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 8.9 (Berkeley) 5/19/95 39 * $Id: kern_synch.c,v 1.24 1996/09/01 10:30:33 davidg Exp $ 40 */ 41 42 #include "opt_ktrace.h" 43 44 #include <sys/param.h> 45 #include <sys/systm.h> 46 #include <sys/proc.h> 47 #include <sys/kernel.h> 48 #include <sys/buf.h> 49 #include <sys/signalvar.h> 50 #include <sys/resourcevar.h> 51 #include <sys/signalvar.h> 52 #include <sys/vmmeter.h> 53 #include <vm/vm.h> 54 #include <vm/vm_param.h> 55 #include <vm/vm_extern.h> 56 #ifdef KTRACE 57 #include <sys/ktrace.h> 58 #endif 59 60 #include <machine/cpu.h> 61 62 static void rqinit __P((void *)); 63 SYSINIT(runqueue, SI_SUB_RUN_QUEUE, SI_ORDER_FIRST, rqinit, NULL) 64 65 u_char curpriority; /* usrpri of curproc */ 66 int lbolt; /* once a second sleep address */ 67 68 extern void endtsleep __P((void *)); 69 extern void updatepri __P((struct proc *p)); 70 71 /* 72 * Force switch among equal priority processes every 100ms. 73 */ 74 /* ARGSUSED */ 75 void 76 roundrobin(arg) 77 void *arg; 78 { 79 80 need_resched(); 81 timeout(roundrobin, NULL, hz / 10); 82 } 83 84 /* 85 * Constants for digital decay and forget: 86 * 90% of (p_estcpu) usage in 5 * loadav time 87 * 95% of (p_pctcpu) usage in 60 seconds (load insensitive) 88 * Note that, as ps(1) mentions, this can let percentages 89 * total over 100% (I've seen 137.9% for 3 processes). 90 * 91 * Note that statclock updates p_estcpu and p_cpticks independently. 92 * 93 * We wish to decay away 90% of p_estcpu in (5 * loadavg) seconds. 94 * That is, the system wants to compute a value of decay such 95 * that the following for loop: 96 * for (i = 0; i < (5 * loadavg); i++) 97 * p_estcpu *= decay; 98 * will compute 99 * p_estcpu *= 0.1; 100 * for all values of loadavg: 101 * 102 * Mathematically this loop can be expressed by saying: 103 * decay ** (5 * loadavg) ~= .1 104 * 105 * The system computes decay as: 106 * decay = (2 * loadavg) / (2 * loadavg + 1) 107 * 108 * We wish to prove that the system's computation of decay 109 * will always fulfill the equation: 110 * decay ** (5 * loadavg) ~= .1 111 * 112 * If we compute b as: 113 * b = 2 * loadavg 114 * then 115 * decay = b / (b + 1) 116 * 117 * We now need to prove two things: 118 * 1) Given factor ** (5 * loadavg) ~= .1, prove factor == b/(b+1) 119 * 2) Given b/(b+1) ** power ~= .1, prove power == (5 * loadavg) 120 * 121 * Facts: 122 * For x close to zero, exp(x) =~ 1 + x, since 123 * exp(x) = 0! + x**1/1! + x**2/2! + ... . 124 * therefore exp(-1/b) =~ 1 - (1/b) = (b-1)/b. 125 * For x close to zero, ln(1+x) =~ x, since 126 * ln(1+x) = x - x**2/2 + x**3/3 - ... -1 < x < 1 127 * therefore ln(b/(b+1)) = ln(1 - 1/(b+1)) =~ -1/(b+1). 128 * ln(.1) =~ -2.30 129 * 130 * Proof of (1): 131 * Solve (factor)**(power) =~ .1 given power (5*loadav): 132 * solving for factor, 133 * ln(factor) =~ (-2.30/5*loadav), or 134 * factor =~ exp(-1/((5/2.30)*loadav)) =~ exp(-1/(2*loadav)) = 135 * exp(-1/b) =~ (b-1)/b =~ b/(b+1). QED 136 * 137 * Proof of (2): 138 * Solve (factor)**(power) =~ .1 given factor == (b/(b+1)): 139 * solving for power, 140 * power*ln(b/(b+1)) =~ -2.30, or 141 * power =~ 2.3 * (b + 1) = 4.6*loadav + 2.3 =~ 5*loadav. QED 142 * 143 * Actual power values for the implemented algorithm are as follows: 144 * loadav: 1 2 3 4 145 * power: 5.68 10.32 14.94 19.55 146 */ 147 148 /* calculations for digital decay to forget 90% of usage in 5*loadav sec */ 149 #define loadfactor(loadav) (2 * (loadav)) 150 #define decay_cpu(loadfac, cpu) (((loadfac) * (cpu)) / ((loadfac) + FSCALE)) 151 152 /* decay 95% of `p_pctcpu' in 60 seconds; see CCPU_SHIFT before changing */ 153 fixpt_t ccpu = 0.95122942450071400909 * FSCALE; /* exp(-1/20) */ 154 155 /* 156 * If `ccpu' is not equal to `exp(-1/20)' and you still want to use the 157 * faster/more-accurate formula, you'll have to estimate CCPU_SHIFT below 158 * and possibly adjust FSHIFT in "param.h" so that (FSHIFT >= CCPU_SHIFT). 159 * 160 * To estimate CCPU_SHIFT for exp(-1/20), the following formula was used: 161 * 1 - exp(-1/20) ~= 0.0487 ~= 0.0488 == 1 (fixed pt, *11* bits). 162 * 163 * If you dont want to bother with the faster/more-accurate formula, you 164 * can set CCPU_SHIFT to (FSHIFT + 1) which will use a slower/less-accurate 165 * (more general) method of calculating the %age of CPU used by a process. 166 */ 167 #define CCPU_SHIFT 11 168 169 /* 170 * Recompute process priorities, every hz ticks. 171 */ 172 /* ARGSUSED */ 173 void 174 schedcpu(arg) 175 void *arg; 176 { 177 register fixpt_t loadfac = loadfactor(averunnable.ldavg[0]); 178 register struct proc *p; 179 register int s; 180 register unsigned int newcpu; 181 182 wakeup((caddr_t)&lbolt); 183 for (p = allproc.lh_first; p != 0; p = p->p_list.le_next) { 184 /* 185 * Increment time in/out of memory and sleep time 186 * (if sleeping). We ignore overflow; with 16-bit int's 187 * (remember them?) overflow takes 45 days. 188 */ 189 p->p_swtime++; 190 if (p->p_stat == SSLEEP || p->p_stat == SSTOP) 191 p->p_slptime++; 192 p->p_pctcpu = (p->p_pctcpu * ccpu) >> FSHIFT; 193 /* 194 * If the process has slept the entire second, 195 * stop recalculating its priority until it wakes up. 196 */ 197 if (p->p_slptime > 1) 198 continue; 199 s = splhigh(); /* prevent state changes and protect run queue */ 200 /* 201 * p_pctcpu is only for ps. 202 */ 203 #if (FSHIFT >= CCPU_SHIFT) 204 p->p_pctcpu += (hz == 100)? 205 ((fixpt_t) p->p_cpticks) << (FSHIFT - CCPU_SHIFT): 206 100 * (((fixpt_t) p->p_cpticks) 207 << (FSHIFT - CCPU_SHIFT)) / hz; 208 #else 209 p->p_pctcpu += ((FSCALE - ccpu) * 210 (p->p_cpticks * FSCALE / hz)) >> FSHIFT; 211 #endif 212 p->p_cpticks = 0; 213 newcpu = (u_int) decay_cpu(loadfac, p->p_estcpu) + p->p_nice; 214 p->p_estcpu = min(newcpu, UCHAR_MAX); 215 resetpriority(p); 216 if (p->p_priority >= PUSER) { 217 #define PPQ (128 / NQS) /* priorities per queue */ 218 if ((p != curproc) && 219 p->p_stat == SRUN && 220 (p->p_flag & P_INMEM) && 221 (p->p_priority / PPQ) != (p->p_usrpri / PPQ)) { 222 remrq(p); 223 p->p_priority = p->p_usrpri; 224 setrunqueue(p); 225 } else 226 p->p_priority = p->p_usrpri; 227 } 228 splx(s); 229 } 230 vmmeter(); 231 timeout(schedcpu, (void *)0, hz); 232 } 233 234 /* 235 * Recalculate the priority of a process after it has slept for a while. 236 * For all load averages >= 1 and max p_estcpu of 255, sleeping for at 237 * least six times the loadfactor will decay p_estcpu to zero. 238 */ 239 void 240 updatepri(p) 241 register struct proc *p; 242 { 243 register unsigned int newcpu = p->p_estcpu; 244 register fixpt_t loadfac = loadfactor(averunnable.ldavg[0]); 245 246 if (p->p_slptime > 5 * loadfac) 247 p->p_estcpu = 0; 248 else { 249 p->p_slptime--; /* the first time was done in schedcpu */ 250 while (newcpu && --p->p_slptime) 251 newcpu = (int) decay_cpu(loadfac, newcpu); 252 p->p_estcpu = min(newcpu, UCHAR_MAX); 253 } 254 resetpriority(p); 255 } 256 257 /* 258 * We're only looking at 7 bits of the address; everything is 259 * aligned to 4, lots of things are aligned to greater powers 260 * of 2. Shift right by 8, i.e. drop the bottom 256 worth. 261 */ 262 #define TABLESIZE 128 263 TAILQ_HEAD(slpquehead, proc) slpque[TABLESIZE]; 264 #define LOOKUP(x) (((long)(x) >> 8) & (TABLESIZE - 1)) 265 266 /* 267 * During autoconfiguration or after a panic, a sleep will simply 268 * lower the priority briefly to allow interrupts, then return. 269 * The priority to be used (safepri) is machine-dependent, thus this 270 * value is initialized and maintained in the machine-dependent layers. 271 * This priority will typically be 0, or the lowest priority 272 * that is safe for use on the interrupt stack; it can be made 273 * higher to block network software interrupts after panics. 274 */ 275 int safepri; 276 277 void 278 sleepinit() 279 { 280 int i; 281 282 for (i = 0; i < TABLESIZE; i++) 283 TAILQ_INIT(&slpque[i]); 284 } 285 286 /* 287 * General sleep call. Suspends the current process until a wakeup is 288 * performed on the specified identifier. The process will then be made 289 * runnable with the specified priority. Sleeps at most timo/hz seconds 290 * (0 means no timeout). If pri includes PCATCH flag, signals are checked 291 * before and after sleeping, else signals are not checked. Returns 0 if 292 * awakened, EWOULDBLOCK if the timeout expires. If PCATCH is set and a 293 * signal needs to be delivered, ERESTART is returned if the current system 294 * call should be restarted if possible, and EINTR is returned if the system 295 * call should be interrupted by the signal (return EINTR). 296 */ 297 int 298 tsleep(ident, priority, wmesg, timo) 299 void *ident; 300 int priority, timo; 301 char *wmesg; 302 { 303 struct proc *p = curproc; 304 int s, sig, catch = priority & PCATCH; 305 306 #ifdef KTRACE 307 if (KTRPOINT(p, KTR_CSW)) 308 ktrcsw(p->p_tracep, 1, 0); 309 #endif 310 s = splhigh(); 311 if (cold || panicstr) { 312 /* 313 * After a panic, or during autoconfiguration, 314 * just give interrupts a chance, then just return; 315 * don't run any other procs or panic below, 316 * in case this is the idle process and already asleep. 317 */ 318 splx(safepri); 319 splx(s); 320 return (0); 321 } 322 #ifdef DIAGNOSTIC 323 if (ident == NULL || p->p_stat != SRUN) 324 panic("tsleep"); 325 #endif 326 p->p_wchan = ident; 327 p->p_wmesg = wmesg; 328 p->p_slptime = 0; 329 p->p_priority = priority & PRIMASK; 330 TAILQ_INSERT_TAIL(&slpque[LOOKUP(ident)], p, p_procq); 331 if (timo) 332 timeout(endtsleep, (void *)p, timo); 333 /* 334 * We put ourselves on the sleep queue and start our timeout 335 * before calling CURSIG, as we could stop there, and a wakeup 336 * or a SIGCONT (or both) could occur while we were stopped. 337 * A SIGCONT would cause us to be marked as SSLEEP 338 * without resuming us, thus we must be ready for sleep 339 * when CURSIG is called. If the wakeup happens while we're 340 * stopped, p->p_wchan will be 0 upon return from CURSIG. 341 */ 342 if (catch) { 343 p->p_flag |= P_SINTR; 344 if ((sig = CURSIG(p))) { 345 if (p->p_wchan) 346 unsleep(p); 347 p->p_stat = SRUN; 348 goto resume; 349 } 350 if (p->p_wchan == 0) { 351 catch = 0; 352 goto resume; 353 } 354 } else 355 sig = 0; 356 p->p_stat = SSLEEP; 357 p->p_stats->p_ru.ru_nvcsw++; 358 mi_switch(); 359 resume: 360 curpriority = p->p_usrpri; 361 splx(s); 362 p->p_flag &= ~P_SINTR; 363 if (p->p_flag & P_TIMEOUT) { 364 p->p_flag &= ~P_TIMEOUT; 365 if (sig == 0) { 366 #ifdef KTRACE 367 if (KTRPOINT(p, KTR_CSW)) 368 ktrcsw(p->p_tracep, 0, 0); 369 #endif 370 return (EWOULDBLOCK); 371 } 372 } else if (timo) 373 untimeout(endtsleep, (void *)p); 374 if (catch && (sig != 0 || (sig = CURSIG(p)))) { 375 #ifdef KTRACE 376 if (KTRPOINT(p, KTR_CSW)) 377 ktrcsw(p->p_tracep, 0, 0); 378 #endif 379 if (p->p_sigacts->ps_sigintr & sigmask(sig)) 380 return (EINTR); 381 return (ERESTART); 382 } 383 #ifdef KTRACE 384 if (KTRPOINT(p, KTR_CSW)) 385 ktrcsw(p->p_tracep, 0, 0); 386 #endif 387 return (0); 388 } 389 390 /* 391 * Implement timeout for tsleep. 392 * If process hasn't been awakened (wchan non-zero), 393 * set timeout flag and undo the sleep. If proc 394 * is stopped, just unsleep so it will remain stopped. 395 */ 396 void 397 endtsleep(arg) 398 void *arg; 399 { 400 register struct proc *p; 401 int s; 402 403 p = (struct proc *)arg; 404 s = splhigh(); 405 if (p->p_wchan) { 406 if (p->p_stat == SSLEEP) 407 setrunnable(p); 408 else 409 unsleep(p); 410 p->p_flag |= P_TIMEOUT; 411 } 412 splx(s); 413 } 414 415 /* 416 * Remove a process from its wait queue 417 */ 418 void 419 unsleep(p) 420 register struct proc *p; 421 { 422 int s; 423 424 s = splhigh(); 425 if (p->p_wchan) { 426 TAILQ_REMOVE(&slpque[LOOKUP(p->p_wchan)], p, p_procq); 427 p->p_wchan = 0; 428 } 429 splx(s); 430 } 431 432 /* 433 * Make all processes sleeping on the specified identifier runnable. 434 */ 435 void 436 wakeup(ident) 437 register void *ident; 438 { 439 register struct slpquehead *qp; 440 register struct proc *p; 441 int s; 442 443 s = splhigh(); 444 qp = &slpque[LOOKUP(ident)]; 445 restart: 446 for (p = qp->tqh_first; p != NULL; p = p->p_procq.tqe_next) { 447 #ifdef DIAGNOSTIC 448 if (p->p_stat != SSLEEP && p->p_stat != SSTOP) 449 panic("wakeup"); 450 #endif 451 if (p->p_wchan == ident) { 452 TAILQ_REMOVE(qp, p, p_procq); 453 p->p_wchan = 0; 454 if (p->p_stat == SSLEEP) { 455 /* OPTIMIZED EXPANSION OF setrunnable(p); */ 456 if (p->p_slptime > 1) 457 updatepri(p); 458 p->p_slptime = 0; 459 p->p_stat = SRUN; 460 if (p->p_flag & P_INMEM) { 461 setrunqueue(p); 462 need_resched(); 463 } else { 464 wakeup((caddr_t)&proc0); 465 } 466 /* END INLINE EXPANSION */ 467 goto restart; 468 } 469 } 470 } 471 splx(s); 472 } 473 474 /* 475 * Make a process sleeping on the specified identifier runnable. 476 * May wake more than one process if a target prcoess is currently 477 * swapped out. 478 */ 479 void 480 wakeup_one(ident) 481 register void *ident; 482 { 483 register struct slpquehead *qp; 484 register struct proc *p; 485 int s; 486 487 s = splhigh(); 488 qp = &slpque[LOOKUP(ident)]; 489 490 for (p = qp->tqh_first; p != NULL; p = p->p_procq.tqe_next) { 491 #ifdef DIAGNOSTIC 492 if (p->p_stat != SSLEEP && p->p_stat != SSTOP) 493 panic("wakeup_one"); 494 #endif 495 if (p->p_wchan == ident) { 496 TAILQ_REMOVE(qp, p, p_procq); 497 p->p_wchan = 0; 498 if (p->p_stat == SSLEEP) { 499 /* OPTIMIZED EXPANSION OF setrunnable(p); */ 500 if (p->p_slptime > 1) 501 updatepri(p); 502 p->p_slptime = 0; 503 p->p_stat = SRUN; 504 if (p->p_flag & P_INMEM) { 505 setrunqueue(p); 506 need_resched(); 507 break; 508 } else { 509 wakeup((caddr_t)&proc0); 510 } 511 /* END INLINE EXPANSION */ 512 } 513 } 514 } 515 splx(s); 516 } 517 518 /* 519 * The machine independent parts of mi_switch(). 520 * Must be called at splstatclock() or higher. 521 */ 522 void 523 mi_switch() 524 { 525 register struct proc *p = curproc; /* XXX */ 526 register struct rlimit *rlim; 527 register long s, u; 528 struct timeval tv; 529 530 #ifdef DEBUG 531 if (p->p_simple_locks) 532 panic("sleep: holding simple lock"); 533 #endif 534 /* 535 * Compute the amount of time during which the current 536 * process was running, and add that to its total so far. 537 */ 538 microtime(&tv); 539 u = p->p_rtime.tv_usec + (tv.tv_usec - runtime.tv_usec); 540 s = p->p_rtime.tv_sec + (tv.tv_sec - runtime.tv_sec); 541 if (u < 0) { 542 u += 1000000; 543 s--; 544 } else if (u >= 1000000) { 545 u -= 1000000; 546 s++; 547 } 548 p->p_rtime.tv_usec = u; 549 p->p_rtime.tv_sec = s; 550 551 /* 552 * Check if the process exceeds its cpu resource allocation. 553 * If over max, kill it. 554 */ 555 if (p->p_stat != SZOMB) { 556 rlim = &p->p_rlimit[RLIMIT_CPU]; 557 if (s >= rlim->rlim_cur) { 558 if (s >= rlim->rlim_max) 559 killproc(p, "exceeded maximum CPU limit"); 560 else { 561 psignal(p, SIGXCPU); 562 if (rlim->rlim_cur < rlim->rlim_max) 563 rlim->rlim_cur += 5; 564 } 565 } 566 } 567 568 /* 569 * Pick a new current process and record its start time. 570 */ 571 cnt.v_swtch++; 572 cpu_switch(p); 573 microtime(&runtime); 574 } 575 576 /* 577 * Initialize the (doubly-linked) run queues 578 * to be empty. 579 */ 580 /* ARGSUSED*/ 581 static void 582 rqinit(dummy) 583 void *dummy; 584 { 585 register int i; 586 587 for (i = 0; i < NQS; i++) { 588 qs[i].ph_link = qs[i].ph_rlink = (struct proc *)&qs[i]; 589 rtqs[i].ph_link = rtqs[i].ph_rlink = (struct proc *)&rtqs[i]; 590 idqs[i].ph_link = idqs[i].ph_rlink = (struct proc *)&idqs[i]; 591 } 592 } 593 594 /* 595 * Change process state to be runnable, 596 * placing it on the run queue if it is in memory, 597 * and awakening the swapper if it isn't in memory. 598 */ 599 void 600 setrunnable(p) 601 register struct proc *p; 602 { 603 register int s; 604 605 s = splhigh(); 606 switch (p->p_stat) { 607 case 0: 608 case SRUN: 609 case SZOMB: 610 default: 611 panic("setrunnable"); 612 case SSTOP: 613 case SSLEEP: 614 unsleep(p); /* e.g. when sending signals */ 615 break; 616 617 case SIDL: 618 break; 619 } 620 p->p_stat = SRUN; 621 if (p->p_flag & P_INMEM) 622 setrunqueue(p); 623 splx(s); 624 if (p->p_slptime > 1) 625 updatepri(p); 626 p->p_slptime = 0; 627 if ((p->p_flag & P_INMEM) == 0) 628 wakeup((caddr_t)&proc0); 629 else if (p->p_priority < curpriority) 630 need_resched(); 631 } 632 633 /* 634 * Compute the priority of a process when running in user mode. 635 * Arrange to reschedule if the resulting priority is better 636 * than that of the current process. 637 */ 638 void 639 resetpriority(p) 640 register struct proc *p; 641 { 642 register unsigned int newpriority; 643 644 if (p->p_rtprio.type == RTP_PRIO_NORMAL) { 645 newpriority = PUSER + p->p_estcpu / 4 + 2 * p->p_nice; 646 newpriority = min(newpriority, MAXPRI); 647 p->p_usrpri = newpriority; 648 if (newpriority < curpriority) 649 need_resched(); 650 } else { 651 need_resched(); 652 } 653 } 654