kern_synch.c revision 74927
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 74927 2001-03-28 11:52:56Z jhb $ 40 */ 41 42#include "opt_ktrace.h" 43 44#include <sys/param.h> 45#include <sys/systm.h> 46#include <sys/condvar.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/proc.h> 53#include <sys/resourcevar.h> 54#include <sys/signalvar.h> 55#include <sys/sx.h> 56#include <sys/sysctl.h> 57#include <sys/sysproto.h> 58#include <sys/vmmeter.h> 59#include <vm/vm.h> 60#include <vm/vm_extern.h> 61#ifdef KTRACE 62#include <sys/uio.h> 63#include <sys/ktrace.h> 64#endif 65 66#include <machine/cpu.h> 67#include <machine/smp.h> 68 69static void sched_setup __P((void *dummy)); 70SYSINIT(sched_setup, SI_SUB_KICK_SCHEDULER, SI_ORDER_FIRST, sched_setup, NULL) 71 72int hogticks; 73int lbolt; 74int sched_quantum; /* Roundrobin scheduling quantum in ticks. */ 75 76static struct callout schedcpu_callout; 77static struct callout roundrobin_callout; 78 79static void endtsleep __P((void *)); 80static void roundrobin __P((void *arg)); 81static void schedcpu __P((void *arg)); 82 83static int 84sysctl_kern_quantum(SYSCTL_HANDLER_ARGS) 85{ 86 int error, new_val; 87 88 new_val = sched_quantum * tick; 89 error = sysctl_handle_int(oidp, &new_val, 0, req); 90 if (error != 0 || req->newptr == NULL) 91 return (error); 92 if (new_val < tick) 93 return (EINVAL); 94 sched_quantum = new_val / tick; 95 hogticks = 2 * sched_quantum; 96 return (0); 97} 98 99SYSCTL_PROC(_kern, OID_AUTO, quantum, CTLTYPE_INT|CTLFLAG_RW, 100 0, sizeof sched_quantum, sysctl_kern_quantum, "I", ""); 101 102/* 103 * Arrange to reschedule if necessary, taking the priorities and 104 * schedulers into account. 105 */ 106void 107maybe_resched(p) 108 struct proc *p; 109{ 110 111 mtx_assert(&sched_lock, MA_OWNED); 112 if (p->p_pri.pri_level < curproc->p_pri.pri_level) 113 need_resched(); 114} 115 116int 117roundrobin_interval(void) 118{ 119 return (sched_quantum); 120} 121 122/* 123 * Force switch among equal priority processes every 100ms. 124 */ 125/* ARGSUSED */ 126static void 127roundrobin(arg) 128 void *arg; 129{ 130 131 mtx_lock_spin(&sched_lock); 132 need_resched(); 133 mtx_unlock_spin(&sched_lock); 134#ifdef SMP 135 forward_roundrobin(); 136#endif 137 138 callout_reset(&roundrobin_callout, sched_quantum, roundrobin, NULL); 139} 140 141/* 142 * Constants for digital decay and forget: 143 * 90% of (p_estcpu) usage in 5 * loadav time 144 * 95% of (p_pctcpu) usage in 60 seconds (load insensitive) 145 * Note that, as ps(1) mentions, this can let percentages 146 * total over 100% (I've seen 137.9% for 3 processes). 147 * 148 * Note that schedclock() updates p_estcpu and p_cpticks asynchronously. 149 * 150 * We wish to decay away 90% of p_estcpu in (5 * loadavg) seconds. 151 * That is, the system wants to compute a value of decay such 152 * that the following for loop: 153 * for (i = 0; i < (5 * loadavg); i++) 154 * p_estcpu *= decay; 155 * will compute 156 * p_estcpu *= 0.1; 157 * for all values of loadavg: 158 * 159 * Mathematically this loop can be expressed by saying: 160 * decay ** (5 * loadavg) ~= .1 161 * 162 * The system computes decay as: 163 * decay = (2 * loadavg) / (2 * loadavg + 1) 164 * 165 * We wish to prove that the system's computation of decay 166 * will always fulfill the equation: 167 * decay ** (5 * loadavg) ~= .1 168 * 169 * If we compute b as: 170 * b = 2 * loadavg 171 * then 172 * decay = b / (b + 1) 173 * 174 * We now need to prove two things: 175 * 1) Given factor ** (5 * loadavg) ~= .1, prove factor == b/(b+1) 176 * 2) Given b/(b+1) ** power ~= .1, prove power == (5 * loadavg) 177 * 178 * Facts: 179 * For x close to zero, exp(x) =~ 1 + x, since 180 * exp(x) = 0! + x**1/1! + x**2/2! + ... . 181 * therefore exp(-1/b) =~ 1 - (1/b) = (b-1)/b. 182 * For x close to zero, ln(1+x) =~ x, since 183 * ln(1+x) = x - x**2/2 + x**3/3 - ... -1 < x < 1 184 * therefore ln(b/(b+1)) = ln(1 - 1/(b+1)) =~ -1/(b+1). 185 * ln(.1) =~ -2.30 186 * 187 * Proof of (1): 188 * Solve (factor)**(power) =~ .1 given power (5*loadav): 189 * solving for factor, 190 * ln(factor) =~ (-2.30/5*loadav), or 191 * factor =~ exp(-1/((5/2.30)*loadav)) =~ exp(-1/(2*loadav)) = 192 * exp(-1/b) =~ (b-1)/b =~ b/(b+1). QED 193 * 194 * Proof of (2): 195 * Solve (factor)**(power) =~ .1 given factor == (b/(b+1)): 196 * solving for power, 197 * power*ln(b/(b+1)) =~ -2.30, or 198 * power =~ 2.3 * (b + 1) = 4.6*loadav + 2.3 =~ 5*loadav. QED 199 * 200 * Actual power values for the implemented algorithm are as follows: 201 * loadav: 1 2 3 4 202 * power: 5.68 10.32 14.94 19.55 203 */ 204 205/* calculations for digital decay to forget 90% of usage in 5*loadav sec */ 206#define loadfactor(loadav) (2 * (loadav)) 207#define decay_cpu(loadfac, cpu) (((loadfac) * (cpu)) / ((loadfac) + FSCALE)) 208 209/* decay 95% of `p_pctcpu' in 60 seconds; see CCPU_SHIFT before changing */ 210static fixpt_t ccpu = 0.95122942450071400909 * FSCALE; /* exp(-1/20) */ 211SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, ""); 212 213/* kernel uses `FSCALE', userland (SHOULD) use kern.fscale */ 214static int fscale __unused = FSCALE; 215SYSCTL_INT(_kern, OID_AUTO, fscale, CTLFLAG_RD, 0, FSCALE, ""); 216 217/* 218 * If `ccpu' is not equal to `exp(-1/20)' and you still want to use the 219 * faster/more-accurate formula, you'll have to estimate CCPU_SHIFT below 220 * and possibly adjust FSHIFT in "param.h" so that (FSHIFT >= CCPU_SHIFT). 221 * 222 * To estimate CCPU_SHIFT for exp(-1/20), the following formula was used: 223 * 1 - exp(-1/20) ~= 0.0487 ~= 0.0488 == 1 (fixed pt, *11* bits). 224 * 225 * If you don't want to bother with the faster/more-accurate formula, you 226 * can set CCPU_SHIFT to (FSHIFT + 1) which will use a slower/less-accurate 227 * (more general) method of calculating the %age of CPU used by a process. 228 */ 229#define CCPU_SHIFT 11 230 231/* 232 * Recompute process priorities, every hz ticks. 233 * MP-safe, called without the Giant mutex. 234 */ 235/* ARGSUSED */ 236static void 237schedcpu(arg) 238 void *arg; 239{ 240 register fixpt_t loadfac = loadfactor(averunnable.ldavg[0]); 241 register struct proc *p; 242 register int realstathz, s; 243 244 realstathz = stathz ? stathz : hz; 245 sx_slock(&allproc_lock); 246 LIST_FOREACH(p, &allproc, p_list) { 247 /* 248 * Increment time in/out of memory and sleep time 249 * (if sleeping). We ignore overflow; with 16-bit int's 250 * (remember them?) overflow takes 45 days. 251 if (p->p_stat == SWAIT) 252 continue; 253 */ 254 mtx_lock_spin(&sched_lock); 255 p->p_swtime++; 256 if (p->p_stat == SSLEEP || p->p_stat == SSTOP) 257 p->p_slptime++; 258 p->p_pctcpu = (p->p_pctcpu * ccpu) >> FSHIFT; 259 /* 260 * If the process has slept the entire second, 261 * stop recalculating its priority until it wakes up. 262 */ 263 if (p->p_slptime > 1) { 264 mtx_unlock_spin(&sched_lock); 265 continue; 266 } 267 268 /* 269 * prevent state changes and protect run queue 270 */ 271 s = splhigh(); 272 273 /* 274 * p_pctcpu is only for ps. 275 */ 276#if (FSHIFT >= CCPU_SHIFT) 277 p->p_pctcpu += (realstathz == 100)? 278 ((fixpt_t) p->p_cpticks) << (FSHIFT - CCPU_SHIFT): 279 100 * (((fixpt_t) p->p_cpticks) 280 << (FSHIFT - CCPU_SHIFT)) / realstathz; 281#else 282 p->p_pctcpu += ((FSCALE - ccpu) * 283 (p->p_cpticks * FSCALE / realstathz)) >> FSHIFT; 284#endif 285 p->p_cpticks = 0; 286 p->p_estcpu = decay_cpu(loadfac, p->p_estcpu); 287 resetpriority(p); 288 if (p->p_pri.pri_level >= PUSER) { 289 if ((p != curproc) && 290#ifdef SMP 291 p->p_oncpu == NOCPU && /* idle */ 292#endif 293 p->p_stat == SRUN && 294 (p->p_sflag & PS_INMEM) && 295 (p->p_pri.pri_level / RQ_PPQ) != 296 (p->p_pri.pri_user / RQ_PPQ)) { 297 remrunqueue(p); 298 p->p_pri.pri_level = p->p_pri.pri_user; 299 setrunqueue(p); 300 } else 301 p->p_pri.pri_level = p->p_pri.pri_user; 302 } 303 mtx_unlock_spin(&sched_lock); 304 splx(s); 305 } 306 sx_sunlock(&allproc_lock); 307 vmmeter(); 308 wakeup((caddr_t)&lbolt); 309 callout_reset(&schedcpu_callout, hz, schedcpu, NULL); 310} 311 312/* 313 * Recalculate the priority of a process after it has slept for a while. 314 * For all load averages >= 1 and max p_estcpu of 255, sleeping for at 315 * least six times the loadfactor will decay p_estcpu to zero. 316 */ 317void 318updatepri(p) 319 register struct proc *p; 320{ 321 register unsigned int newcpu = p->p_estcpu; 322 register fixpt_t loadfac = loadfactor(averunnable.ldavg[0]); 323 324 if (p->p_slptime > 5 * loadfac) 325 p->p_estcpu = 0; 326 else { 327 p->p_slptime--; /* the first time was done in schedcpu */ 328 while (newcpu && --p->p_slptime) 329 newcpu = decay_cpu(loadfac, newcpu); 330 p->p_estcpu = newcpu; 331 } 332 resetpriority(p); 333} 334 335/* 336 * We're only looking at 7 bits of the address; everything is 337 * aligned to 4, lots of things are aligned to greater powers 338 * of 2. Shift right by 8, i.e. drop the bottom 256 worth. 339 */ 340#define TABLESIZE 128 341static TAILQ_HEAD(slpquehead, proc) slpque[TABLESIZE]; 342#define LOOKUP(x) (((intptr_t)(x) >> 8) & (TABLESIZE - 1)) 343 344void 345sleepinit(void) 346{ 347 int i; 348 349 sched_quantum = hz/10; 350 hogticks = 2 * sched_quantum; 351 for (i = 0; i < TABLESIZE; i++) 352 TAILQ_INIT(&slpque[i]); 353} 354 355/* 356 * General sleep call. Suspends the current process until a wakeup is 357 * performed on the specified identifier. The process will then be made 358 * runnable with the specified priority. Sleeps at most timo/hz seconds 359 * (0 means no timeout). If pri includes PCATCH flag, signals are checked 360 * before and after sleeping, else signals are not checked. Returns 0 if 361 * awakened, EWOULDBLOCK if the timeout expires. If PCATCH is set and a 362 * signal needs to be delivered, ERESTART is returned if the current system 363 * call should be restarted if possible, and EINTR is returned if the system 364 * call should be interrupted by the signal (return EINTR). 365 * 366 * The mutex argument is exited before the caller is suspended, and 367 * entered before msleep returns. If priority includes the PDROP 368 * flag the mutex is not entered before returning. 369 */ 370int 371msleep(ident, mtx, priority, wmesg, timo) 372 void *ident; 373 struct mtx *mtx; 374 int priority, timo; 375 const char *wmesg; 376{ 377 struct proc *p = curproc; 378 int sig, catch = priority & PCATCH; 379 int rval = 0; 380 WITNESS_SAVE_DECL(mtx); 381 382#ifdef KTRACE 383 if (p && KTRPOINT(p, KTR_CSW)) 384 ktrcsw(p->p_tracep, 1, 0); 385#endif 386 WITNESS_SLEEP(0, &mtx->mtx_object); 387 mtx_lock_spin(&sched_lock); 388 if (cold || panicstr) { 389 /* 390 * After a panic, or during autoconfiguration, 391 * just give interrupts a chance, then just return; 392 * don't run any other procs or panic below, 393 * in case this is the idle process and already asleep. 394 */ 395 if (mtx != NULL && priority & PDROP) 396 mtx_unlock_flags(mtx, MTX_NOSWITCH); 397 mtx_unlock_spin(&sched_lock); 398 return (0); 399 } 400 401 DROP_GIANT_NOSWITCH(); 402 403 if (mtx != NULL) { 404 mtx_assert(mtx, MA_OWNED | MA_NOTRECURSED); 405 WITNESS_SAVE(&mtx->mtx_object, mtx); 406 mtx_unlock_flags(mtx, MTX_NOSWITCH); 407 if (priority & PDROP) 408 mtx = NULL; 409 } 410 411 KASSERT(p != NULL, ("msleep1")); 412 KASSERT(ident != NULL && p->p_stat == SRUN, ("msleep")); 413 /* 414 * Process may be sitting on a slpque if asleep() was called, remove 415 * it before re-adding. 416 */ 417 if (p->p_wchan != NULL) 418 unsleep(p); 419 420 p->p_wchan = ident; 421 p->p_wmesg = wmesg; 422 p->p_slptime = 0; 423 p->p_pri.pri_level = priority & PRIMASK; 424 CTR4(KTR_PROC, "msleep: proc %p (pid %d, %s), schedlock %p", 425 p, p->p_pid, p->p_comm, (void *) sched_lock.mtx_lock); 426 TAILQ_INSERT_TAIL(&slpque[LOOKUP(ident)], p, p_slpq); 427 if (timo) 428 callout_reset(&p->p_slpcallout, timo, endtsleep, p); 429 /* 430 * We put ourselves on the sleep queue and start our timeout 431 * before calling CURSIG, as we could stop there, and a wakeup 432 * or a SIGCONT (or both) could occur while we were stopped. 433 * A SIGCONT would cause us to be marked as SSLEEP 434 * without resuming us, thus we must be ready for sleep 435 * when CURSIG is called. If the wakeup happens while we're 436 * stopped, p->p_wchan will be 0 upon return from CURSIG. 437 */ 438 if (catch) { 439 CTR4(KTR_PROC, 440 "msleep caught: proc %p (pid %d, %s), schedlock %p", 441 p, p->p_pid, p->p_comm, (void *) sched_lock.mtx_lock); 442 p->p_sflag |= PS_SINTR; 443 mtx_unlock_spin(&sched_lock); 444 if ((sig = CURSIG(p))) { 445 mtx_lock_spin(&sched_lock); 446 if (p->p_wchan) 447 unsleep(p); 448 p->p_stat = SRUN; 449 goto resume; 450 } 451 mtx_lock_spin(&sched_lock); 452 if (p->p_wchan == NULL) { 453 catch = 0; 454 goto resume; 455 } 456 } else 457 sig = 0; 458 p->p_stat = SSLEEP; 459 p->p_stats->p_ru.ru_nvcsw++; 460 mi_switch(); 461 CTR4(KTR_PROC, 462 "msleep resume: proc %p (pid %d, %s), schedlock %p", 463 p, p->p_pid, p->p_comm, (void *) sched_lock.mtx_lock); 464resume: 465 p->p_sflag &= ~PS_SINTR; 466 if (p->p_sflag & PS_TIMEOUT) { 467 p->p_sflag &= ~PS_TIMEOUT; 468 if (sig == 0) { 469#ifdef KTRACE 470 if (KTRPOINT(p, KTR_CSW)) 471 ktrcsw(p->p_tracep, 0, 0); 472#endif 473 rval = EWOULDBLOCK; 474 mtx_unlock_spin(&sched_lock); 475 goto out; 476 } 477 } else if (timo) 478 callout_stop(&p->p_slpcallout); 479 mtx_unlock_spin(&sched_lock); 480 481 if (catch && (sig != 0 || (sig = CURSIG(p)))) { 482#ifdef KTRACE 483 if (KTRPOINT(p, KTR_CSW)) 484 ktrcsw(p->p_tracep, 0, 0); 485#endif 486 PROC_LOCK(p); 487 if (SIGISMEMBER(p->p_sigacts->ps_sigintr, sig)) 488 rval = EINTR; 489 else 490 rval = ERESTART; 491 PROC_UNLOCK(p); 492 goto out; 493 } 494out: 495#ifdef KTRACE 496 if (KTRPOINT(p, KTR_CSW)) 497 ktrcsw(p->p_tracep, 0, 0); 498#endif 499 PICKUP_GIANT(); 500 if (mtx != NULL) { 501 mtx_lock(mtx); 502 WITNESS_RESTORE(&mtx->mtx_object, mtx); 503 } 504 return (rval); 505} 506 507/* 508 * asleep() - async sleep call. Place process on wait queue and return 509 * immediately without blocking. The process stays runnable until mawait() 510 * is called. If ident is NULL, remove process from wait queue if it is still 511 * on one. 512 * 513 * Only the most recent sleep condition is effective when making successive 514 * calls to asleep() or when calling msleep(). 515 * 516 * The timeout, if any, is not initiated until mawait() is called. The sleep 517 * priority, signal, and timeout is specified in the asleep() call but may be 518 * overriden in the mawait() call. 519 * 520 * <<<<<<<< EXPERIMENTAL, UNTESTED >>>>>>>>>> 521 */ 522 523int 524asleep(void *ident, int priority, const char *wmesg, int timo) 525{ 526 struct proc *p = curproc; 527 int s; 528 529 /* 530 * obtain sched_lock while manipulating sleep structures and slpque. 531 * 532 * Remove preexisting wait condition (if any) and place process 533 * on appropriate slpque, but do not put process to sleep. 534 */ 535 536 s = splhigh(); 537 mtx_lock_spin(&sched_lock); 538 539 if (p->p_wchan != NULL) 540 unsleep(p); 541 542 if (ident) { 543 p->p_wchan = ident; 544 p->p_wmesg = wmesg; 545 p->p_slptime = 0; 546 p->p_asleep.as_priority = priority; 547 p->p_asleep.as_timo = timo; 548 TAILQ_INSERT_TAIL(&slpque[LOOKUP(ident)], p, p_slpq); 549 } 550 551 mtx_unlock_spin(&sched_lock); 552 splx(s); 553 554 return(0); 555} 556 557/* 558 * mawait() - wait for async condition to occur. The process blocks until 559 * wakeup() is called on the most recent asleep() address. If wakeup is called 560 * prior to mawait(), mawait() winds up being a NOP. 561 * 562 * If mawait() is called more then once (without an intervening asleep() call), 563 * mawait() is still effectively a NOP but it calls mi_switch() to give other 564 * processes some cpu before returning. The process is left runnable. 565 * 566 * <<<<<<<< EXPERIMENTAL, UNTESTED >>>>>>>>>> 567 */ 568 569int 570mawait(struct mtx *mtx, int priority, int timo) 571{ 572 struct proc *p = curproc; 573 int rval = 0; 574 int s; 575 WITNESS_SAVE_DECL(mtx); 576 577 WITNESS_SLEEP(0, &mtx->mtx_object); 578 mtx_lock_spin(&sched_lock); 579 DROP_GIANT_NOSWITCH(); 580 if (mtx != NULL) { 581 mtx_assert(mtx, MA_OWNED | MA_NOTRECURSED); 582 WITNESS_SAVE(&mtx->mtx_object, mtx); 583 mtx_unlock_flags(mtx, MTX_NOSWITCH); 584 if (priority & PDROP) 585 mtx = NULL; 586 } 587 588 s = splhigh(); 589 590 if (p->p_wchan != NULL) { 591 int sig; 592 int catch; 593 594 /* 595 * The call to mawait() can override defaults specified in 596 * the original asleep(). 597 */ 598 if (priority < 0) 599 priority = p->p_asleep.as_priority; 600 if (timo < 0) 601 timo = p->p_asleep.as_timo; 602 603 /* 604 * Install timeout 605 */ 606 607 if (timo) 608 callout_reset(&p->p_slpcallout, timo, endtsleep, p); 609 610 sig = 0; 611 catch = priority & PCATCH; 612 613 if (catch) { 614 p->p_sflag |= PS_SINTR; 615 mtx_unlock_spin(&sched_lock); 616 if ((sig = CURSIG(p))) { 617 mtx_lock_spin(&sched_lock); 618 if (p->p_wchan) 619 unsleep(p); 620 p->p_stat = SRUN; 621 goto resume; 622 } 623 mtx_lock_spin(&sched_lock); 624 if (p->p_wchan == NULL) { 625 catch = 0; 626 goto resume; 627 } 628 } 629 p->p_stat = SSLEEP; 630 p->p_stats->p_ru.ru_nvcsw++; 631 mi_switch(); 632resume: 633 634 splx(s); 635 p->p_sflag &= ~PS_SINTR; 636 if (p->p_sflag & PS_TIMEOUT) { 637 p->p_sflag &= ~PS_TIMEOUT; 638 if (sig == 0) { 639#ifdef KTRACE 640 if (KTRPOINT(p, KTR_CSW)) 641 ktrcsw(p->p_tracep, 0, 0); 642#endif 643 rval = EWOULDBLOCK; 644 mtx_unlock_spin(&sched_lock); 645 goto out; 646 } 647 } else if (timo) 648 callout_stop(&p->p_slpcallout); 649 mtx_unlock_spin(&sched_lock); 650 651 if (catch && (sig != 0 || (sig = CURSIG(p)))) { 652#ifdef KTRACE 653 if (KTRPOINT(p, KTR_CSW)) 654 ktrcsw(p->p_tracep, 0, 0); 655#endif 656 PROC_LOCK(p); 657 if (SIGISMEMBER(p->p_sigacts->ps_sigintr, sig)) 658 rval = EINTR; 659 else 660 rval = ERESTART; 661 PROC_UNLOCK(p); 662 goto out; 663 } 664#ifdef KTRACE 665 if (KTRPOINT(p, KTR_CSW)) 666 ktrcsw(p->p_tracep, 0, 0); 667#endif 668 } else { 669 /* 670 * If as_priority is 0, mawait() has been called without an 671 * intervening asleep(). We are still effectively a NOP, 672 * but we call mi_switch() for safety. 673 */ 674 675 if (p->p_asleep.as_priority == 0) { 676 p->p_stats->p_ru.ru_nvcsw++; 677 mi_switch(); 678 } 679 mtx_unlock_spin(&sched_lock); 680 splx(s); 681 } 682 683 /* 684 * clear p_asleep.as_priority as an indication that mawait() has been 685 * called. If mawait() is called again without an intervening asleep(), 686 * mawait() is still effectively a NOP but the above mi_switch() code 687 * is triggered as a safety. 688 */ 689 p->p_asleep.as_priority = 0; 690 691out: 692 PICKUP_GIANT(); 693 if (mtx != NULL) { 694 mtx_lock(mtx); 695 WITNESS_RESTORE(&mtx->mtx_object, mtx); 696 } 697 return (rval); 698} 699 700/* 701 * Implement timeout for msleep or asleep()/mawait() 702 * 703 * If process hasn't been awakened (wchan non-zero), 704 * set timeout flag and undo the sleep. If proc 705 * is stopped, just unsleep so it will remain stopped. 706 * MP-safe, called without the Giant mutex. 707 */ 708static void 709endtsleep(arg) 710 void *arg; 711{ 712 register struct proc *p; 713 int s; 714 715 p = (struct proc *)arg; 716 CTR4(KTR_PROC, 717 "endtsleep: proc %p (pid %d, %s), schedlock %p", 718 p, p->p_pid, p->p_comm, (void *) sched_lock.mtx_lock); 719 s = splhigh(); 720 mtx_lock_spin(&sched_lock); 721 if (p->p_wchan) { 722 if (p->p_stat == SSLEEP) 723 setrunnable(p); 724 else 725 unsleep(p); 726 p->p_sflag |= PS_TIMEOUT; 727 } 728 mtx_unlock_spin(&sched_lock); 729 splx(s); 730} 731 732/* 733 * Remove a process from its wait queue 734 */ 735void 736unsleep(p) 737 register struct proc *p; 738{ 739 int s; 740 741 s = splhigh(); 742 mtx_lock_spin(&sched_lock); 743 if (p->p_wchan) { 744 TAILQ_REMOVE(&slpque[LOOKUP(p->p_wchan)], p, p_slpq); 745 p->p_wchan = NULL; 746 } 747 mtx_unlock_spin(&sched_lock); 748 splx(s); 749} 750 751/* 752 * Make all processes sleeping on the specified identifier runnable. 753 */ 754void 755wakeup(ident) 756 register void *ident; 757{ 758 register struct slpquehead *qp; 759 register struct proc *p; 760 int s; 761 762 s = splhigh(); 763 mtx_lock_spin(&sched_lock); 764 qp = &slpque[LOOKUP(ident)]; 765restart: 766 TAILQ_FOREACH(p, qp, p_slpq) { 767 if (p->p_wchan == ident) { 768 TAILQ_REMOVE(qp, p, p_slpq); 769 p->p_wchan = NULL; 770 if (p->p_stat == SSLEEP) { 771 /* OPTIMIZED EXPANSION OF setrunnable(p); */ 772 CTR4(KTR_PROC, 773 "wakeup: proc %p (pid %d, %s), schedlock %p", 774 p, p->p_pid, p->p_comm, (void *) sched_lock.mtx_lock); 775 if (p->p_slptime > 1) 776 updatepri(p); 777 p->p_slptime = 0; 778 p->p_stat = SRUN; 779 if (p->p_sflag & PS_INMEM) { 780 setrunqueue(p); 781 maybe_resched(p); 782 } else { 783 p->p_sflag |= PS_SWAPINREQ; 784 wakeup((caddr_t)&proc0); 785 } 786 /* END INLINE EXPANSION */ 787 goto restart; 788 } 789 } 790 } 791 mtx_unlock_spin(&sched_lock); 792 splx(s); 793} 794 795/* 796 * Make a process sleeping on the specified identifier runnable. 797 * May wake more than one process if a target process is currently 798 * swapped out. 799 */ 800void 801wakeup_one(ident) 802 register void *ident; 803{ 804 register struct slpquehead *qp; 805 register struct proc *p; 806 int s; 807 808 s = splhigh(); 809 mtx_lock_spin(&sched_lock); 810 qp = &slpque[LOOKUP(ident)]; 811 812 TAILQ_FOREACH(p, qp, p_slpq) { 813 if (p->p_wchan == ident) { 814 TAILQ_REMOVE(qp, p, p_slpq); 815 p->p_wchan = NULL; 816 if (p->p_stat == SSLEEP) { 817 /* OPTIMIZED EXPANSION OF setrunnable(p); */ 818 CTR4(KTR_PROC, 819 "wakeup1: proc %p (pid %d, %s), schedlock %p", 820 p, p->p_pid, p->p_comm, (void *) sched_lock.mtx_lock); 821 if (p->p_slptime > 1) 822 updatepri(p); 823 p->p_slptime = 0; 824 p->p_stat = SRUN; 825 if (p->p_sflag & PS_INMEM) { 826 setrunqueue(p); 827 maybe_resched(p); 828 break; 829 } else { 830 p->p_sflag |= PS_SWAPINREQ; 831 wakeup((caddr_t)&proc0); 832 } 833 /* END INLINE EXPANSION */ 834 } 835 } 836 } 837 mtx_unlock_spin(&sched_lock); 838 splx(s); 839} 840 841/* 842 * The machine independent parts of mi_switch(). 843 * Must be called at splstatclock() or higher. 844 */ 845void 846mi_switch() 847{ 848 struct timeval new_switchtime; 849 register struct proc *p = curproc; /* XXX */ 850#if 0 851 register struct rlimit *rlim; 852#endif 853 u_int sched_nest; 854 855 mtx_assert(&sched_lock, MA_OWNED | MA_NOTRECURSED); 856 857 /* 858 * Compute the amount of time during which the current 859 * process was running, and add that to its total so far. 860 */ 861 microuptime(&new_switchtime); 862 if (timevalcmp(&new_switchtime, PCPU_PTR(switchtime), <)) { 863#if 0 864 /* XXX: This doesn't play well with sched_lock right now. */ 865 printf("microuptime() went backwards (%ld.%06ld -> %ld.%06ld)\n", 866 PCPU_GET(switchtime.tv_sec), PCPU_GET(switchtime.tv_usec), 867 new_switchtime.tv_sec, new_switchtime.tv_usec); 868#endif 869 new_switchtime = PCPU_GET(switchtime); 870 } else { 871 p->p_runtime += (new_switchtime.tv_usec - PCPU_GET(switchtime.tv_usec)) + 872 (new_switchtime.tv_sec - PCPU_GET(switchtime.tv_sec)) * 873 (int64_t)1000000; 874 } 875 876#if 0 877 /* 878 * Check if the process exceeds its cpu resource allocation. 879 * If over max, kill it. 880 * 881 * XXX drop sched_lock, pickup Giant 882 */ 883 if (p->p_stat != SZOMB && p->p_limit->p_cpulimit != RLIM_INFINITY && 884 p->p_runtime > p->p_limit->p_cpulimit) { 885 rlim = &p->p_rlimit[RLIMIT_CPU]; 886 if (p->p_runtime / (rlim_t)1000000 >= rlim->rlim_max) { 887 mtx_unlock_spin(&sched_lock); 888 killproc(p, "exceeded maximum CPU limit"); 889 mtx_lock_spin(&sched_lock); 890 } else { 891 mtx_unlock_spin(&sched_lock); 892 PROC_LOCK(p); 893 psignal(p, SIGXCPU); 894 mtx_lock_spin(&sched_lock); 895 PROC_UNLOCK_NOSWITCH(p); 896 if (rlim->rlim_cur < rlim->rlim_max) { 897 /* XXX: we should make a private copy */ 898 rlim->rlim_cur += 5; 899 } 900 } 901 } 902#endif 903 904 /* 905 * Pick a new current process and record its start time. 906 */ 907 cnt.v_swtch++; 908 PCPU_SET(switchtime, new_switchtime); 909 CTR4(KTR_PROC, "mi_switch: old proc %p (pid %d, %s), schedlock %p", 910 p, p->p_pid, p->p_comm, (void *) sched_lock.mtx_lock); 911 sched_nest = sched_lock.mtx_recurse; 912 curproc->p_lastcpu = curproc->p_oncpu; 913 curproc->p_oncpu = NOCPU; 914 clear_resched(); 915 cpu_switch(); 916 curproc->p_oncpu = PCPU_GET(cpuid); 917 sched_lock.mtx_recurse = sched_nest; 918 sched_lock.mtx_lock = (uintptr_t)curproc; 919 CTR4(KTR_PROC, "mi_switch: new proc %p (pid %d, %s), schedlock %p", 920 p, p->p_pid, p->p_comm, (void *) sched_lock.mtx_lock); 921 if (PCPU_GET(switchtime.tv_sec) == 0) 922 microuptime(PCPU_PTR(switchtime)); 923 PCPU_SET(switchticks, ticks); 924} 925 926/* 927 * Change process state to be runnable, 928 * placing it on the run queue if it is in memory, 929 * and awakening the swapper if it isn't in memory. 930 */ 931void 932setrunnable(p) 933 register struct proc *p; 934{ 935 register int s; 936 937 s = splhigh(); 938 mtx_lock_spin(&sched_lock); 939 switch (p->p_stat) { 940 case 0: 941 case SRUN: 942 case SZOMB: 943 case SWAIT: 944 default: 945 panic("setrunnable"); 946 case SSTOP: 947 case SSLEEP: /* e.g. when sending signals */ 948 if (p->p_sflag & PS_CVWAITQ) 949 cv_waitq_remove(p); 950 else 951 unsleep(p); 952 break; 953 954 case SIDL: 955 break; 956 } 957 p->p_stat = SRUN; 958 if (p->p_sflag & PS_INMEM) 959 setrunqueue(p); 960 splx(s); 961 if (p->p_slptime > 1) 962 updatepri(p); 963 p->p_slptime = 0; 964 if ((p->p_sflag & PS_INMEM) == 0) { 965 p->p_sflag |= PS_SWAPINREQ; 966 wakeup((caddr_t)&proc0); 967 } 968 else 969 maybe_resched(p); 970 mtx_unlock_spin(&sched_lock); 971} 972 973/* 974 * Compute the priority of a process when running in user mode. 975 * Arrange to reschedule if the resulting priority is better 976 * than that of the current process. 977 */ 978void 979resetpriority(p) 980 register struct proc *p; 981{ 982 register unsigned int newpriority; 983 984 mtx_lock_spin(&sched_lock); 985 if (p->p_pri.pri_class == PRI_TIMESHARE) { 986 newpriority = PUSER + p->p_estcpu / INVERSE_ESTCPU_WEIGHT + 987 NICE_WEIGHT * (p->p_nice - PRIO_MIN); 988 newpriority = min(max(newpriority, PRI_MIN_TIMESHARE), 989 PRI_MAX_TIMESHARE); 990 p->p_pri.pri_user = newpriority; 991 } 992 maybe_resched(p); 993 mtx_unlock_spin(&sched_lock); 994} 995 996/* ARGSUSED */ 997static void 998sched_setup(dummy) 999 void *dummy; 1000{ 1001 1002 callout_init(&schedcpu_callout, 1); 1003 callout_init(&roundrobin_callout, 0); 1004 1005 /* Kick off timeout driven events by calling first time. */ 1006 roundrobin(NULL); 1007 schedcpu(NULL); 1008} 1009 1010/* 1011 * We adjust the priority of the current process. The priority of 1012 * a process gets worse as it accumulates CPU time. The cpu usage 1013 * estimator (p_estcpu) is increased here. resetpriority() will 1014 * compute a different priority each time p_estcpu increases by 1015 * INVERSE_ESTCPU_WEIGHT 1016 * (until MAXPRI is reached). The cpu usage estimator ramps up 1017 * quite quickly when the process is running (linearly), and decays 1018 * away exponentially, at a rate which is proportionally slower when 1019 * the system is busy. The basic principle is that the system will 1020 * 90% forget that the process used a lot of CPU time in 5 * loadav 1021 * seconds. This causes the system to favor processes which haven't 1022 * run much recently, and to round-robin among other processes. 1023 */ 1024void 1025schedclock(p) 1026 struct proc *p; 1027{ 1028 1029 p->p_cpticks++; 1030 p->p_estcpu = ESTCPULIM(p->p_estcpu + 1); 1031 if ((p->p_estcpu % INVERSE_ESTCPU_WEIGHT) == 0) { 1032 resetpriority(p); 1033 if (p->p_pri.pri_level >= PUSER) 1034 p->p_pri.pri_level = p->p_pri.pri_user; 1035 } 1036} 1037 1038/* 1039 * General purpose yield system call 1040 */ 1041int 1042yield(struct proc *p, struct yield_args *uap) 1043{ 1044 int s; 1045 1046 p->p_retval[0] = 0; 1047 1048 s = splhigh(); 1049 mtx_lock_spin(&sched_lock); 1050 DROP_GIANT_NOSWITCH(); 1051 p->p_pri.pri_level = PRI_MAX_TIMESHARE; 1052 setrunqueue(p); 1053 p->p_stats->p_ru.ru_nvcsw++; 1054 mi_switch(); 1055 mtx_unlock_spin(&sched_lock); 1056 PICKUP_GIANT(); 1057 splx(s); 1058 1059 return (0); 1060} 1061