kern_synch.c revision 72886
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 72886 2001-02-22 13:47:01Z jhb $ 40 */ 41 42#include "opt_ktrace.h" 43 44#include <sys/param.h> 45#include <sys/systm.h> 46#include <sys/proc.h> 47#include <sys/ipl.h> 48#include <sys/kernel.h> 49#include <sys/ktr.h> 50#include <sys/condvar.h> 51#include <sys/lock.h> 52#include <sys/mutex.h> 53#include <sys/signalvar.h> 54#include <sys/resourcevar.h> 55#include <sys/vmmeter.h> 56#include <sys/sysctl.h> 57#include <sys/sysproto.h> 58#include <vm/vm.h> 59#include <vm/vm_extern.h> 60#ifdef KTRACE 61#include <sys/uio.h> 62#include <sys/ktrace.h> 63#endif 64 65#include <machine/cpu.h> 66#include <machine/smp.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(); 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(); 132 mtx_unlock_spin(&sched_lock); 133#ifdef SMP 134 forward_roundrobin(); 135#endif 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 ALLPROC_LOCK(AP_SHARED); 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 ALLPROC_LOCK(AP_RELEASE); 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 s, 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); 386 mtx_lock_spin(&sched_lock); 387 s = splhigh(); 388 if (cold || panicstr) { 389 /* 390 * After a panic, or during autoconfiguration, 391 * just give interrupts a chance, then just return; 392 * don't run any other procs or panic below, 393 * in case this is the idle process and already asleep. 394 */ 395 if (mtx != NULL && priority & PDROP) 396 mtx_unlock_flags(mtx, MTX_NOSWITCH); 397 mtx_unlock_spin(&sched_lock); 398 splx(s); 399 return (0); 400 } 401 402 DROP_GIANT_NOSWITCH(); 403 404 if (mtx != NULL) { 405 mtx_assert(mtx, MA_OWNED | MA_NOTRECURSED); 406 WITNESS_SAVE(mtx, mtx); 407 mtx_unlock_flags(mtx, MTX_NOSWITCH); 408 if (priority & PDROP) 409 mtx = NULL; 410 } 411 412 KASSERT(p != NULL, ("msleep1")); 413 KASSERT(ident != NULL && p->p_stat == SRUN, ("msleep")); 414 /* 415 * Process may be sitting on a slpque if asleep() was called, remove 416 * it before re-adding. 417 */ 418 if (p->p_wchan != NULL) 419 unsleep(p); 420 421 p->p_wchan = ident; 422 p->p_wmesg = wmesg; 423 p->p_slptime = 0; 424 p->p_pri.pri_level = priority & PRIMASK; 425 CTR4(KTR_PROC, "msleep: proc %p (pid %d, %s), schedlock %p", 426 p, p->p_pid, p->p_comm, (void *) sched_lock.mtx_lock); 427 TAILQ_INSERT_TAIL(&slpque[LOOKUP(ident)], p, p_slpq); 428 if (timo) 429 callout_reset(&p->p_slpcallout, timo, endtsleep, p); 430 /* 431 * We put ourselves on the sleep queue and start our timeout 432 * before calling CURSIG, as we could stop there, and a wakeup 433 * or a SIGCONT (or both) could occur while we were stopped. 434 * A SIGCONT would cause us to be marked as SSLEEP 435 * without resuming us, thus we must be ready for sleep 436 * when CURSIG is called. If the wakeup happens while we're 437 * stopped, p->p_wchan will be 0 upon return from CURSIG. 438 */ 439 if (catch) { 440 CTR4(KTR_PROC, 441 "msleep caught: proc %p (pid %d, %s), schedlock %p", 442 p, p->p_pid, p->p_comm, (void *) sched_lock.mtx_lock); 443 p->p_sflag |= PS_SINTR; 444 mtx_unlock_spin(&sched_lock); 445 if ((sig = CURSIG(p))) { 446 mtx_lock_spin(&sched_lock); 447 if (p->p_wchan) 448 unsleep(p); 449 p->p_stat = SRUN; 450 goto resume; 451 } 452 mtx_lock_spin(&sched_lock); 453 if (p->p_wchan == NULL) { 454 catch = 0; 455 goto resume; 456 } 457 } else 458 sig = 0; 459 p->p_stat = SSLEEP; 460 p->p_stats->p_ru.ru_nvcsw++; 461 mi_switch(); 462 CTR4(KTR_PROC, 463 "msleep resume: proc %p (pid %d, %s), schedlock %p", 464 p, p->p_pid, p->p_comm, (void *) sched_lock.mtx_lock); 465resume: 466 splx(s); 467 p->p_sflag &= ~PS_SINTR; 468 if (p->p_sflag & PS_TIMEOUT) { 469 p->p_sflag &= ~PS_TIMEOUT; 470 if (sig == 0) { 471#ifdef KTRACE 472 if (KTRPOINT(p, KTR_CSW)) 473 ktrcsw(p->p_tracep, 0, 0); 474#endif 475 rval = EWOULDBLOCK; 476 mtx_unlock_spin(&sched_lock); 477 goto out; 478 } 479 } else if (timo) 480 callout_stop(&p->p_slpcallout); 481 mtx_unlock_spin(&sched_lock); 482 483 if (catch && (sig != 0 || (sig = CURSIG(p)))) { 484#ifdef KTRACE 485 if (KTRPOINT(p, KTR_CSW)) 486 ktrcsw(p->p_tracep, 0, 0); 487#endif 488 if (SIGISMEMBER(p->p_sigacts->ps_sigintr, sig)) 489 rval = EINTR; 490 else 491 rval = ERESTART; 492 goto out; 493 } 494out: 495#ifdef KTRACE 496 if (KTRPOINT(p, KTR_CSW)) 497 ktrcsw(p->p_tracep, 0, 0); 498#endif 499 PICKUP_GIANT(); 500 if (mtx != NULL) { 501 mtx_lock(mtx); 502 WITNESS_RESTORE(mtx, mtx); 503 } 504 return (rval); 505} 506 507/* 508 * asleep() - async sleep call. Place process on wait queue and return 509 * immediately without blocking. The process stays runnable until mawait() 510 * is called. If ident is NULL, remove process from wait queue if it is still 511 * on one. 512 * 513 * Only the most recent sleep condition is effective when making successive 514 * calls to asleep() or when calling msleep(). 515 * 516 * The timeout, if any, is not initiated until mawait() is called. The sleep 517 * priority, signal, and timeout is specified in the asleep() call but may be 518 * overriden in the mawait() call. 519 * 520 * <<<<<<<< EXPERIMENTAL, UNTESTED >>>>>>>>>> 521 */ 522 523int 524asleep(void *ident, int priority, const char *wmesg, int timo) 525{ 526 struct proc *p = curproc; 527 int s; 528 529 /* 530 * obtain sched_lock while manipulating sleep structures and slpque. 531 * 532 * Remove preexisting wait condition (if any) and place process 533 * on appropriate slpque, but do not put process to sleep. 534 */ 535 536 s = splhigh(); 537 mtx_lock_spin(&sched_lock); 538 539 if (p->p_wchan != NULL) 540 unsleep(p); 541 542 if (ident) { 543 p->p_wchan = ident; 544 p->p_wmesg = wmesg; 545 p->p_slptime = 0; 546 p->p_asleep.as_priority = priority; 547 p->p_asleep.as_timo = timo; 548 TAILQ_INSERT_TAIL(&slpque[LOOKUP(ident)], p, p_slpq); 549 } 550 551 mtx_unlock_spin(&sched_lock); 552 splx(s); 553 554 return(0); 555} 556 557/* 558 * mawait() - wait for async condition to occur. The process blocks until 559 * wakeup() is called on the most recent asleep() address. If wakeup is called 560 * prior to mawait(), mawait() winds up being a NOP. 561 * 562 * If mawait() is called more then once (without an intervening asleep() call), 563 * mawait() is still effectively a NOP but it calls mi_switch() to give other 564 * processes some cpu before returning. The process is left runnable. 565 * 566 * <<<<<<<< EXPERIMENTAL, UNTESTED >>>>>>>>>> 567 */ 568 569int 570mawait(struct mtx *mtx, int priority, int timo) 571{ 572 struct proc *p = curproc; 573 int rval = 0; 574 int s; 575 WITNESS_SAVE_DECL(mtx); 576 577 WITNESS_SLEEP(0, mtx); 578 mtx_lock_spin(&sched_lock); 579 DROP_GIANT_NOSWITCH(); 580 if (mtx != NULL) { 581 mtx_assert(mtx, MA_OWNED | MA_NOTRECURSED); 582 WITNESS_SAVE(mtx, mtx); 583 mtx_unlock_flags(mtx, MTX_NOSWITCH); 584 if (priority & PDROP) 585 mtx = NULL; 586 } 587 588 s = splhigh(); 589 590 if (p->p_wchan != NULL) { 591 int sig; 592 int catch; 593 594 /* 595 * The call to mawait() can override defaults specified in 596 * the original asleep(). 597 */ 598 if (priority < 0) 599 priority = p->p_asleep.as_priority; 600 if (timo < 0) 601 timo = p->p_asleep.as_timo; 602 603 /* 604 * Install timeout 605 */ 606 607 if (timo) 608 callout_reset(&p->p_slpcallout, timo, endtsleep, p); 609 610 sig = 0; 611 catch = priority & PCATCH; 612 613 if (catch) { 614 p->p_sflag |= PS_SINTR; 615 mtx_unlock_spin(&sched_lock); 616 if ((sig = CURSIG(p))) { 617 mtx_lock_spin(&sched_lock); 618 if (p->p_wchan) 619 unsleep(p); 620 p->p_stat = SRUN; 621 goto resume; 622 } 623 mtx_lock_spin(&sched_lock); 624 if (p->p_wchan == NULL) { 625 catch = 0; 626 goto resume; 627 } 628 } 629 p->p_stat = SSLEEP; 630 p->p_stats->p_ru.ru_nvcsw++; 631 mi_switch(); 632resume: 633 634 splx(s); 635 p->p_sflag &= ~PS_SINTR; 636 if (p->p_sflag & PS_TIMEOUT) { 637 p->p_sflag &= ~PS_TIMEOUT; 638 if (sig == 0) { 639#ifdef KTRACE 640 if (KTRPOINT(p, KTR_CSW)) 641 ktrcsw(p->p_tracep, 0, 0); 642#endif 643 rval = EWOULDBLOCK; 644 mtx_unlock_spin(&sched_lock); 645 goto out; 646 } 647 } else if (timo) 648 callout_stop(&p->p_slpcallout); 649 mtx_unlock_spin(&sched_lock); 650 651 if (catch && (sig != 0 || (sig = CURSIG(p)))) { 652#ifdef KTRACE 653 if (KTRPOINT(p, KTR_CSW)) 654 ktrcsw(p->p_tracep, 0, 0); 655#endif 656 if (SIGISMEMBER(p->p_sigacts->ps_sigintr, sig)) 657 rval = EINTR; 658 else 659 rval = ERESTART; 660 goto out; 661 } 662#ifdef KTRACE 663 if (KTRPOINT(p, KTR_CSW)) 664 ktrcsw(p->p_tracep, 0, 0); 665#endif 666 } else { 667 /* 668 * If as_priority is 0, mawait() has been called without an 669 * intervening asleep(). We are still effectively a NOP, 670 * but we call mi_switch() for safety. 671 */ 672 673 if (p->p_asleep.as_priority == 0) { 674 p->p_stats->p_ru.ru_nvcsw++; 675 mi_switch(); 676 } 677 mtx_unlock_spin(&sched_lock); 678 splx(s); 679 } 680 681 /* 682 * clear p_asleep.as_priority as an indication that mawait() has been 683 * called. If mawait() is called again without an intervening asleep(), 684 * mawait() is still effectively a NOP but the above mi_switch() code 685 * is triggered as a safety. 686 */ 687 p->p_asleep.as_priority = 0; 688 689out: 690 PICKUP_GIANT(); 691 if (mtx != NULL) { 692 mtx_lock(mtx); 693 WITNESS_RESTORE(mtx, mtx); 694 } 695 return (rval); 696} 697 698/* 699 * Implement timeout for msleep or asleep()/mawait() 700 * 701 * If process hasn't been awakened (wchan non-zero), 702 * set timeout flag and undo the sleep. If proc 703 * is stopped, just unsleep so it will remain stopped. 704 * MP-safe, called without the Giant mutex. 705 */ 706static void 707endtsleep(arg) 708 void *arg; 709{ 710 register struct proc *p; 711 int s; 712 713 p = (struct proc *)arg; 714 CTR4(KTR_PROC, 715 "endtsleep: proc %p (pid %d, %s), schedlock %p", 716 p, p->p_pid, p->p_comm, (void *) sched_lock.mtx_lock); 717 s = splhigh(); 718 mtx_lock_spin(&sched_lock); 719 if (p->p_wchan) { 720 if (p->p_stat == SSLEEP) 721 setrunnable(p); 722 else 723 unsleep(p); 724 p->p_sflag |= PS_TIMEOUT; 725 } 726 mtx_unlock_spin(&sched_lock); 727 splx(s); 728} 729 730/* 731 * Remove a process from its wait queue 732 */ 733void 734unsleep(p) 735 register struct proc *p; 736{ 737 int s; 738 739 s = splhigh(); 740 mtx_lock_spin(&sched_lock); 741 if (p->p_wchan) { 742 TAILQ_REMOVE(&slpque[LOOKUP(p->p_wchan)], p, p_slpq); 743 p->p_wchan = NULL; 744 } 745 mtx_unlock_spin(&sched_lock); 746 splx(s); 747} 748 749/* 750 * Make all processes sleeping on the specified identifier runnable. 751 */ 752void 753wakeup(ident) 754 register void *ident; 755{ 756 register struct slpquehead *qp; 757 register struct proc *p; 758 int s; 759 760 s = splhigh(); 761 mtx_lock_spin(&sched_lock); 762 qp = &slpque[LOOKUP(ident)]; 763restart: 764 TAILQ_FOREACH(p, qp, p_slpq) { 765 if (p->p_wchan == ident) { 766 TAILQ_REMOVE(qp, p, p_slpq); 767 p->p_wchan = NULL; 768 if (p->p_stat == SSLEEP) { 769 /* OPTIMIZED EXPANSION OF setrunnable(p); */ 770 CTR4(KTR_PROC, 771 "wakeup: proc %p (pid %d, %s), schedlock %p", 772 p, p->p_pid, p->p_comm, (void *) sched_lock.mtx_lock); 773 if (p->p_slptime > 1) 774 updatepri(p); 775 p->p_slptime = 0; 776 p->p_stat = SRUN; 777 if (p->p_sflag & PS_INMEM) { 778 setrunqueue(p); 779 maybe_resched(p); 780 } else { 781 p->p_sflag |= PS_SWAPINREQ; 782 wakeup((caddr_t)&proc0); 783 } 784 /* END INLINE EXPANSION */ 785 goto restart; 786 } 787 } 788 } 789 mtx_unlock_spin(&sched_lock); 790 splx(s); 791} 792 793/* 794 * Make a process sleeping on the specified identifier runnable. 795 * May wake more than one process if a target process is currently 796 * swapped out. 797 */ 798void 799wakeup_one(ident) 800 register void *ident; 801{ 802 register struct slpquehead *qp; 803 register struct proc *p; 804 int s; 805 806 s = splhigh(); 807 mtx_lock_spin(&sched_lock); 808 qp = &slpque[LOOKUP(ident)]; 809 810 TAILQ_FOREACH(p, qp, p_slpq) { 811 if (p->p_wchan == ident) { 812 TAILQ_REMOVE(qp, p, p_slpq); 813 p->p_wchan = NULL; 814 if (p->p_stat == SSLEEP) { 815 /* OPTIMIZED EXPANSION OF setrunnable(p); */ 816 CTR4(KTR_PROC, 817 "wakeup1: proc %p (pid %d, %s), schedlock %p", 818 p, p->p_pid, p->p_comm, (void *) sched_lock.mtx_lock); 819 if (p->p_slptime > 1) 820 updatepri(p); 821 p->p_slptime = 0; 822 p->p_stat = SRUN; 823 if (p->p_sflag & PS_INMEM) { 824 setrunqueue(p); 825 maybe_resched(p); 826 break; 827 } else { 828 p->p_sflag |= PS_SWAPINREQ; 829 wakeup((caddr_t)&proc0); 830 } 831 /* END INLINE EXPANSION */ 832 } 833 } 834 } 835 mtx_unlock_spin(&sched_lock); 836 splx(s); 837} 838 839/* 840 * The machine independent parts of mi_switch(). 841 * Must be called at splstatclock() or higher. 842 */ 843void 844mi_switch() 845{ 846 struct timeval new_switchtime; 847 register struct proc *p = curproc; /* XXX */ 848#if 0 849 register struct rlimit *rlim; 850#endif 851 int x; 852 u_int sched_nest; 853 854 /* 855 * XXX this spl is almost unnecessary. It is partly to allow for 856 * sloppy callers that don't do it (issignal() via CURSIG() is the 857 * main offender). It is partly to work around a bug in the i386 858 * cpu_switch() (the ipl is not preserved). We ran for years 859 * without it. I think there was only a interrupt latency problem. 860 * The main caller, msleep(), does an splx() a couple of instructions 861 * after calling here. The buggy caller, issignal(), usually calls 862 * here at spl0() and sometimes returns at splhigh(). The process 863 * then runs for a little too long at splhigh(). The ipl gets fixed 864 * when the process returns to user mode (or earlier). 865 * 866 * It would probably be better to always call here at spl0(). Callers 867 * are prepared to give up control to another process, so they must 868 * be prepared to be interrupted. The clock stuff here may not 869 * actually need splstatclock(). 870 */ 871 x = splstatclock(); 872 873 mtx_assert(&sched_lock, MA_OWNED | MA_NOTRECURSED); 874 875 /* 876 * Compute the amount of time during which the current 877 * process was running, and add that to its total so far. 878 */ 879 microuptime(&new_switchtime); 880 if (timevalcmp(&new_switchtime, PCPU_PTR(switchtime), <)) { 881#if 0 882 /* XXX: This doesn't play well with sched_lock right now. */ 883 printf("microuptime() went backwards (%ld.%06ld -> %ld.%06ld)\n", 884 PCPU_GET(switchtime.tv_sec), PCPU_GET(switchtime.tv_usec), 885 new_switchtime.tv_sec, new_switchtime.tv_usec); 886#endif 887 new_switchtime = PCPU_GET(switchtime); 888 } else { 889 p->p_runtime += (new_switchtime.tv_usec - PCPU_GET(switchtime.tv_usec)) + 890 (new_switchtime.tv_sec - PCPU_GET(switchtime.tv_sec)) * 891 (int64_t)1000000; 892 } 893 894#if 0 895 /* 896 * Check if the process exceeds its cpu resource allocation. 897 * If over max, kill it. 898 * 899 * XXX drop sched_lock, pickup Giant 900 */ 901 if (p->p_stat != SZOMB && p->p_limit->p_cpulimit != RLIM_INFINITY && 902 p->p_runtime > p->p_limit->p_cpulimit) { 903 rlim = &p->p_rlimit[RLIMIT_CPU]; 904 if (p->p_runtime / (rlim_t)1000000 >= rlim->rlim_max) { 905 mtx_unlock_spin(&sched_lock); 906 killproc(p, "exceeded maximum CPU limit"); 907 mtx_lock_spin(&sched_lock); 908 } else { 909 mtx_unlock_spin(&sched_lock); 910 psignal(p, SIGXCPU); 911 mtx_lock_spin(&sched_lock); 912 if (rlim->rlim_cur < rlim->rlim_max) { 913 /* XXX: we should make a private copy */ 914 rlim->rlim_cur += 5; 915 } 916 } 917 } 918#endif 919 920 /* 921 * Pick a new current process and record its start time. 922 */ 923 cnt.v_swtch++; 924 PCPU_SET(switchtime, new_switchtime); 925 CTR4(KTR_PROC, "mi_switch: old proc %p (pid %d, %s), schedlock %p", 926 p, p->p_pid, p->p_comm, (void *) sched_lock.mtx_lock); 927 sched_nest = sched_lock.mtx_recurse; 928 curproc->p_lastcpu = curproc->p_oncpu; 929 curproc->p_oncpu = NOCPU; 930 clear_resched(); 931 cpu_switch(); 932 curproc->p_oncpu = PCPU_GET(cpuid); 933 sched_lock.mtx_recurse = sched_nest; 934 sched_lock.mtx_lock = (uintptr_t)curproc; 935 CTR4(KTR_PROC, "mi_switch: new proc %p (pid %d, %s), schedlock %p", 936 p, p->p_pid, p->p_comm, (void *) sched_lock.mtx_lock); 937 if (PCPU_GET(switchtime.tv_sec) == 0) 938 microuptime(PCPU_PTR(switchtime)); 939 PCPU_SET(switchticks, ticks); 940 splx(x); 941} 942 943/* 944 * Change process state to be runnable, 945 * placing it on the run queue if it is in memory, 946 * and awakening the swapper if it isn't in memory. 947 */ 948void 949setrunnable(p) 950 register struct proc *p; 951{ 952 register int s; 953 954 s = splhigh(); 955 mtx_lock_spin(&sched_lock); 956 switch (p->p_stat) { 957 case 0: 958 case SRUN: 959 case SZOMB: 960 case SWAIT: 961 default: 962 panic("setrunnable"); 963 case SSTOP: 964 case SSLEEP: /* e.g. when sending signals */ 965 if (p->p_sflag & PS_CVWAITQ) 966 cv_waitq_remove(p); 967 else 968 unsleep(p); 969 break; 970 971 case SIDL: 972 break; 973 } 974 p->p_stat = SRUN; 975 if (p->p_sflag & PS_INMEM) 976 setrunqueue(p); 977 splx(s); 978 if (p->p_slptime > 1) 979 updatepri(p); 980 p->p_slptime = 0; 981 if ((p->p_sflag & PS_INMEM) == 0) { 982 p->p_sflag |= PS_SWAPINREQ; 983 wakeup((caddr_t)&proc0); 984 } 985 else 986 maybe_resched(p); 987 mtx_unlock_spin(&sched_lock); 988} 989 990/* 991 * Compute the priority of a process when running in user mode. 992 * Arrange to reschedule if the resulting priority is better 993 * than that of the current process. 994 */ 995void 996resetpriority(p) 997 register struct proc *p; 998{ 999 register unsigned int newpriority; 1000 1001 mtx_lock_spin(&sched_lock); 1002 if (p->p_pri.pri_class == PRI_TIMESHARE) { 1003 newpriority = PUSER + p->p_estcpu / INVERSE_ESTCPU_WEIGHT + 1004 NICE_WEIGHT * (p->p_nice - PRIO_MIN); 1005 newpriority = min(max(newpriority, PRI_MIN_TIMESHARE), 1006 PRI_MAX_TIMESHARE); 1007 p->p_pri.pri_user = newpriority; 1008 } 1009 maybe_resched(p); 1010 mtx_unlock_spin(&sched_lock); 1011} 1012 1013/* ARGSUSED */ 1014static void 1015sched_setup(dummy) 1016 void *dummy; 1017{ 1018 1019 callout_init(&schedcpu_callout, 1); 1020 callout_init(&roundrobin_callout, 0); 1021 1022 /* Kick off timeout driven events by calling first time. */ 1023 roundrobin(NULL); 1024 schedcpu(NULL); 1025} 1026 1027/* 1028 * We adjust the priority of the current process. The priority of 1029 * a process gets worse as it accumulates CPU time. The cpu usage 1030 * estimator (p_estcpu) is increased here. resetpriority() will 1031 * compute a different priority each time p_estcpu increases by 1032 * INVERSE_ESTCPU_WEIGHT 1033 * (until MAXPRI is reached). The cpu usage estimator ramps up 1034 * quite quickly when the process is running (linearly), and decays 1035 * away exponentially, at a rate which is proportionally slower when 1036 * the system is busy. The basic principle is that the system will 1037 * 90% forget that the process used a lot of CPU time in 5 * loadav 1038 * seconds. This causes the system to favor processes which haven't 1039 * run much recently, and to round-robin among other processes. 1040 */ 1041void 1042schedclock(p) 1043 struct proc *p; 1044{ 1045 1046 p->p_cpticks++; 1047 p->p_estcpu = ESTCPULIM(p->p_estcpu + 1); 1048 if ((p->p_estcpu % INVERSE_ESTCPU_WEIGHT) == 0) { 1049 resetpriority(p); 1050 if (p->p_pri.pri_level >= PUSER) 1051 p->p_pri.pri_level = p->p_pri.pri_user; 1052 } 1053} 1054 1055/* 1056 * General purpose yield system call 1057 */ 1058int 1059yield(struct proc *p, struct yield_args *uap) 1060{ 1061 int s; 1062 1063 p->p_retval[0] = 0; 1064 1065 s = splhigh(); 1066 mtx_lock_spin(&sched_lock); 1067 DROP_GIANT_NOSWITCH(); 1068 p->p_pri.pri_level = PRI_MAX_TIMESHARE; 1069 setrunqueue(p); 1070 p->p_stats->p_ru.ru_nvcsw++; 1071 mi_switch(); 1072 mtx_unlock_spin(&sched_lock); 1073 PICKUP_GIANT(); 1074 splx(s); 1075 1076 return (0); 1077} 1078