kern_synch.c revision 76950
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 76950 2001-05-22 00:58:20Z alfred $ 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 */ 124/* ARGSUSED */ 125static void 126roundrobin(arg) 127 void *arg; 128{ 129 130 mtx_lock_spin(&sched_lock); 131 need_resched(curproc); 132#ifdef SMP 133 forward_roundrobin(); 134#endif 135 mtx_unlock_spin(&sched_lock); 136 137 callout_reset(&roundrobin_callout, sched_quantum, roundrobin, NULL); 138} 139 140/* 141 * Constants for digital decay and forget: 142 * 90% of (p_estcpu) usage in 5 * loadav time 143 * 95% of (p_pctcpu) usage in 60 seconds (load insensitive) 144 * Note that, as ps(1) mentions, this can let percentages 145 * total over 100% (I've seen 137.9% for 3 processes). 146 * 147 * Note that schedclock() updates p_estcpu and p_cpticks asynchronously. 148 * 149 * We wish to decay away 90% of p_estcpu in (5 * loadavg) seconds. 150 * That is, the system wants to compute a value of decay such 151 * that the following for loop: 152 * for (i = 0; i < (5 * loadavg); i++) 153 * p_estcpu *= decay; 154 * will compute 155 * p_estcpu *= 0.1; 156 * for all values of loadavg: 157 * 158 * Mathematically this loop can be expressed by saying: 159 * decay ** (5 * loadavg) ~= .1 160 * 161 * The system computes decay as: 162 * decay = (2 * loadavg) / (2 * loadavg + 1) 163 * 164 * We wish to prove that the system's computation of decay 165 * will always fulfill the equation: 166 * decay ** (5 * loadavg) ~= .1 167 * 168 * If we compute b as: 169 * b = 2 * loadavg 170 * then 171 * decay = b / (b + 1) 172 * 173 * We now need to prove two things: 174 * 1) Given factor ** (5 * loadavg) ~= .1, prove factor == b/(b+1) 175 * 2) Given b/(b+1) ** power ~= .1, prove power == (5 * loadavg) 176 * 177 * Facts: 178 * For x close to zero, exp(x) =~ 1 + x, since 179 * exp(x) = 0! + x**1/1! + x**2/2! + ... . 180 * therefore exp(-1/b) =~ 1 - (1/b) = (b-1)/b. 181 * For x close to zero, ln(1+x) =~ x, since 182 * ln(1+x) = x - x**2/2 + x**3/3 - ... -1 < x < 1 183 * therefore ln(b/(b+1)) = ln(1 - 1/(b+1)) =~ -1/(b+1). 184 * ln(.1) =~ -2.30 185 * 186 * Proof of (1): 187 * Solve (factor)**(power) =~ .1 given power (5*loadav): 188 * solving for factor, 189 * ln(factor) =~ (-2.30/5*loadav), or 190 * factor =~ exp(-1/((5/2.30)*loadav)) =~ exp(-1/(2*loadav)) = 191 * exp(-1/b) =~ (b-1)/b =~ b/(b+1). QED 192 * 193 * Proof of (2): 194 * Solve (factor)**(power) =~ .1 given factor == (b/(b+1)): 195 * solving for power, 196 * power*ln(b/(b+1)) =~ -2.30, or 197 * power =~ 2.3 * (b + 1) = 4.6*loadav + 2.3 =~ 5*loadav. QED 198 * 199 * Actual power values for the implemented algorithm are as follows: 200 * loadav: 1 2 3 4 201 * power: 5.68 10.32 14.94 19.55 202 */ 203 204/* calculations for digital decay to forget 90% of usage in 5*loadav sec */ 205#define loadfactor(loadav) (2 * (loadav)) 206#define decay_cpu(loadfac, cpu) (((loadfac) * (cpu)) / ((loadfac) + FSCALE)) 207 208/* decay 95% of `p_pctcpu' in 60 seconds; see CCPU_SHIFT before changing */ 209static fixpt_t ccpu = 0.95122942450071400909 * FSCALE; /* exp(-1/20) */ 210SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, ""); 211 212/* kernel uses `FSCALE', userland (SHOULD) use kern.fscale */ 213static int fscale __unused = FSCALE; 214SYSCTL_INT(_kern, OID_AUTO, fscale, CTLFLAG_RD, 0, FSCALE, ""); 215 216/* 217 * If `ccpu' is not equal to `exp(-1/20)' and you still want to use the 218 * faster/more-accurate formula, you'll have to estimate CCPU_SHIFT below 219 * and possibly adjust FSHIFT in "param.h" so that (FSHIFT >= CCPU_SHIFT). 220 * 221 * To estimate CCPU_SHIFT for exp(-1/20), the following formula was used: 222 * 1 - exp(-1/20) ~= 0.0487 ~= 0.0488 == 1 (fixed pt, *11* bits). 223 * 224 * If you don't want to bother with the faster/more-accurate formula, you 225 * can set CCPU_SHIFT to (FSHIFT + 1) which will use a slower/less-accurate 226 * (more general) method of calculating the %age of CPU used by a process. 227 */ 228#define CCPU_SHIFT 11 229 230/* 231 * Recompute process priorities, every hz ticks. 232 * MP-safe, called without the Giant mutex. 233 */ 234/* ARGSUSED */ 235static void 236schedcpu(arg) 237 void *arg; 238{ 239 register fixpt_t loadfac = loadfactor(averunnable.ldavg[0]); 240 register struct proc *p; 241 register int realstathz, s; 242 243 realstathz = stathz ? stathz : hz; 244 sx_slock(&allproc_lock); 245 LIST_FOREACH(p, &allproc, p_list) { 246 /* 247 * Increment time in/out of memory and sleep time 248 * (if sleeping). We ignore overflow; with 16-bit int's 249 * (remember them?) overflow takes 45 days. 250 if (p->p_stat == SWAIT) 251 continue; 252 */ 253 mtx_lock_spin(&sched_lock); 254 p->p_swtime++; 255 if (p->p_stat == SSLEEP || p->p_stat == SSTOP) 256 p->p_slptime++; 257 p->p_pctcpu = (p->p_pctcpu * ccpu) >> FSHIFT; 258 /* 259 * If the process has slept the entire second, 260 * stop recalculating its priority until it wakes up. 261 */ 262 if (p->p_slptime > 1) { 263 mtx_unlock_spin(&sched_lock); 264 continue; 265 } 266 267 /* 268 * prevent state changes and protect run queue 269 */ 270 s = splhigh(); 271 272 /* 273 * p_pctcpu is only for ps. 274 */ 275#if (FSHIFT >= CCPU_SHIFT) 276 p->p_pctcpu += (realstathz == 100)? 277 ((fixpt_t) p->p_cpticks) << (FSHIFT - CCPU_SHIFT): 278 100 * (((fixpt_t) p->p_cpticks) 279 << (FSHIFT - CCPU_SHIFT)) / realstathz; 280#else 281 p->p_pctcpu += ((FSCALE - ccpu) * 282 (p->p_cpticks * FSCALE / realstathz)) >> FSHIFT; 283#endif 284 p->p_cpticks = 0; 285 p->p_estcpu = decay_cpu(loadfac, p->p_estcpu); 286 resetpriority(p); 287 if (p->p_pri.pri_level >= PUSER) { 288 if ((p != curproc) && 289#ifdef SMP 290 p->p_oncpu == NOCPU && /* idle */ 291#endif 292 p->p_stat == SRUN && 293 (p->p_sflag & PS_INMEM) && 294 (p->p_pri.pri_level / RQ_PPQ) != 295 (p->p_pri.pri_user / RQ_PPQ)) { 296 remrunqueue(p); 297 p->p_pri.pri_level = p->p_pri.pri_user; 298 setrunqueue(p); 299 } else 300 p->p_pri.pri_level = p->p_pri.pri_user; 301 } 302 mtx_unlock_spin(&sched_lock); 303 splx(s); 304 } 305 sx_sunlock(&allproc_lock); 306 vmmeter(); 307 wakeup((caddr_t)&lbolt); 308 callout_reset(&schedcpu_callout, hz, schedcpu, NULL); 309} 310 311/* 312 * Recalculate the priority of a process after it has slept for a while. 313 * For all load averages >= 1 and max p_estcpu of 255, sleeping for at 314 * least six times the loadfactor will decay p_estcpu to zero. 315 */ 316void 317updatepri(p) 318 register struct proc *p; 319{ 320 register unsigned int newcpu = p->p_estcpu; 321 register fixpt_t loadfac = loadfactor(averunnable.ldavg[0]); 322 323 if (p->p_slptime > 5 * loadfac) 324 p->p_estcpu = 0; 325 else { 326 p->p_slptime--; /* the first time was done in schedcpu */ 327 while (newcpu && --p->p_slptime) 328 newcpu = decay_cpu(loadfac, newcpu); 329 p->p_estcpu = newcpu; 330 } 331 resetpriority(p); 332} 333 334/* 335 * We're only looking at 7 bits of the address; everything is 336 * aligned to 4, lots of things are aligned to greater powers 337 * of 2. Shift right by 8, i.e. drop the bottom 256 worth. 338 */ 339#define TABLESIZE 128 340static TAILQ_HEAD(slpquehead, proc) slpque[TABLESIZE]; 341#define LOOKUP(x) (((intptr_t)(x) >> 8) & (TABLESIZE - 1)) 342 343void 344sleepinit(void) 345{ 346 int i; 347 348 sched_quantum = hz/10; 349 hogticks = 2 * sched_quantum; 350 for (i = 0; i < TABLESIZE; i++) 351 TAILQ_INIT(&slpque[i]); 352} 353 354/* 355 * General sleep call. Suspends the current process until a wakeup is 356 * performed on the specified identifier. The process will then be made 357 * runnable with the specified priority. Sleeps at most timo/hz seconds 358 * (0 means no timeout). If pri includes PCATCH flag, signals are checked 359 * before and after sleeping, else signals are not checked. Returns 0 if 360 * awakened, EWOULDBLOCK if the timeout expires. If PCATCH is set and a 361 * signal needs to be delivered, ERESTART is returned if the current system 362 * call should be restarted if possible, and EINTR is returned if the system 363 * call should be interrupted by the signal (return EINTR). 364 * 365 * The mutex argument is exited before the caller is suspended, and 366 * entered before msleep returns. If priority includes the PDROP 367 * flag the mutex is not entered before returning. 368 */ 369int 370msleep(ident, mtx, priority, wmesg, timo) 371 void *ident; 372 struct mtx *mtx; 373 int priority, timo; 374 const char *wmesg; 375{ 376 struct proc *p = curproc; 377 int sig, catch = priority & PCATCH; 378 int rval = 0; 379 WITNESS_SAVE_DECL(mtx); 380 381#ifdef KTRACE 382 if (p && KTRPOINT(p, KTR_CSW)) 383 ktrcsw(p->p_tracep, 1, 0); 384#endif 385 WITNESS_SLEEP(0, &mtx->mtx_object); 386 mtx_lock_spin(&sched_lock); 387 if (cold || panicstr) { 388 /* 389 * After a panic, or during autoconfiguration, 390 * just give interrupts a chance, then just return; 391 * don't run any other procs or panic below, 392 * in case this is the idle process and already asleep. 393 */ 394 if (mtx != NULL && priority & PDROP) 395 mtx_unlock_flags(mtx, MTX_NOSWITCH); 396 mtx_unlock_spin(&sched_lock); 397 return (0); 398 } 399 400 DROP_GIANT_NOSWITCH(); 401 402 if (mtx != NULL) { 403 mtx_assert(mtx, MA_OWNED | MA_NOTRECURSED); 404 WITNESS_SAVE(&mtx->mtx_object, mtx); 405 mtx_unlock_flags(mtx, MTX_NOSWITCH); 406 if (priority & PDROP) 407 mtx = NULL; 408 } 409 410 KASSERT(p != NULL, ("msleep1")); 411 KASSERT(ident != NULL && p->p_stat == SRUN, ("msleep")); 412 /* 413 * Process may be sitting on a slpque if asleep() was called, remove 414 * it before re-adding. 415 */ 416 if (p->p_wchan != NULL) 417 unsleep(p); 418 419 p->p_wchan = ident; 420 p->p_wmesg = wmesg; 421 p->p_slptime = 0; 422 p->p_pri.pri_level = priority & PRIMASK; 423 CTR4(KTR_PROC, "msleep: proc %p (pid %d, %s), schedlock %p", 424 p, p->p_pid, p->p_comm, (void *) sched_lock.mtx_lock); 425 TAILQ_INSERT_TAIL(&slpque[LOOKUP(ident)], p, p_slpq); 426 if (timo) 427 callout_reset(&p->p_slpcallout, timo, endtsleep, p); 428 /* 429 * We put ourselves on the sleep queue and start our timeout 430 * before calling CURSIG, as we could stop there, and a wakeup 431 * or a SIGCONT (or both) could occur while we were stopped. 432 * A SIGCONT would cause us to be marked as SSLEEP 433 * without resuming us, thus we must be ready for sleep 434 * when CURSIG is called. If the wakeup happens while we're 435 * stopped, p->p_wchan will be 0 upon return from CURSIG. 436 */ 437 if (catch) { 438 CTR4(KTR_PROC, 439 "msleep caught: proc %p (pid %d, %s), schedlock %p", 440 p, p->p_pid, p->p_comm, (void *) sched_lock.mtx_lock); 441 p->p_sflag |= PS_SINTR; 442 mtx_unlock_spin(&sched_lock); 443 if ((sig = CURSIG(p))) { 444 mtx_lock_spin(&sched_lock); 445 if (p->p_wchan) 446 unsleep(p); 447 p->p_stat = SRUN; 448 goto resume; 449 } 450 mtx_lock_spin(&sched_lock); 451 if (p->p_wchan == NULL) { 452 catch = 0; 453 goto resume; 454 } 455 } else 456 sig = 0; 457 p->p_stat = SSLEEP; 458 p->p_stats->p_ru.ru_nvcsw++; 459 mi_switch(); 460 CTR4(KTR_PROC, 461 "msleep resume: proc %p (pid %d, %s), schedlock %p", 462 p, p->p_pid, p->p_comm, (void *) sched_lock.mtx_lock); 463resume: 464 p->p_sflag &= ~PS_SINTR; 465 if (p->p_sflag & PS_TIMEOUT) { 466 p->p_sflag &= ~PS_TIMEOUT; 467 if (sig == 0) { 468#ifdef KTRACE 469 if (KTRPOINT(p, KTR_CSW)) 470 ktrcsw(p->p_tracep, 0, 0); 471#endif 472 rval = EWOULDBLOCK; 473 mtx_unlock_spin(&sched_lock); 474 goto out; 475 } 476 } else if (timo) 477 callout_stop(&p->p_slpcallout); 478 mtx_unlock_spin(&sched_lock); 479 480 if (catch && (sig != 0 || (sig = CURSIG(p)))) { 481#ifdef KTRACE 482 if (KTRPOINT(p, KTR_CSW)) 483 ktrcsw(p->p_tracep, 0, 0); 484#endif 485 PROC_LOCK(p); 486 if (SIGISMEMBER(p->p_sigacts->ps_sigintr, sig)) 487 rval = EINTR; 488 else 489 rval = ERESTART; 490 PROC_UNLOCK(p); 491 goto out; 492 } 493out: 494#ifdef KTRACE 495 if (KTRPOINT(p, KTR_CSW)) 496 ktrcsw(p->p_tracep, 0, 0); 497#endif 498 PICKUP_GIANT(); 499 if (mtx != NULL) { 500 mtx_lock(mtx); 501 WITNESS_RESTORE(&mtx->mtx_object, mtx); 502 } 503 return (rval); 504} 505 506/* 507 * asleep() - async sleep call. Place process on wait queue and return 508 * immediately without blocking. The process stays runnable until mawait() 509 * is called. If ident is NULL, remove process from wait queue if it is still 510 * on one. 511 * 512 * Only the most recent sleep condition is effective when making successive 513 * calls to asleep() or when calling msleep(). 514 * 515 * The timeout, if any, is not initiated until mawait() is called. The sleep 516 * priority, signal, and timeout is specified in the asleep() call but may be 517 * overriden in the mawait() call. 518 * 519 * <<<<<<<< EXPERIMENTAL, UNTESTED >>>>>>>>>> 520 */ 521 522int 523asleep(void *ident, int priority, const char *wmesg, int timo) 524{ 525 struct proc *p = curproc; 526 int s; 527 528 /* 529 * obtain sched_lock while manipulating sleep structures and slpque. 530 * 531 * Remove preexisting wait condition (if any) and place process 532 * on appropriate slpque, but do not put process to sleep. 533 */ 534 535 s = splhigh(); 536 mtx_lock_spin(&sched_lock); 537 538 if (p->p_wchan != NULL) 539 unsleep(p); 540 541 if (ident) { 542 p->p_wchan = ident; 543 p->p_wmesg = wmesg; 544 p->p_slptime = 0; 545 p->p_asleep.as_priority = priority; 546 p->p_asleep.as_timo = timo; 547 TAILQ_INSERT_TAIL(&slpque[LOOKUP(ident)], p, p_slpq); 548 } 549 550 mtx_unlock_spin(&sched_lock); 551 splx(s); 552 553 return(0); 554} 555 556/* 557 * mawait() - wait for async condition to occur. The process blocks until 558 * wakeup() is called on the most recent asleep() address. If wakeup is called 559 * prior to mawait(), mawait() winds up being a NOP. 560 * 561 * If mawait() is called more then once (without an intervening asleep() call), 562 * mawait() is still effectively a NOP but it calls mi_switch() to give other 563 * processes some cpu before returning. The process is left runnable. 564 * 565 * <<<<<<<< EXPERIMENTAL, UNTESTED >>>>>>>>>> 566 */ 567 568int 569mawait(struct mtx *mtx, int priority, int timo) 570{ 571 struct proc *p = curproc; 572 int rval = 0; 573 int s; 574 WITNESS_SAVE_DECL(mtx); 575 576 WITNESS_SLEEP(0, &mtx->mtx_object); 577 mtx_lock_spin(&sched_lock); 578 DROP_GIANT_NOSWITCH(); 579 if (mtx != NULL) { 580 mtx_assert(mtx, MA_OWNED | MA_NOTRECURSED); 581 WITNESS_SAVE(&mtx->mtx_object, mtx); 582 mtx_unlock_flags(mtx, MTX_NOSWITCH); 583 if (priority & PDROP) 584 mtx = NULL; 585 } 586 587 s = splhigh(); 588 589 if (p->p_wchan != NULL) { 590 int sig; 591 int catch; 592 593 /* 594 * The call to mawait() can override defaults specified in 595 * the original asleep(). 596 */ 597 if (priority < 0) 598 priority = p->p_asleep.as_priority; 599 if (timo < 0) 600 timo = p->p_asleep.as_timo; 601 602 /* 603 * Install timeout 604 */ 605 606 if (timo) 607 callout_reset(&p->p_slpcallout, timo, endtsleep, p); 608 609 sig = 0; 610 catch = priority & PCATCH; 611 612 if (catch) { 613 p->p_sflag |= PS_SINTR; 614 mtx_unlock_spin(&sched_lock); 615 if ((sig = CURSIG(p))) { 616 mtx_lock_spin(&sched_lock); 617 if (p->p_wchan) 618 unsleep(p); 619 p->p_stat = SRUN; 620 goto resume; 621 } 622 mtx_lock_spin(&sched_lock); 623 if (p->p_wchan == NULL) { 624 catch = 0; 625 goto resume; 626 } 627 } 628 p->p_stat = SSLEEP; 629 p->p_stats->p_ru.ru_nvcsw++; 630 mi_switch(); 631resume: 632 633 splx(s); 634 p->p_sflag &= ~PS_SINTR; 635 if (p->p_sflag & PS_TIMEOUT) { 636 p->p_sflag &= ~PS_TIMEOUT; 637 if (sig == 0) { 638#ifdef KTRACE 639 if (KTRPOINT(p, KTR_CSW)) 640 ktrcsw(p->p_tracep, 0, 0); 641#endif 642 rval = EWOULDBLOCK; 643 mtx_unlock_spin(&sched_lock); 644 goto out; 645 } 646 } else if (timo) 647 callout_stop(&p->p_slpcallout); 648 mtx_unlock_spin(&sched_lock); 649 650 if (catch && (sig != 0 || (sig = CURSIG(p)))) { 651#ifdef KTRACE 652 if (KTRPOINT(p, KTR_CSW)) 653 ktrcsw(p->p_tracep, 0, 0); 654#endif 655 PROC_LOCK(p); 656 if (SIGISMEMBER(p->p_sigacts->ps_sigintr, sig)) 657 rval = EINTR; 658 else 659 rval = ERESTART; 660 PROC_UNLOCK(p); 661 goto out; 662 } 663#ifdef KTRACE 664 if (KTRPOINT(p, KTR_CSW)) 665 ktrcsw(p->p_tracep, 0, 0); 666#endif 667 } else { 668 /* 669 * If as_priority is 0, mawait() has been called without an 670 * intervening asleep(). We are still effectively a NOP, 671 * but we call mi_switch() for safety. 672 */ 673 674 if (p->p_asleep.as_priority == 0) { 675 p->p_stats->p_ru.ru_nvcsw++; 676 mi_switch(); 677 } 678 mtx_unlock_spin(&sched_lock); 679 splx(s); 680 } 681 682 /* 683 * clear p_asleep.as_priority as an indication that mawait() has been 684 * called. If mawait() is called again without an intervening asleep(), 685 * mawait() is still effectively a NOP but the above mi_switch() code 686 * is triggered as a safety. 687 */ 688 p->p_asleep.as_priority = 0; 689 690out: 691 PICKUP_GIANT(); 692 if (mtx != NULL) { 693 mtx_lock(mtx); 694 WITNESS_RESTORE(&mtx->mtx_object, mtx); 695 } 696 return (rval); 697} 698 699/* 700 * Implement timeout for msleep or asleep()/mawait() 701 * 702 * If process hasn't been awakened (wchan non-zero), 703 * set timeout flag and undo the sleep. If proc 704 * is stopped, just unsleep so it will remain stopped. 705 * MP-safe, called without the Giant mutex. 706 */ 707static void 708endtsleep(arg) 709 void *arg; 710{ 711 register struct proc *p; 712 int s; 713 714 p = (struct proc *)arg; 715 CTR4(KTR_PROC, 716 "endtsleep: proc %p (pid %d, %s), schedlock %p", 717 p, p->p_pid, p->p_comm, (void *) sched_lock.mtx_lock); 718 s = splhigh(); 719 mtx_lock_spin(&sched_lock); 720 if (p->p_wchan) { 721 if (p->p_stat == SSLEEP) 722 setrunnable(p); 723 else 724 unsleep(p); 725 p->p_sflag |= PS_TIMEOUT; 726 } 727 mtx_unlock_spin(&sched_lock); 728 splx(s); 729} 730 731/* 732 * Remove a process from its wait queue 733 */ 734void 735unsleep(p) 736 register struct proc *p; 737{ 738 int s; 739 740 s = splhigh(); 741 mtx_lock_spin(&sched_lock); 742 if (p->p_wchan) { 743 TAILQ_REMOVE(&slpque[LOOKUP(p->p_wchan)], p, p_slpq); 744 p->p_wchan = NULL; 745 } 746 mtx_unlock_spin(&sched_lock); 747 splx(s); 748} 749 750/* 751 * Make all processes sleeping on the specified identifier runnable. 752 */ 753void 754wakeup(ident) 755 register void *ident; 756{ 757 register struct slpquehead *qp; 758 register struct proc *p; 759 int s; 760 761 s = splhigh(); 762 mtx_lock_spin(&sched_lock); 763 qp = &slpque[LOOKUP(ident)]; 764restart: 765 TAILQ_FOREACH(p, qp, p_slpq) { 766 if (p->p_wchan == ident) { 767 TAILQ_REMOVE(qp, p, p_slpq); 768 p->p_wchan = NULL; 769 if (p->p_stat == SSLEEP) { 770 /* OPTIMIZED EXPANSION OF setrunnable(p); */ 771 CTR4(KTR_PROC, 772 "wakeup: proc %p (pid %d, %s), schedlock %p", 773 p, p->p_pid, p->p_comm, (void *) sched_lock.mtx_lock); 774 if (p->p_slptime > 1) 775 updatepri(p); 776 p->p_slptime = 0; 777 p->p_stat = SRUN; 778 if (p->p_sflag & PS_INMEM) { 779 setrunqueue(p); 780 maybe_resched(p); 781 } else { 782 p->p_sflag |= PS_SWAPINREQ; 783 wakeup((caddr_t)&proc0); 784 } 785 /* END INLINE EXPANSION */ 786 goto restart; 787 } 788 } 789 } 790 mtx_unlock_spin(&sched_lock); 791 splx(s); 792} 793 794/* 795 * Make a process sleeping on the specified identifier runnable. 796 * May wake more than one process if a target process is currently 797 * swapped out. 798 */ 799void 800wakeup_one(ident) 801 register void *ident; 802{ 803 register struct slpquehead *qp; 804 register struct proc *p; 805 int s; 806 807 s = splhigh(); 808 mtx_lock_spin(&sched_lock); 809 qp = &slpque[LOOKUP(ident)]; 810 811 TAILQ_FOREACH(p, qp, p_slpq) { 812 if (p->p_wchan == ident) { 813 TAILQ_REMOVE(qp, p, p_slpq); 814 p->p_wchan = NULL; 815 if (p->p_stat == SSLEEP) { 816 /* OPTIMIZED EXPANSION OF setrunnable(p); */ 817 CTR4(KTR_PROC, 818 "wakeup1: proc %p (pid %d, %s), schedlock %p", 819 p, p->p_pid, p->p_comm, (void *) sched_lock.mtx_lock); 820 if (p->p_slptime > 1) 821 updatepri(p); 822 p->p_slptime = 0; 823 p->p_stat = SRUN; 824 if (p->p_sflag & PS_INMEM) { 825 setrunqueue(p); 826 maybe_resched(p); 827 break; 828 } else { 829 p->p_sflag |= PS_SWAPINREQ; 830 wakeup((caddr_t)&proc0); 831 } 832 /* END INLINE EXPANSION */ 833 } 834 } 835 } 836 mtx_unlock_spin(&sched_lock); 837 splx(s); 838} 839 840/* 841 * The machine independent parts of mi_switch(). 842 * Must be called at splstatclock() or higher. 843 */ 844void 845mi_switch() 846{ 847 struct timeval new_switchtime; 848 register struct proc *p = curproc; /* XXX */ 849#if 0 850 register struct rlimit *rlim; 851#endif 852 u_int sched_nest; 853 854 mtx_assert(&sched_lock, MA_OWNED | MA_NOTRECURSED); 855 856 /* 857 * Compute the amount of time during which the current 858 * process was running, and add that to its total so far. 859 */ 860 microuptime(&new_switchtime); 861 if (timevalcmp(&new_switchtime, PCPU_PTR(switchtime), <)) { 862#if 0 863 /* XXX: This doesn't play well with sched_lock right now. */ 864 printf("microuptime() went backwards (%ld.%06ld -> %ld.%06ld)\n", 865 PCPU_GET(switchtime.tv_sec), PCPU_GET(switchtime.tv_usec), 866 new_switchtime.tv_sec, new_switchtime.tv_usec); 867#endif 868 new_switchtime = PCPU_GET(switchtime); 869 } else { 870 p->p_runtime += (new_switchtime.tv_usec - PCPU_GET(switchtime.tv_usec)) + 871 (new_switchtime.tv_sec - PCPU_GET(switchtime.tv_sec)) * 872 (int64_t)1000000; 873 } 874 875#if 0 876 /* 877 * Check if the process exceeds its cpu resource allocation. 878 * If over max, kill it. 879 * 880 * XXX drop sched_lock, pickup Giant 881 */ 882 if (p->p_stat != SZOMB && p->p_limit->p_cpulimit != RLIM_INFINITY && 883 p->p_runtime > p->p_limit->p_cpulimit) { 884 rlim = &p->p_rlimit[RLIMIT_CPU]; 885 if (p->p_runtime / (rlim_t)1000000 >= rlim->rlim_max) { 886 mtx_unlock_spin(&sched_lock); 887 PROC_LOCK(p); 888 killproc(p, "exceeded maximum CPU limit"); 889 mtx_lock_spin(&sched_lock); 890 PROC_UNLOCK_NOSWITCH(p); 891 } else { 892 mtx_unlock_spin(&sched_lock); 893 PROC_LOCK(p); 894 psignal(p, SIGXCPU); 895 mtx_lock_spin(&sched_lock); 896 PROC_UNLOCK_NOSWITCH(p); 897 if (rlim->rlim_cur < rlim->rlim_max) { 898 /* XXX: we should make a private copy */ 899 rlim->rlim_cur += 5; 900 } 901 } 902 } 903#endif 904 905 /* 906 * Pick a new current process and record its start time. 907 */ 908 cnt.v_swtch++; 909 PCPU_SET(switchtime, new_switchtime); 910 CTR4(KTR_PROC, "mi_switch: old proc %p (pid %d, %s), schedlock %p", 911 p, p->p_pid, p->p_comm, (void *) sched_lock.mtx_lock); 912 sched_nest = sched_lock.mtx_recurse; 913 curproc->p_lastcpu = curproc->p_oncpu; 914 curproc->p_oncpu = NOCPU; 915 clear_resched(curproc); 916 cpu_switch(); 917 curproc->p_oncpu = PCPU_GET(cpuid); 918 sched_lock.mtx_recurse = sched_nest; 919 sched_lock.mtx_lock = (uintptr_t)curproc; 920 CTR4(KTR_PROC, "mi_switch: new proc %p (pid %d, %s), schedlock %p", 921 p, p->p_pid, p->p_comm, (void *) sched_lock.mtx_lock); 922 if (PCPU_GET(switchtime.tv_sec) == 0) 923 microuptime(PCPU_PTR(switchtime)); 924 PCPU_SET(switchticks, ticks); 925} 926 927/* 928 * Change process state to be runnable, 929 * placing it on the run queue if it is in memory, 930 * and awakening the swapper if it isn't in memory. 931 */ 932void 933setrunnable(p) 934 register struct proc *p; 935{ 936 register int s; 937 938 s = splhigh(); 939 mtx_lock_spin(&sched_lock); 940 switch (p->p_stat) { 941 case 0: 942 case SRUN: 943 case SZOMB: 944 case SWAIT: 945 default: 946 panic("setrunnable"); 947 case SSTOP: 948 case SSLEEP: /* e.g. when sending signals */ 949 if (p->p_sflag & PS_CVWAITQ) 950 cv_waitq_remove(p); 951 else 952 unsleep(p); 953 break; 954 955 case SIDL: 956 break; 957 } 958 p->p_stat = SRUN; 959 if (p->p_sflag & PS_INMEM) 960 setrunqueue(p); 961 splx(s); 962 if (p->p_slptime > 1) 963 updatepri(p); 964 p->p_slptime = 0; 965 if ((p->p_sflag & PS_INMEM) == 0) { 966 p->p_sflag |= PS_SWAPINREQ; 967 wakeup((caddr_t)&proc0); 968 } 969 else 970 maybe_resched(p); 971 mtx_unlock_spin(&sched_lock); 972} 973 974/* 975 * Compute the priority of a process when running in user mode. 976 * Arrange to reschedule if the resulting priority is better 977 * than that of the current process. 978 */ 979void 980resetpriority(p) 981 register struct proc *p; 982{ 983 register unsigned int newpriority; 984 985 mtx_lock_spin(&sched_lock); 986 if (p->p_pri.pri_class == PRI_TIMESHARE) { 987 newpriority = PUSER + p->p_estcpu / INVERSE_ESTCPU_WEIGHT + 988 NICE_WEIGHT * (p->p_nice - PRIO_MIN); 989 newpriority = min(max(newpriority, PRI_MIN_TIMESHARE), 990 PRI_MAX_TIMESHARE); 991 p->p_pri.pri_user = newpriority; 992 } 993 maybe_resched(p); 994 mtx_unlock_spin(&sched_lock); 995} 996 997/* ARGSUSED */ 998static void 999sched_setup(dummy) 1000 void *dummy; 1001{ 1002 1003 callout_init(&schedcpu_callout, 1); 1004 callout_init(&roundrobin_callout, 0); 1005 1006 /* Kick off timeout driven events by calling first time. */ 1007 roundrobin(NULL); 1008 schedcpu(NULL); 1009} 1010 1011/* 1012 * We adjust the priority of the current process. The priority of 1013 * a process gets worse as it accumulates CPU time. The cpu usage 1014 * estimator (p_estcpu) is increased here. resetpriority() will 1015 * compute a different priority each time p_estcpu increases by 1016 * INVERSE_ESTCPU_WEIGHT 1017 * (until MAXPRI is reached). The cpu usage estimator ramps up 1018 * quite quickly when the process is running (linearly), and decays 1019 * away exponentially, at a rate which is proportionally slower when 1020 * the system is busy. The basic principle is that the system will 1021 * 90% forget that the process used a lot of CPU time in 5 * loadav 1022 * seconds. This causes the system to favor processes which haven't 1023 * run much recently, and to round-robin among other processes. 1024 */ 1025void 1026schedclock(p) 1027 struct proc *p; 1028{ 1029 1030 p->p_cpticks++; 1031 p->p_estcpu = ESTCPULIM(p->p_estcpu + 1); 1032 if ((p->p_estcpu % INVERSE_ESTCPU_WEIGHT) == 0) { 1033 resetpriority(p); 1034 if (p->p_pri.pri_level >= PUSER) 1035 p->p_pri.pri_level = p->p_pri.pri_user; 1036 } 1037} 1038 1039/* 1040 * General purpose yield system call 1041 */ 1042int 1043yield(struct proc *p, struct yield_args *uap) 1044{ 1045 int s; 1046 1047 p->p_retval[0] = 0; 1048 1049 s = splhigh(); 1050 mtx_lock_spin(&sched_lock); 1051 DROP_GIANT_NOSWITCH(); 1052 p->p_pri.pri_level = PRI_MAX_TIMESHARE; 1053 setrunqueue(p); 1054 p->p_stats->p_ru.ru_nvcsw++; 1055 mi_switch(); 1056 mtx_unlock_spin(&sched_lock); 1057 PICKUP_GIANT(); 1058 splx(s); 1059 1060 return (0); 1061} 1062