kern_synch.c revision 65708
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 65708 2000-09-11 00:20:02Z 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/kernel.h> 48#include <sys/ktr.h> 49#include <sys/signalvar.h> 50#include <sys/resourcevar.h> 51#include <sys/vmmeter.h> 52#include <sys/sysctl.h> 53#include <vm/vm.h> 54#include <vm/vm_extern.h> 55#ifdef KTRACE 56#include <sys/uio.h> 57#include <sys/ktrace.h> 58#endif 59 60#include <machine/cpu.h> 61#include <machine/ipl.h> 62#include <machine/smp.h> 63#include <machine/mutex.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 p->p_swtime++; 293 if (p->p_stat == SSLEEP || p->p_stat == SSTOP) 294 p->p_slptime++; 295 p->p_pctcpu = (p->p_pctcpu * ccpu) >> FSHIFT; 296 /* 297 * If the process has slept the entire second, 298 * stop recalculating its priority until it wakes up. 299 */ 300 if (p->p_slptime > 1) 301 continue; 302 /* 303 * prevent state changes and protect run queue 304 */ 305 s = splhigh(); 306 mtx_enter(&sched_lock, MTX_SPIN); 307 308 /* 309 * p_pctcpu is only for ps. 310 */ 311#if (FSHIFT >= CCPU_SHIFT) 312 p->p_pctcpu += (realstathz == 100)? 313 ((fixpt_t) p->p_cpticks) << (FSHIFT - CCPU_SHIFT): 314 100 * (((fixpt_t) p->p_cpticks) 315 << (FSHIFT - CCPU_SHIFT)) / realstathz; 316#else 317 p->p_pctcpu += ((FSCALE - ccpu) * 318 (p->p_cpticks * FSCALE / realstathz)) >> FSHIFT; 319#endif 320 p->p_cpticks = 0; 321 p->p_estcpu = decay_cpu(loadfac, p->p_estcpu); 322 resetpriority(p); 323 if (p->p_priority >= PUSER) { 324 if ((p != curproc) && 325#ifdef SMP 326 p->p_oncpu == 0xff && /* idle */ 327#endif 328 p->p_stat == SRUN && 329 (p->p_flag & P_INMEM) && 330 (p->p_priority / PPQ) != (p->p_usrpri / PPQ)) { 331 remrunqueue(p); 332 p->p_priority = p->p_usrpri; 333 setrunqueue(p); 334 } else 335 p->p_priority = p->p_usrpri; 336 } 337 mtx_exit(&sched_lock, MTX_SPIN); 338 splx(s); 339 } 340 vmmeter(); 341 wakeup((caddr_t)&lbolt); 342 timeout(schedcpu, (void *)0, hz); 343} 344 345/* 346 * Recalculate the priority of a process after it has slept for a while. 347 * For all load averages >= 1 and max p_estcpu of 255, sleeping for at 348 * least six times the loadfactor will decay p_estcpu to zero. 349 */ 350static void 351updatepri(p) 352 register struct proc *p; 353{ 354 register unsigned int newcpu = p->p_estcpu; 355 register fixpt_t loadfac = loadfactor(averunnable.ldavg[0]); 356 357 if (p->p_slptime > 5 * loadfac) 358 p->p_estcpu = 0; 359 else { 360 p->p_slptime--; /* the first time was done in schedcpu */ 361 while (newcpu && --p->p_slptime) 362 newcpu = decay_cpu(loadfac, newcpu); 363 p->p_estcpu = newcpu; 364 } 365 resetpriority(p); 366} 367 368/* 369 * We're only looking at 7 bits of the address; everything is 370 * aligned to 4, lots of things are aligned to greater powers 371 * of 2. Shift right by 8, i.e. drop the bottom 256 worth. 372 */ 373#define TABLESIZE 128 374static TAILQ_HEAD(slpquehead, proc) slpque[TABLESIZE]; 375#define LOOKUP(x) (((intptr_t)(x) >> 8) & (TABLESIZE - 1)) 376 377#if 0 378/* 379 * During autoconfiguration or after a panic, a sleep will simply 380 * lower the priority briefly to allow interrupts, then return. 381 * The priority to be used (safepri) is machine-dependent, thus this 382 * value is initialized and maintained in the machine-dependent layers. 383 * This priority will typically be 0, or the lowest priority 384 * that is safe for use on the interrupt stack; it can be made 385 * higher to block network software interrupts after panics. 386 */ 387int safepri; 388#endif 389 390void 391sleepinit(void) 392{ 393 int i; 394 395 sched_quantum = hz/10; 396 hogticks = 2 * sched_quantum; 397 for (i = 0; i < TABLESIZE; i++) 398 TAILQ_INIT(&slpque[i]); 399} 400 401/* 402 * General sleep call. Suspends the current process until a wakeup is 403 * performed on the specified identifier. The process will then be made 404 * runnable with the specified priority. Sleeps at most timo/hz seconds 405 * (0 means no timeout). If pri includes PCATCH flag, signals are checked 406 * before and after sleeping, else signals are not checked. Returns 0 if 407 * awakened, EWOULDBLOCK if the timeout expires. If PCATCH is set and a 408 * signal needs to be delivered, ERESTART is returned if the current system 409 * call should be restarted if possible, and EINTR is returned if the system 410 * call should be interrupted by the signal (return EINTR). 411 * 412 * The mutex argument is exited before the caller is suspended, and 413 * entered before msleep returns. If priority includes the PDROP 414 * flag the mutex is not entered before returning. 415 */ 416int 417msleep(ident, mtx, priority, wmesg, timo) 418 void *ident; 419 mtx_t *mtx; 420 int priority, timo; 421 const char *wmesg; 422{ 423 struct proc *p = curproc; 424 int s, sig, catch = priority & PCATCH; 425 struct callout_handle thandle; 426 int rval = 0; 427 WITNESS_SAVE_DECL(mtx); 428 429#ifdef KTRACE 430 if (p && KTRPOINT(p, KTR_CSW)) 431 ktrcsw(p->p_tracep, 1, 0); 432#endif 433 WITNESS_SLEEP(0, mtx); 434 mtx_enter(&sched_lock, MTX_SPIN); 435 436 if (mtx != NULL) { 437 WITNESS_SAVE(mtx, mtx); 438 mtx_exit(mtx, MTX_DEF | MTX_NOSWITCH); 439 if (priority & PDROP) 440 mtx = NULL; 441 } 442 443 s = splhigh(); 444 if (cold || panicstr) { 445 /* 446 * After a panic, or during autoconfiguration, 447 * just give interrupts a chance, then just return; 448 * don't run any other procs or panic below, 449 * in case this is the idle process and already asleep. 450 */ 451 mtx_exit(&sched_lock, MTX_SPIN); 452#if 0 453 splx(safepri); 454#endif 455 splx(s); 456 return (0); 457 } 458 459 KASSERT(p != NULL, ("tsleep1")); 460 KASSERT(ident != NULL && p->p_stat == SRUN, ("tsleep")); 461 /* 462 * Process may be sitting on a slpque if asleep() was called, remove 463 * it before re-adding. 464 */ 465 if (p->p_wchan != NULL) 466 unsleep(p); 467 468 p->p_wchan = ident; 469 p->p_wmesg = wmesg; 470 p->p_slptime = 0; 471 p->p_priority = priority & PRIMASK; 472 p->p_nativepri = p->p_priority; 473 CTR4(KTR_PROC, "tsleep: proc %p (pid %d, %s), schedlock %p", 474 p, p->p_pid, p->p_comm, (void *) sched_lock.mtx_lock); 475 TAILQ_INSERT_TAIL(&slpque[LOOKUP(ident)], p, p_procq); 476 if (timo) 477 thandle = timeout(endtsleep, (void *)p, timo); 478 /* 479 * We put ourselves on the sleep queue and start our timeout 480 * before calling CURSIG, as we could stop there, and a wakeup 481 * or a SIGCONT (or both) could occur while we were stopped. 482 * A SIGCONT would cause us to be marked as SSLEEP 483 * without resuming us, thus we must be ready for sleep 484 * when CURSIG is called. If the wakeup happens while we're 485 * stopped, p->p_wchan will be 0 upon return from CURSIG. 486 */ 487 if (catch) { 488 CTR4(KTR_PROC, 489 "tsleep caught: proc %p (pid %d, %s), schedlock %p", 490 p, p->p_pid, p->p_comm, (void *) sched_lock.mtx_lock); 491 p->p_flag |= P_SINTR; 492 if ((sig = CURSIG(p))) { 493 if (p->p_wchan) 494 unsleep(p); 495 p->p_stat = SRUN; 496 goto resume; 497 } 498 if (p->p_wchan == 0) { 499 catch = 0; 500 goto resume; 501 } 502 } else 503 sig = 0; 504 p->p_stat = SSLEEP; 505 p->p_stats->p_ru.ru_nvcsw++; 506 mi_switch(); 507 CTR4(KTR_PROC, 508 "tsleep resume: proc %p (pid %d, %s), schedlock %p", 509 p, p->p_pid, p->p_comm, (void *) sched_lock.mtx_lock); 510resume: 511 curpriority = p->p_usrpri; 512 splx(s); 513 p->p_flag &= ~P_SINTR; 514 if (p->p_flag & P_TIMEOUT) { 515 p->p_flag &= ~P_TIMEOUT; 516 if (sig == 0) { 517#ifdef KTRACE 518 if (KTRPOINT(p, KTR_CSW)) 519 ktrcsw(p->p_tracep, 0, 0); 520#endif 521 rval = EWOULDBLOCK; 522 goto out; 523 } 524 } else if (timo) 525 untimeout(endtsleep, (void *)p, thandle); 526 if (catch && (sig != 0 || (sig = CURSIG(p)))) { 527#ifdef KTRACE 528 if (KTRPOINT(p, KTR_CSW)) 529 ktrcsw(p->p_tracep, 0, 0); 530#endif 531 if (SIGISMEMBER(p->p_sigacts->ps_sigintr, sig)) 532 rval = EINTR; 533 else 534 rval = ERESTART; 535 goto out; 536 } 537out: 538 mtx_exit(&sched_lock, MTX_SPIN); 539#ifdef KTRACE 540 if (KTRPOINT(p, KTR_CSW)) 541 ktrcsw(p->p_tracep, 0, 0); 542#endif 543 if (mtx != NULL) { 544 mtx_enter(mtx, MTX_DEF); 545 WITNESS_RESTORE(mtx, mtx); 546 } 547 return (rval); 548} 549 550/* 551 * asleep() - async sleep call. Place process on wait queue and return 552 * immediately without blocking. The process stays runnable until await() 553 * is called. If ident is NULL, remove process from wait queue if it is still 554 * on one. 555 * 556 * Only the most recent sleep condition is effective when making successive 557 * calls to asleep() or when calling tsleep(). 558 * 559 * The timeout, if any, is not initiated until await() is called. The sleep 560 * priority, signal, and timeout is specified in the asleep() call but may be 561 * overriden in the await() call. 562 * 563 * <<<<<<<< EXPERIMENTAL, UNTESTED >>>>>>>>>> 564 */ 565 566int 567asleep(void *ident, int priority, const char *wmesg, int timo) 568{ 569 struct proc *p = curproc; 570 int s; 571 572 /* 573 * obtain sched_lock while manipulating sleep structures and slpque. 574 * 575 * Remove preexisting wait condition (if any) and place process 576 * on appropriate slpque, but do not put process to sleep. 577 */ 578 579 s = splhigh(); 580 mtx_enter(&sched_lock, MTX_SPIN); 581 582 if (p->p_wchan != NULL) 583 unsleep(p); 584 585 if (ident) { 586 p->p_wchan = ident; 587 p->p_wmesg = wmesg; 588 p->p_slptime = 0; 589 p->p_asleep.as_priority = priority; 590 p->p_asleep.as_timo = timo; 591 TAILQ_INSERT_TAIL(&slpque[LOOKUP(ident)], p, p_procq); 592 } 593 594 mtx_exit(&sched_lock, MTX_SPIN); 595 splx(s); 596 597 return(0); 598} 599 600/* 601 * await() - wait for async condition to occur. The process blocks until 602 * wakeup() is called on the most recent asleep() address. If wakeup is called 603 * priority to await(), await() winds up being a NOP. 604 * 605 * If await() is called more then once (without an intervening asleep() call), 606 * await() is still effectively a NOP but it calls mi_switch() to give other 607 * processes some cpu before returning. The process is left runnable. 608 * 609 * <<<<<<<< EXPERIMENTAL, UNTESTED >>>>>>>>>> 610 */ 611 612int 613await(int priority, int timo) 614{ 615 struct proc *p = curproc; 616 int rval = 0; 617 int s; 618 619 mtx_enter(&sched_lock, MTX_SPIN); 620 621 s = splhigh(); 622 623 if (p->p_wchan != NULL) { 624 struct callout_handle thandle; 625 int sig; 626 int catch; 627 628 /* 629 * The call to await() can override defaults specified in 630 * the original asleep(). 631 */ 632 if (priority < 0) 633 priority = p->p_asleep.as_priority; 634 if (timo < 0) 635 timo = p->p_asleep.as_timo; 636 637 /* 638 * Install timeout 639 */ 640 641 if (timo) 642 thandle = timeout(endtsleep, (void *)p, timo); 643 644 sig = 0; 645 catch = priority & PCATCH; 646 647 if (catch) { 648 p->p_flag |= P_SINTR; 649 if ((sig = CURSIG(p))) { 650 if (p->p_wchan) 651 unsleep(p); 652 p->p_stat = SRUN; 653 goto resume; 654 } 655 if (p->p_wchan == NULL) { 656 catch = 0; 657 goto resume; 658 } 659 } 660 p->p_stat = SSLEEP; 661 p->p_stats->p_ru.ru_nvcsw++; 662 mi_switch(); 663resume: 664 curpriority = p->p_usrpri; 665 666 splx(s); 667 p->p_flag &= ~P_SINTR; 668 if (p->p_flag & P_TIMEOUT) { 669 p->p_flag &= ~P_TIMEOUT; 670 if (sig == 0) { 671#ifdef KTRACE 672 if (KTRPOINT(p, KTR_CSW)) 673 ktrcsw(p->p_tracep, 0, 0); 674#endif 675 rval = EWOULDBLOCK; 676 goto out; 677 } 678 } else if (timo) 679 untimeout(endtsleep, (void *)p, thandle); 680 if (catch && (sig != 0 || (sig = CURSIG(p)))) { 681#ifdef KTRACE 682 if (KTRPOINT(p, KTR_CSW)) 683 ktrcsw(p->p_tracep, 0, 0); 684#endif 685 if (SIGISMEMBER(p->p_sigacts->ps_sigintr, sig)) 686 rval = EINTR; 687 else 688 rval = ERESTART; 689 goto out; 690 } 691#ifdef KTRACE 692 if (KTRPOINT(p, KTR_CSW)) 693 ktrcsw(p->p_tracep, 0, 0); 694#endif 695 } else { 696 /* 697 * If as_priority is 0, await() has been called without an 698 * intervening asleep(). We are still effectively a NOP, 699 * but we call mi_switch() for safety. 700 */ 701 702 if (p->p_asleep.as_priority == 0) { 703 p->p_stats->p_ru.ru_nvcsw++; 704 mi_switch(); 705 } 706 splx(s); 707 } 708 709 /* 710 * clear p_asleep.as_priority as an indication that await() has been 711 * called. If await() is called again without an intervening asleep(), 712 * await() is still effectively a NOP but the above mi_switch() code 713 * is triggered as a safety. 714 */ 715 p->p_asleep.as_priority = 0; 716 717out: 718 mtx_exit(&sched_lock, MTX_SPIN); 719 720 return (rval); 721} 722 723/* 724 * Implement timeout for tsleep or asleep()/await() 725 * 726 * If process hasn't been awakened (wchan non-zero), 727 * set timeout flag and undo the sleep. If proc 728 * is stopped, just unsleep so it will remain stopped. 729 */ 730static void 731endtsleep(arg) 732 void *arg; 733{ 734 register struct proc *p; 735 int s; 736 737 p = (struct proc *)arg; 738 CTR4(KTR_PROC, 739 "endtsleep: proc %p (pid %d, %s), schedlock %p", 740 p, p->p_pid, p->p_comm, (void *) sched_lock.mtx_lock); 741 s = splhigh(); 742 mtx_enter(&sched_lock, MTX_SPIN); 743 if (p->p_wchan) { 744 if (p->p_stat == SSLEEP) 745 setrunnable(p); 746 else 747 unsleep(p); 748 p->p_flag |= P_TIMEOUT; 749 } 750 mtx_exit(&sched_lock, MTX_SPIN); 751 splx(s); 752} 753 754/* 755 * Remove a process from its wait queue 756 */ 757void 758unsleep(p) 759 register struct proc *p; 760{ 761 int s; 762 763 s = splhigh(); 764 mtx_enter(&sched_lock, MTX_SPIN); 765 if (p->p_wchan) { 766 TAILQ_REMOVE(&slpque[LOOKUP(p->p_wchan)], p, p_procq); 767 p->p_wchan = 0; 768 } 769 mtx_exit(&sched_lock, MTX_SPIN); 770 splx(s); 771} 772 773/* 774 * Make all processes sleeping on the specified identifier runnable. 775 */ 776void 777wakeup(ident) 778 register void *ident; 779{ 780 register struct slpquehead *qp; 781 register struct proc *p; 782 int s; 783 784 s = splhigh(); 785 mtx_enter(&sched_lock, MTX_SPIN); 786 qp = &slpque[LOOKUP(ident)]; 787restart: 788 TAILQ_FOREACH(p, qp, p_procq) { 789 if (p->p_wchan == ident) { 790 TAILQ_REMOVE(qp, p, p_procq); 791 p->p_wchan = 0; 792 if (p->p_stat == SSLEEP) { 793 /* OPTIMIZED EXPANSION OF setrunnable(p); */ 794 CTR4(KTR_PROC, 795 "wakeup: proc %p (pid %d, %s), schedlock %p", 796 p, p->p_pid, p->p_comm, (void *) sched_lock.mtx_lock); 797 if (p->p_slptime > 1) 798 updatepri(p); 799 p->p_slptime = 0; 800 p->p_stat = SRUN; 801 if (p->p_flag & P_INMEM) { 802 setrunqueue(p); 803 maybe_resched(p); 804 } else { 805 p->p_flag |= P_SWAPINREQ; 806 wakeup((caddr_t)&proc0); 807 } 808 /* END INLINE EXPANSION */ 809 goto restart; 810 } 811 } 812 } 813 mtx_exit(&sched_lock, MTX_SPIN); 814 splx(s); 815} 816 817/* 818 * Make a process sleeping on the specified identifier runnable. 819 * May wake more than one process if a target process is currently 820 * swapped out. 821 */ 822void 823wakeup_one(ident) 824 register void *ident; 825{ 826 register struct slpquehead *qp; 827 register struct proc *p; 828 int s; 829 830 s = splhigh(); 831 mtx_enter(&sched_lock, MTX_SPIN); 832 qp = &slpque[LOOKUP(ident)]; 833 834 TAILQ_FOREACH(p, qp, p_procq) { 835 if (p->p_wchan == ident) { 836 TAILQ_REMOVE(qp, p, p_procq); 837 p->p_wchan = 0; 838 if (p->p_stat == SSLEEP) { 839 /* OPTIMIZED EXPANSION OF setrunnable(p); */ 840 CTR4(KTR_PROC, 841 "wakeup1: proc %p (pid %d, %s), schedlock %p", 842 p, p->p_pid, p->p_comm, (void *) sched_lock.mtx_lock); 843 if (p->p_slptime > 1) 844 updatepri(p); 845 p->p_slptime = 0; 846 p->p_stat = SRUN; 847 if (p->p_flag & P_INMEM) { 848 setrunqueue(p); 849 maybe_resched(p); 850 break; 851 } else { 852 p->p_flag |= P_SWAPINREQ; 853 wakeup((caddr_t)&proc0); 854 } 855 /* END INLINE EXPANSION */ 856 } 857 } 858 } 859 mtx_exit(&sched_lock, MTX_SPIN); 860 splx(s); 861} 862 863/* 864 * The machine independent parts of mi_switch(). 865 * Must be called at splstatclock() or higher. 866 */ 867void 868mi_switch() 869{ 870 struct timeval new_switchtime; 871 register struct proc *p = curproc; /* XXX */ 872 register struct rlimit *rlim; 873 int giantreleased; 874 int x; 875 WITNESS_SAVE_DECL(Giant); 876 877 /* 878 * XXX this spl is almost unnecessary. It is partly to allow for 879 * sloppy callers that don't do it (issignal() via CURSIG() is the 880 * main offender). It is partly to work around a bug in the i386 881 * cpu_switch() (the ipl is not preserved). We ran for years 882 * without it. I think there was only a interrupt latency problem. 883 * The main caller, tsleep(), does an splx() a couple of instructions 884 * after calling here. The buggy caller, issignal(), usually calls 885 * here at spl0() and sometimes returns at splhigh(). The process 886 * then runs for a little too long at splhigh(). The ipl gets fixed 887 * when the process returns to user mode (or earlier). 888 * 889 * It would probably be better to always call here at spl0(). Callers 890 * are prepared to give up control to another process, so they must 891 * be prepared to be interrupted. The clock stuff here may not 892 * actually need splstatclock(). 893 */ 894 x = splstatclock(); 895 896 CTR4(KTR_PROC, "mi_switch: old proc %p (pid %d, %s), schedlock %p", 897 p, p->p_pid, p->p_comm, (void *) sched_lock.mtx_lock); 898 mtx_enter(&sched_lock, MTX_SPIN | MTX_RLIKELY); 899 900 WITNESS_SAVE(&Giant, Giant); 901 for (giantreleased = 0; mtx_owned(&Giant); giantreleased++) 902 mtx_exit(&Giant, MTX_DEF | MTX_NOSWITCH); 903 904#ifdef SIMPLELOCK_DEBUG 905 if (p->p_simple_locks) 906 printf("sleep: holding simple lock\n"); 907#endif 908 /* 909 * Compute the amount of time during which the current 910 * process was running, and add that to its total so far. 911 */ 912 microuptime(&new_switchtime); 913 if (timevalcmp(&new_switchtime, &switchtime, <)) { 914 printf("microuptime() went backwards (%ld.%06ld -> %ld.%06ld)\n", 915 switchtime.tv_sec, switchtime.tv_usec, 916 new_switchtime.tv_sec, new_switchtime.tv_usec); 917 new_switchtime = switchtime; 918 } else { 919 p->p_runtime += (new_switchtime.tv_usec - switchtime.tv_usec) + 920 (new_switchtime.tv_sec - switchtime.tv_sec) * (int64_t)1000000; 921 } 922 923 /* 924 * Check if the process exceeds its cpu resource allocation. 925 * If over max, kill it. 926 * 927 * XXX drop sched_lock, pickup Giant 928 */ 929 if (p->p_stat != SZOMB && p->p_limit->p_cpulimit != RLIM_INFINITY && 930 p->p_runtime > p->p_limit->p_cpulimit) { 931 rlim = &p->p_rlimit[RLIMIT_CPU]; 932 if (p->p_runtime / (rlim_t)1000000 >= rlim->rlim_max) { 933 killproc(p, "exceeded maximum CPU limit"); 934 } else { 935 psignal(p, SIGXCPU); 936 if (rlim->rlim_cur < rlim->rlim_max) { 937 /* XXX: we should make a private copy */ 938 rlim->rlim_cur += 5; 939 } 940 } 941 } 942 943 /* 944 * Pick a new current process and record its start time. 945 */ 946 cnt.v_swtch++; 947 switchtime = new_switchtime; 948 CTR4(KTR_PROC, "mi_switch: old proc %p (pid %d, %s), schedlock %p", 949 p, p->p_pid, p->p_comm, (void *) sched_lock.mtx_lock); 950 cpu_switch(); 951 CTR4(KTR_PROC, "mi_switch: new proc %p (pid %d, %s), schedlock %p", 952 p, p->p_pid, p->p_comm, (void *) sched_lock.mtx_lock); 953 if (switchtime.tv_sec == 0) 954 microuptime(&switchtime); 955 switchticks = ticks; 956 mtx_exit(&sched_lock, MTX_SPIN); 957 while (giantreleased--) 958 mtx_enter(&Giant, MTX_DEF); 959 WITNESS_RESTORE(&Giant, Giant); 960 961 splx(x); 962} 963 964/* 965 * Change process state to be runnable, 966 * placing it on the run queue if it is in memory, 967 * and awakening the swapper if it isn't in memory. 968 */ 969void 970setrunnable(p) 971 register struct proc *p; 972{ 973 register int s; 974 975 s = splhigh(); 976 mtx_enter(&sched_lock, MTX_SPIN); 977 switch (p->p_stat) { 978 case 0: 979 case SRUN: 980 case SZOMB: 981 case SWAIT: 982 default: 983 panic("setrunnable"); 984 case SSTOP: 985 case SSLEEP: 986 unsleep(p); /* e.g. when sending signals */ 987 break; 988 989 case SIDL: 990 break; 991 } 992 p->p_stat = SRUN; 993 if (p->p_flag & P_INMEM) 994 setrunqueue(p); 995 mtx_exit(&sched_lock, MTX_SPIN); 996 splx(s); 997 if (p->p_slptime > 1) 998 updatepri(p); 999 p->p_slptime = 0; 1000 if ((p->p_flag & P_INMEM) == 0) { 1001 p->p_flag |= P_SWAPINREQ; 1002 wakeup((caddr_t)&proc0); 1003 } 1004 else 1005 maybe_resched(p); 1006} 1007 1008/* 1009 * Compute the priority of a process when running in user mode. 1010 * Arrange to reschedule if the resulting priority is better 1011 * than that of the current process. 1012 */ 1013void 1014resetpriority(p) 1015 register struct proc *p; 1016{ 1017 register unsigned int newpriority; 1018 1019 if (p->p_rtprio.type == RTP_PRIO_NORMAL) { 1020 newpriority = PUSER + p->p_estcpu / INVERSE_ESTCPU_WEIGHT + 1021 NICE_WEIGHT * (p->p_nice - PRIO_MIN); 1022 newpriority = min(newpriority, MAXPRI); 1023 p->p_usrpri = newpriority; 1024 } 1025 maybe_resched(p); 1026} 1027 1028/* ARGSUSED */ 1029static void 1030sched_setup(dummy) 1031 void *dummy; 1032{ 1033 /* Kick off timeout driven events by calling first time. */ 1034 roundrobin(NULL); 1035 schedcpu(NULL); 1036} 1037 1038/* 1039 * We adjust the priority of the current process. The priority of 1040 * a process gets worse as it accumulates CPU time. The cpu usage 1041 * estimator (p_estcpu) is increased here. resetpriority() will 1042 * compute a different priority each time p_estcpu increases by 1043 * INVERSE_ESTCPU_WEIGHT 1044 * (until MAXPRI is reached). The cpu usage estimator ramps up 1045 * quite quickly when the process is running (linearly), and decays 1046 * away exponentially, at a rate which is proportionally slower when 1047 * the system is busy. The basic principle is that the system will 1048 * 90% forget that the process used a lot of CPU time in 5 * loadav 1049 * seconds. This causes the system to favor processes which haven't 1050 * run much recently, and to round-robin among other processes. 1051 */ 1052void 1053schedclock(p) 1054 struct proc *p; 1055{ 1056 1057 p->p_cpticks++; 1058 p->p_estcpu = ESTCPULIM(p->p_estcpu + 1); 1059 if ((p->p_estcpu % INVERSE_ESTCPU_WEIGHT) == 0) { 1060 resetpriority(p); 1061 if (p->p_priority >= PUSER) 1062 p->p_priority = p->p_usrpri; 1063 } 1064} 1065