kern_synch.c revision 79131
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 79131 2001-07-03 08:00:57Z 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->p_oncpu == NOCPU && /* idle */ 287 p->p_stat == SRUN && 288 (p->p_sflag & PS_INMEM) && 289 (p->p_pri.pri_level / RQ_PPQ) != 290 (p->p_pri.pri_user / RQ_PPQ)) { 291 remrunqueue(p); 292 p->p_pri.pri_level = p->p_pri.pri_user; 293 setrunqueue(p); 294 } else 295 p->p_pri.pri_level = p->p_pri.pri_user; 296 } 297 mtx_unlock_spin(&sched_lock); 298 } 299 sx_sunlock(&allproc_lock); 300 vmmeter(); 301 wakeup((caddr_t)&lbolt); 302 callout_reset(&schedcpu_callout, hz, schedcpu, NULL); 303} 304 305/* 306 * Recalculate the priority of a process after it has slept for a while. 307 * For all load averages >= 1 and max p_estcpu of 255, sleeping for at 308 * least six times the loadfactor will decay p_estcpu to zero. 309 */ 310void 311updatepri(p) 312 register struct proc *p; 313{ 314 register unsigned int newcpu = p->p_estcpu; 315 register fixpt_t loadfac = loadfactor(averunnable.ldavg[0]); 316 317 if (p->p_slptime > 5 * loadfac) 318 p->p_estcpu = 0; 319 else { 320 p->p_slptime--; /* the first time was done in schedcpu */ 321 while (newcpu && --p->p_slptime) 322 newcpu = decay_cpu(loadfac, newcpu); 323 p->p_estcpu = newcpu; 324 } 325 resetpriority(p); 326} 327 328/* 329 * We're only looking at 7 bits of the address; everything is 330 * aligned to 4, lots of things are aligned to greater powers 331 * of 2. Shift right by 8, i.e. drop the bottom 256 worth. 332 */ 333#define TABLESIZE 128 334static TAILQ_HEAD(slpquehead, proc) slpque[TABLESIZE]; 335#define LOOKUP(x) (((intptr_t)(x) >> 8) & (TABLESIZE - 1)) 336 337void 338sleepinit(void) 339{ 340 int i; 341 342 sched_quantum = hz/10; 343 hogticks = 2 * sched_quantum; 344 for (i = 0; i < TABLESIZE; i++) 345 TAILQ_INIT(&slpque[i]); 346} 347 348/* 349 * General sleep call. Suspends the current process until a wakeup is 350 * performed on the specified identifier. The process will then be made 351 * runnable with the specified priority. Sleeps at most timo/hz seconds 352 * (0 means no timeout). If pri includes PCATCH flag, signals are checked 353 * before and after sleeping, else signals are not checked. Returns 0 if 354 * awakened, EWOULDBLOCK if the timeout expires. If PCATCH is set and a 355 * signal needs to be delivered, ERESTART is returned if the current system 356 * call should be restarted if possible, and EINTR is returned if the system 357 * call should be interrupted by the signal (return EINTR). 358 * 359 * The mutex argument is exited before the caller is suspended, and 360 * entered before msleep returns. If priority includes the PDROP 361 * flag the mutex is not entered before returning. 362 */ 363int 364msleep(ident, mtx, priority, wmesg, timo) 365 void *ident; 366 struct mtx *mtx; 367 int priority, timo; 368 const char *wmesg; 369{ 370 struct proc *p = curproc; 371 int sig, catch = priority & PCATCH; 372 int rval = 0; 373 WITNESS_SAVE_DECL(mtx); 374 375#ifdef KTRACE 376 if (p && KTRPOINT(p, KTR_CSW)) 377 ktrcsw(p->p_tracep, 1, 0); 378#endif 379 WITNESS_SLEEP(0, &mtx->mtx_object); 380 KASSERT(timo != 0 || mtx_owned(&Giant) || mtx != NULL, 381 ("sleeping without a mutex")); 382 mtx_lock_spin(&sched_lock); 383 if (cold || panicstr) { 384 /* 385 * After a panic, or during autoconfiguration, 386 * just give interrupts a chance, then just return; 387 * don't run any other procs or panic below, 388 * in case this is the idle process and already asleep. 389 */ 390 if (mtx != NULL && priority & PDROP) 391 mtx_unlock_flags(mtx, MTX_NOSWITCH); 392 mtx_unlock_spin(&sched_lock); 393 return (0); 394 } 395 396 DROP_GIANT_NOSWITCH(); 397 398 if (mtx != NULL) { 399 mtx_assert(mtx, MA_OWNED | MA_NOTRECURSED); 400 WITNESS_SAVE(&mtx->mtx_object, mtx); 401 mtx_unlock_flags(mtx, MTX_NOSWITCH); 402 if (priority & PDROP) 403 mtx = NULL; 404 } 405 406 KASSERT(p != NULL, ("msleep1")); 407 KASSERT(ident != NULL && p->p_stat == SRUN, ("msleep")); 408 /* 409 * Process may be sitting on a slpque if asleep() was called, remove 410 * it before re-adding. 411 */ 412 if (p->p_wchan != NULL) 413 unsleep(p); 414 415 p->p_wchan = ident; 416 p->p_wmesg = wmesg; 417 p->p_slptime = 0; 418 p->p_pri.pri_level = priority & PRIMASK; 419 CTR5(KTR_PROC, "msleep: proc %p (pid %d, %s) on %s (%p)", p, p->p_pid, 420 p->p_comm, wmesg, ident); 421 TAILQ_INSERT_TAIL(&slpque[LOOKUP(ident)], p, p_slpq); 422 if (timo) 423 callout_reset(&p->p_slpcallout, timo, endtsleep, p); 424 /* 425 * We put ourselves on the sleep queue and start our timeout 426 * before calling CURSIG, as we could stop there, and a wakeup 427 * or a SIGCONT (or both) could occur while we were stopped. 428 * A SIGCONT would cause us to be marked as SSLEEP 429 * without resuming us, thus we must be ready for sleep 430 * when CURSIG is called. If the wakeup happens while we're 431 * stopped, p->p_wchan will be 0 upon return from CURSIG. 432 */ 433 if (catch) { 434 CTR3(KTR_PROC, "msleep caught: proc %p (pid %d, %s)", p, 435 p->p_pid, p->p_comm); 436 p->p_sflag |= PS_SINTR; 437 mtx_unlock_spin(&sched_lock); 438 PROC_LOCK(p); 439 sig = CURSIG(p); 440 mtx_lock_spin(&sched_lock); 441 PROC_UNLOCK_NOSWITCH(p); 442 if (sig != 0) { 443 if (p->p_wchan) 444 unsleep(p); 445 } else if (p->p_wchan == NULL) 446 catch = 0; 447 } else 448 sig = 0; 449 if (p->p_wchan != NULL) { 450 p->p_stat = SSLEEP; 451 p->p_stats->p_ru.ru_nvcsw++; 452 mi_switch(); 453 } 454 CTR3(KTR_PROC, "msleep resume: proc %p (pid %d, %s)", p, p->p_pid, 455 p->p_comm); 456 KASSERT(p->p_stat == SRUN, ("running but not SRUN")); 457 p->p_sflag &= ~PS_SINTR; 458 if (p->p_sflag & PS_TIMEOUT) { 459 p->p_sflag &= ~PS_TIMEOUT; 460 if (sig == 0) 461 rval = EWOULDBLOCK; 462 } else if (timo) 463 callout_stop(&p->p_slpcallout); 464 mtx_unlock_spin(&sched_lock); 465 466 if (rval == 0 && catch) { 467 PROC_LOCK(p); 468 /* XXX: shouldn't we always be calling CURSIG() */ 469 if (sig != 0 || (sig = CURSIG(p))) { 470 if (SIGISMEMBER(p->p_sigacts->ps_sigintr, sig)) 471 rval = EINTR; 472 else 473 rval = ERESTART; 474 } 475 PROC_UNLOCK(p); 476 } 477 PICKUP_GIANT(); 478#ifdef KTRACE 479 mtx_lock(&Giant); 480 if (KTRPOINT(p, KTR_CSW)) 481 ktrcsw(p->p_tracep, 0, 0); 482 mtx_unlock(&Giant); 483#endif 484 if (mtx != NULL) { 485 mtx_lock(mtx); 486 WITNESS_RESTORE(&mtx->mtx_object, mtx); 487 } 488 return (rval); 489} 490 491/* 492 * asleep() - async sleep call. Place process on wait queue and return 493 * immediately without blocking. The process stays runnable until mawait() 494 * is called. If ident is NULL, remove process from wait queue if it is still 495 * on one. 496 * 497 * Only the most recent sleep condition is effective when making successive 498 * calls to asleep() or when calling msleep(). 499 * 500 * The timeout, if any, is not initiated until mawait() is called. The sleep 501 * priority, signal, and timeout is specified in the asleep() call but may be 502 * overriden in the mawait() call. 503 * 504 * <<<<<<<< EXPERIMENTAL, UNTESTED >>>>>>>>>> 505 */ 506 507int 508asleep(void *ident, int priority, const char *wmesg, int timo) 509{ 510 struct proc *p = curproc; 511 512 /* 513 * Remove preexisting wait condition (if any) and place process 514 * on appropriate slpque, but do not put process to sleep. 515 */ 516 517 mtx_lock_spin(&sched_lock); 518 519 if (p->p_wchan != NULL) 520 unsleep(p); 521 522 if (ident) { 523 p->p_wchan = ident; 524 p->p_wmesg = wmesg; 525 p->p_slptime = 0; 526 p->p_asleep.as_priority = priority; 527 p->p_asleep.as_timo = timo; 528 TAILQ_INSERT_TAIL(&slpque[LOOKUP(ident)], p, p_slpq); 529 } 530 531 mtx_unlock_spin(&sched_lock); 532 533 return(0); 534} 535 536/* 537 * mawait() - wait for async condition to occur. The process blocks until 538 * wakeup() is called on the most recent asleep() address. If wakeup is called 539 * prior to mawait(), mawait() winds up being a NOP. 540 * 541 * If mawait() is called more then once (without an intervening asleep() call), 542 * mawait() is still effectively a NOP but it calls mi_switch() to give other 543 * processes some cpu before returning. The process is left runnable. 544 * 545 * <<<<<<<< EXPERIMENTAL, UNTESTED >>>>>>>>>> 546 */ 547 548int 549mawait(struct mtx *mtx, int priority, int timo) 550{ 551 struct proc *p = curproc; 552 int rval = 0; 553 WITNESS_SAVE_DECL(mtx); 554 555 WITNESS_SLEEP(0, &mtx->mtx_object); 556 KASSERT(timo > 0 || mtx_owned(&Giant) || mtx != NULL, 557 ("sleeping without a mutex")); 558 mtx_lock_spin(&sched_lock); 559 DROP_GIANT_NOSWITCH(); 560 if (mtx != NULL) { 561 mtx_assert(mtx, MA_OWNED | MA_NOTRECURSED); 562 WITNESS_SAVE(&mtx->mtx_object, mtx); 563 mtx_unlock_flags(mtx, MTX_NOSWITCH); 564 if (priority & PDROP) 565 mtx = NULL; 566 } 567 568 if (p->p_wchan != NULL) { 569 int sig; 570 int catch; 571 572#ifdef KTRACE 573 if (p && KTRPOINT(p, KTR_CSW)) 574 ktrcsw(p->p_tracep, 1, 0); 575#endif 576 /* 577 * The call to mawait() can override defaults specified in 578 * the original asleep(). 579 */ 580 if (priority < 0) 581 priority = p->p_asleep.as_priority; 582 if (timo < 0) 583 timo = p->p_asleep.as_timo; 584 585 /* 586 * Install timeout 587 */ 588 589 if (timo) 590 callout_reset(&p->p_slpcallout, timo, endtsleep, p); 591 592 sig = 0; 593 catch = priority & PCATCH; 594 595 if (catch) { 596 p->p_sflag |= PS_SINTR; 597 mtx_unlock_spin(&sched_lock); 598 PROC_LOCK(p); 599 sig = CURSIG(p); 600 mtx_lock_spin(&sched_lock); 601 PROC_UNLOCK_NOSWITCH(p); 602 if (sig != 0) { 603 if (p->p_wchan) 604 unsleep(p); 605 } else if (p->p_wchan == NULL) 606 catch = 0; 607 } 608 if (p->p_wchan != NULL) { 609 p->p_stat = SSLEEP; 610 p->p_stats->p_ru.ru_nvcsw++; 611 mi_switch(); 612 } 613 KASSERT(p->p_stat == SRUN, ("running but not SRUN")); 614 p->p_sflag &= ~PS_SINTR; 615 if (p->p_sflag & PS_TIMEOUT) { 616 p->p_sflag &= ~PS_TIMEOUT; 617 if (sig == 0) 618 rval = EWOULDBLOCK; 619 } else if (timo) 620 callout_stop(&p->p_slpcallout); 621 mtx_unlock_spin(&sched_lock); 622 if (rval == 0 && catch) { 623 PROC_LOCK(p); 624 if (sig != 0 || (sig = CURSIG(p))) { 625 if (SIGISMEMBER(p->p_sigacts->ps_sigintr, sig)) 626 rval = EINTR; 627 else 628 rval = ERESTART; 629 } 630 PROC_UNLOCK(p); 631 } 632#ifdef KTRACE 633 mtx_lock(&Giant); 634 if (KTRPOINT(p, KTR_CSW)) 635 ktrcsw(p->p_tracep, 0, 0); 636 mtx_unlock(&Giant); 637#endif 638 } else { 639 /* 640 * If as_priority is 0, mawait() has been called without an 641 * intervening asleep(). We are still effectively a NOP, 642 * but we call mi_switch() for safety. 643 */ 644 645 if (p->p_asleep.as_priority == 0) { 646 p->p_stats->p_ru.ru_nvcsw++; 647 mi_switch(); 648 } 649 mtx_unlock_spin(&sched_lock); 650 } 651 652 /* 653 * clear p_asleep.as_priority as an indication that mawait() has been 654 * called. If mawait() is called again without an intervening asleep(), 655 * mawait() is still effectively a NOP but the above mi_switch() code 656 * is triggered as a safety. 657 */ 658 if (rval == 0) 659 p->p_asleep.as_priority = 0; 660 661 PICKUP_GIANT(); 662 if (mtx != NULL) { 663 mtx_lock(mtx); 664 WITNESS_RESTORE(&mtx->mtx_object, mtx); 665 } 666 return (rval); 667} 668 669/* 670 * Implement timeout for msleep or asleep()/mawait() 671 * 672 * If process hasn't been awakened (wchan non-zero), 673 * set timeout flag and undo the sleep. If proc 674 * is stopped, just unsleep so it will remain stopped. 675 * MP-safe, called without the Giant mutex. 676 */ 677static void 678endtsleep(arg) 679 void *arg; 680{ 681 register struct proc *p; 682 683 p = (struct proc *)arg; 684 CTR3(KTR_PROC, "endtsleep: proc %p (pid %d, %s)", p, p->p_pid, 685 p->p_comm); 686 mtx_lock_spin(&sched_lock); 687 if (p->p_wchan) { 688 if (p->p_stat == SSLEEP) 689 setrunnable(p); 690 else 691 unsleep(p); 692 p->p_sflag |= PS_TIMEOUT; 693 } 694 mtx_unlock_spin(&sched_lock); 695} 696 697/* 698 * Remove a process from its wait queue 699 */ 700void 701unsleep(p) 702 register struct proc *p; 703{ 704 705 mtx_lock_spin(&sched_lock); 706 if (p->p_wchan) { 707 TAILQ_REMOVE(&slpque[LOOKUP(p->p_wchan)], p, p_slpq); 708 p->p_wchan = NULL; 709 } 710 mtx_unlock_spin(&sched_lock); 711} 712 713/* 714 * Make all processes sleeping on the specified identifier runnable. 715 */ 716void 717wakeup(ident) 718 register void *ident; 719{ 720 register struct slpquehead *qp; 721 register struct proc *p; 722 723 mtx_lock_spin(&sched_lock); 724 qp = &slpque[LOOKUP(ident)]; 725restart: 726 TAILQ_FOREACH(p, qp, p_slpq) { 727 if (p->p_wchan == ident) { 728 TAILQ_REMOVE(qp, p, p_slpq); 729 p->p_wchan = NULL; 730 if (p->p_stat == SSLEEP) { 731 /* OPTIMIZED EXPANSION OF setrunnable(p); */ 732 CTR3(KTR_PROC, "wakeup: proc %p (pid %d, %s)", 733 p, p->p_pid, p->p_comm); 734 if (p->p_slptime > 1) 735 updatepri(p); 736 p->p_slptime = 0; 737 p->p_stat = SRUN; 738 if (p->p_sflag & PS_INMEM) { 739 setrunqueue(p); 740 maybe_resched(p); 741 } else { 742 p->p_sflag |= PS_SWAPINREQ; 743 wakeup((caddr_t)&proc0); 744 } 745 /* END INLINE EXPANSION */ 746 goto restart; 747 } 748 } 749 } 750 mtx_unlock_spin(&sched_lock); 751} 752 753/* 754 * Make a process sleeping on the specified identifier runnable. 755 * May wake more than one process if a target process is currently 756 * swapped out. 757 */ 758void 759wakeup_one(ident) 760 register void *ident; 761{ 762 register struct slpquehead *qp; 763 register struct proc *p; 764 765 mtx_lock_spin(&sched_lock); 766 qp = &slpque[LOOKUP(ident)]; 767 768 TAILQ_FOREACH(p, qp, p_slpq) { 769 if (p->p_wchan == ident) { 770 TAILQ_REMOVE(qp, p, p_slpq); 771 p->p_wchan = NULL; 772 if (p->p_stat == SSLEEP) { 773 /* OPTIMIZED EXPANSION OF setrunnable(p); */ 774 CTR3(KTR_PROC, "wakeup1: proc %p (pid %d, %s)", 775 p, p->p_pid, p->p_comm); 776 if (p->p_slptime > 1) 777 updatepri(p); 778 p->p_slptime = 0; 779 p->p_stat = SRUN; 780 if (p->p_sflag & PS_INMEM) { 781 setrunqueue(p); 782 maybe_resched(p); 783 break; 784 } else { 785 p->p_sflag |= PS_SWAPINREQ; 786 wakeup((caddr_t)&proc0); 787 } 788 /* END INLINE EXPANSION */ 789 } 790 } 791 } 792 mtx_unlock_spin(&sched_lock); 793} 794 795/* 796 * The machine independent parts of mi_switch(). 797 */ 798void 799mi_switch() 800{ 801 struct timeval new_switchtime; 802 register struct proc *p = curproc; /* XXX */ 803#if 0 804 register struct rlimit *rlim; 805#endif 806 critical_t sched_crit; 807 u_int sched_nest; 808 809 mtx_assert(&sched_lock, MA_OWNED | MA_NOTRECURSED); 810 811 /* 812 * Compute the amount of time during which the current 813 * process was running, and add that to its total so far. 814 */ 815 microuptime(&new_switchtime); 816 if (timevalcmp(&new_switchtime, PCPU_PTR(switchtime), <)) { 817#if 0 818 /* XXX: This doesn't play well with sched_lock right now. */ 819 printf("microuptime() went backwards (%ld.%06ld -> %ld.%06ld)\n", 820 PCPU_GET(switchtime.tv_sec), PCPU_GET(switchtime.tv_usec), 821 new_switchtime.tv_sec, new_switchtime.tv_usec); 822#endif 823 new_switchtime = PCPU_GET(switchtime); 824 } else { 825 p->p_runtime += (new_switchtime.tv_usec - PCPU_GET(switchtime.tv_usec)) + 826 (new_switchtime.tv_sec - PCPU_GET(switchtime.tv_sec)) * 827 (int64_t)1000000; 828 } 829 830#if 0 831 /* 832 * Check if the process exceeds its cpu resource allocation. 833 * If over max, kill it. 834 * 835 * XXX drop sched_lock, pickup Giant 836 */ 837 if (p->p_stat != SZOMB && p->p_limit->p_cpulimit != RLIM_INFINITY && 838 p->p_runtime > p->p_limit->p_cpulimit) { 839 rlim = &p->p_rlimit[RLIMIT_CPU]; 840 if (p->p_runtime / (rlim_t)1000000 >= rlim->rlim_max) { 841 mtx_unlock_spin(&sched_lock); 842 PROC_LOCK(p); 843 killproc(p, "exceeded maximum CPU limit"); 844 mtx_lock_spin(&sched_lock); 845 PROC_UNLOCK_NOSWITCH(p); 846 } else { 847 mtx_unlock_spin(&sched_lock); 848 PROC_LOCK(p); 849 psignal(p, SIGXCPU); 850 mtx_lock_spin(&sched_lock); 851 PROC_UNLOCK_NOSWITCH(p); 852 if (rlim->rlim_cur < rlim->rlim_max) { 853 /* XXX: we should make a private copy */ 854 rlim->rlim_cur += 5; 855 } 856 } 857 } 858#endif 859 860 /* 861 * Pick a new current process and record its start time. 862 */ 863 cnt.v_swtch++; 864 PCPU_SET(switchtime, new_switchtime); 865 CTR3(KTR_PROC, "mi_switch: old proc %p (pid %d, %s)", p, p->p_pid, 866 p->p_comm); 867 sched_crit = sched_lock.mtx_savecrit; 868 sched_nest = sched_lock.mtx_recurse; 869 curproc->p_lastcpu = curproc->p_oncpu; 870 curproc->p_oncpu = NOCPU; 871 clear_resched(curproc); 872 cpu_switch(); 873 curproc->p_oncpu = PCPU_GET(cpuid); 874 sched_lock.mtx_savecrit = sched_crit; 875 sched_lock.mtx_recurse = sched_nest; 876 sched_lock.mtx_lock = (uintptr_t)curproc; 877 CTR3(KTR_PROC, "mi_switch: new proc %p (pid %d, %s)", p, p->p_pid, 878 p->p_comm); 879 if (PCPU_GET(switchtime.tv_sec) == 0) 880 microuptime(PCPU_PTR(switchtime)); 881 PCPU_SET(switchticks, ticks); 882} 883 884/* 885 * Change process state to be runnable, 886 * placing it on the run queue if it is in memory, 887 * and awakening the swapper if it isn't in memory. 888 */ 889void 890setrunnable(p) 891 register struct proc *p; 892{ 893 894 mtx_lock_spin(&sched_lock); 895 switch (p->p_stat) { 896 case 0: 897 case SRUN: 898 case SZOMB: 899 case SWAIT: 900 default: 901 panic("setrunnable"); 902 case SSTOP: 903 case SSLEEP: /* e.g. when sending signals */ 904 if (p->p_sflag & PS_CVWAITQ) 905 cv_waitq_remove(p); 906 else 907 unsleep(p); 908 break; 909 910 case SIDL: 911 break; 912 } 913 p->p_stat = SRUN; 914 if (p->p_slptime > 1) 915 updatepri(p); 916 p->p_slptime = 0; 917 if ((p->p_sflag & PS_INMEM) == 0) { 918 p->p_sflag |= PS_SWAPINREQ; 919 wakeup((caddr_t)&proc0); 920 } else { 921 setrunqueue(p); 922 maybe_resched(p); 923 } 924 mtx_unlock_spin(&sched_lock); 925} 926 927/* 928 * Compute the priority of a process when running in user mode. 929 * Arrange to reschedule if the resulting priority is better 930 * than that of the current process. 931 */ 932void 933resetpriority(p) 934 register struct proc *p; 935{ 936 register unsigned int newpriority; 937 938 mtx_lock_spin(&sched_lock); 939 if (p->p_pri.pri_class == PRI_TIMESHARE) { 940 newpriority = PUSER + p->p_estcpu / INVERSE_ESTCPU_WEIGHT + 941 NICE_WEIGHT * (p->p_nice - PRIO_MIN); 942 newpriority = min(max(newpriority, PRI_MIN_TIMESHARE), 943 PRI_MAX_TIMESHARE); 944 p->p_pri.pri_user = newpriority; 945 } 946 maybe_resched(p); 947 mtx_unlock_spin(&sched_lock); 948} 949 950/* ARGSUSED */ 951static void 952sched_setup(dummy) 953 void *dummy; 954{ 955 956 callout_init(&schedcpu_callout, 1); 957 callout_init(&roundrobin_callout, 0); 958 959 /* Kick off timeout driven events by calling first time. */ 960 roundrobin(NULL); 961 schedcpu(NULL); 962} 963 964/* 965 * We adjust the priority of the current process. The priority of 966 * a process gets worse as it accumulates CPU time. The cpu usage 967 * estimator (p_estcpu) is increased here. resetpriority() will 968 * compute a different priority each time p_estcpu increases by 969 * INVERSE_ESTCPU_WEIGHT 970 * (until MAXPRI is reached). The cpu usage estimator ramps up 971 * quite quickly when the process is running (linearly), and decays 972 * away exponentially, at a rate which is proportionally slower when 973 * the system is busy. The basic principle is that the system will 974 * 90% forget that the process used a lot of CPU time in 5 * loadav 975 * seconds. This causes the system to favor processes which haven't 976 * run much recently, and to round-robin among other processes. 977 */ 978void 979schedclock(p) 980 struct proc *p; 981{ 982 983 p->p_cpticks++; 984 p->p_estcpu = ESTCPULIM(p->p_estcpu + 1); 985 if ((p->p_estcpu % INVERSE_ESTCPU_WEIGHT) == 0) { 986 resetpriority(p); 987 if (p->p_pri.pri_level >= PUSER) 988 p->p_pri.pri_level = p->p_pri.pri_user; 989 } 990} 991 992/* 993 * General purpose yield system call 994 */ 995int 996yield(struct proc *p, struct yield_args *uap) 997{ 998 999 p->p_retval[0] = 0; 1000 1001 mtx_lock_spin(&sched_lock); 1002 DROP_GIANT_NOSWITCH(); 1003 p->p_pri.pri_level = PRI_MAX_TIMESHARE; 1004 setrunqueue(p); 1005 p->p_stats->p_ru.ru_nvcsw++; 1006 mi_switch(); 1007 mtx_unlock_spin(&sched_lock); 1008 PICKUP_GIANT(); 1009 1010 return (0); 1011} 1012