kern_synch.c revision 7090
1/*- 2 * Copyright (c) 1982, 1986, 1990, 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 * 3. All advertising materials mentioning features or use of this software 19 * must display the following acknowledgement: 20 * This product includes software developed by the University of 21 * California, Berkeley and its contributors. 22 * 4. Neither the name of the University nor the names of its contributors 23 * may be used to endorse or promote products derived from this software 24 * without specific prior written permission. 25 * 26 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND 27 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 28 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 29 * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE 30 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 31 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 32 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 33 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 34 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 35 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 36 * SUCH DAMAGE. 37 * 38 * @(#)kern_synch.c 8.6 (Berkeley) 1/21/94 39 * $Id: kern_synch.c,v 1.9 1994/12/12 06:04:27 davidg Exp $ 40 */ 41 42#include <sys/param.h> 43#include <sys/systm.h> 44#include <sys/proc.h> 45#include <sys/kernel.h> 46#include <sys/buf.h> 47#include <sys/signalvar.h> 48#include <sys/resourcevar.h> 49#include <sys/signalvar.h> 50#include <vm/vm.h> 51#ifdef KTRACE 52#include <sys/ktrace.h> 53#endif 54 55#include <machine/cpu.h> 56 57u_char curpriority; /* usrpri of curproc */ 58int lbolt; /* once a second sleep address */ 59 60void endtsleep __P((void *)); 61 62/* 63 * Force switch among equal priority processes every 100ms. 64 */ 65/* ARGSUSED */ 66void 67roundrobin(arg) 68 void *arg; 69{ 70 71 need_resched(); 72 timeout(roundrobin, NULL, hz / 10); 73} 74 75/* 76 * Constants for digital decay and forget: 77 * 90% of (p_estcpu) usage in 5 * loadav time 78 * 95% of (p_pctcpu) usage in 60 seconds (load insensitive) 79 * Note that, as ps(1) mentions, this can let percentages 80 * total over 100% (I've seen 137.9% for 3 processes). 81 * 82 * Note that hardclock updates p_estcpu and p_cpticks independently. 83 * 84 * We wish to decay away 90% of p_estcpu in (5 * loadavg) seconds. 85 * That is, the system wants to compute a value of decay such 86 * that the following for loop: 87 * for (i = 0; i < (5 * loadavg); i++) 88 * p_estcpu *= decay; 89 * will compute 90 * p_estcpu *= 0.1; 91 * for all values of loadavg: 92 * 93 * Mathematically this loop can be expressed by saying: 94 * decay ** (5 * loadavg) ~= .1 95 * 96 * The system computes decay as: 97 * decay = (2 * loadavg) / (2 * loadavg + 1) 98 * 99 * We wish to prove that the system's computation of decay 100 * will always fulfill the equation: 101 * decay ** (5 * loadavg) ~= .1 102 * 103 * If we compute b as: 104 * b = 2 * loadavg 105 * then 106 * decay = b / (b + 1) 107 * 108 * We now need to prove two things: 109 * 1) Given factor ** (5 * loadavg) ~= .1, prove factor == b/(b+1) 110 * 2) Given b/(b+1) ** power ~= .1, prove power == (5 * loadavg) 111 * 112 * Facts: 113 * For x close to zero, exp(x) =~ 1 + x, since 114 * exp(x) = 0! + x**1/1! + x**2/2! + ... . 115 * therefore exp(-1/b) =~ 1 - (1/b) = (b-1)/b. 116 * For x close to zero, ln(1+x) =~ x, since 117 * ln(1+x) = x - x**2/2 + x**3/3 - ... -1 < x < 1 118 * therefore ln(b/(b+1)) = ln(1 - 1/(b+1)) =~ -1/(b+1). 119 * ln(.1) =~ -2.30 120 * 121 * Proof of (1): 122 * Solve (factor)**(power) =~ .1 given power (5*loadav): 123 * solving for factor, 124 * ln(factor) =~ (-2.30/5*loadav), or 125 * factor =~ exp(-1/((5/2.30)*loadav)) =~ exp(-1/(2*loadav)) = 126 * exp(-1/b) =~ (b-1)/b =~ b/(b+1). QED 127 * 128 * Proof of (2): 129 * Solve (factor)**(power) =~ .1 given factor == (b/(b+1)): 130 * solving for power, 131 * power*ln(b/(b+1)) =~ -2.30, or 132 * power =~ 2.3 * (b + 1) = 4.6*loadav + 2.3 =~ 5*loadav. QED 133 * 134 * Actual power values for the implemented algorithm are as follows: 135 * loadav: 1 2 3 4 136 * power: 5.68 10.32 14.94 19.55 137 */ 138 139/* calculations for digital decay to forget 90% of usage in 5*loadav sec */ 140#define loadfactor(loadav) (2 * (loadav)) 141#define decay_cpu(loadfac, cpu) (((loadfac) * (cpu)) / ((loadfac) + FSCALE)) 142 143/* decay 95% of `p_pctcpu' in 60 seconds; see CCPU_SHIFT before changing */ 144fixpt_t ccpu = 0.95122942450071400909 * FSCALE; /* exp(-1/20) */ 145 146/* 147 * If `ccpu' is not equal to `exp(-1/20)' and you still want to use the 148 * faster/more-accurate formula, you'll have to estimate CCPU_SHIFT below 149 * and possibly adjust FSHIFT in "param.h" so that (FSHIFT >= CCPU_SHIFT). 150 * 151 * To estimate CCPU_SHIFT for exp(-1/20), the following formula was used: 152 * 1 - exp(-1/20) ~= 0.0487 ~= 0.0488 == 1 (fixed pt, *11* bits). 153 * 154 * If you dont want to bother with the faster/more-accurate formula, you 155 * can set CCPU_SHIFT to (FSHIFT + 1) which will use a slower/less-accurate 156 * (more general) method of calculating the %age of CPU used by a process. 157 */ 158#define CCPU_SHIFT 11 159 160/* 161 * Recompute process priorities, every hz ticks. 162 */ 163/* ARGSUSED */ 164void 165schedcpu(arg) 166 void *arg; 167{ 168 register fixpt_t loadfac = loadfactor(averunnable.ldavg[0]); 169 register struct proc *p; 170 register int s; 171 register unsigned int newcpu; 172 173 wakeup((caddr_t)&lbolt); 174 for (p = (struct proc *)allproc; p != NULL; p = p->p_next) { 175 /* 176 * Increment time in/out of memory and sleep time 177 * (if sleeping). We ignore overflow; with 16-bit int's 178 * (remember them?) overflow takes 45 days. 179 */ 180 p->p_swtime++; 181 if (p->p_stat == SSLEEP || p->p_stat == SSTOP) 182 p->p_slptime++; 183 p->p_pctcpu = (p->p_pctcpu * ccpu) >> FSHIFT; 184 /* 185 * If the process has slept the entire second, 186 * stop recalculating its priority until it wakes up. 187 */ 188 if (p->p_slptime > 1) 189 continue; 190 s = splstatclock(); /* prevent state changes */ 191 /* 192 * p_pctcpu is only for ps. 193 */ 194#if (FSHIFT >= CCPU_SHIFT) 195 p->p_pctcpu += (hz == 100)? 196 ((fixpt_t) p->p_cpticks) << (FSHIFT - CCPU_SHIFT): 197 100 * (((fixpt_t) p->p_cpticks) 198 << (FSHIFT - CCPU_SHIFT)) / hz; 199#else 200 p->p_pctcpu += ((FSCALE - ccpu) * 201 (p->p_cpticks * FSCALE / hz)) >> FSHIFT; 202#endif 203 p->p_cpticks = 0; 204 newcpu = (u_int) decay_cpu(loadfac, p->p_estcpu) + p->p_nice; 205 p->p_estcpu = min(newcpu, UCHAR_MAX); 206 resetpriority(p); 207 if (p->p_priority >= PUSER) { 208#define PPQ (128 / NQS) /* priorities per queue */ 209 if ((p != curproc) && 210 p->p_stat == SRUN && 211 (p->p_flag & P_INMEM) && 212 (p->p_priority / PPQ) != (p->p_usrpri / PPQ)) { 213 remrq(p); 214 p->p_priority = p->p_usrpri; 215 setrunqueue(p); 216 } else 217 p->p_priority = p->p_usrpri; 218 } 219 splx(s); 220 } 221 vmmeter(); 222 timeout(schedcpu, (void *)0, hz); 223} 224 225/* 226 * Recalculate the priority of a process after it has slept for a while. 227 * For all load averages >= 1 and max p_estcpu of 255, sleeping for at 228 * least six times the loadfactor will decay p_estcpu to zero. 229 */ 230void 231updatepri(p) 232 register struct proc *p; 233{ 234 register unsigned int newcpu = p->p_estcpu; 235 register fixpt_t loadfac = loadfactor(averunnable.ldavg[0]); 236 237 if (p->p_slptime > 5 * loadfac) 238 p->p_estcpu = 0; 239 else { 240 p->p_slptime--; /* the first time was done in schedcpu */ 241 while (newcpu && --p->p_slptime) 242 newcpu = (int) decay_cpu(loadfac, newcpu); 243 p->p_estcpu = min(newcpu, UCHAR_MAX); 244 } 245 resetpriority(p); 246} 247 248/* 249 * We're only looking at 7 bits of the address; everything is 250 * aligned to 4, lots of things are aligned to greater powers 251 * of 2. Shift right by 8, i.e. drop the bottom 256 worth. 252 */ 253#define TABLESIZE 128 254#define LOOKUP(x) (((int)(x) >> 8) & (TABLESIZE - 1)) 255struct slpque { 256 struct proc *sq_head; 257 struct proc **sq_tailp; 258} slpque[TABLESIZE]; 259 260/* 261 * During autoconfiguration or after a panic, a sleep will simply 262 * lower the priority briefly to allow interrupts, then return. 263 * The priority to be used (safepri) is machine-dependent, thus this 264 * value is initialized and maintained in the machine-dependent layers. 265 * This priority will typically be 0, or the lowest priority 266 * that is safe for use on the interrupt stack; it can be made 267 * higher to block network software interrupts after panics. 268 */ 269int safepri; 270 271/* 272 * General sleep call. Suspends the current process until a wakeup is 273 * performed on the specified identifier. The process will then be made 274 * runnable with the specified priority. Sleeps at most timo/hz seconds 275 * (0 means no timeout). If pri includes PCATCH flag, signals are checked 276 * before and after sleeping, else signals are not checked. Returns 0 if 277 * awakened, EWOULDBLOCK if the timeout expires. If PCATCH is set and a 278 * signal needs to be delivered, ERESTART is returned if the current system 279 * call should be restarted if possible, and EINTR is returned if the system 280 * call should be interrupted by the signal (return EINTR). 281 */ 282int 283tsleep(ident, priority, wmesg, timo) 284 void *ident; 285 int priority, timo; 286 char *wmesg; 287{ 288 register struct proc *p = curproc; 289 register struct slpque *qp; 290 register s; 291 int sig, catch = priority & PCATCH; 292 293#ifdef KTRACE 294 if (KTRPOINT(p, KTR_CSW)) 295 ktrcsw(p->p_tracep, 1, 0); 296#endif 297 s = splhigh(); 298 if (cold || panicstr) { 299 /* 300 * After a panic, or during autoconfiguration, 301 * just give interrupts a chance, then just return; 302 * don't run any other procs or panic below, 303 * in case this is the idle process and already asleep. 304 */ 305 splx(safepri); 306 splx(s); 307 return (0); 308 } 309#ifdef DIAGNOSTIC 310 if (ident == NULL || p->p_stat != SRUN || p->p_back) 311 panic("tsleep"); 312#endif 313 p->p_wchan = ident; 314 p->p_wmesg = wmesg; 315 p->p_slptime = 0; 316 p->p_priority = priority & PRIMASK; 317 qp = &slpque[LOOKUP(ident)]; 318 if (qp->sq_head == 0) 319 qp->sq_head = p; 320 else 321 *qp->sq_tailp = p; 322 *(qp->sq_tailp = &p->p_forw) = 0; 323 if (timo) 324 timeout(endtsleep, (void *)p, timo); 325 /* 326 * We put ourselves on the sleep queue and start our timeout 327 * before calling CURSIG, as we could stop there, and a wakeup 328 * or a SIGCONT (or both) could occur while we were stopped. 329 * A SIGCONT would cause us to be marked as SSLEEP 330 * without resuming us, thus we must be ready for sleep 331 * when CURSIG is called. If the wakeup happens while we're 332 * stopped, p->p_wchan will be 0 upon return from CURSIG. 333 */ 334 if (catch) { 335 p->p_flag |= P_SINTR; 336 if ((sig = CURSIG(p))) { 337 if (p->p_wchan) 338 unsleep(p); 339 p->p_stat = SRUN; 340 goto resume; 341 } 342 if (p->p_wchan == 0) { 343 catch = 0; 344 goto resume; 345 } 346 } else 347 sig = 0; 348 p->p_stat = SSLEEP; 349 p->p_stats->p_ru.ru_nvcsw++; 350 mi_switch(); 351resume: 352 curpriority = p->p_usrpri; 353 splx(s); 354 p->p_flag &= ~P_SINTR; 355 if (p->p_flag & P_TIMEOUT) { 356 p->p_flag &= ~P_TIMEOUT; 357 if (sig == 0) { 358#ifdef KTRACE 359 if (KTRPOINT(p, KTR_CSW)) 360 ktrcsw(p->p_tracep, 0, 0); 361#endif 362 return (EWOULDBLOCK); 363 } 364 } else if (timo) 365 untimeout(endtsleep, (void *)p); 366 if (catch && (sig != 0 || (sig = CURSIG(p)))) { 367#ifdef KTRACE 368 if (KTRPOINT(p, KTR_CSW)) 369 ktrcsw(p->p_tracep, 0, 0); 370#endif 371 if (p->p_sigacts->ps_sigintr & sigmask(sig)) 372 return (EINTR); 373 return (ERESTART); 374 } 375#ifdef KTRACE 376 if (KTRPOINT(p, KTR_CSW)) 377 ktrcsw(p->p_tracep, 0, 0); 378#endif 379 return (0); 380} 381 382/* 383 * Implement timeout for tsleep. 384 * If process hasn't been awakened (wchan non-zero), 385 * set timeout flag and undo the sleep. If proc 386 * is stopped, just unsleep so it will remain stopped. 387 */ 388void 389endtsleep(arg) 390 void *arg; 391{ 392 register struct proc *p; 393 int s; 394 395 p = (struct proc *)arg; 396 s = splhigh(); 397 if (p->p_wchan) { 398 if (p->p_stat == SSLEEP) 399 setrunnable(p); 400 else 401 unsleep(p); 402 p->p_flag |= P_TIMEOUT; 403 } 404 splx(s); 405} 406 407/* 408 * Short-term, non-interruptable sleep. 409 */ 410void 411sleep(ident, priority) 412 void *ident; 413 int priority; 414{ 415 register struct proc *p = curproc; 416 register struct slpque *qp; 417 register s; 418 419#ifdef DIAGNOSTIC 420 if (priority > PZERO) { 421 printf("sleep called with priority %d > PZERO, wchan: %p\n", 422 priority, ident); 423 panic("old sleep"); 424 } 425#endif 426 s = splhigh(); 427 if (cold || panicstr) { 428 /* 429 * After a panic, or during autoconfiguration, 430 * just give interrupts a chance, then just return; 431 * don't run any other procs or panic below, 432 * in case this is the idle process and already asleep. 433 */ 434 splx(safepri); 435 splx(s); 436 return; 437 } 438#ifdef DIAGNOSTIC 439 if (ident == NULL || p->p_stat != SRUN || p->p_back) 440 panic("sleep"); 441#endif 442 p->p_wchan = ident; 443 p->p_wmesg = NULL; 444 p->p_slptime = 0; 445 p->p_priority = priority; 446 qp = &slpque[LOOKUP(ident)]; 447 if (qp->sq_head == 0) 448 qp->sq_head = p; 449 else 450 *qp->sq_tailp = p; 451 *(qp->sq_tailp = &p->p_forw) = 0; 452 p->p_stat = SSLEEP; 453 p->p_stats->p_ru.ru_nvcsw++; 454#ifdef KTRACE 455 if (KTRPOINT(p, KTR_CSW)) 456 ktrcsw(p->p_tracep, 1, 0); 457#endif 458 mi_switch(); 459#ifdef KTRACE 460 if (KTRPOINT(p, KTR_CSW)) 461 ktrcsw(p->p_tracep, 0, 0); 462#endif 463 curpriority = p->p_usrpri; 464 splx(s); 465} 466 467/* 468 * Remove a process from its wait queue 469 */ 470void 471unsleep(p) 472 register struct proc *p; 473{ 474 register struct slpque *qp; 475 register struct proc **hp; 476 int s; 477 478 s = splhigh(); 479 if (p->p_wchan) { 480 hp = &(qp = &slpque[LOOKUP(p->p_wchan)])->sq_head; 481 while (*hp != p) 482 hp = &(*hp)->p_forw; 483 *hp = p->p_forw; 484 if (qp->sq_tailp == &p->p_forw) 485 qp->sq_tailp = hp; 486 p->p_wchan = 0; 487 } 488 splx(s); 489} 490 491/* 492 * Make all processes sleeping on the specified identifier runnable. 493 */ 494void 495wakeup(ident) 496 register void *ident; 497{ 498 register struct slpque *qp; 499 register struct proc *p, **q; 500 int s; 501 502 s = splhigh(); 503 qp = &slpque[LOOKUP(ident)]; 504restart: 505 for (q = &qp->sq_head; *q; ) { 506 p = *q; 507#ifdef DIAGNOSTIC 508 if (p->p_back || (p->p_stat != SSLEEP && p->p_stat != SSTOP)) 509 panic("wakeup"); 510#endif 511 if (p->p_wchan == ident) { 512 p->p_wchan = 0; 513 *q = p->p_forw; 514 if (qp->sq_tailp == &p->p_forw) 515 qp->sq_tailp = q; 516 if (p->p_stat == SSLEEP) { 517 /* OPTIMIZED EXPANSION OF setrunnable(p); */ 518 if (p->p_slptime > 1) 519 updatepri(p); 520 p->p_slptime = 0; 521 p->p_stat = SRUN; 522 if (p->p_flag & P_INMEM) 523 setrunqueue(p); 524 /* 525 * Since curpriority is a user priority, 526 * p->p_priority is always better than 527 * curpriority. 528 */ 529 if ((p->p_flag & P_INMEM) == 0) 530 wakeup((caddr_t)&proc0); 531 else 532 need_resched(); 533 /* END INLINE EXPANSION */ 534 goto restart; 535 } 536 } else 537 q = &p->p_forw; 538 } 539 splx(s); 540} 541 542/* 543 * The machine independent parts of mi_switch(). 544 * Must be called at splstatclock() or higher. 545 */ 546void 547mi_switch() 548{ 549 register struct proc *p = curproc; /* XXX */ 550 register struct rlimit *rlim; 551 register long s, u; 552 struct timeval tv; 553 554 /* 555 * Compute the amount of time during which the current 556 * process was running, and add that to its total so far. 557 */ 558 microtime(&tv); 559 u = p->p_rtime.tv_usec + (tv.tv_usec - runtime.tv_usec); 560 s = p->p_rtime.tv_sec + (tv.tv_sec - runtime.tv_sec); 561 if (u < 0) { 562 u += 1000000; 563 s--; 564 } else if (u >= 1000000) { 565 u -= 1000000; 566 s++; 567 } 568 p->p_rtime.tv_usec = u; 569 p->p_rtime.tv_sec = s; 570 571 /* 572 * Check if the process exceeds its cpu resource allocation. 573 * If over max, kill it. In any case, if it has run for more 574 * than 10 minutes, reduce priority to give others a chance. 575 */ 576 if (p->p_stat != SZOMB) { 577 rlim = &p->p_rlimit[RLIMIT_CPU]; 578 if (s >= rlim->rlim_cur) { 579 if (s >= rlim->rlim_max) 580 psignal(p, SIGKILL); 581 else { 582 psignal(p, SIGXCPU); 583 if (rlim->rlim_cur < rlim->rlim_max) 584 rlim->rlim_cur += 5; 585 } 586 } 587 if (s > 10 * 60 && p->p_ucred->cr_uid && p->p_nice == NZERO) { 588 p->p_nice = NZERO + 4; 589 resetpriority(p); 590 } 591 } 592 593 /* 594 * Pick a new current process and record its start time. 595 */ 596 cnt.v_swtch++; 597 cpu_switch(p); 598 microtime(&runtime); 599} 600 601/* 602 * Initialize the (doubly-linked) run queues 603 * to be empty. 604 */ 605void 606rqinit() 607{ 608 register int i; 609 610 for (i = 0; i < NQS; i++) { 611 qs[i].ph_link = qs[i].ph_rlink = (struct proc *)&qs[i]; 612 rtqs[i].ph_link = rtqs[i].ph_rlink = (struct proc *)&rtqs[i]; 613 idqs[i].ph_link = idqs[i].ph_rlink = (struct proc *)&idqs[i]; 614 } 615} 616 617/* 618 * Change process state to be runnable, 619 * placing it on the run queue if it is in memory, 620 * and awakening the swapper if it isn't in memory. 621 */ 622void 623setrunnable(p) 624 register struct proc *p; 625{ 626 register int s; 627 628 s = splhigh(); 629 switch (p->p_stat) { 630 case 0: 631 case SRUN: 632 case SZOMB: 633 default: 634 panic("setrunnable"); 635 case SSTOP: 636 case SSLEEP: 637 unsleep(p); /* e.g. when sending signals */ 638 break; 639 640 case SIDL: 641 break; 642 } 643 p->p_stat = SRUN; 644 if (p->p_flag & P_INMEM) 645 setrunqueue(p); 646 splx(s); 647 if (p->p_slptime > 1) 648 updatepri(p); 649 p->p_slptime = 0; 650 if ((p->p_flag & P_INMEM) == 0) 651 wakeup((caddr_t)&proc0); 652 else if (p->p_priority < curpriority) 653 need_resched(); 654} 655 656/* 657 * Compute the priority of a process when running in user mode. 658 * Arrange to reschedule if the resulting priority is better 659 * than that of the current process. 660 */ 661void 662resetpriority(p) 663 register struct proc *p; 664{ 665 register unsigned int newpriority; 666 667 if (p->p_rtprio.type == RTP_PRIO_NORMAL) { 668 newpriority = PUSER + p->p_estcpu / 4 + 2 * p->p_nice; 669 newpriority = min(newpriority, MAXPRI); 670 p->p_usrpri = newpriority; 671 if (newpriority < curpriority) 672 need_resched(); 673 } else { 674 need_resched(); 675 } 676} 677