_estcpu; 461 } 462 463 void 464 sched_lwp_exit(struct lwp *l) 465 { 466 467 } 468 469 void 470 sched_lwp_collect(struct lwp *t) 471 { 472 lwp_t *l; 473 474 /* Absorb estcpu value of collected LWP. */ 475 l = curlwp; 476 lwp_lock(l); 477 l->l_estcpu += t->l_estcpu; 478 lwp_unlock(l); 479 } 480 481 void 482 sched_oncpu(lwp_t *l) 483 { 484 485 } 486 487 void 488 sched_newts(lwp_t *l) 489 { 490 491 } 492 493 /* 494 * Sysctl nodes and initialization. 495 */ 496 497 static int 498 sysctl_sched_rtts(SYSCTLFN_ARGS) 499 { 500 struct sysctlnode node; 501 int rttsms = hztoms(rrticks); 502 503 node = *rnode; 504 node.sysctl_data = &rttsms; 505 return sysctl_lookup(SYSCTLFN_CALL(&node)); 506 } 507 508 SYSCTL_SETUP(sysctl_sched_4bsd_setup, "sysctl sched setup") 509 { 510 const struct sysctlnode *node = NULL; 511 512 sysctl_createv(clog, 0, NULL, NULL, 513 CTLFLAG_PERMANENT, 514 CTLTYPE_NODE, "kern", NULL, 515 NULL, 0, NULL, 0, 516 CTL_KERN, CTL_EOL); 517 sysctl_createv(clog, 0, NULL, &node, 518 CTLFLAG_PERMANENT, 519 CTLTYPE_NODE, "sched", 520 SYSCTL_DESCR("Scheduler options"), 521 NULL, 0, NULL, 0, 522 CTL_KERN, CTL_CREATE, CTL_EOL); 523 524 if (node == NULL) 525 return; 526 527 rrticks = hz / 10; 528 529 sysctl_createv(NULL, 0, &node, NULL, 530 CTLFLAG_PERMANENT, 531 CTLTYPE_STRING, "name", NULL, 532 NULL, 0, __UNCONST("4.4BSD"), 0, 533 CTL_CREATE, CTL_EOL); 534 sysctl_createv(NULL, 0, &node, NULL, 535 CTLFLAG_PERMANENT, 536 CTLTYPE_INT, "rtts", 537 SYSCTL_DESCR("Round-robin time quantum (in miliseconds)"), 538 sysctl_sched_rtts, 0, NULL, 0, 539 CTL_CREATE, CTL_EOL); 540 } 541
Indexes created Tue Sep 23 07:09:52 GMT 2025