kern_synch.c revision 57701
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 * $FreeBSD: head/sys/kern/kern_synch.c 57701 2000-03-02 16:20:07Z dufault $ 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/signalvar.h> 49#include <sys/resourcevar.h> 50#include <sys/vmmeter.h> 51#include <sys/sysctl.h> 52#include <vm/vm.h> 53#include <vm/vm_extern.h> 54#ifdef KTRACE 55#include <sys/uio.h> 56#include <sys/ktrace.h> 57#endif 58 59#include <machine/cpu.h> 60#ifdef SMP 61#include <machine/smp.h> 62#endif 63 64static void sched_setup __P((void *dummy)); 65SYSINIT(sched_setup, SI_SUB_KICK_SCHEDULER, SI_ORDER_FIRST, sched_setup, NULL) 66 67u_char curpriority; 68u_char currtpriority; 69int hogticks; 70int lbolt; 71int sched_quantum; /* Roundrobin scheduling quantum in ticks. */ 72 73static int curpriority_cmp __P((struct proc *p)); 74static void endtsleep __P((void *)); 75static void maybe_resched __P((struct proc *chk)); 76static void roundrobin __P((void *arg)); 77static void schedcpu __P((void *arg)); 78static void updatepri __P((struct proc *p)); 79 80static int 81sysctl_kern_quantum SYSCTL_HANDLER_ARGS 82{ 83 int error, new_val; 84 85 new_val = sched_quantum * tick; 86 error = sysctl_handle_int(oidp, &new_val, 0, req); 87 if (error != 0 || req->newptr == NULL) 88 return (error); 89 if (new_val < tick) 90 return (EINVAL); 91 sched_quantum = new_val / tick; 92 hogticks = 2 * sched_quantum; 93 return (0); 94} 95 96SYSCTL_PROC(_kern, OID_AUTO, quantum, CTLTYPE_INT|CTLFLAG_RW, 97 0, sizeof sched_quantum, sysctl_kern_quantum, "I", ""); 98 99/*- 100 * Compare priorities. Return: 101 * <0: priority of p < current priority 102 * 0: priority of p == current priority 103 * >0: priority of p > current priority 104 * The priorities are the normal priorities or the normal realtime priorities 105 * if p is on the same scheduler as curproc. Otherwise the process on the 106 * more realtimeish scheduler has lowest priority. As usual, a higher 107 * priority really means a lower priority. 108 */ 109static int 110curpriority_cmp(p) 111 struct proc *p; 112{ 113 int c_class, p_class; 114 115 if (curproc->p_rtprio.prio != currtpriority) 116 Debugger("currtprio"); 117 c_class = RTP_PRIO_BASE(curproc->p_rtprio.type); 118 p_class = RTP_PRIO_BASE(p->p_rtprio.type); 119 if (p_class != c_class) 120 return (p_class - c_class); 121 if (p_class == RTP_PRIO_NORMAL) 122 return (((int)p->p_priority - (int)curpriority) / PPQ); 123 return ((int)p->p_rtprio.prio - (int)currtpriority); 124} 125 126/* 127 * Arrange to reschedule if necessary, taking the priorities and 128 * schedulers into account. 129 */ 130static void 131maybe_resched(chk) 132 struct proc *chk; 133{ 134 struct proc *p = curproc; /* XXX */ 135 136 /* 137 * XXX idle scheduler still broken because proccess stays on idle 138 * scheduler during waits (such as when getting FS locks). If a 139 * standard process becomes runaway cpu-bound, the system can lockup 140 * due to idle-scheduler processes in wakeup never getting any cpu. 141 */ 142 if (p == NULL) { 143#if 0 144 need_resched(); 145#endif 146 } else if (chk == p) { 147 /* We may need to yield if our priority has been raised. */ 148 if (curpriority_cmp(chk) > 0) 149 need_resched(); 150 } else if (curpriority_cmp(chk) < 0) 151 need_resched(); 152} 153 154int 155roundrobin_interval(void) 156{ 157 return (sched_quantum); 158} 159 160/* 161 * Force switch among equal priority processes every 100ms. 162 */ 163/* ARGSUSED */ 164static void 165roundrobin(arg) 166 void *arg; 167{ 168#ifndef SMP 169 struct proc *p = curproc; /* XXX */ 170#endif 171 172#ifdef SMP 173 need_resched(); 174 forward_roundrobin(); 175#else 176 if (p == 0 || RTP_PRIO_NEED_RR(p->p_rtprio.type)) 177 need_resched(); 178#endif 179 180 timeout(roundrobin, NULL, sched_quantum); 181} 182 183/* 184 * Constants for digital decay and forget: 185 * 90% of (p_estcpu) usage in 5 * loadav time 186 * 95% of (p_pctcpu) usage in 60 seconds (load insensitive) 187 * Note that, as ps(1) mentions, this can let percentages 188 * total over 100% (I've seen 137.9% for 3 processes). 189 * 190 * Note that schedclock() updates p_estcpu and p_cpticks asynchronously. 191 * 192 * We wish to decay away 90% of p_estcpu in (5 * loadavg) seconds. 193 * That is, the system wants to compute a value of decay such 194 * that the following for loop: 195 * for (i = 0; i < (5 * loadavg); i++) 196 * p_estcpu *= decay; 197 * will compute 198 * p_estcpu *= 0.1; 199 * for all values of loadavg: 200 * 201 * Mathematically this loop can be expressed by saying: 202 * decay ** (5 * loadavg) ~= .1 203 * 204 * The system computes decay as: 205 * decay = (2 * loadavg) / (2 * loadavg + 1) 206 * 207 * We wish to prove that the system's computation of decay 208 * will always fulfill the equation: 209 * decay ** (5 * loadavg) ~= .1 210 * 211 * If we compute b as: 212 * b = 2 * loadavg 213 * then 214 * decay = b / (b + 1) 215 * 216 * We now need to prove two things: 217 * 1) Given factor ** (5 * loadavg) ~= .1, prove factor == b/(b+1) 218 * 2) Given b/(b+1) ** power ~= .1, prove power == (5 * loadavg) 219 * 220 * Facts: 221 * For x close to zero, exp(x) =~ 1 + x, since 222 * exp(x) = 0! + x**1/1! + x**2/2! + ... . 223 * therefore exp(-1/b) =~ 1 - (1/b) = (b-1)/b. 224 * For x close to zero, ln(1+x) =~ x, since 225 * ln(1+x) = x - x**2/2 + x**3/3 - ... -1 < x < 1 226 * therefore ln(b/(b+1)) = ln(1 - 1/(b+1)) =~ -1/(b+1). 227 * ln(.1) =~ -2.30 228 * 229 * Proof of (1): 230 * Solve (factor)**(power) =~ .1 given power (5*loadav): 231 * solving for factor, 232 * ln(factor) =~ (-2.30/5*loadav), or 233 * factor =~ exp(-1/((5/2.30)*loadav)) =~ exp(-1/(2*loadav)) = 234 * exp(-1/b) =~ (b-1)/b =~ b/(b+1). QED 235 * 236 * Proof of (2): 237 * Solve (factor)**(power) =~ .1 given factor == (b/(b+1)): 238 * solving for power, 239 * power*ln(b/(b+1)) =~ -2.30, or 240 * power =~ 2.3 * (b + 1) = 4.6*loadav + 2.3 =~ 5*loadav. QED 241 * 242 * Actual power values for the implemented algorithm are as follows: 243 * loadav: 1 2 3 4 244 * power: 5.68 10.32 14.94 19.55 245 */ 246 247/* calculations for digital decay to forget 90% of usage in 5*loadav sec */ 248#define loadfactor(loadav) (2 * (loadav)) 249#define decay_cpu(loadfac, cpu) (((loadfac) * (cpu)) / ((loadfac) + FSCALE)) 250 251/* decay 95% of `p_pctcpu' in 60 seconds; see CCPU_SHIFT before changing */ 252static fixpt_t ccpu = 0.95122942450071400909 * FSCALE; /* exp(-1/20) */ 253SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, ""); 254 255/* kernel uses `FSCALE', userland (SHOULD) use kern.fscale */ 256static int fscale __unused = FSCALE; 257SYSCTL_INT(_kern, OID_AUTO, fscale, CTLFLAG_RD, 0, FSCALE, ""); 258 259/* 260 * If `ccpu' is not equal to `exp(-1/20)' and you still want to use the 261 * faster/more-accurate formula, you'll have to estimate CCPU_SHIFT below 262 * and possibly adjust FSHIFT in "param.h" so that (FSHIFT >= CCPU_SHIFT). 263 * 264 * To estimate CCPU_SHIFT for exp(-1/20), the following formula was used: 265 * 1 - exp(-1/20) ~= 0.0487 ~= 0.0488 == 1 (fixed pt, *11* bits). 266 * 267 * If you don't want to bother with the faster/more-accurate formula, you 268 * can set CCPU_SHIFT to (FSHIFT + 1) which will use a slower/less-accurate 269 * (more general) method of calculating the %age of CPU used by a process. 270 */ 271#define CCPU_SHIFT 11 272 273/* 274 * Recompute process priorities, every hz ticks. 275 */ 276/* ARGSUSED */ 277static void 278schedcpu(arg) 279 void *arg; 280{ 281 register fixpt_t loadfac = loadfactor(averunnable.ldavg[0]); 282 register struct proc *p; 283 register int realstathz, s; 284 285 realstathz = stathz ? stathz : hz; 286 LIST_FOREACH(p, &allproc, p_list) { 287 /* 288 * Increment time in/out of memory and sleep time 289 * (if sleeping). We ignore overflow; with 16-bit int's 290 * (remember them?) overflow takes 45 days. 291 */ 292 p->p_swtime++; 293 if (p->p_stat == SSLEEP || p->p_stat == SSTOP) 294 p->p_slptime++; 295 p->p_pctcpu = (p->p_pctcpu * ccpu) >> FSHIFT; 296 /* 297 * If the process has slept the entire second, 298 * stop recalculating its priority until it wakes up. 299 */ 300 if (p->p_slptime > 1) 301 continue; 302 s = splhigh(); /* prevent state changes and protect run queue */ 303 /* 304 * p_pctcpu is only for ps. 305 */ 306#if (FSHIFT >= CCPU_SHIFT) 307 p->p_pctcpu += (realstathz == 100)? 308 ((fixpt_t) p->p_cpticks) << (FSHIFT - CCPU_SHIFT): 309 100 * (((fixpt_t) p->p_cpticks) 310 << (FSHIFT - CCPU_SHIFT)) / realstathz; 311#else 312 p->p_pctcpu += ((FSCALE - ccpu) * 313 (p->p_cpticks * FSCALE / realstathz)) >> FSHIFT; 314#endif 315 p->p_cpticks = 0; 316 p->p_estcpu = decay_cpu(loadfac, p->p_estcpu); 317 resetpriority(p); 318 if (p->p_priority >= PUSER) { 319 if ((p != curproc) && 320#ifdef SMP 321 p->p_oncpu == 0xff && /* idle */ 322#endif 323 p->p_stat == SRUN && 324 (p->p_flag & P_INMEM) && 325 (p->p_priority / PPQ) != (p->p_usrpri / PPQ)) { 326 remrunqueue(p); 327 p->p_priority = p->p_usrpri; 328 setrunqueue(p); 329 } else 330 p->p_priority = p->p_usrpri; 331 } 332 splx(s); 333 } 334 vmmeter(); 335 wakeup((caddr_t)&lbolt); 336 timeout(schedcpu, (void *)0, hz); 337} 338 339/* 340 * Recalculate the priority of a process after it has slept for a while. 341 * For all load averages >= 1 and max p_estcpu of 255, sleeping for at 342 * least six times the loadfactor will decay p_estcpu to zero. 343 */ 344static void 345updatepri(p) 346 register struct proc *p; 347{ 348 register unsigned int newcpu = p->p_estcpu; 349 register fixpt_t loadfac = loadfactor(averunnable.ldavg[0]); 350 351 if (p->p_slptime > 5 * loadfac) 352 p->p_estcpu = 0; 353 else { 354 p->p_slptime--; /* the first time was done in schedcpu */ 355 while (newcpu && --p->p_slptime) 356 newcpu = decay_cpu(loadfac, newcpu); 357 p->p_estcpu = newcpu; 358 } 359 resetpriority(p); 360} 361 362/* 363 * We're only looking at 7 bits of the address; everything is 364 * aligned to 4, lots of things are aligned to greater powers 365 * of 2. Shift right by 8, i.e. drop the bottom 256 worth. 366 */ 367#define TABLESIZE 128 368static TAILQ_HEAD(slpquehead, proc) slpque[TABLESIZE]; 369#define LOOKUP(x) (((intptr_t)(x) >> 8) & (TABLESIZE - 1)) 370 371/* 372 * During autoconfiguration or after a panic, a sleep will simply 373 * lower the priority briefly to allow interrupts, then return. 374 * The priority to be used (safepri) is machine-dependent, thus this 375 * value is initialized and maintained in the machine-dependent layers. 376 * This priority will typically be 0, or the lowest priority 377 * that is safe for use on the interrupt stack; it can be made 378 * higher to block network software interrupts after panics. 379 */ 380int safepri; 381 382void 383sleepinit(void) 384{ 385 int i; 386 387 sched_quantum = hz/10; 388 hogticks = 2 * sched_quantum; 389 for (i = 0; i < TABLESIZE; i++) 390 TAILQ_INIT(&slpque[i]); 391} 392 393/* 394 * General sleep call. Suspends the current process until a wakeup is 395 * performed on the specified identifier. The process will then be made 396 * runnable with the specified priority. Sleeps at most timo/hz seconds 397 * (0 means no timeout). If pri includes PCATCH flag, signals are checked 398 * before and after sleeping, else signals are not checked. Returns 0 if 399 * awakened, EWOULDBLOCK if the timeout expires. If PCATCH is set and a 400 * signal needs to be delivered, ERESTART is returned if the current system 401 * call should be restarted if possible, and EINTR is returned if the system 402 * call should be interrupted by the signal (return EINTR). 403 */ 404int 405tsleep(ident, priority, wmesg, timo) 406 void *ident; 407 int priority, timo; 408 const char *wmesg; 409{ 410 struct proc *p = curproc; 411 int s, sig, catch = priority & PCATCH; 412 struct callout_handle thandle; 413 414#ifdef KTRACE 415 if (p && KTRPOINT(p, KTR_CSW)) 416 ktrcsw(p->p_tracep, 1, 0); 417#endif 418 s = splhigh(); 419 if (cold || panicstr) { 420 /* 421 * After a panic, or during autoconfiguration, 422 * just give interrupts a chance, then just return; 423 * don't run any other procs or panic below, 424 * in case this is the idle process and already asleep. 425 */ 426 splx(safepri); 427 splx(s); 428 return (0); 429 } 430 KASSERT(p != NULL, ("tsleep1")); 431 KASSERT(ident != NULL && p->p_stat == SRUN, ("tsleep")); 432 /* 433 * Process may be sitting on a slpque if asleep() was called, remove 434 * it before re-adding. 435 */ 436 if (p->p_wchan != NULL) 437 unsleep(p); 438 439 p->p_wchan = ident; 440 p->p_wmesg = wmesg; 441 p->p_slptime = 0; 442 p->p_priority = priority & PRIMASK; 443 TAILQ_INSERT_TAIL(&slpque[LOOKUP(ident)], p, p_procq); 444 if (timo) 445 thandle = timeout(endtsleep, (void *)p, timo); 446 /* 447 * We put ourselves on the sleep queue and start our timeout 448 * before calling CURSIG, as we could stop there, and a wakeup 449 * or a SIGCONT (or both) could occur while we were stopped. 450 * A SIGCONT would cause us to be marked as SSLEEP 451 * without resuming us, thus we must be ready for sleep 452 * when CURSIG is called. If the wakeup happens while we're 453 * stopped, p->p_wchan will be 0 upon return from CURSIG. 454 */ 455 if (catch) { 456 p->p_flag |= P_SINTR; 457 if ((sig = CURSIG(p))) { 458 if (p->p_wchan) 459 unsleep(p); 460 p->p_stat = SRUN; 461 goto resume; 462 } 463 if (p->p_wchan == 0) { 464 catch = 0; 465 goto resume; 466 } 467 } else 468 sig = 0; 469 p->p_stat = SSLEEP; 470 p->p_stats->p_ru.ru_nvcsw++; 471 mi_switch(); 472resume: 473 curpriority = p->p_usrpri; 474 currtpriority = p->p_rtprio.prio; 475 splx(s); 476 p->p_flag &= ~P_SINTR; 477 if (p->p_flag & P_TIMEOUT) { 478 p->p_flag &= ~P_TIMEOUT; 479 if (sig == 0) { 480#ifdef KTRACE 481 if (KTRPOINT(p, KTR_CSW)) 482 ktrcsw(p->p_tracep, 0, 0); 483#endif 484 return (EWOULDBLOCK); 485 } 486 } else if (timo) 487 untimeout(endtsleep, (void *)p, thandle); 488 if (catch && (sig != 0 || (sig = CURSIG(p)))) { 489#ifdef KTRACE 490 if (KTRPOINT(p, KTR_CSW)) 491 ktrcsw(p->p_tracep, 0, 0); 492#endif 493 if (SIGISMEMBER(p->p_sigacts->ps_sigintr, sig)) 494 return (EINTR); 495 return (ERESTART); 496 } 497#ifdef KTRACE 498 if (KTRPOINT(p, KTR_CSW)) 499 ktrcsw(p->p_tracep, 0, 0); 500#endif 501 return (0); 502} 503 504/* 505 * asleep() - async sleep call. Place process on wait queue and return 506 * immediately without blocking. The process stays runnable until await() 507 * is called. If ident is NULL, remove process from wait queue if it is still 508 * on one. 509 * 510 * Only the most recent sleep condition is effective when making successive 511 * calls to asleep() or when calling tsleep(). 512 * 513 * The timeout, if any, is not initiated until await() is called. The sleep 514 * priority, signal, and timeout is specified in the asleep() call but may be 515 * overriden in the await() call. 516 * 517 * <<<<<<<< EXPERIMENTAL, UNTESTED >>>>>>>>>> 518 */ 519 520int 521asleep(void *ident, int priority, const char *wmesg, int timo) 522{ 523 struct proc *p = curproc; 524 int s; 525 526 /* 527 * splhigh() while manipulating sleep structures and slpque. 528 * 529 * Remove preexisting wait condition (if any) and place process 530 * on appropriate slpque, but do not put process to sleep. 531 */ 532 533 s = splhigh(); 534 535 if (p->p_wchan != NULL) 536 unsleep(p); 537 538 if (ident) { 539 p->p_wchan = ident; 540 p->p_wmesg = wmesg; 541 p->p_slptime = 0; 542 p->p_asleep.as_priority = priority; 543 p->p_asleep.as_timo = timo; 544 TAILQ_INSERT_TAIL(&slpque[LOOKUP(ident)], p, p_procq); 545 } 546 547 splx(s); 548 549 return(0); 550} 551 552/* 553 * await() - wait for async condition to occur. The process blocks until 554 * wakeup() is called on the most recent asleep() address. If wakeup is called 555 * priority to await(), await() winds up being a NOP. 556 * 557 * If await() is called more then once (without an intervening asleep() call), 558 * await() is still effectively a NOP but it calls mi_switch() to give other 559 * processes some cpu before returning. The process is left runnable. 560 * 561 * <<<<<<<< EXPERIMENTAL, UNTESTED >>>>>>>>>> 562 */ 563 564int 565await(int priority, int timo) 566{ 567 struct proc *p = curproc; 568 int s; 569 570 s = splhigh(); 571 572 if (p->p_wchan != NULL) { 573 struct callout_handle thandle; 574 int sig; 575 int catch; 576 577 /* 578 * The call to await() can override defaults specified in 579 * the original asleep(). 580 */ 581 if (priority < 0) 582 priority = p->p_asleep.as_priority; 583 if (timo < 0) 584 timo = p->p_asleep.as_timo; 585 586 /* 587 * Install timeout 588 */ 589 590 if (timo) 591 thandle = timeout(endtsleep, (void *)p, timo); 592 593 sig = 0; 594 catch = priority & PCATCH; 595 596 if (catch) { 597 p->p_flag |= P_SINTR; 598 if ((sig = CURSIG(p))) { 599 if (p->p_wchan) 600 unsleep(p); 601 p->p_stat = SRUN; 602 goto resume; 603 } 604 if (p->p_wchan == NULL) { 605 catch = 0; 606 goto resume; 607 } 608 } 609 p->p_stat = SSLEEP; 610 p->p_stats->p_ru.ru_nvcsw++; 611 mi_switch(); 612resume: 613 curpriority = p->p_usrpri; 614 currtpriority = p->p_rtprio.prio; 615 616 splx(s); 617 p->p_flag &= ~P_SINTR; 618 if (p->p_flag & P_TIMEOUT) { 619 p->p_flag &= ~P_TIMEOUT; 620 if (sig == 0) { 621#ifdef KTRACE 622 if (KTRPOINT(p, KTR_CSW)) 623 ktrcsw(p->p_tracep, 0, 0); 624#endif 625 return (EWOULDBLOCK); 626 } 627 } else if (timo) 628 untimeout(endtsleep, (void *)p, thandle); 629 if (catch && (sig != 0 || (sig = CURSIG(p)))) { 630#ifdef KTRACE 631 if (KTRPOINT(p, KTR_CSW)) 632 ktrcsw(p->p_tracep, 0, 0); 633#endif 634 if (SIGISMEMBER(p->p_sigacts->ps_sigintr, sig)) 635 return (EINTR); 636 return (ERESTART); 637 } 638#ifdef KTRACE 639 if (KTRPOINT(p, KTR_CSW)) 640 ktrcsw(p->p_tracep, 0, 0); 641#endif 642 } else { 643 /* 644 * If as_priority is 0, await() has been called without an 645 * intervening asleep(). We are still effectively a NOP, 646 * but we call mi_switch() for safety. 647 */ 648 649 if (p->p_asleep.as_priority == 0) { 650 p->p_stats->p_ru.ru_nvcsw++; 651 mi_switch(); 652 } 653 splx(s); 654 } 655 656 /* 657 * clear p_asleep.as_priority as an indication that await() has been 658 * called. If await() is called again without an intervening asleep(), 659 * await() is still effectively a NOP but the above mi_switch() code 660 * is triggered as a safety. 661 */ 662 p->p_asleep.as_priority = 0; 663 664 return (0); 665} 666 667/* 668 * Implement timeout for tsleep or asleep()/await() 669 * 670 * If process hasn't been awakened (wchan non-zero), 671 * set timeout flag and undo the sleep. If proc 672 * is stopped, just unsleep so it will remain stopped. 673 */ 674static void 675endtsleep(arg) 676 void *arg; 677{ 678 register struct proc *p; 679 int s; 680 681 p = (struct proc *)arg; 682 s = splhigh(); 683 if (p->p_wchan) { 684 if (p->p_stat == SSLEEP) 685 setrunnable(p); 686 else 687 unsleep(p); 688 p->p_flag |= P_TIMEOUT; 689 } 690 splx(s); 691} 692 693/* 694 * Remove a process from its wait queue 695 */ 696void 697unsleep(p) 698 register struct proc *p; 699{ 700 int s; 701 702 s = splhigh(); 703 if (p->p_wchan) { 704 TAILQ_REMOVE(&slpque[LOOKUP(p->p_wchan)], p, p_procq); 705 p->p_wchan = 0; 706 } 707 splx(s); 708} 709 710/* 711 * Make all processes sleeping on the specified identifier runnable. 712 */ 713void 714wakeup(ident) 715 register void *ident; 716{ 717 register struct slpquehead *qp; 718 register struct proc *p; 719 int s; 720 721 s = splhigh(); 722 qp = &slpque[LOOKUP(ident)]; 723restart: 724 TAILQ_FOREACH(p, qp, p_procq) { 725 if (p->p_wchan == ident) { 726 TAILQ_REMOVE(qp, p, p_procq); 727 p->p_wchan = 0; 728 if (p->p_stat == SSLEEP) { 729 /* OPTIMIZED EXPANSION OF setrunnable(p); */ 730 if (p->p_slptime > 1) 731 updatepri(p); 732 p->p_slptime = 0; 733 p->p_stat = SRUN; 734 if (p->p_flag & P_INMEM) { 735 setrunqueue(p); 736 maybe_resched(p); 737 } else { 738 p->p_flag |= P_SWAPINREQ; 739 wakeup((caddr_t)&proc0); 740 } 741 /* END INLINE EXPANSION */ 742 goto restart; 743 } 744 } 745 } 746 splx(s); 747} 748 749/* 750 * Make a process sleeping on the specified identifier runnable. 751 * May wake more than one process if a target prcoess is currently 752 * swapped out. 753 */ 754void 755wakeup_one(ident) 756 register void *ident; 757{ 758 register struct slpquehead *qp; 759 register struct proc *p; 760 int s; 761 762 s = splhigh(); 763 qp = &slpque[LOOKUP(ident)]; 764 765 TAILQ_FOREACH(p, qp, p_procq) { 766 if (p->p_wchan == ident) { 767 TAILQ_REMOVE(qp, p, p_procq); 768 p->p_wchan = 0; 769 if (p->p_stat == SSLEEP) { 770 /* OPTIMIZED EXPANSION OF setrunnable(p); */ 771 if (p->p_slptime > 1) 772 updatepri(p); 773 p->p_slptime = 0; 774 p->p_stat = SRUN; 775 if (p->p_flag & P_INMEM) { 776 setrunqueue(p); 777 maybe_resched(p); 778 break; 779 } else { 780 p->p_flag |= P_SWAPINREQ; 781 wakeup((caddr_t)&proc0); 782 } 783 /* END INLINE EXPANSION */ 784 } 785 } 786 } 787 splx(s); 788} 789 790/* 791 * The machine independent parts of mi_switch(). 792 * Must be called at splstatclock() or higher. 793 */ 794void 795mi_switch() 796{ 797 struct timeval new_switchtime; 798 register struct proc *p = curproc; /* XXX */ 799 register struct rlimit *rlim; 800 int x; 801 802 /* 803 * XXX this spl is almost unnecessary. It is partly to allow for 804 * sloppy callers that don't do it (issignal() via CURSIG() is the 805 * main offender). It is partly to work around a bug in the i386 806 * cpu_switch() (the ipl is not preserved). We ran for years 807 * without it. I think there was only a interrupt latency problem. 808 * The main caller, tsleep(), does an splx() a couple of instructions 809 * after calling here. The buggy caller, issignal(), usually calls 810 * here at spl0() and sometimes returns at splhigh(). The process 811 * then runs for a little too long at splhigh(). The ipl gets fixed 812 * when the process returns to user mode (or earlier). 813 * 814 * It would probably be better to always call here at spl0(). Callers 815 * are prepared to give up control to another process, so they must 816 * be prepared to be interrupted. The clock stuff here may not 817 * actually need splstatclock(). 818 */ 819 x = splstatclock(); 820 821#ifdef SIMPLELOCK_DEBUG 822 if (p->p_simple_locks) 823 printf("sleep: holding simple lock\n"); 824#endif 825 /* 826 * Compute the amount of time during which the current 827 * process was running, and add that to its total so far. 828 */ 829 microuptime(&new_switchtime); 830 if (timevalcmp(&new_switchtime, &switchtime, <)) { 831 printf("microuptime() went backwards (%ld.%06ld -> %ld,%06ld)\n", 832 switchtime.tv_sec, switchtime.tv_usec, 833 new_switchtime.tv_sec, new_switchtime.tv_usec); 834 new_switchtime = switchtime; 835 } else { 836 p->p_runtime += (new_switchtime.tv_usec - switchtime.tv_usec) + 837 (new_switchtime.tv_sec - switchtime.tv_sec) * (int64_t)1000000; 838 } 839 840 /* 841 * Check if the process exceeds its cpu resource allocation. 842 * If over max, kill it. 843 */ 844 if (p->p_stat != SZOMB && p->p_limit->p_cpulimit != RLIM_INFINITY && 845 p->p_runtime > p->p_limit->p_cpulimit) { 846 rlim = &p->p_rlimit[RLIMIT_CPU]; 847 if (p->p_runtime / (rlim_t)1000000 >= rlim->rlim_max) { 848 killproc(p, "exceeded maximum CPU limit"); 849 } else { 850 psignal(p, SIGXCPU); 851 if (rlim->rlim_cur < rlim->rlim_max) { 852 /* XXX: we should make a private copy */ 853 rlim->rlim_cur += 5; 854 } 855 } 856 } 857 858 /* 859 * Pick a new current process and record its start time. 860 */ 861 cnt.v_swtch++; 862 switchtime = new_switchtime; 863 cpu_switch(p); 864 if (switchtime.tv_sec == 0) 865 microuptime(&switchtime); 866 switchticks = ticks; 867 868 splx(x); 869} 870 871/* 872 * Change process state to be runnable, 873 * placing it on the run queue if it is in memory, 874 * and awakening the swapper if it isn't in memory. 875 */ 876void 877setrunnable(p) 878 register struct proc *p; 879{ 880 register int s; 881 882 s = splhigh(); 883 switch (p->p_stat) { 884 case 0: 885 case SRUN: 886 case SZOMB: 887 default: 888 panic("setrunnable"); 889 case SSTOP: 890 case SSLEEP: 891 unsleep(p); /* e.g. when sending signals */ 892 break; 893 894 case SIDL: 895 break; 896 } 897 p->p_stat = SRUN; 898 if (p->p_flag & P_INMEM) 899 setrunqueue(p); 900 splx(s); 901 if (p->p_slptime > 1) 902 updatepri(p); 903 p->p_slptime = 0; 904 if ((p->p_flag & P_INMEM) == 0) { 905 p->p_flag |= P_SWAPINREQ; 906 wakeup((caddr_t)&proc0); 907 } 908 else 909 maybe_resched(p); 910} 911 912/* 913 * Compute the priority of a process when running in user mode. 914 * Arrange to reschedule if the resulting priority is better 915 * than that of the current process. 916 */ 917void 918resetpriority(p) 919 register struct proc *p; 920{ 921 register unsigned int newpriority; 922 923 if (p->p_rtprio.type == RTP_PRIO_NORMAL) { 924 newpriority = PUSER + p->p_estcpu / INVERSE_ESTCPU_WEIGHT + 925 NICE_WEIGHT * p->p_nice; 926 newpriority = min(newpriority, MAXPRI); 927 p->p_usrpri = newpriority; 928 } 929 maybe_resched(p); 930} 931 932/* ARGSUSED */ 933static void 934sched_setup(dummy) 935 void *dummy; 936{ 937 /* Kick off timeout driven events by calling first time. */ 938 roundrobin(NULL); 939 schedcpu(NULL); 940} 941 942/* 943 * We adjust the priority of the current process. The priority of 944 * a process gets worse as it accumulates CPU time. The cpu usage 945 * estimator (p_estcpu) is increased here. resetpriority() will 946 * compute a different priority each time p_estcpu increases by 947 * INVERSE_ESTCPU_WEIGHT 948 * (until MAXPRI is reached). The cpu usage estimator ramps up 949 * quite quickly when the process is running (linearly), and decays 950 * away exponentially, at a rate which is proportionally slower when 951 * the system is busy. The basic principle is that the system will 952 * 90% forget that the process used a lot of CPU time in 5 * loadav 953 * seconds. This causes the system to favor processes which haven't 954 * run much recently, and to round-robin among other processes. 955 */ 956void 957schedclock(p) 958 struct proc *p; 959{ 960 961 p->p_cpticks++; 962 p->p_estcpu = ESTCPULIM(p->p_estcpu + 1); 963 if ((p->p_estcpu % INVERSE_ESTCPU_WEIGHT) == 0) { 964 resetpriority(p); 965 if (p->p_priority >= PUSER) 966 p->p_priority = p->p_usrpri; 967 } 968} 969