kern_synch.c revision 79130
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 79130 2001-07-03 07:53:35Z 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/kernel.h> 48#include <sys/ktr.h> 49#include <sys/lock.h> 50#include <sys/mutex.h> 51#include <sys/proc.h> 52#include <sys/resourcevar.h> 53#include <sys/signalvar.h> 54#include <sys/smp.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 68static void sched_setup __P((void *dummy)); 69SYSINIT(sched_setup, SI_SUB_KICK_SCHEDULER, SI_ORDER_FIRST, sched_setup, NULL) 70 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 void endtsleep __P((void *)); 79static void roundrobin __P((void *arg)); 80static void schedcpu __P((void *arg)); 81 82static int 83sysctl_kern_quantum(SYSCTL_HANDLER_ARGS) 84{ 85 int error, new_val; 86 87 new_val = sched_quantum * tick; 88 error = sysctl_handle_int(oidp, &new_val, 0, req); 89 if (error != 0 || req->newptr == NULL) 90 return (error); 91 if (new_val < tick) 92 return (EINVAL); 93 sched_quantum = new_val / tick; 94 hogticks = 2 * sched_quantum; 95 return (0); 96} 97 98SYSCTL_PROC(_kern, OID_AUTO, quantum, CTLTYPE_INT|CTLFLAG_RW, 99 0, sizeof sched_quantum, sysctl_kern_quantum, "I", ""); 100 101/* 102 * Arrange to reschedule if necessary, taking the priorities and 103 * schedulers into account. 104 */ 105void 106maybe_resched(p) 107 struct proc *p; 108{ 109 110 mtx_assert(&sched_lock, MA_OWNED); 111 if (p->p_pri.pri_level < curproc->p_pri.pri_level) 112 need_resched(curproc); 113} 114 115int 116roundrobin_interval(void) 117{ 118 return (sched_quantum); 119} 120 121/* 122 * Force switch among equal priority processes every 100ms. 123 * We don't actually need to force a context switch of the current process. 124 * The act of firing the event triggers a context switch to softclock() and 125 * then switching back out again which is equivalent to a preemption, thus 126 * no further work is needed on the local CPU. 127 */ 128/* ARGSUSED */ 129static void 130roundrobin(arg) 131 void *arg; 132{ 133 134#ifdef SMP 135 mtx_lock_spin(&sched_lock); 136 forward_roundrobin(); 137 mtx_unlock_spin(&sched_lock); 138#endif 139 140 callout_reset(&roundrobin_callout, sched_quantum, roundrobin, NULL); 141} 142 143/* 144 * Constants for digital decay and forget: 145 * 90% of (p_estcpu) usage in 5 * loadav time 146 * 95% of (p_pctcpu) usage in 60 seconds (load insensitive) 147 * Note that, as ps(1) mentions, this can let percentages 148 * total over 100% (I've seen 137.9% for 3 processes). 149 * 150 * Note that schedclock() updates p_estcpu and p_cpticks asynchronously. 151 * 152 * We wish to decay away 90% of p_estcpu in (5 * loadavg) seconds. 153 * That is, the system wants to compute a value of decay such 154 * that the following for loop: 155 * for (i = 0; i < (5 * loadavg); i++) 156 * p_estcpu *= decay; 157 * will compute 158 * p_estcpu *= 0.1; 159 * for all values of loadavg: 160 * 161 * Mathematically this loop can be expressed by saying: 162 * decay ** (5 * loadavg) ~= .1 163 * 164 * The system computes decay as: 165 * decay = (2 * loadavg) / (2 * loadavg + 1) 166 * 167 * We wish to prove that the system's computation of decay 168 * will always fulfill the equation: 169 * decay ** (5 * loadavg) ~= .1 170 * 171 * If we compute b as: 172 * b = 2 * loadavg 173 * then 174 * decay = b / (b + 1) 175 * 176 * We now need to prove two things: 177 * 1) Given factor ** (5 * loadavg) ~= .1, prove factor == b/(b+1) 178 * 2) Given b/(b+1) ** power ~= .1, prove power == (5 * loadavg) 179 * 180 * Facts: 181 * For x close to zero, exp(x) =~ 1 + x, since 182 * exp(x) = 0! + x**1/1! + x**2/2! + ... . 183 * therefore exp(-1/b) =~ 1 - (1/b) = (b-1)/b. 184 * For x close to zero, ln(1+x) =~ x, since 185 * ln(1+x) = x - x**2/2 + x**3/3 - ... -1 < x < 1 186 * therefore ln(b/(b+1)) = ln(1 - 1/(b+1)) =~ -1/(b+1). 187 * ln(.1) =~ -2.30 188 * 189 * Proof of (1): 190 * Solve (factor)**(power) =~ .1 given power (5*loadav): 191 * solving for factor, 192 * ln(factor) =~ (-2.30/5*loadav), or 193 * factor =~ exp(-1/((5/2.30)*loadav)) =~ exp(-1/(2*loadav)) = 194 * exp(-1/b) =~ (b-1)/b =~ b/(b+1). QED 195 * 196 * Proof of (2): 197 * Solve (factor)**(power) =~ .1 given factor == (b/(b+1)): 198 * solving for power, 199 * power*ln(b/(b+1)) =~ -2.30, or 200 * power =~ 2.3 * (b + 1) = 4.6*loadav + 2.3 =~ 5*loadav. QED 201 * 202 * Actual power values for the implemented algorithm are as follows: 203 * loadav: 1 2 3 4 204 * power: 5.68 10.32 14.94 19.55 205 */ 206 207/* calculations for digital decay to forget 90% of usage in 5*loadav sec */ 208#define loadfactor(loadav) (2 * (loadav)) 209#define decay_cpu(loadfac, cpu) (((loadfac) * (cpu)) / ((loadfac) + FSCALE)) 210 211/* decay 95% of `p_pctcpu' in 60 seconds; see CCPU_SHIFT before changing */ 212static fixpt_t ccpu = 0.95122942450071400909 * FSCALE; /* exp(-1/20) */ 213SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, ""); 214 215/* kernel uses `FSCALE', userland (SHOULD) use kern.fscale */ 216static int fscale __unused = FSCALE; 217SYSCTL_INT(_kern, OID_AUTO, fscale, CTLFLAG_RD, 0, FSCALE, ""); 218 219/* 220 * If `ccpu' is not equal to `exp(-1/20)' and you still want to use the 221 * faster/more-accurate formula, you'll have to estimate CCPU_SHIFT below 222 * and possibly adjust FSHIFT in "param.h" so that (FSHIFT >= CCPU_SHIFT). 223 * 224 * To estimate CCPU_SHIFT for exp(-1/20), the following formula was used: 225 * 1 - exp(-1/20) ~= 0.0487 ~= 0.0488 == 1 (fixed pt, *11* bits). 226 * 227 * If you don't want to bother with the faster/more-accurate formula, you 228 * can set CCPU_SHIFT to (FSHIFT + 1) which will use a slower/less-accurate 229 * (more general) method of calculating the %age of CPU used by a process. 230 */ 231#define CCPU_SHIFT 11 232 233/* 234 * Recompute process priorities, every hz ticks. 235 * MP-safe, called without the Giant mutex. 236 */ 237/* ARGSUSED */ 238static void 239schedcpu(arg) 240 void *arg; 241{ 242 register fixpt_t loadfac = loadfactor(averunnable.ldavg[0]); 243 register struct proc *p; 244 register int realstathz; 245 246 realstathz = stathz ? stathz : hz; 247 sx_slock(&allproc_lock); 248 LIST_FOREACH(p, &allproc, p_list) { 249 /* 250 * Increment time in/out of memory and sleep time 251 * (if sleeping). We ignore overflow; with 16-bit int's 252 * (remember them?) overflow takes 45 days. 253 if (p->p_stat == SWAIT) 254 continue; 255 */ 256 mtx_lock_spin(&sched_lock); 257 p->p_swtime++; 258 if (p->p_stat == SSLEEP || p->p_stat == SSTOP) 259 p->p_slptime++; 260 p->p_pctcpu = (p->p_pctcpu * ccpu) >> FSHIFT; 261 /* 262 * If the process has slept the entire second, 263 * stop recalculating its priority until it wakes up. 264 */ 265 if (p->p_slptime > 1) { 266 mtx_unlock_spin(&sched_lock); 267 continue; 268 } 269 270 /* 271 * p_pctcpu is only for ps. 272 */ 273#if (FSHIFT >= CCPU_SHIFT) 274 p->p_pctcpu += (realstathz == 100)? 275 ((fixpt_t) p->p_cpticks) << (FSHIFT - CCPU_SHIFT): 276 100 * (((fixpt_t) p->p_cpticks) 277 << (FSHIFT - CCPU_SHIFT)) / realstathz; 278#else 279 p->p_pctcpu += ((FSCALE - ccpu) * 280 (p->p_cpticks * FSCALE / realstathz)) >> FSHIFT; 281#endif 282 p->p_cpticks = 0; 283 p->p_estcpu = decay_cpu(loadfac, p->p_estcpu); 284 resetpriority(p); 285 if (p->p_pri.pri_level >= PUSER) { 286 if ((p != curproc) && 287#ifdef SMP 288 p->p_oncpu == NOCPU && /* idle */ 289#endif 290 p->p_stat == SRUN && 291 (p->p_sflag & PS_INMEM) && 292 (p->p_pri.pri_level / RQ_PPQ) != 293 (p->p_pri.pri_user / RQ_PPQ)) { 294 remrunqueue(p); 295 p->p_pri.pri_level = p->p_pri.pri_user; 296 setrunqueue(p); 297 } else 298 p->p_pri.pri_level = p->p_pri.pri_user; 299 } 300 mtx_unlock_spin(&sched_lock); 301 } 302 sx_sunlock(&allproc_lock); 303 vmmeter(); 304 wakeup((caddr_t)&lbolt); 305 callout_reset(&schedcpu_callout, hz, schedcpu, NULL); 306} 307 308/* 309 * Recalculate the priority of a process after it has slept for a while. 310 * For all load averages >= 1 and max p_estcpu of 255, sleeping for at 311 * least six times the loadfactor will decay p_estcpu to zero. 312 */ 313void 314updatepri(p) 315 register struct proc *p; 316{ 317 register unsigned int newcpu = p->p_estcpu; 318 register fixpt_t loadfac = loadfactor(averunnable.ldavg[0]); 319 320 if (p->p_slptime > 5 * loadfac) 321 p->p_estcpu = 0; 322 else { 323 p->p_slptime--; /* the first time was done in schedcpu */ 324 while (newcpu && --p->p_slptime) 325 newcpu = decay_cpu(loadfac, newcpu); 326 p->p_estcpu = newcpu; 327 } 328 resetpriority(p); 329} 330 331/* 332 * We're only looking at 7 bits of the address; everything is 333 * aligned to 4, lots of things are aligned to greater powers 334 * of 2. Shift right by 8, i.e. drop the bottom 256 worth. 335 */ 336#define TABLESIZE 128 337static TAILQ_HEAD(slpquehead, proc) slpque[TABLESIZE]; 338#define LOOKUP(x) (((intptr_t)(x) >> 8) & (TABLESIZE - 1)) 339 340void 341sleepinit(void) 342{ 343 int i; 344 345 sched_quantum = hz/10; 346 hogticks = 2 * sched_quantum; 347 for (i = 0; i < TABLESIZE; i++) 348 TAILQ_INIT(&slpque[i]); 349} 350 351/* 352 * General sleep call. Suspends the current process until a wakeup is 353 * performed on the specified identifier. The process will then be made 354 * runnable with the specified priority. Sleeps at most timo/hz seconds 355 * (0 means no timeout). If pri includes PCATCH flag, signals are checked 356 * before and after sleeping, else signals are not checked. Returns 0 if 357 * awakened, EWOULDBLOCK if the timeout expires. If PCATCH is set and a 358 * signal needs to be delivered, ERESTART is returned if the current system 359 * call should be restarted if possible, and EINTR is returned if the system 360 * call should be interrupted by the signal (return EINTR). 361 * 362 * The mutex argument is exited before the caller is suspended, and 363 * entered before msleep returns. If priority includes the PDROP 364 * flag the mutex is not entered before returning. 365 */ 366int 367msleep(ident, mtx, priority, wmesg, timo) 368 void *ident; 369 struct mtx *mtx; 370 int priority, timo; 371 const char *wmesg; 372{ 373 struct proc *p = curproc; 374 int sig, catch = priority & PCATCH; 375 int rval = 0; 376 WITNESS_SAVE_DECL(mtx); 377 378#ifdef KTRACE 379 if (p && KTRPOINT(p, KTR_CSW)) 380 ktrcsw(p->p_tracep, 1, 0); 381#endif 382 WITNESS_SLEEP(0, &mtx->mtx_object); 383 KASSERT(timo != 0 || mtx_owned(&Giant) || mtx != NULL, 384 ("sleeping without a mutex")); 385 mtx_lock_spin(&sched_lock); 386 if (cold || panicstr) { 387 /* 388 * After a panic, or during autoconfiguration, 389 * just give interrupts a chance, then just return; 390 * don't run any other procs or panic below, 391 * in case this is the idle process and already asleep. 392 */ 393 if (mtx != NULL && priority & PDROP) 394 mtx_unlock_flags(mtx, MTX_NOSWITCH); 395 mtx_unlock_spin(&sched_lock); 396 return (0); 397 } 398 399 DROP_GIANT_NOSWITCH(); 400 401 if (mtx != NULL) { 402 mtx_assert(mtx, MA_OWNED | MA_NOTRECURSED); 403 WITNESS_SAVE(&mtx->mtx_object, mtx); 404 mtx_unlock_flags(mtx, MTX_NOSWITCH); 405 if (priority & PDROP) 406 mtx = NULL; 407 } 408 409 KASSERT(p != NULL, ("msleep1")); 410 KASSERT(ident != NULL && p->p_stat == SRUN, ("msleep")); 411 /* 412 * Process may be sitting on a slpque if asleep() was called, remove 413 * it before re-adding. 414 */ 415 if (p->p_wchan != NULL) 416 unsleep(p); 417 418 p->p_wchan = ident; 419 p->p_wmesg = wmesg; 420 p->p_slptime = 0; 421 p->p_pri.pri_level = priority & PRIMASK; 422 CTR5(KTR_PROC, "msleep: proc %p (pid %d, %s) on %s (%p)", p, p->p_pid, 423 p->p_comm, wmesg, ident); 424 TAILQ_INSERT_TAIL(&slpque[LOOKUP(ident)], p, p_slpq); 425 if (timo) 426 callout_reset(&p->p_slpcallout, timo, endtsleep, p); 427 /* 428 * We put ourselves on the sleep queue and start our timeout 429 * before calling CURSIG, as we could stop there, and a wakeup 430 * or a SIGCONT (or both) could occur while we were stopped. 431 * A SIGCONT would cause us to be marked as SSLEEP 432 * without resuming us, thus we must be ready for sleep 433 * when CURSIG is called. If the wakeup happens while we're 434 * stopped, p->p_wchan will be 0 upon return from CURSIG. 435 */ 436 if (catch) { 437 CTR3(KTR_PROC, "msleep caught: proc %p (pid %d, %s)", p, 438 p->p_pid, p->p_comm); 439 p->p_sflag |= PS_SINTR; 440 mtx_unlock_spin(&sched_lock); 441 PROC_LOCK(p); 442 sig = CURSIG(p); 443 mtx_lock_spin(&sched_lock); 444 PROC_UNLOCK_NOSWITCH(p); 445 if (sig != 0) { 446 if (p->p_wchan) 447 unsleep(p); 448 } else if (p->p_wchan == NULL) 449 catch = 0; 450 } else 451 sig = 0; 452 if (p->p_wchan != NULL) { 453 p->p_stat = SSLEEP; 454 p->p_stats->p_ru.ru_nvcsw++; 455 mi_switch(); 456 } 457 CTR3(KTR_PROC, "msleep resume: proc %p (pid %d, %s)", p, p->p_pid, 458 p->p_comm); 459 KASSERT(p->p_stat == SRUN, ("running but not SRUN")); 460 p->p_sflag &= ~PS_SINTR; 461 if (p->p_sflag & PS_TIMEOUT) { 462 p->p_sflag &= ~PS_TIMEOUT; 463 if (sig == 0) 464 rval = EWOULDBLOCK; 465 } else if (timo) 466 callout_stop(&p->p_slpcallout); 467 mtx_unlock_spin(&sched_lock); 468 469 if (rval == 0 && catch) { 470 PROC_LOCK(p); 471 /* XXX: shouldn't we always be calling CURSIG() */ 472 if (sig != 0 || (sig = CURSIG(p))) { 473 if (SIGISMEMBER(p->p_sigacts->ps_sigintr, sig)) 474 rval = EINTR; 475 else 476 rval = ERESTART; 477 } 478 PROC_UNLOCK(p); 479 } 480 PICKUP_GIANT(); 481#ifdef KTRACE 482 mtx_lock(&Giant); 483 if (KTRPOINT(p, KTR_CSW)) 484 ktrcsw(p->p_tracep, 0, 0); 485 mtx_unlock(&Giant); 486#endif 487 if (mtx != NULL) { 488 mtx_lock(mtx); 489 WITNESS_RESTORE(&mtx->mtx_object, mtx); 490 } 491 return (rval); 492} 493 494/* 495 * asleep() - async sleep call. Place process on wait queue and return 496 * immediately without blocking. The process stays runnable until mawait() 497 * is called. If ident is NULL, remove process from wait queue if it is still 498 * on one. 499 * 500 * Only the most recent sleep condition is effective when making successive 501 * calls to asleep() or when calling msleep(). 502 * 503 * The timeout, if any, is not initiated until mawait() is called. The sleep 504 * priority, signal, and timeout is specified in the asleep() call but may be 505 * overriden in the mawait() call. 506 * 507 * <<<<<<<< EXPERIMENTAL, UNTESTED >>>>>>>>>> 508 */ 509 510int 511asleep(void *ident, int priority, const char *wmesg, int timo) 512{ 513 struct proc *p = curproc; 514 515 /* 516 * Remove preexisting wait condition (if any) and place process 517 * on appropriate slpque, but do not put process to sleep. 518 */ 519 520 mtx_lock_spin(&sched_lock); 521 522 if (p->p_wchan != NULL) 523 unsleep(p); 524 525 if (ident) { 526 p->p_wchan = ident; 527 p->p_wmesg = wmesg; 528 p->p_slptime = 0; 529 p->p_asleep.as_priority = priority; 530 p->p_asleep.as_timo = timo; 531 TAILQ_INSERT_TAIL(&slpque[LOOKUP(ident)], p, p_slpq); 532 } 533 534 mtx_unlock_spin(&sched_lock); 535 536 return(0); 537} 538 539/* 540 * mawait() - wait for async condition to occur. The process blocks until 541 * wakeup() is called on the most recent asleep() address. If wakeup is called 542 * prior to mawait(), mawait() winds up being a NOP. 543 * 544 * If mawait() is called more then once (without an intervening asleep() call), 545 * mawait() is still effectively a NOP but it calls mi_switch() to give other 546 * processes some cpu before returning. The process is left runnable. 547 * 548 * <<<<<<<< EXPERIMENTAL, UNTESTED >>>>>>>>>> 549 */ 550 551int 552mawait(struct mtx *mtx, int priority, int timo) 553{ 554 struct proc *p = curproc; 555 int rval = 0; 556 WITNESS_SAVE_DECL(mtx); 557 558 WITNESS_SLEEP(0, &mtx->mtx_object); 559 KASSERT(timo > 0 || mtx_owned(&Giant) || mtx != NULL, 560 ("sleeping without a mutex")); 561 mtx_lock_spin(&sched_lock); 562 DROP_GIANT_NOSWITCH(); 563 if (mtx != NULL) { 564 mtx_assert(mtx, MA_OWNED | MA_NOTRECURSED); 565 WITNESS_SAVE(&mtx->mtx_object, mtx); 566 mtx_unlock_flags(mtx, MTX_NOSWITCH); 567 if (priority & PDROP) 568 mtx = NULL; 569 } 570 571 if (p->p_wchan != NULL) { 572 int sig; 573 int catch; 574 575#ifdef KTRACE 576 if (p && KTRPOINT(p, KTR_CSW)) 577 ktrcsw(p->p_tracep, 1, 0); 578#endif 579 /* 580 * The call to mawait() can override defaults specified in 581 * the original asleep(). 582 */ 583 if (priority < 0) 584 priority = p->p_asleep.as_priority; 585 if (timo < 0) 586 timo = p->p_asleep.as_timo; 587 588 /* 589 * Install timeout 590 */ 591 592 if (timo) 593 callout_reset(&p->p_slpcallout, timo, endtsleep, p); 594 595 sig = 0; 596 catch = priority & PCATCH; 597 598 if (catch) { 599 p->p_sflag |= PS_SINTR; 600 mtx_unlock_spin(&sched_lock); 601 PROC_LOCK(p); 602 sig = CURSIG(p); 603 mtx_lock_spin(&sched_lock); 604 PROC_UNLOCK_NOSWITCH(p); 605 if (sig != 0) { 606 if (p->p_wchan) 607 unsleep(p); 608 } else if (p->p_wchan == NULL) 609 catch = 0; 610 } 611 if (p->p_wchan != NULL) { 612 p->p_stat = SSLEEP; 613 p->p_stats->p_ru.ru_nvcsw++; 614 mi_switch(); 615 } 616 KASSERT(p->p_stat == SRUN, ("running but not SRUN")); 617 p->p_sflag &= ~PS_SINTR; 618 if (p->p_sflag & PS_TIMEOUT) { 619 p->p_sflag &= ~PS_TIMEOUT; 620 if (sig == 0) 621 rval = EWOULDBLOCK; 622 } else if (timo) 623 callout_stop(&p->p_slpcallout); 624 mtx_unlock_spin(&sched_lock); 625 if (rval == 0 && catch) { 626 PROC_LOCK(p); 627 if (sig != 0 || (sig = CURSIG(p))) { 628 if (SIGISMEMBER(p->p_sigacts->ps_sigintr, sig)) 629 rval = EINTR; 630 else 631 rval = ERESTART; 632 } 633 PROC_UNLOCK(p); 634 } 635#ifdef KTRACE 636 mtx_lock(&Giant); 637 if (KTRPOINT(p, KTR_CSW)) 638 ktrcsw(p->p_tracep, 0, 0); 639 mtx_unlock(&Giant); 640#endif 641 } else { 642 /* 643 * If as_priority is 0, mawait() has been called without an 644 * intervening asleep(). We are still effectively a NOP, 645 * but we call mi_switch() for safety. 646 */ 647 648 if (p->p_asleep.as_priority == 0) { 649 p->p_stats->p_ru.ru_nvcsw++; 650 mi_switch(); 651 } 652 mtx_unlock_spin(&sched_lock); 653 } 654 655 /* 656 * clear p_asleep.as_priority as an indication that mawait() has been 657 * called. If mawait() is called again without an intervening asleep(), 658 * mawait() is still effectively a NOP but the above mi_switch() code 659 * is triggered as a safety. 660 */ 661 if (rval == 0) 662 p->p_asleep.as_priority = 0; 663 664 PICKUP_GIANT(); 665 if (mtx != NULL) { 666 mtx_lock(mtx); 667 WITNESS_RESTORE(&mtx->mtx_object, mtx); 668 } 669 return (rval); 670} 671 672/* 673 * Implement timeout for msleep or asleep()/mawait() 674 * 675 * If process hasn't been awakened (wchan non-zero), 676 * set timeout flag and undo the sleep. If proc 677 * is stopped, just unsleep so it will remain stopped. 678 * MP-safe, called without the Giant mutex. 679 */ 680static void 681endtsleep(arg) 682 void *arg; 683{ 684 register struct proc *p; 685 686 p = (struct proc *)arg; 687 CTR3(KTR_PROC, "endtsleep: proc %p (pid %d, %s)", p, p->p_pid, 688 p->p_comm); 689 mtx_lock_spin(&sched_lock); 690 if (p->p_wchan) { 691 if (p->p_stat == SSLEEP) 692 setrunnable(p); 693 else 694 unsleep(p); 695 p->p_sflag |= PS_TIMEOUT; 696 } 697 mtx_unlock_spin(&sched_lock); 698} 699 700/* 701 * Remove a process from its wait queue 702 */ 703void 704unsleep(p) 705 register struct proc *p; 706{ 707 708 mtx_lock_spin(&sched_lock); 709 if (p->p_wchan) { 710 TAILQ_REMOVE(&slpque[LOOKUP(p->p_wchan)], p, p_slpq); 711 p->p_wchan = NULL; 712 } 713 mtx_unlock_spin(&sched_lock); 714} 715 716/* 717 * Make all processes sleeping on the specified identifier runnable. 718 */ 719void 720wakeup(ident) 721 register void *ident; 722{ 723 register struct slpquehead *qp; 724 register struct proc *p; 725 726 mtx_lock_spin(&sched_lock); 727 qp = &slpque[LOOKUP(ident)]; 728restart: 729 TAILQ_FOREACH(p, qp, p_slpq) { 730 if (p->p_wchan == ident) { 731 TAILQ_REMOVE(qp, p, p_slpq); 732 p->p_wchan = NULL; 733 if (p->p_stat == SSLEEP) { 734 /* OPTIMIZED EXPANSION OF setrunnable(p); */ 735 CTR3(KTR_PROC, "wakeup: proc %p (pid %d, %s)", 736 p, p->p_pid, p->p_comm); 737 if (p->p_slptime > 1) 738 updatepri(p); 739 p->p_slptime = 0; 740 p->p_stat = SRUN; 741 if (p->p_sflag & PS_INMEM) { 742 setrunqueue(p); 743 maybe_resched(p); 744 } else { 745 p->p_sflag |= PS_SWAPINREQ; 746 wakeup((caddr_t)&proc0); 747 } 748 /* END INLINE EXPANSION */ 749 goto restart; 750 } 751 } 752 } 753 mtx_unlock_spin(&sched_lock); 754} 755 756/* 757 * Make a process sleeping on the specified identifier runnable. 758 * May wake more than one process if a target process is currently 759 * swapped out. 760 */ 761void 762wakeup_one(ident) 763 register void *ident; 764{ 765 register struct slpquehead *qp; 766 register struct proc *p; 767 768 mtx_lock_spin(&sched_lock); 769 qp = &slpque[LOOKUP(ident)]; 770 771 TAILQ_FOREACH(p, qp, p_slpq) { 772 if (p->p_wchan == ident) { 773 TAILQ_REMOVE(qp, p, p_slpq); 774 p->p_wchan = NULL; 775 if (p->p_stat == SSLEEP) { 776 /* OPTIMIZED EXPANSION OF setrunnable(p); */ 777 CTR3(KTR_PROC, "wakeup1: proc %p (pid %d, %s)", 778 p, p->p_pid, p->p_comm); 779 if (p->p_slptime > 1) 780 updatepri(p); 781 p->p_slptime = 0; 782 p->p_stat = SRUN; 783 if (p->p_sflag & PS_INMEM) { 784 setrunqueue(p); 785 maybe_resched(p); 786 break; 787 } else { 788 p->p_sflag |= PS_SWAPINREQ; 789 wakeup((caddr_t)&proc0); 790 } 791 /* END INLINE EXPANSION */ 792 } 793 } 794 } 795 mtx_unlock_spin(&sched_lock); 796} 797 798/* 799 * The machine independent parts of mi_switch(). 800 */ 801void 802mi_switch() 803{ 804 struct timeval new_switchtime; 805 register struct proc *p = curproc; /* XXX */ 806#if 0 807 register struct rlimit *rlim; 808#endif 809 critical_t sched_crit; 810 u_int sched_nest; 811 812 mtx_assert(&sched_lock, MA_OWNED | MA_NOTRECURSED); 813 814 /* 815 * Compute the amount of time during which the current 816 * process was running, and add that to its total so far. 817 */ 818 microuptime(&new_switchtime); 819 if (timevalcmp(&new_switchtime, PCPU_PTR(switchtime), <)) { 820#if 0 821 /* XXX: This doesn't play well with sched_lock right now. */ 822 printf("microuptime() went backwards (%ld.%06ld -> %ld.%06ld)\n", 823 PCPU_GET(switchtime.tv_sec), PCPU_GET(switchtime.tv_usec), 824 new_switchtime.tv_sec, new_switchtime.tv_usec); 825#endif 826 new_switchtime = PCPU_GET(switchtime); 827 } else { 828 p->p_runtime += (new_switchtime.tv_usec - PCPU_GET(switchtime.tv_usec)) + 829 (new_switchtime.tv_sec - PCPU_GET(switchtime.tv_sec)) * 830 (int64_t)1000000; 831 } 832 833#if 0 834 /* 835 * Check if the process exceeds its cpu resource allocation. 836 * If over max, kill it. 837 * 838 * XXX drop sched_lock, pickup Giant 839 */ 840 if (p->p_stat != SZOMB && p->p_limit->p_cpulimit != RLIM_INFINITY && 841 p->p_runtime > p->p_limit->p_cpulimit) { 842 rlim = &p->p_rlimit[RLIMIT_CPU]; 843 if (p->p_runtime / (rlim_t)1000000 >= rlim->rlim_max) { 844 mtx_unlock_spin(&sched_lock); 845 PROC_LOCK(p); 846 killproc(p, "exceeded maximum CPU limit"); 847 mtx_lock_spin(&sched_lock); 848 PROC_UNLOCK_NOSWITCH(p); 849 } else { 850 mtx_unlock_spin(&sched_lock); 851 PROC_LOCK(p); 852 psignal(p, SIGXCPU); 853 mtx_lock_spin(&sched_lock); 854 PROC_UNLOCK_NOSWITCH(p); 855 if (rlim->rlim_cur < rlim->rlim_max) { 856 /* XXX: we should make a private copy */ 857 rlim->rlim_cur += 5; 858 } 859 } 860 } 861#endif 862 863 /* 864 * Pick a new current process and record its start time. 865 */ 866 cnt.v_swtch++; 867 PCPU_SET(switchtime, new_switchtime); 868 CTR3(KTR_PROC, "mi_switch: old proc %p (pid %d, %s)", p, p->p_pid, 869 p->p_comm); 870 sched_crit = sched_lock.mtx_savecrit; 871 sched_nest = sched_lock.mtx_recurse; 872 curproc->p_lastcpu = curproc->p_oncpu; 873 curproc->p_oncpu = NOCPU; 874 clear_resched(curproc); 875 cpu_switch(); 876 curproc->p_oncpu = PCPU_GET(cpuid); 877 sched_lock.mtx_savecrit = sched_crit; 878 sched_lock.mtx_recurse = sched_nest; 879 sched_lock.mtx_lock = (uintptr_t)curproc; 880 CTR3(KTR_PROC, "mi_switch: new proc %p (pid %d, %s)", p, p->p_pid, 881 p->p_comm); 882 if (PCPU_GET(switchtime.tv_sec) == 0) 883 microuptime(PCPU_PTR(switchtime)); 884 PCPU_SET(switchticks, ticks); 885} 886 887/* 888 * Change process state to be runnable, 889 * placing it on the run queue if it is in memory, 890 * and awakening the swapper if it isn't in memory. 891 */ 892void 893setrunnable(p) 894 register struct proc *p; 895{ 896 897 mtx_lock_spin(&sched_lock); 898 switch (p->p_stat) { 899 case 0: 900 case SRUN: 901 case SZOMB: 902 case SWAIT: 903 default: 904 panic("setrunnable"); 905 case SSTOP: 906 case SSLEEP: /* e.g. when sending signals */ 907 if (p->p_sflag & PS_CVWAITQ) 908 cv_waitq_remove(p); 909 else 910 unsleep(p); 911 break; 912 913 case SIDL: 914 break; 915 } 916 p->p_stat = SRUN; 917 if (p->p_slptime > 1) 918 updatepri(p); 919 p->p_slptime = 0; 920 if ((p->p_sflag & PS_INMEM) == 0) { 921 p->p_sflag |= PS_SWAPINREQ; 922 wakeup((caddr_t)&proc0); 923 } else { 924 setrunqueue(p); 925 maybe_resched(p); 926 } 927 mtx_unlock_spin(&sched_lock); 928} 929 930/* 931 * Compute the priority of a process when running in user mode. 932 * Arrange to reschedule if the resulting priority is better 933 * than that of the current process. 934 */ 935void 936resetpriority(p) 937 register struct proc *p; 938{ 939 register unsigned int newpriority; 940 941 mtx_lock_spin(&sched_lock); 942 if (p->p_pri.pri_class == PRI_TIMESHARE) { 943 newpriority = PUSER + p->p_estcpu / INVERSE_ESTCPU_WEIGHT + 944 NICE_WEIGHT * (p->p_nice - PRIO_MIN); 945 newpriority = min(max(newpriority, PRI_MIN_TIMESHARE), 946 PRI_MAX_TIMESHARE); 947 p->p_pri.pri_user = newpriority; 948 } 949 maybe_resched(p); 950 mtx_unlock_spin(&sched_lock); 951} 952 953/* ARGSUSED */ 954static void 955sched_setup(dummy) 956 void *dummy; 957{ 958 959 callout_init(&schedcpu_callout, 1); 960 callout_init(&roundrobin_callout, 0); 961 962 /* Kick off timeout driven events by calling first time. */ 963 roundrobin(NULL); 964 schedcpu(NULL); 965} 966 967/* 968 * We adjust the priority of the current process. The priority of 969 * a process gets worse as it accumulates CPU time. The cpu usage 970 * estimator (p_estcpu) is increased here. resetpriority() will 971 * compute a different priority each time p_estcpu increases by 972 * INVERSE_ESTCPU_WEIGHT 973 * (until MAXPRI is reached). The cpu usage estimator ramps up 974 * quite quickly when the process is running (linearly), and decays 975 * away exponentially, at a rate which is proportionally slower when 976 * the system is busy. The basic principle is that the system will 977 * 90% forget that the process used a lot of CPU time in 5 * loadav 978 * seconds. This causes the system to favor processes which haven't 979 * run much recently, and to round-robin among other processes. 980 */ 981void 982schedclock(p) 983 struct proc *p; 984{ 985 986 p->p_cpticks++; 987 p->p_estcpu = ESTCPULIM(p->p_estcpu + 1); 988 if ((p->p_estcpu % INVERSE_ESTCPU_WEIGHT) == 0) { 989 resetpriority(p); 990 if (p->p_pri.pri_level >= PUSER) 991 p->p_pri.pri_level = p->p_pri.pri_user; 992 } 993} 994 995/* 996 * General purpose yield system call 997 */ 998int 999yield(struct proc *p, struct yield_args *uap) 1000{ 1001 1002 p->p_retval[0] = 0; 1003 1004 mtx_lock_spin(&sched_lock); 1005 DROP_GIANT_NOSWITCH(); 1006 p->p_pri.pri_level = PRI_MAX_TIMESHARE; 1007 setrunqueue(p); 1008 p->p_stats->p_ru.ru_nvcsw++; 1009 mi_switch(); 1010 mtx_unlock_spin(&sched_lock); 1011 PICKUP_GIANT(); 1012 1013 return (0); 1014} 1015