kern_synch.c revision 76830
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 76830 2001-05-19 01:40:48Z 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(!mtx_owned(&vm_mtx) || mtx == &vm_mtx, 382 ("sleeping with vm_mtx held.")); 383#ifdef KTRACE 384 if (p && KTRPOINT(p, KTR_CSW)) 385 ktrcsw(p->p_tracep, 1, 0); 386#endif 387 WITNESS_SLEEP(0, &mtx->mtx_object); 388 mtx_lock_spin(&sched_lock); 389 if (cold || panicstr) { 390 /* 391 * After a panic, or during autoconfiguration, 392 * just give interrupts a chance, then just return; 393 * don't run any other procs or panic below, 394 * in case this is the idle process and already asleep. 395 */ 396 if (mtx != NULL && priority & PDROP) 397 mtx_unlock_flags(mtx, MTX_NOSWITCH); 398 mtx_unlock_spin(&sched_lock); 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_object, 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 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 PROC_LOCK(p); 488 if (SIGISMEMBER(p->p_sigacts->ps_sigintr, sig)) 489 rval = EINTR; 490 else 491 rval = ERESTART; 492 PROC_UNLOCK(p); 493 goto out; 494 } 495out: 496#ifdef KTRACE 497 if (KTRPOINT(p, KTR_CSW)) 498 ktrcsw(p->p_tracep, 0, 0); 499#endif 500 PICKUP_GIANT(); 501 if (mtx != NULL) { 502 mtx_lock(mtx); 503 WITNESS_RESTORE(&mtx->mtx_object, mtx); 504 } 505 return (rval); 506} 507 508/* 509 * asleep() - async sleep call. Place process on wait queue and return 510 * immediately without blocking. The process stays runnable until mawait() 511 * is called. If ident is NULL, remove process from wait queue if it is still 512 * on one. 513 * 514 * Only the most recent sleep condition is effective when making successive 515 * calls to asleep() or when calling msleep(). 516 * 517 * The timeout, if any, is not initiated until mawait() is called. The sleep 518 * priority, signal, and timeout is specified in the asleep() call but may be 519 * overriden in the mawait() call. 520 * 521 * <<<<<<<< EXPERIMENTAL, UNTESTED >>>>>>>>>> 522 */ 523 524int 525asleep(void *ident, int priority, const char *wmesg, int timo) 526{ 527 struct proc *p = curproc; 528 int s; 529 530 /* 531 * obtain sched_lock while manipulating sleep structures and slpque. 532 * 533 * Remove preexisting wait condition (if any) and place process 534 * on appropriate slpque, but do not put process to sleep. 535 */ 536 537 s = splhigh(); 538 mtx_lock_spin(&sched_lock); 539 540 if (p->p_wchan != NULL) 541 unsleep(p); 542 543 if (ident) { 544 p->p_wchan = ident; 545 p->p_wmesg = wmesg; 546 p->p_slptime = 0; 547 p->p_asleep.as_priority = priority; 548 p->p_asleep.as_timo = timo; 549 TAILQ_INSERT_TAIL(&slpque[LOOKUP(ident)], p, p_slpq); 550 } 551 552 mtx_unlock_spin(&sched_lock); 553 splx(s); 554 555 return(0); 556} 557 558/* 559 * mawait() - wait for async condition to occur. The process blocks until 560 * wakeup() is called on the most recent asleep() address. If wakeup is called 561 * prior to mawait(), mawait() winds up being a NOP. 562 * 563 * If mawait() is called more then once (without an intervening asleep() call), 564 * mawait() is still effectively a NOP but it calls mi_switch() to give other 565 * processes some cpu before returning. The process is left runnable. 566 * 567 * <<<<<<<< EXPERIMENTAL, UNTESTED >>>>>>>>>> 568 */ 569 570int 571mawait(struct mtx *mtx, int priority, int timo) 572{ 573 struct proc *p = curproc; 574 int rval = 0; 575 int s; 576 WITNESS_SAVE_DECL(mtx); 577 578 WITNESS_SLEEP(0, &mtx->mtx_object); 579 mtx_lock_spin(&sched_lock); 580 DROP_GIANT_NOSWITCH(); 581 if (mtx != NULL) { 582 mtx_assert(mtx, MA_OWNED | MA_NOTRECURSED); 583 WITNESS_SAVE(&mtx->mtx_object, mtx); 584 mtx_unlock_flags(mtx, MTX_NOSWITCH); 585 if (priority & PDROP) 586 mtx = NULL; 587 } 588 589 s = splhigh(); 590 591 if (p->p_wchan != NULL) { 592 int sig; 593 int catch; 594 595 /* 596 * The call to mawait() can override defaults specified in 597 * the original asleep(). 598 */ 599 if (priority < 0) 600 priority = p->p_asleep.as_priority; 601 if (timo < 0) 602 timo = p->p_asleep.as_timo; 603 604 /* 605 * Install timeout 606 */ 607 608 if (timo) 609 callout_reset(&p->p_slpcallout, timo, endtsleep, p); 610 611 sig = 0; 612 catch = priority & PCATCH; 613 614 if (catch) { 615 p->p_sflag |= PS_SINTR; 616 mtx_unlock_spin(&sched_lock); 617 if ((sig = CURSIG(p))) { 618 mtx_lock_spin(&sched_lock); 619 if (p->p_wchan) 620 unsleep(p); 621 p->p_stat = SRUN; 622 goto resume; 623 } 624 mtx_lock_spin(&sched_lock); 625 if (p->p_wchan == NULL) { 626 catch = 0; 627 goto resume; 628 } 629 } 630 p->p_stat = SSLEEP; 631 p->p_stats->p_ru.ru_nvcsw++; 632 mi_switch(); 633resume: 634 635 splx(s); 636 p->p_sflag &= ~PS_SINTR; 637 if (p->p_sflag & PS_TIMEOUT) { 638 p->p_sflag &= ~PS_TIMEOUT; 639 if (sig == 0) { 640#ifdef KTRACE 641 if (KTRPOINT(p, KTR_CSW)) 642 ktrcsw(p->p_tracep, 0, 0); 643#endif 644 rval = EWOULDBLOCK; 645 mtx_unlock_spin(&sched_lock); 646 goto out; 647 } 648 } else if (timo) 649 callout_stop(&p->p_slpcallout); 650 mtx_unlock_spin(&sched_lock); 651 652 if (catch && (sig != 0 || (sig = CURSIG(p)))) { 653#ifdef KTRACE 654 if (KTRPOINT(p, KTR_CSW)) 655 ktrcsw(p->p_tracep, 0, 0); 656#endif 657 PROC_LOCK(p); 658 if (SIGISMEMBER(p->p_sigacts->ps_sigintr, sig)) 659 rval = EINTR; 660 else 661 rval = ERESTART; 662 PROC_UNLOCK(p); 663 goto out; 664 } 665#ifdef KTRACE 666 if (KTRPOINT(p, KTR_CSW)) 667 ktrcsw(p->p_tracep, 0, 0); 668#endif 669 } else { 670 /* 671 * If as_priority is 0, mawait() has been called without an 672 * intervening asleep(). We are still effectively a NOP, 673 * but we call mi_switch() for safety. 674 */ 675 676 if (p->p_asleep.as_priority == 0) { 677 p->p_stats->p_ru.ru_nvcsw++; 678 mi_switch(); 679 } 680 mtx_unlock_spin(&sched_lock); 681 splx(s); 682 } 683 684 /* 685 * clear p_asleep.as_priority as an indication that mawait() has been 686 * called. If mawait() is called again without an intervening asleep(), 687 * mawait() is still effectively a NOP but the above mi_switch() code 688 * is triggered as a safety. 689 */ 690 p->p_asleep.as_priority = 0; 691 692out: 693 PICKUP_GIANT(); 694 if (mtx != NULL) { 695 mtx_lock(mtx); 696 WITNESS_RESTORE(&mtx->mtx_object, mtx); 697 } 698 return (rval); 699} 700 701/* 702 * Implement timeout for msleep or asleep()/mawait() 703 * 704 * If process hasn't been awakened (wchan non-zero), 705 * set timeout flag and undo the sleep. If proc 706 * is stopped, just unsleep so it will remain stopped. 707 * MP-safe, called without the Giant mutex. 708 */ 709static void 710endtsleep(arg) 711 void *arg; 712{ 713 register struct proc *p; 714 int s; 715 716 p = (struct proc *)arg; 717 CTR4(KTR_PROC, 718 "endtsleep: proc %p (pid %d, %s), schedlock %p", 719 p, p->p_pid, p->p_comm, (void *) sched_lock.mtx_lock); 720 s = splhigh(); 721 mtx_lock_spin(&sched_lock); 722 if (p->p_wchan) { 723 if (p->p_stat == SSLEEP) 724 setrunnable(p); 725 else 726 unsleep(p); 727 p->p_sflag |= PS_TIMEOUT; 728 } 729 mtx_unlock_spin(&sched_lock); 730 splx(s); 731} 732 733/* 734 * Remove a process from its wait queue 735 */ 736void 737unsleep(p) 738 register struct proc *p; 739{ 740 int s; 741 742 s = splhigh(); 743 mtx_lock_spin(&sched_lock); 744 if (p->p_wchan) { 745 TAILQ_REMOVE(&slpque[LOOKUP(p->p_wchan)], p, p_slpq); 746 p->p_wchan = NULL; 747 } 748 mtx_unlock_spin(&sched_lock); 749 splx(s); 750} 751 752/* 753 * Make all processes sleeping on the specified identifier runnable. 754 */ 755void 756wakeup(ident) 757 register void *ident; 758{ 759 register struct slpquehead *qp; 760 register struct proc *p; 761 int s; 762 763 s = splhigh(); 764 mtx_lock_spin(&sched_lock); 765 qp = &slpque[LOOKUP(ident)]; 766restart: 767 TAILQ_FOREACH(p, qp, p_slpq) { 768 if (p->p_wchan == ident) { 769 TAILQ_REMOVE(qp, p, p_slpq); 770 p->p_wchan = NULL; 771 if (p->p_stat == SSLEEP) { 772 /* OPTIMIZED EXPANSION OF setrunnable(p); */ 773 CTR4(KTR_PROC, 774 "wakeup: proc %p (pid %d, %s), schedlock %p", 775 p, p->p_pid, p->p_comm, (void *) sched_lock.mtx_lock); 776 if (p->p_slptime > 1) 777 updatepri(p); 778 p->p_slptime = 0; 779 p->p_stat = SRUN; 780 if (p->p_sflag & PS_INMEM) { 781 setrunqueue(p); 782 maybe_resched(p); 783 } else { 784 p->p_sflag |= PS_SWAPINREQ; 785 wakeup((caddr_t)&proc0); 786 } 787 /* END INLINE EXPANSION */ 788 goto restart; 789 } 790 } 791 } 792 mtx_unlock_spin(&sched_lock); 793 splx(s); 794} 795 796/* 797 * Make a process sleeping on the specified identifier runnable. 798 * May wake more than one process if a target process is currently 799 * swapped out. 800 */ 801void 802wakeup_one(ident) 803 register void *ident; 804{ 805 register struct slpquehead *qp; 806 register struct proc *p; 807 int s; 808 809 s = splhigh(); 810 mtx_lock_spin(&sched_lock); 811 qp = &slpque[LOOKUP(ident)]; 812 813 TAILQ_FOREACH(p, qp, p_slpq) { 814 if (p->p_wchan == ident) { 815 TAILQ_REMOVE(qp, p, p_slpq); 816 p->p_wchan = NULL; 817 if (p->p_stat == SSLEEP) { 818 /* OPTIMIZED EXPANSION OF setrunnable(p); */ 819 CTR4(KTR_PROC, 820 "wakeup1: proc %p (pid %d, %s), schedlock %p", 821 p, p->p_pid, p->p_comm, (void *) sched_lock.mtx_lock); 822 if (p->p_slptime > 1) 823 updatepri(p); 824 p->p_slptime = 0; 825 p->p_stat = SRUN; 826 if (p->p_sflag & PS_INMEM) { 827 setrunqueue(p); 828 maybe_resched(p); 829 break; 830 } else { 831 p->p_sflag |= PS_SWAPINREQ; 832 wakeup((caddr_t)&proc0); 833 } 834 /* END INLINE EXPANSION */ 835 } 836 } 837 } 838 mtx_unlock_spin(&sched_lock); 839 splx(s); 840} 841 842/* 843 * The machine independent parts of mi_switch(). 844 * Must be called at splstatclock() or higher. 845 */ 846void 847mi_switch() 848{ 849 struct timeval new_switchtime; 850 register struct proc *p = curproc; /* XXX */ 851#if 0 852 register struct rlimit *rlim; 853#endif 854 u_int sched_nest; 855 856 mtx_assert(&sched_lock, MA_OWNED | MA_NOTRECURSED); 857 858 /* 859 * Compute the amount of time during which the current 860 * process was running, and add that to its total so far. 861 */ 862 microuptime(&new_switchtime); 863 if (timevalcmp(&new_switchtime, PCPU_PTR(switchtime), <)) { 864#if 0 865 /* XXX: This doesn't play well with sched_lock right now. */ 866 printf("microuptime() went backwards (%ld.%06ld -> %ld.%06ld)\n", 867 PCPU_GET(switchtime.tv_sec), PCPU_GET(switchtime.tv_usec), 868 new_switchtime.tv_sec, new_switchtime.tv_usec); 869#endif 870 new_switchtime = PCPU_GET(switchtime); 871 } else { 872 p->p_runtime += (new_switchtime.tv_usec - PCPU_GET(switchtime.tv_usec)) + 873 (new_switchtime.tv_sec - PCPU_GET(switchtime.tv_sec)) * 874 (int64_t)1000000; 875 } 876 877#if 0 878 /* 879 * Check if the process exceeds its cpu resource allocation. 880 * If over max, kill it. 881 * 882 * XXX drop sched_lock, pickup Giant 883 */ 884 if (p->p_stat != SZOMB && p->p_limit->p_cpulimit != RLIM_INFINITY && 885 p->p_runtime > p->p_limit->p_cpulimit) { 886 rlim = &p->p_rlimit[RLIMIT_CPU]; 887 if (p->p_runtime / (rlim_t)1000000 >= rlim->rlim_max) { 888 mtx_unlock_spin(&sched_lock); 889 PROC_LOCK(p); 890 killproc(p, "exceeded maximum CPU limit"); 891 mtx_lock_spin(&sched_lock); 892 PROC_UNLOCK_NOSWITCH(p); 893 } else { 894 mtx_unlock_spin(&sched_lock); 895 PROC_LOCK(p); 896 psignal(p, SIGXCPU); 897 mtx_lock_spin(&sched_lock); 898 PROC_UNLOCK_NOSWITCH(p); 899 if (rlim->rlim_cur < rlim->rlim_max) { 900 /* XXX: we should make a private copy */ 901 rlim->rlim_cur += 5; 902 } 903 } 904 } 905#endif 906 907 /* 908 * Pick a new current process and record its start time. 909 */ 910 cnt.v_swtch++; 911 PCPU_SET(switchtime, new_switchtime); 912 CTR4(KTR_PROC, "mi_switch: old proc %p (pid %d, %s), schedlock %p", 913 p, p->p_pid, p->p_comm, (void *) sched_lock.mtx_lock); 914 sched_nest = sched_lock.mtx_recurse; 915 curproc->p_lastcpu = curproc->p_oncpu; 916 curproc->p_oncpu = NOCPU; 917 clear_resched(curproc); 918 cpu_switch(); 919 curproc->p_oncpu = PCPU_GET(cpuid); 920 sched_lock.mtx_recurse = sched_nest; 921 sched_lock.mtx_lock = (uintptr_t)curproc; 922 CTR4(KTR_PROC, "mi_switch: new proc %p (pid %d, %s), schedlock %p", 923 p, p->p_pid, p->p_comm, (void *) sched_lock.mtx_lock); 924 if (PCPU_GET(switchtime.tv_sec) == 0) 925 microuptime(PCPU_PTR(switchtime)); 926 PCPU_SET(switchticks, ticks); 927} 928 929/* 930 * Change process state to be runnable, 931 * placing it on the run queue if it is in memory, 932 * and awakening the swapper if it isn't in memory. 933 */ 934void 935setrunnable(p) 936 register struct proc *p; 937{ 938 register int s; 939 940 s = splhigh(); 941 mtx_lock_spin(&sched_lock); 942 switch (p->p_stat) { 943 case 0: 944 case SRUN: 945 case SZOMB: 946 case SWAIT: 947 default: 948 panic("setrunnable"); 949 case SSTOP: 950 case SSLEEP: /* e.g. when sending signals */ 951 if (p->p_sflag & PS_CVWAITQ) 952 cv_waitq_remove(p); 953 else 954 unsleep(p); 955 break; 956 957 case SIDL: 958 break; 959 } 960 p->p_stat = SRUN; 961 if (p->p_sflag & PS_INMEM) 962 setrunqueue(p); 963 splx(s); 964 if (p->p_slptime > 1) 965 updatepri(p); 966 p->p_slptime = 0; 967 if ((p->p_sflag & PS_INMEM) == 0) { 968 p->p_sflag |= PS_SWAPINREQ; 969 wakeup((caddr_t)&proc0); 970 } 971 else 972 maybe_resched(p); 973 mtx_unlock_spin(&sched_lock); 974} 975 976/* 977 * Compute the priority of a process when running in user mode. 978 * Arrange to reschedule if the resulting priority is better 979 * than that of the current process. 980 */ 981void 982resetpriority(p) 983 register struct proc *p; 984{ 985 register unsigned int newpriority; 986 987 mtx_lock_spin(&sched_lock); 988 if (p->p_pri.pri_class == PRI_TIMESHARE) { 989 newpriority = PUSER + p->p_estcpu / INVERSE_ESTCPU_WEIGHT + 990 NICE_WEIGHT * (p->p_nice - PRIO_MIN); 991 newpriority = min(max(newpriority, PRI_MIN_TIMESHARE), 992 PRI_MAX_TIMESHARE); 993 p->p_pri.pri_user = newpriority; 994 } 995 maybe_resched(p); 996 mtx_unlock_spin(&sched_lock); 997} 998 999/* ARGSUSED */ 1000static void 1001sched_setup(dummy) 1002 void *dummy; 1003{ 1004 1005 callout_init(&schedcpu_callout, 1); 1006 callout_init(&roundrobin_callout, 0); 1007 1008 /* Kick off timeout driven events by calling first time. */ 1009 roundrobin(NULL); 1010 schedcpu(NULL); 1011} 1012 1013/* 1014 * We adjust the priority of the current process. The priority of 1015 * a process gets worse as it accumulates CPU time. The cpu usage 1016 * estimator (p_estcpu) is increased here. resetpriority() will 1017 * compute a different priority each time p_estcpu increases by 1018 * INVERSE_ESTCPU_WEIGHT 1019 * (until MAXPRI is reached). The cpu usage estimator ramps up 1020 * quite quickly when the process is running (linearly), and decays 1021 * away exponentially, at a rate which is proportionally slower when 1022 * the system is busy. The basic principle is that the system will 1023 * 90% forget that the process used a lot of CPU time in 5 * loadav 1024 * seconds. This causes the system to favor processes which haven't 1025 * run much recently, and to round-robin among other processes. 1026 */ 1027void 1028schedclock(p) 1029 struct proc *p; 1030{ 1031 1032 p->p_cpticks++; 1033 p->p_estcpu = ESTCPULIM(p->p_estcpu + 1); 1034 if ((p->p_estcpu % INVERSE_ESTCPU_WEIGHT) == 0) { 1035 resetpriority(p); 1036 if (p->p_pri.pri_level >= PUSER) 1037 p->p_pri.pri_level = p->p_pri.pri_user; 1038 } 1039} 1040 1041/* 1042 * General purpose yield system call 1043 */ 1044int 1045yield(struct proc *p, struct yield_args *uap) 1046{ 1047 int s; 1048 1049 p->p_retval[0] = 0; 1050 1051 s = splhigh(); 1052 mtx_lock_spin(&sched_lock); 1053 DROP_GIANT_NOSWITCH(); 1054 p->p_pri.pri_level = PRI_MAX_TIMESHARE; 1055 setrunqueue(p); 1056 p->p_stats->p_ru.ru_nvcsw++; 1057 mi_switch(); 1058 mtx_unlock_spin(&sched_lock); 1059 PICKUP_GIANT(); 1060 splx(s); 1061 1062 return (0); 1063} 1064