kern_synch.c revision 69437
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 69437 2000-12-01 04:55:52Z jake $ 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/mutex.h> 51#include <sys/signalvar.h> 52#include <sys/resourcevar.h> 53#include <sys/vmmeter.h> 54#include <sys/sysctl.h> 55#include <vm/vm.h> 56#include <vm/vm_extern.h> 57#ifdef KTRACE 58#include <sys/uio.h> 59#include <sys/ktrace.h> 60#endif 61 62#include <machine/cpu.h> 63#include <machine/smp.h> 64 65static void sched_setup __P((void *dummy)); 66SYSINIT(sched_setup, SI_SUB_KICK_SCHEDULER, SI_ORDER_FIRST, sched_setup, NULL) 67 68u_char curpriority; 69int hogticks; 70int lbolt; 71int sched_quantum; /* Roundrobin scheduling quantum in ticks. */ 72 73static struct callout schedcpu_callout; 74static struct callout roundrobin_callout; 75 76static int curpriority_cmp __P((struct proc *p)); 77static void endtsleep __P((void *)); 78static void maybe_resched __P((struct proc *chk)); 79static void roundrobin __P((void *arg)); 80static void schedcpu __P((void *arg)); 81static void updatepri __P((struct proc *p)); 82 83static int 84sysctl_kern_quantum(SYSCTL_HANDLER_ARGS) 85{ 86 int error, new_val; 87 88 new_val = sched_quantum * tick; 89 error = sysctl_handle_int(oidp, &new_val, 0, req); 90 if (error != 0 || req->newptr == NULL) 91 return (error); 92 if (new_val < tick) 93 return (EINVAL); 94 sched_quantum = new_val / tick; 95 hogticks = 2 * sched_quantum; 96 return (0); 97} 98 99SYSCTL_PROC(_kern, OID_AUTO, quantum, CTLTYPE_INT|CTLFLAG_RW, 100 0, sizeof sched_quantum, sysctl_kern_quantum, "I", ""); 101 102/*- 103 * Compare priorities. Return: 104 * <0: priority of p < current priority 105 * 0: priority of p == current priority 106 * >0: priority of p > current priority 107 * The priorities are the normal priorities or the normal realtime priorities 108 * if p is on the same scheduler as curproc. Otherwise the process on the 109 * more realtimeish scheduler has lowest priority. As usual, a higher 110 * priority really means a lower priority. 111 */ 112static int 113curpriority_cmp(p) 114 struct proc *p; 115{ 116 int c_class, p_class; 117 118 c_class = RTP_PRIO_BASE(curproc->p_rtprio.type); 119 p_class = RTP_PRIO_BASE(p->p_rtprio.type); 120 if (p_class != c_class) 121 return (p_class - c_class); 122 if (p_class == RTP_PRIO_NORMAL) 123 return (((int)p->p_priority - (int)curpriority) / PPQ); 124 return ((int)p->p_rtprio.prio - (int)curproc->p_rtprio.prio); 125} 126 127/* 128 * Arrange to reschedule if necessary, taking the priorities and 129 * schedulers into account. 130 */ 131static void 132maybe_resched(chk) 133 struct proc *chk; 134{ 135 struct proc *p = curproc; /* XXX */ 136 137 /* 138 * XXX idle scheduler still broken because proccess stays on idle 139 * scheduler during waits (such as when getting FS locks). If a 140 * standard process becomes runaway cpu-bound, the system can lockup 141 * due to idle-scheduler processes in wakeup never getting any cpu. 142 */ 143 if (p == idleproc) { 144#if 0 145 need_resched(); 146#endif 147 } else if (chk == p) { 148 /* We may need to yield if our priority has been raised. */ 149 if (curpriority_cmp(chk) > 0) 150 need_resched(); 151 } else if (curpriority_cmp(chk) < 0) 152 need_resched(); 153} 154 155int 156roundrobin_interval(void) 157{ 158 return (sched_quantum); 159} 160 161/* 162 * Force switch among equal priority processes every 100ms. 163 */ 164/* ARGSUSED */ 165static void 166roundrobin(arg) 167 void *arg; 168{ 169#ifndef SMP 170 struct proc *p = curproc; /* XXX */ 171#endif 172 173#ifdef SMP 174 need_resched(); 175 forward_roundrobin(); 176#else 177 if (p == idleproc || RTP_PRIO_NEED_RR(p->p_rtprio.type)) 178 need_resched(); 179#endif 180 181 callout_reset(&roundrobin_callout, sched_quantum, roundrobin, NULL); 182} 183 184/* 185 * Constants for digital decay and forget: 186 * 90% of (p_estcpu) usage in 5 * loadav time 187 * 95% of (p_pctcpu) usage in 60 seconds (load insensitive) 188 * Note that, as ps(1) mentions, this can let percentages 189 * total over 100% (I've seen 137.9% for 3 processes). 190 * 191 * Note that schedclock() updates p_estcpu and p_cpticks asynchronously. 192 * 193 * We wish to decay away 90% of p_estcpu in (5 * loadavg) seconds. 194 * That is, the system wants to compute a value of decay such 195 * that the following for loop: 196 * for (i = 0; i < (5 * loadavg); i++) 197 * p_estcpu *= decay; 198 * will compute 199 * p_estcpu *= 0.1; 200 * for all values of loadavg: 201 * 202 * Mathematically this loop can be expressed by saying: 203 * decay ** (5 * loadavg) ~= .1 204 * 205 * The system computes decay as: 206 * decay = (2 * loadavg) / (2 * loadavg + 1) 207 * 208 * We wish to prove that the system's computation of decay 209 * will always fulfill the equation: 210 * decay ** (5 * loadavg) ~= .1 211 * 212 * If we compute b as: 213 * b = 2 * loadavg 214 * then 215 * decay = b / (b + 1) 216 * 217 * We now need to prove two things: 218 * 1) Given factor ** (5 * loadavg) ~= .1, prove factor == b/(b+1) 219 * 2) Given b/(b+1) ** power ~= .1, prove power == (5 * loadavg) 220 * 221 * Facts: 222 * For x close to zero, exp(x) =~ 1 + x, since 223 * exp(x) = 0! + x**1/1! + x**2/2! + ... . 224 * therefore exp(-1/b) =~ 1 - (1/b) = (b-1)/b. 225 * For x close to zero, ln(1+x) =~ x, since 226 * ln(1+x) = x - x**2/2 + x**3/3 - ... -1 < x < 1 227 * therefore ln(b/(b+1)) = ln(1 - 1/(b+1)) =~ -1/(b+1). 228 * ln(.1) =~ -2.30 229 * 230 * Proof of (1): 231 * Solve (factor)**(power) =~ .1 given power (5*loadav): 232 * solving for factor, 233 * ln(factor) =~ (-2.30/5*loadav), or 234 * factor =~ exp(-1/((5/2.30)*loadav)) =~ exp(-1/(2*loadav)) = 235 * exp(-1/b) =~ (b-1)/b =~ b/(b+1). QED 236 * 237 * Proof of (2): 238 * Solve (factor)**(power) =~ .1 given factor == (b/(b+1)): 239 * solving for power, 240 * power*ln(b/(b+1)) =~ -2.30, or 241 * power =~ 2.3 * (b + 1) = 4.6*loadav + 2.3 =~ 5*loadav. QED 242 * 243 * Actual power values for the implemented algorithm are as follows: 244 * loadav: 1 2 3 4 245 * power: 5.68 10.32 14.94 19.55 246 */ 247 248/* calculations for digital decay to forget 90% of usage in 5*loadav sec */ 249#define loadfactor(loadav) (2 * (loadav)) 250#define decay_cpu(loadfac, cpu) (((loadfac) * (cpu)) / ((loadfac) + FSCALE)) 251 252/* decay 95% of `p_pctcpu' in 60 seconds; see CCPU_SHIFT before changing */ 253static fixpt_t ccpu = 0.95122942450071400909 * FSCALE; /* exp(-1/20) */ 254SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, ""); 255 256/* kernel uses `FSCALE', userland (SHOULD) use kern.fscale */ 257static int fscale __unused = FSCALE; 258SYSCTL_INT(_kern, OID_AUTO, fscale, CTLFLAG_RD, 0, FSCALE, ""); 259 260/* 261 * If `ccpu' is not equal to `exp(-1/20)' and you still want to use the 262 * faster/more-accurate formula, you'll have to estimate CCPU_SHIFT below 263 * and possibly adjust FSHIFT in "param.h" so that (FSHIFT >= CCPU_SHIFT). 264 * 265 * To estimate CCPU_SHIFT for exp(-1/20), the following formula was used: 266 * 1 - exp(-1/20) ~= 0.0487 ~= 0.0488 == 1 (fixed pt, *11* bits). 267 * 268 * If you don't want to bother with the faster/more-accurate formula, you 269 * can set CCPU_SHIFT to (FSHIFT + 1) which will use a slower/less-accurate 270 * (more general) method of calculating the %age of CPU used by a process. 271 */ 272#define CCPU_SHIFT 11 273 274/* 275 * Recompute process priorities, every hz ticks. 276 * MP-safe, called without the Giant mutex. 277 */ 278/* ARGSUSED */ 279static void 280schedcpu(arg) 281 void *arg; 282{ 283 register fixpt_t loadfac = loadfactor(averunnable.ldavg[0]); 284 register struct proc *p; 285 register int realstathz, s; 286 287 realstathz = stathz ? stathz : hz; 288 lockmgr(&allproc_lock, LK_SHARED, NULL, CURPROC); 289 LIST_FOREACH(p, &allproc, p_list) { 290 /* 291 * Increment time in/out of memory and sleep time 292 * (if sleeping). We ignore overflow; with 16-bit int's 293 * (remember them?) overflow takes 45 days. 294 if (p->p_stat == SWAIT) 295 continue; 296 */ 297 mtx_enter(&sched_lock, MTX_SPIN); 298 p->p_swtime++; 299 if (p->p_stat == SSLEEP || p->p_stat == SSTOP) 300 p->p_slptime++; 301 p->p_pctcpu = (p->p_pctcpu * ccpu) >> FSHIFT; 302 /* 303 * If the process has slept the entire second, 304 * stop recalculating its priority until it wakes up. 305 */ 306 if (p->p_slptime > 1) { 307 mtx_exit(&sched_lock, MTX_SPIN); 308 continue; 309 } 310 311 /* 312 * prevent state changes and protect run queue 313 */ 314 s = splhigh(); 315 316 /* 317 * p_pctcpu is only for ps. 318 */ 319#if (FSHIFT >= CCPU_SHIFT) 320 p->p_pctcpu += (realstathz == 100)? 321 ((fixpt_t) p->p_cpticks) << (FSHIFT - CCPU_SHIFT): 322 100 * (((fixpt_t) p->p_cpticks) 323 << (FSHIFT - CCPU_SHIFT)) / realstathz; 324#else 325 p->p_pctcpu += ((FSCALE - ccpu) * 326 (p->p_cpticks * FSCALE / realstathz)) >> FSHIFT; 327#endif 328 p->p_cpticks = 0; 329 p->p_estcpu = decay_cpu(loadfac, p->p_estcpu); 330 resetpriority(p); 331 if (p->p_priority >= PUSER) { 332 if ((p != curproc) && 333#ifdef SMP 334 p->p_oncpu == 0xff && /* idle */ 335#endif 336 p->p_stat == SRUN && 337 (p->p_flag & P_INMEM) && 338 (p->p_priority / PPQ) != (p->p_usrpri / PPQ)) { 339 remrunqueue(p); 340 p->p_priority = p->p_usrpri; 341 setrunqueue(p); 342 } else 343 p->p_priority = p->p_usrpri; 344 } 345 mtx_exit(&sched_lock, MTX_SPIN); 346 splx(s); 347 } 348 lockmgr(&allproc_lock, LK_RELEASE, NULL, CURPROC); 349 vmmeter(); 350 wakeup((caddr_t)&lbolt); 351 callout_reset(&schedcpu_callout, hz, schedcpu, NULL); 352} 353 354/* 355 * Recalculate the priority of a process after it has slept for a while. 356 * For all load averages >= 1 and max p_estcpu of 255, sleeping for at 357 * least six times the loadfactor will decay p_estcpu to zero. 358 */ 359static void 360updatepri(p) 361 register struct proc *p; 362{ 363 register unsigned int newcpu = p->p_estcpu; 364 register fixpt_t loadfac = loadfactor(averunnable.ldavg[0]); 365 366 if (p->p_slptime > 5 * loadfac) 367 p->p_estcpu = 0; 368 else { 369 p->p_slptime--; /* the first time was done in schedcpu */ 370 while (newcpu && --p->p_slptime) 371 newcpu = decay_cpu(loadfac, newcpu); 372 p->p_estcpu = newcpu; 373 } 374 resetpriority(p); 375} 376 377/* 378 * We're only looking at 7 bits of the address; everything is 379 * aligned to 4, lots of things are aligned to greater powers 380 * of 2. Shift right by 8, i.e. drop the bottom 256 worth. 381 */ 382#define TABLESIZE 128 383static TAILQ_HEAD(slpquehead, proc) slpque[TABLESIZE]; 384#define LOOKUP(x) (((intptr_t)(x) >> 8) & (TABLESIZE - 1)) 385 386void 387sleepinit(void) 388{ 389 int i; 390 391 sched_quantum = hz/10; 392 hogticks = 2 * sched_quantum; 393 for (i = 0; i < TABLESIZE; i++) 394 TAILQ_INIT(&slpque[i]); 395} 396 397/* 398 * General sleep call. Suspends the current process until a wakeup is 399 * performed on the specified identifier. The process will then be made 400 * runnable with the specified priority. Sleeps at most timo/hz seconds 401 * (0 means no timeout). If pri includes PCATCH flag, signals are checked 402 * before and after sleeping, else signals are not checked. Returns 0 if 403 * awakened, EWOULDBLOCK if the timeout expires. If PCATCH is set and a 404 * signal needs to be delivered, ERESTART is returned if the current system 405 * call should be restarted if possible, and EINTR is returned if the system 406 * call should be interrupted by the signal (return EINTR). 407 * 408 * The mutex argument is exited before the caller is suspended, and 409 * entered before msleep returns. If priority includes the PDROP 410 * flag the mutex is not entered before returning. 411 */ 412int 413msleep(ident, mtx, priority, wmesg, timo) 414 void *ident; 415 struct mtx *mtx; 416 int priority, timo; 417 const char *wmesg; 418{ 419 struct proc *p = curproc; 420 int s, sig, catch = priority & PCATCH; 421 int rval = 0; 422 WITNESS_SAVE_DECL(mtx); 423 424#ifdef KTRACE 425 if (p && KTRPOINT(p, KTR_CSW)) 426 ktrcsw(p->p_tracep, 1, 0); 427#endif 428 WITNESS_SLEEP(0, mtx); 429 mtx_enter(&sched_lock, MTX_SPIN); 430 s = splhigh(); 431 if (cold || panicstr) { 432 /* 433 * After a panic, or during autoconfiguration, 434 * just give interrupts a chance, then just return; 435 * don't run any other procs or panic below, 436 * in case this is the idle process and already asleep. 437 */ 438 if (mtx != NULL && priority & PDROP) 439 mtx_exit(mtx, MTX_DEF | MTX_NOSWITCH); 440 mtx_exit(&sched_lock, MTX_SPIN); 441 splx(s); 442 return (0); 443 } 444 445 DROP_GIANT_NOSWITCH(); 446 447 if (mtx != NULL) { 448 mtx_assert(mtx, MA_OWNED | MA_NOTRECURSED); 449 WITNESS_SAVE(mtx, mtx); 450 mtx_exit(mtx, MTX_DEF | MTX_NOSWITCH); 451 if (priority & PDROP) 452 mtx = NULL; 453 } 454 455 KASSERT(p != NULL, ("msleep1")); 456 KASSERT(ident != NULL && p->p_stat == SRUN, ("msleep")); 457 /* 458 * Process may be sitting on a slpque if asleep() was called, remove 459 * it before re-adding. 460 */ 461 if (p->p_wchan != NULL) 462 unsleep(p); 463 464 p->p_wchan = ident; 465 p->p_wmesg = wmesg; 466 p->p_slptime = 0; 467 p->p_priority = priority & PRIMASK; 468 CTR4(KTR_PROC, "msleep: proc %p (pid %d, %s), schedlock %p", 469 p, p->p_pid, p->p_comm, (void *) sched_lock.mtx_lock); 470 TAILQ_INSERT_TAIL(&slpque[LOOKUP(ident)], p, p_slpq); 471 if (timo) 472 callout_reset(&p->p_slpcallout, timo, endtsleep, p); 473 /* 474 * We put ourselves on the sleep queue and start our timeout 475 * before calling CURSIG, as we could stop there, and a wakeup 476 * or a SIGCONT (or both) could occur while we were stopped. 477 * A SIGCONT would cause us to be marked as SSLEEP 478 * without resuming us, thus we must be ready for sleep 479 * when CURSIG is called. If the wakeup happens while we're 480 * stopped, p->p_wchan will be 0 upon return from CURSIG. 481 */ 482 if (catch) { 483 CTR4(KTR_PROC, 484 "msleep caught: proc %p (pid %d, %s), schedlock %p", 485 p, p->p_pid, p->p_comm, (void *) sched_lock.mtx_lock); 486 p->p_flag |= P_SINTR; 487 mtx_exit(&sched_lock, MTX_SPIN); 488 if ((sig = CURSIG(p))) { 489 mtx_enter(&sched_lock, MTX_SPIN); 490 if (p->p_wchan) 491 unsleep(p); 492 p->p_stat = SRUN; 493 goto resume; 494 } 495 mtx_enter(&sched_lock, MTX_SPIN); 496 if (p->p_wchan == 0) { 497 catch = 0; 498 goto resume; 499 } 500 } else 501 sig = 0; 502 p->p_stat = SSLEEP; 503 p->p_stats->p_ru.ru_nvcsw++; 504 mi_switch(); 505 CTR4(KTR_PROC, 506 "msleep resume: proc %p (pid %d, %s), schedlock %p", 507 p, p->p_pid, p->p_comm, (void *) sched_lock.mtx_lock); 508resume: 509 curpriority = p->p_usrpri; 510 splx(s); 511 p->p_flag &= ~P_SINTR; 512 if (p->p_flag & P_TIMEOUT) { 513 p->p_flag &= ~P_TIMEOUT; 514 if (sig == 0) { 515#ifdef KTRACE 516 if (KTRPOINT(p, KTR_CSW)) 517 ktrcsw(p->p_tracep, 0, 0); 518#endif 519 rval = EWOULDBLOCK; 520 mtx_exit(&sched_lock, MTX_SPIN); 521 goto out; 522 } 523 } else if (timo) 524 callout_stop(&p->p_slpcallout); 525 mtx_exit(&sched_lock, MTX_SPIN); 526 527 if (catch && (sig != 0 || (sig = CURSIG(p)))) { 528#ifdef KTRACE 529 if (KTRPOINT(p, KTR_CSW)) 530 ktrcsw(p->p_tracep, 0, 0); 531#endif 532 if (SIGISMEMBER(p->p_sigacts->ps_sigintr, sig)) 533 rval = EINTR; 534 else 535 rval = ERESTART; 536 goto out; 537 } 538out: 539#ifdef KTRACE 540 if (KTRPOINT(p, KTR_CSW)) 541 ktrcsw(p->p_tracep, 0, 0); 542#endif 543 PICKUP_GIANT(); 544 if (mtx != NULL) { 545 mtx_enter(mtx, MTX_DEF); 546 WITNESS_RESTORE(mtx, mtx); 547 } 548 return (rval); 549} 550 551/* 552 * asleep() - async sleep call. Place process on wait queue and return 553 * immediately without blocking. The process stays runnable until mawait() 554 * is called. If ident is NULL, remove process from wait queue if it is still 555 * on one. 556 * 557 * Only the most recent sleep condition is effective when making successive 558 * calls to asleep() or when calling msleep(). 559 * 560 * The timeout, if any, is not initiated until mawait() is called. The sleep 561 * priority, signal, and timeout is specified in the asleep() call but may be 562 * overriden in the mawait() call. 563 * 564 * <<<<<<<< EXPERIMENTAL, UNTESTED >>>>>>>>>> 565 */ 566 567int 568asleep(void *ident, int priority, const char *wmesg, int timo) 569{ 570 struct proc *p = curproc; 571 int s; 572 573 /* 574 * obtain sched_lock while manipulating sleep structures and slpque. 575 * 576 * Remove preexisting wait condition (if any) and place process 577 * on appropriate slpque, but do not put process to sleep. 578 */ 579 580 s = splhigh(); 581 mtx_enter(&sched_lock, MTX_SPIN); 582 583 if (p->p_wchan != NULL) 584 unsleep(p); 585 586 if (ident) { 587 p->p_wchan = ident; 588 p->p_wmesg = wmesg; 589 p->p_slptime = 0; 590 p->p_asleep.as_priority = priority; 591 p->p_asleep.as_timo = timo; 592 TAILQ_INSERT_TAIL(&slpque[LOOKUP(ident)], p, p_slpq); 593 } 594 595 mtx_exit(&sched_lock, MTX_SPIN); 596 splx(s); 597 598 return(0); 599} 600 601/* 602 * mawait() - wait for async condition to occur. The process blocks until 603 * wakeup() is called on the most recent asleep() address. If wakeup is called 604 * prior to mawait(), mawait() winds up being a NOP. 605 * 606 * If mawait() is called more then once (without an intervening asleep() call), 607 * mawait() is still effectively a NOP but it calls mi_switch() to give other 608 * processes some cpu before returning. The process is left runnable. 609 * 610 * <<<<<<<< EXPERIMENTAL, UNTESTED >>>>>>>>>> 611 */ 612 613int 614mawait(struct mtx *mtx, int priority, int timo) 615{ 616 struct proc *p = curproc; 617 int rval = 0; 618 int s; 619 WITNESS_SAVE_DECL(mtx); 620 621 WITNESS_SLEEP(0, mtx); 622 mtx_enter(&sched_lock, MTX_SPIN); 623 DROP_GIANT_NOSWITCH(); 624 if (mtx != NULL) { 625 mtx_assert(mtx, MA_OWNED | MA_NOTRECURSED); 626 WITNESS_SAVE(mtx, mtx); 627 mtx_exit(mtx, MTX_DEF | MTX_NOSWITCH); 628 if (priority & PDROP) 629 mtx = NULL; 630 } 631 632 s = splhigh(); 633 634 if (p->p_wchan != NULL) { 635 int sig; 636 int catch; 637 638 /* 639 * The call to mawait() can override defaults specified in 640 * the original asleep(). 641 */ 642 if (priority < 0) 643 priority = p->p_asleep.as_priority; 644 if (timo < 0) 645 timo = p->p_asleep.as_timo; 646 647 /* 648 * Install timeout 649 */ 650 651 if (timo) 652 callout_reset(&p->p_slpcallout, timo, endtsleep, p); 653 654 sig = 0; 655 catch = priority & PCATCH; 656 657 if (catch) { 658 p->p_flag |= P_SINTR; 659 mtx_exit(&sched_lock, MTX_SPIN); 660 if ((sig = CURSIG(p))) { 661 mtx_enter(&sched_lock, MTX_SPIN); 662 if (p->p_wchan) 663 unsleep(p); 664 p->p_stat = SRUN; 665 goto resume; 666 } 667 mtx_enter(&sched_lock, MTX_SPIN); 668 if (p->p_wchan == NULL) { 669 catch = 0; 670 goto resume; 671 } 672 } 673 p->p_stat = SSLEEP; 674 p->p_stats->p_ru.ru_nvcsw++; 675 mi_switch(); 676resume: 677 curpriority = p->p_usrpri; 678 679 splx(s); 680 p->p_flag &= ~P_SINTR; 681 if (p->p_flag & P_TIMEOUT) { 682 p->p_flag &= ~P_TIMEOUT; 683 if (sig == 0) { 684#ifdef KTRACE 685 if (KTRPOINT(p, KTR_CSW)) 686 ktrcsw(p->p_tracep, 0, 0); 687#endif 688 rval = EWOULDBLOCK; 689 mtx_exit(&sched_lock, MTX_SPIN); 690 goto out; 691 } 692 } else if (timo) 693 callout_stop(&p->p_slpcallout); 694 mtx_exit(&sched_lock, MTX_SPIN); 695 696 if (catch && (sig != 0 || (sig = CURSIG(p)))) { 697#ifdef KTRACE 698 if (KTRPOINT(p, KTR_CSW)) 699 ktrcsw(p->p_tracep, 0, 0); 700#endif 701 if (SIGISMEMBER(p->p_sigacts->ps_sigintr, sig)) 702 rval = EINTR; 703 else 704 rval = ERESTART; 705 goto out; 706 } 707#ifdef KTRACE 708 if (KTRPOINT(p, KTR_CSW)) 709 ktrcsw(p->p_tracep, 0, 0); 710#endif 711 } else { 712 /* 713 * If as_priority is 0, mawait() has been called without an 714 * intervening asleep(). We are still effectively a NOP, 715 * but we call mi_switch() for safety. 716 */ 717 718 if (p->p_asleep.as_priority == 0) { 719 p->p_stats->p_ru.ru_nvcsw++; 720 mi_switch(); 721 } 722 mtx_exit(&sched_lock, MTX_SPIN); 723 splx(s); 724 } 725 726 /* 727 * clear p_asleep.as_priority as an indication that mawait() has been 728 * called. If mawait() is called again without an intervening asleep(), 729 * mawait() is still effectively a NOP but the above mi_switch() code 730 * is triggered as a safety. 731 */ 732 p->p_asleep.as_priority = 0; 733 734out: 735 PICKUP_GIANT(); 736 if (mtx != NULL) { 737 mtx_enter(mtx, MTX_DEF); 738 WITNESS_RESTORE(mtx, mtx); 739 } 740 return (rval); 741} 742 743/* 744 * Implement timeout for msleep or asleep()/mawait() 745 * 746 * If process hasn't been awakened (wchan non-zero), 747 * set timeout flag and undo the sleep. If proc 748 * is stopped, just unsleep so it will remain stopped. 749 * MP-safe, called without the Giant mutex. 750 */ 751static void 752endtsleep(arg) 753 void *arg; 754{ 755 register struct proc *p; 756 int s; 757 758 p = (struct proc *)arg; 759 CTR4(KTR_PROC, 760 "endtsleep: proc %p (pid %d, %s), schedlock %p", 761 p, p->p_pid, p->p_comm, (void *) sched_lock.mtx_lock); 762 s = splhigh(); 763 mtx_enter(&sched_lock, MTX_SPIN); 764 if (p->p_wchan) { 765 if (p->p_stat == SSLEEP) 766 setrunnable(p); 767 else 768 unsleep(p); 769 p->p_flag |= P_TIMEOUT; 770 } 771 mtx_exit(&sched_lock, MTX_SPIN); 772 splx(s); 773} 774 775/* 776 * Remove a process from its wait queue 777 */ 778void 779unsleep(p) 780 register struct proc *p; 781{ 782 int s; 783 784 s = splhigh(); 785 mtx_enter(&sched_lock, MTX_SPIN); 786 if (p->p_wchan) { 787 TAILQ_REMOVE(&slpque[LOOKUP(p->p_wchan)], p, p_slpq); 788 p->p_wchan = 0; 789 } 790 mtx_exit(&sched_lock, MTX_SPIN); 791 splx(s); 792} 793 794/* 795 * Make all processes sleeping on the specified identifier runnable. 796 */ 797void 798wakeup(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_enter(&sched_lock, MTX_SPIN); 807 qp = &slpque[LOOKUP(ident)]; 808restart: 809 TAILQ_FOREACH(p, qp, p_slpq) { 810 if (p->p_wchan == ident) { 811 TAILQ_REMOVE(qp, p, p_slpq); 812 p->p_wchan = 0; 813 if (p->p_stat == SSLEEP) { 814 /* OPTIMIZED EXPANSION OF setrunnable(p); */ 815 CTR4(KTR_PROC, 816 "wakeup: 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_flag & P_INMEM) { 823 setrunqueue(p); 824 maybe_resched(p); 825 } else { 826 p->p_flag |= P_SWAPINREQ; 827 wakeup((caddr_t)&proc0); 828 } 829 /* END INLINE EXPANSION */ 830 goto restart; 831 } 832 } 833 } 834 mtx_exit(&sched_lock, MTX_SPIN); 835 splx(s); 836} 837 838/* 839 * Make a process sleeping on the specified identifier runnable. 840 * May wake more than one process if a target process is currently 841 * swapped out. 842 */ 843void 844wakeup_one(ident) 845 register void *ident; 846{ 847 register struct slpquehead *qp; 848 register struct proc *p; 849 int s; 850 851 s = splhigh(); 852 mtx_enter(&sched_lock, MTX_SPIN); 853 qp = &slpque[LOOKUP(ident)]; 854 855 TAILQ_FOREACH(p, qp, p_slpq) { 856 if (p->p_wchan == ident) { 857 TAILQ_REMOVE(qp, p, p_slpq); 858 p->p_wchan = 0; 859 if (p->p_stat == SSLEEP) { 860 /* OPTIMIZED EXPANSION OF setrunnable(p); */ 861 CTR4(KTR_PROC, 862 "wakeup1: proc %p (pid %d, %s), schedlock %p", 863 p, p->p_pid, p->p_comm, (void *) sched_lock.mtx_lock); 864 if (p->p_slptime > 1) 865 updatepri(p); 866 p->p_slptime = 0; 867 p->p_stat = SRUN; 868 if (p->p_flag & P_INMEM) { 869 setrunqueue(p); 870 maybe_resched(p); 871 break; 872 } else { 873 p->p_flag |= P_SWAPINREQ; 874 wakeup((caddr_t)&proc0); 875 } 876 /* END INLINE EXPANSION */ 877 } 878 } 879 } 880 mtx_exit(&sched_lock, MTX_SPIN); 881 splx(s); 882} 883 884/* 885 * The machine independent parts of mi_switch(). 886 * Must be called at splstatclock() or higher. 887 */ 888void 889mi_switch() 890{ 891 struct timeval new_switchtime; 892 register struct proc *p = curproc; /* XXX */ 893 register struct rlimit *rlim; 894 int x; 895 896 /* 897 * XXX this spl is almost unnecessary. It is partly to allow for 898 * sloppy callers that don't do it (issignal() via CURSIG() is the 899 * main offender). It is partly to work around a bug in the i386 900 * cpu_switch() (the ipl is not preserved). We ran for years 901 * without it. I think there was only a interrupt latency problem. 902 * The main caller, msleep(), does an splx() a couple of instructions 903 * after calling here. The buggy caller, issignal(), usually calls 904 * here at spl0() and sometimes returns at splhigh(). The process 905 * then runs for a little too long at splhigh(). The ipl gets fixed 906 * when the process returns to user mode (or earlier). 907 * 908 * It would probably be better to always call here at spl0(). Callers 909 * are prepared to give up control to another process, so they must 910 * be prepared to be interrupted. The clock stuff here may not 911 * actually need splstatclock(). 912 */ 913 x = splstatclock(); 914 915 mtx_assert(&sched_lock, MA_OWNED); 916 917#ifdef SIMPLELOCK_DEBUG 918 if (p->p_simple_locks) 919 printf("sleep: holding simple lock\n"); 920#endif 921 /* 922 * Compute the amount of time during which the current 923 * process was running, and add that to its total so far. 924 */ 925 microuptime(&new_switchtime); 926 if (timevalcmp(&new_switchtime, &switchtime, <)) { 927 printf("microuptime() went backwards (%ld.%06ld -> %ld.%06ld)\n", 928 switchtime.tv_sec, switchtime.tv_usec, 929 new_switchtime.tv_sec, new_switchtime.tv_usec); 930 new_switchtime = switchtime; 931 } else { 932 p->p_runtime += (new_switchtime.tv_usec - switchtime.tv_usec) + 933 (new_switchtime.tv_sec - switchtime.tv_sec) * (int64_t)1000000; 934 } 935 936 /* 937 * Check if the process exceeds its cpu resource allocation. 938 * If over max, kill it. 939 * 940 * XXX drop sched_lock, pickup Giant 941 */ 942 if (p->p_stat != SZOMB && p->p_limit->p_cpulimit != RLIM_INFINITY && 943 p->p_runtime > p->p_limit->p_cpulimit) { 944 rlim = &p->p_rlimit[RLIMIT_CPU]; 945 if (p->p_runtime / (rlim_t)1000000 >= rlim->rlim_max) { 946 killproc(p, "exceeded maximum CPU limit"); 947 } else { 948 psignal(p, SIGXCPU); 949 if (rlim->rlim_cur < rlim->rlim_max) { 950 /* XXX: we should make a private copy */ 951 rlim->rlim_cur += 5; 952 } 953 } 954 } 955 956 /* 957 * Pick a new current process and record its start time. 958 */ 959 cnt.v_swtch++; 960 switchtime = new_switchtime; 961 CTR4(KTR_PROC, "mi_switch: old proc %p (pid %d, %s), schedlock %p", 962 p, p->p_pid, p->p_comm, (void *) sched_lock.mtx_lock); 963 cpu_switch(); 964 CTR4(KTR_PROC, "mi_switch: new proc %p (pid %d, %s), schedlock %p", 965 p, p->p_pid, p->p_comm, (void *) sched_lock.mtx_lock); 966 if (switchtime.tv_sec == 0) 967 microuptime(&switchtime); 968 switchticks = ticks; 969 splx(x); 970} 971 972/* 973 * Change process state to be runnable, 974 * placing it on the run queue if it is in memory, 975 * and awakening the swapper if it isn't in memory. 976 */ 977void 978setrunnable(p) 979 register struct proc *p; 980{ 981 register int s; 982 983 s = splhigh(); 984 mtx_enter(&sched_lock, MTX_SPIN); 985 switch (p->p_stat) { 986 case 0: 987 case SRUN: 988 case SZOMB: 989 case SWAIT: 990 default: 991 panic("setrunnable"); 992 case SSTOP: 993 case SSLEEP: 994 unsleep(p); /* e.g. when sending signals */ 995 break; 996 997 case SIDL: 998 break; 999 } 1000 p->p_stat = SRUN; 1001 if (p->p_flag & P_INMEM) 1002 setrunqueue(p); 1003 splx(s); 1004 if (p->p_slptime > 1) 1005 updatepri(p); 1006 p->p_slptime = 0; 1007 if ((p->p_flag & P_INMEM) == 0) { 1008 p->p_flag |= P_SWAPINREQ; 1009 wakeup((caddr_t)&proc0); 1010 } 1011 else 1012 maybe_resched(p); 1013 mtx_exit(&sched_lock, MTX_SPIN); 1014} 1015 1016/* 1017 * Compute the priority of a process when running in user mode. 1018 * Arrange to reschedule if the resulting priority is better 1019 * than that of the current process. 1020 */ 1021void 1022resetpriority(p) 1023 register struct proc *p; 1024{ 1025 register unsigned int newpriority; 1026 1027 mtx_enter(&sched_lock, MTX_SPIN); 1028 if (p->p_rtprio.type == RTP_PRIO_NORMAL) { 1029 newpriority = PUSER + p->p_estcpu / INVERSE_ESTCPU_WEIGHT + 1030 NICE_WEIGHT * (p->p_nice - PRIO_MIN); 1031 newpriority = min(newpriority, MAXPRI); 1032 p->p_usrpri = newpriority; 1033 } 1034 maybe_resched(p); 1035 mtx_exit(&sched_lock, MTX_SPIN); 1036} 1037 1038/* ARGSUSED */ 1039static void 1040sched_setup(dummy) 1041 void *dummy; 1042{ 1043 1044 callout_init(&schedcpu_callout, 1); 1045 callout_init(&roundrobin_callout, 0); 1046 1047 /* Kick off timeout driven events by calling first time. */ 1048 roundrobin(NULL); 1049 schedcpu(NULL); 1050} 1051 1052/* 1053 * We adjust the priority of the current process. The priority of 1054 * a process gets worse as it accumulates CPU time. The cpu usage 1055 * estimator (p_estcpu) is increased here. resetpriority() will 1056 * compute a different priority each time p_estcpu increases by 1057 * INVERSE_ESTCPU_WEIGHT 1058 * (until MAXPRI is reached). The cpu usage estimator ramps up 1059 * quite quickly when the process is running (linearly), and decays 1060 * away exponentially, at a rate which is proportionally slower when 1061 * the system is busy. The basic principle is that the system will 1062 * 90% forget that the process used a lot of CPU time in 5 * loadav 1063 * seconds. This causes the system to favor processes which haven't 1064 * run much recently, and to round-robin among other processes. 1065 */ 1066void 1067schedclock(p) 1068 struct proc *p; 1069{ 1070 1071 p->p_cpticks++; 1072 p->p_estcpu = ESTCPULIM(p->p_estcpu + 1); 1073 if ((p->p_estcpu % INVERSE_ESTCPU_WEIGHT) == 0) { 1074 resetpriority(p); 1075 if (p->p_priority >= PUSER) 1076 p->p_priority = p->p_usrpri; 1077 } 1078} 1079