kern_synch.c revision 71564
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 71564 2001-01-24 11:10:55Z 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/condvar.h> 51#include <sys/lock.h> 52#include <sys/mutex.h> 53#include <sys/signalvar.h> 54#include <sys/resourcevar.h> 55#include <sys/vmmeter.h> 56#include <sys/sysctl.h> 57#include <sys/sysproto.h> 58#include <vm/vm.h> 59#include <vm/vm_extern.h> 60#ifdef KTRACE 61#include <sys/uio.h> 62#include <sys/ktrace.h> 63#endif 64 65#include <machine/cpu.h> 66#include <machine/smp.h> 67 68static void sched_setup __P((void *dummy)); 69SYSINIT(sched_setup, SI_SUB_KICK_SCHEDULER, SI_ORDER_FIRST, sched_setup, NULL) 70 71u_char curpriority; 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 int curpriority_cmp __P((struct proc *p)); 80static void endtsleep __P((void *)); 81static void roundrobin __P((void *arg)); 82static void schedcpu __P((void *arg)); 83 84static int 85sysctl_kern_quantum(SYSCTL_HANDLER_ARGS) 86{ 87 int error, new_val; 88 89 new_val = sched_quantum * tick; 90 error = sysctl_handle_int(oidp, &new_val, 0, req); 91 if (error != 0 || req->newptr == NULL) 92 return (error); 93 if (new_val < tick) 94 return (EINVAL); 95 sched_quantum = new_val / tick; 96 hogticks = 2 * sched_quantum; 97 return (0); 98} 99 100SYSCTL_PROC(_kern, OID_AUTO, quantum, CTLTYPE_INT|CTLFLAG_RW, 101 0, sizeof sched_quantum, sysctl_kern_quantum, "I", ""); 102 103/*- 104 * Compare priorities. Return: 105 * <0: priority of p < current priority 106 * 0: priority of p == current priority 107 * >0: priority of p > current priority 108 * The priorities are the normal priorities or the normal realtime priorities 109 * if p is on the same scheduler as curproc. Otherwise the process on the 110 * more realtimeish scheduler has lowest priority. As usual, a higher 111 * priority really means a lower priority. 112 */ 113static int 114curpriority_cmp(p) 115 struct proc *p; 116{ 117 int c_class, p_class; 118 119 c_class = RTP_PRIO_BASE(curproc->p_rtprio.type); 120 p_class = RTP_PRIO_BASE(p->p_rtprio.type); 121 if (p_class != c_class) 122 return (p_class - c_class); 123 if (p_class == RTP_PRIO_NORMAL) 124 return (((int)p->p_priority - (int)curpriority) / PPQ); 125 return ((int)p->p_rtprio.prio - (int)curproc->p_rtprio.prio); 126} 127 128/* 129 * Arrange to reschedule if necessary, taking the priorities and 130 * schedulers into account. 131 */ 132void 133maybe_resched(chk) 134 struct proc *chk; 135{ 136 struct proc *p = curproc; /* XXX */ 137 138 /* 139 * XXX idle scheduler still broken because proccess stays on idle 140 * scheduler during waits (such as when getting FS locks). If a 141 * standard process becomes runaway cpu-bound, the system can lockup 142 * due to idle-scheduler processes in wakeup never getting any cpu. 143 */ 144 if (p == PCPU_GET(idleproc)) { 145#if 0 146 need_resched(); 147#endif 148 } else if (chk == p) { 149 /* We may need to yield if our priority has been raised. */ 150 if (curpriority_cmp(chk) > 0) 151 need_resched(); 152 } else if (curpriority_cmp(chk) < 0) 153 need_resched(); 154} 155 156int 157roundrobin_interval(void) 158{ 159 return (sched_quantum); 160} 161 162/* 163 * Force switch among equal priority processes every 100ms. 164 */ 165/* ARGSUSED */ 166static void 167roundrobin(arg) 168 void *arg; 169{ 170#ifndef SMP 171 struct proc *p = curproc; /* XXX */ 172#endif 173 174#ifdef SMP 175 need_resched(); 176 forward_roundrobin(); 177#else 178 if (p == PCPU_GET(idleproc) || RTP_PRIO_NEED_RR(p->p_rtprio.type)) 179 need_resched(); 180#endif 181 182 callout_reset(&roundrobin_callout, sched_quantum, roundrobin, NULL); 183} 184 185/* 186 * Constants for digital decay and forget: 187 * 90% of (p_estcpu) usage in 5 * loadav time 188 * 95% of (p_pctcpu) usage in 60 seconds (load insensitive) 189 * Note that, as ps(1) mentions, this can let percentages 190 * total over 100% (I've seen 137.9% for 3 processes). 191 * 192 * Note that schedclock() updates p_estcpu and p_cpticks asynchronously. 193 * 194 * We wish to decay away 90% of p_estcpu in (5 * loadavg) seconds. 195 * That is, the system wants to compute a value of decay such 196 * that the following for loop: 197 * for (i = 0; i < (5 * loadavg); i++) 198 * p_estcpu *= decay; 199 * will compute 200 * p_estcpu *= 0.1; 201 * for all values of loadavg: 202 * 203 * Mathematically this loop can be expressed by saying: 204 * decay ** (5 * loadavg) ~= .1 205 * 206 * The system computes decay as: 207 * decay = (2 * loadavg) / (2 * loadavg + 1) 208 * 209 * We wish to prove that the system's computation of decay 210 * will always fulfill the equation: 211 * decay ** (5 * loadavg) ~= .1 212 * 213 * If we compute b as: 214 * b = 2 * loadavg 215 * then 216 * decay = b / (b + 1) 217 * 218 * We now need to prove two things: 219 * 1) Given factor ** (5 * loadavg) ~= .1, prove factor == b/(b+1) 220 * 2) Given b/(b+1) ** power ~= .1, prove power == (5 * loadavg) 221 * 222 * Facts: 223 * For x close to zero, exp(x) =~ 1 + x, since 224 * exp(x) = 0! + x**1/1! + x**2/2! + ... . 225 * therefore exp(-1/b) =~ 1 - (1/b) = (b-1)/b. 226 * For x close to zero, ln(1+x) =~ x, since 227 * ln(1+x) = x - x**2/2 + x**3/3 - ... -1 < x < 1 228 * therefore ln(b/(b+1)) = ln(1 - 1/(b+1)) =~ -1/(b+1). 229 * ln(.1) =~ -2.30 230 * 231 * Proof of (1): 232 * Solve (factor)**(power) =~ .1 given power (5*loadav): 233 * solving for factor, 234 * ln(factor) =~ (-2.30/5*loadav), or 235 * factor =~ exp(-1/((5/2.30)*loadav)) =~ exp(-1/(2*loadav)) = 236 * exp(-1/b) =~ (b-1)/b =~ b/(b+1). QED 237 * 238 * Proof of (2): 239 * Solve (factor)**(power) =~ .1 given factor == (b/(b+1)): 240 * solving for power, 241 * power*ln(b/(b+1)) =~ -2.30, or 242 * power =~ 2.3 * (b + 1) = 4.6*loadav + 2.3 =~ 5*loadav. QED 243 * 244 * Actual power values for the implemented algorithm are as follows: 245 * loadav: 1 2 3 4 246 * power: 5.68 10.32 14.94 19.55 247 */ 248 249/* calculations for digital decay to forget 90% of usage in 5*loadav sec */ 250#define loadfactor(loadav) (2 * (loadav)) 251#define decay_cpu(loadfac, cpu) (((loadfac) * (cpu)) / ((loadfac) + FSCALE)) 252 253/* decay 95% of `p_pctcpu' in 60 seconds; see CCPU_SHIFT before changing */ 254static fixpt_t ccpu = 0.95122942450071400909 * FSCALE; /* exp(-1/20) */ 255SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, ""); 256 257/* kernel uses `FSCALE', userland (SHOULD) use kern.fscale */ 258static int fscale __unused = FSCALE; 259SYSCTL_INT(_kern, OID_AUTO, fscale, CTLFLAG_RD, 0, FSCALE, ""); 260 261/* 262 * If `ccpu' is not equal to `exp(-1/20)' and you still want to use the 263 * faster/more-accurate formula, you'll have to estimate CCPU_SHIFT below 264 * and possibly adjust FSHIFT in "param.h" so that (FSHIFT >= CCPU_SHIFT). 265 * 266 * To estimate CCPU_SHIFT for exp(-1/20), the following formula was used: 267 * 1 - exp(-1/20) ~= 0.0487 ~= 0.0488 == 1 (fixed pt, *11* bits). 268 * 269 * If you don't want to bother with the faster/more-accurate formula, you 270 * can set CCPU_SHIFT to (FSHIFT + 1) which will use a slower/less-accurate 271 * (more general) method of calculating the %age of CPU used by a process. 272 */ 273#define CCPU_SHIFT 11 274 275/* 276 * Recompute process priorities, every hz ticks. 277 * MP-safe, called without the Giant mutex. 278 */ 279/* ARGSUSED */ 280static void 281schedcpu(arg) 282 void *arg; 283{ 284 register fixpt_t loadfac = loadfactor(averunnable.ldavg[0]); 285 register struct proc *p; 286 register int realstathz, s; 287 288 realstathz = stathz ? stathz : hz; 289 ALLPROC_LOCK(AP_SHARED); 290 LIST_FOREACH(p, &allproc, p_list) { 291 /* 292 * Increment time in/out of memory and sleep time 293 * (if sleeping). We ignore overflow; with 16-bit int's 294 * (remember them?) overflow takes 45 days. 295 if (p->p_stat == SWAIT) 296 continue; 297 */ 298 mtx_enter(&sched_lock, MTX_SPIN); 299 p->p_swtime++; 300 if (p->p_stat == SSLEEP || p->p_stat == SSTOP) 301 p->p_slptime++; 302 p->p_pctcpu = (p->p_pctcpu * ccpu) >> FSHIFT; 303 /* 304 * If the process has slept the entire second, 305 * stop recalculating its priority until it wakes up. 306 */ 307 if (p->p_slptime > 1) { 308 mtx_exit(&sched_lock, MTX_SPIN); 309 continue; 310 } 311 312 /* 313 * prevent state changes and protect run queue 314 */ 315 s = splhigh(); 316 317 /* 318 * p_pctcpu is only for ps. 319 */ 320#if (FSHIFT >= CCPU_SHIFT) 321 p->p_pctcpu += (realstathz == 100)? 322 ((fixpt_t) p->p_cpticks) << (FSHIFT - CCPU_SHIFT): 323 100 * (((fixpt_t) p->p_cpticks) 324 << (FSHIFT - CCPU_SHIFT)) / realstathz; 325#else 326 p->p_pctcpu += ((FSCALE - ccpu) * 327 (p->p_cpticks * FSCALE / realstathz)) >> FSHIFT; 328#endif 329 p->p_cpticks = 0; 330 p->p_estcpu = decay_cpu(loadfac, p->p_estcpu); 331 resetpriority(p); 332 if (p->p_priority >= PUSER) { 333 if ((p != curproc) && 334#ifdef SMP 335 p->p_oncpu == 0xff && /* idle */ 336#endif 337 p->p_stat == SRUN && 338 (p->p_sflag & PS_INMEM) && 339 (p->p_priority / PPQ) != (p->p_usrpri / PPQ)) { 340 remrunqueue(p); 341 p->p_priority = p->p_usrpri; 342 setrunqueue(p); 343 } else 344 p->p_priority = p->p_usrpri; 345 } 346 mtx_exit(&sched_lock, MTX_SPIN); 347 splx(s); 348 } 349 ALLPROC_LOCK(AP_RELEASE); 350 vmmeter(); 351 wakeup((caddr_t)&lbolt); 352 callout_reset(&schedcpu_callout, hz, schedcpu, NULL); 353} 354 355/* 356 * Recalculate the priority of a process after it has slept for a while. 357 * For all load averages >= 1 and max p_estcpu of 255, sleeping for at 358 * least six times the loadfactor will decay p_estcpu to zero. 359 */ 360void 361updatepri(p) 362 register struct proc *p; 363{ 364 register unsigned int newcpu = p->p_estcpu; 365 register fixpt_t loadfac = loadfactor(averunnable.ldavg[0]); 366 367 if (p->p_slptime > 5 * loadfac) 368 p->p_estcpu = 0; 369 else { 370 p->p_slptime--; /* the first time was done in schedcpu */ 371 while (newcpu && --p->p_slptime) 372 newcpu = decay_cpu(loadfac, newcpu); 373 p->p_estcpu = newcpu; 374 } 375 resetpriority(p); 376} 377 378/* 379 * We're only looking at 7 bits of the address; everything is 380 * aligned to 4, lots of things are aligned to greater powers 381 * of 2. Shift right by 8, i.e. drop the bottom 256 worth. 382 */ 383#define TABLESIZE 128 384static TAILQ_HEAD(slpquehead, proc) slpque[TABLESIZE]; 385#define LOOKUP(x) (((intptr_t)(x) >> 8) & (TABLESIZE - 1)) 386 387void 388sleepinit(void) 389{ 390 int i; 391 392 sched_quantum = hz/10; 393 hogticks = 2 * sched_quantum; 394 for (i = 0; i < TABLESIZE; i++) 395 TAILQ_INIT(&slpque[i]); 396} 397 398/* 399 * General sleep call. Suspends the current process until a wakeup is 400 * performed on the specified identifier. The process will then be made 401 * runnable with the specified priority. Sleeps at most timo/hz seconds 402 * (0 means no timeout). If pri includes PCATCH flag, signals are checked 403 * before and after sleeping, else signals are not checked. Returns 0 if 404 * awakened, EWOULDBLOCK if the timeout expires. If PCATCH is set and a 405 * signal needs to be delivered, ERESTART is returned if the current system 406 * call should be restarted if possible, and EINTR is returned if the system 407 * call should be interrupted by the signal (return EINTR). 408 * 409 * The mutex argument is exited before the caller is suspended, and 410 * entered before msleep returns. If priority includes the PDROP 411 * flag the mutex is not entered before returning. 412 */ 413int 414msleep(ident, mtx, priority, wmesg, timo) 415 void *ident; 416 struct mtx *mtx; 417 int priority, timo; 418 const char *wmesg; 419{ 420 struct proc *p = curproc; 421 int s, sig, catch = priority & PCATCH; 422 int rval = 0; 423 WITNESS_SAVE_DECL(mtx); 424 425#ifdef KTRACE 426 if (p && KTRPOINT(p, KTR_CSW)) 427 ktrcsw(p->p_tracep, 1, 0); 428#endif 429 WITNESS_SLEEP(0, mtx); 430 mtx_enter(&sched_lock, MTX_SPIN); 431 s = splhigh(); 432 if (cold || panicstr) { 433 /* 434 * After a panic, or during autoconfiguration, 435 * just give interrupts a chance, then just return; 436 * don't run any other procs or panic below, 437 * in case this is the idle process and already asleep. 438 */ 439 if (mtx != NULL && priority & PDROP) 440 mtx_exit(mtx, MTX_DEF | MTX_NOSWITCH); 441 mtx_exit(&sched_lock, MTX_SPIN); 442 splx(s); 443 return (0); 444 } 445 446 DROP_GIANT_NOSWITCH(); 447 448 if (mtx != NULL) { 449 mtx_assert(mtx, MA_OWNED | MA_NOTRECURSED); 450 WITNESS_SAVE(mtx, mtx); 451 mtx_exit(mtx, MTX_DEF | MTX_NOSWITCH); 452 if (priority & PDROP) 453 mtx = NULL; 454 } 455 456 KASSERT(p != NULL, ("msleep1")); 457 KASSERT(ident != NULL && p->p_stat == SRUN, ("msleep")); 458 /* 459 * Process may be sitting on a slpque if asleep() was called, remove 460 * it before re-adding. 461 */ 462 if (p->p_wchan != NULL) 463 unsleep(p); 464 465 p->p_wchan = ident; 466 p->p_wmesg = wmesg; 467 p->p_slptime = 0; 468 p->p_priority = priority & PRIMASK; 469 CTR4(KTR_PROC, "msleep: proc %p (pid %d, %s), schedlock %p", 470 p, p->p_pid, p->p_comm, (void *) sched_lock.mtx_lock); 471 TAILQ_INSERT_TAIL(&slpque[LOOKUP(ident)], p, p_slpq); 472 if (timo) 473 callout_reset(&p->p_slpcallout, timo, endtsleep, p); 474 /* 475 * We put ourselves on the sleep queue and start our timeout 476 * before calling CURSIG, as we could stop there, and a wakeup 477 * or a SIGCONT (or both) could occur while we were stopped. 478 * A SIGCONT would cause us to be marked as SSLEEP 479 * without resuming us, thus we must be ready for sleep 480 * when CURSIG is called. If the wakeup happens while we're 481 * stopped, p->p_wchan will be 0 upon return from CURSIG. 482 */ 483 if (catch) { 484 CTR4(KTR_PROC, 485 "msleep caught: proc %p (pid %d, %s), schedlock %p", 486 p, p->p_pid, p->p_comm, (void *) sched_lock.mtx_lock); 487 p->p_sflag |= PS_SINTR; 488 mtx_exit(&sched_lock, MTX_SPIN); 489 if ((sig = CURSIG(p))) { 490 mtx_enter(&sched_lock, MTX_SPIN); 491 if (p->p_wchan) 492 unsleep(p); 493 p->p_stat = SRUN; 494 goto resume; 495 } 496 mtx_enter(&sched_lock, MTX_SPIN); 497 if (p->p_wchan == NULL) { 498 catch = 0; 499 goto resume; 500 } 501 } else 502 sig = 0; 503 p->p_stat = SSLEEP; 504 p->p_stats->p_ru.ru_nvcsw++; 505 mi_switch(); 506 CTR4(KTR_PROC, 507 "msleep resume: proc %p (pid %d, %s), schedlock %p", 508 p, p->p_pid, p->p_comm, (void *) sched_lock.mtx_lock); 509resume: 510 curpriority = p->p_usrpri; 511 splx(s); 512 p->p_sflag &= ~PS_SINTR; 513 if (p->p_sflag & PS_TIMEOUT) { 514 p->p_sflag &= ~PS_TIMEOUT; 515 if (sig == 0) { 516#ifdef KTRACE 517 if (KTRPOINT(p, KTR_CSW)) 518 ktrcsw(p->p_tracep, 0, 0); 519#endif 520 rval = EWOULDBLOCK; 521 mtx_exit(&sched_lock, MTX_SPIN); 522 goto out; 523 } 524 } else if (timo) 525 callout_stop(&p->p_slpcallout); 526 mtx_exit(&sched_lock, MTX_SPIN); 527 528 if (catch && (sig != 0 || (sig = CURSIG(p)))) { 529#ifdef KTRACE 530 if (KTRPOINT(p, KTR_CSW)) 531 ktrcsw(p->p_tracep, 0, 0); 532#endif 533 if (SIGISMEMBER(p->p_sigacts->ps_sigintr, sig)) 534 rval = EINTR; 535 else 536 rval = ERESTART; 537 goto out; 538 } 539out: 540#ifdef KTRACE 541 if (KTRPOINT(p, KTR_CSW)) 542 ktrcsw(p->p_tracep, 0, 0); 543#endif 544 PICKUP_GIANT(); 545 if (mtx != NULL) { 546 mtx_enter(mtx, MTX_DEF); 547 WITNESS_RESTORE(mtx, mtx); 548 } 549 return (rval); 550} 551 552/* 553 * asleep() - async sleep call. Place process on wait queue and return 554 * immediately without blocking. The process stays runnable until mawait() 555 * is called. If ident is NULL, remove process from wait queue if it is still 556 * on one. 557 * 558 * Only the most recent sleep condition is effective when making successive 559 * calls to asleep() or when calling msleep(). 560 * 561 * The timeout, if any, is not initiated until mawait() is called. The sleep 562 * priority, signal, and timeout is specified in the asleep() call but may be 563 * overriden in the mawait() call. 564 * 565 * <<<<<<<< EXPERIMENTAL, UNTESTED >>>>>>>>>> 566 */ 567 568int 569asleep(void *ident, int priority, const char *wmesg, int timo) 570{ 571 struct proc *p = curproc; 572 int s; 573 574 /* 575 * obtain sched_lock while manipulating sleep structures and slpque. 576 * 577 * Remove preexisting wait condition (if any) and place process 578 * on appropriate slpque, but do not put process to sleep. 579 */ 580 581 s = splhigh(); 582 mtx_enter(&sched_lock, MTX_SPIN); 583 584 if (p->p_wchan != NULL) 585 unsleep(p); 586 587 if (ident) { 588 p->p_wchan = ident; 589 p->p_wmesg = wmesg; 590 p->p_slptime = 0; 591 p->p_asleep.as_priority = priority; 592 p->p_asleep.as_timo = timo; 593 TAILQ_INSERT_TAIL(&slpque[LOOKUP(ident)], p, p_slpq); 594 } 595 596 mtx_exit(&sched_lock, MTX_SPIN); 597 splx(s); 598 599 return(0); 600} 601 602/* 603 * mawait() - wait for async condition to occur. The process blocks until 604 * wakeup() is called on the most recent asleep() address. If wakeup is called 605 * prior to mawait(), mawait() winds up being a NOP. 606 * 607 * If mawait() is called more then once (without an intervening asleep() call), 608 * mawait() is still effectively a NOP but it calls mi_switch() to give other 609 * processes some cpu before returning. The process is left runnable. 610 * 611 * <<<<<<<< EXPERIMENTAL, UNTESTED >>>>>>>>>> 612 */ 613 614int 615mawait(struct mtx *mtx, int priority, int timo) 616{ 617 struct proc *p = curproc; 618 int rval = 0; 619 int s; 620 WITNESS_SAVE_DECL(mtx); 621 622 WITNESS_SLEEP(0, mtx); 623 mtx_enter(&sched_lock, MTX_SPIN); 624 DROP_GIANT_NOSWITCH(); 625 if (mtx != NULL) { 626 mtx_assert(mtx, MA_OWNED | MA_NOTRECURSED); 627 WITNESS_SAVE(mtx, mtx); 628 mtx_exit(mtx, MTX_DEF | MTX_NOSWITCH); 629 if (priority & PDROP) 630 mtx = NULL; 631 } 632 633 s = splhigh(); 634 635 if (p->p_wchan != NULL) { 636 int sig; 637 int catch; 638 639 /* 640 * The call to mawait() can override defaults specified in 641 * the original asleep(). 642 */ 643 if (priority < 0) 644 priority = p->p_asleep.as_priority; 645 if (timo < 0) 646 timo = p->p_asleep.as_timo; 647 648 /* 649 * Install timeout 650 */ 651 652 if (timo) 653 callout_reset(&p->p_slpcallout, timo, endtsleep, p); 654 655 sig = 0; 656 catch = priority & PCATCH; 657 658 if (catch) { 659 p->p_sflag |= PS_SINTR; 660 mtx_exit(&sched_lock, MTX_SPIN); 661 if ((sig = CURSIG(p))) { 662 mtx_enter(&sched_lock, MTX_SPIN); 663 if (p->p_wchan) 664 unsleep(p); 665 p->p_stat = SRUN; 666 goto resume; 667 } 668 mtx_enter(&sched_lock, MTX_SPIN); 669 if (p->p_wchan == NULL) { 670 catch = 0; 671 goto resume; 672 } 673 } 674 p->p_stat = SSLEEP; 675 p->p_stats->p_ru.ru_nvcsw++; 676 mi_switch(); 677resume: 678 curpriority = p->p_usrpri; 679 680 splx(s); 681 p->p_sflag &= ~PS_SINTR; 682 if (p->p_sflag & PS_TIMEOUT) { 683 p->p_sflag &= ~PS_TIMEOUT; 684 if (sig == 0) { 685#ifdef KTRACE 686 if (KTRPOINT(p, KTR_CSW)) 687 ktrcsw(p->p_tracep, 0, 0); 688#endif 689 rval = EWOULDBLOCK; 690 mtx_exit(&sched_lock, MTX_SPIN); 691 goto out; 692 } 693 } else if (timo) 694 callout_stop(&p->p_slpcallout); 695 mtx_exit(&sched_lock, MTX_SPIN); 696 697 if (catch && (sig != 0 || (sig = CURSIG(p)))) { 698#ifdef KTRACE 699 if (KTRPOINT(p, KTR_CSW)) 700 ktrcsw(p->p_tracep, 0, 0); 701#endif 702 if (SIGISMEMBER(p->p_sigacts->ps_sigintr, sig)) 703 rval = EINTR; 704 else 705 rval = ERESTART; 706 goto out; 707 } 708#ifdef KTRACE 709 if (KTRPOINT(p, KTR_CSW)) 710 ktrcsw(p->p_tracep, 0, 0); 711#endif 712 } else { 713 /* 714 * If as_priority is 0, mawait() has been called without an 715 * intervening asleep(). We are still effectively a NOP, 716 * but we call mi_switch() for safety. 717 */ 718 719 if (p->p_asleep.as_priority == 0) { 720 p->p_stats->p_ru.ru_nvcsw++; 721 mi_switch(); 722 } 723 mtx_exit(&sched_lock, MTX_SPIN); 724 splx(s); 725 } 726 727 /* 728 * clear p_asleep.as_priority as an indication that mawait() has been 729 * called. If mawait() is called again without an intervening asleep(), 730 * mawait() is still effectively a NOP but the above mi_switch() code 731 * is triggered as a safety. 732 */ 733 p->p_asleep.as_priority = 0; 734 735out: 736 PICKUP_GIANT(); 737 if (mtx != NULL) { 738 mtx_enter(mtx, MTX_DEF); 739 WITNESS_RESTORE(mtx, mtx); 740 } 741 return (rval); 742} 743 744/* 745 * Implement timeout for msleep or asleep()/mawait() 746 * 747 * If process hasn't been awakened (wchan non-zero), 748 * set timeout flag and undo the sleep. If proc 749 * is stopped, just unsleep so it will remain stopped. 750 * MP-safe, called without the Giant mutex. 751 */ 752static void 753endtsleep(arg) 754 void *arg; 755{ 756 register struct proc *p; 757 int s; 758 759 p = (struct proc *)arg; 760 CTR4(KTR_PROC, 761 "endtsleep: proc %p (pid %d, %s), schedlock %p", 762 p, p->p_pid, p->p_comm, (void *) sched_lock.mtx_lock); 763 s = splhigh(); 764 mtx_enter(&sched_lock, MTX_SPIN); 765 if (p->p_wchan) { 766 if (p->p_stat == SSLEEP) 767 setrunnable(p); 768 else 769 unsleep(p); 770 p->p_sflag |= PS_TIMEOUT; 771 } 772 mtx_exit(&sched_lock, MTX_SPIN); 773 splx(s); 774} 775 776/* 777 * Remove a process from its wait queue 778 */ 779void 780unsleep(p) 781 register struct proc *p; 782{ 783 int s; 784 785 s = splhigh(); 786 mtx_enter(&sched_lock, MTX_SPIN); 787 if (p->p_wchan) { 788 TAILQ_REMOVE(&slpque[LOOKUP(p->p_wchan)], p, p_slpq); 789 p->p_wchan = NULL; 790 } 791 mtx_exit(&sched_lock, MTX_SPIN); 792 splx(s); 793} 794 795/* 796 * Make all processes sleeping on the specified identifier runnable. 797 */ 798void 799wakeup(ident) 800 register void *ident; 801{ 802 register struct slpquehead *qp; 803 register struct proc *p; 804 int s; 805 806 s = splhigh(); 807 mtx_enter(&sched_lock, MTX_SPIN); 808 qp = &slpque[LOOKUP(ident)]; 809restart: 810 TAILQ_FOREACH(p, qp, p_slpq) { 811 if (p->p_wchan == ident) { 812 TAILQ_REMOVE(qp, p, p_slpq); 813 p->p_wchan = NULL; 814 if (p->p_stat == SSLEEP) { 815 /* OPTIMIZED EXPANSION OF setrunnable(p); */ 816 CTR4(KTR_PROC, 817 "wakeup: proc %p (pid %d, %s), schedlock %p", 818 p, p->p_pid, p->p_comm, (void *) sched_lock.mtx_lock); 819 if (p->p_slptime > 1) 820 updatepri(p); 821 p->p_slptime = 0; 822 p->p_stat = SRUN; 823 if (p->p_sflag & PS_INMEM) { 824 setrunqueue(p); 825 maybe_resched(p); 826 } else { 827 p->p_sflag |= PS_SWAPINREQ; 828 wakeup((caddr_t)&proc0); 829 } 830 /* END INLINE EXPANSION */ 831 goto restart; 832 } 833 } 834 } 835 mtx_exit(&sched_lock, MTX_SPIN); 836 splx(s); 837} 838 839/* 840 * Make a process sleeping on the specified identifier runnable. 841 * May wake more than one process if a target process is currently 842 * swapped out. 843 */ 844void 845wakeup_one(ident) 846 register void *ident; 847{ 848 register struct slpquehead *qp; 849 register struct proc *p; 850 int s; 851 852 s = splhigh(); 853 mtx_enter(&sched_lock, MTX_SPIN); 854 qp = &slpque[LOOKUP(ident)]; 855 856 TAILQ_FOREACH(p, qp, p_slpq) { 857 if (p->p_wchan == ident) { 858 TAILQ_REMOVE(qp, p, p_slpq); 859 p->p_wchan = NULL; 860 if (p->p_stat == SSLEEP) { 861 /* OPTIMIZED EXPANSION OF setrunnable(p); */ 862 CTR4(KTR_PROC, 863 "wakeup1: proc %p (pid %d, %s), schedlock %p", 864 p, p->p_pid, p->p_comm, (void *) sched_lock.mtx_lock); 865 if (p->p_slptime > 1) 866 updatepri(p); 867 p->p_slptime = 0; 868 p->p_stat = SRUN; 869 if (p->p_sflag & PS_INMEM) { 870 setrunqueue(p); 871 maybe_resched(p); 872 break; 873 } else { 874 p->p_sflag |= PS_SWAPINREQ; 875 wakeup((caddr_t)&proc0); 876 } 877 /* END INLINE EXPANSION */ 878 } 879 } 880 } 881 mtx_exit(&sched_lock, MTX_SPIN); 882 splx(s); 883} 884 885/* 886 * The machine independent parts of mi_switch(). 887 * Must be called at splstatclock() or higher. 888 */ 889void 890mi_switch() 891{ 892 struct timeval new_switchtime; 893 register struct proc *p = curproc; /* XXX */ 894#if 0 895 register struct rlimit *rlim; 896#endif 897 int x; 898 899 /* 900 * XXX this spl is almost unnecessary. It is partly to allow for 901 * sloppy callers that don't do it (issignal() via CURSIG() is the 902 * main offender). It is partly to work around a bug in the i386 903 * cpu_switch() (the ipl is not preserved). We ran for years 904 * without it. I think there was only a interrupt latency problem. 905 * The main caller, msleep(), does an splx() a couple of instructions 906 * after calling here. The buggy caller, issignal(), usually calls 907 * here at spl0() and sometimes returns at splhigh(). The process 908 * then runs for a little too long at splhigh(). The ipl gets fixed 909 * when the process returns to user mode (or earlier). 910 * 911 * It would probably be better to always call here at spl0(). Callers 912 * are prepared to give up control to another process, so they must 913 * be prepared to be interrupted. The clock stuff here may not 914 * actually need splstatclock(). 915 */ 916 x = splstatclock(); 917 918 mtx_assert(&sched_lock, MA_OWNED | MA_NOTRECURSED); 919 920#ifdef SIMPLELOCK_DEBUG 921 if (p->p_simple_locks) 922 printf("sleep: holding simple lock\n"); 923#endif 924 /* 925 * Compute the amount of time during which the current 926 * process was running, and add that to its total so far. 927 */ 928 microuptime(&new_switchtime); 929 if (timevalcmp(&new_switchtime, PCPU_PTR(switchtime), <)) { 930#if 0 931 /* XXX: This doesn't play well with sched_lock right now. */ 932 printf("microuptime() went backwards (%ld.%06ld -> %ld.%06ld)\n", 933 PCPU_GET(switchtime.tv_sec), PCPU_GET(switchtime.tv_usec), 934 new_switchtime.tv_sec, new_switchtime.tv_usec); 935#endif 936 new_switchtime = PCPU_GET(switchtime); 937 } else { 938 p->p_runtime += (new_switchtime.tv_usec - PCPU_GET(switchtime.tv_usec)) + 939 (new_switchtime.tv_sec - PCPU_GET(switchtime.tv_sec)) * 940 (int64_t)1000000; 941 } 942 943#if 0 944 /* 945 * Check if the process exceeds its cpu resource allocation. 946 * If over max, kill it. 947 * 948 * XXX drop sched_lock, pickup Giant 949 */ 950 if (p->p_stat != SZOMB && p->p_limit->p_cpulimit != RLIM_INFINITY && 951 p->p_runtime > p->p_limit->p_cpulimit) { 952 rlim = &p->p_rlimit[RLIMIT_CPU]; 953 if (p->p_runtime / (rlim_t)1000000 >= rlim->rlim_max) { 954 mtx_exit(&sched_lock, MTX_SPIN); 955 killproc(p, "exceeded maximum CPU limit"); 956 mtx_enter(&sched_lock, MTX_SPIN); 957 } else { 958 mtx_exit(&sched_lock, MTX_SPIN); 959 psignal(p, SIGXCPU); 960 mtx_enter(&sched_lock, MTX_SPIN); 961 if (rlim->rlim_cur < rlim->rlim_max) { 962 /* XXX: we should make a private copy */ 963 rlim->rlim_cur += 5; 964 } 965 } 966 } 967#endif 968 969 /* 970 * Pick a new current process and record its start time. 971 */ 972 cnt.v_swtch++; 973 PCPU_SET(switchtime, new_switchtime); 974 CTR4(KTR_PROC, "mi_switch: old proc %p (pid %d, %s), schedlock %p", 975 p, p->p_pid, p->p_comm, (void *) sched_lock.mtx_lock); 976 cpu_switch(); 977 CTR4(KTR_PROC, "mi_switch: new proc %p (pid %d, %s), schedlock %p", 978 p, p->p_pid, p->p_comm, (void *) sched_lock.mtx_lock); 979 if (PCPU_GET(switchtime.tv_sec) == 0) 980 microuptime(PCPU_PTR(switchtime)); 981 PCPU_SET(switchticks, ticks); 982 splx(x); 983} 984 985/* 986 * Change process state to be runnable, 987 * placing it on the run queue if it is in memory, 988 * and awakening the swapper if it isn't in memory. 989 */ 990void 991setrunnable(p) 992 register struct proc *p; 993{ 994 register int s; 995 996 s = splhigh(); 997 mtx_enter(&sched_lock, MTX_SPIN); 998 switch (p->p_stat) { 999 case 0: 1000 case SRUN: 1001 case SZOMB: 1002 case SWAIT: 1003 default: 1004 panic("setrunnable"); 1005 case SSTOP: 1006 case SSLEEP: /* e.g. when sending signals */ 1007 if (p->p_sflag & PS_CVWAITQ) 1008 cv_waitq_remove(p); 1009 else 1010 unsleep(p); 1011 break; 1012 1013 case SIDL: 1014 break; 1015 } 1016 p->p_stat = SRUN; 1017 if (p->p_sflag & PS_INMEM) 1018 setrunqueue(p); 1019 splx(s); 1020 if (p->p_slptime > 1) 1021 updatepri(p); 1022 p->p_slptime = 0; 1023 if ((p->p_sflag & PS_INMEM) == 0) { 1024 p->p_sflag |= PS_SWAPINREQ; 1025 wakeup((caddr_t)&proc0); 1026 } 1027 else 1028 maybe_resched(p); 1029 mtx_exit(&sched_lock, MTX_SPIN); 1030} 1031 1032/* 1033 * Compute the priority of a process when running in user mode. 1034 * Arrange to reschedule if the resulting priority is better 1035 * than that of the current process. 1036 */ 1037void 1038resetpriority(p) 1039 register struct proc *p; 1040{ 1041 register unsigned int newpriority; 1042 1043 mtx_enter(&sched_lock, MTX_SPIN); 1044 if (p->p_rtprio.type == RTP_PRIO_NORMAL) { 1045 newpriority = PUSER + p->p_estcpu / INVERSE_ESTCPU_WEIGHT + 1046 NICE_WEIGHT * (p->p_nice - PRIO_MIN); 1047 newpriority = min(newpriority, MAXPRI); 1048 p->p_usrpri = newpriority; 1049 } 1050 maybe_resched(p); 1051 mtx_exit(&sched_lock, MTX_SPIN); 1052} 1053 1054/* ARGSUSED */ 1055static void 1056sched_setup(dummy) 1057 void *dummy; 1058{ 1059 1060 callout_init(&schedcpu_callout, 1); 1061 callout_init(&roundrobin_callout, 0); 1062 1063 /* Kick off timeout driven events by calling first time. */ 1064 roundrobin(NULL); 1065 schedcpu(NULL); 1066} 1067 1068/* 1069 * We adjust the priority of the current process. The priority of 1070 * a process gets worse as it accumulates CPU time. The cpu usage 1071 * estimator (p_estcpu) is increased here. resetpriority() will 1072 * compute a different priority each time p_estcpu increases by 1073 * INVERSE_ESTCPU_WEIGHT 1074 * (until MAXPRI is reached). The cpu usage estimator ramps up 1075 * quite quickly when the process is running (linearly), and decays 1076 * away exponentially, at a rate which is proportionally slower when 1077 * the system is busy. The basic principle is that the system will 1078 * 90% forget that the process used a lot of CPU time in 5 * loadav 1079 * seconds. This causes the system to favor processes which haven't 1080 * run much recently, and to round-robin among other processes. 1081 */ 1082void 1083schedclock(p) 1084 struct proc *p; 1085{ 1086 1087 p->p_cpticks++; 1088 p->p_estcpu = ESTCPULIM(p->p_estcpu + 1); 1089 if ((p->p_estcpu % INVERSE_ESTCPU_WEIGHT) == 0) { 1090 resetpriority(p); 1091 if (p->p_priority >= PUSER) 1092 p->p_priority = p->p_usrpri; 1093 } 1094} 1095 1096/* 1097 * General purpose yield system call 1098 */ 1099int 1100yield(struct proc *p, struct yield_args *uap) 1101{ 1102 int s; 1103 1104 p->p_retval[0] = 0; 1105 1106 s = splhigh(); 1107 mtx_enter(&sched_lock, MTX_SPIN); 1108 DROP_GIANT_NOSWITCH(); 1109 p->p_priority = MAXPRI; 1110 setrunqueue(p); 1111 p->p_stats->p_ru.ru_nvcsw++; 1112 mi_switch(); 1113 mtx_exit(&sched_lock, MTX_SPIN); 1114 PICKUP_GIANT(); 1115 splx(s); 1116 1117 return (0); 1118} 1119