kern_synch.c revision 72828
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 72828 2001-02-22 00:32:13Z 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 if (p->p_pri.pri_level < curproc->p_pri.pri_level) 111 need_resched(); 112} 113 114int 115roundrobin_interval(void) 116{ 117 return (sched_quantum); 118} 119 120/* 121 * Force switch among equal priority processes every 100ms. 122 */ 123/* ARGSUSED */ 124static void 125roundrobin(arg) 126 void *arg; 127{ 128 129 mtx_lock_spin(&sched_lock); 130 need_resched(); 131 mtx_unlock_spin(&sched_lock); 132#ifdef SMP 133 forward_roundrobin(); 134#endif 135 136 callout_reset(&roundrobin_callout, sched_quantum, roundrobin, NULL); 137} 138 139/* 140 * Constants for digital decay and forget: 141 * 90% of (p_estcpu) usage in 5 * loadav time 142 * 95% of (p_pctcpu) usage in 60 seconds (load insensitive) 143 * Note that, as ps(1) mentions, this can let percentages 144 * total over 100% (I've seen 137.9% for 3 processes). 145 * 146 * Note that schedclock() updates p_estcpu and p_cpticks asynchronously. 147 * 148 * We wish to decay away 90% of p_estcpu in (5 * loadavg) seconds. 149 * That is, the system wants to compute a value of decay such 150 * that the following for loop: 151 * for (i = 0; i < (5 * loadavg); i++) 152 * p_estcpu *= decay; 153 * will compute 154 * p_estcpu *= 0.1; 155 * for all values of loadavg: 156 * 157 * Mathematically this loop can be expressed by saying: 158 * decay ** (5 * loadavg) ~= .1 159 * 160 * The system computes decay as: 161 * decay = (2 * loadavg) / (2 * loadavg + 1) 162 * 163 * We wish to prove that the system's computation of decay 164 * will always fulfill the equation: 165 * decay ** (5 * loadavg) ~= .1 166 * 167 * If we compute b as: 168 * b = 2 * loadavg 169 * then 170 * decay = b / (b + 1) 171 * 172 * We now need to prove two things: 173 * 1) Given factor ** (5 * loadavg) ~= .1, prove factor == b/(b+1) 174 * 2) Given b/(b+1) ** power ~= .1, prove power == (5 * loadavg) 175 * 176 * Facts: 177 * For x close to zero, exp(x) =~ 1 + x, since 178 * exp(x) = 0! + x**1/1! + x**2/2! + ... . 179 * therefore exp(-1/b) =~ 1 - (1/b) = (b-1)/b. 180 * For x close to zero, ln(1+x) =~ x, since 181 * ln(1+x) = x - x**2/2 + x**3/3 - ... -1 < x < 1 182 * therefore ln(b/(b+1)) = ln(1 - 1/(b+1)) =~ -1/(b+1). 183 * ln(.1) =~ -2.30 184 * 185 * Proof of (1): 186 * Solve (factor)**(power) =~ .1 given power (5*loadav): 187 * solving for factor, 188 * ln(factor) =~ (-2.30/5*loadav), or 189 * factor =~ exp(-1/((5/2.30)*loadav)) =~ exp(-1/(2*loadav)) = 190 * exp(-1/b) =~ (b-1)/b =~ b/(b+1). QED 191 * 192 * Proof of (2): 193 * Solve (factor)**(power) =~ .1 given factor == (b/(b+1)): 194 * solving for power, 195 * power*ln(b/(b+1)) =~ -2.30, or 196 * power =~ 2.3 * (b + 1) = 4.6*loadav + 2.3 =~ 5*loadav. QED 197 * 198 * Actual power values for the implemented algorithm are as follows: 199 * loadav: 1 2 3 4 200 * power: 5.68 10.32 14.94 19.55 201 */ 202 203/* calculations for digital decay to forget 90% of usage in 5*loadav sec */ 204#define loadfactor(loadav) (2 * (loadav)) 205#define decay_cpu(loadfac, cpu) (((loadfac) * (cpu)) / ((loadfac) + FSCALE)) 206 207/* decay 95% of `p_pctcpu' in 60 seconds; see CCPU_SHIFT before changing */ 208static fixpt_t ccpu = 0.95122942450071400909 * FSCALE; /* exp(-1/20) */ 209SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, ""); 210 211/* kernel uses `FSCALE', userland (SHOULD) use kern.fscale */ 212static int fscale __unused = FSCALE; 213SYSCTL_INT(_kern, OID_AUTO, fscale, CTLFLAG_RD, 0, FSCALE, ""); 214 215/* 216 * If `ccpu' is not equal to `exp(-1/20)' and you still want to use the 217 * faster/more-accurate formula, you'll have to estimate CCPU_SHIFT below 218 * and possibly adjust FSHIFT in "param.h" so that (FSHIFT >= CCPU_SHIFT). 219 * 220 * To estimate CCPU_SHIFT for exp(-1/20), the following formula was used: 221 * 1 - exp(-1/20) ~= 0.0487 ~= 0.0488 == 1 (fixed pt, *11* bits). 222 * 223 * If you don't want to bother with the faster/more-accurate formula, you 224 * can set CCPU_SHIFT to (FSHIFT + 1) which will use a slower/less-accurate 225 * (more general) method of calculating the %age of CPU used by a process. 226 */ 227#define CCPU_SHIFT 11 228 229/* 230 * Recompute process priorities, every hz ticks. 231 * MP-safe, called without the Giant mutex. 232 */ 233/* ARGSUSED */ 234static void 235schedcpu(arg) 236 void *arg; 237{ 238 register fixpt_t loadfac = loadfactor(averunnable.ldavg[0]); 239 register struct proc *p; 240 register int realstathz, s; 241 242 realstathz = stathz ? stathz : hz; 243 ALLPROC_LOCK(AP_SHARED); 244 LIST_FOREACH(p, &allproc, p_list) { 245 /* 246 * Increment time in/out of memory and sleep time 247 * (if sleeping). We ignore overflow; with 16-bit int's 248 * (remember them?) overflow takes 45 days. 249 if (p->p_stat == SWAIT) 250 continue; 251 */ 252 mtx_lock_spin(&sched_lock); 253 p->p_swtime++; 254 if (p->p_stat == SSLEEP || p->p_stat == SSTOP) 255 p->p_slptime++; 256 p->p_pctcpu = (p->p_pctcpu * ccpu) >> FSHIFT; 257 /* 258 * If the process has slept the entire second, 259 * stop recalculating its priority until it wakes up. 260 */ 261 if (p->p_slptime > 1) { 262 mtx_unlock_spin(&sched_lock); 263 continue; 264 } 265 266 /* 267 * prevent state changes and protect run queue 268 */ 269 s = splhigh(); 270 271 /* 272 * p_pctcpu is only for ps. 273 */ 274#if (FSHIFT >= CCPU_SHIFT) 275 p->p_pctcpu += (realstathz == 100)? 276 ((fixpt_t) p->p_cpticks) << (FSHIFT - CCPU_SHIFT): 277 100 * (((fixpt_t) p->p_cpticks) 278 << (FSHIFT - CCPU_SHIFT)) / realstathz; 279#else 280 p->p_pctcpu += ((FSCALE - ccpu) * 281 (p->p_cpticks * FSCALE / realstathz)) >> FSHIFT; 282#endif 283 p->p_cpticks = 0; 284 p->p_estcpu = decay_cpu(loadfac, p->p_estcpu); 285 resetpriority(p); 286 if (p->p_pri.pri_level >= PUSER) { 287 if ((p != curproc) && 288#ifdef SMP 289 p->p_oncpu == NOCPU && /* idle */ 290#endif 291 p->p_stat == SRUN && 292 (p->p_sflag & PS_INMEM) && 293 (p->p_pri.pri_level / RQ_PPQ) != 294 (p->p_pri.pri_user / RQ_PPQ)) { 295 remrunqueue(p); 296 p->p_pri.pri_level = p->p_pri.pri_user; 297 setrunqueue(p); 298 } else 299 p->p_pri.pri_level = p->p_pri.pri_user; 300 } 301 mtx_unlock_spin(&sched_lock); 302 splx(s); 303 } 304 ALLPROC_LOCK(AP_RELEASE); 305 vmmeter(); 306 wakeup((caddr_t)&lbolt); 307 callout_reset(&schedcpu_callout, hz, schedcpu, NULL); 308} 309 310/* 311 * Recalculate the priority of a process after it has slept for a while. 312 * For all load averages >= 1 and max p_estcpu of 255, sleeping for at 313 * least six times the loadfactor will decay p_estcpu to zero. 314 */ 315void 316updatepri(p) 317 register struct proc *p; 318{ 319 register unsigned int newcpu = p->p_estcpu; 320 register fixpt_t loadfac = loadfactor(averunnable.ldavg[0]); 321 322 if (p->p_slptime > 5 * loadfac) 323 p->p_estcpu = 0; 324 else { 325 p->p_slptime--; /* the first time was done in schedcpu */ 326 while (newcpu && --p->p_slptime) 327 newcpu = decay_cpu(loadfac, newcpu); 328 p->p_estcpu = newcpu; 329 } 330 resetpriority(p); 331} 332 333/* 334 * We're only looking at 7 bits of the address; everything is 335 * aligned to 4, lots of things are aligned to greater powers 336 * of 2. Shift right by 8, i.e. drop the bottom 256 worth. 337 */ 338#define TABLESIZE 128 339static TAILQ_HEAD(slpquehead, proc) slpque[TABLESIZE]; 340#define LOOKUP(x) (((intptr_t)(x) >> 8) & (TABLESIZE - 1)) 341 342void 343sleepinit(void) 344{ 345 int i; 346 347 sched_quantum = hz/10; 348 hogticks = 2 * sched_quantum; 349 for (i = 0; i < TABLESIZE; i++) 350 TAILQ_INIT(&slpque[i]); 351} 352 353/* 354 * General sleep call. Suspends the current process until a wakeup is 355 * performed on the specified identifier. The process will then be made 356 * runnable with the specified priority. Sleeps at most timo/hz seconds 357 * (0 means no timeout). If pri includes PCATCH flag, signals are checked 358 * before and after sleeping, else signals are not checked. Returns 0 if 359 * awakened, EWOULDBLOCK if the timeout expires. If PCATCH is set and a 360 * signal needs to be delivered, ERESTART is returned if the current system 361 * call should be restarted if possible, and EINTR is returned if the system 362 * call should be interrupted by the signal (return EINTR). 363 * 364 * The mutex argument is exited before the caller is suspended, and 365 * entered before msleep returns. If priority includes the PDROP 366 * flag the mutex is not entered before returning. 367 */ 368int 369msleep(ident, mtx, priority, wmesg, timo) 370 void *ident; 371 struct mtx *mtx; 372 int priority, timo; 373 const char *wmesg; 374{ 375 struct proc *p = curproc; 376 int s, sig, catch = priority & PCATCH; 377 int rval = 0; 378 WITNESS_SAVE_DECL(mtx); 379 380#ifdef KTRACE 381 if (p && KTRPOINT(p, KTR_CSW)) 382 ktrcsw(p->p_tracep, 1, 0); 383#endif 384 WITNESS_SLEEP(0, mtx); 385 mtx_lock_spin(&sched_lock); 386 s = splhigh(); 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 splx(s); 398 return (0); 399 } 400 401 DROP_GIANT_NOSWITCH(); 402 403 if (mtx != NULL) { 404 mtx_assert(mtx, MA_OWNED | MA_NOTRECURSED); 405 WITNESS_SAVE(mtx, mtx); 406 mtx_unlock_flags(mtx, MTX_NOSWITCH); 407 if (priority & PDROP) 408 mtx = NULL; 409 } 410 411 KASSERT(p != NULL, ("msleep1")); 412 KASSERT(ident != NULL && p->p_stat == SRUN, ("msleep")); 413 /* 414 * Process may be sitting on a slpque if asleep() was called, remove 415 * it before re-adding. 416 */ 417 if (p->p_wchan != NULL) 418 unsleep(p); 419 420 p->p_wchan = ident; 421 p->p_wmesg = wmesg; 422 p->p_slptime = 0; 423 p->p_pri.pri_level = priority & PRIMASK; 424 CTR4(KTR_PROC, "msleep: proc %p (pid %d, %s), schedlock %p", 425 p, p->p_pid, p->p_comm, (void *) sched_lock.mtx_lock); 426 TAILQ_INSERT_TAIL(&slpque[LOOKUP(ident)], p, p_slpq); 427 if (timo) 428 callout_reset(&p->p_slpcallout, timo, endtsleep, p); 429 /* 430 * We put ourselves on the sleep queue and start our timeout 431 * before calling CURSIG, as we could stop there, and a wakeup 432 * or a SIGCONT (or both) could occur while we were stopped. 433 * A SIGCONT would cause us to be marked as SSLEEP 434 * without resuming us, thus we must be ready for sleep 435 * when CURSIG is called. If the wakeup happens while we're 436 * stopped, p->p_wchan will be 0 upon return from CURSIG. 437 */ 438 if (catch) { 439 CTR4(KTR_PROC, 440 "msleep caught: proc %p (pid %d, %s), schedlock %p", 441 p, p->p_pid, p->p_comm, (void *) sched_lock.mtx_lock); 442 p->p_sflag |= PS_SINTR; 443 mtx_unlock_spin(&sched_lock); 444 if ((sig = CURSIG(p))) { 445 mtx_lock_spin(&sched_lock); 446 if (p->p_wchan) 447 unsleep(p); 448 p->p_stat = SRUN; 449 goto resume; 450 } 451 mtx_lock_spin(&sched_lock); 452 if (p->p_wchan == NULL) { 453 catch = 0; 454 goto resume; 455 } 456 } else 457 sig = 0; 458 p->p_stat = SSLEEP; 459 p->p_stats->p_ru.ru_nvcsw++; 460 mi_switch(); 461 CTR4(KTR_PROC, 462 "msleep resume: proc %p (pid %d, %s), schedlock %p", 463 p, p->p_pid, p->p_comm, (void *) sched_lock.mtx_lock); 464resume: 465 splx(s); 466 p->p_sflag &= ~PS_SINTR; 467 if (p->p_sflag & PS_TIMEOUT) { 468 p->p_sflag &= ~PS_TIMEOUT; 469 if (sig == 0) { 470#ifdef KTRACE 471 if (KTRPOINT(p, KTR_CSW)) 472 ktrcsw(p->p_tracep, 0, 0); 473#endif 474 rval = EWOULDBLOCK; 475 mtx_unlock_spin(&sched_lock); 476 goto out; 477 } 478 } else if (timo) 479 callout_stop(&p->p_slpcallout); 480 mtx_unlock_spin(&sched_lock); 481 482 if (catch && (sig != 0 || (sig = CURSIG(p)))) { 483#ifdef KTRACE 484 if (KTRPOINT(p, KTR_CSW)) 485 ktrcsw(p->p_tracep, 0, 0); 486#endif 487 if (SIGISMEMBER(p->p_sigacts->ps_sigintr, sig)) 488 rval = EINTR; 489 else 490 rval = ERESTART; 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); 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); 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); 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 if (SIGISMEMBER(p->p_sigacts->ps_sigintr, sig)) 656 rval = EINTR; 657 else 658 rval = ERESTART; 659 goto out; 660 } 661#ifdef KTRACE 662 if (KTRPOINT(p, KTR_CSW)) 663 ktrcsw(p->p_tracep, 0, 0); 664#endif 665 } else { 666 /* 667 * If as_priority is 0, mawait() has been called without an 668 * intervening asleep(). We are still effectively a NOP, 669 * but we call mi_switch() for safety. 670 */ 671 672 if (p->p_asleep.as_priority == 0) { 673 p->p_stats->p_ru.ru_nvcsw++; 674 mi_switch(); 675 } 676 mtx_unlock_spin(&sched_lock); 677 splx(s); 678 } 679 680 /* 681 * clear p_asleep.as_priority as an indication that mawait() has been 682 * called. If mawait() is called again without an intervening asleep(), 683 * mawait() is still effectively a NOP but the above mi_switch() code 684 * is triggered as a safety. 685 */ 686 p->p_asleep.as_priority = 0; 687 688out: 689 PICKUP_GIANT(); 690 if (mtx != NULL) { 691 mtx_lock(mtx); 692 WITNESS_RESTORE(mtx, mtx); 693 } 694 return (rval); 695} 696 697/* 698 * Implement timeout for msleep or asleep()/mawait() 699 * 700 * If process hasn't been awakened (wchan non-zero), 701 * set timeout flag and undo the sleep. If proc 702 * is stopped, just unsleep so it will remain stopped. 703 * MP-safe, called without the Giant mutex. 704 */ 705static void 706endtsleep(arg) 707 void *arg; 708{ 709 register struct proc *p; 710 int s; 711 712 p = (struct proc *)arg; 713 CTR4(KTR_PROC, 714 "endtsleep: proc %p (pid %d, %s), schedlock %p", 715 p, p->p_pid, p->p_comm, (void *) sched_lock.mtx_lock); 716 s = splhigh(); 717 mtx_lock_spin(&sched_lock); 718 if (p->p_wchan) { 719 if (p->p_stat == SSLEEP) 720 setrunnable(p); 721 else 722 unsleep(p); 723 p->p_sflag |= PS_TIMEOUT; 724 } 725 mtx_unlock_spin(&sched_lock); 726 splx(s); 727} 728 729/* 730 * Remove a process from its wait queue 731 */ 732void 733unsleep(p) 734 register struct proc *p; 735{ 736 int s; 737 738 s = splhigh(); 739 mtx_lock_spin(&sched_lock); 740 if (p->p_wchan) { 741 TAILQ_REMOVE(&slpque[LOOKUP(p->p_wchan)], p, p_slpq); 742 p->p_wchan = NULL; 743 } 744 mtx_unlock_spin(&sched_lock); 745 splx(s); 746} 747 748/* 749 * Make all processes sleeping on the specified identifier runnable. 750 */ 751void 752wakeup(ident) 753 register void *ident; 754{ 755 register struct slpquehead *qp; 756 register struct proc *p; 757 int s; 758 759 s = splhigh(); 760 mtx_lock_spin(&sched_lock); 761 qp = &slpque[LOOKUP(ident)]; 762restart: 763 TAILQ_FOREACH(p, qp, p_slpq) { 764 if (p->p_wchan == ident) { 765 TAILQ_REMOVE(qp, p, p_slpq); 766 p->p_wchan = NULL; 767 if (p->p_stat == SSLEEP) { 768 /* OPTIMIZED EXPANSION OF setrunnable(p); */ 769 CTR4(KTR_PROC, 770 "wakeup: proc %p (pid %d, %s), schedlock %p", 771 p, p->p_pid, p->p_comm, (void *) sched_lock.mtx_lock); 772 if (p->p_slptime > 1) 773 updatepri(p); 774 p->p_slptime = 0; 775 p->p_stat = SRUN; 776 if (p->p_sflag & PS_INMEM) { 777 setrunqueue(p); 778 maybe_resched(p); 779 } else { 780 p->p_sflag |= PS_SWAPINREQ; 781 wakeup((caddr_t)&proc0); 782 } 783 /* END INLINE EXPANSION */ 784 goto restart; 785 } 786 } 787 } 788 mtx_unlock_spin(&sched_lock); 789 splx(s); 790} 791 792/* 793 * Make a process sleeping on the specified identifier runnable. 794 * May wake more than one process if a target process is currently 795 * swapped out. 796 */ 797void 798wakeup_one(ident) 799 register void *ident; 800{ 801 register struct slpquehead *qp; 802 register struct proc *p; 803 int s; 804 805 s = splhigh(); 806 mtx_lock_spin(&sched_lock); 807 qp = &slpque[LOOKUP(ident)]; 808 809 TAILQ_FOREACH(p, qp, p_slpq) { 810 if (p->p_wchan == ident) { 811 TAILQ_REMOVE(qp, p, p_slpq); 812 p->p_wchan = NULL; 813 if (p->p_stat == SSLEEP) { 814 /* OPTIMIZED EXPANSION OF setrunnable(p); */ 815 CTR4(KTR_PROC, 816 "wakeup1: proc %p (pid %d, %s), schedlock %p", 817 p, p->p_pid, p->p_comm, (void *) sched_lock.mtx_lock); 818 if (p->p_slptime > 1) 819 updatepri(p); 820 p->p_slptime = 0; 821 p->p_stat = SRUN; 822 if (p->p_sflag & PS_INMEM) { 823 setrunqueue(p); 824 maybe_resched(p); 825 break; 826 } else { 827 p->p_sflag |= PS_SWAPINREQ; 828 wakeup((caddr_t)&proc0); 829 } 830 /* END INLINE EXPANSION */ 831 } 832 } 833 } 834 mtx_unlock_spin(&sched_lock); 835 splx(s); 836} 837 838/* 839 * The machine independent parts of mi_switch(). 840 * Must be called at splstatclock() or higher. 841 */ 842void 843mi_switch() 844{ 845 struct timeval new_switchtime; 846 register struct proc *p = curproc; /* XXX */ 847#if 0 848 register struct rlimit *rlim; 849#endif 850 int x; 851 u_int sched_nest; 852 853 /* 854 * XXX this spl is almost unnecessary. It is partly to allow for 855 * sloppy callers that don't do it (issignal() via CURSIG() is the 856 * main offender). It is partly to work around a bug in the i386 857 * cpu_switch() (the ipl is not preserved). We ran for years 858 * without it. I think there was only a interrupt latency problem. 859 * The main caller, msleep(), does an splx() a couple of instructions 860 * after calling here. The buggy caller, issignal(), usually calls 861 * here at spl0() and sometimes returns at splhigh(). The process 862 * then runs for a little too long at splhigh(). The ipl gets fixed 863 * when the process returns to user mode (or earlier). 864 * 865 * It would probably be better to always call here at spl0(). Callers 866 * are prepared to give up control to another process, so they must 867 * be prepared to be interrupted. The clock stuff here may not 868 * actually need splstatclock(). 869 */ 870 x = splstatclock(); 871 872 mtx_assert(&sched_lock, MA_OWNED | MA_NOTRECURSED); 873 874 /* 875 * Compute the amount of time during which the current 876 * process was running, and add that to its total so far. 877 */ 878 microuptime(&new_switchtime); 879 if (timevalcmp(&new_switchtime, PCPU_PTR(switchtime), <)) { 880#if 0 881 /* XXX: This doesn't play well with sched_lock right now. */ 882 printf("microuptime() went backwards (%ld.%06ld -> %ld.%06ld)\n", 883 PCPU_GET(switchtime.tv_sec), PCPU_GET(switchtime.tv_usec), 884 new_switchtime.tv_sec, new_switchtime.tv_usec); 885#endif 886 new_switchtime = PCPU_GET(switchtime); 887 } else { 888 p->p_runtime += (new_switchtime.tv_usec - PCPU_GET(switchtime.tv_usec)) + 889 (new_switchtime.tv_sec - PCPU_GET(switchtime.tv_sec)) * 890 (int64_t)1000000; 891 } 892 893#if 0 894 /* 895 * Check if the process exceeds its cpu resource allocation. 896 * If over max, kill it. 897 * 898 * XXX drop sched_lock, pickup Giant 899 */ 900 if (p->p_stat != SZOMB && p->p_limit->p_cpulimit != RLIM_INFINITY && 901 p->p_runtime > p->p_limit->p_cpulimit) { 902 rlim = &p->p_rlimit[RLIMIT_CPU]; 903 if (p->p_runtime / (rlim_t)1000000 >= rlim->rlim_max) { 904 mtx_unlock_spin(&sched_lock); 905 killproc(p, "exceeded maximum CPU limit"); 906 mtx_lock_spin(&sched_lock); 907 } else { 908 mtx_unlock_spin(&sched_lock); 909 psignal(p, SIGXCPU); 910 mtx_lock_spin(&sched_lock); 911 if (rlim->rlim_cur < rlim->rlim_max) { 912 /* XXX: we should make a private copy */ 913 rlim->rlim_cur += 5; 914 } 915 } 916 } 917#endif 918 919 /* 920 * Pick a new current process and record its start time. 921 */ 922 cnt.v_swtch++; 923 PCPU_SET(switchtime, new_switchtime); 924 CTR4(KTR_PROC, "mi_switch: old proc %p (pid %d, %s), schedlock %p", 925 p, p->p_pid, p->p_comm, (void *) sched_lock.mtx_lock); 926 sched_nest = sched_lock.mtx_recurse; 927 curproc->p_lastcpu = curproc->p_oncpu; 928 curproc->p_oncpu = NOCPU; 929 clear_resched(); 930 cpu_switch(); 931 curproc->p_oncpu = PCPU_GET(cpuid); 932 sched_lock.mtx_recurse = sched_nest; 933 sched_lock.mtx_lock = (uintptr_t)curproc; 934 CTR4(KTR_PROC, "mi_switch: new proc %p (pid %d, %s), schedlock %p", 935 p, p->p_pid, p->p_comm, (void *) sched_lock.mtx_lock); 936 if (PCPU_GET(switchtime.tv_sec) == 0) 937 microuptime(PCPU_PTR(switchtime)); 938 PCPU_SET(switchticks, ticks); 939 splx(x); 940} 941 942/* 943 * Change process state to be runnable, 944 * placing it on the run queue if it is in memory, 945 * and awakening the swapper if it isn't in memory. 946 */ 947void 948setrunnable(p) 949 register struct proc *p; 950{ 951 register int s; 952 953 s = splhigh(); 954 mtx_lock_spin(&sched_lock); 955 switch (p->p_stat) { 956 case 0: 957 case SRUN: 958 case SZOMB: 959 case SWAIT: 960 default: 961 panic("setrunnable"); 962 case SSTOP: 963 case SSLEEP: /* e.g. when sending signals */ 964 if (p->p_sflag & PS_CVWAITQ) 965 cv_waitq_remove(p); 966 else 967 unsleep(p); 968 break; 969 970 case SIDL: 971 break; 972 } 973 p->p_stat = SRUN; 974 if (p->p_sflag & PS_INMEM) 975 setrunqueue(p); 976 splx(s); 977 if (p->p_slptime > 1) 978 updatepri(p); 979 p->p_slptime = 0; 980 if ((p->p_sflag & PS_INMEM) == 0) { 981 p->p_sflag |= PS_SWAPINREQ; 982 wakeup((caddr_t)&proc0); 983 } 984 else 985 maybe_resched(p); 986 mtx_unlock_spin(&sched_lock); 987} 988 989/* 990 * Compute the priority of a process when running in user mode. 991 * Arrange to reschedule if the resulting priority is better 992 * than that of the current process. 993 */ 994void 995resetpriority(p) 996 register struct proc *p; 997{ 998 register unsigned int newpriority; 999 1000 mtx_lock_spin(&sched_lock); 1001 if (p->p_pri.pri_class == PRI_TIMESHARE) { 1002 newpriority = PUSER + p->p_estcpu / INVERSE_ESTCPU_WEIGHT + 1003 NICE_WEIGHT * (p->p_nice - PRIO_MIN); 1004 newpriority = min(max(newpriority, PRI_MIN_TIMESHARE), 1005 PRI_MAX_TIMESHARE); 1006 p->p_pri.pri_user = newpriority; 1007 } 1008 maybe_resched(p); 1009 mtx_unlock_spin(&sched_lock); 1010} 1011 1012/* ARGSUSED */ 1013static void 1014sched_setup(dummy) 1015 void *dummy; 1016{ 1017 1018 callout_init(&schedcpu_callout, 1); 1019 callout_init(&roundrobin_callout, 0); 1020 1021 /* Kick off timeout driven events by calling first time. */ 1022 roundrobin(NULL); 1023 schedcpu(NULL); 1024} 1025 1026/* 1027 * We adjust the priority of the current process. The priority of 1028 * a process gets worse as it accumulates CPU time. The cpu usage 1029 * estimator (p_estcpu) is increased here. resetpriority() will 1030 * compute a different priority each time p_estcpu increases by 1031 * INVERSE_ESTCPU_WEIGHT 1032 * (until MAXPRI is reached). The cpu usage estimator ramps up 1033 * quite quickly when the process is running (linearly), and decays 1034 * away exponentially, at a rate which is proportionally slower when 1035 * the system is busy. The basic principle is that the system will 1036 * 90% forget that the process used a lot of CPU time in 5 * loadav 1037 * seconds. This causes the system to favor processes which haven't 1038 * run much recently, and to round-robin among other processes. 1039 */ 1040void 1041schedclock(p) 1042 struct proc *p; 1043{ 1044 1045 p->p_cpticks++; 1046 p->p_estcpu = ESTCPULIM(p->p_estcpu + 1); 1047 if ((p->p_estcpu % INVERSE_ESTCPU_WEIGHT) == 0) { 1048 resetpriority(p); 1049 if (p->p_pri.pri_level >= PUSER) 1050 p->p_pri.pri_level = p->p_pri.pri_user; 1051 } 1052} 1053 1054/* 1055 * General purpose yield system call 1056 */ 1057int 1058yield(struct proc *p, struct yield_args *uap) 1059{ 1060 int s; 1061 1062 p->p_retval[0] = 0; 1063 1064 s = splhigh(); 1065 mtx_lock_spin(&sched_lock); 1066 DROP_GIANT_NOSWITCH(); 1067 p->p_pri.pri_level = PRI_MAX_TIMESHARE; 1068 setrunqueue(p); 1069 p->p_stats->p_ru.ru_nvcsw++; 1070 mi_switch(); 1071 mtx_unlock_spin(&sched_lock); 1072 PICKUP_GIANT(); 1073 splx(s); 1074 1075 return (0); 1076} 1077