kern_synch.c revision 79128
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 79128 2001-07-03 07:39:06Z 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 CTR5(KTR_PROC, "msleep: proc %p (pid %d, %s) on %s (%p)", p, p->p_pid, 429 p->p_comm, wmesg, ident); 430 TAILQ_INSERT_TAIL(&slpque[LOOKUP(ident)], p, p_slpq); 431 if (timo) 432 callout_reset(&p->p_slpcallout, timo, endtsleep, p); 433 /* 434 * We put ourselves on the sleep queue and start our timeout 435 * before calling CURSIG, as we could stop there, and a wakeup 436 * or a SIGCONT (or both) could occur while we were stopped. 437 * A SIGCONT would cause us to be marked as SSLEEP 438 * without resuming us, thus we must be ready for sleep 439 * when CURSIG is called. If the wakeup happens while we're 440 * stopped, p->p_wchan will be 0 upon return from CURSIG. 441 */ 442 if (catch) { 443 CTR3(KTR_PROC, "msleep caught: proc %p (pid %d, %s)", p, 444 p->p_pid, p->p_comm); 445 p->p_sflag |= PS_SINTR; 446 mtx_unlock_spin(&sched_lock); 447 PROC_LOCK(p); 448 sig = CURSIG(p); 449 mtx_lock_spin(&sched_lock); 450 PROC_UNLOCK_NOSWITCH(p); 451 if (sig != 0) { 452 if (p->p_wchan) 453 unsleep(p); 454 } else if (p->p_wchan == NULL) 455 catch = 0; 456 } else 457 sig = 0; 458 if (p->p_wchan != NULL) { 459 p->p_stat = SSLEEP; 460 p->p_stats->p_ru.ru_nvcsw++; 461 mi_switch(); 462 } 463 CTR3(KTR_PROC, "msleep resume: proc %p (pid %d, %s)", p, p->p_pid, 464 p->p_comm); 465 KASSERT(p->p_stat == SRUN, ("running but not SRUN")); 466 p->p_sflag &= ~PS_SINTR; 467 if (p->p_sflag & PS_TIMEOUT) { 468 p->p_sflag &= ~PS_TIMEOUT; 469 if (sig == 0) 470 rval = EWOULDBLOCK; 471 } else if (timo) 472 callout_stop(&p->p_slpcallout); 473 mtx_unlock_spin(&sched_lock); 474 475 if (rval == 0 && catch) { 476 PROC_LOCK(p); 477 /* XXX: shouldn't we always be calling CURSIG() */ 478 if (sig != 0 || (sig = CURSIG(p))) { 479 if (SIGISMEMBER(p->p_sigacts->ps_sigintr, sig)) 480 rval = EINTR; 481 else 482 rval = ERESTART; 483 } 484 PROC_UNLOCK(p); 485 } 486 PICKUP_GIANT(); 487#ifdef KTRACE 488 mtx_lock(&Giant); 489 if (KTRPOINT(p, KTR_CSW)) 490 ktrcsw(p->p_tracep, 0, 0); 491 mtx_unlock(&Giant); 492#endif 493 if (mtx != NULL) { 494 mtx_lock(mtx); 495 WITNESS_RESTORE(&mtx->mtx_object, mtx); 496 } 497 return (rval); 498} 499 500/* 501 * asleep() - async sleep call. Place process on wait queue and return 502 * immediately without blocking. The process stays runnable until mawait() 503 * is called. If ident is NULL, remove process from wait queue if it is still 504 * on one. 505 * 506 * Only the most recent sleep condition is effective when making successive 507 * calls to asleep() or when calling msleep(). 508 * 509 * The timeout, if any, is not initiated until mawait() is called. The sleep 510 * priority, signal, and timeout is specified in the asleep() call but may be 511 * overriden in the mawait() call. 512 * 513 * <<<<<<<< EXPERIMENTAL, UNTESTED >>>>>>>>>> 514 */ 515 516int 517asleep(void *ident, int priority, const char *wmesg, int timo) 518{ 519 struct proc *p = curproc; 520 521 /* 522 * Remove preexisting wait condition (if any) and place process 523 * on appropriate slpque, but do not put process to sleep. 524 */ 525 526 mtx_lock_spin(&sched_lock); 527 528 if (p->p_wchan != NULL) 529 unsleep(p); 530 531 if (ident) { 532 p->p_wchan = ident; 533 p->p_wmesg = wmesg; 534 p->p_slptime = 0; 535 p->p_asleep.as_priority = priority; 536 p->p_asleep.as_timo = timo; 537 TAILQ_INSERT_TAIL(&slpque[LOOKUP(ident)], p, p_slpq); 538 } 539 540 mtx_unlock_spin(&sched_lock); 541 542 return(0); 543} 544 545/* 546 * mawait() - wait for async condition to occur. The process blocks until 547 * wakeup() is called on the most recent asleep() address. If wakeup is called 548 * prior to mawait(), mawait() winds up being a NOP. 549 * 550 * If mawait() is called more then once (without an intervening asleep() call), 551 * mawait() is still effectively a NOP but it calls mi_switch() to give other 552 * processes some cpu before returning. The process is left runnable. 553 * 554 * <<<<<<<< EXPERIMENTAL, UNTESTED >>>>>>>>>> 555 */ 556 557int 558mawait(struct mtx *mtx, int priority, int timo) 559{ 560 struct proc *p = curproc; 561 int rval = 0; 562 WITNESS_SAVE_DECL(mtx); 563 564 WITNESS_SLEEP(0, &mtx->mtx_object); 565 KASSERT(timo > 0 || mtx_owned(&Giant) || mtx != NULL, 566 ("sleeping without a mutex")); 567 mtx_lock_spin(&sched_lock); 568 DROP_GIANT_NOSWITCH(); 569 if (mtx != NULL) { 570 mtx_assert(mtx, MA_OWNED | MA_NOTRECURSED); 571 WITNESS_SAVE(&mtx->mtx_object, mtx); 572 mtx_unlock_flags(mtx, MTX_NOSWITCH); 573 if (priority & PDROP) 574 mtx = NULL; 575 } 576 577 if (p->p_wchan != NULL) { 578 int sig; 579 int catch; 580 581#ifdef KTRACE 582 if (p && KTRPOINT(p, KTR_CSW)) 583 ktrcsw(p->p_tracep, 1, 0); 584#endif 585 /* 586 * The call to mawait() can override defaults specified in 587 * the original asleep(). 588 */ 589 if (priority < 0) 590 priority = p->p_asleep.as_priority; 591 if (timo < 0) 592 timo = p->p_asleep.as_timo; 593 594 /* 595 * Install timeout 596 */ 597 598 if (timo) 599 callout_reset(&p->p_slpcallout, timo, endtsleep, p); 600 601 sig = 0; 602 catch = priority & PCATCH; 603 604 if (catch) { 605 p->p_sflag |= PS_SINTR; 606 mtx_unlock_spin(&sched_lock); 607 PROC_LOCK(p); 608 sig = CURSIG(p); 609 mtx_lock_spin(&sched_lock); 610 PROC_UNLOCK_NOSWITCH(p); 611 if (sig != 0) { 612 if (p->p_wchan) 613 unsleep(p); 614 } else if (p->p_wchan == NULL) 615 catch = 0; 616 } 617 if (p->p_wchan != NULL) { 618 p->p_stat = SSLEEP; 619 p->p_stats->p_ru.ru_nvcsw++; 620 mi_switch(); 621 } 622 KASSERT(p->p_stat == SRUN, ("running but not SRUN")); 623 p->p_sflag &= ~PS_SINTR; 624 if (p->p_sflag & PS_TIMEOUT) { 625 p->p_sflag &= ~PS_TIMEOUT; 626 if (sig == 0) 627 rval = EWOULDBLOCK; 628 } else if (timo) 629 callout_stop(&p->p_slpcallout); 630 mtx_unlock_spin(&sched_lock); 631 if (rval == 0 && catch) { 632 PROC_LOCK(p); 633 if (sig != 0 || (sig = CURSIG(p))) { 634 if (SIGISMEMBER(p->p_sigacts->ps_sigintr, sig)) 635 rval = EINTR; 636 else 637 rval = ERESTART; 638 } 639 PROC_UNLOCK(p); 640 } 641#ifdef KTRACE 642 mtx_lock(&Giant); 643 if (KTRPOINT(p, KTR_CSW)) 644 ktrcsw(p->p_tracep, 0, 0); 645 mtx_unlock(&Giant); 646#endif 647 } else { 648 /* 649 * If as_priority is 0, mawait() has been called without an 650 * intervening asleep(). We are still effectively a NOP, 651 * but we call mi_switch() for safety. 652 */ 653 654 if (p->p_asleep.as_priority == 0) { 655 p->p_stats->p_ru.ru_nvcsw++; 656 mi_switch(); 657 } 658 mtx_unlock_spin(&sched_lock); 659 } 660 661 /* 662 * clear p_asleep.as_priority as an indication that mawait() has been 663 * called. If mawait() is called again without an intervening asleep(), 664 * mawait() is still effectively a NOP but the above mi_switch() code 665 * is triggered as a safety. 666 */ 667 if (rval == 0) 668 p->p_asleep.as_priority = 0; 669 670 PICKUP_GIANT(); 671 if (mtx != NULL) { 672 mtx_lock(mtx); 673 WITNESS_RESTORE(&mtx->mtx_object, mtx); 674 } 675 return (rval); 676} 677 678/* 679 * Implement timeout for msleep or asleep()/mawait() 680 * 681 * If process hasn't been awakened (wchan non-zero), 682 * set timeout flag and undo the sleep. If proc 683 * is stopped, just unsleep so it will remain stopped. 684 * MP-safe, called without the Giant mutex. 685 */ 686static void 687endtsleep(arg) 688 void *arg; 689{ 690 register struct proc *p; 691 int s; 692 693 p = (struct proc *)arg; 694 CTR3(KTR_PROC, "endtsleep: proc %p (pid %d, %s)", p, p->p_pid, 695 p->p_comm); 696 s = splhigh(); 697 mtx_lock_spin(&sched_lock); 698 if (p->p_wchan) { 699 if (p->p_stat == SSLEEP) 700 setrunnable(p); 701 else 702 unsleep(p); 703 p->p_sflag |= PS_TIMEOUT; 704 } 705 mtx_unlock_spin(&sched_lock); 706 splx(s); 707} 708 709/* 710 * Remove a process from its wait queue 711 */ 712void 713unsleep(p) 714 register struct proc *p; 715{ 716 int s; 717 718 s = splhigh(); 719 mtx_lock_spin(&sched_lock); 720 if (p->p_wchan) { 721 TAILQ_REMOVE(&slpque[LOOKUP(p->p_wchan)], p, p_slpq); 722 p->p_wchan = NULL; 723 } 724 mtx_unlock_spin(&sched_lock); 725 splx(s); 726} 727 728/* 729 * Make all processes sleeping on the specified identifier runnable. 730 */ 731void 732wakeup(ident) 733 register void *ident; 734{ 735 register struct slpquehead *qp; 736 register struct proc *p; 737 int s; 738 739 s = splhigh(); 740 mtx_lock_spin(&sched_lock); 741 qp = &slpque[LOOKUP(ident)]; 742restart: 743 TAILQ_FOREACH(p, qp, p_slpq) { 744 if (p->p_wchan == ident) { 745 TAILQ_REMOVE(qp, p, p_slpq); 746 p->p_wchan = NULL; 747 if (p->p_stat == SSLEEP) { 748 /* OPTIMIZED EXPANSION OF setrunnable(p); */ 749 CTR3(KTR_PROC, "wakeup: proc %p (pid %d, %s)", 750 p, p->p_pid, p->p_comm); 751 if (p->p_slptime > 1) 752 updatepri(p); 753 p->p_slptime = 0; 754 p->p_stat = SRUN; 755 if (p->p_sflag & PS_INMEM) { 756 setrunqueue(p); 757 maybe_resched(p); 758 } else { 759 p->p_sflag |= PS_SWAPINREQ; 760 wakeup((caddr_t)&proc0); 761 } 762 /* END INLINE EXPANSION */ 763 goto restart; 764 } 765 } 766 } 767 mtx_unlock_spin(&sched_lock); 768 splx(s); 769} 770 771/* 772 * Make a process sleeping on the specified identifier runnable. 773 * May wake more than one process if a target process is currently 774 * swapped out. 775 */ 776void 777wakeup_one(ident) 778 register void *ident; 779{ 780 register struct slpquehead *qp; 781 register struct proc *p; 782 int s; 783 784 s = splhigh(); 785 mtx_lock_spin(&sched_lock); 786 qp = &slpque[LOOKUP(ident)]; 787 788 TAILQ_FOREACH(p, qp, p_slpq) { 789 if (p->p_wchan == ident) { 790 TAILQ_REMOVE(qp, p, p_slpq); 791 p->p_wchan = NULL; 792 if (p->p_stat == SSLEEP) { 793 /* OPTIMIZED EXPANSION OF setrunnable(p); */ 794 CTR3(KTR_PROC, "wakeup1: proc %p (pid %d, %s)", 795 p, p->p_pid, p->p_comm); 796 if (p->p_slptime > 1) 797 updatepri(p); 798 p->p_slptime = 0; 799 p->p_stat = SRUN; 800 if (p->p_sflag & PS_INMEM) { 801 setrunqueue(p); 802 maybe_resched(p); 803 break; 804 } else { 805 p->p_sflag |= PS_SWAPINREQ; 806 wakeup((caddr_t)&proc0); 807 } 808 /* END INLINE EXPANSION */ 809 } 810 } 811 } 812 mtx_unlock_spin(&sched_lock); 813 splx(s); 814} 815 816/* 817 * The machine independent parts of mi_switch(). 818 * Must be called at splstatclock() or higher. 819 */ 820void 821mi_switch() 822{ 823 struct timeval new_switchtime; 824 register struct proc *p = curproc; /* XXX */ 825#if 0 826 register struct rlimit *rlim; 827#endif 828 critical_t sched_crit; 829 u_int sched_nest; 830 831 mtx_assert(&sched_lock, MA_OWNED | MA_NOTRECURSED); 832 833 /* 834 * Compute the amount of time during which the current 835 * process was running, and add that to its total so far. 836 */ 837 microuptime(&new_switchtime); 838 if (timevalcmp(&new_switchtime, PCPU_PTR(switchtime), <)) { 839#if 0 840 /* XXX: This doesn't play well with sched_lock right now. */ 841 printf("microuptime() went backwards (%ld.%06ld -> %ld.%06ld)\n", 842 PCPU_GET(switchtime.tv_sec), PCPU_GET(switchtime.tv_usec), 843 new_switchtime.tv_sec, new_switchtime.tv_usec); 844#endif 845 new_switchtime = PCPU_GET(switchtime); 846 } else { 847 p->p_runtime += (new_switchtime.tv_usec - PCPU_GET(switchtime.tv_usec)) + 848 (new_switchtime.tv_sec - PCPU_GET(switchtime.tv_sec)) * 849 (int64_t)1000000; 850 } 851 852#if 0 853 /* 854 * Check if the process exceeds its cpu resource allocation. 855 * If over max, kill it. 856 * 857 * XXX drop sched_lock, pickup Giant 858 */ 859 if (p->p_stat != SZOMB && p->p_limit->p_cpulimit != RLIM_INFINITY && 860 p->p_runtime > p->p_limit->p_cpulimit) { 861 rlim = &p->p_rlimit[RLIMIT_CPU]; 862 if (p->p_runtime / (rlim_t)1000000 >= rlim->rlim_max) { 863 mtx_unlock_spin(&sched_lock); 864 PROC_LOCK(p); 865 killproc(p, "exceeded maximum CPU limit"); 866 mtx_lock_spin(&sched_lock); 867 PROC_UNLOCK_NOSWITCH(p); 868 } else { 869 mtx_unlock_spin(&sched_lock); 870 PROC_LOCK(p); 871 psignal(p, SIGXCPU); 872 mtx_lock_spin(&sched_lock); 873 PROC_UNLOCK_NOSWITCH(p); 874 if (rlim->rlim_cur < rlim->rlim_max) { 875 /* XXX: we should make a private copy */ 876 rlim->rlim_cur += 5; 877 } 878 } 879 } 880#endif 881 882 /* 883 * Pick a new current process and record its start time. 884 */ 885 cnt.v_swtch++; 886 PCPU_SET(switchtime, new_switchtime); 887 CTR3(KTR_PROC, "mi_switch: old proc %p (pid %d, %s)", p, p->p_pid, 888 p->p_comm); 889 sched_crit = sched_lock.mtx_savecrit; 890 sched_nest = sched_lock.mtx_recurse; 891 curproc->p_lastcpu = curproc->p_oncpu; 892 curproc->p_oncpu = NOCPU; 893 clear_resched(curproc); 894 cpu_switch(); 895 curproc->p_oncpu = PCPU_GET(cpuid); 896 sched_lock.mtx_savecrit = sched_crit; 897 sched_lock.mtx_recurse = sched_nest; 898 sched_lock.mtx_lock = (uintptr_t)curproc; 899 CTR3(KTR_PROC, "mi_switch: new proc %p (pid %d, %s)", p, p->p_pid, 900 p->p_comm); 901 if (PCPU_GET(switchtime.tv_sec) == 0) 902 microuptime(PCPU_PTR(switchtime)); 903 PCPU_SET(switchticks, ticks); 904} 905 906/* 907 * Change process state to be runnable, 908 * placing it on the run queue if it is in memory, 909 * and awakening the swapper if it isn't in memory. 910 */ 911void 912setrunnable(p) 913 register struct proc *p; 914{ 915 register int s; 916 917 s = splhigh(); 918 mtx_lock_spin(&sched_lock); 919 switch (p->p_stat) { 920 case 0: 921 case SRUN: 922 case SZOMB: 923 case SWAIT: 924 default: 925 panic("setrunnable"); 926 case SSTOP: 927 case SSLEEP: /* e.g. when sending signals */ 928 if (p->p_sflag & PS_CVWAITQ) 929 cv_waitq_remove(p); 930 else 931 unsleep(p); 932 break; 933 934 case SIDL: 935 break; 936 } 937 p->p_stat = SRUN; 938 if (p->p_sflag & PS_INMEM) 939 setrunqueue(p); 940 splx(s); 941 if (p->p_slptime > 1) 942 updatepri(p); 943 p->p_slptime = 0; 944 if ((p->p_sflag & PS_INMEM) == 0) { 945 p->p_sflag |= PS_SWAPINREQ; 946 wakeup((caddr_t)&proc0); 947 } 948 else 949 maybe_resched(p); 950 mtx_unlock_spin(&sched_lock); 951} 952 953/* 954 * Compute the priority of a process when running in user mode. 955 * Arrange to reschedule if the resulting priority is better 956 * than that of the current process. 957 */ 958void 959resetpriority(p) 960 register struct proc *p; 961{ 962 register unsigned int newpriority; 963 964 mtx_lock_spin(&sched_lock); 965 if (p->p_pri.pri_class == PRI_TIMESHARE) { 966 newpriority = PUSER + p->p_estcpu / INVERSE_ESTCPU_WEIGHT + 967 NICE_WEIGHT * (p->p_nice - PRIO_MIN); 968 newpriority = min(max(newpriority, PRI_MIN_TIMESHARE), 969 PRI_MAX_TIMESHARE); 970 p->p_pri.pri_user = newpriority; 971 } 972 maybe_resched(p); 973 mtx_unlock_spin(&sched_lock); 974} 975 976/* ARGSUSED */ 977static void 978sched_setup(dummy) 979 void *dummy; 980{ 981 982 callout_init(&schedcpu_callout, 1); 983 callout_init(&roundrobin_callout, 0); 984 985 /* Kick off timeout driven events by calling first time. */ 986 roundrobin(NULL); 987 schedcpu(NULL); 988} 989 990/* 991 * We adjust the priority of the current process. The priority of 992 * a process gets worse as it accumulates CPU time. The cpu usage 993 * estimator (p_estcpu) is increased here. resetpriority() will 994 * compute a different priority each time p_estcpu increases by 995 * INVERSE_ESTCPU_WEIGHT 996 * (until MAXPRI is reached). The cpu usage estimator ramps up 997 * quite quickly when the process is running (linearly), and decays 998 * away exponentially, at a rate which is proportionally slower when 999 * the system is busy. The basic principle is that the system will 1000 * 90% forget that the process used a lot of CPU time in 5 * loadav 1001 * seconds. This causes the system to favor processes which haven't 1002 * run much recently, and to round-robin among other processes. 1003 */ 1004void 1005schedclock(p) 1006 struct proc *p; 1007{ 1008 1009 p->p_cpticks++; 1010 p->p_estcpu = ESTCPULIM(p->p_estcpu + 1); 1011 if ((p->p_estcpu % INVERSE_ESTCPU_WEIGHT) == 0) { 1012 resetpriority(p); 1013 if (p->p_pri.pri_level >= PUSER) 1014 p->p_pri.pri_level = p->p_pri.pri_user; 1015 } 1016} 1017 1018/* 1019 * General purpose yield system call 1020 */ 1021int 1022yield(struct proc *p, struct yield_args *uap) 1023{ 1024 int s; 1025 1026 p->p_retval[0] = 0; 1027 1028 s = splhigh(); 1029 mtx_lock_spin(&sched_lock); 1030 DROP_GIANT_NOSWITCH(); 1031 p->p_pri.pri_level = PRI_MAX_TIMESHARE; 1032 setrunqueue(p); 1033 p->p_stats->p_ru.ru_nvcsw++; 1034 mi_switch(); 1035 mtx_unlock_spin(&sched_lock); 1036 PICKUP_GIANT(); 1037 splx(s); 1038 1039 return (0); 1040} 1041