kern_synch.c revision 68794
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 68794 2000-11-15 22:39:35Z jhb $ 40 */ 41 42#include "opt_ktrace.h" 43 44#include <sys/param.h> 45#include <sys/systm.h> 46#include <sys/proc.h> 47#include <sys/ipl.h> 48#include <sys/kernel.h> 49#include <sys/ktr.h> 50#include <sys/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); 425 426 if (mtx != NULL) { 427 mtx_assert(mtx, MA_OWNED | MA_NOTRECURSED); 428 WITNESS_SAVE(mtx, mtx); 429 mtx_exit(mtx, MTX_DEF | MTX_NOSWITCH); 430 if (priority & PDROP) 431 mtx = NULL; 432 } 433 434 s = splhigh(); 435 if (cold || panicstr) { 436 /* 437 * After a panic, or during autoconfiguration, 438 * just give interrupts a chance, then just return; 439 * don't run any other procs or panic below, 440 * in case this is the idle process and already asleep. 441 */ 442 mtx_exit(&sched_lock, MTX_SPIN); 443 splx(s); 444 return (0); 445 } 446 447 KASSERT(p != NULL, ("msleep1")); 448 KASSERT(ident != NULL && p->p_stat == SRUN, ("msleep")); 449 /* 450 * Process may be sitting on a slpque if asleep() was called, remove 451 * it before re-adding. 452 */ 453 if (p->p_wchan != NULL) 454 unsleep(p); 455 456 p->p_wchan = ident; 457 p->p_wmesg = wmesg; 458 p->p_slptime = 0; 459 p->p_priority = priority & PRIMASK; 460 p->p_nativepri = p->p_priority; 461 CTR4(KTR_PROC, "msleep: proc %p (pid %d, %s), schedlock %p", 462 p, p->p_pid, p->p_comm, (void *) sched_lock.mtx_lock); 463 TAILQ_INSERT_TAIL(&slpque[LOOKUP(ident)], p, p_procq); 464 if (timo) 465 thandle = timeout(endtsleep, (void *)p, timo); 466 /* 467 * We put ourselves on the sleep queue and start our timeout 468 * before calling CURSIG, as we could stop there, and a wakeup 469 * or a SIGCONT (or both) could occur while we were stopped. 470 * A SIGCONT would cause us to be marked as SSLEEP 471 * without resuming us, thus we must be ready for sleep 472 * when CURSIG is called. If the wakeup happens while we're 473 * stopped, p->p_wchan will be 0 upon return from CURSIG. 474 */ 475 if (catch) { 476 CTR4(KTR_PROC, 477 "msleep caught: proc %p (pid %d, %s), schedlock %p", 478 p, p->p_pid, p->p_comm, (void *) sched_lock.mtx_lock); 479 p->p_flag |= P_SINTR; 480 if ((sig = CURSIG(p))) { 481 if (p->p_wchan) 482 unsleep(p); 483 p->p_stat = SRUN; 484 goto resume; 485 } 486 if (p->p_wchan == 0) { 487 catch = 0; 488 goto resume; 489 } 490 } else 491 sig = 0; 492 p->p_stat = SSLEEP; 493 p->p_stats->p_ru.ru_nvcsw++; 494 mi_switch(); 495 CTR4(KTR_PROC, 496 "msleep resume: proc %p (pid %d, %s), schedlock %p", 497 p, p->p_pid, p->p_comm, (void *) sched_lock.mtx_lock); 498resume: 499 curpriority = p->p_usrpri; 500 splx(s); 501 p->p_flag &= ~P_SINTR; 502 if (p->p_flag & P_TIMEOUT) { 503 p->p_flag &= ~P_TIMEOUT; 504 if (sig == 0) { 505#ifdef KTRACE 506 if (KTRPOINT(p, KTR_CSW)) 507 ktrcsw(p->p_tracep, 0, 0); 508#endif 509 rval = EWOULDBLOCK; 510 goto out; 511 } 512 } else if (timo) 513 untimeout(endtsleep, (void *)p, thandle); 514 if (catch && (sig != 0 || (sig = CURSIG(p)))) { 515#ifdef KTRACE 516 if (KTRPOINT(p, KTR_CSW)) 517 ktrcsw(p->p_tracep, 0, 0); 518#endif 519 if (SIGISMEMBER(p->p_sigacts->ps_sigintr, sig)) 520 rval = EINTR; 521 else 522 rval = ERESTART; 523 goto out; 524 } 525out: 526 mtx_exit(&sched_lock, MTX_SPIN); 527#ifdef KTRACE 528 if (KTRPOINT(p, KTR_CSW)) 529 ktrcsw(p->p_tracep, 0, 0); 530#endif 531 if (mtx != NULL) { 532 mtx_enter(mtx, MTX_DEF); 533 WITNESS_RESTORE(mtx, mtx); 534 } 535 return (rval); 536} 537 538/* 539 * asleep() - async sleep call. Place process on wait queue and return 540 * immediately without blocking. The process stays runnable until mawait() 541 * is called. If ident is NULL, remove process from wait queue if it is still 542 * on one. 543 * 544 * Only the most recent sleep condition is effective when making successive 545 * calls to asleep() or when calling msleep(). 546 * 547 * The timeout, if any, is not initiated until mawait() is called. The sleep 548 * priority, signal, and timeout is specified in the asleep() call but may be 549 * overriden in the mawait() call. 550 * 551 * <<<<<<<< EXPERIMENTAL, UNTESTED >>>>>>>>>> 552 */ 553 554int 555asleep(void *ident, int priority, const char *wmesg, int timo) 556{ 557 struct proc *p = curproc; 558 int s; 559 560 /* 561 * obtain sched_lock while manipulating sleep structures and slpque. 562 * 563 * Remove preexisting wait condition (if any) and place process 564 * on appropriate slpque, but do not put process to sleep. 565 */ 566 567 s = splhigh(); 568 mtx_enter(&sched_lock, MTX_SPIN); 569 570 if (p->p_wchan != NULL) 571 unsleep(p); 572 573 if (ident) { 574 p->p_wchan = ident; 575 p->p_wmesg = wmesg; 576 p->p_slptime = 0; 577 p->p_asleep.as_priority = priority; 578 p->p_asleep.as_timo = timo; 579 TAILQ_INSERT_TAIL(&slpque[LOOKUP(ident)], p, p_procq); 580 } 581 582 mtx_exit(&sched_lock, MTX_SPIN); 583 splx(s); 584 585 return(0); 586} 587 588/* 589 * mawait() - wait for async condition to occur. The process blocks until 590 * wakeup() is called on the most recent asleep() address. If wakeup is called 591 * prior to mawait(), mawait() winds up being a NOP. 592 * 593 * If mawait() is called more then once (without an intervening asleep() call), 594 * mawait() is still effectively a NOP but it calls mi_switch() to give other 595 * processes some cpu before returning. The process is left runnable. 596 * 597 * <<<<<<<< EXPERIMENTAL, UNTESTED >>>>>>>>>> 598 */ 599 600int 601mawait(struct mtx *mtx, int priority, int timo) 602{ 603 struct proc *p = curproc; 604 int rval = 0; 605 int s; 606 WITNESS_SAVE_DECL(mtx); 607 608 WITNESS_SLEEP(0, mtx); 609 mtx_enter(&sched_lock, MTX_SPIN); 610 if (mtx != NULL) { 611 mtx_assert(mtx, MA_OWNED | MA_NOTRECURSED); 612 WITNESS_SAVE(mtx, mtx); 613 mtx_exit(mtx, MTX_DEF | MTX_NOSWITCH); 614 if (priority & PDROP) 615 mtx = NULL; 616 } 617 618 s = splhigh(); 619 620 if (p->p_wchan != NULL) { 621 struct callout_handle thandle; 622 int sig; 623 int catch; 624 625 /* 626 * The call to mawait() can override defaults specified in 627 * the original asleep(). 628 */ 629 if (priority < 0) 630 priority = p->p_asleep.as_priority; 631 if (timo < 0) 632 timo = p->p_asleep.as_timo; 633 634 /* 635 * Install timeout 636 */ 637 638 if (timo) 639 thandle = timeout(endtsleep, (void *)p, timo); 640 641 sig = 0; 642 catch = priority & PCATCH; 643 644 if (catch) { 645 p->p_flag |= P_SINTR; 646 if ((sig = CURSIG(p))) { 647 if (p->p_wchan) 648 unsleep(p); 649 p->p_stat = SRUN; 650 goto resume; 651 } 652 if (p->p_wchan == NULL) { 653 catch = 0; 654 goto resume; 655 } 656 } 657 p->p_stat = SSLEEP; 658 p->p_stats->p_ru.ru_nvcsw++; 659 mi_switch(); 660resume: 661 curpriority = p->p_usrpri; 662 663 splx(s); 664 p->p_flag &= ~P_SINTR; 665 if (p->p_flag & P_TIMEOUT) { 666 p->p_flag &= ~P_TIMEOUT; 667 if (sig == 0) { 668#ifdef KTRACE 669 if (KTRPOINT(p, KTR_CSW)) 670 ktrcsw(p->p_tracep, 0, 0); 671#endif 672 rval = EWOULDBLOCK; 673 goto out; 674 } 675 } else if (timo) 676 untimeout(endtsleep, (void *)p, thandle); 677 if (catch && (sig != 0 || (sig = CURSIG(p)))) { 678#ifdef KTRACE 679 if (KTRPOINT(p, KTR_CSW)) 680 ktrcsw(p->p_tracep, 0, 0); 681#endif 682 if (SIGISMEMBER(p->p_sigacts->ps_sigintr, sig)) 683 rval = EINTR; 684 else 685 rval = ERESTART; 686 goto out; 687 } 688#ifdef KTRACE 689 if (KTRPOINT(p, KTR_CSW)) 690 ktrcsw(p->p_tracep, 0, 0); 691#endif 692 } else { 693 /* 694 * If as_priority is 0, mawait() has been called without an 695 * intervening asleep(). We are still effectively a NOP, 696 * but we call mi_switch() for safety. 697 */ 698 699 if (p->p_asleep.as_priority == 0) { 700 p->p_stats->p_ru.ru_nvcsw++; 701 mi_switch(); 702 } 703 splx(s); 704 } 705 706 /* 707 * clear p_asleep.as_priority as an indication that mawait() has been 708 * called. If mawait() is called again without an intervening asleep(), 709 * mawait() is still effectively a NOP but the above mi_switch() code 710 * is triggered as a safety. 711 */ 712 p->p_asleep.as_priority = 0; 713 714out: 715 mtx_exit(&sched_lock, MTX_SPIN); 716 717 if (mtx != NULL) { 718 mtx_enter(mtx, MTX_DEF); 719 WITNESS_RESTORE(mtx, mtx); 720 } 721 return (rval); 722} 723 724/* 725 * Implement timeout for msleep or asleep()/mawait() 726 * 727 * If process hasn't been awakened (wchan non-zero), 728 * set timeout flag and undo the sleep. If proc 729 * is stopped, just unsleep so it will remain stopped. 730 */ 731static void 732endtsleep(arg) 733 void *arg; 734{ 735 register struct proc *p; 736 int s; 737 738 p = (struct proc *)arg; 739 CTR4(KTR_PROC, 740 "endtsleep: proc %p (pid %d, %s), schedlock %p", 741 p, p->p_pid, p->p_comm, (void *) sched_lock.mtx_lock); 742 s = splhigh(); 743 mtx_enter(&sched_lock, MTX_SPIN); 744 if (p->p_wchan) { 745 if (p->p_stat == SSLEEP) 746 setrunnable(p); 747 else 748 unsleep(p); 749 p->p_flag |= P_TIMEOUT; 750 } 751 mtx_exit(&sched_lock, MTX_SPIN); 752 splx(s); 753} 754 755/* 756 * Remove a process from its wait queue 757 */ 758void 759unsleep(p) 760 register struct proc *p; 761{ 762 int s; 763 764 s = splhigh(); 765 mtx_enter(&sched_lock, MTX_SPIN); 766 if (p->p_wchan) { 767 TAILQ_REMOVE(&slpque[LOOKUP(p->p_wchan)], p, p_procq); 768 p->p_wchan = 0; 769 } 770 mtx_exit(&sched_lock, MTX_SPIN); 771 splx(s); 772} 773 774/* 775 * Make all processes sleeping on the specified identifier runnable. 776 */ 777void 778wakeup(ident) 779 register void *ident; 780{ 781 register struct slpquehead *qp; 782 register struct proc *p; 783 int s; 784 785 s = splhigh(); 786 mtx_enter(&sched_lock, MTX_SPIN); 787 qp = &slpque[LOOKUP(ident)]; 788restart: 789 TAILQ_FOREACH(p, qp, p_procq) { 790 if (p->p_wchan == ident) { 791 TAILQ_REMOVE(qp, p, p_procq); 792 p->p_wchan = 0; 793 if (p->p_stat == SSLEEP) { 794 /* OPTIMIZED EXPANSION OF setrunnable(p); */ 795 CTR4(KTR_PROC, 796 "wakeup: proc %p (pid %d, %s), schedlock %p", 797 p, p->p_pid, p->p_comm, (void *) sched_lock.mtx_lock); 798 if (p->p_slptime > 1) 799 updatepri(p); 800 p->p_slptime = 0; 801 p->p_stat = SRUN; 802 if (p->p_flag & P_INMEM) { 803 setrunqueue(p); 804 maybe_resched(p); 805 } else { 806 p->p_flag |= P_SWAPINREQ; 807 wakeup((caddr_t)&proc0); 808 } 809 /* END INLINE EXPANSION */ 810 goto restart; 811 } 812 } 813 } 814 mtx_exit(&sched_lock, MTX_SPIN); 815 splx(s); 816} 817 818/* 819 * Make a process sleeping on the specified identifier runnable. 820 * May wake more than one process if a target process is currently 821 * swapped out. 822 */ 823void 824wakeup_one(ident) 825 register void *ident; 826{ 827 register struct slpquehead *qp; 828 register struct proc *p; 829 int s; 830 831 s = splhigh(); 832 mtx_enter(&sched_lock, MTX_SPIN); 833 qp = &slpque[LOOKUP(ident)]; 834 835 TAILQ_FOREACH(p, qp, p_procq) { 836 if (p->p_wchan == ident) { 837 TAILQ_REMOVE(qp, p, p_procq); 838 p->p_wchan = 0; 839 if (p->p_stat == SSLEEP) { 840 /* OPTIMIZED EXPANSION OF setrunnable(p); */ 841 CTR4(KTR_PROC, 842 "wakeup1: proc %p (pid %d, %s), schedlock %p", 843 p, p->p_pid, p->p_comm, (void *) sched_lock.mtx_lock); 844 if (p->p_slptime > 1) 845 updatepri(p); 846 p->p_slptime = 0; 847 p->p_stat = SRUN; 848 if (p->p_flag & P_INMEM) { 849 setrunqueue(p); 850 maybe_resched(p); 851 break; 852 } else { 853 p->p_flag |= P_SWAPINREQ; 854 wakeup((caddr_t)&proc0); 855 } 856 /* END INLINE EXPANSION */ 857 } 858 } 859 } 860 mtx_exit(&sched_lock, MTX_SPIN); 861 splx(s); 862} 863 864/* 865 * The machine independent parts of mi_switch(). 866 * Must be called at splstatclock() or higher. 867 */ 868void 869mi_switch() 870{ 871 struct timeval new_switchtime; 872 register struct proc *p = curproc; /* XXX */ 873 register struct rlimit *rlim; 874 int giantreleased; 875 int x; 876 WITNESS_SAVE_DECL(Giant); 877 878 /* 879 * XXX this spl is almost unnecessary. It is partly to allow for 880 * sloppy callers that don't do it (issignal() via CURSIG() is the 881 * main offender). It is partly to work around a bug in the i386 882 * cpu_switch() (the ipl is not preserved). We ran for years 883 * without it. I think there was only a interrupt latency problem. 884 * The main caller, msleep(), does an splx() a couple of instructions 885 * after calling here. The buggy caller, issignal(), usually calls 886 * here at spl0() and sometimes returns at splhigh(). The process 887 * then runs for a little too long at splhigh(). The ipl gets fixed 888 * when the process returns to user mode (or earlier). 889 * 890 * It would probably be better to always call here at spl0(). Callers 891 * are prepared to give up control to another process, so they must 892 * be prepared to be interrupted. The clock stuff here may not 893 * actually need splstatclock(). 894 */ 895 x = splstatclock(); 896 897 CTR4(KTR_PROC, "mi_switch: old proc %p (pid %d, %s), schedlock %p", 898 p, p->p_pid, p->p_comm, (void *) sched_lock.mtx_lock); 899 mtx_enter(&sched_lock, MTX_SPIN | MTX_RLIKELY); 900 901 if (mtx_owned(&Giant)) 902 WITNESS_SAVE(&Giant, Giant); 903 for (giantreleased = 0; mtx_owned(&Giant); giantreleased++) 904 mtx_exit(&Giant, MTX_DEF | MTX_NOSWITCH); 905 906#ifdef SIMPLELOCK_DEBUG 907 if (p->p_simple_locks) 908 printf("sleep: holding simple lock\n"); 909#endif 910 /* 911 * Compute the amount of time during which the current 912 * process was running, and add that to its total so far. 913 */ 914 microuptime(&new_switchtime); 915 if (timevalcmp(&new_switchtime, &switchtime, <)) { 916 printf("microuptime() went backwards (%ld.%06ld -> %ld.%06ld)\n", 917 switchtime.tv_sec, switchtime.tv_usec, 918 new_switchtime.tv_sec, new_switchtime.tv_usec); 919 new_switchtime = switchtime; 920 } else { 921 p->p_runtime += (new_switchtime.tv_usec - switchtime.tv_usec) + 922 (new_switchtime.tv_sec - switchtime.tv_sec) * (int64_t)1000000; 923 } 924 925 /* 926 * Check if the process exceeds its cpu resource allocation. 927 * If over max, kill it. 928 * 929 * XXX drop sched_lock, pickup Giant 930 */ 931 if (p->p_stat != SZOMB && p->p_limit->p_cpulimit != RLIM_INFINITY && 932 p->p_runtime > p->p_limit->p_cpulimit) { 933 rlim = &p->p_rlimit[RLIMIT_CPU]; 934 if (p->p_runtime / (rlim_t)1000000 >= rlim->rlim_max) { 935 killproc(p, "exceeded maximum CPU limit"); 936 } else { 937 psignal(p, SIGXCPU); 938 if (rlim->rlim_cur < rlim->rlim_max) { 939 /* XXX: we should make a private copy */ 940 rlim->rlim_cur += 5; 941 } 942 } 943 } 944 945 /* 946 * Pick a new current process and record its start time. 947 */ 948 cnt.v_swtch++; 949 switchtime = new_switchtime; 950 CTR4(KTR_PROC, "mi_switch: old proc %p (pid %d, %s), schedlock %p", 951 p, p->p_pid, p->p_comm, (void *) sched_lock.mtx_lock); 952 cpu_switch(); 953 CTR4(KTR_PROC, "mi_switch: new proc %p (pid %d, %s), schedlock %p", 954 p, p->p_pid, p->p_comm, (void *) sched_lock.mtx_lock); 955 if (switchtime.tv_sec == 0) 956 microuptime(&switchtime); 957 switchticks = ticks; 958 mtx_exit(&sched_lock, MTX_SPIN); 959 while (giantreleased--) 960 mtx_enter(&Giant, MTX_DEF); 961 if (mtx_owned(&Giant)) 962 WITNESS_RESTORE(&Giant, Giant); 963 964 splx(x); 965} 966 967/* 968 * Change process state to be runnable, 969 * placing it on the run queue if it is in memory, 970 * and awakening the swapper if it isn't in memory. 971 */ 972void 973setrunnable(p) 974 register struct proc *p; 975{ 976 register int s; 977 978 s = splhigh(); 979 mtx_enter(&sched_lock, MTX_SPIN); 980 switch (p->p_stat) { 981 case 0: 982 case SRUN: 983 case SZOMB: 984 case SWAIT: 985 default: 986 panic("setrunnable"); 987 case SSTOP: 988 case SSLEEP: 989 unsleep(p); /* e.g. when sending signals */ 990 break; 991 992 case SIDL: 993 break; 994 } 995 p->p_stat = SRUN; 996 if (p->p_flag & P_INMEM) 997 setrunqueue(p); 998 splx(s); 999 if (p->p_slptime > 1) 1000 updatepri(p); 1001 p->p_slptime = 0; 1002 if ((p->p_flag & P_INMEM) == 0) { 1003 p->p_flag |= P_SWAPINREQ; 1004 wakeup((caddr_t)&proc0); 1005 } 1006 else 1007 maybe_resched(p); 1008 mtx_exit(&sched_lock, MTX_SPIN); 1009} 1010 1011/* 1012 * Compute the priority of a process when running in user mode. 1013 * Arrange to reschedule if the resulting priority is better 1014 * than that of the current process. 1015 */ 1016void 1017resetpriority(p) 1018 register struct proc *p; 1019{ 1020 register unsigned int newpriority; 1021 1022 mtx_enter(&sched_lock, MTX_SPIN); 1023 if (p->p_rtprio.type == RTP_PRIO_NORMAL) { 1024 newpriority = PUSER + p->p_estcpu / INVERSE_ESTCPU_WEIGHT + 1025 NICE_WEIGHT * (p->p_nice - PRIO_MIN); 1026 newpriority = min(newpriority, MAXPRI); 1027 p->p_usrpri = newpriority; 1028 } 1029 maybe_resched(p); 1030 mtx_exit(&sched_lock, MTX_SPIN); 1031} 1032 1033/* ARGSUSED */ 1034static void 1035sched_setup(dummy) 1036 void *dummy; 1037{ 1038 /* Kick off timeout driven events by calling first time. */ 1039 roundrobin(NULL); 1040 schedcpu(NULL); 1041} 1042 1043/* 1044 * We adjust the priority of the current process. The priority of 1045 * a process gets worse as it accumulates CPU time. The cpu usage 1046 * estimator (p_estcpu) is increased here. resetpriority() will 1047 * compute a different priority each time p_estcpu increases by 1048 * INVERSE_ESTCPU_WEIGHT 1049 * (until MAXPRI is reached). The cpu usage estimator ramps up 1050 * quite quickly when the process is running (linearly), and decays 1051 * away exponentially, at a rate which is proportionally slower when 1052 * the system is busy. The basic principle is that the system will 1053 * 90% forget that the process used a lot of CPU time in 5 * loadav 1054 * seconds. This causes the system to favor processes which haven't 1055 * run much recently, and to round-robin among other processes. 1056 */ 1057void 1058schedclock(p) 1059 struct proc *p; 1060{ 1061 1062 p->p_cpticks++; 1063 p->p_estcpu = ESTCPULIM(p->p_estcpu + 1); 1064 if ((p->p_estcpu % INVERSE_ESTCPU_WEIGHT) == 0) { 1065 resetpriority(p); 1066 if (p->p_priority >= PUSER) 1067 p->p_priority = p->p_usrpri; 1068 } 1069} 1070