kern_synch.c revision 58717
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 58717 2000-03-28 07:16:37Z dillon $ 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#include <machine/ipl.h> 61#ifdef SMP 62#include <machine/smp.h> 63#endif 64 65static void sched_setup __P((void *dummy)); 66SYSINIT(sched_setup, SI_SUB_KICK_SCHEDULER, SI_ORDER_FIRST, sched_setup, NULL) 67 68u_char curpriority; 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 c_class = RTP_PRIO_BASE(curproc->p_rtprio.type); 116 p_class = RTP_PRIO_BASE(p->p_rtprio.type); 117 if (p_class != c_class) 118 return (p_class - c_class); 119 if (p_class == RTP_PRIO_NORMAL) 120 return (((int)p->p_priority - (int)curpriority) / PPQ); 121 return ((int)p->p_rtprio.prio - (int)curproc->p_rtprio.prio); 122} 123 124/* 125 * Arrange to reschedule if necessary, taking the priorities and 126 * schedulers into account. 127 */ 128static void 129maybe_resched(chk) 130 struct proc *chk; 131{ 132 struct proc *p = curproc; /* XXX */ 133 134 /* 135 * XXX idle scheduler still broken because proccess stays on idle 136 * scheduler during waits (such as when getting FS locks). If a 137 * standard process becomes runaway cpu-bound, the system can lockup 138 * due to idle-scheduler processes in wakeup never getting any cpu. 139 */ 140 if (p == NULL) { 141#if 0 142 need_resched(); 143#endif 144 } else if (chk == p) { 145 /* We may need to yield if our priority has been raised. */ 146 if (curpriority_cmp(chk) > 0) 147 need_resched(); 148 } else if (curpriority_cmp(chk) < 0) 149 need_resched(); 150} 151 152int 153roundrobin_interval(void) 154{ 155 return (sched_quantum); 156} 157 158/* 159 * Force switch among equal priority processes every 100ms. 160 */ 161/* ARGSUSED */ 162static void 163roundrobin(arg) 164 void *arg; 165{ 166#ifndef SMP 167 struct proc *p = curproc; /* XXX */ 168#endif 169 170#ifdef SMP 171 need_resched(); 172 forward_roundrobin(); 173#else 174 if (p == 0 || RTP_PRIO_NEED_RR(p->p_rtprio.type)) 175 need_resched(); 176#endif 177 178 timeout(roundrobin, NULL, sched_quantum); 179} 180 181/* 182 * Constants for digital decay and forget: 183 * 90% of (p_estcpu) usage in 5 * loadav time 184 * 95% of (p_pctcpu) usage in 60 seconds (load insensitive) 185 * Note that, as ps(1) mentions, this can let percentages 186 * total over 100% (I've seen 137.9% for 3 processes). 187 * 188 * Note that schedclock() updates p_estcpu and p_cpticks asynchronously. 189 * 190 * We wish to decay away 90% of p_estcpu in (5 * loadavg) seconds. 191 * That is, the system wants to compute a value of decay such 192 * that the following for loop: 193 * for (i = 0; i < (5 * loadavg); i++) 194 * p_estcpu *= decay; 195 * will compute 196 * p_estcpu *= 0.1; 197 * for all values of loadavg: 198 * 199 * Mathematically this loop can be expressed by saying: 200 * decay ** (5 * loadavg) ~= .1 201 * 202 * The system computes decay as: 203 * decay = (2 * loadavg) / (2 * loadavg + 1) 204 * 205 * We wish to prove that the system's computation of decay 206 * will always fulfill the equation: 207 * decay ** (5 * loadavg) ~= .1 208 * 209 * If we compute b as: 210 * b = 2 * loadavg 211 * then 212 * decay = b / (b + 1) 213 * 214 * We now need to prove two things: 215 * 1) Given factor ** (5 * loadavg) ~= .1, prove factor == b/(b+1) 216 * 2) Given b/(b+1) ** power ~= .1, prove power == (5 * loadavg) 217 * 218 * Facts: 219 * For x close to zero, exp(x) =~ 1 + x, since 220 * exp(x) = 0! + x**1/1! + x**2/2! + ... . 221 * therefore exp(-1/b) =~ 1 - (1/b) = (b-1)/b. 222 * For x close to zero, ln(1+x) =~ x, since 223 * ln(1+x) = x - x**2/2 + x**3/3 - ... -1 < x < 1 224 * therefore ln(b/(b+1)) = ln(1 - 1/(b+1)) =~ -1/(b+1). 225 * ln(.1) =~ -2.30 226 * 227 * Proof of (1): 228 * Solve (factor)**(power) =~ .1 given power (5*loadav): 229 * solving for factor, 230 * ln(factor) =~ (-2.30/5*loadav), or 231 * factor =~ exp(-1/((5/2.30)*loadav)) =~ exp(-1/(2*loadav)) = 232 * exp(-1/b) =~ (b-1)/b =~ b/(b+1). QED 233 * 234 * Proof of (2): 235 * Solve (factor)**(power) =~ .1 given factor == (b/(b+1)): 236 * solving for power, 237 * power*ln(b/(b+1)) =~ -2.30, or 238 * power =~ 2.3 * (b + 1) = 4.6*loadav + 2.3 =~ 5*loadav. QED 239 * 240 * Actual power values for the implemented algorithm are as follows: 241 * loadav: 1 2 3 4 242 * power: 5.68 10.32 14.94 19.55 243 */ 244 245/* calculations for digital decay to forget 90% of usage in 5*loadav sec */ 246#define loadfactor(loadav) (2 * (loadav)) 247#define decay_cpu(loadfac, cpu) (((loadfac) * (cpu)) / ((loadfac) + FSCALE)) 248 249/* decay 95% of `p_pctcpu' in 60 seconds; see CCPU_SHIFT before changing */ 250static fixpt_t ccpu = 0.95122942450071400909 * FSCALE; /* exp(-1/20) */ 251SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, ""); 252 253/* kernel uses `FSCALE', userland (SHOULD) use kern.fscale */ 254static int fscale __unused = FSCALE; 255SYSCTL_INT(_kern, OID_AUTO, fscale, CTLFLAG_RD, 0, FSCALE, ""); 256 257/* 258 * If `ccpu' is not equal to `exp(-1/20)' and you still want to use the 259 * faster/more-accurate formula, you'll have to estimate CCPU_SHIFT below 260 * and possibly adjust FSHIFT in "param.h" so that (FSHIFT >= CCPU_SHIFT). 261 * 262 * To estimate CCPU_SHIFT for exp(-1/20), the following formula was used: 263 * 1 - exp(-1/20) ~= 0.0487 ~= 0.0488 == 1 (fixed pt, *11* bits). 264 * 265 * If you don't want to bother with the faster/more-accurate formula, you 266 * can set CCPU_SHIFT to (FSHIFT + 1) which will use a slower/less-accurate 267 * (more general) method of calculating the %age of CPU used by a process. 268 */ 269#define CCPU_SHIFT 11 270 271/* 272 * Recompute process priorities, every hz ticks. 273 */ 274/* ARGSUSED */ 275static void 276schedcpu(arg) 277 void *arg; 278{ 279 register fixpt_t loadfac = loadfactor(averunnable.ldavg[0]); 280 register struct proc *p; 281 register int realstathz, s; 282 283 realstathz = stathz ? stathz : hz; 284 LIST_FOREACH(p, &allproc, p_list) { 285 /* 286 * Increment time in/out of memory and sleep time 287 * (if sleeping). We ignore overflow; with 16-bit int's 288 * (remember them?) overflow takes 45 days. 289 */ 290 p->p_swtime++; 291 if (p->p_stat == SSLEEP || p->p_stat == SSTOP) 292 p->p_slptime++; 293 p->p_pctcpu = (p->p_pctcpu * ccpu) >> FSHIFT; 294 /* 295 * If the process has slept the entire second, 296 * stop recalculating its priority until it wakes up. 297 */ 298 if (p->p_slptime > 1) 299 continue; 300 s = splhigh(); /* prevent state changes and protect run queue */ 301 /* 302 * p_pctcpu is only for ps. 303 */ 304#if (FSHIFT >= CCPU_SHIFT) 305 p->p_pctcpu += (realstathz == 100)? 306 ((fixpt_t) p->p_cpticks) << (FSHIFT - CCPU_SHIFT): 307 100 * (((fixpt_t) p->p_cpticks) 308 << (FSHIFT - CCPU_SHIFT)) / realstathz; 309#else 310 p->p_pctcpu += ((FSCALE - ccpu) * 311 (p->p_cpticks * FSCALE / realstathz)) >> FSHIFT; 312#endif 313 p->p_cpticks = 0; 314 p->p_estcpu = decay_cpu(loadfac, p->p_estcpu); 315 resetpriority(p); 316 if (p->p_priority >= PUSER) { 317 if ((p != curproc) && 318#ifdef SMP 319 p->p_oncpu == 0xff && /* idle */ 320#endif 321 p->p_stat == SRUN && 322 (p->p_flag & P_INMEM) && 323 (p->p_priority / PPQ) != (p->p_usrpri / PPQ)) { 324 remrunqueue(p); 325 p->p_priority = p->p_usrpri; 326 setrunqueue(p); 327 } else 328 p->p_priority = p->p_usrpri; 329 } 330 splx(s); 331 } 332 vmmeter(); 333 wakeup((caddr_t)&lbolt); 334 timeout(schedcpu, (void *)0, hz); 335} 336 337/* 338 * Recalculate the priority of a process after it has slept for a while. 339 * For all load averages >= 1 and max p_estcpu of 255, sleeping for at 340 * least six times the loadfactor will decay p_estcpu to zero. 341 */ 342static void 343updatepri(p) 344 register struct proc *p; 345{ 346 register unsigned int newcpu = p->p_estcpu; 347 register fixpt_t loadfac = loadfactor(averunnable.ldavg[0]); 348 349 if (p->p_slptime > 5 * loadfac) 350 p->p_estcpu = 0; 351 else { 352 p->p_slptime--; /* the first time was done in schedcpu */ 353 while (newcpu && --p->p_slptime) 354 newcpu = decay_cpu(loadfac, newcpu); 355 p->p_estcpu = newcpu; 356 } 357 resetpriority(p); 358} 359 360/* 361 * We're only looking at 7 bits of the address; everything is 362 * aligned to 4, lots of things are aligned to greater powers 363 * of 2. Shift right by 8, i.e. drop the bottom 256 worth. 364 */ 365#define TABLESIZE 128 366static TAILQ_HEAD(slpquehead, proc) slpque[TABLESIZE]; 367#define LOOKUP(x) (((intptr_t)(x) >> 8) & (TABLESIZE - 1)) 368 369/* 370 * During autoconfiguration or after a panic, a sleep will simply 371 * lower the priority briefly to allow interrupts, then return. 372 * The priority to be used (safepri) is machine-dependent, thus this 373 * value is initialized and maintained in the machine-dependent layers. 374 * This priority will typically be 0, or the lowest priority 375 * that is safe for use on the interrupt stack; it can be made 376 * higher to block network software interrupts after panics. 377 */ 378int safepri; 379 380void 381sleepinit(void) 382{ 383 int i; 384 385 sched_quantum = hz/10; 386 hogticks = 2 * sched_quantum; 387 for (i = 0; i < TABLESIZE; i++) 388 TAILQ_INIT(&slpque[i]); 389} 390 391/* 392 * General sleep call. Suspends the current process until a wakeup is 393 * performed on the specified identifier. The process will then be made 394 * runnable with the specified priority. Sleeps at most timo/hz seconds 395 * (0 means no timeout). If pri includes PCATCH flag, signals are checked 396 * before and after sleeping, else signals are not checked. Returns 0 if 397 * awakened, EWOULDBLOCK if the timeout expires. If PCATCH is set and a 398 * signal needs to be delivered, ERESTART is returned if the current system 399 * call should be restarted if possible, and EINTR is returned if the system 400 * call should be interrupted by the signal (return EINTR). 401 */ 402int 403tsleep(ident, priority, wmesg, timo) 404 void *ident; 405 int priority, timo; 406 const char *wmesg; 407{ 408 struct proc *p = curproc; 409 int s, sig, catch = priority & PCATCH; 410 struct callout_handle thandle; 411 412#ifdef KTRACE 413 if (p && KTRPOINT(p, KTR_CSW)) 414 ktrcsw(p->p_tracep, 1, 0); 415#endif 416 s = splhigh(); 417 if (cold || panicstr) { 418 /* 419 * After a panic, or during autoconfiguration, 420 * just give interrupts a chance, then just return; 421 * don't run any other procs or panic below, 422 * in case this is the idle process and already asleep. 423 */ 424 splx(safepri); 425 splx(s); 426 return (0); 427 } 428 KASSERT(p != NULL, ("tsleep1")); 429 KASSERT(ident != NULL && p->p_stat == SRUN, ("tsleep")); 430 /* 431 * Process may be sitting on a slpque if asleep() was called, remove 432 * it before re-adding. 433 */ 434 if (p->p_wchan != NULL) 435 unsleep(p); 436 437 p->p_wchan = ident; 438 p->p_wmesg = wmesg; 439 p->p_slptime = 0; 440 p->p_priority = priority & PRIMASK; 441 TAILQ_INSERT_TAIL(&slpque[LOOKUP(ident)], p, p_procq); 442 if (timo) 443 thandle = timeout(endtsleep, (void *)p, timo); 444 /* 445 * We put ourselves on the sleep queue and start our timeout 446 * before calling CURSIG, as we could stop there, and a wakeup 447 * or a SIGCONT (or both) could occur while we were stopped. 448 * A SIGCONT would cause us to be marked as SSLEEP 449 * without resuming us, thus we must be ready for sleep 450 * when CURSIG is called. If the wakeup happens while we're 451 * stopped, p->p_wchan will be 0 upon return from CURSIG. 452 */ 453 if (catch) { 454 p->p_flag |= P_SINTR; 455 if ((sig = CURSIG(p))) { 456 if (p->p_wchan) 457 unsleep(p); 458 p->p_stat = SRUN; 459 goto resume; 460 } 461 if (p->p_wchan == 0) { 462 catch = 0; 463 goto resume; 464 } 465 } else 466 sig = 0; 467 p->p_stat = SSLEEP; 468 p->p_stats->p_ru.ru_nvcsw++; 469 mi_switch(); 470resume: 471 curpriority = p->p_usrpri; 472 splx(s); 473 p->p_flag &= ~P_SINTR; 474 if (p->p_flag & P_TIMEOUT) { 475 p->p_flag &= ~P_TIMEOUT; 476 if (sig == 0) { 477#ifdef KTRACE 478 if (KTRPOINT(p, KTR_CSW)) 479 ktrcsw(p->p_tracep, 0, 0); 480#endif 481 return (EWOULDBLOCK); 482 } 483 } else if (timo) 484 untimeout(endtsleep, (void *)p, thandle); 485 if (catch && (sig != 0 || (sig = CURSIG(p)))) { 486#ifdef KTRACE 487 if (KTRPOINT(p, KTR_CSW)) 488 ktrcsw(p->p_tracep, 0, 0); 489#endif 490 if (SIGISMEMBER(p->p_sigacts->ps_sigintr, sig)) 491 return (EINTR); 492 return (ERESTART); 493 } 494#ifdef KTRACE 495 if (KTRPOINT(p, KTR_CSW)) 496 ktrcsw(p->p_tracep, 0, 0); 497#endif 498 return (0); 499} 500 501/* 502 * asleep() - async sleep call. Place process on wait queue and return 503 * immediately without blocking. The process stays runnable until await() 504 * is called. If ident is NULL, remove process from wait queue if it is still 505 * on one. 506 * 507 * Only the most recent sleep condition is effective when making successive 508 * calls to asleep() or when calling tsleep(). 509 * 510 * The timeout, if any, is not initiated until await() is called. The sleep 511 * priority, signal, and timeout is specified in the asleep() call but may be 512 * overriden in the await() call. 513 * 514 * <<<<<<<< EXPERIMENTAL, UNTESTED >>>>>>>>>> 515 */ 516 517int 518asleep(void *ident, int priority, const char *wmesg, int timo) 519{ 520 struct proc *p = curproc; 521 int s; 522 523 /* 524 * splhigh() while manipulating sleep structures and slpque. 525 * 526 * Remove preexisting wait condition (if any) and place process 527 * on appropriate slpque, but do not put process to sleep. 528 */ 529 530 s = splhigh(); 531 532 if (p->p_wchan != NULL) 533 unsleep(p); 534 535 if (ident) { 536 p->p_wchan = ident; 537 p->p_wmesg = wmesg; 538 p->p_slptime = 0; 539 p->p_asleep.as_priority = priority; 540 p->p_asleep.as_timo = timo; 541 TAILQ_INSERT_TAIL(&slpque[LOOKUP(ident)], p, p_procq); 542 } 543 544 splx(s); 545 546 return(0); 547} 548 549/* 550 * await() - wait for async condition to occur. The process blocks until 551 * wakeup() is called on the most recent asleep() address. If wakeup is called 552 * priority to await(), await() winds up being a NOP. 553 * 554 * If await() is called more then once (without an intervening asleep() call), 555 * await() is still effectively a NOP but it calls mi_switch() to give other 556 * processes some cpu before returning. The process is left runnable. 557 * 558 * <<<<<<<< EXPERIMENTAL, UNTESTED >>>>>>>>>> 559 */ 560 561int 562await(int priority, int timo) 563{ 564 struct proc *p = curproc; 565 int s; 566 567 s = splhigh(); 568 569 if (p->p_wchan != NULL) { 570 struct callout_handle thandle; 571 int sig; 572 int catch; 573 574 /* 575 * The call to await() can override defaults specified in 576 * the original asleep(). 577 */ 578 if (priority < 0) 579 priority = p->p_asleep.as_priority; 580 if (timo < 0) 581 timo = p->p_asleep.as_timo; 582 583 /* 584 * Install timeout 585 */ 586 587 if (timo) 588 thandle = timeout(endtsleep, (void *)p, timo); 589 590 sig = 0; 591 catch = priority & PCATCH; 592 593 if (catch) { 594 p->p_flag |= P_SINTR; 595 if ((sig = CURSIG(p))) { 596 if (p->p_wchan) 597 unsleep(p); 598 p->p_stat = SRUN; 599 goto resume; 600 } 601 if (p->p_wchan == NULL) { 602 catch = 0; 603 goto resume; 604 } 605 } 606 p->p_stat = SSLEEP; 607 p->p_stats->p_ru.ru_nvcsw++; 608 mi_switch(); 609resume: 610 curpriority = p->p_usrpri; 611 612 splx(s); 613 p->p_flag &= ~P_SINTR; 614 if (p->p_flag & P_TIMEOUT) { 615 p->p_flag &= ~P_TIMEOUT; 616 if (sig == 0) { 617#ifdef KTRACE 618 if (KTRPOINT(p, KTR_CSW)) 619 ktrcsw(p->p_tracep, 0, 0); 620#endif 621 return (EWOULDBLOCK); 622 } 623 } else if (timo) 624 untimeout(endtsleep, (void *)p, thandle); 625 if (catch && (sig != 0 || (sig = CURSIG(p)))) { 626#ifdef KTRACE 627 if (KTRPOINT(p, KTR_CSW)) 628 ktrcsw(p->p_tracep, 0, 0); 629#endif 630 if (SIGISMEMBER(p->p_sigacts->ps_sigintr, sig)) 631 return (EINTR); 632 return (ERESTART); 633 } 634#ifdef KTRACE 635 if (KTRPOINT(p, KTR_CSW)) 636 ktrcsw(p->p_tracep, 0, 0); 637#endif 638 } else { 639 /* 640 * If as_priority is 0, await() has been called without an 641 * intervening asleep(). We are still effectively a NOP, 642 * but we call mi_switch() for safety. 643 */ 644 645 if (p->p_asleep.as_priority == 0) { 646 p->p_stats->p_ru.ru_nvcsw++; 647 mi_switch(); 648 } 649 splx(s); 650 } 651 652 /* 653 * clear p_asleep.as_priority as an indication that await() has been 654 * called. If await() is called again without an intervening asleep(), 655 * await() is still effectively a NOP but the above mi_switch() code 656 * is triggered as a safety. 657 */ 658 p->p_asleep.as_priority = 0; 659 660 return (0); 661} 662 663/* 664 * Implement timeout for tsleep or asleep()/await() 665 * 666 * If process hasn't been awakened (wchan non-zero), 667 * set timeout flag and undo the sleep. If proc 668 * is stopped, just unsleep so it will remain stopped. 669 */ 670static void 671endtsleep(arg) 672 void *arg; 673{ 674 register struct proc *p; 675 int s; 676 677 p = (struct proc *)arg; 678 s = splhigh(); 679 if (p->p_wchan) { 680 if (p->p_stat == SSLEEP) 681 setrunnable(p); 682 else 683 unsleep(p); 684 p->p_flag |= P_TIMEOUT; 685 } 686 splx(s); 687} 688 689/* 690 * Remove a process from its wait queue 691 */ 692void 693unsleep(p) 694 register struct proc *p; 695{ 696 int s; 697 698 s = splhigh(); 699 if (p->p_wchan) { 700 TAILQ_REMOVE(&slpque[LOOKUP(p->p_wchan)], p, p_procq); 701 p->p_wchan = 0; 702 } 703 splx(s); 704} 705 706/* 707 * Make all processes sleeping on the specified identifier runnable. 708 */ 709void 710wakeup(ident) 711 register void *ident; 712{ 713 register struct slpquehead *qp; 714 register struct proc *p; 715 int s; 716 717 s = splhigh(); 718 qp = &slpque[LOOKUP(ident)]; 719restart: 720 TAILQ_FOREACH(p, qp, p_procq) { 721 if (p->p_wchan == ident) { 722 TAILQ_REMOVE(qp, p, p_procq); 723 p->p_wchan = 0; 724 if (p->p_stat == SSLEEP) { 725 /* OPTIMIZED EXPANSION OF setrunnable(p); */ 726 if (p->p_slptime > 1) 727 updatepri(p); 728 p->p_slptime = 0; 729 p->p_stat = SRUN; 730 if (p->p_flag & P_INMEM) { 731 setrunqueue(p); 732 maybe_resched(p); 733 } else { 734 p->p_flag |= P_SWAPINREQ; 735 wakeup((caddr_t)&proc0); 736 } 737 /* END INLINE EXPANSION */ 738 goto restart; 739 } 740 } 741 } 742 splx(s); 743} 744 745/* 746 * Make a process sleeping on the specified identifier runnable. 747 * May wake more than one process if a target prcoess is currently 748 * swapped out. 749 */ 750void 751wakeup_one(ident) 752 register void *ident; 753{ 754 register struct slpquehead *qp; 755 register struct proc *p; 756 int s; 757 758 s = splhigh(); 759 qp = &slpque[LOOKUP(ident)]; 760 761 TAILQ_FOREACH(p, qp, p_procq) { 762 if (p->p_wchan == ident) { 763 TAILQ_REMOVE(qp, p, p_procq); 764 p->p_wchan = 0; 765 if (p->p_stat == SSLEEP) { 766 /* OPTIMIZED EXPANSION OF setrunnable(p); */ 767 if (p->p_slptime > 1) 768 updatepri(p); 769 p->p_slptime = 0; 770 p->p_stat = SRUN; 771 if (p->p_flag & P_INMEM) { 772 setrunqueue(p); 773 maybe_resched(p); 774 break; 775 } else { 776 p->p_flag |= P_SWAPINREQ; 777 wakeup((caddr_t)&proc0); 778 } 779 /* END INLINE EXPANSION */ 780 } 781 } 782 } 783 splx(s); 784} 785 786/* 787 * The machine independent parts of mi_switch(). 788 * Must be called at splstatclock() or higher. 789 */ 790void 791mi_switch() 792{ 793 struct timeval new_switchtime; 794 register struct proc *p = curproc; /* XXX */ 795 register struct rlimit *rlim; 796 int x; 797 798 /* 799 * XXX this spl is almost unnecessary. It is partly to allow for 800 * sloppy callers that don't do it (issignal() via CURSIG() is the 801 * main offender). It is partly to work around a bug in the i386 802 * cpu_switch() (the ipl is not preserved). We ran for years 803 * without it. I think there was only a interrupt latency problem. 804 * The main caller, tsleep(), does an splx() a couple of instructions 805 * after calling here. The buggy caller, issignal(), usually calls 806 * here at spl0() and sometimes returns at splhigh(). The process 807 * then runs for a little too long at splhigh(). The ipl gets fixed 808 * when the process returns to user mode (or earlier). 809 * 810 * It would probably be better to always call here at spl0(). Callers 811 * are prepared to give up control to another process, so they must 812 * be prepared to be interrupted. The clock stuff here may not 813 * actually need splstatclock(). 814 */ 815 x = splstatclock(); 816 817#ifdef SIMPLELOCK_DEBUG 818 if (p->p_simple_locks) 819 printf("sleep: holding simple lock\n"); 820#endif 821 /* 822 * Compute the amount of time during which the current 823 * process was running, and add that to its total so far. 824 */ 825 microuptime(&new_switchtime); 826 if (timevalcmp(&new_switchtime, &switchtime, <)) { 827 printf("microuptime() went backwards (%ld.%06ld -> %ld,%06ld)\n", 828 switchtime.tv_sec, switchtime.tv_usec, 829 new_switchtime.tv_sec, new_switchtime.tv_usec); 830 new_switchtime = switchtime; 831 } else { 832 p->p_runtime += (new_switchtime.tv_usec - switchtime.tv_usec) + 833 (new_switchtime.tv_sec - switchtime.tv_sec) * (int64_t)1000000; 834 } 835 836 /* 837 * Check if the process exceeds its cpu resource allocation. 838 * If over max, kill it. 839 */ 840 if (p->p_stat != SZOMB && p->p_limit->p_cpulimit != RLIM_INFINITY && 841 p->p_runtime > p->p_limit->p_cpulimit) { 842 rlim = &p->p_rlimit[RLIMIT_CPU]; 843 if (p->p_runtime / (rlim_t)1000000 >= rlim->rlim_max) { 844 killproc(p, "exceeded maximum CPU limit"); 845 } else { 846 psignal(p, SIGXCPU); 847 if (rlim->rlim_cur < rlim->rlim_max) { 848 /* XXX: we should make a private copy */ 849 rlim->rlim_cur += 5; 850 } 851 } 852 } 853 854 /* 855 * Pick a new current process and record its start time. 856 */ 857 cnt.v_swtch++; 858 switchtime = new_switchtime; 859 cpu_switch(p); 860 if (switchtime.tv_sec == 0) 861 microuptime(&switchtime); 862 switchticks = ticks; 863 864 splx(x); 865} 866 867/* 868 * Change process state to be runnable, 869 * placing it on the run queue if it is in memory, 870 * and awakening the swapper if it isn't in memory. 871 */ 872void 873setrunnable(p) 874 register struct proc *p; 875{ 876 register int s; 877 878 s = splhigh(); 879 switch (p->p_stat) { 880 case 0: 881 case SRUN: 882 case SZOMB: 883 default: 884 panic("setrunnable"); 885 case SSTOP: 886 case SSLEEP: 887 unsleep(p); /* e.g. when sending signals */ 888 break; 889 890 case SIDL: 891 break; 892 } 893 p->p_stat = SRUN; 894 if (p->p_flag & P_INMEM) 895 setrunqueue(p); 896 splx(s); 897 if (p->p_slptime > 1) 898 updatepri(p); 899 p->p_slptime = 0; 900 if ((p->p_flag & P_INMEM) == 0) { 901 p->p_flag |= P_SWAPINREQ; 902 wakeup((caddr_t)&proc0); 903 } 904 else 905 maybe_resched(p); 906} 907 908/* 909 * Compute the priority of a process when running in user mode. 910 * Arrange to reschedule if the resulting priority is better 911 * than that of the current process. 912 */ 913void 914resetpriority(p) 915 register struct proc *p; 916{ 917 register unsigned int newpriority; 918 919 if (p->p_rtprio.type == RTP_PRIO_NORMAL) { 920 newpriority = PUSER + p->p_estcpu / INVERSE_ESTCPU_WEIGHT + 921 NICE_WEIGHT * p->p_nice; 922 newpriority = min(newpriority, MAXPRI); 923 p->p_usrpri = newpriority; 924 } 925 maybe_resched(p); 926} 927 928/* ARGSUSED */ 929static void 930sched_setup(dummy) 931 void *dummy; 932{ 933 /* Kick off timeout driven events by calling first time. */ 934 roundrobin(NULL); 935 schedcpu(NULL); 936} 937 938/* 939 * We adjust the priority of the current process. The priority of 940 * a process gets worse as it accumulates CPU time. The cpu usage 941 * estimator (p_estcpu) is increased here. resetpriority() will 942 * compute a different priority each time p_estcpu increases by 943 * INVERSE_ESTCPU_WEIGHT 944 * (until MAXPRI is reached). The cpu usage estimator ramps up 945 * quite quickly when the process is running (linearly), and decays 946 * away exponentially, at a rate which is proportionally slower when 947 * the system is busy. The basic principle is that the system will 948 * 90% forget that the process used a lot of CPU time in 5 * loadav 949 * seconds. This causes the system to favor processes which haven't 950 * run much recently, and to round-robin among other processes. 951 */ 952void 953schedclock(p) 954 struct proc *p; 955{ 956 957 p->p_cpticks++; 958 p->p_estcpu = ESTCPULIM(p->p_estcpu + 1); 959 if ((p->p_estcpu % INVERSE_ESTCPU_WEIGHT) == 0) { 960 resetpriority(p); 961 if (p->p_priority >= PUSER) 962 p->p_priority = p->p_usrpri; 963 } 964} 965