kern_synch.c revision 69645
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 69645 2000-12-06 00:33:58Z jhb $ 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/ipl.h> 48#include <sys/kernel.h> 49#include <sys/ktr.h> 50#include <sys/lock.h> 51#include <sys/mutex.h> 52#include <sys/signalvar.h> 53#include <sys/resourcevar.h> 54#include <sys/vmmeter.h> 55#include <sys/sysctl.h> 56#include <sys/sysproto.h> 57#include <vm/vm.h> 58#include <vm/vm_extern.h> 59#ifdef KTRACE 60#include <sys/uio.h> 61#include <sys/ktrace.h> 62#endif 63 64#include <machine/cpu.h> 65#include <machine/smp.h> 66 67static void sched_setup __P((void *dummy)); 68SYSINIT(sched_setup, SI_SUB_KICK_SCHEDULER, SI_ORDER_FIRST, sched_setup, NULL) 69 70u_char curpriority; 71int hogticks; 72int lbolt; 73int sched_quantum; /* Roundrobin scheduling quantum in ticks. */ 74 75static struct callout schedcpu_callout; 76static struct callout roundrobin_callout; 77 78static int curpriority_cmp __P((struct proc *p)); 79static void endtsleep __P((void *)); 80static void maybe_resched __P((struct proc *chk)); 81static void roundrobin __P((void *arg)); 82static void schedcpu __P((void *arg)); 83static void updatepri __P((struct proc *p)); 84 85static int 86sysctl_kern_quantum(SYSCTL_HANDLER_ARGS) 87{ 88 int error, new_val; 89 90 new_val = sched_quantum * tick; 91 error = sysctl_handle_int(oidp, &new_val, 0, req); 92 if (error != 0 || req->newptr == NULL) 93 return (error); 94 if (new_val < tick) 95 return (EINVAL); 96 sched_quantum = new_val / tick; 97 hogticks = 2 * sched_quantum; 98 return (0); 99} 100 101SYSCTL_PROC(_kern, OID_AUTO, quantum, CTLTYPE_INT|CTLFLAG_RW, 102 0, sizeof sched_quantum, sysctl_kern_quantum, "I", ""); 103 104/*- 105 * Compare priorities. Return: 106 * <0: priority of p < current priority 107 * 0: priority of p == current priority 108 * >0: priority of p > current priority 109 * The priorities are the normal priorities or the normal realtime priorities 110 * if p is on the same scheduler as curproc. Otherwise the process on the 111 * more realtimeish scheduler has lowest priority. As usual, a higher 112 * priority really means a lower priority. 113 */ 114static int 115curpriority_cmp(p) 116 struct proc *p; 117{ 118 int c_class, p_class; 119 120 c_class = RTP_PRIO_BASE(curproc->p_rtprio.type); 121 p_class = RTP_PRIO_BASE(p->p_rtprio.type); 122 if (p_class != c_class) 123 return (p_class - c_class); 124 if (p_class == RTP_PRIO_NORMAL) 125 return (((int)p->p_priority - (int)curpriority) / PPQ); 126 return ((int)p->p_rtprio.prio - (int)curproc->p_rtprio.prio); 127} 128 129/* 130 * Arrange to reschedule if necessary, taking the priorities and 131 * schedulers into account. 132 */ 133static void 134maybe_resched(chk) 135 struct proc *chk; 136{ 137 struct proc *p = curproc; /* XXX */ 138 139 /* 140 * XXX idle scheduler still broken because proccess stays on idle 141 * scheduler during waits (such as when getting FS locks). If a 142 * standard process becomes runaway cpu-bound, the system can lockup 143 * due to idle-scheduler processes in wakeup never getting any cpu. 144 */ 145 if (p == idleproc) { 146#if 0 147 need_resched(); 148#endif 149 } else if (chk == p) { 150 /* We may need to yield if our priority has been raised. */ 151 if (curpriority_cmp(chk) > 0) 152 need_resched(); 153 } else if (curpriority_cmp(chk) < 0) 154 need_resched(); 155} 156 157int 158roundrobin_interval(void) 159{ 160 return (sched_quantum); 161} 162 163/* 164 * Force switch among equal priority processes every 100ms. 165 */ 166/* ARGSUSED */ 167static void 168roundrobin(arg) 169 void *arg; 170{ 171#ifndef SMP 172 struct proc *p = curproc; /* XXX */ 173#endif 174 175#ifdef SMP 176 need_resched(); 177 forward_roundrobin(); 178#else 179 if (p == idleproc || RTP_PRIO_NEED_RR(p->p_rtprio.type)) 180 need_resched(); 181#endif 182 183 callout_reset(&roundrobin_callout, sched_quantum, roundrobin, NULL); 184} 185 186/* 187 * Constants for digital decay and forget: 188 * 90% of (p_estcpu) usage in 5 * loadav time 189 * 95% of (p_pctcpu) usage in 60 seconds (load insensitive) 190 * Note that, as ps(1) mentions, this can let percentages 191 * total over 100% (I've seen 137.9% for 3 processes). 192 * 193 * Note that schedclock() updates p_estcpu and p_cpticks asynchronously. 194 * 195 * We wish to decay away 90% of p_estcpu in (5 * loadavg) seconds. 196 * That is, the system wants to compute a value of decay such 197 * that the following for loop: 198 * for (i = 0; i < (5 * loadavg); i++) 199 * p_estcpu *= decay; 200 * will compute 201 * p_estcpu *= 0.1; 202 * for all values of loadavg: 203 * 204 * Mathematically this loop can be expressed by saying: 205 * decay ** (5 * loadavg) ~= .1 206 * 207 * The system computes decay as: 208 * decay = (2 * loadavg) / (2 * loadavg + 1) 209 * 210 * We wish to prove that the system's computation of decay 211 * will always fulfill the equation: 212 * decay ** (5 * loadavg) ~= .1 213 * 214 * If we compute b as: 215 * b = 2 * loadavg 216 * then 217 * decay = b / (b + 1) 218 * 219 * We now need to prove two things: 220 * 1) Given factor ** (5 * loadavg) ~= .1, prove factor == b/(b+1) 221 * 2) Given b/(b+1) ** power ~= .1, prove power == (5 * loadavg) 222 * 223 * Facts: 224 * For x close to zero, exp(x) =~ 1 + x, since 225 * exp(x) = 0! + x**1/1! + x**2/2! + ... . 226 * therefore exp(-1/b) =~ 1 - (1/b) = (b-1)/b. 227 * For x close to zero, ln(1+x) =~ x, since 228 * ln(1+x) = x - x**2/2 + x**3/3 - ... -1 < x < 1 229 * therefore ln(b/(b+1)) = ln(1 - 1/(b+1)) =~ -1/(b+1). 230 * ln(.1) =~ -2.30 231 * 232 * Proof of (1): 233 * Solve (factor)**(power) =~ .1 given power (5*loadav): 234 * solving for factor, 235 * ln(factor) =~ (-2.30/5*loadav), or 236 * factor =~ exp(-1/((5/2.30)*loadav)) =~ exp(-1/(2*loadav)) = 237 * exp(-1/b) =~ (b-1)/b =~ b/(b+1). QED 238 * 239 * Proof of (2): 240 * Solve (factor)**(power) =~ .1 given factor == (b/(b+1)): 241 * solving for power, 242 * power*ln(b/(b+1)) =~ -2.30, or 243 * power =~ 2.3 * (b + 1) = 4.6*loadav + 2.3 =~ 5*loadav. QED 244 * 245 * Actual power values for the implemented algorithm are as follows: 246 * loadav: 1 2 3 4 247 * power: 5.68 10.32 14.94 19.55 248 */ 249 250/* calculations for digital decay to forget 90% of usage in 5*loadav sec */ 251#define loadfactor(loadav) (2 * (loadav)) 252#define decay_cpu(loadfac, cpu) (((loadfac) * (cpu)) / ((loadfac) + FSCALE)) 253 254/* decay 95% of `p_pctcpu' in 60 seconds; see CCPU_SHIFT before changing */ 255static fixpt_t ccpu = 0.95122942450071400909 * FSCALE; /* exp(-1/20) */ 256SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, ""); 257 258/* kernel uses `FSCALE', userland (SHOULD) use kern.fscale */ 259static int fscale __unused = FSCALE; 260SYSCTL_INT(_kern, OID_AUTO, fscale, CTLFLAG_RD, 0, FSCALE, ""); 261 262/* 263 * If `ccpu' is not equal to `exp(-1/20)' and you still want to use the 264 * faster/more-accurate formula, you'll have to estimate CCPU_SHIFT below 265 * and possibly adjust FSHIFT in "param.h" so that (FSHIFT >= CCPU_SHIFT). 266 * 267 * To estimate CCPU_SHIFT for exp(-1/20), the following formula was used: 268 * 1 - exp(-1/20) ~= 0.0487 ~= 0.0488 == 1 (fixed pt, *11* bits). 269 * 270 * If you don't want to bother with the faster/more-accurate formula, you 271 * can set CCPU_SHIFT to (FSHIFT + 1) which will use a slower/less-accurate 272 * (more general) method of calculating the %age of CPU used by a process. 273 */ 274#define CCPU_SHIFT 11 275 276/* 277 * Recompute process priorities, every hz ticks. 278 * MP-safe, called without the Giant mutex. 279 */ 280/* ARGSUSED */ 281static void 282schedcpu(arg) 283 void *arg; 284{ 285 register fixpt_t loadfac = loadfactor(averunnable.ldavg[0]); 286 register struct proc *p; 287 register int realstathz, s; 288 289 realstathz = stathz ? stathz : hz; 290 lockmgr(&allproc_lock, LK_SHARED, NULL, CURPROC); 291 LIST_FOREACH(p, &allproc, p_list) { 292 /* 293 * Increment time in/out of memory and sleep time 294 * (if sleeping). We ignore overflow; with 16-bit int's 295 * (remember them?) overflow takes 45 days. 296 if (p->p_stat == SWAIT) 297 continue; 298 */ 299 mtx_enter(&sched_lock, MTX_SPIN); 300 p->p_swtime++; 301 if (p->p_stat == SSLEEP || p->p_stat == SSTOP) 302 p->p_slptime++; 303 p->p_pctcpu = (p->p_pctcpu * ccpu) >> FSHIFT; 304 /* 305 * If the process has slept the entire second, 306 * stop recalculating its priority until it wakes up. 307 */ 308 if (p->p_slptime > 1) { 309 mtx_exit(&sched_lock, MTX_SPIN); 310 continue; 311 } 312 313 /* 314 * prevent state changes and protect run queue 315 */ 316 s = splhigh(); 317 318 /* 319 * p_pctcpu is only for ps. 320 */ 321#if (FSHIFT >= CCPU_SHIFT) 322 p->p_pctcpu += (realstathz == 100)? 323 ((fixpt_t) p->p_cpticks) << (FSHIFT - CCPU_SHIFT): 324 100 * (((fixpt_t) p->p_cpticks) 325 << (FSHIFT - CCPU_SHIFT)) / realstathz; 326#else 327 p->p_pctcpu += ((FSCALE - ccpu) * 328 (p->p_cpticks * FSCALE / realstathz)) >> FSHIFT; 329#endif 330 p->p_cpticks = 0; 331 p->p_estcpu = decay_cpu(loadfac, p->p_estcpu); 332 resetpriority(p); 333 if (p->p_priority >= PUSER) { 334 if ((p != curproc) && 335#ifdef SMP 336 p->p_oncpu == 0xff && /* idle */ 337#endif 338 p->p_stat == SRUN && 339 (p->p_flag & P_INMEM) && 340 (p->p_priority / PPQ) != (p->p_usrpri / PPQ)) { 341 remrunqueue(p); 342 p->p_priority = p->p_usrpri; 343 setrunqueue(p); 344 } else 345 p->p_priority = p->p_usrpri; 346 } 347 mtx_exit(&sched_lock, MTX_SPIN); 348 splx(s); 349 } 350 lockmgr(&allproc_lock, LK_RELEASE, NULL, CURPROC); 351 vmmeter(); 352 wakeup((caddr_t)&lbolt); 353 callout_reset(&schedcpu_callout, hz, schedcpu, NULL); 354} 355 356/* 357 * Recalculate the priority of a process after it has slept for a while. 358 * For all load averages >= 1 and max p_estcpu of 255, sleeping for at 359 * least six times the loadfactor will decay p_estcpu to zero. 360 */ 361static void 362updatepri(p) 363 register struct proc *p; 364{ 365 register unsigned int newcpu = p->p_estcpu; 366 register fixpt_t loadfac = loadfactor(averunnable.ldavg[0]); 367 368 if (p->p_slptime > 5 * loadfac) 369 p->p_estcpu = 0; 370 else { 371 p->p_slptime--; /* the first time was done in schedcpu */ 372 while (newcpu && --p->p_slptime) 373 newcpu = decay_cpu(loadfac, newcpu); 374 p->p_estcpu = newcpu; 375 } 376 resetpriority(p); 377} 378 379/* 380 * We're only looking at 7 bits of the address; everything is 381 * aligned to 4, lots of things are aligned to greater powers 382 * of 2. Shift right by 8, i.e. drop the bottom 256 worth. 383 */ 384#define TABLESIZE 128 385static TAILQ_HEAD(slpquehead, proc) slpque[TABLESIZE]; 386#define LOOKUP(x) (((intptr_t)(x) >> 8) & (TABLESIZE - 1)) 387 388void 389sleepinit(void) 390{ 391 int i; 392 393 sched_quantum = hz/10; 394 hogticks = 2 * sched_quantum; 395 for (i = 0; i < TABLESIZE; i++) 396 TAILQ_INIT(&slpque[i]); 397} 398 399/* 400 * General sleep call. Suspends the current process until a wakeup is 401 * performed on the specified identifier. The process will then be made 402 * runnable with the specified priority. Sleeps at most timo/hz seconds 403 * (0 means no timeout). If pri includes PCATCH flag, signals are checked 404 * before and after sleeping, else signals are not checked. Returns 0 if 405 * awakened, EWOULDBLOCK if the timeout expires. If PCATCH is set and a 406 * signal needs to be delivered, ERESTART is returned if the current system 407 * call should be restarted if possible, and EINTR is returned if the system 408 * call should be interrupted by the signal (return EINTR). 409 * 410 * The mutex argument is exited before the caller is suspended, and 411 * entered before msleep returns. If priority includes the PDROP 412 * flag the mutex is not entered before returning. 413 */ 414int 415msleep(ident, mtx, priority, wmesg, timo) 416 void *ident; 417 struct mtx *mtx; 418 int priority, timo; 419 const char *wmesg; 420{ 421 struct proc *p = curproc; 422 int s, sig, catch = priority & PCATCH; 423 int rval = 0; 424 WITNESS_SAVE_DECL(mtx); 425 426#ifdef KTRACE 427 if (p && KTRPOINT(p, KTR_CSW)) 428 ktrcsw(p->p_tracep, 1, 0); 429#endif 430 WITNESS_SLEEP(0, mtx); 431 mtx_enter(&sched_lock, MTX_SPIN); 432 s = splhigh(); 433 if (cold || panicstr) { 434 /* 435 * After a panic, or during autoconfiguration, 436 * just give interrupts a chance, then just return; 437 * don't run any other procs or panic below, 438 * in case this is the idle process and already asleep. 439 */ 440 if (mtx != NULL && priority & PDROP) 441 mtx_exit(mtx, MTX_DEF | MTX_NOSWITCH); 442 mtx_exit(&sched_lock, MTX_SPIN); 443 splx(s); 444 return (0); 445 } 446 447 DROP_GIANT_NOSWITCH(); 448 449 if (mtx != NULL) { 450 mtx_assert(mtx, MA_OWNED | MA_NOTRECURSED); 451 WITNESS_SAVE(mtx, mtx); 452 mtx_exit(mtx, MTX_DEF | MTX_NOSWITCH); 453 if (priority & PDROP) 454 mtx = NULL; 455 } 456 457 KASSERT(p != NULL, ("msleep1")); 458 KASSERT(ident != NULL && p->p_stat == SRUN, ("msleep")); 459 /* 460 * Process may be sitting on a slpque if asleep() was called, remove 461 * it before re-adding. 462 */ 463 if (p->p_wchan != NULL) 464 unsleep(p); 465 466 p->p_wchan = ident; 467 p->p_wmesg = wmesg; 468 p->p_slptime = 0; 469 p->p_priority = priority & PRIMASK; 470 CTR4(KTR_PROC, "msleep: proc %p (pid %d, %s), schedlock %p", 471 p, p->p_pid, p->p_comm, (void *) sched_lock.mtx_lock); 472 TAILQ_INSERT_TAIL(&slpque[LOOKUP(ident)], p, p_slpq); 473 if (timo) 474 callout_reset(&p->p_slpcallout, timo, endtsleep, p); 475 /* 476 * We put ourselves on the sleep queue and start our timeout 477 * before calling CURSIG, as we could stop there, and a wakeup 478 * or a SIGCONT (or both) could occur while we were stopped. 479 * A SIGCONT would cause us to be marked as SSLEEP 480 * without resuming us, thus we must be ready for sleep 481 * when CURSIG is called. If the wakeup happens while we're 482 * stopped, p->p_wchan will be 0 upon return from CURSIG. 483 */ 484 if (catch) { 485 CTR4(KTR_PROC, 486 "msleep caught: proc %p (pid %d, %s), schedlock %p", 487 p, p->p_pid, p->p_comm, (void *) sched_lock.mtx_lock); 488 p->p_flag |= P_SINTR; 489 mtx_exit(&sched_lock, MTX_SPIN); 490 if ((sig = CURSIG(p))) { 491 mtx_enter(&sched_lock, MTX_SPIN); 492 if (p->p_wchan) 493 unsleep(p); 494 p->p_stat = SRUN; 495 goto resume; 496 } 497 mtx_enter(&sched_lock, MTX_SPIN); 498 if (p->p_wchan == 0) { 499 catch = 0; 500 goto resume; 501 } 502 } else 503 sig = 0; 504 p->p_stat = SSLEEP; 505 p->p_stats->p_ru.ru_nvcsw++; 506 mi_switch(); 507 CTR4(KTR_PROC, 508 "msleep resume: proc %p (pid %d, %s), schedlock %p", 509 p, p->p_pid, p->p_comm, (void *) sched_lock.mtx_lock); 510resume: 511 curpriority = p->p_usrpri; 512 splx(s); 513 p->p_flag &= ~P_SINTR; 514 if (p->p_flag & P_TIMEOUT) { 515 p->p_flag &= ~P_TIMEOUT; 516 if (sig == 0) { 517#ifdef KTRACE 518 if (KTRPOINT(p, KTR_CSW)) 519 ktrcsw(p->p_tracep, 0, 0); 520#endif 521 rval = EWOULDBLOCK; 522 mtx_exit(&sched_lock, MTX_SPIN); 523 goto out; 524 } 525 } else if (timo) 526 callout_stop(&p->p_slpcallout); 527 mtx_exit(&sched_lock, MTX_SPIN); 528 529 if (catch && (sig != 0 || (sig = CURSIG(p)))) { 530#ifdef KTRACE 531 if (KTRPOINT(p, KTR_CSW)) 532 ktrcsw(p->p_tracep, 0, 0); 533#endif 534 if (SIGISMEMBER(p->p_sigacts->ps_sigintr, sig)) 535 rval = EINTR; 536 else 537 rval = ERESTART; 538 goto out; 539 } 540out: 541#ifdef KTRACE 542 if (KTRPOINT(p, KTR_CSW)) 543 ktrcsw(p->p_tracep, 0, 0); 544#endif 545 PICKUP_GIANT(); 546 if (mtx != NULL) { 547 mtx_enter(mtx, MTX_DEF); 548 WITNESS_RESTORE(mtx, mtx); 549 } 550 return (rval); 551} 552 553/* 554 * asleep() - async sleep call. Place process on wait queue and return 555 * immediately without blocking. The process stays runnable until mawait() 556 * is called. If ident is NULL, remove process from wait queue if it is still 557 * on one. 558 * 559 * Only the most recent sleep condition is effective when making successive 560 * calls to asleep() or when calling msleep(). 561 * 562 * The timeout, if any, is not initiated until mawait() is called. The sleep 563 * priority, signal, and timeout is specified in the asleep() call but may be 564 * overriden in the mawait() call. 565 * 566 * <<<<<<<< EXPERIMENTAL, UNTESTED >>>>>>>>>> 567 */ 568 569int 570asleep(void *ident, int priority, const char *wmesg, int timo) 571{ 572 struct proc *p = curproc; 573 int s; 574 575 /* 576 * obtain sched_lock while manipulating sleep structures and slpque. 577 * 578 * Remove preexisting wait condition (if any) and place process 579 * on appropriate slpque, but do not put process to sleep. 580 */ 581 582 s = splhigh(); 583 mtx_enter(&sched_lock, MTX_SPIN); 584 585 if (p->p_wchan != NULL) 586 unsleep(p); 587 588 if (ident) { 589 p->p_wchan = ident; 590 p->p_wmesg = wmesg; 591 p->p_slptime = 0; 592 p->p_asleep.as_priority = priority; 593 p->p_asleep.as_timo = timo; 594 TAILQ_INSERT_TAIL(&slpque[LOOKUP(ident)], p, p_slpq); 595 } 596 597 mtx_exit(&sched_lock, MTX_SPIN); 598 splx(s); 599 600 return(0); 601} 602 603/* 604 * mawait() - wait for async condition to occur. The process blocks until 605 * wakeup() is called on the most recent asleep() address. If wakeup is called 606 * prior to mawait(), mawait() winds up being a NOP. 607 * 608 * If mawait() is called more then once (without an intervening asleep() call), 609 * mawait() is still effectively a NOP but it calls mi_switch() to give other 610 * processes some cpu before returning. The process is left runnable. 611 * 612 * <<<<<<<< EXPERIMENTAL, UNTESTED >>>>>>>>>> 613 */ 614 615int 616mawait(struct mtx *mtx, int priority, int timo) 617{ 618 struct proc *p = curproc; 619 int rval = 0; 620 int s; 621 WITNESS_SAVE_DECL(mtx); 622 623 WITNESS_SLEEP(0, mtx); 624 mtx_enter(&sched_lock, MTX_SPIN); 625 DROP_GIANT_NOSWITCH(); 626 if (mtx != NULL) { 627 mtx_assert(mtx, MA_OWNED | MA_NOTRECURSED); 628 WITNESS_SAVE(mtx, mtx); 629 mtx_exit(mtx, MTX_DEF | MTX_NOSWITCH); 630 if (priority & PDROP) 631 mtx = NULL; 632 } 633 634 s = splhigh(); 635 636 if (p->p_wchan != NULL) { 637 int sig; 638 int catch; 639 640 /* 641 * The call to mawait() can override defaults specified in 642 * the original asleep(). 643 */ 644 if (priority < 0) 645 priority = p->p_asleep.as_priority; 646 if (timo < 0) 647 timo = p->p_asleep.as_timo; 648 649 /* 650 * Install timeout 651 */ 652 653 if (timo) 654 callout_reset(&p->p_slpcallout, timo, endtsleep, p); 655 656 sig = 0; 657 catch = priority & PCATCH; 658 659 if (catch) { 660 p->p_flag |= P_SINTR; 661 mtx_exit(&sched_lock, MTX_SPIN); 662 if ((sig = CURSIG(p))) { 663 mtx_enter(&sched_lock, MTX_SPIN); 664 if (p->p_wchan) 665 unsleep(p); 666 p->p_stat = SRUN; 667 goto resume; 668 } 669 mtx_enter(&sched_lock, MTX_SPIN); 670 if (p->p_wchan == NULL) { 671 catch = 0; 672 goto resume; 673 } 674 } 675 p->p_stat = SSLEEP; 676 p->p_stats->p_ru.ru_nvcsw++; 677 mi_switch(); 678resume: 679 curpriority = p->p_usrpri; 680 681 splx(s); 682 p->p_flag &= ~P_SINTR; 683 if (p->p_flag & P_TIMEOUT) { 684 p->p_flag &= ~P_TIMEOUT; 685 if (sig == 0) { 686#ifdef KTRACE 687 if (KTRPOINT(p, KTR_CSW)) 688 ktrcsw(p->p_tracep, 0, 0); 689#endif 690 rval = EWOULDBLOCK; 691 mtx_exit(&sched_lock, MTX_SPIN); 692 goto out; 693 } 694 } else if (timo) 695 callout_stop(&p->p_slpcallout); 696 mtx_exit(&sched_lock, MTX_SPIN); 697 698 if (catch && (sig != 0 || (sig = CURSIG(p)))) { 699#ifdef KTRACE 700 if (KTRPOINT(p, KTR_CSW)) 701 ktrcsw(p->p_tracep, 0, 0); 702#endif 703 if (SIGISMEMBER(p->p_sigacts->ps_sigintr, sig)) 704 rval = EINTR; 705 else 706 rval = ERESTART; 707 goto out; 708 } 709#ifdef KTRACE 710 if (KTRPOINT(p, KTR_CSW)) 711 ktrcsw(p->p_tracep, 0, 0); 712#endif 713 } else { 714 /* 715 * If as_priority is 0, mawait() has been called without an 716 * intervening asleep(). We are still effectively a NOP, 717 * but we call mi_switch() for safety. 718 */ 719 720 if (p->p_asleep.as_priority == 0) { 721 p->p_stats->p_ru.ru_nvcsw++; 722 mi_switch(); 723 } 724 mtx_exit(&sched_lock, MTX_SPIN); 725 splx(s); 726 } 727 728 /* 729 * clear p_asleep.as_priority as an indication that mawait() has been 730 * called. If mawait() is called again without an intervening asleep(), 731 * mawait() is still effectively a NOP but the above mi_switch() code 732 * is triggered as a safety. 733 */ 734 p->p_asleep.as_priority = 0; 735 736out: 737 PICKUP_GIANT(); 738 if (mtx != NULL) { 739 mtx_enter(mtx, MTX_DEF); 740 WITNESS_RESTORE(mtx, mtx); 741 } 742 return (rval); 743} 744 745/* 746 * Implement timeout for msleep or asleep()/mawait() 747 * 748 * If process hasn't been awakened (wchan non-zero), 749 * set timeout flag and undo the sleep. If proc 750 * is stopped, just unsleep so it will remain stopped. 751 * MP-safe, called without the Giant mutex. 752 */ 753static void 754endtsleep(arg) 755 void *arg; 756{ 757 register struct proc *p; 758 int s; 759 760 p = (struct proc *)arg; 761 CTR4(KTR_PROC, 762 "endtsleep: proc %p (pid %d, %s), schedlock %p", 763 p, p->p_pid, p->p_comm, (void *) sched_lock.mtx_lock); 764 s = splhigh(); 765 mtx_enter(&sched_lock, MTX_SPIN); 766 if (p->p_wchan) { 767 if (p->p_stat == SSLEEP) 768 setrunnable(p); 769 else 770 unsleep(p); 771 p->p_flag |= P_TIMEOUT; 772 } 773 mtx_exit(&sched_lock, MTX_SPIN); 774 splx(s); 775} 776 777/* 778 * Remove a process from its wait queue 779 */ 780void 781unsleep(p) 782 register struct proc *p; 783{ 784 int s; 785 786 s = splhigh(); 787 mtx_enter(&sched_lock, MTX_SPIN); 788 if (p->p_wchan) { 789 TAILQ_REMOVE(&slpque[LOOKUP(p->p_wchan)], p, p_slpq); 790 p->p_wchan = 0; 791 } 792 mtx_exit(&sched_lock, MTX_SPIN); 793 splx(s); 794} 795 796/* 797 * Make all processes sleeping on the specified identifier runnable. 798 */ 799void 800wakeup(ident) 801 register void *ident; 802{ 803 register struct slpquehead *qp; 804 register struct proc *p; 805 int s; 806 807 s = splhigh(); 808 mtx_enter(&sched_lock, MTX_SPIN); 809 qp = &slpque[LOOKUP(ident)]; 810restart: 811 TAILQ_FOREACH(p, qp, p_slpq) { 812 if (p->p_wchan == ident) { 813 TAILQ_REMOVE(qp, p, p_slpq); 814 p->p_wchan = 0; 815 if (p->p_stat == SSLEEP) { 816 /* OPTIMIZED EXPANSION OF setrunnable(p); */ 817 CTR4(KTR_PROC, 818 "wakeup: proc %p (pid %d, %s), schedlock %p", 819 p, p->p_pid, p->p_comm, (void *) sched_lock.mtx_lock); 820 if (p->p_slptime > 1) 821 updatepri(p); 822 p->p_slptime = 0; 823 p->p_stat = SRUN; 824 if (p->p_flag & P_INMEM) { 825 setrunqueue(p); 826 maybe_resched(p); 827 } else { 828 p->p_flag |= P_SWAPINREQ; 829 wakeup((caddr_t)&proc0); 830 } 831 /* END INLINE EXPANSION */ 832 goto restart; 833 } 834 } 835 } 836 mtx_exit(&sched_lock, MTX_SPIN); 837 splx(s); 838} 839 840/* 841 * Make a process sleeping on the specified identifier runnable. 842 * May wake more than one process if a target process is currently 843 * swapped out. 844 */ 845void 846wakeup_one(ident) 847 register void *ident; 848{ 849 register struct slpquehead *qp; 850 register struct proc *p; 851 int s; 852 853 s = splhigh(); 854 mtx_enter(&sched_lock, MTX_SPIN); 855 qp = &slpque[LOOKUP(ident)]; 856 857 TAILQ_FOREACH(p, qp, p_slpq) { 858 if (p->p_wchan == ident) { 859 TAILQ_REMOVE(qp, p, p_slpq); 860 p->p_wchan = 0; 861 if (p->p_stat == SSLEEP) { 862 /* OPTIMIZED EXPANSION OF setrunnable(p); */ 863 CTR4(KTR_PROC, 864 "wakeup1: proc %p (pid %d, %s), schedlock %p", 865 p, p->p_pid, p->p_comm, (void *) sched_lock.mtx_lock); 866 if (p->p_slptime > 1) 867 updatepri(p); 868 p->p_slptime = 0; 869 p->p_stat = SRUN; 870 if (p->p_flag & P_INMEM) { 871 setrunqueue(p); 872 maybe_resched(p); 873 break; 874 } else { 875 p->p_flag |= P_SWAPINREQ; 876 wakeup((caddr_t)&proc0); 877 } 878 /* END INLINE EXPANSION */ 879 } 880 } 881 } 882 mtx_exit(&sched_lock, MTX_SPIN); 883 splx(s); 884} 885 886/* 887 * The machine independent parts of mi_switch(). 888 * Must be called at splstatclock() or higher. 889 */ 890void 891mi_switch() 892{ 893 struct timeval new_switchtime; 894 register struct proc *p = curproc; /* XXX */ 895 register struct rlimit *rlim; 896 int x; 897 898 /* 899 * XXX this spl is almost unnecessary. It is partly to allow for 900 * sloppy callers that don't do it (issignal() via CURSIG() is the 901 * main offender). It is partly to work around a bug in the i386 902 * cpu_switch() (the ipl is not preserved). We ran for years 903 * without it. I think there was only a interrupt latency problem. 904 * The main caller, msleep(), does an splx() a couple of instructions 905 * after calling here. The buggy caller, issignal(), usually calls 906 * here at spl0() and sometimes returns at splhigh(). The process 907 * then runs for a little too long at splhigh(). The ipl gets fixed 908 * when the process returns to user mode (or earlier). 909 * 910 * It would probably be better to always call here at spl0(). Callers 911 * are prepared to give up control to another process, so they must 912 * be prepared to be interrupted. The clock stuff here may not 913 * actually need splstatclock(). 914 */ 915 x = splstatclock(); 916 917 mtx_assert(&sched_lock, MA_OWNED); 918 919#ifdef SIMPLELOCK_DEBUG 920 if (p->p_simple_locks) 921 printf("sleep: holding simple lock\n"); 922#endif 923 /* 924 * Compute the amount of time during which the current 925 * process was running, and add that to its total so far. 926 */ 927 microuptime(&new_switchtime); 928 if (timevalcmp(&new_switchtime, &switchtime, <)) { 929 printf("microuptime() went backwards (%ld.%06ld -> %ld.%06ld)\n", 930 switchtime.tv_sec, switchtime.tv_usec, 931 new_switchtime.tv_sec, new_switchtime.tv_usec); 932 new_switchtime = switchtime; 933 } else { 934 p->p_runtime += (new_switchtime.tv_usec - switchtime.tv_usec) + 935 (new_switchtime.tv_sec - switchtime.tv_sec) * (int64_t)1000000; 936 } 937 938 /* 939 * Check if the process exceeds its cpu resource allocation. 940 * If over max, kill it. 941 * 942 * XXX drop sched_lock, pickup Giant 943 */ 944 if (p->p_stat != SZOMB && p->p_limit->p_cpulimit != RLIM_INFINITY && 945 p->p_runtime > p->p_limit->p_cpulimit) { 946 rlim = &p->p_rlimit[RLIMIT_CPU]; 947 if (p->p_runtime / (rlim_t)1000000 >= rlim->rlim_max) { 948 killproc(p, "exceeded maximum CPU limit"); 949 } else { 950 psignal(p, SIGXCPU); 951 if (rlim->rlim_cur < rlim->rlim_max) { 952 /* XXX: we should make a private copy */ 953 rlim->rlim_cur += 5; 954 } 955 } 956 } 957 958 /* 959 * Pick a new current process and record its start time. 960 */ 961 cnt.v_swtch++; 962 switchtime = new_switchtime; 963 CTR4(KTR_PROC, "mi_switch: old proc %p (pid %d, %s), schedlock %p", 964 p, p->p_pid, p->p_comm, (void *) sched_lock.mtx_lock); 965 cpu_switch(); 966 CTR4(KTR_PROC, "mi_switch: new proc %p (pid %d, %s), schedlock %p", 967 p, p->p_pid, p->p_comm, (void *) sched_lock.mtx_lock); 968 if (switchtime.tv_sec == 0) 969 microuptime(&switchtime); 970 switchticks = ticks; 971 splx(x); 972} 973 974/* 975 * Change process state to be runnable, 976 * placing it on the run queue if it is in memory, 977 * and awakening the swapper if it isn't in memory. 978 */ 979void 980setrunnable(p) 981 register struct proc *p; 982{ 983 register int s; 984 985 s = splhigh(); 986 mtx_enter(&sched_lock, MTX_SPIN); 987 switch (p->p_stat) { 988 case 0: 989 case SRUN: 990 case SZOMB: 991 case SWAIT: 992 default: 993 panic("setrunnable"); 994 case SSTOP: 995 case SSLEEP: 996 unsleep(p); /* e.g. when sending signals */ 997 break; 998 999 case SIDL: 1000 break; 1001 } 1002 p->p_stat = SRUN; 1003 if (p->p_flag & P_INMEM) 1004 setrunqueue(p); 1005 splx(s); 1006 if (p->p_slptime > 1) 1007 updatepri(p); 1008 p->p_slptime = 0; 1009 if ((p->p_flag & P_INMEM) == 0) { 1010 p->p_flag |= P_SWAPINREQ; 1011 wakeup((caddr_t)&proc0); 1012 } 1013 else 1014 maybe_resched(p); 1015 mtx_exit(&sched_lock, MTX_SPIN); 1016} 1017 1018/* 1019 * Compute the priority of a process when running in user mode. 1020 * Arrange to reschedule if the resulting priority is better 1021 * than that of the current process. 1022 */ 1023void 1024resetpriority(p) 1025 register struct proc *p; 1026{ 1027 register unsigned int newpriority; 1028 1029 mtx_enter(&sched_lock, MTX_SPIN); 1030 if (p->p_rtprio.type == RTP_PRIO_NORMAL) { 1031 newpriority = PUSER + p->p_estcpu / INVERSE_ESTCPU_WEIGHT + 1032 NICE_WEIGHT * (p->p_nice - PRIO_MIN); 1033 newpriority = min(newpriority, MAXPRI); 1034 p->p_usrpri = newpriority; 1035 } 1036 maybe_resched(p); 1037 mtx_exit(&sched_lock, MTX_SPIN); 1038} 1039 1040/* ARGSUSED */ 1041static void 1042sched_setup(dummy) 1043 void *dummy; 1044{ 1045 1046 callout_init(&schedcpu_callout, 1); 1047 callout_init(&roundrobin_callout, 0); 1048 1049 /* Kick off timeout driven events by calling first time. */ 1050 roundrobin(NULL); 1051 schedcpu(NULL); 1052} 1053 1054/* 1055 * We adjust the priority of the current process. The priority of 1056 * a process gets worse as it accumulates CPU time. The cpu usage 1057 * estimator (p_estcpu) is increased here. resetpriority() will 1058 * compute a different priority each time p_estcpu increases by 1059 * INVERSE_ESTCPU_WEIGHT 1060 * (until MAXPRI is reached). The cpu usage estimator ramps up 1061 * quite quickly when the process is running (linearly), and decays 1062 * away exponentially, at a rate which is proportionally slower when 1063 * the system is busy. The basic principle is that the system will 1064 * 90% forget that the process used a lot of CPU time in 5 * loadav 1065 * seconds. This causes the system to favor processes which haven't 1066 * run much recently, and to round-robin among other processes. 1067 */ 1068void 1069schedclock(p) 1070 struct proc *p; 1071{ 1072 1073 p->p_cpticks++; 1074 p->p_estcpu = ESTCPULIM(p->p_estcpu + 1); 1075 if ((p->p_estcpu % INVERSE_ESTCPU_WEIGHT) == 0) { 1076 resetpriority(p); 1077 if (p->p_priority >= PUSER) 1078 p->p_priority = p->p_usrpri; 1079 } 1080} 1081 1082/* 1083 * General purpose yield system call 1084 */ 1085int 1086yield(struct proc *p, struct yield_args *uap) 1087{ 1088 int s; 1089 1090 p->p_retval[0] = 0; 1091 1092 s = splhigh(); 1093 mtx_enter(&sched_lock, MTX_SPIN); 1094 DROP_GIANT_NOSWITCH(); 1095 p->p_priority = MAXPRI; 1096 setrunqueue(p); 1097 p->p_stats->p_ru.ru_nvcsw++; 1098 mi_switch(); 1099 mtx_exit(&sched_lock, MTX_SPIN); 1100 PICKUP_GIANT(); 1101 splx(s); 1102 1103 return (0); 1104} 1105