kern_synch.c revision 3308
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 1.6 1994/10/02 04:45:50 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 60/* 61 * Force switch among equal priority processes every 100ms. 62 */ 63/* ARGSUSED */ 64void 65roundrobin(arg) 66 void *arg; 67{ 68 69 need_resched(); 70 timeout(roundrobin, NULL, hz / 10); 71} 72 73/* 74 * Constants for digital decay and forget: 75 * 90% of (p_estcpu) usage in 5 * loadav time 76 * 95% of (p_pctcpu) usage in 60 seconds (load insensitive) 77 * Note that, as ps(1) mentions, this can let percentages 78 * total over 100% (I've seen 137.9% for 3 processes). 79 * 80 * Note that hardclock updates p_estcpu and p_cpticks independently. 81 * 82 * We wish to decay away 90% of p_estcpu in (5 * loadavg) seconds. 83 * That is, the system wants to compute a value of decay such 84 * that the following for loop: 85 * for (i = 0; i < (5 * loadavg); i++) 86 * p_estcpu *= decay; 87 * will compute 88 * p_estcpu *= 0.1; 89 * for all values of loadavg: 90 * 91 * Mathematically this loop can be expressed by saying: 92 * decay ** (5 * loadavg) ~= .1 93 * 94 * The system computes decay as: 95 * decay = (2 * loadavg) / (2 * loadavg + 1) 96 * 97 * We wish to prove that the system's computation of decay 98 * will always fulfill the equation: 99 * decay ** (5 * loadavg) ~= .1 100 * 101 * If we compute b as: 102 * b = 2 * loadavg 103 * then 104 * decay = b / (b + 1) 105 * 106 * We now need to prove two things: 107 * 1) Given factor ** (5 * loadavg) ~= .1, prove factor == b/(b+1) 108 * 2) Given b/(b+1) ** power ~= .1, prove power == (5 * loadavg) 109 * 110 * Facts: 111 * For x close to zero, exp(x) =~ 1 + x, since 112 * exp(x) = 0! + x**1/1! + x**2/2! + ... . 113 * therefore exp(-1/b) =~ 1 - (1/b) = (b-1)/b. 114 * For x close to zero, ln(1+x) =~ x, since 115 * ln(1+x) = x - x**2/2 + x**3/3 - ... -1 < x < 1 116 * therefore ln(b/(b+1)) = ln(1 - 1/(b+1)) =~ -1/(b+1). 117 * ln(.1) =~ -2.30 118 * 119 * Proof of (1): 120 * Solve (factor)**(power) =~ .1 given power (5*loadav): 121 * solving for factor, 122 * ln(factor) =~ (-2.30/5*loadav), or 123 * factor =~ exp(-1/((5/2.30)*loadav)) =~ exp(-1/(2*loadav)) = 124 * exp(-1/b) =~ (b-1)/b =~ b/(b+1). QED 125 * 126 * Proof of (2): 127 * Solve (factor)**(power) =~ .1 given factor == (b/(b+1)): 128 * solving for power, 129 * power*ln(b/(b+1)) =~ -2.30, or 130 * power =~ 2.3 * (b + 1) = 4.6*loadav + 2.3 =~ 5*loadav. QED 131 * 132 * Actual power values for the implemented algorithm are as follows: 133 * loadav: 1 2 3 4 134 * power: 5.68 10.32 14.94 19.55 135 */ 136 137/* calculations for digital decay to forget 90% of usage in 5*loadav sec */ 138#define loadfactor(loadav) (2 * (loadav)) 139#define decay_cpu(loadfac, cpu) (((loadfac) * (cpu)) / ((loadfac) + FSCALE)) 140 141/* decay 95% of `p_pctcpu' in 60 seconds; see CCPU_SHIFT before changing */ 142fixpt_t ccpu = 0.95122942450071400909 * FSCALE; /* exp(-1/20) */ 143 144/* 145 * If `ccpu' is not equal to `exp(-1/20)' and you still want to use the 146 * faster/more-accurate formula, you'll have to estimate CCPU_SHIFT below 147 * and possibly adjust FSHIFT in "param.h" so that (FSHIFT >= CCPU_SHIFT). 148 * 149 * To estimate CCPU_SHIFT for exp(-1/20), the following formula was used: 150 * 1 - exp(-1/20) ~= 0.0487 ~= 0.0488 == 1 (fixed pt, *11* bits). 151 * 152 * If you dont want to bother with the faster/more-accurate formula, you 153 * can set CCPU_SHIFT to (FSHIFT + 1) which will use a slower/less-accurate 154 * (more general) method of calculating the %age of CPU used by a process. 155 */ 156#define CCPU_SHIFT 11 157 158/* 159 * Recompute process priorities, every hz ticks. 160 */ 161/* ARGSUSED */ 162void 163schedcpu(arg) 164 void *arg; 165{ 166 register fixpt_t loadfac = loadfactor(averunnable.ldavg[0]); 167 register struct proc *p; 168 register int s; 169 register unsigned int newcpu; 170 171 wakeup((caddr_t)&lbolt); 172 for (p = (struct proc *)allproc; p != NULL; p = p->p_next) { 173 /* 174 * Increment time in/out of memory and sleep time 175 * (if sleeping). We ignore overflow; with 16-bit int's 176 * (remember them?) overflow takes 45 days. 177 */ 178 p->p_swtime++; 179 if (p->p_stat == SSLEEP || p->p_stat == SSTOP) 180 p->p_slptime++; 181 p->p_pctcpu = (p->p_pctcpu * ccpu) >> FSHIFT; 182 /* 183 * If the process has slept the entire second, 184 * stop recalculating its priority until it wakes up. 185 */ 186 if (p->p_slptime > 1) 187 continue; 188 s = splstatclock(); /* prevent state changes */ 189 /* 190 * p_pctcpu is only for ps. 191 */ 192#if (FSHIFT >= CCPU_SHIFT) 193 p->p_pctcpu += (hz == 100)? 194 ((fixpt_t) p->p_cpticks) << (FSHIFT - CCPU_SHIFT): 195 100 * (((fixpt_t) p->p_cpticks) 196 << (FSHIFT - CCPU_SHIFT)) / hz; 197#else 198 p->p_pctcpu += ((FSCALE - ccpu) * 199 (p->p_cpticks * FSCALE / hz)) >> FSHIFT; 200#endif 201 p->p_cpticks = 0; 202 newcpu = (u_int) decay_cpu(loadfac, p->p_estcpu) + p->p_nice; 203 p->p_estcpu = min(newcpu, UCHAR_MAX); 204 resetpriority(p); 205 if (p->p_priority >= PUSER) { 206#define PPQ (128 / NQS) /* priorities per queue */ 207 if ((p != curproc) && 208 p->p_stat == SRUN && 209 (p->p_flag & P_INMEM) && 210 (p->p_priority / PPQ) != (p->p_usrpri / PPQ)) { 211 remrq(p); 212 p->p_priority = p->p_usrpri; 213 setrunqueue(p); 214 } else 215 p->p_priority = p->p_usrpri; 216 } 217 splx(s); 218 } 219 vmmeter(); 220 if (bclnlist != NULL) 221 wakeup((caddr_t)pageproc); 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 extern int cold; 293 void endtsleep __P((void *)); 294 295#ifdef KTRACE 296 if (KTRPOINT(p, KTR_CSW)) 297 ktrcsw(p->p_tracep, 1, 0); 298#endif 299 s = splhigh(); 300 if (cold || panicstr) { 301 /* 302 * After a panic, or during autoconfiguration, 303 * just give interrupts a chance, then just return; 304 * don't run any other procs or panic below, 305 * in case this is the idle process and already asleep. 306 */ 307 splx(safepri); 308 splx(s); 309 return (0); 310 } 311#ifdef DIAGNOSTIC 312 if (ident == NULL || p->p_stat != SRUN || p->p_back) 313 panic("tsleep"); 314#endif 315 p->p_wchan = ident; 316 p->p_wmesg = wmesg; 317 p->p_slptime = 0; 318 p->p_priority = priority & PRIMASK; 319 qp = &slpque[LOOKUP(ident)]; 320 if (qp->sq_head == 0) 321 qp->sq_head = p; 322 else 323 *qp->sq_tailp = p; 324 *(qp->sq_tailp = &p->p_forw) = 0; 325 if (timo) 326 timeout(endtsleep, (void *)p, timo); 327 /* 328 * We put ourselves on the sleep queue and start our timeout 329 * before calling CURSIG, as we could stop there, and a wakeup 330 * or a SIGCONT (or both) could occur while we were stopped. 331 * A SIGCONT would cause us to be marked as SSLEEP 332 * without resuming us, thus we must be ready for sleep 333 * when CURSIG is called. If the wakeup happens while we're 334 * stopped, p->p_wchan will be 0 upon return from CURSIG. 335 */ 336 if (catch) { 337 p->p_flag |= P_SINTR; 338 if ((sig = CURSIG(p))) { 339 if (p->p_wchan) 340 unsleep(p); 341 p->p_stat = SRUN; 342 goto resume; 343 } 344 if (p->p_wchan == 0) { 345 catch = 0; 346 goto resume; 347 } 348 } else 349 sig = 0; 350 p->p_stat = SSLEEP; 351 p->p_stats->p_ru.ru_nvcsw++; 352 mi_switch(); 353resume: 354 curpriority = p->p_usrpri; 355 splx(s); 356 p->p_flag &= ~P_SINTR; 357 if (p->p_flag & P_TIMEOUT) { 358 p->p_flag &= ~P_TIMEOUT; 359 if (sig == 0) { 360#ifdef KTRACE 361 if (KTRPOINT(p, KTR_CSW)) 362 ktrcsw(p->p_tracep, 0, 0); 363#endif 364 return (EWOULDBLOCK); 365 } 366 } else if (timo) 367 untimeout(endtsleep, (void *)p); 368 if (catch && (sig != 0 || (sig = CURSIG(p)))) { 369#ifdef KTRACE 370 if (KTRPOINT(p, KTR_CSW)) 371 ktrcsw(p->p_tracep, 0, 0); 372#endif 373 if (p->p_sigacts->ps_sigintr & sigmask(sig)) 374 return (EINTR); 375 return (ERESTART); 376 } 377#ifdef KTRACE 378 if (KTRPOINT(p, KTR_CSW)) 379 ktrcsw(p->p_tracep, 0, 0); 380#endif 381 return (0); 382} 383 384/* 385 * Implement timeout for tsleep. 386 * If process hasn't been awakened (wchan non-zero), 387 * set timeout flag and undo the sleep. If proc 388 * is stopped, just unsleep so it will remain stopped. 389 */ 390void 391endtsleep(arg) 392 void *arg; 393{ 394 register struct proc *p; 395 int s; 396 397 p = (struct proc *)arg; 398 s = splhigh(); 399 if (p->p_wchan) { 400 if (p->p_stat == SSLEEP) 401 setrunnable(p); 402 else 403 unsleep(p); 404 p->p_flag |= P_TIMEOUT; 405 } 406 splx(s); 407} 408 409/* 410 * Short-term, non-interruptable sleep. 411 */ 412void 413sleep(ident, priority) 414 void *ident; 415 int priority; 416{ 417 register struct proc *p = curproc; 418 register struct slpque *qp; 419 register s; 420 extern int cold; 421 422#ifdef DIAGNOSTIC 423 if (priority > PZERO) { 424 printf("sleep called with priority %d > PZERO, wchan: %p\n", 425 priority, ident); 426 panic("old sleep"); 427 } 428#endif 429 s = splhigh(); 430 if (cold || panicstr) { 431 /* 432 * After a panic, or during autoconfiguration, 433 * just give interrupts a chance, then just return; 434 * don't run any other procs or panic below, 435 * in case this is the idle process and already asleep. 436 */ 437 splx(safepri); 438 splx(s); 439 return; 440 } 441#ifdef DIAGNOSTIC 442 if (ident == NULL || p->p_stat != SRUN || p->p_back) 443 panic("sleep"); 444#endif 445 p->p_wchan = ident; 446 p->p_wmesg = NULL; 447 p->p_slptime = 0; 448 p->p_priority = priority; 449 qp = &slpque[LOOKUP(ident)]; 450 if (qp->sq_head == 0) 451 qp->sq_head = p; 452 else 453 *qp->sq_tailp = p; 454 *(qp->sq_tailp = &p->p_forw) = 0; 455 p->p_stat = SSLEEP; 456 p->p_stats->p_ru.ru_nvcsw++; 457#ifdef KTRACE 458 if (KTRPOINT(p, KTR_CSW)) 459 ktrcsw(p->p_tracep, 1, 0); 460#endif 461 mi_switch(); 462#ifdef KTRACE 463 if (KTRPOINT(p, KTR_CSW)) 464 ktrcsw(p->p_tracep, 0, 0); 465#endif 466 curpriority = p->p_usrpri; 467 splx(s); 468} 469 470/* 471 * Remove a process from its wait queue 472 */ 473void 474unsleep(p) 475 register struct proc *p; 476{ 477 register struct slpque *qp; 478 register struct proc **hp; 479 int s; 480 481 s = splhigh(); 482 if (p->p_wchan) { 483 hp = &(qp = &slpque[LOOKUP(p->p_wchan)])->sq_head; 484 while (*hp != p) 485 hp = &(*hp)->p_forw; 486 *hp = p->p_forw; 487 if (qp->sq_tailp == &p->p_forw) 488 qp->sq_tailp = hp; 489 p->p_wchan = 0; 490 } 491 splx(s); 492} 493 494/* 495 * Make all processes sleeping on the specified identifier runnable. 496 */ 497void 498wakeup(ident) 499 register void *ident; 500{ 501 register struct slpque *qp; 502 register struct proc *p, **q; 503 int s; 504 505 s = splhigh(); 506 qp = &slpque[LOOKUP(ident)]; 507restart: 508 for (q = &qp->sq_head; *q; ) { 509 p = *q; 510#ifdef DIAGNOSTIC 511 if (p->p_back || (p->p_stat != SSLEEP && p->p_stat != SSTOP)) 512 panic("wakeup"); 513#endif 514 if (p->p_wchan == ident) { 515 p->p_wchan = 0; 516 *q = p->p_forw; 517 if (qp->sq_tailp == &p->p_forw) 518 qp->sq_tailp = q; 519 if (p->p_stat == SSLEEP) { 520 /* OPTIMIZED EXPANSION OF setrunnable(p); */ 521 if (p->p_slptime > 1) 522 updatepri(p); 523 p->p_slptime = 0; 524 p->p_stat = SRUN; 525 if (p->p_flag & P_INMEM) 526 setrunqueue(p); 527 /* 528 * Since curpriority is a user priority, 529 * p->p_priority is always better than 530 * curpriority. 531 */ 532 if ((p->p_flag & P_INMEM) == 0) 533 wakeup((caddr_t)&proc0); 534 else 535 need_resched(); 536 /* END INLINE EXPANSION */ 537 goto restart; 538 } 539 } else 540 q = &p->p_forw; 541 } 542 splx(s); 543} 544 545/* 546 * The machine independent parts of mi_switch(). 547 * Must be called at splstatclock() or higher. 548 */ 549void 550mi_switch() 551{ 552 register struct proc *p = curproc; /* XXX */ 553 register struct rlimit *rlim; 554 register long s, u; 555 struct timeval tv; 556 557 /* 558 * Compute the amount of time during which the current 559 * process was running, and add that to its total so far. 560 */ 561 microtime(&tv); 562 u = p->p_rtime.tv_usec + (tv.tv_usec - runtime.tv_usec); 563 s = p->p_rtime.tv_sec + (tv.tv_sec - runtime.tv_sec); 564 if (u < 0) { 565 u += 1000000; 566 s--; 567 } else if (u >= 1000000) { 568 u -= 1000000; 569 s++; 570 } 571 p->p_rtime.tv_usec = u; 572 p->p_rtime.tv_sec = s; 573 574 /* 575 * Check if the process exceeds its cpu resource allocation. 576 * If over max, kill it. In any case, if it has run for more 577 * than 10 minutes, reduce priority to give others a chance. 578 */ 579 rlim = &p->p_rlimit[RLIMIT_CPU]; 580 if (s >= rlim->rlim_cur) { 581 if (s >= rlim->rlim_max) 582 psignal(p, SIGKILL); 583 else { 584 psignal(p, SIGXCPU); 585 if (rlim->rlim_cur < rlim->rlim_max) 586 rlim->rlim_cur += 5; 587 } 588 } 589 if (s > 10 * 60 && p->p_ucred->cr_uid && p->p_nice == NZERO) { 590 p->p_nice = NZERO + 4; 591 resetpriority(p); 592 } 593 594 /* 595 * Pick a new current process and record its start time. 596 */ 597 cnt.v_swtch++; 598 cpu_switch(p); 599 microtime(&runtime); 600} 601 602/* 603 * Initialize the (doubly-linked) run queues 604 * to be empty. 605 */ 606void 607rqinit() 608{ 609 register int i; 610 611 for (i = 0; i < NQS; i++) { 612 qs[i].ph_link = qs[i].ph_rlink = (struct proc *)&qs[i]; 613 rtqs[i].ph_link = rtqs[i].ph_rlink = (struct proc *)&rtqs[i]; 614 idqs[i].ph_link = idqs[i].ph_rlink = (struct proc *)&idqs[i]; 615 } 616} 617 618/* 619 * Change process state to be runnable, 620 * placing it on the run queue if it is in memory, 621 * and awakening the swapper if it isn't in memory. 622 */ 623void 624setrunnable(p) 625 register struct proc *p; 626{ 627 register int s; 628 629 s = splhigh(); 630 switch (p->p_stat) { 631 case 0: 632 case SRUN: 633 case SZOMB: 634 default: 635 panic("setrunnable"); 636 case SSTOP: 637 case SSLEEP: 638 unsleep(p); /* e.g. when sending signals */ 639 break; 640 641 case SIDL: 642 break; 643 } 644 p->p_stat = SRUN; 645 if (p->p_flag & P_INMEM) 646 setrunqueue(p); 647 splx(s); 648 if (p->p_slptime > 1) 649 updatepri(p); 650 p->p_slptime = 0; 651 if ((p->p_flag & P_INMEM) == 0) 652 wakeup((caddr_t)&proc0); 653 else if (p->p_priority < curpriority) 654 need_resched(); 655} 656 657/* 658 * Compute the priority of a process when running in user mode. 659 * Arrange to reschedule if the resulting priority is better 660 * than that of the current process. 661 */ 662void 663resetpriority(p) 664 register struct proc *p; 665{ 666 register unsigned int newpriority; 667 668 if (p->p_rtprio.type == RTP_PRIO_NORMAL) { 669 newpriority = PUSER + p->p_estcpu / 4 + 2 * p->p_nice; 670 newpriority = min(newpriority, MAXPRI); 671 p->p_usrpri = newpriority; 672 if (newpriority < curpriority) 673 need_resched(); 674 } else { 675 need_resched(); 676 } 677} 678