kern_synch.c revision 79126
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 79126 2001-07-03 05:33:09Z 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, s; 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 * prevent state changes and protect run queue 272 */ 273 s = splhigh(); 274 275 /* 276 * p_pctcpu is only for ps. 277 */ 278#if (FSHIFT >= CCPU_SHIFT) 279 p->p_pctcpu += (realstathz == 100)? 280 ((fixpt_t) p->p_cpticks) << (FSHIFT - CCPU_SHIFT): 281 100 * (((fixpt_t) p->p_cpticks) 282 << (FSHIFT - CCPU_SHIFT)) / realstathz; 283#else 284 p->p_pctcpu += ((FSCALE - ccpu) * 285 (p->p_cpticks * FSCALE / realstathz)) >> FSHIFT; 286#endif 287 p->p_cpticks = 0; 288 p->p_estcpu = decay_cpu(loadfac, p->p_estcpu); 289 resetpriority(p); 290 if (p->p_pri.pri_level >= PUSER) { 291 if ((p != curproc) && 292#ifdef SMP 293 p->p_oncpu == NOCPU && /* idle */ 294#endif 295 p->p_stat == SRUN && 296 (p->p_sflag & PS_INMEM) && 297 (p->p_pri.pri_level / RQ_PPQ) != 298 (p->p_pri.pri_user / RQ_PPQ)) { 299 remrunqueue(p); 300 p->p_pri.pri_level = p->p_pri.pri_user; 301 setrunqueue(p); 302 } else 303 p->p_pri.pri_level = p->p_pri.pri_user; 304 } 305 mtx_unlock_spin(&sched_lock); 306 splx(s); 307 } 308 sx_sunlock(&allproc_lock); 309 vmmeter(); 310 wakeup((caddr_t)&lbolt); 311 callout_reset(&schedcpu_callout, hz, schedcpu, NULL); 312} 313 314/* 315 * Recalculate the priority of a process after it has slept for a while. 316 * For all load averages >= 1 and max p_estcpu of 255, sleeping for at 317 * least six times the loadfactor will decay p_estcpu to zero. 318 */ 319void 320updatepri(p) 321 register struct proc *p; 322{ 323 register unsigned int newcpu = p->p_estcpu; 324 register fixpt_t loadfac = loadfactor(averunnable.ldavg[0]); 325 326 if (p->p_slptime > 5 * loadfac) 327 p->p_estcpu = 0; 328 else { 329 p->p_slptime--; /* the first time was done in schedcpu */ 330 while (newcpu && --p->p_slptime) 331 newcpu = decay_cpu(loadfac, newcpu); 332 p->p_estcpu = newcpu; 333 } 334 resetpriority(p); 335} 336 337/* 338 * We're only looking at 7 bits of the address; everything is 339 * aligned to 4, lots of things are aligned to greater powers 340 * of 2. Shift right by 8, i.e. drop the bottom 256 worth. 341 */ 342#define TABLESIZE 128 343static TAILQ_HEAD(slpquehead, proc) slpque[TABLESIZE]; 344#define LOOKUP(x) (((intptr_t)(x) >> 8) & (TABLESIZE - 1)) 345 346void 347sleepinit(void) 348{ 349 int i; 350 351 sched_quantum = hz/10; 352 hogticks = 2 * sched_quantum; 353 for (i = 0; i < TABLESIZE; i++) 354 TAILQ_INIT(&slpque[i]); 355} 356 357/* 358 * General sleep call. Suspends the current process until a wakeup is 359 * performed on the specified identifier. The process will then be made 360 * runnable with the specified priority. Sleeps at most timo/hz seconds 361 * (0 means no timeout). If pri includes PCATCH flag, signals are checked 362 * before and after sleeping, else signals are not checked. Returns 0 if 363 * awakened, EWOULDBLOCK if the timeout expires. If PCATCH is set and a 364 * signal needs to be delivered, ERESTART is returned if the current system 365 * call should be restarted if possible, and EINTR is returned if the system 366 * call should be interrupted by the signal (return EINTR). 367 * 368 * The mutex argument is exited before the caller is suspended, and 369 * entered before msleep returns. If priority includes the PDROP 370 * flag the mutex is not entered before returning. 371 */ 372int 373msleep(ident, mtx, priority, wmesg, timo) 374 void *ident; 375 struct mtx *mtx; 376 int priority, timo; 377 const char *wmesg; 378{ 379 struct proc *p = curproc; 380 int sig, catch = priority & PCATCH; 381 int rval = 0; 382 WITNESS_SAVE_DECL(mtx); 383 384#ifdef KTRACE 385 if (p && KTRPOINT(p, KTR_CSW)) 386 ktrcsw(p->p_tracep, 1, 0); 387#endif 388 WITNESS_SLEEP(0, &mtx->mtx_object); 389 KASSERT(timo != 0 || mtx_owned(&Giant) || mtx != NULL, 390 ("sleeping without a mutex")); 391 mtx_lock_spin(&sched_lock); 392 if (cold || panicstr) { 393 /* 394 * After a panic, or during autoconfiguration, 395 * just give interrupts a chance, then just return; 396 * don't run any other procs or panic below, 397 * in case this is the idle process and already asleep. 398 */ 399 if (mtx != NULL && priority & PDROP) 400 mtx_unlock_flags(mtx, MTX_NOSWITCH); 401 mtx_unlock_spin(&sched_lock); 402 return (0); 403 } 404 405 DROP_GIANT_NOSWITCH(); 406 407 if (mtx != NULL) { 408 mtx_assert(mtx, MA_OWNED | MA_NOTRECURSED); 409 WITNESS_SAVE(&mtx->mtx_object, mtx); 410 mtx_unlock_flags(mtx, MTX_NOSWITCH); 411 if (priority & PDROP) 412 mtx = NULL; 413 } 414 415 KASSERT(p != NULL, ("msleep1")); 416 KASSERT(ident != NULL && p->p_stat == SRUN, ("msleep")); 417 /* 418 * Process may be sitting on a slpque if asleep() was called, remove 419 * it before re-adding. 420 */ 421 if (p->p_wchan != NULL) 422 unsleep(p); 423 424 p->p_wchan = ident; 425 p->p_wmesg = wmesg; 426 p->p_slptime = 0; 427 p->p_pri.pri_level = priority & PRIMASK; 428 CTR3(KTR_PROC, "msleep: proc %p (pid %d, %s)", p, p->p_pid, p->p_comm); 429 TAILQ_INSERT_TAIL(&slpque[LOOKUP(ident)], p, p_slpq); 430 if (timo) 431 callout_reset(&p->p_slpcallout, timo, endtsleep, p); 432 /* 433 * We put ourselves on the sleep queue and start our timeout 434 * before calling CURSIG, as we could stop there, and a wakeup 435 * or a SIGCONT (or both) could occur while we were stopped. 436 * A SIGCONT would cause us to be marked as SSLEEP 437 * without resuming us, thus we must be ready for sleep 438 * when CURSIG is called. If the wakeup happens while we're 439 * stopped, p->p_wchan will be 0 upon return from CURSIG. 440 */ 441 if (catch) { 442 CTR3(KTR_PROC, "msleep caught: proc %p (pid %d, %s)", p, 443 p->p_pid, p->p_comm); 444 p->p_sflag |= PS_SINTR; 445 mtx_unlock_spin(&sched_lock); 446 PROC_LOCK(p); 447 sig = CURSIG(p); 448 mtx_lock_spin(&sched_lock); 449 PROC_UNLOCK_NOSWITCH(p); 450 if (sig != 0) { 451 if (p->p_wchan) 452 unsleep(p); 453 } else if (p->p_wchan == NULL) 454 catch = 0; 455 } else 456 sig = 0; 457 if (p->p_wchan != NULL) { 458 p->p_stat = SSLEEP; 459 p->p_stats->p_ru.ru_nvcsw++; 460 mi_switch(); 461 } 462 CTR3(KTR_PROC, "msleep resume: proc %p (pid %d, %s)", p, p->p_pid, 463 p->p_comm); 464 KASSERT(p->p_stat == SRUN, ("running but not SRUN")); 465 p->p_sflag &= ~PS_SINTR; 466 if (p->p_sflag & PS_TIMEOUT) { 467 p->p_sflag &= ~PS_TIMEOUT; 468 if (sig == 0) 469 rval = EWOULDBLOCK; 470 } else if (timo) 471 callout_stop(&p->p_slpcallout); 472 mtx_unlock_spin(&sched_lock); 473 474 if (rval == 0 && catch) { 475 PROC_LOCK(p); 476 /* XXX: shouldn't we always be calling CURSIG() */ 477 if (sig != 0 || (sig = CURSIG(p))) { 478 if (SIGISMEMBER(p->p_sigacts->ps_sigintr, sig)) 479 rval = EINTR; 480 else 481 rval = ERESTART; 482 } 483 PROC_UNLOCK(p); 484 } 485 PICKUP_GIANT(); 486#ifdef KTRACE 487 mtx_lock(&Giant); 488 if (KTRPOINT(p, KTR_CSW)) 489 ktrcsw(p->p_tracep, 0, 0); 490 mtx_unlock(&Giant); 491#endif 492 if (mtx != NULL) { 493 mtx_lock(mtx); 494 WITNESS_RESTORE(&mtx->mtx_object, mtx); 495 } 496 return (rval); 497} 498 499/* 500 * asleep() - async sleep call. Place process on wait queue and return 501 * immediately without blocking. The process stays runnable until mawait() 502 * is called. If ident is NULL, remove process from wait queue if it is still 503 * on one. 504 * 505 * Only the most recent sleep condition is effective when making successive 506 * calls to asleep() or when calling msleep(). 507 * 508 * The timeout, if any, is not initiated until mawait() is called. The sleep 509 * priority, signal, and timeout is specified in the asleep() call but may be 510 * overriden in the mawait() call. 511 * 512 * <<<<<<<< EXPERIMENTAL, UNTESTED >>>>>>>>>> 513 */ 514 515int 516asleep(void *ident, int priority, const char *wmesg, int timo) 517{ 518 struct proc *p = curproc; 519 520 /* 521 * Remove preexisting wait condition (if any) and place process 522 * on appropriate slpque, but do not put process to sleep. 523 */ 524 525 mtx_lock_spin(&sched_lock); 526 527 if (p->p_wchan != NULL) 528 unsleep(p); 529 530 if (ident) { 531 p->p_wchan = ident; 532 p->p_wmesg = wmesg; 533 p->p_slptime = 0; 534 p->p_asleep.as_priority = priority; 535 p->p_asleep.as_timo = timo; 536 TAILQ_INSERT_TAIL(&slpque[LOOKUP(ident)], p, p_slpq); 537 } 538 539 mtx_unlock_spin(&sched_lock); 540 541 return(0); 542} 543 544/* 545 * mawait() - wait for async condition to occur. The process blocks until 546 * wakeup() is called on the most recent asleep() address. If wakeup is called 547 * prior to mawait(), mawait() winds up being a NOP. 548 * 549 * If mawait() is called more then once (without an intervening asleep() call), 550 * mawait() is still effectively a NOP but it calls mi_switch() to give other 551 * processes some cpu before returning. The process is left runnable. 552 * 553 * <<<<<<<< EXPERIMENTAL, UNTESTED >>>>>>>>>> 554 */ 555 556int 557mawait(struct mtx *mtx, int priority, int timo) 558{ 559 struct proc *p = curproc; 560 int rval = 0; 561 WITNESS_SAVE_DECL(mtx); 562 563 WITNESS_SLEEP(0, &mtx->mtx_object); 564 KASSERT(timo > 0 || mtx_owned(&Giant) || mtx != NULL, 565 ("sleeping without a mutex")); 566 mtx_lock_spin(&sched_lock); 567 DROP_GIANT_NOSWITCH(); 568 if (mtx != NULL) { 569 mtx_assert(mtx, MA_OWNED | MA_NOTRECURSED); 570 WITNESS_SAVE(&mtx->mtx_object, mtx); 571 mtx_unlock_flags(mtx, MTX_NOSWITCH); 572 if (priority & PDROP) 573 mtx = NULL; 574 } 575 576 if (p->p_wchan != NULL) { 577 int sig; 578 int catch; 579 580#ifdef KTRACE 581 if (p && KTRPOINT(p, KTR_CSW)) 582 ktrcsw(p->p_tracep, 1, 0); 583#endif 584 /* 585 * The call to mawait() can override defaults specified in 586 * the original asleep(). 587 */ 588 if (priority < 0) 589 priority = p->p_asleep.as_priority; 590 if (timo < 0) 591 timo = p->p_asleep.as_timo; 592 593 /* 594 * Install timeout 595 */ 596 597 if (timo) 598 callout_reset(&p->p_slpcallout, timo, endtsleep, p); 599 600 sig = 0; 601 catch = priority & PCATCH; 602 603 if (catch) { 604 p->p_sflag |= PS_SINTR; 605 mtx_unlock_spin(&sched_lock); 606 PROC_LOCK(p); 607 sig = CURSIG(p); 608 mtx_lock_spin(&sched_lock); 609 PROC_UNLOCK_NOSWITCH(p); 610 if (sig != 0) { 611 if (p->p_wchan) 612 unsleep(p); 613 } else if (p->p_wchan == NULL) 614 catch = 0; 615 } 616 if (p->p_wchan != NULL) { 617 p->p_stat = SSLEEP; 618 p->p_stats->p_ru.ru_nvcsw++; 619 mi_switch(); 620 } 621 KASSERT(p->p_stat == SRUN, ("running but not SRUN")); 622 p->p_sflag &= ~PS_SINTR; 623 if (p->p_sflag & PS_TIMEOUT) { 624 p->p_sflag &= ~PS_TIMEOUT; 625 if (sig == 0) 626 rval = EWOULDBLOCK; 627 } else if (timo) 628 callout_stop(&p->p_slpcallout); 629 mtx_unlock_spin(&sched_lock); 630 if (rval == 0 && catch) { 631 PROC_LOCK(p); 632 if (sig != 0 || (sig = CURSIG(p))) { 633 if (SIGISMEMBER(p->p_sigacts->ps_sigintr, sig)) 634 rval = EINTR; 635 else 636 rval = ERESTART; 637 } 638 PROC_UNLOCK(p); 639 } 640#ifdef KTRACE 641 mtx_lock(&Giant); 642 if (KTRPOINT(p, KTR_CSW)) 643 ktrcsw(p->p_tracep, 0, 0); 644 mtx_unlock(&Giant); 645#endif 646 } else { 647 /* 648 * If as_priority is 0, mawait() has been called without an 649 * intervening asleep(). We are still effectively a NOP, 650 * but we call mi_switch() for safety. 651 */ 652 653 if (p->p_asleep.as_priority == 0) { 654 p->p_stats->p_ru.ru_nvcsw++; 655 mi_switch(); 656 } 657 mtx_unlock_spin(&sched_lock); 658 } 659 660 /* 661 * clear p_asleep.as_priority as an indication that mawait() has been 662 * called. If mawait() is called again without an intervening asleep(), 663 * mawait() is still effectively a NOP but the above mi_switch() code 664 * is triggered as a safety. 665 */ 666 if (rval == 0) 667 p->p_asleep.as_priority = 0; 668 669 PICKUP_GIANT(); 670 if (mtx != NULL) { 671 mtx_lock(mtx); 672 WITNESS_RESTORE(&mtx->mtx_object, mtx); 673 } 674 return (rval); 675} 676 677/* 678 * Implement timeout for msleep or asleep()/mawait() 679 * 680 * If process hasn't been awakened (wchan non-zero), 681 * set timeout flag and undo the sleep. If proc 682 * is stopped, just unsleep so it will remain stopped. 683 * MP-safe, called without the Giant mutex. 684 */ 685static void 686endtsleep(arg) 687 void *arg; 688{ 689 register struct proc *p; 690 int s; 691 692 p = (struct proc *)arg; 693 CTR3(KTR_PROC, "endtsleep: proc %p (pid %d, %s)", p, p->p_pid, 694 p->p_comm); 695 s = splhigh(); 696 mtx_lock_spin(&sched_lock); 697 if (p->p_wchan) { 698 if (p->p_stat == SSLEEP) 699 setrunnable(p); 700 else 701 unsleep(p); 702 p->p_sflag |= PS_TIMEOUT; 703 } 704 mtx_unlock_spin(&sched_lock); 705 splx(s); 706} 707 708/* 709 * Remove a process from its wait queue 710 */ 711void 712unsleep(p) 713 register struct proc *p; 714{ 715 int s; 716 717 s = splhigh(); 718 mtx_lock_spin(&sched_lock); 719 if (p->p_wchan) { 720 TAILQ_REMOVE(&slpque[LOOKUP(p->p_wchan)], p, p_slpq); 721 p->p_wchan = NULL; 722 } 723 mtx_unlock_spin(&sched_lock); 724 splx(s); 725} 726 727/* 728 * Make all processes sleeping on the specified identifier runnable. 729 */ 730void 731wakeup(ident) 732 register void *ident; 733{ 734 register struct slpquehead *qp; 735 register struct proc *p; 736 int s; 737 738 s = splhigh(); 739 mtx_lock_spin(&sched_lock); 740 qp = &slpque[LOOKUP(ident)]; 741restart: 742 TAILQ_FOREACH(p, qp, p_slpq) { 743 if (p->p_wchan == ident) { 744 TAILQ_REMOVE(qp, p, p_slpq); 745 p->p_wchan = NULL; 746 if (p->p_stat == SSLEEP) { 747 /* OPTIMIZED EXPANSION OF setrunnable(p); */ 748 CTR3(KTR_PROC, "wakeup: proc %p (pid %d, %s)", 749 p, p->p_pid, p->p_comm); 750 if (p->p_slptime > 1) 751 updatepri(p); 752 p->p_slptime = 0; 753 p->p_stat = SRUN; 754 if (p->p_sflag & PS_INMEM) { 755 setrunqueue(p); 756 maybe_resched(p); 757 } else { 758 p->p_sflag |= PS_SWAPINREQ; 759 wakeup((caddr_t)&proc0); 760 } 761 /* END INLINE EXPANSION */ 762 goto restart; 763 } 764 } 765 } 766 mtx_unlock_spin(&sched_lock); 767 splx(s); 768} 769 770/* 771 * Make a process sleeping on the specified identifier runnable. 772 * May wake more than one process if a target process is currently 773 * swapped out. 774 */ 775void 776wakeup_one(ident) 777 register void *ident; 778{ 779 register struct slpquehead *qp; 780 register struct proc *p; 781 int s; 782 783 s = splhigh(); 784 mtx_lock_spin(&sched_lock); 785 qp = &slpque[LOOKUP(ident)]; 786 787 TAILQ_FOREACH(p, qp, p_slpq) { 788 if (p->p_wchan == ident) { 789 TAILQ_REMOVE(qp, p, p_slpq); 790 p->p_wchan = NULL; 791 if (p->p_stat == SSLEEP) { 792 /* OPTIMIZED EXPANSION OF setrunnable(p); */ 793 CTR3(KTR_PROC, "wakeup1: proc %p (pid %d, %s)", 794 p, p->p_pid, p->p_comm); 795 if (p->p_slptime > 1) 796 updatepri(p); 797 p->p_slptime = 0; 798 p->p_stat = SRUN; 799 if (p->p_sflag & PS_INMEM) { 800 setrunqueue(p); 801 maybe_resched(p); 802 break; 803 } else { 804 p->p_sflag |= PS_SWAPINREQ; 805 wakeup((caddr_t)&proc0); 806 } 807 /* END INLINE EXPANSION */ 808 } 809 } 810 } 811 mtx_unlock_spin(&sched_lock); 812 splx(s); 813} 814 815/* 816 * The machine independent parts of mi_switch(). 817 * Must be called at splstatclock() or higher. 818 */ 819void 820mi_switch() 821{ 822 struct timeval new_switchtime; 823 register struct proc *p = curproc; /* XXX */ 824#if 0 825 register struct rlimit *rlim; 826#endif 827 critical_t sched_crit; 828 u_int sched_nest; 829 830 mtx_assert(&sched_lock, MA_OWNED | MA_NOTRECURSED); 831 832 /* 833 * Compute the amount of time during which the current 834 * process was running, and add that to its total so far. 835 */ 836 microuptime(&new_switchtime); 837 if (timevalcmp(&new_switchtime, PCPU_PTR(switchtime), <)) { 838#if 0 839 /* XXX: This doesn't play well with sched_lock right now. */ 840 printf("microuptime() went backwards (%ld.%06ld -> %ld.%06ld)\n", 841 PCPU_GET(switchtime.tv_sec), PCPU_GET(switchtime.tv_usec), 842 new_switchtime.tv_sec, new_switchtime.tv_usec); 843#endif 844 new_switchtime = PCPU_GET(switchtime); 845 } else { 846 p->p_runtime += (new_switchtime.tv_usec - PCPU_GET(switchtime.tv_usec)) + 847 (new_switchtime.tv_sec - PCPU_GET(switchtime.tv_sec)) * 848 (int64_t)1000000; 849 } 850 851#if 0 852 /* 853 * Check if the process exceeds its cpu resource allocation. 854 * If over max, kill it. 855 * 856 * XXX drop sched_lock, pickup Giant 857 */ 858 if (p->p_stat != SZOMB && p->p_limit->p_cpulimit != RLIM_INFINITY && 859 p->p_runtime > p->p_limit->p_cpulimit) { 860 rlim = &p->p_rlimit[RLIMIT_CPU]; 861 if (p->p_runtime / (rlim_t)1000000 >= rlim->rlim_max) { 862 mtx_unlock_spin(&sched_lock); 863 PROC_LOCK(p); 864 killproc(p, "exceeded maximum CPU limit"); 865 mtx_lock_spin(&sched_lock); 866 PROC_UNLOCK_NOSWITCH(p); 867 } else { 868 mtx_unlock_spin(&sched_lock); 869 PROC_LOCK(p); 870 psignal(p, SIGXCPU); 871 mtx_lock_spin(&sched_lock); 872 PROC_UNLOCK_NOSWITCH(p); 873 if (rlim->rlim_cur < rlim->rlim_max) { 874 /* XXX: we should make a private copy */ 875 rlim->rlim_cur += 5; 876 } 877 } 878 } 879#endif 880 881 /* 882 * Pick a new current process and record its start time. 883 */ 884 cnt.v_swtch++; 885 PCPU_SET(switchtime, new_switchtime); 886 CTR3(KTR_PROC, "mi_switch: old proc %p (pid %d, %s)", p, p->p_pid, 887 p->p_comm); 888 sched_crit = sched_lock.mtx_savecrit; 889 sched_nest = sched_lock.mtx_recurse; 890 curproc->p_lastcpu = curproc->p_oncpu; 891 curproc->p_oncpu = NOCPU; 892 clear_resched(curproc); 893 cpu_switch(); 894 curproc->p_oncpu = PCPU_GET(cpuid); 895 sched_lock.mtx_savecrit = sched_crit; 896 sched_lock.mtx_recurse = sched_nest; 897 sched_lock.mtx_lock = (uintptr_t)curproc; 898 CTR3(KTR_PROC, "mi_switch: new proc %p (pid %d, %s)", p, p->p_pid, 899 p->p_comm); 900 if (PCPU_GET(switchtime.tv_sec) == 0) 901 microuptime(PCPU_PTR(switchtime)); 902 PCPU_SET(switchticks, ticks); 903} 904 905/* 906 * Change process state to be runnable, 907 * placing it on the run queue if it is in memory, 908 * and awakening the swapper if it isn't in memory. 909 */ 910void 911setrunnable(p) 912 register struct proc *p; 913{ 914 register int s; 915 916 s = splhigh(); 917 mtx_lock_spin(&sched_lock); 918 switch (p->p_stat) { 919 case 0: 920 case SRUN: 921 case SZOMB: 922 case SWAIT: 923 default: 924 panic("setrunnable"); 925 case SSTOP: 926 case SSLEEP: /* e.g. when sending signals */ 927 if (p->p_sflag & PS_CVWAITQ) 928 cv_waitq_remove(p); 929 else 930 unsleep(p); 931 break; 932 933 case SIDL: 934 break; 935 } 936 p->p_stat = SRUN; 937 if (p->p_sflag & PS_INMEM) 938 setrunqueue(p); 939 splx(s); 940 if (p->p_slptime > 1) 941 updatepri(p); 942 p->p_slptime = 0; 943 if ((p->p_sflag & PS_INMEM) == 0) { 944 p->p_sflag |= PS_SWAPINREQ; 945 wakeup((caddr_t)&proc0); 946 } 947 else 948 maybe_resched(p); 949 mtx_unlock_spin(&sched_lock); 950} 951 952/* 953 * Compute the priority of a process when running in user mode. 954 * Arrange to reschedule if the resulting priority is better 955 * than that of the current process. 956 */ 957void 958resetpriority(p) 959 register struct proc *p; 960{ 961 register unsigned int newpriority; 962 963 mtx_lock_spin(&sched_lock); 964 if (p->p_pri.pri_class == PRI_TIMESHARE) { 965 newpriority = PUSER + p->p_estcpu / INVERSE_ESTCPU_WEIGHT + 966 NICE_WEIGHT * (p->p_nice - PRIO_MIN); 967 newpriority = min(max(newpriority, PRI_MIN_TIMESHARE), 968 PRI_MAX_TIMESHARE); 969 p->p_pri.pri_user = newpriority; 970 } 971 maybe_resched(p); 972 mtx_unlock_spin(&sched_lock); 973} 974 975/* ARGSUSED */ 976static void 977sched_setup(dummy) 978 void *dummy; 979{ 980 981 callout_init(&schedcpu_callout, 1); 982 callout_init(&roundrobin_callout, 0); 983 984 /* Kick off timeout driven events by calling first time. */ 985 roundrobin(NULL); 986 schedcpu(NULL); 987} 988 989/* 990 * We adjust the priority of the current process. The priority of 991 * a process gets worse as it accumulates CPU time. The cpu usage 992 * estimator (p_estcpu) is increased here. resetpriority() will 993 * compute a different priority each time p_estcpu increases by 994 * INVERSE_ESTCPU_WEIGHT 995 * (until MAXPRI is reached). The cpu usage estimator ramps up 996 * quite quickly when the process is running (linearly), and decays 997 * away exponentially, at a rate which is proportionally slower when 998 * the system is busy. The basic principle is that the system will 999 * 90% forget that the process used a lot of CPU time in 5 * loadav 1000 * seconds. This causes the system to favor processes which haven't 1001 * run much recently, and to round-robin among other processes. 1002 */ 1003void 1004schedclock(p) 1005 struct proc *p; 1006{ 1007 1008 p->p_cpticks++; 1009 p->p_estcpu = ESTCPULIM(p->p_estcpu + 1); 1010 if ((p->p_estcpu % INVERSE_ESTCPU_WEIGHT) == 0) { 1011 resetpriority(p); 1012 if (p->p_pri.pri_level >= PUSER) 1013 p->p_pri.pri_level = p->p_pri.pri_user; 1014 } 1015} 1016 1017/* 1018 * General purpose yield system call 1019 */ 1020int 1021yield(struct proc *p, struct yield_args *uap) 1022{ 1023 int s; 1024 1025 p->p_retval[0] = 0; 1026 1027 s = splhigh(); 1028 mtx_lock_spin(&sched_lock); 1029 DROP_GIANT_NOSWITCH(); 1030 p->p_pri.pri_level = PRI_MAX_TIMESHARE; 1031 setrunqueue(p); 1032 p->p_stats->p_ru.ru_nvcsw++; 1033 mi_switch(); 1034 mtx_unlock_spin(&sched_lock); 1035 PICKUP_GIANT(); 1036 splx(s); 1037 1038 return (0); 1039} 1040