kern_synch.c revision 76827
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 76827 2001-05-19 01:28:09Z 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 KASSERT(ident == &proc0 || /* XXX: swapper */ 382 timo != 0 || /* XXX: we might still miss a wakeup */ 383 mtx_owned(&Giant) || mtx != NULL, 384 ("indefinite sleep without mutex, wmesg: \"%s\" ident: %p", 385 wmesg, ident)); 386 if (mtx_owned(&vm_mtx) && mtx != &vm_mtx) 387 panic("sleeping with vm_mtx held."); 388#ifdef KTRACE 389 if (p && KTRPOINT(p, KTR_CSW)) 390 ktrcsw(p->p_tracep, 1, 0); 391#endif 392 WITNESS_SLEEP(0, &mtx->mtx_object); 393 mtx_lock_spin(&sched_lock); 394 if (cold || panicstr) { 395 /* 396 * After a panic, or during autoconfiguration, 397 * just give interrupts a chance, then just return; 398 * don't run any other procs or panic below, 399 * in case this is the idle process and already asleep. 400 */ 401 if (mtx != NULL && priority & PDROP) 402 mtx_unlock_flags(mtx, MTX_NOSWITCH); 403 mtx_unlock_spin(&sched_lock); 404 return (0); 405 } 406 407 DROP_GIANT_NOSWITCH(); 408 409 if (mtx != NULL) { 410 mtx_assert(mtx, MA_OWNED | MA_NOTRECURSED); 411 WITNESS_SAVE(&mtx->mtx_object, mtx); 412 mtx_unlock_flags(mtx, MTX_NOSWITCH); 413 if (priority & PDROP) 414 mtx = NULL; 415 } 416 417 KASSERT(p != NULL, ("msleep1")); 418 KASSERT(ident != NULL && p->p_stat == SRUN, ("msleep")); 419 /* 420 * Process may be sitting on a slpque if asleep() was called, remove 421 * it before re-adding. 422 */ 423 if (p->p_wchan != NULL) 424 unsleep(p); 425 426 p->p_wchan = ident; 427 p->p_wmesg = wmesg; 428 p->p_slptime = 0; 429 p->p_pri.pri_level = priority & PRIMASK; 430 CTR4(KTR_PROC, "msleep: proc %p (pid %d, %s), schedlock %p", 431 p, p->p_pid, p->p_comm, (void *) sched_lock.mtx_lock); 432 TAILQ_INSERT_TAIL(&slpque[LOOKUP(ident)], p, p_slpq); 433 if (timo) 434 callout_reset(&p->p_slpcallout, timo, endtsleep, p); 435 /* 436 * We put ourselves on the sleep queue and start our timeout 437 * before calling CURSIG, as we could stop there, and a wakeup 438 * or a SIGCONT (or both) could occur while we were stopped. 439 * A SIGCONT would cause us to be marked as SSLEEP 440 * without resuming us, thus we must be ready for sleep 441 * when CURSIG is called. If the wakeup happens while we're 442 * stopped, p->p_wchan will be 0 upon return from CURSIG. 443 */ 444 if (catch) { 445 CTR4(KTR_PROC, 446 "msleep caught: proc %p (pid %d, %s), schedlock %p", 447 p, p->p_pid, p->p_comm, (void *) sched_lock.mtx_lock); 448 p->p_sflag |= PS_SINTR; 449 mtx_unlock_spin(&sched_lock); 450 if ((sig = CURSIG(p))) { 451 mtx_lock_spin(&sched_lock); 452 if (p->p_wchan) 453 unsleep(p); 454 p->p_stat = SRUN; 455 goto resume; 456 } 457 mtx_lock_spin(&sched_lock); 458 if (p->p_wchan == NULL) { 459 catch = 0; 460 goto resume; 461 } 462 } else 463 sig = 0; 464 p->p_stat = SSLEEP; 465 p->p_stats->p_ru.ru_nvcsw++; 466 mi_switch(); 467 CTR4(KTR_PROC, 468 "msleep resume: proc %p (pid %d, %s), schedlock %p", 469 p, p->p_pid, p->p_comm, (void *) sched_lock.mtx_lock); 470resume: 471 p->p_sflag &= ~PS_SINTR; 472 if (p->p_sflag & PS_TIMEOUT) { 473 p->p_sflag &= ~PS_TIMEOUT; 474 if (sig == 0) { 475#ifdef KTRACE 476 if (KTRPOINT(p, KTR_CSW)) 477 ktrcsw(p->p_tracep, 0, 0); 478#endif 479 rval = EWOULDBLOCK; 480 mtx_unlock_spin(&sched_lock); 481 goto out; 482 } 483 } else if (timo) 484 callout_stop(&p->p_slpcallout); 485 mtx_unlock_spin(&sched_lock); 486 487 if (catch && (sig != 0 || (sig = CURSIG(p)))) { 488#ifdef KTRACE 489 if (KTRPOINT(p, KTR_CSW)) 490 ktrcsw(p->p_tracep, 0, 0); 491#endif 492 PROC_LOCK(p); 493 if (SIGISMEMBER(p->p_sigacts->ps_sigintr, sig)) 494 rval = EINTR; 495 else 496 rval = ERESTART; 497 PROC_UNLOCK(p); 498 goto out; 499 } 500out: 501#ifdef KTRACE 502 if (KTRPOINT(p, KTR_CSW)) 503 ktrcsw(p->p_tracep, 0, 0); 504#endif 505 PICKUP_GIANT(); 506 if (mtx != NULL) { 507 mtx_lock(mtx); 508 WITNESS_RESTORE(&mtx->mtx_object, mtx); 509 } 510 return (rval); 511} 512 513/* 514 * asleep() - async sleep call. Place process on wait queue and return 515 * immediately without blocking. The process stays runnable until mawait() 516 * is called. If ident is NULL, remove process from wait queue if it is still 517 * on one. 518 * 519 * Only the most recent sleep condition is effective when making successive 520 * calls to asleep() or when calling msleep(). 521 * 522 * The timeout, if any, is not initiated until mawait() is called. The sleep 523 * priority, signal, and timeout is specified in the asleep() call but may be 524 * overriden in the mawait() call. 525 * 526 * <<<<<<<< EXPERIMENTAL, UNTESTED >>>>>>>>>> 527 */ 528 529int 530asleep(void *ident, int priority, const char *wmesg, int timo) 531{ 532 struct proc *p = curproc; 533 int s; 534 535 /* 536 * obtain sched_lock while manipulating sleep structures and slpque. 537 * 538 * Remove preexisting wait condition (if any) and place process 539 * on appropriate slpque, but do not put process to sleep. 540 */ 541 542 s = splhigh(); 543 mtx_lock_spin(&sched_lock); 544 545 if (p->p_wchan != NULL) 546 unsleep(p); 547 548 if (ident) { 549 p->p_wchan = ident; 550 p->p_wmesg = wmesg; 551 p->p_slptime = 0; 552 p->p_asleep.as_priority = priority; 553 p->p_asleep.as_timo = timo; 554 TAILQ_INSERT_TAIL(&slpque[LOOKUP(ident)], p, p_slpq); 555 } 556 557 mtx_unlock_spin(&sched_lock); 558 splx(s); 559 560 return(0); 561} 562 563/* 564 * mawait() - wait for async condition to occur. The process blocks until 565 * wakeup() is called on the most recent asleep() address. If wakeup is called 566 * prior to mawait(), mawait() winds up being a NOP. 567 * 568 * If mawait() is called more then once (without an intervening asleep() call), 569 * mawait() is still effectively a NOP but it calls mi_switch() to give other 570 * processes some cpu before returning. The process is left runnable. 571 * 572 * <<<<<<<< EXPERIMENTAL, UNTESTED >>>>>>>>>> 573 */ 574 575int 576mawait(struct mtx *mtx, int priority, int timo) 577{ 578 struct proc *p = curproc; 579 int rval = 0; 580 int s; 581 WITNESS_SAVE_DECL(mtx); 582 583 WITNESS_SLEEP(0, &mtx->mtx_object); 584 mtx_lock_spin(&sched_lock); 585 DROP_GIANT_NOSWITCH(); 586 if (mtx != NULL) { 587 mtx_assert(mtx, MA_OWNED | MA_NOTRECURSED); 588 WITNESS_SAVE(&mtx->mtx_object, mtx); 589 mtx_unlock_flags(mtx, MTX_NOSWITCH); 590 if (priority & PDROP) 591 mtx = NULL; 592 } 593 594 s = splhigh(); 595 596 if (p->p_wchan != NULL) { 597 int sig; 598 int catch; 599 600 /* 601 * The call to mawait() can override defaults specified in 602 * the original asleep(). 603 */ 604 if (priority < 0) 605 priority = p->p_asleep.as_priority; 606 if (timo < 0) 607 timo = p->p_asleep.as_timo; 608 609 /* 610 * Install timeout 611 */ 612 613 if (timo) 614 callout_reset(&p->p_slpcallout, timo, endtsleep, p); 615 616 sig = 0; 617 catch = priority & PCATCH; 618 619 if (catch) { 620 p->p_sflag |= PS_SINTR; 621 mtx_unlock_spin(&sched_lock); 622 if ((sig = CURSIG(p))) { 623 mtx_lock_spin(&sched_lock); 624 if (p->p_wchan) 625 unsleep(p); 626 p->p_stat = SRUN; 627 goto resume; 628 } 629 mtx_lock_spin(&sched_lock); 630 if (p->p_wchan == NULL) { 631 catch = 0; 632 goto resume; 633 } 634 } 635 p->p_stat = SSLEEP; 636 p->p_stats->p_ru.ru_nvcsw++; 637 mi_switch(); 638resume: 639 640 splx(s); 641 p->p_sflag &= ~PS_SINTR; 642 if (p->p_sflag & PS_TIMEOUT) { 643 p->p_sflag &= ~PS_TIMEOUT; 644 if (sig == 0) { 645#ifdef KTRACE 646 if (KTRPOINT(p, KTR_CSW)) 647 ktrcsw(p->p_tracep, 0, 0); 648#endif 649 rval = EWOULDBLOCK; 650 mtx_unlock_spin(&sched_lock); 651 goto out; 652 } 653 } else if (timo) 654 callout_stop(&p->p_slpcallout); 655 mtx_unlock_spin(&sched_lock); 656 657 if (catch && (sig != 0 || (sig = CURSIG(p)))) { 658#ifdef KTRACE 659 if (KTRPOINT(p, KTR_CSW)) 660 ktrcsw(p->p_tracep, 0, 0); 661#endif 662 PROC_LOCK(p); 663 if (SIGISMEMBER(p->p_sigacts->ps_sigintr, sig)) 664 rval = EINTR; 665 else 666 rval = ERESTART; 667 PROC_UNLOCK(p); 668 goto out; 669 } 670#ifdef KTRACE 671 if (KTRPOINT(p, KTR_CSW)) 672 ktrcsw(p->p_tracep, 0, 0); 673#endif 674 } else { 675 /* 676 * If as_priority is 0, mawait() has been called without an 677 * intervening asleep(). We are still effectively a NOP, 678 * but we call mi_switch() for safety. 679 */ 680 681 if (p->p_asleep.as_priority == 0) { 682 p->p_stats->p_ru.ru_nvcsw++; 683 mi_switch(); 684 } 685 mtx_unlock_spin(&sched_lock); 686 splx(s); 687 } 688 689 /* 690 * clear p_asleep.as_priority as an indication that mawait() has been 691 * called. If mawait() is called again without an intervening asleep(), 692 * mawait() is still effectively a NOP but the above mi_switch() code 693 * is triggered as a safety. 694 */ 695 p->p_asleep.as_priority = 0; 696 697out: 698 PICKUP_GIANT(); 699 if (mtx != NULL) { 700 mtx_lock(mtx); 701 WITNESS_RESTORE(&mtx->mtx_object, mtx); 702 } 703 return (rval); 704} 705 706/* 707 * Implement timeout for msleep or asleep()/mawait() 708 * 709 * If process hasn't been awakened (wchan non-zero), 710 * set timeout flag and undo the sleep. If proc 711 * is stopped, just unsleep so it will remain stopped. 712 * MP-safe, called without the Giant mutex. 713 */ 714static void 715endtsleep(arg) 716 void *arg; 717{ 718 register struct proc *p; 719 int s; 720 721 p = (struct proc *)arg; 722 CTR4(KTR_PROC, 723 "endtsleep: proc %p (pid %d, %s), schedlock %p", 724 p, p->p_pid, p->p_comm, (void *) sched_lock.mtx_lock); 725 s = splhigh(); 726 mtx_lock_spin(&sched_lock); 727 if (p->p_wchan) { 728 if (p->p_stat == SSLEEP) 729 setrunnable(p); 730 else 731 unsleep(p); 732 p->p_sflag |= PS_TIMEOUT; 733 } 734 mtx_unlock_spin(&sched_lock); 735 splx(s); 736} 737 738/* 739 * Remove a process from its wait queue 740 */ 741void 742unsleep(p) 743 register struct proc *p; 744{ 745 int s; 746 747 s = splhigh(); 748 mtx_lock_spin(&sched_lock); 749 if (p->p_wchan) { 750 TAILQ_REMOVE(&slpque[LOOKUP(p->p_wchan)], p, p_slpq); 751 p->p_wchan = NULL; 752 } 753 mtx_unlock_spin(&sched_lock); 754 splx(s); 755} 756 757/* 758 * Make all processes sleeping on the specified identifier runnable. 759 */ 760void 761wakeup(ident) 762 register void *ident; 763{ 764 register struct slpquehead *qp; 765 register struct proc *p; 766 int s; 767 768 s = splhigh(); 769 mtx_lock_spin(&sched_lock); 770 qp = &slpque[LOOKUP(ident)]; 771restart: 772 TAILQ_FOREACH(p, qp, p_slpq) { 773 if (p->p_wchan == ident) { 774 TAILQ_REMOVE(qp, p, p_slpq); 775 p->p_wchan = NULL; 776 if (p->p_stat == SSLEEP) { 777 /* OPTIMIZED EXPANSION OF setrunnable(p); */ 778 CTR4(KTR_PROC, 779 "wakeup: proc %p (pid %d, %s), schedlock %p", 780 p, p->p_pid, p->p_comm, (void *) sched_lock.mtx_lock); 781 if (p->p_slptime > 1) 782 updatepri(p); 783 p->p_slptime = 0; 784 p->p_stat = SRUN; 785 if (p->p_sflag & PS_INMEM) { 786 setrunqueue(p); 787 maybe_resched(p); 788 } else { 789 p->p_sflag |= PS_SWAPINREQ; 790 wakeup((caddr_t)&proc0); 791 } 792 /* END INLINE EXPANSION */ 793 goto restart; 794 } 795 } 796 } 797 mtx_unlock_spin(&sched_lock); 798 splx(s); 799} 800 801/* 802 * Make a process sleeping on the specified identifier runnable. 803 * May wake more than one process if a target process is currently 804 * swapped out. 805 */ 806void 807wakeup_one(ident) 808 register void *ident; 809{ 810 register struct slpquehead *qp; 811 register struct proc *p; 812 int s; 813 814 s = splhigh(); 815 mtx_lock_spin(&sched_lock); 816 qp = &slpque[LOOKUP(ident)]; 817 818 TAILQ_FOREACH(p, qp, p_slpq) { 819 if (p->p_wchan == ident) { 820 TAILQ_REMOVE(qp, p, p_slpq); 821 p->p_wchan = NULL; 822 if (p->p_stat == SSLEEP) { 823 /* OPTIMIZED EXPANSION OF setrunnable(p); */ 824 CTR4(KTR_PROC, 825 "wakeup1: proc %p (pid %d, %s), schedlock %p", 826 p, p->p_pid, p->p_comm, (void *) sched_lock.mtx_lock); 827 if (p->p_slptime > 1) 828 updatepri(p); 829 p->p_slptime = 0; 830 p->p_stat = SRUN; 831 if (p->p_sflag & PS_INMEM) { 832 setrunqueue(p); 833 maybe_resched(p); 834 break; 835 } else { 836 p->p_sflag |= PS_SWAPINREQ; 837 wakeup((caddr_t)&proc0); 838 } 839 /* END INLINE EXPANSION */ 840 } 841 } 842 } 843 mtx_unlock_spin(&sched_lock); 844 splx(s); 845} 846 847/* 848 * The machine independent parts of mi_switch(). 849 * Must be called at splstatclock() or higher. 850 */ 851void 852mi_switch() 853{ 854 struct timeval new_switchtime; 855 register struct proc *p = curproc; /* XXX */ 856#if 0 857 register struct rlimit *rlim; 858#endif 859 u_int sched_nest; 860 861 mtx_assert(&sched_lock, MA_OWNED | MA_NOTRECURSED); 862 863 /* 864 * Compute the amount of time during which the current 865 * process was running, and add that to its total so far. 866 */ 867 microuptime(&new_switchtime); 868 if (timevalcmp(&new_switchtime, PCPU_PTR(switchtime), <)) { 869#if 0 870 /* XXX: This doesn't play well with sched_lock right now. */ 871 printf("microuptime() went backwards (%ld.%06ld -> %ld.%06ld)\n", 872 PCPU_GET(switchtime.tv_sec), PCPU_GET(switchtime.tv_usec), 873 new_switchtime.tv_sec, new_switchtime.tv_usec); 874#endif 875 new_switchtime = PCPU_GET(switchtime); 876 } else { 877 p->p_runtime += (new_switchtime.tv_usec - PCPU_GET(switchtime.tv_usec)) + 878 (new_switchtime.tv_sec - PCPU_GET(switchtime.tv_sec)) * 879 (int64_t)1000000; 880 } 881 882#if 0 883 /* 884 * Check if the process exceeds its cpu resource allocation. 885 * If over max, kill it. 886 * 887 * XXX drop sched_lock, pickup Giant 888 */ 889 if (p->p_stat != SZOMB && p->p_limit->p_cpulimit != RLIM_INFINITY && 890 p->p_runtime > p->p_limit->p_cpulimit) { 891 rlim = &p->p_rlimit[RLIMIT_CPU]; 892 if (p->p_runtime / (rlim_t)1000000 >= rlim->rlim_max) { 893 mtx_unlock_spin(&sched_lock); 894 PROC_LOCK(p); 895 killproc(p, "exceeded maximum CPU limit"); 896 mtx_lock_spin(&sched_lock); 897 PROC_UNLOCK_NOSWITCH(p); 898 } else { 899 mtx_unlock_spin(&sched_lock); 900 PROC_LOCK(p); 901 psignal(p, SIGXCPU); 902 mtx_lock_spin(&sched_lock); 903 PROC_UNLOCK_NOSWITCH(p); 904 if (rlim->rlim_cur < rlim->rlim_max) { 905 /* XXX: we should make a private copy */ 906 rlim->rlim_cur += 5; 907 } 908 } 909 } 910#endif 911 912 /* 913 * Pick a new current process and record its start time. 914 */ 915 cnt.v_swtch++; 916 PCPU_SET(switchtime, new_switchtime); 917 CTR4(KTR_PROC, "mi_switch: old proc %p (pid %d, %s), schedlock %p", 918 p, p->p_pid, p->p_comm, (void *) sched_lock.mtx_lock); 919 sched_nest = sched_lock.mtx_recurse; 920 curproc->p_lastcpu = curproc->p_oncpu; 921 curproc->p_oncpu = NOCPU; 922 clear_resched(curproc); 923 cpu_switch(); 924 curproc->p_oncpu = PCPU_GET(cpuid); 925 sched_lock.mtx_recurse = sched_nest; 926 sched_lock.mtx_lock = (uintptr_t)curproc; 927 CTR4(KTR_PROC, "mi_switch: new proc %p (pid %d, %s), schedlock %p", 928 p, p->p_pid, p->p_comm, (void *) sched_lock.mtx_lock); 929 if (PCPU_GET(switchtime.tv_sec) == 0) 930 microuptime(PCPU_PTR(switchtime)); 931 PCPU_SET(switchticks, ticks); 932} 933 934/* 935 * Change process state to be runnable, 936 * placing it on the run queue if it is in memory, 937 * and awakening the swapper if it isn't in memory. 938 */ 939void 940setrunnable(p) 941 register struct proc *p; 942{ 943 register int s; 944 945 s = splhigh(); 946 mtx_lock_spin(&sched_lock); 947 switch (p->p_stat) { 948 case 0: 949 case SRUN: 950 case SZOMB: 951 case SWAIT: 952 default: 953 panic("setrunnable"); 954 case SSTOP: 955 case SSLEEP: /* e.g. when sending signals */ 956 if (p->p_sflag & PS_CVWAITQ) 957 cv_waitq_remove(p); 958 else 959 unsleep(p); 960 break; 961 962 case SIDL: 963 break; 964 } 965 p->p_stat = SRUN; 966 if (p->p_sflag & PS_INMEM) 967 setrunqueue(p); 968 splx(s); 969 if (p->p_slptime > 1) 970 updatepri(p); 971 p->p_slptime = 0; 972 if ((p->p_sflag & PS_INMEM) == 0) { 973 p->p_sflag |= PS_SWAPINREQ; 974 wakeup((caddr_t)&proc0); 975 } 976 else 977 maybe_resched(p); 978 mtx_unlock_spin(&sched_lock); 979} 980 981/* 982 * Compute the priority of a process when running in user mode. 983 * Arrange to reschedule if the resulting priority is better 984 * than that of the current process. 985 */ 986void 987resetpriority(p) 988 register struct proc *p; 989{ 990 register unsigned int newpriority; 991 992 mtx_lock_spin(&sched_lock); 993 if (p->p_pri.pri_class == PRI_TIMESHARE) { 994 newpriority = PUSER + p->p_estcpu / INVERSE_ESTCPU_WEIGHT + 995 NICE_WEIGHT * (p->p_nice - PRIO_MIN); 996 newpriority = min(max(newpriority, PRI_MIN_TIMESHARE), 997 PRI_MAX_TIMESHARE); 998 p->p_pri.pri_user = newpriority; 999 } 1000 maybe_resched(p); 1001 mtx_unlock_spin(&sched_lock); 1002} 1003 1004/* ARGSUSED */ 1005static void 1006sched_setup(dummy) 1007 void *dummy; 1008{ 1009 1010 callout_init(&schedcpu_callout, 1); 1011 callout_init(&roundrobin_callout, 0); 1012 1013 /* Kick off timeout driven events by calling first time. */ 1014 roundrobin(NULL); 1015 schedcpu(NULL); 1016} 1017 1018/* 1019 * We adjust the priority of the current process. The priority of 1020 * a process gets worse as it accumulates CPU time. The cpu usage 1021 * estimator (p_estcpu) is increased here. resetpriority() will 1022 * compute a different priority each time p_estcpu increases by 1023 * INVERSE_ESTCPU_WEIGHT 1024 * (until MAXPRI is reached). The cpu usage estimator ramps up 1025 * quite quickly when the process is running (linearly), and decays 1026 * away exponentially, at a rate which is proportionally slower when 1027 * the system is busy. The basic principle is that the system will 1028 * 90% forget that the process used a lot of CPU time in 5 * loadav 1029 * seconds. This causes the system to favor processes which haven't 1030 * run much recently, and to round-robin among other processes. 1031 */ 1032void 1033schedclock(p) 1034 struct proc *p; 1035{ 1036 1037 p->p_cpticks++; 1038 p->p_estcpu = ESTCPULIM(p->p_estcpu + 1); 1039 if ((p->p_estcpu % INVERSE_ESTCPU_WEIGHT) == 0) { 1040 resetpriority(p); 1041 if (p->p_pri.pri_level >= PUSER) 1042 p->p_pri.pri_level = p->p_pri.pri_user; 1043 } 1044} 1045 1046/* 1047 * General purpose yield system call 1048 */ 1049int 1050yield(struct proc *p, struct yield_args *uap) 1051{ 1052 int s; 1053 1054 p->p_retval[0] = 0; 1055 1056 s = splhigh(); 1057 mtx_lock_spin(&sched_lock); 1058 DROP_GIANT_NOSWITCH(); 1059 p->p_pri.pri_level = PRI_MAX_TIMESHARE; 1060 setrunqueue(p); 1061 p->p_stats->p_ru.ru_nvcsw++; 1062 mi_switch(); 1063 mtx_unlock_spin(&sched_lock); 1064 PICKUP_GIANT(); 1065 splx(s); 1066 1067 return (0); 1068} 1069