kern_synch.c revision 68862
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 68862 2000-11-17 18:09:18Z 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 int curpriority_cmp __P((struct proc *p)); 74static void endtsleep __P((void *)); 75static void maybe_resched __P((struct proc *chk)); 76static void roundrobin __P((void *arg)); 77static void schedcpu __P((void *arg)); 78static void updatepri __P((struct proc *p)); 79 80static int 81sysctl_kern_quantum(SYSCTL_HANDLER_ARGS) 82{ 83 int error, new_val; 84 85 new_val = sched_quantum * tick; 86 error = sysctl_handle_int(oidp, &new_val, 0, req); 87 if (error != 0 || req->newptr == NULL) 88 return (error); 89 if (new_val < tick) 90 return (EINVAL); 91 sched_quantum = new_val / tick; 92 hogticks = 2 * sched_quantum; 93 return (0); 94} 95 96SYSCTL_PROC(_kern, OID_AUTO, quantum, CTLTYPE_INT|CTLFLAG_RW, 97 0, sizeof sched_quantum, sysctl_kern_quantum, "I", ""); 98 99/*- 100 * Compare priorities. Return: 101 * <0: priority of p < current priority 102 * 0: priority of p == current priority 103 * >0: priority of p > current priority 104 * The priorities are the normal priorities or the normal realtime priorities 105 * if p is on the same scheduler as curproc. Otherwise the process on the 106 * more realtimeish scheduler has lowest priority. As usual, a higher 107 * priority really means a lower priority. 108 */ 109static int 110curpriority_cmp(p) 111 struct proc *p; 112{ 113 int c_class, p_class; 114 115 c_class = RTP_PRIO_BASE(curproc->p_rtprio.type); 116 p_class = RTP_PRIO_BASE(p->p_rtprio.type); 117 if (p_class != c_class) 118 return (p_class - c_class); 119 if (p_class == RTP_PRIO_NORMAL) 120 return (((int)p->p_priority - (int)curpriority) / PPQ); 121 return ((int)p->p_rtprio.prio - (int)curproc->p_rtprio.prio); 122} 123 124/* 125 * Arrange to reschedule if necessary, taking the priorities and 126 * schedulers into account. 127 */ 128static void 129maybe_resched(chk) 130 struct proc *chk; 131{ 132 struct proc *p = curproc; /* XXX */ 133 134 /* 135 * XXX idle scheduler still broken because proccess stays on idle 136 * scheduler during waits (such as when getting FS locks). If a 137 * standard process becomes runaway cpu-bound, the system can lockup 138 * due to idle-scheduler processes in wakeup never getting any cpu. 139 */ 140 if (p == idleproc) { 141#if 0 142 need_resched(); 143#endif 144 } else if (chk == p) { 145 /* We may need to yield if our priority has been raised. */ 146 if (curpriority_cmp(chk) > 0) 147 need_resched(); 148 } else if (curpriority_cmp(chk) < 0) 149 need_resched(); 150} 151 152int 153roundrobin_interval(void) 154{ 155 return (sched_quantum); 156} 157 158/* 159 * Force switch among equal priority processes every 100ms. 160 */ 161/* ARGSUSED */ 162static void 163roundrobin(arg) 164 void *arg; 165{ 166#ifndef SMP 167 struct proc *p = curproc; /* XXX */ 168#endif 169 170#ifdef SMP 171 need_resched(); 172 forward_roundrobin(); 173#else 174 if (p == idleproc || RTP_PRIO_NEED_RR(p->p_rtprio.type)) 175 need_resched(); 176#endif 177 178 timeout(roundrobin, NULL, sched_quantum); 179} 180 181/* 182 * Constants for digital decay and forget: 183 * 90% of (p_estcpu) usage in 5 * loadav time 184 * 95% of (p_pctcpu) usage in 60 seconds (load insensitive) 185 * Note that, as ps(1) mentions, this can let percentages 186 * total over 100% (I've seen 137.9% for 3 processes). 187 * 188 * Note that schedclock() updates p_estcpu and p_cpticks asynchronously. 189 * 190 * We wish to decay away 90% of p_estcpu in (5 * loadavg) seconds. 191 * That is, the system wants to compute a value of decay such 192 * that the following for loop: 193 * for (i = 0; i < (5 * loadavg); i++) 194 * p_estcpu *= decay; 195 * will compute 196 * p_estcpu *= 0.1; 197 * for all values of loadavg: 198 * 199 * Mathematically this loop can be expressed by saying: 200 * decay ** (5 * loadavg) ~= .1 201 * 202 * The system computes decay as: 203 * decay = (2 * loadavg) / (2 * loadavg + 1) 204 * 205 * We wish to prove that the system's computation of decay 206 * will always fulfill the equation: 207 * decay ** (5 * loadavg) ~= .1 208 * 209 * If we compute b as: 210 * b = 2 * loadavg 211 * then 212 * decay = b / (b + 1) 213 * 214 * We now need to prove two things: 215 * 1) Given factor ** (5 * loadavg) ~= .1, prove factor == b/(b+1) 216 * 2) Given b/(b+1) ** power ~= .1, prove power == (5 * loadavg) 217 * 218 * Facts: 219 * For x close to zero, exp(x) =~ 1 + x, since 220 * exp(x) = 0! + x**1/1! + x**2/2! + ... . 221 * therefore exp(-1/b) =~ 1 - (1/b) = (b-1)/b. 222 * For x close to zero, ln(1+x) =~ x, since 223 * ln(1+x) = x - x**2/2 + x**3/3 - ... -1 < x < 1 224 * therefore ln(b/(b+1)) = ln(1 - 1/(b+1)) =~ -1/(b+1). 225 * ln(.1) =~ -2.30 226 * 227 * Proof of (1): 228 * Solve (factor)**(power) =~ .1 given power (5*loadav): 229 * solving for factor, 230 * ln(factor) =~ (-2.30/5*loadav), or 231 * factor =~ exp(-1/((5/2.30)*loadav)) =~ exp(-1/(2*loadav)) = 232 * exp(-1/b) =~ (b-1)/b =~ b/(b+1). QED 233 * 234 * Proof of (2): 235 * Solve (factor)**(power) =~ .1 given factor == (b/(b+1)): 236 * solving for power, 237 * power*ln(b/(b+1)) =~ -2.30, or 238 * power =~ 2.3 * (b + 1) = 4.6*loadav + 2.3 =~ 5*loadav. QED 239 * 240 * Actual power values for the implemented algorithm are as follows: 241 * loadav: 1 2 3 4 242 * power: 5.68 10.32 14.94 19.55 243 */ 244 245/* calculations for digital decay to forget 90% of usage in 5*loadav sec */ 246#define loadfactor(loadav) (2 * (loadav)) 247#define decay_cpu(loadfac, cpu) (((loadfac) * (cpu)) / ((loadfac) + FSCALE)) 248 249/* decay 95% of `p_pctcpu' in 60 seconds; see CCPU_SHIFT before changing */ 250static fixpt_t ccpu = 0.95122942450071400909 * FSCALE; /* exp(-1/20) */ 251SYSCTL_INT(_kern, OID_AUTO, ccpu, CTLFLAG_RD, &ccpu, 0, ""); 252 253/* kernel uses `FSCALE', userland (SHOULD) use kern.fscale */ 254static int fscale __unused = FSCALE; 255SYSCTL_INT(_kern, OID_AUTO, fscale, CTLFLAG_RD, 0, FSCALE, ""); 256 257/* 258 * If `ccpu' is not equal to `exp(-1/20)' and you still want to use the 259 * faster/more-accurate formula, you'll have to estimate CCPU_SHIFT below 260 * and possibly adjust FSHIFT in "param.h" so that (FSHIFT >= CCPU_SHIFT). 261 * 262 * To estimate CCPU_SHIFT for exp(-1/20), the following formula was used: 263 * 1 - exp(-1/20) ~= 0.0487 ~= 0.0488 == 1 (fixed pt, *11* bits). 264 * 265 * If you don't want to bother with the faster/more-accurate formula, you 266 * can set CCPU_SHIFT to (FSHIFT + 1) which will use a slower/less-accurate 267 * (more general) method of calculating the %age of CPU used by a process. 268 */ 269#define CCPU_SHIFT 11 270 271/* 272 * Recompute process priorities, every hz ticks. 273 */ 274/* ARGSUSED */ 275static void 276schedcpu(arg) 277 void *arg; 278{ 279 register fixpt_t loadfac = loadfactor(averunnable.ldavg[0]); 280 register struct proc *p; 281 register int realstathz, s; 282 283 realstathz = stathz ? stathz : hz; 284 LIST_FOREACH(p, &allproc, p_list) { 285 /* 286 * Increment time in/out of memory and sleep time 287 * (if sleeping). We ignore overflow; with 16-bit int's 288 * (remember them?) overflow takes 45 days. 289 if (p->p_stat == SWAIT) 290 continue; 291 */ 292 mtx_enter(&sched_lock, MTX_SPIN); 293 p->p_swtime++; 294 if (p->p_stat == SSLEEP || p->p_stat == SSTOP) 295 p->p_slptime++; 296 p->p_pctcpu = (p->p_pctcpu * ccpu) >> FSHIFT; 297 /* 298 * If the process has slept the entire second, 299 * stop recalculating its priority until it wakes up. 300 */ 301 if (p->p_slptime > 1) { 302 mtx_exit(&sched_lock, MTX_SPIN); 303 continue; 304 } 305 306 /* 307 * prevent state changes and protect run queue 308 */ 309 s = splhigh(); 310 311 /* 312 * p_pctcpu is only for ps. 313 */ 314#if (FSHIFT >= CCPU_SHIFT) 315 p->p_pctcpu += (realstathz == 100)? 316 ((fixpt_t) p->p_cpticks) << (FSHIFT - CCPU_SHIFT): 317 100 * (((fixpt_t) p->p_cpticks) 318 << (FSHIFT - CCPU_SHIFT)) / realstathz; 319#else 320 p->p_pctcpu += ((FSCALE - ccpu) * 321 (p->p_cpticks * FSCALE / realstathz)) >> FSHIFT; 322#endif 323 p->p_cpticks = 0; 324 p->p_estcpu = decay_cpu(loadfac, p->p_estcpu); 325 resetpriority(p); 326 if (p->p_priority >= PUSER) { 327 if ((p != curproc) && 328#ifdef SMP 329 p->p_oncpu == 0xff && /* idle */ 330#endif 331 p->p_stat == SRUN && 332 (p->p_flag & P_INMEM) && 333 (p->p_priority / PPQ) != (p->p_usrpri / PPQ)) { 334 remrunqueue(p); 335 p->p_priority = p->p_usrpri; 336 setrunqueue(p); 337 } else 338 p->p_priority = p->p_usrpri; 339 } 340 mtx_exit(&sched_lock, MTX_SPIN); 341 splx(s); 342 } 343 vmmeter(); 344 wakeup((caddr_t)&lbolt); 345 timeout(schedcpu, (void *)0, hz); 346} 347 348/* 349 * Recalculate the priority of a process after it has slept for a while. 350 * For all load averages >= 1 and max p_estcpu of 255, sleeping for at 351 * least six times the loadfactor will decay p_estcpu to zero. 352 */ 353static void 354updatepri(p) 355 register struct proc *p; 356{ 357 register unsigned int newcpu = p->p_estcpu; 358 register fixpt_t loadfac = loadfactor(averunnable.ldavg[0]); 359 360 if (p->p_slptime > 5 * loadfac) 361 p->p_estcpu = 0; 362 else { 363 p->p_slptime--; /* the first time was done in schedcpu */ 364 while (newcpu && --p->p_slptime) 365 newcpu = decay_cpu(loadfac, newcpu); 366 p->p_estcpu = newcpu; 367 } 368 resetpriority(p); 369} 370 371/* 372 * We're only looking at 7 bits of the address; everything is 373 * aligned to 4, lots of things are aligned to greater powers 374 * of 2. Shift right by 8, i.e. drop the bottom 256 worth. 375 */ 376#define TABLESIZE 128 377static TAILQ_HEAD(slpquehead, proc) slpque[TABLESIZE]; 378#define LOOKUP(x) (((intptr_t)(x) >> 8) & (TABLESIZE - 1)) 379 380void 381sleepinit(void) 382{ 383 int i; 384 385 sched_quantum = hz/10; 386 hogticks = 2 * sched_quantum; 387 for (i = 0; i < TABLESIZE; i++) 388 TAILQ_INIT(&slpque[i]); 389} 390 391/* 392 * General sleep call. Suspends the current process until a wakeup is 393 * performed on the specified identifier. The process will then be made 394 * runnable with the specified priority. Sleeps at most timo/hz seconds 395 * (0 means no timeout). If pri includes PCATCH flag, signals are checked 396 * before and after sleeping, else signals are not checked. Returns 0 if 397 * awakened, EWOULDBLOCK if the timeout expires. If PCATCH is set and a 398 * signal needs to be delivered, ERESTART is returned if the current system 399 * call should be restarted if possible, and EINTR is returned if the system 400 * call should be interrupted by the signal (return EINTR). 401 * 402 * The mutex argument is exited before the caller is suspended, and 403 * entered before msleep returns. If priority includes the PDROP 404 * flag the mutex is not entered before returning. 405 */ 406int 407msleep(ident, mtx, priority, wmesg, timo) 408 void *ident; 409 struct mtx *mtx; 410 int priority, timo; 411 const char *wmesg; 412{ 413 struct proc *p = curproc; 414 int s, sig, catch = priority & PCATCH; 415 struct callout_handle thandle; 416 int rval = 0; 417 WITNESS_SAVE_DECL(mtx); 418 419#ifdef KTRACE 420 if (p && KTRPOINT(p, KTR_CSW)) 421 ktrcsw(p->p_tracep, 1, 0); 422#endif 423 WITNESS_SLEEP(0, mtx); 424 mtx_enter(&sched_lock, MTX_SPIN); 425 DROP_GIANT_NOSWITCH(); 426 427 if (mtx != NULL) { 428 mtx_assert(mtx, MA_OWNED | MA_NOTRECURSED); 429 WITNESS_SAVE(mtx, mtx); 430 mtx_exit(mtx, MTX_DEF | MTX_NOSWITCH); 431 if (priority & PDROP) 432 mtx = NULL; 433 } 434 435 s = splhigh(); 436 if (cold || panicstr) { 437 /* 438 * After a panic, or during autoconfiguration, 439 * just give interrupts a chance, then just return; 440 * don't run any other procs or panic below, 441 * in case this is the idle process and already asleep. 442 */ 443 mtx_exit(&sched_lock, MTX_SPIN); 444 splx(s); 445 return (0); 446 } 447 448 KASSERT(p != NULL, ("msleep1")); 449 KASSERT(ident != NULL && p->p_stat == SRUN, ("msleep")); 450 /* 451 * Process may be sitting on a slpque if asleep() was called, remove 452 * it before re-adding. 453 */ 454 if (p->p_wchan != NULL) 455 unsleep(p); 456 457 p->p_wchan = ident; 458 p->p_wmesg = wmesg; 459 p->p_slptime = 0; 460 p->p_priority = priority & PRIMASK; 461 p->p_nativepri = p->p_priority; 462 CTR4(KTR_PROC, "msleep: proc %p (pid %d, %s), schedlock %p", 463 p, p->p_pid, p->p_comm, (void *) sched_lock.mtx_lock); 464 TAILQ_INSERT_TAIL(&slpque[LOOKUP(ident)], p, p_slpq); 465 if (timo) 466 thandle = timeout(endtsleep, (void *)p, timo); 467 /* 468 * We put ourselves on the sleep queue and start our timeout 469 * before calling CURSIG, as we could stop there, and a wakeup 470 * or a SIGCONT (or both) could occur while we were stopped. 471 * A SIGCONT would cause us to be marked as SSLEEP 472 * without resuming us, thus we must be ready for sleep 473 * when CURSIG is called. If the wakeup happens while we're 474 * stopped, p->p_wchan will be 0 upon return from CURSIG. 475 */ 476 if (catch) { 477 CTR4(KTR_PROC, 478 "msleep caught: proc %p (pid %d, %s), schedlock %p", 479 p, p->p_pid, p->p_comm, (void *) sched_lock.mtx_lock); 480 p->p_flag |= P_SINTR; 481 mtx_exit(&sched_lock, MTX_SPIN); 482 if ((sig = CURSIG(p))) { 483 mtx_enter(&sched_lock, MTX_SPIN); 484 if (p->p_wchan) 485 unsleep(p); 486 p->p_stat = SRUN; 487 goto resume; 488 } 489 mtx_enter(&sched_lock, MTX_SPIN); 490 if (p->p_wchan == 0) { 491 catch = 0; 492 goto resume; 493 } 494 } else 495 sig = 0; 496 p->p_stat = SSLEEP; 497 p->p_stats->p_ru.ru_nvcsw++; 498 mi_switch(); 499 CTR4(KTR_PROC, 500 "msleep resume: proc %p (pid %d, %s), schedlock %p", 501 p, p->p_pid, p->p_comm, (void *) sched_lock.mtx_lock); 502resume: 503 curpriority = p->p_usrpri; 504 splx(s); 505 p->p_flag &= ~P_SINTR; 506 if (p->p_flag & P_TIMEOUT) { 507 p->p_flag &= ~P_TIMEOUT; 508 if (sig == 0) { 509#ifdef KTRACE 510 if (KTRPOINT(p, KTR_CSW)) 511 ktrcsw(p->p_tracep, 0, 0); 512#endif 513 rval = EWOULDBLOCK; 514 mtx_exit(&sched_lock, MTX_SPIN); 515 goto out; 516 } 517 } else if (timo) 518 untimeout(endtsleep, (void *)p, thandle); 519 mtx_exit(&sched_lock, MTX_SPIN); 520 521 if (catch && (sig != 0 || (sig = CURSIG(p)))) { 522#ifdef KTRACE 523 if (KTRPOINT(p, KTR_CSW)) 524 ktrcsw(p->p_tracep, 0, 0); 525#endif 526 if (SIGISMEMBER(p->p_sigacts->ps_sigintr, sig)) 527 rval = EINTR; 528 else 529 rval = ERESTART; 530 goto out; 531 } 532out: 533#ifdef KTRACE 534 if (KTRPOINT(p, KTR_CSW)) 535 ktrcsw(p->p_tracep, 0, 0); 536#endif 537 PICKUP_GIANT(); 538 if (mtx != NULL) { 539 mtx_enter(mtx, MTX_DEF); 540 WITNESS_RESTORE(mtx, mtx); 541 } 542 return (rval); 543} 544 545/* 546 * asleep() - async sleep call. Place process on wait queue and return 547 * immediately without blocking. The process stays runnable until mawait() 548 * is called. If ident is NULL, remove process from wait queue if it is still 549 * on one. 550 * 551 * Only the most recent sleep condition is effective when making successive 552 * calls to asleep() or when calling msleep(). 553 * 554 * The timeout, if any, is not initiated until mawait() is called. The sleep 555 * priority, signal, and timeout is specified in the asleep() call but may be 556 * overriden in the mawait() call. 557 * 558 * <<<<<<<< EXPERIMENTAL, UNTESTED >>>>>>>>>> 559 */ 560 561int 562asleep(void *ident, int priority, const char *wmesg, int timo) 563{ 564 struct proc *p = curproc; 565 int s; 566 567 /* 568 * obtain sched_lock while manipulating sleep structures and slpque. 569 * 570 * Remove preexisting wait condition (if any) and place process 571 * on appropriate slpque, but do not put process to sleep. 572 */ 573 574 s = splhigh(); 575 mtx_enter(&sched_lock, MTX_SPIN); 576 577 if (p->p_wchan != NULL) 578 unsleep(p); 579 580 if (ident) { 581 p->p_wchan = ident; 582 p->p_wmesg = wmesg; 583 p->p_slptime = 0; 584 p->p_asleep.as_priority = priority; 585 p->p_asleep.as_timo = timo; 586 TAILQ_INSERT_TAIL(&slpque[LOOKUP(ident)], p, p_slpq); 587 } 588 589 mtx_exit(&sched_lock, MTX_SPIN); 590 splx(s); 591 592 return(0); 593} 594 595/* 596 * mawait() - wait for async condition to occur. The process blocks until 597 * wakeup() is called on the most recent asleep() address. If wakeup is called 598 * prior to mawait(), mawait() winds up being a NOP. 599 * 600 * If mawait() is called more then once (without an intervening asleep() call), 601 * mawait() is still effectively a NOP but it calls mi_switch() to give other 602 * processes some cpu before returning. The process is left runnable. 603 * 604 * <<<<<<<< EXPERIMENTAL, UNTESTED >>>>>>>>>> 605 */ 606 607int 608mawait(struct mtx *mtx, int priority, int timo) 609{ 610 struct proc *p = curproc; 611 int rval = 0; 612 int s; 613 WITNESS_SAVE_DECL(mtx); 614 615 WITNESS_SLEEP(0, mtx); 616 mtx_enter(&sched_lock, MTX_SPIN); 617 DROP_GIANT_NOSWITCH(); 618 if (mtx != NULL) { 619 mtx_assert(mtx, MA_OWNED | MA_NOTRECURSED); 620 WITNESS_SAVE(mtx, mtx); 621 mtx_exit(mtx, MTX_DEF | MTX_NOSWITCH); 622 if (priority & PDROP) 623 mtx = NULL; 624 } 625 626 s = splhigh(); 627 628 if (p->p_wchan != NULL) { 629 struct callout_handle thandle; 630 int sig; 631 int catch; 632 633 /* 634 * The call to mawait() can override defaults specified in 635 * the original asleep(). 636 */ 637 if (priority < 0) 638 priority = p->p_asleep.as_priority; 639 if (timo < 0) 640 timo = p->p_asleep.as_timo; 641 642 /* 643 * Install timeout 644 */ 645 646 if (timo) 647 thandle = timeout(endtsleep, (void *)p, timo); 648 649 sig = 0; 650 catch = priority & PCATCH; 651 652 if (catch) { 653 p->p_flag |= P_SINTR; 654 mtx_exit(&sched_lock, MTX_SPIN); 655 if ((sig = CURSIG(p))) { 656 mtx_enter(&sched_lock, MTX_SPIN); 657 if (p->p_wchan) 658 unsleep(p); 659 p->p_stat = SRUN; 660 goto resume; 661 } 662 mtx_enter(&sched_lock, MTX_SPIN); 663 if (p->p_wchan == NULL) { 664 catch = 0; 665 goto resume; 666 } 667 } 668 p->p_stat = SSLEEP; 669 p->p_stats->p_ru.ru_nvcsw++; 670 mi_switch(); 671resume: 672 curpriority = p->p_usrpri; 673 674 splx(s); 675 p->p_flag &= ~P_SINTR; 676 if (p->p_flag & P_TIMEOUT) { 677 p->p_flag &= ~P_TIMEOUT; 678 if (sig == 0) { 679#ifdef KTRACE 680 if (KTRPOINT(p, KTR_CSW)) 681 ktrcsw(p->p_tracep, 0, 0); 682#endif 683 rval = EWOULDBLOCK; 684 mtx_exit(&sched_lock, MTX_SPIN); 685 goto out; 686 } 687 } else if (timo) 688 untimeout(endtsleep, (void *)p, thandle); 689 mtx_exit(&sched_lock, MTX_SPIN); 690 691 if (catch && (sig != 0 || (sig = CURSIG(p)))) { 692#ifdef KTRACE 693 if (KTRPOINT(p, KTR_CSW)) 694 ktrcsw(p->p_tracep, 0, 0); 695#endif 696 if (SIGISMEMBER(p->p_sigacts->ps_sigintr, sig)) 697 rval = EINTR; 698 else 699 rval = ERESTART; 700 goto out; 701 } 702#ifdef KTRACE 703 if (KTRPOINT(p, KTR_CSW)) 704 ktrcsw(p->p_tracep, 0, 0); 705#endif 706 } else { 707 /* 708 * If as_priority is 0, mawait() has been called without an 709 * intervening asleep(). We are still effectively a NOP, 710 * but we call mi_switch() for safety. 711 */ 712 713 if (p->p_asleep.as_priority == 0) { 714 p->p_stats->p_ru.ru_nvcsw++; 715 mi_switch(); 716 } 717 mtx_exit(&sched_lock, MTX_SPIN); 718 splx(s); 719 } 720 721 /* 722 * clear p_asleep.as_priority as an indication that mawait() has been 723 * called. If mawait() is called again without an intervening asleep(), 724 * mawait() is still effectively a NOP but the above mi_switch() code 725 * is triggered as a safety. 726 */ 727 p->p_asleep.as_priority = 0; 728 729out: 730 PICKUP_GIANT(); 731 if (mtx != NULL) { 732 mtx_enter(mtx, MTX_DEF); 733 WITNESS_RESTORE(mtx, mtx); 734 } 735 return (rval); 736} 737 738/* 739 * Implement timeout for msleep or asleep()/mawait() 740 * 741 * If process hasn't been awakened (wchan non-zero), 742 * set timeout flag and undo the sleep. If proc 743 * is stopped, just unsleep so it will remain stopped. 744 */ 745static void 746endtsleep(arg) 747 void *arg; 748{ 749 register struct proc *p; 750 int s; 751 752 p = (struct proc *)arg; 753 CTR4(KTR_PROC, 754 "endtsleep: proc %p (pid %d, %s), schedlock %p", 755 p, p->p_pid, p->p_comm, (void *) sched_lock.mtx_lock); 756 s = splhigh(); 757 mtx_enter(&sched_lock, MTX_SPIN); 758 if (p->p_wchan) { 759 if (p->p_stat == SSLEEP) 760 setrunnable(p); 761 else 762 unsleep(p); 763 p->p_flag |= P_TIMEOUT; 764 } 765 mtx_exit(&sched_lock, MTX_SPIN); 766 splx(s); 767} 768 769/* 770 * Remove a process from its wait queue 771 */ 772void 773unsleep(p) 774 register struct proc *p; 775{ 776 int s; 777 778 s = splhigh(); 779 mtx_enter(&sched_lock, MTX_SPIN); 780 if (p->p_wchan) { 781 TAILQ_REMOVE(&slpque[LOOKUP(p->p_wchan)], p, p_slpq); 782 p->p_wchan = 0; 783 } 784 mtx_exit(&sched_lock, MTX_SPIN); 785 splx(s); 786} 787 788/* 789 * Make all processes sleeping on the specified identifier runnable. 790 */ 791void 792wakeup(ident) 793 register void *ident; 794{ 795 register struct slpquehead *qp; 796 register struct proc *p; 797 int s; 798 799 s = splhigh(); 800 mtx_enter(&sched_lock, MTX_SPIN); 801 qp = &slpque[LOOKUP(ident)]; 802restart: 803 TAILQ_FOREACH(p, qp, p_slpq) { 804 if (p->p_wchan == ident) { 805 TAILQ_REMOVE(qp, p, p_slpq); 806 p->p_wchan = 0; 807 if (p->p_stat == SSLEEP) { 808 /* OPTIMIZED EXPANSION OF setrunnable(p); */ 809 CTR4(KTR_PROC, 810 "wakeup: proc %p (pid %d, %s), schedlock %p", 811 p, p->p_pid, p->p_comm, (void *) sched_lock.mtx_lock); 812 if (p->p_slptime > 1) 813 updatepri(p); 814 p->p_slptime = 0; 815 p->p_stat = SRUN; 816 if (p->p_flag & P_INMEM) { 817 setrunqueue(p); 818 maybe_resched(p); 819 } else { 820 p->p_flag |= P_SWAPINREQ; 821 wakeup((caddr_t)&proc0); 822 } 823 /* END INLINE EXPANSION */ 824 goto restart; 825 } 826 } 827 } 828 mtx_exit(&sched_lock, MTX_SPIN); 829 splx(s); 830} 831 832/* 833 * Make a process sleeping on the specified identifier runnable. 834 * May wake more than one process if a target process is currently 835 * swapped out. 836 */ 837void 838wakeup_one(ident) 839 register void *ident; 840{ 841 register struct slpquehead *qp; 842 register struct proc *p; 843 int s; 844 845 s = splhigh(); 846 mtx_enter(&sched_lock, MTX_SPIN); 847 qp = &slpque[LOOKUP(ident)]; 848 849 TAILQ_FOREACH(p, qp, p_slpq) { 850 if (p->p_wchan == ident) { 851 TAILQ_REMOVE(qp, p, p_slpq); 852 p->p_wchan = 0; 853 if (p->p_stat == SSLEEP) { 854 /* OPTIMIZED EXPANSION OF setrunnable(p); */ 855 CTR4(KTR_PROC, 856 "wakeup1: proc %p (pid %d, %s), schedlock %p", 857 p, p->p_pid, p->p_comm, (void *) sched_lock.mtx_lock); 858 if (p->p_slptime > 1) 859 updatepri(p); 860 p->p_slptime = 0; 861 p->p_stat = SRUN; 862 if (p->p_flag & P_INMEM) { 863 setrunqueue(p); 864 maybe_resched(p); 865 break; 866 } else { 867 p->p_flag |= P_SWAPINREQ; 868 wakeup((caddr_t)&proc0); 869 } 870 /* END INLINE EXPANSION */ 871 } 872 } 873 } 874 mtx_exit(&sched_lock, MTX_SPIN); 875 splx(s); 876} 877 878/* 879 * The machine independent parts of mi_switch(). 880 * Must be called at splstatclock() or higher. 881 */ 882void 883mi_switch() 884{ 885 struct timeval new_switchtime; 886 register struct proc *p = curproc; /* XXX */ 887 register struct rlimit *rlim; 888 int x; 889 890 /* 891 * XXX this spl is almost unnecessary. It is partly to allow for 892 * sloppy callers that don't do it (issignal() via CURSIG() is the 893 * main offender). It is partly to work around a bug in the i386 894 * cpu_switch() (the ipl is not preserved). We ran for years 895 * without it. I think there was only a interrupt latency problem. 896 * The main caller, msleep(), does an splx() a couple of instructions 897 * after calling here. The buggy caller, issignal(), usually calls 898 * here at spl0() and sometimes returns at splhigh(). The process 899 * then runs for a little too long at splhigh(). The ipl gets fixed 900 * when the process returns to user mode (or earlier). 901 * 902 * It would probably be better to always call here at spl0(). Callers 903 * are prepared to give up control to another process, so they must 904 * be prepared to be interrupted. The clock stuff here may not 905 * actually need splstatclock(). 906 */ 907 x = splstatclock(); 908 909 mtx_assert(&sched_lock, MA_OWNED); 910 911#ifdef SIMPLELOCK_DEBUG 912 if (p->p_simple_locks) 913 printf("sleep: holding simple lock\n"); 914#endif 915 /* 916 * Compute the amount of time during which the current 917 * process was running, and add that to its total so far. 918 */ 919 microuptime(&new_switchtime); 920 if (timevalcmp(&new_switchtime, &switchtime, <)) { 921 printf("microuptime() went backwards (%ld.%06ld -> %ld.%06ld)\n", 922 switchtime.tv_sec, switchtime.tv_usec, 923 new_switchtime.tv_sec, new_switchtime.tv_usec); 924 new_switchtime = switchtime; 925 } else { 926 p->p_runtime += (new_switchtime.tv_usec - switchtime.tv_usec) + 927 (new_switchtime.tv_sec - switchtime.tv_sec) * (int64_t)1000000; 928 } 929 930 /* 931 * Check if the process exceeds its cpu resource allocation. 932 * If over max, kill it. 933 * 934 * XXX drop sched_lock, pickup Giant 935 */ 936 if (p->p_stat != SZOMB && p->p_limit->p_cpulimit != RLIM_INFINITY && 937 p->p_runtime > p->p_limit->p_cpulimit) { 938 rlim = &p->p_rlimit[RLIMIT_CPU]; 939 if (p->p_runtime / (rlim_t)1000000 >= rlim->rlim_max) { 940 killproc(p, "exceeded maximum CPU limit"); 941 } else { 942 psignal(p, SIGXCPU); 943 if (rlim->rlim_cur < rlim->rlim_max) { 944 /* XXX: we should make a private copy */ 945 rlim->rlim_cur += 5; 946 } 947 } 948 } 949 950 /* 951 * Pick a new current process and record its start time. 952 */ 953 cnt.v_swtch++; 954 switchtime = new_switchtime; 955 CTR4(KTR_PROC, "mi_switch: old proc %p (pid %d, %s), schedlock %p", 956 p, p->p_pid, p->p_comm, (void *) sched_lock.mtx_lock); 957 cpu_switch(); 958 CTR4(KTR_PROC, "mi_switch: new proc %p (pid %d, %s), schedlock %p", 959 p, p->p_pid, p->p_comm, (void *) sched_lock.mtx_lock); 960 if (switchtime.tv_sec == 0) 961 microuptime(&switchtime); 962 switchticks = ticks; 963 splx(x); 964} 965 966/* 967 * Change process state to be runnable, 968 * placing it on the run queue if it is in memory, 969 * and awakening the swapper if it isn't in memory. 970 */ 971void 972setrunnable(p) 973 register struct proc *p; 974{ 975 register int s; 976 977 s = splhigh(); 978 mtx_enter(&sched_lock, MTX_SPIN); 979 switch (p->p_stat) { 980 case 0: 981 case SRUN: 982 case SZOMB: 983 case SWAIT: 984 default: 985 panic("setrunnable"); 986 case SSTOP: 987 case SSLEEP: 988 unsleep(p); /* e.g. when sending signals */ 989 break; 990 991 case SIDL: 992 break; 993 } 994 p->p_stat = SRUN; 995 if (p->p_flag & P_INMEM) 996 setrunqueue(p); 997 splx(s); 998 if (p->p_slptime > 1) 999 updatepri(p); 1000 p->p_slptime = 0; 1001 if ((p->p_flag & P_INMEM) == 0) { 1002 p->p_flag |= P_SWAPINREQ; 1003 wakeup((caddr_t)&proc0); 1004 } 1005 else 1006 maybe_resched(p); 1007 mtx_exit(&sched_lock, MTX_SPIN); 1008} 1009 1010/* 1011 * Compute the priority of a process when running in user mode. 1012 * Arrange to reschedule if the resulting priority is better 1013 * than that of the current process. 1014 */ 1015void 1016resetpriority(p) 1017 register struct proc *p; 1018{ 1019 register unsigned int newpriority; 1020 1021 mtx_enter(&sched_lock, MTX_SPIN); 1022 if (p->p_rtprio.type == RTP_PRIO_NORMAL) { 1023 newpriority = PUSER + p->p_estcpu / INVERSE_ESTCPU_WEIGHT + 1024 NICE_WEIGHT * (p->p_nice - PRIO_MIN); 1025 newpriority = min(newpriority, MAXPRI); 1026 p->p_usrpri = newpriority; 1027 } 1028 maybe_resched(p); 1029 mtx_exit(&sched_lock, MTX_SPIN); 1030} 1031 1032/* ARGSUSED */ 1033static void 1034sched_setup(dummy) 1035 void *dummy; 1036{ 1037 /* Kick off timeout driven events by calling first time. */ 1038 roundrobin(NULL); 1039 schedcpu(NULL); 1040} 1041 1042/* 1043 * We adjust the priority of the current process. The priority of 1044 * a process gets worse as it accumulates CPU time. The cpu usage 1045 * estimator (p_estcpu) is increased here. resetpriority() will 1046 * compute a different priority each time p_estcpu increases by 1047 * INVERSE_ESTCPU_WEIGHT 1048 * (until MAXPRI is reached). The cpu usage estimator ramps up 1049 * quite quickly when the process is running (linearly), and decays 1050 * away exponentially, at a rate which is proportionally slower when 1051 * the system is busy. The basic principle is that the system will 1052 * 90% forget that the process used a lot of CPU time in 5 * loadav 1053 * seconds. This causes the system to favor processes which haven't 1054 * run much recently, and to round-robin among other processes. 1055 */ 1056void 1057schedclock(p) 1058 struct proc *p; 1059{ 1060 1061 p->p_cpticks++; 1062 p->p_estcpu = ESTCPULIM(p->p_estcpu + 1); 1063 if ((p->p_estcpu % INVERSE_ESTCPU_WEIGHT) == 0) { 1064 resetpriority(p); 1065 if (p->p_priority >= PUSER) 1066 p->p_priority = p->p_usrpri; 1067 } 1068} 1069